Proceedings of Papers

Serbia, Niš, September 10 - 12, 2020

2020 55th International Scientific Conference on Information, Communication and
Energy Systems and Technologies (ICEST)

Serbia, Niš, September 10 - 12, 2020

Proceeding of Papers

Editors: Prof. Dr. Nebojša S. Dončov

Prof. Dr. Zoran Ž. Stanković
Dr. Biljana P. Stošić

Technical Support: Dr. Biljana P. Stošić

Published by: Institute of Electrical and Electronics Engineers (IEEE)

Faculty of Electronic Engineering, University of Niš, Serbia

Printing of this edition has been financially supported by

Ministry of Education, Science and Technological Development of Republic of Serbia

Number of copies printed: 50

IEEE Catalog Number: CFP20UWE-PRT

ISBN: 978-1-7281-7142-5 (IEEE)

Copyright and Reprint Permission: Abstracting is permitted with credit to the source. Libraries are
permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in
this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the
code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For reprint
or republication permission, email to IEEE Copyrights Manager at pubs-permissions@ieee.org. All rights
reserved. Copyright ©2020 by IEEE.
Dear Colleagues and ICEST Friends,

On behalf of the Technical Program Committee and Organizing Committee, it is a great pleasure for me to
announce that the 55th International Scientific Conference on Information, Communication and Energy
Systems and Technologies - ICEST 2020 conference was held on-line from 10th to 12th September,
2020. The Conference is jointly organized by the Faculty of Electronic Engineering, Niš, Serbia; the
Faculty of Telecommunications, Sofia, Bulgaria, and by the Faculty of Technical Sciences, Bitola,
Macedonia.

Due to COVID-19 issue ICEST 2020 Conference was moved from beginning of July to the beginning of
September hoping that the situation should resolve by then. After deep considerations and wishes of the
whole organization team for doing it live at least in some form, we found that the only responsible
behavior of organizers is to move ICEST 2020 to a fully virtual on-line conference.

The decision was based not only on safety issue, as we believed that we could handle it by using the
measures to reduce the risks to minimum, it is rather based on the uncertainty where the situation may
escalate and cause various and difficult problems for attendees.

As to the earlier ICEST Conferences, many authors from institutions all over the Europe submitted their
papers. This year, 60 papers have been accepted for on-line presentation.

I hope that all participants had taken opportunities not only to exchange their knowledge, experiences and
ideas but also to make contacts and establish further collaboration.

I hope that we will meet live again at the next ICEST Conference.

Prof. Dr. Nebojša Dončov,

ICEST 2020 Conference Chair
organized by

University of Niš
Faculty of Electronic Engineering
Serbia

Technical University of Sofia

Faculty of Telecommunications
Bulgaria

University "St. Kliment Ohridski", Bitola

Faculty of Technical Sciences
North Macedonia

under auspices of
• Ministry of Education, Science and Technological Development of the
Republic of Serbia

in cooperation with
• Serbia and Montenegro IEEE Section
• IEEE Bulgarian Section
• IEEE Republic of Macedonia Section
 TECHNICAL PROGRAM COMMITTEE

Chair:
Nebojša Dočov University of Niš, Serbia

Vice-chairs:
Kalin Dimitrov Technical University of Sofia, Bulgaria
Mitko Kostov University St.Kliment Ohridski, Bitola, Macedonia

Honorary Chairs:
Bratislav Milovanović Singidunum University, Niš, Serbia
Rumen Arnaudov Technical University of Sofia, Bulgaria
Cvetko Mitrovski St. Kliment Ohridski University, Bitola, Macedonia

Local Coordinators:
Dragan Mančić Dean of the Faculty of Electronic Engineering, University of Niš, Serbia
Ilia Iliev Dean of the Faculty of Telecommunications, Technical University of Sofia,
Bulgaria
Stojanche Nusev Dean of the Faculty of Technical Sciences, St. Kliment Ohridski University,
Bitola, Macedonia

Members:
Agata Manolova Technical University of Sofia, Bulgaria
Biljana Stošić University of Niš, Serbia
Boyanka Nikolova Technical University of Sofia, Bulgaria
Brian Barsky University of California, Berkeley, USA
Chrysostomos Stylios Technological Education Institute of Epiros, Greece
Đani Juričić Jožef Stefan Institute, Slovenia
D.D. Rasolomampionona Warsaw University of Technology, Poland
Dimitar Taskovski Ss. Cyril and Methodius University Skopje, Macedonia
Dimiter Alexandrov Lakehead University, Canada
Dragan Denić University of Niš, Serbia
Dragan Stojanović University of Niš, Serbia
Dušanka Bošković University of Sarajevo, Bosnia and Herzegovina
Eduard Ivanjko University of Zagreb, Croatia
Evelin Krmac University of Ljubljana, Slovenia
Evelina Pencheva Technical University of Sofia, Bulgaria
Fernando Alvarez University of Extremadura, Spain
Filiz Güneş Yildiz Technical University, Turkey
Ismail Altas Karadeniz Technical University, Turkey
Ivan Corbev Ss. Cyril and Methodius University Skopje, Macedonia
J.A.R.P. de Carvalho University Beira Interior, Portugal
Jair Minoro Abe Paulista University, Brazil
Jan Machac Czech Technical University, Prague, Czech Republic
Jugoslav Joković University of Niš, Serbia
Karel Raz University of West Bohemia, Czech Republic
Kazumi Nakamatsu University of Hyogo, Japan
Ljupčo Trpezanovski St. Kliment Ohridski University Bitola, Macedonia
Mariofanna Milanova University of Arkansas at Little Rock, USA
Metodija Atanasovski St. Kliment Ohridski University Bitola, Macedonia
Mikolaj Bartłomiejczyk Gdańsk University of Technology, Poland
Mikolaj Leszczuk AGH University of Science and Technology, Poland
Nikola Donato University of Messina, Italy
Pece Mitrevski St. Kliment Ohridski University Bitola, Macedonia
Pero Latkoski Ss. Cyril and Methodius University Skopje, Macedonia
Petar Popovski Aalborg University, Denmark
Peter Planinšič University of Maribor, Slovenia
Slavko Rupčić University of Osijek, Croatia
Stanimir Valtchev New University of Lisbon, Portugal
Stefan Stancescu Polytechnic University of Bucharest, Romania
Stevo Bozinovski South Carolina State University, USA
Szilvia Nagy University of Gyor, Hungary
Urban Burnik University of Ljubljana, Slovenia
Uwe Siart Technical University of Munich, Germany
Vincenzo Piuri University of Milan, Italy
Wojciech Krzysztofik Wroclaw University of Technology, Poland
Wojtek Bock University of Ottawa, Canada
Zdenka Babić University of Banja Luka, Bosnia and Herzegovina
Zoran Jovanović University of Niš, Serbia
Zoran Stanković University of Niš, Serbia

CONFERENCE ORGANIZING COMMITTEE

Chairs:
N. Dončov University of Niš, Serbia
K. Dimitrov Technical University of Sofia, Bulgaria
M. Kostov St. Kliment Ohridski University, Bitola, Macedonia

Local Coordinators:
D. Mančić University of Niš, Serbia
I. Iliev Technical University of Sofia, Bulgaria
S. Nusev St. Kliment Ohridski University, Bitola, Macedonia

Technical Editor:
Z. Stanković University of Niš, Serbia

Secretary:
B. Stošić University of Niš, Serbia

Members:
A. Atanasković University of Niš, Serbia
T. Dimitrijević University of Niš, Serbia
J. Joković University of Niš, Serbia
D. Mančić University of Niš, Serbia
N. Maleš-Ilić University of Niš, Serbia
Z. Marinković University of Niš, Serbia
V. Marković University of Niš, Serbia
 M. Milijić University of Niš, Serbia
O. Pronić-Rančić University of Niš, Serbia
B. Veselić University of Niš, Serbia
L. Stoimenov University of Niš, Serbia
J. Milošević University of Niš, Serbia

L. Laskov Technical University of Sofia, Bulgaria

A. Mihaylova Technical University of Sofia, Bulgaria
S. Antonov Technical University of Sofia, Bulgaria
V. Stoynov Technical University of Sofia, Bulgaria
D. Mihaylova Technical University of Sofia, Bulgaria
K. Valkov Technical University of Sofia, Bulgaria
N. Hristova Technical University of Sofia, Bulgaria
D. Kireva-Mihova Technical University of Sofia, Bulgaria
T. Brusev Technical University of Sofia, Bulgaria
M. Todorov Technical University of Sofia, Bulgaria
K. Raynova Technical University of Sofia, Bulgaria
K. Stoyanova Technical University of Sofia, Bulgaria
M. Stoyanova Technical University of Sofia, Bulgaria
M. Nedyalkova Technical University of Sofia, Bulgaria
V. Dolapchieva Technical University of Sofia, Bulgaria

M. Atanasovski St. Kliment Ohridski University, Bitola, Macedonia

B. Arapinoski St. Kliment Ohridski University, Bitola, Macedonia
G. Janevska St. Kliment Ohridski University, Bitola, Macedonia
V. Dancevska St. Kliment Ohridski University, Bitola, Macedonia
N. Mojsovska St. Kliment Ohridski University, Bitola, Macedonia
J. Pargovski St. Kliment Ohridski University, Bitola, Macedonia
M. Spirovski St. Kliment Ohridski University, Bitola, Macedonia
I. Andreevski St. Kliment Ohridski University, Bitola, Macedonia
D. Koltovska St. Kliment Ohridski University, Bitola, Macedonia

CONFERENCE SECRETARIAT
Address:
ICEST 2020 Conference
University of Niš
Faculty of Electronic Engineering
Aleksandra Medvedeva 14
18000 Niš, Serbia
phone: +381 18 529 105, 529 302, 529 303
fax: +381 18 588 399
e-mail: icest@elfak.ni.ac.rs

CONFERENCE INTERNET SITE

For further information, please visit the Conference Internet Site: http://www.icestconf.org
 LIST OF ICEST 2020 REVIEWERS

Dr. Aleksandar Atanasković Dr. Marija Milijić

University of Niš, Serbia University of Niš, Serbia

Prof. Dr. Blagoja Arapinoski Prof. Dr. Nenad Milošević

University "St. Kliment Ohridski", Bitola, University of Niš, Serbia
North Macedonia
Prof. Dr. Nebojša Mitrović
Prof. Dr. Metodija Atanasovski University of Niš, Serbia
University "St. Kliment Ohridski", Bitola,
North Macedonia Assist. Prof. Jelena Nikolić
University of Niš, Serbia
Prof. Dr. Milan Bebić
University of Belgrade, Serbia Prof. Dr. Tatjana Nikolić
University of Niš, Serbia
Prof. Dr. Dejan Ćirić
University of Niš, Serbia Nenad Petrović
University of Niš, Serbia
Dr. Tijana Dimitrijević
University of Niš, Serbia Assist. Prof. Bane Popadić
University of Novi Sad, Serbia
Prof. Dr. Nebojša Dončov
University of Niš, Serbia Prof. Dr. Nikola Rendevski
University "St. Kliment Ohridski", Bitola,
Dr. Danijel Djošić North Macedonia
University of Priština, Kosovska Mitrovica, Serbia
Prof. Dr. Blagoj Ristevski
Prof. Dr. Vassil Guliashki University "St. Kliment Ohridski", Bitola,
Bulgarian Academy of Sciences, Sofia, Bulgaria North Macedonia

Filip Filipović Assist. Prof. Ivanovitch Silva

University of Niš, Serbia Universidade Federal do Rio Grande do Norte,
Natal, Brazil
Prof. Dr. Siniša Ilić
University of Priština, Kosovska Mitrovica, Serbia Prof. Dr. Emiliano Sisinni
University of Brescia, Brescia, Italy
Prof. Dr. Aleksandar Jevremović
Singidunum University, Belgrade, Serbia Prof. Dr. Zoran Stanković
University of Niš, Serbia
Prof. Dr. Zoran Jovanović
University of Niš, Serbia Prof. Dr. Dragan Stojanović
University of Niš, Serbia
Prof. Dr. Mitko Kostov
University “St. Kliment Ohridski", Bitola, Dr. Biljana Stošić
North Macedonia University of Niš, Serbia

Prof. Dr. Zlatica Marinković Prof. Dr. Dragan Tasić

University of Niš, Serbia University of Niš, Serbia

Prof. Dr. Vera Marković Prof. Dr. Mladen Veinović

University of Niš, Serbia Singidunum University, Belgrade, Serbia

Prof. Dr. Dejan Milić Assist. Prof. Ivan Vulić

University of Niš, Serbia University of Defence, Belgrade, serbia
 TABLE OF CONTENTS
COMPUTER SYSTEMS AND INTERNET TECHNOLOGIES I

Regular papers:
CIT1.1 PIC Microcontroller Module in a Flexible Microprocessor Development System ................................... 3
Valentina Rankovska and Goran Goranov
Technical University of Gabrovo, Faculty of Electrical Engineering and Electronics, Bulgaria
CIT1.2 Efficient Discovery of Bent Function using Reed-Muller Subsets ............................................................ 7
Miloš Radmanović
University of Niš, Faculty of Electronic Engineering, Serbia
CIT1.3 Multilayer Long Short-Term Memory (LSTM) Neural Networks in Time Series Analysis...................... 11
Nemanja Malinović, Bratislav Predić and Miloš Roganović
University of Niš, Faculty of Electronic Engineering, Serbia
CIT1.4 Computer Visualization of Geometry of Polynomial Roots ..................................................................... 15
Ivan Petković and Petar Rajković
University of Niš, Faculty of Electronic Engineering, Serbia
CIT1.5 Continuous 2D Maps Based on Spherical Cube Datasets ....................................................................... 19
Aleksandar Dimitrijević and Peter Strobl*
University of Niš, Faculty of Electronic Engineering, Serbia
*European Commission, DG Joint Research Centre, Ispra, Italy
CIT1.6 Simulation Environment for Optimal Resource Planning during COVID-19 Crisis ............................... 23
Nenad Petrović
University of Niš, Faculty of Electronic Engineering, Serbia
CIT1.7 Tackling the COVID-19 Conspiracies: The Data-Driven Approach.......................................................... 27
Nenad Petrović
University of Niš, Faculty of Electronic Engineering, Serbia

COMPUTER SYSTEMS AND INTERNET TECHNOLOGIES II

Regular papers:
CIT2.1 Augmented and Virtual Reality Web Applications for Music Stage Performance ................................. 33
Nenad Petrović
University of Niš, Faculty of Electronic Engineering, Serbia
CIT2.2 The Application of Search Engine Optimization in Internet Marketing ................................................... 37
Aleksandar Andonov
University of Library Studies and Information Technologies, Faculty of Information Studies, Sofia, Bulgaria
CIT2.3 Improving the Concept of Medication Vending Machine in the Light of COVID-19 and other
Pandemics ..................................................................................................................................................... 42
Dragan Janković, Aleksandar Milenković and Andjelija Djordjević
University of Niš, Faculty of Electronic Engineering, Serbia
CIT2.4 Smart Health Services for Epidemic Control ............................................................................................. 46
Aldina Avdić, Ulfeta Marovac and Dragan Janković*
State University of Novi Pazar, Serbia
*University of Niš, Faculty of Electronic Engineering, Serbia
CIT2.5 Syntax Analysis of Serbian Language using Context-free Grammars.................................................... 50
Teodora Djordjević and Suzana Stojković
University of Niš, Faculty of Electronic Engineering, Serbia

I
CIT2.6 Signaling in 4G/5G with NB-IoT Support in 5G Option 3 .......................................................................... 54
Martina Janakieska, Pero Latkoski and Vladimir Atanasovski
Ss. Cyril and Methodius University, Faculty of Electrical Engineering and Information Technology, Skopje,
North Macedonia
CIT2.7 Supervision and Monitoring System of the Power Line Poles using IIoT Technology ......................... 58
Miloš Djordjević, Jana Vračar and Aleksandra Stojković
University of Niš, Faculty of Electronic Engineering, Serbia

ENGINEERING EDUCATION

Regular papers:
EE.1 Low Cost Laboratory Environment for the Use of Optical Methods for Transmission of Audio
Signals ........................................................................................................................................................... 65
Tsvetan Valkovski and Kalin Dimitrov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
EE.2 IMA-NET: Innovative e-Management Platform for Academic Network ................................................... 69
Galia Marinova and Krisel Tola
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
EE.3 Expert Tool for Filter Design Program Selection in Online-CADCOM Platform ..................................... 73
Ognyan Chikov and Galia Marinova
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
EE.4 Analysis of Pulse Code Modulation in MATLAB/Octave Environment ................................................... 77
Lyubomir Laskov, Veska Georgieva and Kalin Dimitrov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
EE.5 Model of Project-Based Learning Platform ................................................................................................ 81
Ava Chikurteva and Denis Chikurtev
Institute of Information and Communication Technologies, Bulgarian Academy of Sciences, Sofia, Bulgaria
EE.6 Practical Methods for Teaching of Solar Cell Characterization ............................................................... 85
Nikola Mitrović and Milan Stojanović
University of Niš, Faculty of Electronic Engineering, Serbia
EE.7 Servo-system Design for Arduino Due Based Laboratory Platform ....................................................... 89
Marija Filipović, Snežana Djordjević, Vladimir Mitić and Boban Veselić
University of Niš, Faculty of Electronic Engineering, Serbia
EE.8 Reception of Audio Signals Received from Different Angles by Optical Methods ................................ 93
Tsvetan Valkovski
Technical University of Sofia, Faculty of Telecommunications, Bulgaria

ENERGY SYSTEMS AND EFFICIENCY I

Regular papers:
ESE1.1 Application of Neural Networks for Short Term Load Forecasting in Power System of North
Macedonia ..................................................................................................................................................... 99
Goran Veljanovski, Metodija Atanasovski, Mitko Kostov and Pande Popovski
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ESE1.2 Fuzzy Multiple Linear Regression for the Short Term Load Forecast .................................................. 102
Andrija Petrušić and Aleksandar Janjić
University of Niš, Faculty of Electronic Engineering, Serbia
ESE1.3 Power Load Forecast for North Macedonia using Machine Learning................................................... 106
Pande Popovski, Mitko Kostov, Metodija Atanasovski and Goran Veljanovski
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia

II
ESE1.4 K-Nearest Neighbor Regression for Forecasting Electricity Demand .................................................. 110
Metodija Atanasovski, Mitko Kostov, Blagoja Arapinoski and Mile Spirovski
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ESE1.5 MATLAB Simulation of Grid-Connected Photovoltaic System .............................................................. 114
Pande Popovski, Natasa Mojsoska and Blagoja Stevanovski
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ESE1.6 Influence of Reduction Factors on High Voltage Cables on the Transfer of Potentials in the
Network ....................................................................................................................................................... 119
Nikolce Acevski, Angela Micevska and Ana Manivilovska
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ESE1.7 Evaluation of Reliability of Power Substations Considering the Uncertainty of the Distributed
Generation ................................................................................................................................................... 123
Atanas Iliev, Dimitar Dimitrov and Natasha Dimishkovska
Ss. Cyril and Methodius University, Faculty of Electrical Engineering and Information Technologies, Skopje,
North Macedonia
ESE1.8 Reduction of Energy and Power Losses in Distribution Network Using Energy Storage
Systems ....................................................................................................................................................... 127
Nikola Krstić
University of Niš, Faculty of Electronic Engineering, Serbia

ENERGY SYSTEMS AND EFFICIENCY II

Regular papers:
ESE2.1 Design and Regulation of the HF Inverter for Contactless Power Transfer Systems ......................... 133
Nikolay Madzharov and Desislav Iliev
Technical University of Gabrovo, Faculty of Electrical Engineering, Bulgaria
ESE2.2 Analysis of Electromagnetic Characteristics of Asynchronous Induction Machine ........................... 137
Igor Sterjovski, Blagoja Arapinoski, Metodija Atanasovski and Vesna Ceselkovska
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ESE2.3 Comparison of Methods for Control of Linear Induction Motors .......................................................... 141
Bojidar Markov and Diana Markova*
University of Food Technologies, Plovdiv, Bulgaria
*IT Academy, Plovdiv, Bulgaria
ESE2.4 The Cost of Indecision in Energy Efficiency. A Cost of Opportunity Analysis for an Industrial
Consumer .................................................................................................................................................... 145
Ioan-Sevastian Bitir Istrate, Cristian Gheorghiu and Miruna Castravete
POLITEHNICA University of Bucharest, Faculty of Power Engineering, Romania

SIGNAL PROCESSING I
Regular papers:
SP1.1 The Effects of Coefficient Quantization on New CIC FIR Filters ........................................................... 151
Biljana Stošić
University of Niš, Faculty of Electronic Engineering, Serbia
SP1.2 Improvement of Energy Efficiency of PAM Constellation by Applying Optimal Companding
Quantization in Constellation Designing ................................................................................................. 155
Slobodan Vlajkov, Aleksandra Jovanović and Zoran Perić
University of Niš, Faculty of Electronic Engineering, Serbia
SP1.3 The Influence of Compression Parameter μ on the Energy Efficiency of PAM Constellation
Based on μ-law Companding Quantization ............................................................................................. 159
Slobodan Vlajkov, Aleksandra Jovanović and Zoran Perić
University of Niš, Faculty of Electronic Engineering, Serbia

III
SP1.4 How Initialization is Related to Deep Neural Networks Generalization Capability:
Experimental Study .................................................................................................................................... 163
Ljubinka Sandjakoska and Frosina Stojanovska*
University of Information Science and Technology “St. Paul the Apostle”, Faculty of Computer Science and
Engineering, Ohrid, North Macedonia
*Ss. Cyril and Methodius University, Faculty of Computer Science and Engineering, Skopje, North
Macedonia
SP1.5 Using Neural Networks Application for the Font Recognition Task Solution ...................................... 167
Oleksii Bychkov, Kateryna Merkulova, George Dimitrov*, Yelyzaveta Zhabska, Iva Kostadinova*,
Pepa Petrova*, Pavel Petrov**, Iglika Getova* and Galina Panayotova*
Taras Shevchenko National University of Kyiv, Ukraine
*University of Library Studies and Information Technologies, Sofia, Bulgaria
**University of Economics, Varna, Bulgaria

SIGNAL PROCESSING II

Regular papers:
SP2.1 Adaptive 3D Interpolation of Images ........................................................................................................ 173
Rumen Mironov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
SP2.2 Noise Adaptive Wiener Filtering of Images.............................................................................................. 177
Lora Petkova and Ivo Draganov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
SP2.3 Along-Track and Cross-Track Noise Analysis of Minimal Curvature on Mars Orbiter Laser
Altimeter Data ............................................................................................................................................. 181
Nicole Christoff and Agata Manolova
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
SP2.4 Analysis for Watermark in Medical Image using Watermarking with Wavelet Transform and
DCT .............................................................................................................................................................. 185
Stoyan Kushlev and Rumen Mironov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria

MICROWAVES, ANTENNAS AND PROPAGATION

Regular papers:
MAP.1 On the Gas Sensing Properties of Microwave Transducers .................................................................. 191
Giovanni Gugliandolo, Mariangela Latino*, Giuseppe Campobello**, Zlatica Marinković***,
Giovanni Crupi**** and Nicola Donato**
Istituto Nazionale di Ricerca Metrologica - INRIM, and Politecnico di Torino, Italy
*University of Messina, MIFT Department, Italy
**University of Messina, Department of Engineering, Italy
***University of Niš, Faculty of Electronic Engineering, Serbia
****University of Messina, BIOMORF Department, Italy
MAP.2 Artificial Neural Network Modeling of Interdigital Capacitor Sensor for Oxygen Detection ............... 195
Zlatica Marinković, Giovanni Gugliandolo*, Mariangela Latino**, Giuseppe Campobello***,
Giovanni Crupi**** and Nicola Donato***
University of Niš, Faculty of Electronic Engineering, Serbia
*Istituto Nazionale di Ricerca Metrologica - INRIM, and Politecnico di Torino, Italy
** University of Messina, MIFT Department, Italy
***University of Messina, Department of Engineering, Italy
****University of Messina, BIOMORF Department, Italy

IV
MAP.3 Analysis of Dual Band PA Nonlinear Performances by Application of Digital Linearisation
Model ........................................................................................................................................................... 199
Aleksandar Atanasković, Aleksandra Djorić and Natasa Maleš-Ilić
University of Niš, Faculty of Electronic Engineering, Serbia
MAP.4 Design of Asymmetrical Slot Antenna Array in Corner Reflector .......................................................... 203
Marija Milijić and Branka Jokanović*
University of Niš, Faculty of Electronic Engineering, Serbia
*Institute of Physics, University of Belgrade and the Academy of Engineering Sciences of Serbia, Serbia
MAP.5 Radar Antenna Array Designing for Unmanned Aerial Vehicles (UAV)................................................. 207
Ivaylo Nachev and Peter Petkov
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
MAP.6 Design of well-matched Microwave Slot Antenna on a Flat Metal Grounded Plate using Neural
Model ........................................................................................................................................................... 211
Zoran Stanković, Nebojša Dončov, Biljana Stošić, Maja Sarevska*, Ivan Milovanović**
University of Niš, Faculty of Electronic Engineering, Serbia
*AUE-FON University Skopje, North Macedonia
**Singidunum University, Educational Center Niš, Serbia
MAP.7 An Automatic System for Antenna Radiation Pattern Measurement .................................................... 216
Ilia Iliev and Ivaylo Nachev
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
MAP.8 Inverse Reading of ITU-R P.1546 Propagation Curves for the Broadcast Transmitter EM Field
Level ............................................................................................................................................................ 220
Miloš Radojković, Olivera Pronić Rančić* and Zlatica Marinković*
Energoprojekt Entel, Belgrade, Serbia
*University of Niš, Faculty of Electronic Engineering, Serbia
MAP.9 GPU-Enabled Simulation of a Combined Macrodiversity System in Gamma-Shadowed
Nakagami-m Fading Channel .................................................................................................................... 224
Nenad Petrović, Dejan Milić, Suad Suljović, Samir Koničanin and Selena Vasić*
University of Niš, Faculty of Electronic Engineering, Serbia
* Metropolitan University, Faculty of Information Technology, Belgrade, Serbia

MEASUREMENTS, CONTROL SYSTEMS AND ICT

Regular papers:
ECI.1 Measurement and Analysis of PM Particles Emitted by Automotive Brakes ....................................... 231
Emil Iontchev, Rosen Miletiev*, Rumen Yordanov* and Ilian Damyanov*
Higher School of Transport, Bulgaria
*Technical University of Sofia, Faculty of Telecommunications, Bulgaria
ECI.2 Investigation of the Method of RMS Measurement Based on Moving Averaging ............................... 235
Anna Kostina, Plamen Tzvetkov* and Andrey Serov
National Research University “Moscow Power Engineering Institute”, Russia
*Technical University of Sofia, Faculty of Automatics, Bulgaria
ECI.3 Monitoring of Air Pollution in Skopje in order to Determine Possible Causes .................................... 239
Vladimir Mijakovski and Monika Lutovska*
University "St. Kliment Ohridski", Faculty of Technical Sciences, Bitola, North Macedonia
*"Mother Teresa" University, Faculty of Technical Sciences, Skopje, North Macedonia
ECI.4 Load Disturbance Rejection in Hydro Turbine Control Based on Wiener Filtering ............................. 244
Teofana Puleva and Elena Haralanova
Technical University of Sofia, Faculty of Automatics, Bulgaria
ECI.5 Multi-Objective LMI Based Robust State-Feedback Design of a Hydro Generator Speed
Controller .................................................................................................................................................... 248
Andrey Yonchev and Teofana Puleva
Technical University of Sofia, Faculty of Automatics, Bulgaria

V
ECI.6 Impact of Microclimate Changes and Noise to Workers as a Part of an Implemented Health
and Safety System ...................................................................................................................................... 252
Ivo Kuzmanov and Roberto Pašić
St. Kliment Ohridski University, Faculty of Technical Sciences, Bitola, North Macedonia
ECI.7 Hardware Implementation and Comparison of CRA and TRA when Trying to Recover the
AES-128 Key ................................................................................................................................................ 256
Ivaylo Vladimirov, Desislava Nikolova and Zornitsa Terneva
Technical University of Sofia, Faculty of Telecommunications, Bulgaria
ECI.8 Software Implementation of CRA and TRA to Recover the AES-128 Key using Side-Channel
Signals with Python3 ................................................................................................................................. 259
Desislava Nikolova, Ivaylo Vladimirov and Zornitsa Terneva
Technical University of Sofia, Faculty of Telecommunications, Bulgaria

VI
 COMPUTER SYSTEMS
AND INTERNET TECHNOLOGIES I
PIC Microcontroller Module in a Flexible Microprocessor
Development System
Valentina V. Rankovska1 and Goran D. Goranov2
Abstract – A microcontroller module with a PIC18
microcontroller has been designed and implemented. It contains
all the components in addition to the microcontroller, necessary
for its operation as a microprocessor core in embedded systems
for real applications.
The microcontroller module is the smaller size and different
microcontroller compared to the widely used Arduino Uno.
Sample algorithms and software has been designed to test the
module within the microprocessor system for which it was
primary intended in order to prove its correct operation.
Keywords – Microprocessor development system,
Microcontroller module, PIC18, Development tools.
I. INTRODUCTION
Nowadays microcontrollers and embedded systems are
almost everywhere in our everyday life. Modern
microcontrollers are powerful, versatile, low-cost, reliable,
low-consumption tools for solving various tasks in the field of
automation, control, entertainment etc.
That is the reason to increase the number of the courses
related to microcontrollers in the new Bachelor degree
curriculum of "Industrial and Automotive Electronics" at the
Technical University of Gabrovo: Microprocessor Circuits,
Industrial and Automotive Interfaces, Embedded Systems,
Intelligent Electronic Systems. Furthermore various types of
microcontrollers according to their features, efficiency,
parameters, etc. are studied. This necessitates the use of a
variety of hardware and software development tools, which in
turn complicates and makes it more expensive the learning
process. That is why an approach is preferred applying unified
and flexible tools that can be easily reconfigured according to
the specific application.
As a part of that process, a development microprocessor
system for educational purposes was designed [1]. After that,
it was completed with other tools, among which the
demonstration module, presented in [2].
Section II describes briefly the above two boards. In
Section III the designed microprocessor module is presented.
It contains a PIC18 microcontroller and is intended to be used
with the development system. Some test results are also given.
1Valentina V. Rankovska is with Technical University of
Gabrovo, Faculty of Electrical Engineering and Electronics, Hadzhi
Dimitar 4, 5300 Gabrovo, Bulgaria, E-mail: rankovska@tugab.bg
2Goran D. Goranov is with Technical University of Gabrovo,
Faculty of Electrical Engineering and Electronics, Hadzhi Dimitar 4,
5300 Gabrovo, Bulgaria, E-mail: g_goranov@tugab.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
II. MICROPROCESSOR DEVELOPMENT SYSTEM
The main purpose in designing the microprocessor
development system has been for it to be flexible and
relatively versatile, because various microcontrollers with
different characteristics, functionalities and microprocessor
cores are studied: 8- and 32-bit; with simpler or advanced
capabilities; with PIC, AVR, ARM processor cores.
That is why it provides the possibility various replaceable
microcontroller modules to be used, which are with a
minimum set of peripherals (preferably open source) and one
and the same form factor (Fig. 1). That is possible because,
due to the rapid spread of the Arduino platform, presently
there is a large variety of development boards compatible with
it. This minimalist approach is even adopted by companies
that manufacture microcontrollers and development tools
based on ARM cores.
Fig. 1. Block diagram of the microprocessor development system
uCM - Microcontroller Module; DM - Demonstration Module; M -
Motors; LM - LED Matrix; К - Keypad; PS - Power Supply
The other main part of the development system, in addition
to the microcontroller module, is the multifunctional
demonstration module, which includes a set of peripheral
devices (input/ output, actuators, etc.), often applied in
embedded systems: buttons, LEDs, seven-segment and liquid
crystal displays, drivers for low power and stepper motors,
relays, etc. The block diagram of the demonstration module is
shown in Fig. 2.
III. THE DESIGNED PIC18 MICROCONTROLLER
MODULE
Various development boards can be mounted on the
development system, for instance Arduino Uno (with 8-bit
microcontroller with AVR processor core), PIC32-Pinguino-
OTG (with 32-bit microcontroller with MIPS processor core),
Nucleo L476RGT6 (with 32-bit microcontroller with ARM
processor core), etc.
3

Fig. 2. Block diagram of the demonstration module (DM in Fig. 1)
PS - Power Supply; B - Breadboard; - D&I - Display and
Indication; ID - Input Devices; D - Drivers; R - Relays;
EC - Expansion Connectors
Much less common are similar development boards with
PIC microcontrollers, mostly 32-bit ones. That is why a
microcontroller board, based on 8-bit PIC18 microcontroller
is designed. Figs. 3 and 4 show respectively its electrical
circuit and a photograph.
Fig. 3. Electrical circuit of PIC18 microcontroller module
Fig. 4. Photograph of PIC18 microcontroller module
The suggested module represents a microcontrollercore
with the minimum necessary components to operate.
The main component is PIC18F25K50. It is an 8-bit
microcontroller with advanced functionality and improved
architecture (in comparison to PIC16 and older PIC18). It is
suitable for educational and development purposes because of
a variety of peripheral blocks. In addition to the “traditional”
and often used modules, like analog to digital converter,
timers, capture/comare/PWM modules, serial interface blocks,
etc., it has also a Digital-to-Analog Converter (DAC), USB
module, SR latch, a set of internal oscillators, etc.
The main features of PIC18(L)F2X/45K50 microcontrollers
are summarized in Table I [3].
4
Table I
MAIN FEATURES OF PIC18(L)F2X/45K50 MICROCONTROLLERS
A crystal oscillator 20MHz supplies an external clock,
although the microcontroller can operate with an internal
oscillator too (31 kHz to 16 MHz).
A reset button is connected to the MCLR pin, the latter also
used for programming voltage.
As the board is designed for educational, not hobby
purposes, it is not provided for it to work with simple
integrated environments such as Arduino IDE, Pinguino IDE,
etc. That is why there is not a bootloader in the
microcontroller and programming could be made only using
common in-circuit debugger/programmer, such as PICkit3/4
via a special single row female header, connected to the
corresponding pins of the microcontroller.
V-USB is the internal power voltage regulator output,
supplying a positive voltage for the USB module. It should be
connected to ground through a ceramic capacitor with a low
value and operating voltage not less than 16V.
The module uses stable 5V power supply, provided either
via dedicated pins, or through the USB-micro B connector.
The latter option is more suitable for the lab classes
(depending on the overall load): we could supply the system
from a PC, a battery, an adapter, provided they supply at least
500mA DC current.
All the general purpose pins of the microcontroller are
connected to two single row female headers, thus making
them accessible by the external parts of the system.
IV. TESTING THE OPERATION OF THE
MICROCONTROLLER MODULE
Sample algorithms and programs have been designed in
order to test the operation of the microcontroller unit. They
will be used also in laboratory classes by the students to study
the operation of the internal microcontroller peripheral
modules and their implementation in embedded systems.
The Microchip’s integrated development environment
MPLAB X IDE v5.20 and the free version of the C compiler
ХС8 v2.05 have been used for the design of the software.
The programming has been made with the in-circuit
debugger PICkit4 [4].
The Saleae logic analyzer, operating with free software, has
been used to test the sample programs in real time (Fig. 5). It
is 8-channel, with sample frequency of up to 24 MHz, and
connects to a PC via USB [5].
Some of the programs use the CCP modules in various
modes of operation; the ADC; the serial interface blocks, etc.
For instance, Fig. 6 presents the format of the sent ASCII
codes of the letters of the word “Hello”, by EUSART module
in asynchronous mode. The algorithm is simple, and interrupts
are used to receive and transmit data.
 Fig. 5. Logic Analyzer Saleae

In other example, the voltage on a desired analog pin of the

microcontroller is measured and the value is converted and
sent to a 128x64 Dot Matrix OLED display (Organic Light-
Emitting Diode) [6]. The microcontroller communicates with
the displayvia I2C interface, implemented by the MSSP
(Master Synchronous Serial Port) module.
The overall algorithm includes the following user functions:
• The main function, which algorithm is shown in Fig.7;
• ADC_Init (Fig. 8), ADC_Start, ADC_AutoTrigger,
ADC_Interrupt (Fig. 9), which configure the ADC,
initiate the conversion, define an automatic conversion,
and read the converted value in an interrupt;
• OLED_Init, OLED_Command, OLED_Clear,
OLED_Send (Fig. 10): configure the display, send a
command to it, clean the display and send data/
command to it;
• I2C_Init, I2C_Start, I2C_Stop, I2C_Transmit,
I2C_Wait: configure I2C module, start and stop its
operation, send data, wait to complete sending of the
data;
• Interrupt_Enable and __interrupt(low_priority)
IRQ_Handler: first enables the interrupt and the second
is the handler for all the types of interrupts.
Figs. 11 and 12 illustrate a part of the communication Fig. 7. Main function algorithm Fig. 8. ADC_Init function
between the microcontroller and OLED display. The green algorithm
dots in Fig. 11 denote the “Start” conditions, initiating the
communication, and the red dot denotes the “Stop” condition.
A larger scale of the data framing is shown in Fig. 12. ACKNOWLEDGEMENT
The present paper is supported by the Science Research
V. CONCLUSION Fund at the Ministry of Education and Science under contract
№2002E/2020.
A microcontroller module has been designed and
implemented, which, in addition to the microcontroller itself,
contains also all the components necessary for its operation as REFERENCES
a microprocessor core in an embedded system.
The microcontroller module is the smaller size compared to [1] V. Rankovska, S. Rankovski, „Microprocessor Development
the widely used Arduino Uno. However, it is based on widely System Applying a Multifunctional Demonstration Board“,
spread PIC18 microcontroller. ICEST 2019, Conference Proceedings, Ohrid, North Macedonia,
pp. 414-417, 2019.
It is primary intended to be used as a part of the [2] V. Rankovska, S. Rankovski, „Multi-functional Demo Module
microprocessor development system, presented in [2]. The for Microprocessor Development Systems for Educational
proper operation of the module is tested using sample Purposes“, ET2018, Conference Proceedings, Sozopol,
algorithms and software in C programming language and Bulgaria, pp. 55-58, 2018.
software and hardware development tools, applied in teaching [3] PIC18(L)F2X/45K50. 28/40/44-Pin, Low-Power, High-
Microprocessor circuits for Bachelor degree students in Performance Microcontrollers with XLP Technology. Microchp
Industrial and automotive electronics. Technology Inc., 2014.
[4] https://www.microchip.com/developmenttools/ProductDetails/P
G164140
[5] https://www.saleae.com/
[6] 128x64 Dot Matrix OLED/PLED Segment/Common Driver
with Controller. Solomon Systech Limited, 2008.

5
Fig. 9. Algorithm of the function, servicing ADC Interrupts

Fig. 10. Algorithm of the function, sending the data to the OLED

Fig. 6. Waveforms of ЕUSART operation in asynchronous mode

Fig. 11. Waveforms of MSSP operation in I2C mode at communication between the microcontroller and the OLED display

Fig. 12. A larger scale of the waveforms, shown in Fig. 11, showing the parameters of the clock

6
 Efficient Discovery of Bent Function Using
Reed-Muller Subsets
Miloš Radmanović
Abstract – Bent functions are functions that have the biggest
distance from all linear binary functions. Attacks against
cryptosystems can be defeated when these systems use bent
functions. Generalisation of bent functions is still unknown and
when they have many variables, it is very difficult to find them.
Therefore, the construction of bent functions is done by using
discovery of them within a subset of functions. Thus, this paper
describes a method for creating groups of functions in Reed-
Muller (RM) domain within which efficient detection of bent
functions is performed. As follows, the search space for discovery
is limited, and it can be defined for various subsets of Boolean
functions. Experimental results confirm that the time required
for discovery of bent function when using RM subsets are
smaller than when using general random discovery.
Keywords – Cryptography, Binary function, Bent construction,
Reed-Muller subsets.
I. INTRODUCTION
Bent functions are binary functions having the biggest
possible Hamming distance from the set of linear functions.
Thus, bent functions have the highest possible nonlinearity.
These functions are the core of many cryptographic systems
thanks to their ability to provide secure communication. They
have applications in a variety of such systems, like hash
functions and block and stream chippers. Therefore, the study
of bent functions for cryptography is essential for secure
communication in the future.
The discovery of cryptographically useful bent functions is a
difficult task. As the number of variables in a binary function
increases, bent functions become extremely rare in the set of
all possible binary functions. Bent functions are possible only
when the number of variables is even. Furthermore, the precise
generalization of bent functions is still unknown. Also, the
formal method for their construction does not exist. Thus, in
the last years, many researches are developed a lot of methods
for construction of bent functions using algebraic [1],
combinatorial [2], and heuristic techniques [3]. These methods
can be complex, computationally intensive, difficult to
implement, and also do not always produce a sufficient variety
of bent functions.
The absolute values of all Walsh coefficients of bent
functions are equal and it is the most common characterization
for them. All coefficients have the absolute value 2 n / 2 , where
n is number of binary function variables. Consequently, bent
function testing requires computation and comparison of all
The author is with the University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
milos.radmanovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
2 n Walsh coefficients. Even for small number of function
variables, this computation requires a lot of processing time
[4]. Therefore, different techniques for discovery of bent
functions become very popular. They are mostly based on the
reduction of the discovery time [5]. In the present paper I study
bent functions using a representation by Reed-Muller (RM)
spectral coefficients.
Reed-Muller spectral transforms have many applications in
synthesis, verification, and testing of logic circuits, processing
techniques and signal encoding [6], and in many other areas
[7]. For practical applications, the computation of RM spectral
transform needs to be efficient. The fast Reed-Muller spectral
transform algorithm reduces computational complexity from
O(N2) to O(Nlog2N). This algorithm is derived from Fast
Fourier Transform (FFT) algorithm and it is very effective tool
in various applications. This algorithm has also been
developed for parallel computations with shared memory and
with distributed memory.
Using restrictions of bent function algebraic degree, some
authors [8], and [9] proposed bent functions discovery by
using the RM spectrum and related normal form which is in
related literature also called the algebraic normal form [10].
Previous research on discovering of these functions in RM
domain is discussed for example in [11], [12]. The efficiency
of a multicore CPU parallel algorithm for random bent
function discovery in RM domain is presented in [13]. Parallel
computing on manycore GPUs also enables parallel
processing on commodity hardware. GPU parallel algorithm
for bent function discovery in RM domain is presented in
[14].
Note that, for functions with 4 variables, number of
possible non-zero coefficients in RM spectra varies from 2 to
11, and there are 896 bent functions., Also, for binary
functions with 4 variables, the search space length is 55
functions for RM spectra with 2 non-zero coefficients, and
462 for RM spectra with 6 non-zero coefficients. For binary
functions with 6 variables, the search space length is
538,257,874,440 for RM spectra with 21 non-zero
coefficients. (the number of the possible non-zero coeffi-
cients). Computation time is often a limiting parameter for
practical applications of discovering of bent functions in RM
domain.
Therefore, discovery of bent functions certainly needs
reduction of the search space by using some restrictions. Thus,
it is presented an efficient technique for discovery of bent
functions using RM subsets. RM subsets are based on the
restrictions of the number and order of RM coefficients for
bent functions. There are vertical, horizonal and grid subsets
of binary functions depending on different restrictions of bent
functions in RM domain, [15]. As follows the search space for
7
discovery is limited and it is different for various predefined
subsets of Boolean functions.
Experimental results confirm that the described technique
for efficient discovery of bent functions using RM subsets
shows better computational performance over performing the
random discovering. This method is investigated using
different subsets in RM domain. The number of computations
exponentially increases in relation to the function size.
Accordingly, experiments were done with benchmark
functions from 6 to 10 variables.
II. BACKGROUND THEORY
If a function f and its Reed-Muller spectrum S f , in matrix
notation, are represented by vectors F = [ f 2 n −1 ,, f1 , f 0, ]T
and S f = [ s 2n −1 ,, s1 , s0 ]T respectively, the positive polarity
RM transform is defined by the positive polarity RM matrix
RM (n) [16]:
S f = RM (n) F ,
f ( x0 , x1 ,, xn ) = X (n) S f ,
where
n
X (n) = ⊗[ 1 xi ]
i =1
and
n
RM (n) = ⊗ RM (1) ,
i=1
1 0
RM (1) =  ,
1 1
where ⊗ is the Kronecker product, addition and multiplication
are in modulo 2, RM (1) represents the basic positive RM
transform matrix. Note that RM (1) is self-inverse in nature.
The FFT computing algorithm can be used to compute the
coefficients in RM spectrum. This algorithm is composed of
the “butterfly” operations which are repeated and have
structure derived from the basic transform matrices [16].
Similar, in the Hadamard ordering, the Walsh spectral
transform matrix of order n, denoted by W (n) , is defined as
[16]:
n 1 1 
W ( n) = ⊗ W (1), W (1) =  .
i =1
1 − 1
The set of bent functions is very small compared to the set of
all Boolean functions. As noticed above, bent functions exist
for an even number of variables, and the exact number of bent
functions with -variables is known only for small n. In
literature the number of bent functions was computed using
super-computers only for . The general number of bent
functions is an open problem and it is known the lower and
upper bounds in respect to [3].
A n-variable Boolean function in (1,−1) encoding is called
bent if all elements of the Walsh spectrum vector S f ,W have
the equal absolute value 2 n / 2 [16].
For example, for a bent function f of 4 variables, given by
truth vector:
F = [1,1,0,0,1,1,1,1,1,0,0,1,0,1,0,1] .
T
(7)
The Walsh spectrum, in matrix notation, with (1,-1) encoding
of function truth vector, can be calculated as S f ,W = W (4) F ,
where W ( 4) is Walsh transform matrix of order 4, and result
is:
S f ,W = [- 4,4,-4,-4,4,-4,-4,-4,-4,-4,-4,4,4,4,-4,4] ,
T
(8)
It should be noticed that absolute values of all Walsh
coefficients is equal to 4.
RM form of previous example of bent function is:
f = 1 ⊕ x3 ⊕ x2 x3 ⊕ x`1 x4 ⊕ x`1 x2 . (9)
(1)
Note that the algebraic degree is less or equal to value 2.
(2) The orders of RM coefficients are related with the positions of
these coefficients in the RM spectrum [5]. This characteristic
is widely used in discovery of bent function in RM domain.
There are various subsets of bent functions whose
(3) formation is based on function characteristics and additional
conditions. For example, one of the most important subset of
bent functions is subset of quadratic bent functions. All
quadratic bent functions are known and they can be
(4) constructed by using the affine transformation of the function
variables [6]. The general drawback of this subset is that this
subset contains very small number compared to the total
number of bent functions. Many subsets of bent functions
(5)
have this drawback. Therefore, in this paper it expermented
with subsets in RM domain.
In RM domain, set of all bent functions can be splitted
relative to 3 different criteria using different properties of
RM-spectra [17]. The number of non-zero RM coefficients
defines vertical RM subset. The order of nonlinearity defined
as maximaum number of variables in a product term
corresponding to a non-zero RM coefficient and it defines the
horizontal RM subset. Both, the number of non-zero RM
coefficients and the order of nonliearity defines grid RM
subsets [15]. The main objective is to investigate the
discovery of bent functions in each of these subsets.
A subset V(k) of bent functions is called Vertical k-subset
if all bent functions in this subset has k RM coefficients not
equal to 0 [15]. The number of variables defines possible V
(6)
subsets [15] for bent functions.
For example, for given 4-variables bent functions:
f1 = 1 ⊕ x1 x2 ⊕ x3 x4 , (10)
and
f 2 = x4 ⊕ x1 x2 ⊕ x1 x4 , (11)
have same number of non-zero Reed-Muller coefficients and
they are in the same subset (3).
8
 A subset H(kmin, kmax) of bent functions is called the
Horizontal (kmin, kmax)-subset if all bent functions in this subset
has order of RM coefficients not equal to 0 between values
kmin and kmax [15]. The values kmin and kmax need to be less than
possible orders of non-zero RM spectrum coefficients for bent
functions. The number of variables defines possible H subsets
for bent functions [15].
For example, for given 4-variables bent functions:
f1 = x1 x2 ⊕ x2 x4 ⊕ x1 x4 , (12)
and
f 2 = x1 x3 ⊕ x2 x4 , (13)
have only coefficients with order 2 and consequently they
belong to the same subset (22).
A subset () of bent functions is called the
Grid()-subset if all bent functions in this subset
has RM-coefficients not equal to 0 and has order of RM
coefficients not equal to 0 between values kmin and kmax [15].
The subsets of bent functions are the intersection of and
subsets [15]. Consequentially, the number of variables
defines possible G subsets for bent functions [15].
For example, for given 4-variables bent functions:
f1 = x1 x2 ⊕ x3 x4 , (14)
and
f 2 = x1 x3 ⊕ x2 x4 , (15)
have same number of coefficients and all with order 2 and
consequently they belong to the same subset (2, 2, 2).
III. DISCOVERY OF BENT FUNCTIONS USING
REED-MULLER SUBSETS
The algorithm for the discovery of bent functions using RM
subsets is given as Algorithm 1.
Algorithm 1
1: Set function size
2: Set the vertical subset parameters (min. and max number
of the non-zero RM spectrum coefficients).
3: Set the horiznatal subsets (min and max order of the non-
zero RM spectrum coefficients).
4: Set randomly the number of the non-zero RM
coefficients using vertical subsets restrictions.
5: Set randomly the positions of the non-zero RM
coefficients using horizontal subsets restrictions.
6: Create binary function.
7: Bent testing of all Walsh coefficients of the created
function, if not passed go to the step 2.
8: Return bent function found.
The algorithm for the discovery of bent functions using RM
subsets takes the number of function variables n (function
size) as its input. Optionally it takes as its input the min and
max parameter as a range of number of non-zero RM
spectrum coefficients. This max parameter should be less then
maximal number of coefficients that is allowed for bent
9
functions. Optionally it also takes the min and max parameter
as range for order of the non-zero RM spectrum coefficients.
This max parameter should be less then maximal order of
coefficients that is allowed for bent functions. The orders of
RM coefficients restrict the their positions in RM spectrum
[15].
These restrictions ensure discovery possibility and
conseqetily reduce the possible search space in RM domain
for bent function random discovery.
The bent testing of Walsh coefficient includes checking if
the coefficient has the absolute value 2n/2, where n is function
size. Computation of Walsh coefficients in this algorithm use
fast Walsh spectral transform.
Note that, in the Boolean domain, bent testing is much
faster than in the RM domain since it does not include
creation of binary function from RM spectrum (step 6). The
method for computation of binary truth vector from RM
spectrum is also use the fast Reed-Muller transform.
IV. EXPERIMENTAL RESULTS
For experimental analysis, it is developed an
implementation using C++ programming language of method
for the random discovery of bent functions in RM domain.
Also it is developed an implementation of the discovery of
bent functions using RM subsets. These experiments are done
on standard PC platform (CPU: 3.6 GHz, RAM: 12 GB).
Table 1 shows computation performance of algorithm for
random discovery of bent function using the referent C++
implementation and the C++ implementation using RM
subsets. The presented computational times are given in
milliseconds. It is compared the computation time needed for
discovery of one bent function for entered number and/or
order of non-zero coefficients in RM spectrum, in respect to
discovery of one bent function for random number and order
of non-zero coefficients.
Comparison of these numbers is motivated by decreasing
computation time when discovery use RM subsets. When the
number of non-zero coefficients is relatively small,
discovering of bent function is faster The proposed algorithm
is implemented and experimentally tested for discovering bent
function of 6, 8 and 10 variables in the RM domain. It is
compared with our previous algorithm for random discovering
of bent functions in RM domain [12], [13], [14]. Note that, for
the Boolean functions with large number of variables,
experiments were not presented, because the computation
time limit of 30 minutes is exceeded.
Bent functions need to be non-deterministic if they are used
for cryptographic applications. In this case, when the number
and/or order are predefined, bent functions are not completely
non-deterministic. Therefore, in practical applications, there is
need for using various RM subsets with different number of
non-zero coefficients, as well as different orders.
It should be noticed that the computation times for
discovering bent functions using RM subsets of 6, 8, and 10
variables for all 5 different tests are always smaller than using
the random discovery. Note that for the larger binary
functions, relative difference between computation times
between tests is smaller in percentage.
 TABLE I ACKNOWLEDGEMENTS
COMPUTATION TIME OF ALGORITHM FOR DISCOVERY OF BENT
FUNCTIONS USING REED-MULLER SUBSETS I would like to acknowledge to the Ministry of Education
and Science of the Republic of Serbia for funded the research
Num. of Computation time [ms] work in this paper, as a part of project III44006 (2011-2020).
function
exhaustive V(6) H(3,6) G(6,3,6)
variables
6 1.532 1.190 982 782 REFERENCES
6 1.512 1.244 876 791
[1] N. Y. Yu, G. Gong, “Constructions of Quadratic Bent Functions
6 1.526 1.116 855 804
in Polynomial Forms”, IEEE Trans. on Information Theory, vol.
6 1.529 1.321 943 756 52, no. 7, pp. 3291-3299, 2006.
6 1.538 1.181 911 772 [2] J. Seberry, X. Zhang, “Constructions of bent functions from two
8 1437 1015 739 648 known bent functions”, Australasian Journal of Combinatorics.
vol. 9, pp. 21-34, 1994.
8 1441 1047 832 633 [3] A. Grocholewska-Czurylo, J. Stoklosa, “Generating Bent
8 1439 986 815 696 Functions”, Proc. Advanced Computer Systems Eighth
8 1443 1211 794 642 International Conference, ACS 2001, pp. 361-370, Springer US,
Mielno, Poland, 2002.
8 1436 973 716 623
[4] P. Langevin, G. Leander, “Counting all Bent Functions in
10 25516 21233 19365 18745 Dimension Eight 99270589265934370305785861242880”,
10 25429 22648 19583 18573 Designs, Codes and Cryptography vol. 59, pp. 193-201, 2011.
10 25678 22751 19976 18459 [5] N. Tokareva, Bent Functions, Results and Applications to
Cryptography, Academic Press, 2015.
10 25577 21177 19023 18011
[6] M. A. Thornton, R. Drechsler, and D. M. Miller, Spectral
10 25433 21375 19856 19683 Techniques in VLSI CAD, Springer, 2001.
[7] T. Sasao, J. Butler, and M. Thornton, Progress in Applications
of Boolean Functions, Morgan and Claypool Publishers, 2010.
V. CONCLUSION [8] S. Mesnager, Bent Functions, Fundamentals and Results,
Springer International Publishing, 2016.
This paper proposes a technique for efficient discovery of [9] J. Climent, F. Garcia, and V. Requena, “On the Iterative
Boolean bent functions in Reed-Muller domain. The Construction of Bent Functions”, in: N. Mastorakis, A. Cecchi,
(eds.), Proc. 5th WSEAS Int. Conf. on Inf. Security and Privacy,
algorithm is based on the random discovery of RM spectrum
pp. 15-18. Stevens Point, Wisconsin, United States, 2006.
of bent function by using RM subsets. The technique is mainly [10] O. Logachev, A. Salnikov, and V. Yashchenko, Boolean
focused on the reduction of the discovery search time. RM Functions in Coding Theory and Cryptography, American
subsets are based on the restrictions of the number and/or Mathematical Society, 2012.
order of RM coefficients for bent functions. Using these [11] M. Radmanović, R. Stanković, “Discovery of Bent Functions in
restrictions in RM domain, there are vertical, horizonal and Fixed Polarity Reed-Muller Domain Using Dual Polarity
grid subsets of binary functions. Restrictions of number and/or Route”, Proc. 13th Int. Workshop on Boolean Problems, pp.43-
order of RM coefficients influence on the size of the search 59, Bremen, Germany, 2018.
space for bent functions in RM domain. [12] M. Radmanović, “Efficient Bentness Testing for Cryptographic
Function using Statistical Analysis of Binary Functions”, Proc.
The proposed algorithm is implemented and experimentally
ICEST 2019, vol. 1, pp. 187-190, Ohrid, Macedonia, 2019.
tested for discovering bent function of 4, 6 and 10 variables. It [13] M. Radmanović, R. Stanković, “Random generation of bent
is compared with our previous algorithm for random functions on multicore CPU platform”. Proc. ICEST 2016, pp.
discovering of bent functions in RM domain. Comparation 239-242. Ohrid, Macedonia, 2016.
includes the computation time needed for discovery of one [14] M. Radmanović, “Efficient Random Generation of Bent
bent function for the predefined number and/or order of non- Functions Using GPU Platform”, Proc. 12th Int. Workshop on
zero coefficients in RM spectrum (using vertical, horizonal Boolean Problems, IWSBP2016, Freiberg, Germany, 2016, pp.
and grid subsets), and time needed for discovery of one bent 167-173.
function for random number and order of non-zero [15] M. Radmanović, R. Stanković, “Construction of Subsets of Bent
Functions Satisfying Restrictions in the Reed-Muller Domain”,
coefficients
Facta Universitatis - Series: Electronics and Energetics, vol. 31,
Experimental results confirm that the time required for no. 2, pp. 207-222, 2018.
discovery of bent function when using RM subsets are smaller [16] M. G. Karpovsky, R. S. Stanković, and J. T. Astola, Spectral
than when using random discovery. It can be also seen that Logic and its Applications for the Design of Digital Devices,
proposed algorithm in RM domain significantly reduces Wiley, 2008.
search space for bent function discovery, which lead to [17] M. Radmanović, R. S. Stanković, “Subsets of Boolean Bent
discovery of bent functions that are not completely non- Functions in the Spectral Reed-Muller Domain”, Proc. 13-th
deterministic. Therefore, in practical application there is need Reed-Muller 2017 Workshop, pp. 33-39, Novi Sad, Serbia,
for using various RM subsets. 2017.
Future work will be on research for discovery of bent
functions using various other spectral domains, like Haar, or
Kronecker domains.

10
 Multilayer Long Short-Term Memory (LSTM) Neural
Networks in Time Series Analysis
Nemanja S. Malinović1, Bratislav B. Predić2 and Miloš Roganović3
Abstract – Traditional methods that deal with time series sequence analysis and sequence generating are. Fig. 1
analysis problems are based on statistical models. The latest illustrates a multilayer RNN. An input vector sequence
methods that solve time series analysis problems are based on the = ( , ,…, passes through the stack of hidden layers
use of a special architecture of recurrent neural networks known with neurons first to compute the hidden vector sequence
as Long Short-Term Memory Networks. This paper presents
solutions that address the prediction of product sales on a
denoted as , ,…, where represents an
monthly basis. output of T-th neuron is n-th network layer. After passing the
last layer, on the output is generated predicted sequence
Keywords – Time series, Neural networks, Prediction, LSTM. = ( , ,…, . Each output vector is used to
parameterize a predictive distribution Pr( | ) for the next
possible outputs .
I. INTRODUCTION

Time series represents a phenomenon that describes events

of the same type that recur over a period of time with certain
characteristics. Some of the goals of a time series analysis are
sequence prediction, sequence classification and sequence
generating. Models used for describing time series are
numerous starting from simple statistical methods to the
modern neural network methods. This work will describe the
transition from the basic methods to the state of the art
methods like LSTM neural network. This work will give a
review of neural networks as a model for time series
forecasting focusing on LSTM neural networks architecture
and represent results of sales prediction of a product on a
product type and product category levels.
Fig. 1. Multilayer RNN (taken from [1])
II. NEURAL NETWORKS The skip connection from the inputs to all hidden layers
and from the all hidden layers to the output make a training
In analyzing time series with traditional statistical models as faster by reducing the number of processing steps between the
stochastic processes, there are problems such as missing bottom of the network and the top, and thereby mitigating the
values in a dataset, models often work only with univariate vanishing gradient problem [2] Activation of hidden layers can
data and data are assumed to have a linear dependence. For be calculated as
this purpose particularly good results are shown by a special
class of recurrent networks known as Long-Short Term neural , (1)
networks (LSTM). , (2)
where terms denote weight matrices, is hidden layer
A. Recurrent Neural Networks activation function, and Bias vectors are denoted as b. The
output sequence is computed as
Recurrent neural networks (RNN) represent a feed-forward
class of neural networks that find application in problems as ӯ ∑ , (3)
ӯ , (4)
1Nemanja S. Malinović is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš, where is the activation function of the output layer.
Serbia, E-mail: malinovicnemanja@elfak.rs Adjustment of connections weights can be performed using
2Bratislav B. Predić is with University of Niš, Faculty of Back Propagation (BP) algorithm or its advanced version Back
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš, Propagation Through Time (BPTT) [3].
Serbia, E-mail: bratislav.predic@elfak.ni.ac.rs
3Miloš B. Roganović is with University of Niš, Faculty of

Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,

Serbia, E-mail: milos.roganovic@elfak.ni.ac.rs

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 11

 B. Exploding Gradient Problems

A Back Propagation is actually a propagated gradient

descent algorithm, where gradients are propagated backward.
Recurrent neural networks have exploding gradient problems
in learning a long time series. As explained by Hochreiter and
Schmidhuber [4], error propagation is determined by:
∑ …∑ ∏ . (5)
In Eq. (5), n represents number of non-input units (neurons),
q denotes the number of time steps for how far the error is
propagated from an arbitrary unit u to an arbitrary unit v, w is
a weight on the connection between connected units and
, f is activation function and is an input on
the unit . In case

1 (6) Fig. 2. Basic LSTM cell (taken from [4])

for all m the largest product increases exponentially with q and
index of memory cell,
exploding gradient problem appears. In contrast, when
block and
0 (7) l. Cell input is defined as
a vanishing gradient problem appears. Those problems cause
difficulties and weights adjustment of neural network
connections, especially for long time series with the great
value of q.
C. Basic Architecture of Long Short-Term Memory
Recurrent neural networks behave badly with data that
contains long time dependencies. Standard RNN is not able to
remember old inputs. Long Short-Term Memory (LSTM)
neural networks overcome vanishing and exploding gradients
problems and have a memory which is used to deal with
the problem of long time-dependent inputs.
For the first time, this type of architecture was presented in
the paper [4]. The basic design of memory cells contains input
and output gate as shown in Fig. 2.
With this cell vanishing gradient problem is solved by
keeping constant error flow in case of new signal absence. A
cell contains state denoted as is affected by the cell input
signal and two input signals of gates, an input signal of
the input gate denoted as and signal which
represents input signal for output gate.
Input gate activation for cell j at the moment t can be
calculated as
= (8)
with input of gate calculated as
∑ 1 . (9)
In Eqs. (8)-(11), j represents index of memory block, v is
denotes v-th cell in j-th memory
is the weight of connection from unit m to unit
∑ 1 . (12)
With cell input and gate outputs defined in this way,
functions for cell state change and state output can be
represented as
0 0, (13)
1 , t>0 , (14)
. (15)
Eq. (13) shows initial cell state and Eq. (14) is a function of
later changes. The output of the cell is explained in Eq. (15).
In calculation of next cell state and cell output g(x) and h(x)
represent a central logistic sigmoid function. This design deals
fine with vanishing gradient problem, but the output may
become constant if the same sequence continues to come on
the cell entrance, which happens because of h(x) saturation
that starts with the same value repeating on the cell exit.
D. LSTM with Forget Gate
Paper [5] represents a solution to this problem by
introducing the new gate, forget gate. A function of this gate is
to reset memory blocks once their contents are out of date and
hence useless. The architecture of the LSTM cell with forget
gate is shown in Fig. 3.
Input and output of the forget gate are calculated
as

Similarly, input and activation for the output gate are shown ∑ 1 , (16)
in Eqs. (10) and (11), respectively:
= . (17)
∑ 1 , (10)

= . (11)

12
 A. Individual product sales

Data used for individual product sales forecasting contains

daily product sales in a shop for a period of 30 months. Sales
from the first 30 months are used as a training set and the last
three months make up the training set. The goal of a prediction
was to find a number of potentially sold product items for a
future month.
The architecture of the used neural network includes two
layers, the first ConvLSTM and the simple LSTM layer as the
second. Table I represents the loss and accuracy values of
individual product sales on a monthly base obtained on
training and validation set for different training parameters.

TABLE I
PRODUCT SALES PREDICTION
Fig. 3. LSTM cell with forget gate (taken from [4])
Time Tr. Tr. Val Val
Epochs Batch
steps Acc. Loss Acc Loss
A. ConvLSTM Cells 1000 6 32 96.77% 1.53 100% 0.02
1000 12 32 96.77% 1.53 100% 0.02
This type of cell is first shown in the paper [6]. The
4000 6 32 96.77% 1.52 100% 0.20
ConvLSTM determines the future state of a certain cell in the
4000 12 32 96.77% 1.53 100% 0.20
grid by the inputs and past states of its local neighbors. This
can easily be achieved by using a convolution operator in the 10000 6 32 96.77% 1.51 100% 0.20
state-to-state and input-to-state transitions. Gate outputs are 10000 12 32 96.77% 1.51 100% 0.20
now changed and before activation function is applied gate
inputs pass through convolution operator function Used time steps values are chosen with the assumption that
data owns periodicity on a semi-annual and annual basis.
= , (18) Higher values of training epochs are used to compare training
time values with time values of the network for category sales
= , (19)
prediction.
= . (20)
In Eqs. (18)-(20), function of convolution operator is A. Category Product Sales
represented as . According to [7], this modification ensures
spatial invariance of the unit. Considering the number of layers and type of layers, the
same architecture is used for a prediction of monthly sales on
III. FORECASTING OF SALES TIME SERIES a product category base. Table II shows the same model
parameters for monthly sales of the product category.
An application of LSTM neural networks for prediction of
sales time series is given in this section. Prediction is made on TABLE II
CATEGORY SALES PREDICTION
a dataset that can be found on Kaggle site [8]. The dataset
contains product sales information of several shops in the Epochs Steps Batch T. Ac. T.Los V.Ac. V.Los
period between January 2013 and November 2015. Sales
1000 6 32 86.67% 2.46 33.33% 0.15
information consists of a number of individual products sold
on a daily level in each of the shops. Products are also 1000 12 32 86.67% 2.82 33.33% 0.23
categorized into 84 different categories and there are 1256 4000 6 32 92.77% 2.41 66.66% 0.25
different products. Time series prediction solves problems of 4000 12 32 93.33% 1.54 66.66% 0.22
monthly sales for individual products and product categories 10000 6 32 96.67% 1.82 66.66% 0.31
in the single shop. 10000 12 32 96.67% 1.51 100% 0.21
Since the data are originally given in the form of daily sales
on item level, before training and testing processes data are Since data are aggregated on category level, bigger
grouped on a monthly level for a specific product or product fluctuations are present in monthly sale. More epochs are
category dependent on a problem that is solved. So the time needed by the network to achieve satisfactory prediction
step used in all architectures described below is a month. accuracy.

13
 B. Models Parameters another proof that the LSTM network expresses especially
good results in solving this type of problem.
Parameters of trained models will be given in this section. This paper represents a rarely used architecture that
Training times are given in Table III. Since training time is includes the ConvLSTM layer as a part of the network. This
invariant with respect to the number of time steps, table layer is included with the idea to capture relations between
contains only results for 12 time steps. Time is measured in neighbors’ values which are given in successive monthly
seconds. sales before the LSTM layer. Also, the aim of this paper was
In Table IV are shown number of parameters per layer and to show the differences in the way of training models for
time as well as total number of parameters in network. The sales prediction at the product and category level. Since
batch size used in training was 32, as an optimizer is used fluctuations in the number of items sold on category level
Adam optimizer and mean squared error as a loss function. are greater than fluctuations in the product level conclusion
Created two layered LSTM network shows good prediction is that a larger number of epochs are necessary to cover and
results with high accuracy. The used dataset does not show learn all relations.
adrastic difference in semi-annual and annual seasonality
according to results.
REFERENCES
TABLE III
COMPARISON OF TRAINING TIMES [1] A. Graves, “Generating Sequences With Recurrent Neural
Networks”, ArXiv13080850 Cs, Jun. 2014, Accessed: Apr. 13,
Epochs Prediction type Time 2020. [Online]. Available: http://arxiv.org/abs/1308.0850
1000 Product 22.17s [2] Y. Bengio, P. Simard, and P. Frasconi, “Learning Long-Term
Dependencies with Gradient Descent is Difficult”, IEEE Trans.
1000 Category 12.63s Neural Netw., vol. 5, no. 2, pp. 157–166, Mar. 1994, doi:
4000 Product 76.26s 10.1109/72.279181
4000 Category 42.26s [3] J. Guo, “BackPropagation Through Time,” 2013.
[4] S. Hochreiter, J. Schmidhuber, “Long Short-Term Memory”,
10000 Product 205.74s MIT Press, Nov. 01, 1997, Accessed: Apr. 13, 2020. [Online].
10000 Category 114.76s Available:https://doi.org/10.1162/neco.1997.9.8.1735
[5] F. Gers, J. Schmidhuber, and F. Cummins, “Learning to Forget:
TABLE IV Continual Prediction with LSTM”, Neural Comput., vol. 12, pp.
NUMBER OF PARAMETERS PER LAYER 2451-2471, Oct. 2000, doi: 10.1162/089976600300015015
[6] S. Agethen, W. H. Hsu, “Deep Multi-Kernel Convolutional
ConvLSTM LSTM Dense Total LSTM Networks and an Attention-Based Mechanism for
layer layer layer Videos”, ArXiv190808990 Cs Stat, Jul. 2019, Accessed: Apr. 14,
25 4864 33 4922 2020. [Online]. Available: http://arxiv.org/abs/1908.08990
[7] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W. Wong, and W.
Woo, “Convolutional LSTM Network: A Machine Learning
IV. CONCLUSION Approach for Precipitation Nowcasting”, ArXiv150604214 Cs,
Sep. 2015, Accessed: Apr. 14, 2020. [Online]. Available:
http://arxiv.org/abs/1506.04214
Compared to the solution given on Kaggle site [8],
[8] “Sales Forecast LSTM - 67% (Beginner-Friendly).”
the network presented in this paper has better accuracy. The https://kaggle.com/carmnejsu/sales-forecast-lstm-67-beginner-
time necessary for network training is longer and that can be friendly (accessed Apr. 14, 2020).
explained by network complexity. This neural network is

14
Computer Visualization of Geometry of Polynomial Roots
Ivan Petković1 and Petar Rajković2
Abstract – Using computer visualization, several results in the
field of geometry of polynomial roots are presented. This
computational tool yields a new dimension and a deep insight
into properties of the geometry of polynomial roots and iterative
processes for finding polynomial roots, discovering at the same
time new properties, quite beyond human abilities. More
generally, the presented communication demonstrates that
sophisticated computer graphics, transforming numerical data
into images, brings modern and novel perspectives of some
mathematical topics.
Keywords – Computer visualization, Computer graphics,
Geometry of polynomial zeros, Analysis of root-solvers.
I. INTRODUCTION
The aim of this paper is to present results of author's
investigation in the field of geometry of polynomial roots and
root-solvers using computer visualization. This is certainly
interesting and challenging theme since polynomials are a part
of many problems of numerous scientific branches, such as
applied mathematics, engineering disciplines, computer
science, physics, chemistry, astronomy, finance, health care,
social science, as well as many other fields of human
activities. Undoubtedly, the use of computer visualization,
that actually transforms numerical data into 2D or 3D
pictures, provides modern and novel approach to some
mathematical topics yielding a new dimension and a deep
insight into properties and behaviour of iterative processes for
finding polynomial roots as well as distribution of the roots.
II. COMPUTER VISUALIZATION
The advance of digital computer hardware and software has
had a dramatic influence on many scientific disciplines; at
present, computers play an increasingly central role in almost
every scientific discipline, including mathematics as well. Let
us emphasize that many difficult problems in applied
mathematics and other scientific branches cannot be solved
without powerful computational tools, such as symbolic
computation and computer graphics.
Rapid development of computer graphics during the last
two decades provides scientists in many disciplines to use
sophisticated computer techniques in order to study and
visualize phenomena which are not observable directly. High
1Ivan Petković is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
ivan.petkovic@elfak.ni.ac.rs
2Petar Rajković is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
petar.rajkovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
quality computer visualization enables entirely new and
remarkable insights into a wide variety of disciplines, to
mention some of them: architecture and engineering,
biomedical applications (X-ray Computed Tomography (CT),
Magnetic Resonance Imaging (MRI), above all), mathematics,
physics, business and management graphics, education and
learning, user interface design, image syntheses, mapping and
cartography, video and multimedia technology (animation,
movies, video games), vector graphics, 3D modeling, digital
image processing, visual arts and design, and so on. Let us
emphasize that many problems in the mentioned topics are
still mission impossible for humans.
Computer graphics provides very fast and fruitful bridge
between computers and human beings. Simply, the
combination of technology and mathematics, supplied by
computer graphics, has numerous applications in theory and
practice at present and, certainly, a huge perspective in the
future. G. David Kerlick [1] said that “... scientific (or
mathematical, or engineering) visualization uses computer
graphics to make the results of analysis, experiment, and
simulation accessible to humans via the visual system, thereby
aiding insight into data. A series of case studies illustrate
situations in which visualization has led to new insights. In
particular, mathematicians got a powerful tool to visualize the
abstract objects and processes that arise in all branches of
mathematical research.” Today, using computer graphics,
mathematicians are able not only to study the geometric
aspects of many mathematical and engineering branches but
also to discover new properties on various topics of
mathematics and construct new very efficient algorithms.
We end this section with the announcement of computer
scientists from Carnegie Mellon University (Pittsburgh) that
they developed a new type of computer visualization which
automatically turns complex mathematical equations and
expressions into pictures. This is certainly a further important
advance in the topic. The researchers will present their
software in July of 2020 at the SIGGRAPH (virtually)
Conference on Computer Graphics.
III. DISTRIBUTION OF ROOTS OF RANDOM
POLYNOMIALS
Algebraic polynomials whose coefficients are random
numbers are of great importance since they appear in various
problems of physics, engineering and economics such as
filtering theory, spectral analysis of random matrices,
statistical communication, regression curves in statistics,
characteristic equations of random matrices, the study of
random difference equations, the analysis of capital and
investment in mathematical economics. etc. For these reasons,
a number of books and papers have been devoted to the study
15
of random polynomials, see, e.g., [2], [3]. Working in the area this purpose we have performed a great number of numerical
of Experimental Mathematics, a new scientific discipline that experiments. Using graphical tools of Wolfram computational
spans numerical mathematics and computational techniques, system Mathematica, we have tested numerous random
we made use of graphical methods to visualize an important polynomials of various degrees. Our experiments confirmed
theorem on distribution of roots and pose a conjecture of the the assertion of Theorem 1.
symmetry of complex roots of random polynomials.
Denote a sequence of independently identically distributed Example 1. For demonstration, we present one experiment
(real or complex valued) random variables with } . Let involving a polynomial of degree 2000 with random
( )= + + ⋯+ + coefficients belonging to the interval −2,2 . From Fig. 1 we
be a random polynomial of degree with the roots a ,…, a . observe that almost all roots are located in the ring
Furthermore, for , (0 ≤ ≤ < ∞) introduce a | 1 − < | | < 1 + } where ≈ 0.01 which empirically
probability measure on the complex plane confirms Theorem 1 to a good extent. As an additional
( , ) = ( | ≤ |a | ≤ }), where (∙) denotes the consequence of our extensive experiment, we arrived at the
number of roots that belong to the ring | ≤ |a | ≤ } in assumption (in terms of convergence in probability) that the
the complex plane. Then / defines the empirical roots of random polynomials are symmetrically located with
distribution of roots of . If ∈ (0,1), respect to the coordinate origin. For now, this assumption
we define = 1 − , = 1 + , then is called delta remains an open problem.
measure in the empirical distribution.
IV. DYNAMIC STUDY OF ROOT-FINDING METHODS
The following theorem has been proved in [3]:
The growing development of computer graphics at the end
Theorem 1. If and only if
of the twentieth century have led to the advance of efficient
E (1 + | |) < ∞, methodology for visual study of convergence behaviour of
iterative methods for solving nonlinear equations. It turned out
then the sequence of the empirical distributions / that dynamic study, relied on basins of attraction, is a
converges to the delta measure at 1 almost surely, that is, powerful tool for realistic quality study of root-finding
1 methods. The basin of attraction is a kind of computer
(1 − , 1 + ) → 1, → ∞
visualization that provides visual insight into the convergence
holds for any ∈ (0,1). behaviour of root-finding methods. Actually, basins of
attraction are the sets of points in the complex plane which
In the above theorem E is mathematical expectation while the simulate the convergence to the roots of a given function
denotation → denotes so-called convergence in probability. under the chosen iterative process. They have brought
This theorem asserts that, under some weak constraints on the significant contribution to the topic due to the important data
coefficients of a random polynomial, almost all its roots and clear visual insight into the basic features of considered
“concentrate uniformly” close to the unit circle with high iterative methods, convergence behaviour and domain of
probability. convergence above all. In this way, we arrive at a fruitful
methodology for estimating quality of root-finding methods
and their ranking, which successfully replaces unreliable
methodologyfor ranking iterative methods based on
numerical examples.
First we give the definition of a basin of attraction.

Definition 1. Define a (convergent) root-finding iteration by

= ( )( = 0,1,2, … )
that involves sufficiently many timesdifferentiable function
in some complex domain ⊆ ℂ having simple or multiple
roots , , … , ∈ . Then the basin of attraction for the
root is defined as follows:

, ( )= ∈ |ℎ = ( ) ℎ
= }
Assume that we have found a rectangle
= − , } × − , } centered at the origin that contains all
roots of a given polynomial . The studied iterative method is
Fig. 1. Location of roots of random polynomial of degree 2000 tested on each of the × equally spaced points in the
rectangle which form an equidistant lattice . First, we
It is of interest to visualize the assertion of Theorem 1, define the limit number of iterations and then run iterative
which is typical subject of Experimental Mathematics. For process = ( ) starting from an initial point ∈ . If

16
the iterative process does not satisfy the given stopping
criterion | − | < ( defines the accuracy of
approximations, say, = 10 or = 10 ) in ≤
iterations, then such starting point is proclaimed “divergent.”
For each basin we record (i) average number of iterations (for
all points belonging to the lattice ) required to satisfy the
stopping criterion, (ii) CPU time in seconds for all
× points and (iii) the number of (proclaimed)
“divergent” points. The data (i)-(iii) provide deeper insight
into the behavior of an iterative method and its domain of
convergence. Obviously, a method is as better as the
consumed CPU time, the average number of iterations and
function evaluations per point, and the number of “divergent”
points are smaller. The best outcome is that the number of
divergent point is 0, which indicates the global convergence
of the considered method.
Furthermore, we associate different colors to different
basins of attraction in such a way that the shading is darker if
Fig. 2. Newton’s method Eq. (1)
the number of iterations is higher. Besides, we color starting
points which do not satisfy the stopping criterion after IT
iterations in black (so-called “divergent” points).
Convergence behavior of any method can be also estimated
to a certain extent according to the shape of basins of
attraction for the tested example.It is always preferable that
the basins of attraction for the roots have as large as possible
unvaried contiguous areas, separated by the boundaries that
have (approximately) straight-line form. The fewer blobs and
fractals on the boundaries also point to good convergence
properties.
To demonstrate the dynamic study of two iterative methods
for finding simple roots, we present a typical example using
equidistant lattice consisting of 360000 points, that is, the
resolution is 600×600. In this example the permitted number
of iterations is = 40, and the stopping criterion is given by
| − | < 10 = . Our aim is to compare two the most
frequently used root-solvers for approximating simple zeros,
Newton's method of the second order
( ) Fig. 3. Halley’s method Eq. (2)
= − (
( = 0,1, … ) (1)
)
According to the above data we conclude that Newton's
and Halley's method of the third order method Eq. (1) has considerably more divergent points (about
( ) ( ) 1.25%) than Halley's method Eq. (2) (less than 0.006%
= − ( ) ( ) ( )
( = 0,1, … ) (2) starting points). Halley's method Eq. (2) reaches the stopping
criterion using only 5.2 iterations (in average) against 10.5 for
Example 2. We have plotted two basins of attraction applying Newton's method Eq. (1) and consumes 58.16 seconds for all
the methods Eq. (1) and Eq. (2) to the polynomial 360000 starting points, which is considerably less than
( )= –1 Newton's method (110.6 seconds). Considering shapes of
and the square = = + | − 3 ≤ ≤ 3, −3 ≤ ≤ 3}. basins we observe that in both cases particular basins are of
The basins are given in Figs. 2 and 3. large unvaried size. However, the basins of Halley's method
Plotting these two basins of attraction we have recorded the Eq. (2) have mild advantage since their boundaries are almost
following useful data: straight lines (desirable property) and contains only few small
Newton's method Eq. (1) Halley's method Eq. (2) blobs, while the boundaries of basins of Newton's method Eq.
divergent point 4508 12 (1) have not only larger blobs but also fractal parts
average number of iteration 10.5 5.2 (undesirable property). The presented results clearly show that
CPU time (in sec) 110.6 58.16 Halley's method Eq. (2) is better than Newton's method
Eq. (1) regarding the most important convergence properties.
This fact could be surprising facts for many numerical
analysts who consider that Newton’s method is most suitable.

17
 V. COMPUTER VISUALIZATION OF THE FLOW OF
ITERATIVE PROCESS

In most cases the most powerful methods for approximating

all polynomial roots are based on simultaneous calculation of
all roots. Estimating the quality of simultaneous root-finding
methods it turned out that so called empirical methodology
gives the most realistic ranking, see
Kalantari [4], Andreev et al. [6], and I. Petković and
Herceg [7]. In [7] the authors used empirical methodology
reducing it to the following quality criterion:“Among the set
of iterative methods for the simultaneous approximation of
polynomial roots, the best method is the one that attains the
required accuracy of approximations using the smallest CPU
time starting from the same set of initial approximations”.
Empirical methodology for simultaneous methods is based
on a huge number of experiments by measuring CPU times in
the presence of real conditions. These conditions assume the
choice of arbitrary initial approximations equidistantly spaced Fig. 4. Trajectories of approximations of polynomial roots
on a circle of radius of a disk that contains all roots of a
given polynomial. For demonstration, we present the
following example. VI. CONCLUSION
Example 3. We have considered Ehrlich-Aberth methods Using computer visualization, some results of author's
with Newton's corrections [8] for simultaneous determination investigation in the area of root-finding methods and
of simple zeros of a given polynomial P of degree distribution of polynomial roots are presented. They provide a
new and fruitful deep insight into the geometry of polynomial
( ) ( )
= − ( ) , (3) zeros and iterative processes, simultaneous discovering new
( ) ∑ ( ) ( ) ( ) ( )
properties of the considered topic and ranking roots-solvers.
/ The presented miraculous properties would be impossible for
humans without use of computer visualization.
where ∈ 1, … , } and = 0,1, … is the iteration index. It
was proved in [7] that this method is one of the best methods
for finding polynomial roots simultaneously. The method
REFERENCES
Eq. (3) has been applied to the “two-ring” polynomial
[1] G. David Kerlick, “Four Lectures on Computer Graphics and
( ) = ( + 7)( − 1)( − 5 ) of degree 21. Data Visualization”, Ch. 15 in: Relativity and Scientific
The complete flow of the iterative process is described by Computing, Springer, pp. 287-313, 1996.
trajectories composed of current approximations (marked by [2] A. T. Bharucha-Reid, M. Sambandham, Random Polynomials,
black points) starting from initial approximations Academic Press, 1986.
equidistantly spaced on a circle of a radius of a disk [3] I. Ibragimov, D. Zaporozhets, “On Distribution of Roots of
containing all roots of the given polynomial, see Fig. 4. Red Random Polynomials in Complex Plane”, arXiv:1102.3517v1
points in Fig. 4 denote the targeting roots. [math.PR] 17 Feb 2011.]
From Fig. 4 one can observe that all trajectories possess [4] B. Kalantari, Polynomial Root-Finding and Polynomiography,
World Scientific, New Jersey, 2009.
mainly regular paths; besides,approximations are directly
[5] I. Petković, Đ. Herceg, “Symbolic Computation and Computer
approaching the roots. They are almost radially distributed Graphics as Tools for Developing and Studying New Root-
and continuously gravitate towards the roots in the course of Finding Methods”, Appl. Math. Comput., vol. 295, no. 2, pp.
iterating. Starting from initial points (equidistantly spaced on 95-113, 2017.
the circle), the considered method converge (super) linearly [6] F. Andreev, B. Kalantari, and I. Kalantari, “Measuring the
but almost straightforwardly toward the roots, with rapid Average Performance of Root-Finding Algorithms and Imaging
accelerations at the end of iterative process (due to the it Through Polynomiography”, Proceedings of 17th IMACS
closeness to the target - the exact root). At the moment, there World Congress, Scientific Computation, Applied Mathematics
is no proof that these trajectories always possess this and Simulation, Paris, 2005.
[7] I. Petković, Đ. Herceg, “Computer Methodologies for
desirable straightforward behaviour. The proof is a
Comparison of Computational Efficiency of Simultaneous
challenging but very difficult open problem. Actually, the Methods for Finding Polynomial Zeros”, J. Comp. Appl. Math.,
proof would show that the considered method Eq. (3) is vol. 368, page 19, \# 112513, 2020.
globally convergent. The comparison and analysis of the form [8] A. W.M. Nourein, “An Improvement on two Iteration Methods
of trajectories of tested simultaneous methods can serve for for Simultaneous Determination of the Zeros of a Polynomial”,
the estimation of the quality of these methods. Internat. J. Comput. Math., vol. 6, no 3, pp. 241-252, 1977.

18
 Continuous 2D Maps Based on Spherical Cube Datasets
Aleksandar M. Dimitrijević1 and Peter A. Strobl2
Abstract – This paper is one in a series of papers attempting to
popularize the application of cube-based Discrete Global Grid
Systems (DGGSs). To increase acceptability, a simple but
effective implementation is proposed. The paper discusses the
fundamental aspects of such an implementation and presents a
continuous 2D map creation process, based on such systems.
Keywords – Map, Discrete Global Grid System, Geographic
Information Systems, Spherical Cube.
I. INTRODUCTION
People have always needed to georeference, to determine
their position in the space that surrounds them. Traditionally,
parts of the Earth's surface have been transformed into planar
maps because these were easier to produce and store and have
proven to be very effective until there was a need to apply
them to the entire planet. The modern economy, trade,
science, and not the least military applications, with the use of
high-resolution global data sets, require a much more efficient
global framework. In recent decades, the amount of collected
data, obtained by satellite and aerial imaging, has further
increased the need for efficient organization and standardized
representation to facilitate their exchange. These are the
conditions that led to the development of Discrete Global Grid
Systems.
Discrete Global Grid Systems are a class of spatial
reference systems that use hierarchical tessellation of cells to
partition and address the globe without gaps or overlaps. Most
of such systems are based on regular, multiresolution
partitions of polyhedra, called Geodesic DGGSs (GDGGS)
[1], with various methods for achieving the proper tessellation
of the surface [2]. To foster the interoperability of geospatial
data, Open Geospatial Consortium (OGC) has founded the
DGGS Standard and Domain Working Groups and published
the abstract specification [3] and extension mechanisms for
DGGSs.
This paper is a continuation of the previous work [4] on the
popularization of cube-based GDGGSs, through examining
their properties and optimizing the mapping process. Here the
focus is on continuous 2D map creation and axis-based
indexing. A cube-based dataset consists of six separate data
collections, each associated with an appropriate cube face.
However, cube edges are artificially imposed delimiters that
should not constrain browsing or data manipulation. Also,
some algorithms applied in Geographic Information Systems
(GIS) are developed for planar geometry and may require
1 orientation relative to the Earth, an indexing method for cell
Aleksandar M. Dimitrijević is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: aleksandar.dimitrijevic@elfak.ni.ac.rs
2Peter A. Strobl is with European Commission, DG Joint Research
Centre, Ispra, Italy, E-mail: peter.strobl@ec.europa.eu
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
collecting all relevant features in a single 2D reference
system. We propose a method for collecting and combining
all relevant map sections, regardless of the cube face they
belong, according to the map creation parameters. The created
2D map is continuous, which means it can be traversed in all
directions without reaching the boundaries.
The discussion starts with a brief history of cube-based
maps and GDGGSs, in the next section.
II. RELATED WORK
The oldest preserved cube-based map dates back more than
two centuries ago. C. G. Reichard created a six-sheet atlas of
the whole world [5] in 1803, using a gnomonic projection.
Cube-based projections of the Earth returned to focus in the
early 1970s. Sadourny [6] defined a class of conservative
finite-difference approximations of the primitive equations for
quasi-uniform spherical grids derived from regular
polyhedrons in 1972. The first feasibility study of a
quadrilateralized spherical cube Earth database system [7] was
carried out by Chan and O’Neill only a few years later, for the
US Navy Department. They proposed a spherical cube map
projection that, with some modifications, was extensively
used for the Cosmic Background Explorer (COBE) project at
NASA, more than a decade later. O'Neill and Laubscher [8]
derived the exact equal-area sphere-to-cube mapping in 1976,
but the proposed formulation was thought to be
computationally intractable at that time [9]. Increase of the
computation power revived this projection in the last decade
when e.g. Lambers and Kolb used it for the accurate rendering
of planetary-scale terrain data [10].
At the turn of the century, the sphere-to-cube mapping was
used as a simple method to extend existing 2D spatial
databases and information browsing systems to the 3D
spherical data model [11]. Some modern open-source
libraries, like S2 [12], made use of spherical cubes to facilitate
seamless geographic databases with low distortion. The
numerous spherical cube maps proposed in recent years,
highlight the relevance and popularity of this approach.
In the following section, we discuss various aspects of
cube-based GDDGs and design choices that influence their
implementation.
III. CUBE-BASED GDGGS
There are several design choices that fully specify a
GDGGS [1]. Among them are a base polyhedron and its fixed
addressing and polyhedron-to-sphere mapping. In this section,
we discuss our design choices, and how each of them
contributes to the quality and efficiency of the implementation
process.
19
A. Base Polyhedron
The choice of the base polyhedron crucially influences the
characteristics of GDGGS. We have chosen a cube for several
reasons. First, the faces of the cube, as well as the cells that
build them, are quadrilaterals, which is the optimal storage
form. They are the most user-friendly, easily mapped to
standard data structures, and displayed efficiently.
Second, quadrilateral cells allow congruent nesting,
meaning that the lower-level cell completely contains higher-
level cells that cover the same area. This feature is consistent
with many spatial partitioning data structures, such as
quadtrees.
Third, the coordinate axes are orthogonal, and all cells at
the same level of detail are identical. These properties are
usually taken for granted, but cannot be achieved with other
grids, like hexagonal.
Unfortunately, the use of the cube, as a base polyhedron,
also has some disadvantages. One of them is relatively large
distortion. In general, the fewer and larger faces a polyhedron
has, the worse it approximates the Earth's surface.
Also, quadrilateral grids do not exhibit uniform adjacency.
Each cell has four edge-sharing neighbors with centers on a
certain distance, and four vertex-sharing neighbors on greater
distance. This property complicates the use of such grids for
discrete simulations [1].
Cube-based GDGGSs have their advantages and
disadvantages. What prevails depends on the application.
Considering that our goal is to popularize the application of
GDGGSs, we chose the cube as the base polyhedron to
simplify the implementation and increase acceptability. To
reduce the effect of distortion, to some extent, we proposed an
appropriate mapping of the ellipsoid to the sphere and an
optimised orientation of the base cube [4].
Fig. 1. Optimal base cube orientation for the landmass distortion
minimization and the cube-map faces for such orientation
B. Base Cube Orientation
The simplest orientation of the base cube aligns one
Cartesian axis of the cube with the Earth’s axis of rotation and
points another toward the prime meridian [8]. This orientation
is common to most implementations [7, 12]. However, some
20
properties can be improved if the cube is rotated. Even the
oldest known cube-based map [5] rotates a cube by 27°
around the Earth’s axis of rotation. One of the first discussions
on the orientation of the base cube is given in [8], as a way to
maximize the usefulness and efficiency of the spherical cube
database, accompanying with appropriate transformations.
Considering the base cube orientation, our goal is to
minimize the distortion of the landmass and to reduce the
continental plates splitting by the cube edges. Confining the
rotation angles to integer numbers, the goal is achieved by
rotating about all thee Cartesian axis for the following angles:
ϕ = 17°, θ = -10° and ρ = 32° [4]. Fig.1 displays the
orientation of the cube and shows the position of the
continental plates on the cube faces for the optimised
orientation of the base cube.
C. Spatial Partitioning and Indexing
To create a multiresolution discrete grid, each face of the
polyhedron has to be appropriately subdivided into cells. The
partitioning method is defined by topology and refinement
ratio. Modern GDGGSs use one of the four partition
topologies [1]: squares, triangles, diamonds, or hexagons.
Cube-based GDGGSs, almost exclusively, use square
partition topology as a natural way to subdivide square faces.
The refinement ratio is defined as a ratio of the number of
planar polygon cells at two consecutive resolutions. Small
refinement ratios are preferred since they allow more grid
resolutions, and, hence, enable a smoother transition between
levels of detail. Simple and popular refinement ratios are 4
and 9 but other possibilities exist [13].
Dealing with an extremely huge number of cells imposes
the need for an efficient addressing scheme, usually referred
to as indexing. Many indexing methods have been proposed
for DGGSs, but all of them can be categorized into three types
[14]: hierarchy-based, space-filling curve and axis-based
indexing.
Hierarchy-based indexing follows the refinement process,
usually imposed by a space-partitioning data-structure in the
background. This index is typically represented with a string,
which length corresponds to the resolution and each character
(or number) represents the spatial position of the cell within
its parent. Space-filling curve indexing is based on recursively
defined 1D curves that systematically traverse a space (a unit
square in case of 2D space). This indexing scheme is popular
because it serializes a multidimensional address into a very
compact single number and reduces access time to storage; by
preserving near-neighbor relationship (cells with consecutive
indices are mainly spatially close to each other). The two most
popular space-filling cures, applied in cube-based GDGGSs
indexing, are Hilbert and Morton curves. For example, a
quadrilateralized spherical cube Earth database [7] uses the
Morton curve, while S2 Geometry Library [12] uses the
Hilbert curve.
Axis-based indexing is a natural way of cell addressing
using an m-dimensional position vector. By unfolding the
base cube into a plane, a unique 2D coordinate system can be
used for all its faces. However, most of the implementations
utilize separate coordinate systems for each face, which
introduces a face dereferencing coordinate (f) in the index.
Also, dealing with multiresolution data adds another
coordinate referring to the resolution (r). Axis-based indexing
allows trivial neighbors determination, for the cells within the
same cube face, but also easy adjacent cells addressing across
the edges of the faces, which qualifies it as preferred indexing
method in our continuous 2D map creation process
To enable multilayer data organization, we use a 5-
coordinate index (l,r,f,x,y), where l is a layer, r is resolution, f
is a face, and x and y are integer cell coordinates within the
face. The order of the coordinates corresponds to the
frequency of their change to alleviate a section name creation
by concatenating more frequently changed substrings at the
end of the section name. The layer and the resolution are fixed
for a map. As such, the procedure for adjacent cell name
generation recalculates only (f,x,y) part of the index.
Coordinate f changes only on the edges of the faces, which is
mostly infrequent, while (x,y) part is changed with each cell.
D. Sphere-to-Cube Mapping
Generally, there are two types of mapping the base
polyhedron to a sphere or ellipsoid [1]: direct spherical
subdivision and map projection. The first approach creates
partitions directly on the surface of the sphere or ellipsoid that
correspond to the polyhedral face partitions, while the second
approach uses some inverse map projection.
The second approach is more applicable for the cube-based
GDGGSs, with dozens unique projections being used in recent
decades for geodesy but also in computer graphics [15, 16].
Since a cube can only be projected to a sphere or ellipsoid
preserving either area (equal-area projection) or shape
(conformal projections), distortions are unavoidable. Because
trying to preserve either property, significantly deteriorates
the other, many projections are compromises that are neither
equal-area, nor conformal.
Choosing the right projection depends on the application,
and it beyond the scope of this paper. To illustrate the process
of the continuous 2D map creation, we use adjusted spherical
cube [17]. This spherical cube map projection is neither equal-
area nor conformal, with maximum-to-minimum area
distortion 1.414, maximum angular distortion 31.167°, and
average angular distortion 11.57°. By rotating the base cube,
the average angular distortion of the continental plates can be
reduced by almost 20%, from 11.21° to 9.03° [4].
IV. CONTINUOUS MAP GENERATION
The first step in the map-generation process is the input
parameters parsing. The map is specified by the center of the
viewport (in geographic coordinates), the level of detail (or
resolution) and the window size (width and height) in map
units (i.e. pixels). Although it is a common practice to define
map space with a bounding box, we use the center of the
viewport to enable usage of geographic coordinates and hide
the cube-based DGGS’s coordinate system from the users.
The entire map is created as a mosaic of individual parts,
called sections. Sections are bitmaps that correspond to cells
21
in DGGS nomenclature. However, we are using term section
to emphasize their compound structure. Even though the size
of a section can be relatively large, sometimes larger than the
viewport, the viewport may intersect several sections.
The second step in the map-generation process is to
determine the central section, i.e. the section which contains
the center of the viewport. To get the coordinates of the
viewport center for chosen spherical cube projection, we need
to apply the forward pipeline [4]. The forward pipeline, based
on the geographic coordinates in the WGS84 system,
determines the cube face (f) and the relative coordinates
within the cube face (x,y) ∈ [-1,1].
Depending on the resolution (r) and the section size, the
coordinates are determined, i.e. indices of the central section
(is,js) and the relative displacement of the center of the whole
map from the corner of the central section (xs,ys) ∈ [0,1].
Based on the map name, resolution (r), cube face (f) and the
section indices (is,js), the name of the given map section is
generated. Moreover, it is now possible to specify the names
of the sections that make up the entire map. The axis-based
indexing, proposed in the previous section, significantly
alleviates the process of section names generation.
Fig. 3. Orienting sections to fit the neighbor - the upper section
rotates by 90° with a change of the lower (central) section
The entire map is generated as a tile matrix of sections. For
systems with enough memory and larger viewports, a 3×3
matrix or larger is recommended, while for systems with less
memory and smaller viewports, a 2×2 matrix is
recommended. A tile matrix with an odd number of sections
along each axis is easier to implement, because its contents do
not change as long as the center of the viewport is within the
central section. In a matrix with an even number of sections,
the map modification is triggered by changing the quadrant in
which the center of the viewport is, and the central section is
not actually in the middle of the viewport, but is moved to
some corner of the map. The matrix size (N×N), sections size
(Ssize), and the viewport diagonal (Dsize) are bound by the
following relation: Dsize ≤ (N-1) Ssize.
For each position in the tile matrix, it is necessary to
determine the section that should be drawn in it, as well as the
section’s orientation. Fig.3. shows how the upper section
changes orientation to fit the lower one. When defining the
map, it is considered that the central section has a correct
orientation, while the others may be rotated. The rotated
sections only change the order of reading their pixels and do
not require resampling.
 If the central section has both x and y coordinates with
minimum or maximum values for the cube face, at least one
vertex sharing neighbor is an empty section. In the special
case, when just a single section creates the whole cube face,
all four vertex sharing neighbors are empty sections. Fig.4a.
depicts such a case.
a) c)
b) d)
Fig. 4. The process of map creation: a) smallest scale initial
proto-map, b) determining the angle of rotation, c) rotated
proto-map, d) clipped map
If the adjacent sections belong to the same cube face, they
are simply drawn, without any transformations. If they are
not, it is necessary to determine whether they are empty or
belong to the other face. One way to determine this is to test
the index values, which are obtained as a displacement in the
grid relative to the coordinates of the central section. If both x
and y coordinates of the section are out of range for the face,
the section is empty. If only one coordinate is out of range,
section belongs to the adjacent face. Sections from adjacent
faces change orientation according on the neighborhood
relationship. Fig. 4a shows a proto-map of the smallest scale,
with New York City being the center of the viewport.
It is a common practice to orient the map up to the north.
Therefore, the next step is to determine the angle of rotation,
to achieve that property. The angle for which the map should
be rotated is the angle between the line passing through the
center of the viewport and the point having the same longitude
and infinitesimally increased latitude, and the Y-axis on the
map (Fig. 4b).
In addition to the rotation, it is necessary to move the center
of the viewport to the center of the entire map (Fig. 4c). The
last step is clipping to the defined viewport (Fig. 4d).
V. CONCLUSION
Nowadays enormous amounts of geospatial data, collected
from various sources, pose new challenges for their efficient
management. GDDGSs offer a solution, but are often seen as
a complex mechanism that is not yet widely accepted by end-
users. To overcome this, we presented the basic concepts of
cube-based GDGGSs and some respective design choices
making these easy to implement and with acceptable
distortions for many applications.
22
In order to provide a simple interface for classical
geospatial data processing, we proposed a procedure for
generating continuous 2D maps. This way data can be
accessed in native format without the need to know the
coordinate system or other GDGGS implementation details in
which they are originally defined.
REFERENCES
[1] K. Sahr, D. White, and A. J. Kimerling, “Geodesic Discrete
Global Grid Systems”, Cartogr Geogr Inf Sci, vol. 30, no. 2, pp.
121-134, 2003.
[2] A. J. Kimerling, K. Sahr, D. White, and L. Song, “Comparing
Geometrical Properties of Global Grids”, Cartogr Geogr Inf Sci,
vol. 26, no. 4, pp. 271-288, 1999.
[3] Open Geospatial Consortium, “Topic 21: Discrete Global Grid
Systems Abstract Specification”, 2017.
[4] A. Dimitrijević, P. Strobl, M. Lambers, A. Milosavljević, and D.
Rančić, “Distortion Optimized Spherical Cube Mapping for
Discrete Global Grid Systems”, In: Zdravković, M., Konjović,
Z., Trajanović, M. (Eds.) ICIST 2020 Proceedings, vol. 1, pp.
109-113, 2020.
[5] C. G. Reichard, “Atlas des ganzen Erdkreises in der Central-
Projection auf fechs Tafeln entworfen”, Weimar, 1803.
https://haab-digital.klassik-stiftung.de/viewer/image/926414453
[6] R. Sadourney, “Conservative Finite-Difference Approximations
of the Primitive Equations on Quasi-Uniform Spherical Grids”,
Monthly Weather Review, vol. 100, no. 2, February 1972, pp.
136-144.
[7] F. Chan, E. O’Neill, “Feasibility Study of a Quadrilateralized
Spherical Cube Earth Data Base”, Tech. Report EPRF 2-75
(CSC), Environmental Prediction Research Facility, April 1975.
[8] E. O’Neill, R. Laubscher, “Extended Studies of a
Quadrilateralized Spherical Cube Earth Data Base”, Tech.
Report NEPRF 3-76 (CSC), Naval Environmental Prediction
Research Facility, May 1976.
[9] M. Calabretta, E. Greisen, “Representations of Celestial
Coordinates in FITS”, Astronomy & Astrophysics, vol. 395, no.
3, pp. 1077-1122, 2002.
[10] M. Lambers, A. Kolb, “Ellipsoidal Cube Maps for Accurate
Rendering of Planetary-Scale Terrain Data”, in Proc. Pacific
Graphics (Short Papers), pp. 5-10, 2012.
[11] H. Alborzi, H. Samet, “Augmenting SAND with a Spherical
Data Model”, In: Proc. First Int. Conf. on Discrete Global Grids,
Santa Barbara, CA, USA, 26-28. March 2000.
[12] S2 Geometry Library, https://s2geometry.io
[13] A. M. Amiri, F. Bhojani, and F. Samavati, “One-to-two Digital
Earth”, In: Bebis G. et al. (eds) Advances in Visual Computing,
ISVC 2013, LNSC 8034, pp. 681-692, 2013.
[14] A. M. Amiri, F. Samavati, and P. Peterson, “Categorization and
Conversions for Indexing Methods of Discrete Global Grid
Systems”, ISPRS Int. J. Geo-Inf., vol. 4, no. 1, pp. 320-336,
2015.
[15] A. Dimitrijevic, M. Lambers, and D. Rancic, “Comparison of
Spherical Cube Map Projections Used in Planet-Sized Terrain
Rendering”, Facta Universitatis, Series: Mathematics and
Informatics, vol. 31, no. 2, pp. 259-297, 2016.
[16] M. Lambers, “Survey of Cube Mapping Methods in Interactive
Computer Graphics”, The Visual Computer, vol. 36, no. 5, pp.
1043-1051, 2020.
[17] R. Lerbour, J.-E. Marvie, and P. Gautron, “Adaptive Real-Time
Rendering of Planetary Terrains”, In: Proc. Int. Conf. Computer
Graphics, Visualization and Computer Vision (WSCG), pp. 89-
96, 2010.
 Simulation Environment for Optimal Resource Planning
During COVID-19 Crisis
Nenad Petrović
Abstract – Recent COVID-19 outbreak has affected both 19 crisis.
human resources and economy around the world dramatically In pandemic, it is of utmost importance to plan resources
leading towards pandemic and serious global crisis. In this efficiently, but timely – either human, material resources or
paper, a simulation environment leveraging the synergy of deep equipment. The protection of health and the economy of a
learning-based predictions and linear optimization for efficient
resource planning is presented. The proposed solution shows
country are linked tightly [6]. On the other side, it was shown
satisfiable prediction performance and enables efficient resource that simulation is an effective approach contributing to the
exchange by reducing the trading costs. reduction of pandemic’s consequences, especially at
provincial and local level [7]. Therefore, in this paper, a
Keywords – Deep Learning, Coronavirus, COVID-19, Linear software simulation environment is proposed aiming the
optimization, Simulation. efficient resource planning in current pandemic COVID-19
crisis.
In [8], an approach to prediction of COVID-19 spread
I. INTRODUCTION
based on simulation in Iran, which has been one of the
disease’s epicenters in the beginning, was presented.
The hottest topic since the first days this year was the
However, the work presented in this paper builds upon the
discovery of new infectious disease COVID-19 caused by
approach from [9], that has shown as effective, leveraging
SARS-Cov-2 virus (also known as 2019-nCoV or simply
prediction and linear optimization for optimal blockchain-
coronavirus) in China, which was behind several pneumonia
based energy trading in smart grids. Analogously, in [10],
cases in Wuhan’s Seafood Wholesale Market [1]. One of the
linear optimization was used within smart city simulation
main differences from well-known influenza-alike viruses is
environment for network resource planning together with deep
the fact that even persons without any symptoms might be
learning and fading effect calculation.
potential sources of infection [2, 3], which led to fast outbreak
in China. A long and varying incubation period (2-14 days,
exceptions of up to 27 days) [3] and high death rate in elder II. BACKGROUND
population [1, 2], making the situation even more difficult.
However, after just several weeks, the first cases were
A. Linear Optimization
reported in other continents. It is believed that coronavirus
reached Europe in the end of January, with the first case
Linear optimization (known as linear programming) refers
reported in France [4]. In next several days, the cases were
to mathematical method whose goal is to achieve the best
reported in several other countries as well (Germany, Finland,
possible outcome, relying on a mathematical model, where the
Italy, Sweden) [4]. However, there were still just few cases in
requirements are represented by linear relationships [11]. It
each of these countries and still did not seem so critical. Later,
covers the techniques for optimization of a linear objective
in mid-February, the number of cases in Italy has started to
function, subject to linear equality and inequality constraints.
increase dramatically making it one of the world’s worst-
The vector of variables determined as output of optimization
affected countries. Towards the middle of March, the COVID-
is decision variable.
19 was also detected in most parts of Europe which led to
In this paper, linear programming is leveraged for optimal
limited movement, lockdowns and closing borders in some of
resource planning and exchange between cities during
the countries. Therefore, the new disease turned from Chinese
COVID-19 crisis. The implementation is based on AMPL1, an
outbreak to worldwide pandemic in short time, which led to
algebraic modeling language for mathematical programming
tragic consequences in many countries, huge number of lost
which enables writing linear programs using the syntax
lives, on the first place, but catastrophic financial losses and
inspired by traditional algebraic notation. While it is not used
stagnation of economy as well [2]. Moreover, there is no
for solving optimization problems, it provides interface to
specific treatment for COVID-19 proven by clinical trials yet
other programs responsible for that task. In this paper,
[5]. For all these reasons, the world has reached the COVID-
CPLEX2 based on simplex method [12] was used as a solver.

Nenad Petrović is with University of Niš, Faculty of Electronic

Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
nenad.petrovic@elfak.ni.ac.rs

1
https://ampl.com/
2
https://www.ibm.com/analytics/cplex-optimizer

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 23

 Fig. 2. Modelling environment notation as UML class diagram
B. Deep Learning
Deep learning includes various machine learning techniques
based on artificial neural networks. Artificial Neural Network
(ANN) is a group of computational units interconnected via
weighted links known as perceptrons. Each unit contains one
or more weighted input connections, a transfer function
combining the inputs and an output connection [13].
Perceptron receives a signal, processes it and forwards the
result to other connected nodes. Three types of layers exist in
neural networks [13, 14]: 1) input - corresponds to the input
variables 2) hidden – resides between the input and output
layers 3) output - produces the output variables. A deep neural
network (DNN) is an artificial neural network (ANN) which
has multiple layers placed between the input and output layers
[14].
In this paper, we make use of deep learning in order to
predict number of new cases and deaths within city, necessary
for resource demand estimation. The prediction model is
trained on publicly available online data about COVID-19
spread and death cases from previous period. TensorFlow3 for
Python programming language with GPU-powered execution
on CUDA-compatible graphics hardware was used for
implementation.
III. IMPLEMENTATION
A. Solution Overview
First, user creates models of cities in terms of available
resources (medical workers, equipment, vehicles, patients,
3
https://www.tensorflow.org/
24
disinfection) using a visual modelling environment run in web
browser. A domain-specific notation is used for that purpose.
The modelling tool implementation is based on JavaScript,
HTML and CSS. Once the model is completed, it is
augmented by the data about expected resource demands
based on the results achieved relying on deep learning
prediction model. Online public data about number of new
infected people and deaths in countries and cities around the
world on daily basis from various sources is aggregated,
filtered and prepared for making new predictions. Moreover,
AMPL .dat file generator constructs appropriate input for
linear optimization process taking the previously created
model and predictions that are produced as result of deep
learning. After that the linear optimization according to the
AMPL optimization model is performed, giving the output
that represents optimal resource exchange which does not
only help the affected cities from unnecessary distribution
costs and spending, but also could lead to more lives saved. In
Fig. 1, an illustration of the proposed solution is shown.
Online data sources
3
Prediction
1
Modelling
2
4 Linear
6
environment optimization
AMPL .dat
Users generator
5
Savings
Fig. 1. Solution overview: 1-Model creation 2-Resource model 3-
Aggregated and filtered data from online data sources 4-Predictions
based on past data 5-AMPL data file 6-Optimization results
B. Modelling and Simulation Environment
A UML class diagram of the metamodel used for city
resource modeling notation is given in Fig. 2. A map consists
of cities that are able to exchange resources during the crisis.
The parameter describing the city connections is transport
cost. Relevant parameters for cities are the number of citizens,
its area size, number of hours allowed for movement.
Moreover, relevant parameters for pandemic situation are the
total number of cases and number of critical cases. Each city
has six types of resources: 1) hospital buildings 2) ambulance
vehicles 3) medical devices 4) medical workers 5) disinfection
products 6) food. An important resource feature is its price. At
each hospital, there is a limited capacity. At each moment, a
number of places within the hospital can be occupied by
patients. In case of COVID-19, respiratory equipment is
crucial for saving lives in critical cases. For that reason, it is
one of the most important resources exchanged between the
cities, apart from disinfection products which can lead to
reduction of new cases [15]. There is also possibility to set
weather parameters like temperature, rain and snow.
The implementation of modelling and simulation
environment is based on engine used for [9] and [16]. There
are two views within the tool: (a) map view and (b) city view.
Map view gives a global overview of the affected cities. City
view is available for each city from the map, by double-
clicking on it. In this view, it is possible to perform the
resource allocation for further exchange on a local or global
level. In case of foreign cities or countries included as
donation/help or resource exchange partners, they can be
added to the map and their resources set. In Fig. 3, a
screenshot of the proposed tool is given.
City: Pirot
Edit hospital_1 parameters
capacity 50
patients 7
CONFIRM
Fig. 3. Modelling and simulation tool screenshot: (a) map view
(b) city view
Moreover, several simulation mechanisms are
implemented. First, it is checked whether the number of
respiratory devices is greater or equal to the number of critical
patients:
num _ critical ≤ num _ respiratory
After that, it is checked if the number of medical workers is
enough to handle the number of expected new cases based on
estimation of how many patients each medic can handle:
expected _ cases
≤ num _ medics
patients _ daily _ per _ medic
A similar check like (2) is also done for ambulance
vehicles.
Finally, it is checked if there is enough free space to place
the patients in hospital:
expected _ cases ≤ hospital _ free _ places
If one these conditions denoted fails, then an alert occurs in
the simulation environment and corresponding resource
exchange process based on the selected linear optimization
strategy will be triggered.
if (conditioni is false) then optimization( strategyi ) (4)
C. Predictions Based on Deep Learning
The data aggregated from online sources was split into two
disjoint sets – the training set (75%) and test set (25%). It
consists of total eight variables, as it is shown in Table 1.
TABLE I
SPREAD PREDICTION DATASET
Id Density Age Day T. Mh New D.
[p/km2] [y] [0..366] [C] [h] [p] [p]
City id represents the identification number of the
considered area within the country. Number of persons per
km2 provides the information about population density within
the area. Median age is the age that divides a population into
two numerically equally sized sets, which might be of huge
importance for prediction of death cases. Day refers to the
ordinal number of a day (from 1 to 365 or 366). Temperature
is the average daily temperature that day which could affect
the movement of citizens that day. Instead of temperature,
other weather parameters might be leveraged, such as rainfall.
Movement hours is time frame during the day when the
citizens are allowed to go outside and may vary across
countries and cities, ranging from 0 (total quarantine) to 24
hours (regular situation). New cases refer to the number of
new infected people identified within the area, as an average
of 7 days which is approximately the average length of the
incubation period [3]. Deaths is the number of people that
died as consequence of COVID-19 disease that day. City id,
density, age, day number, temperature and movement hours
are treated as independent variables, while the number of new
cases and deaths are dependent variables (predicted values).
However, the information about population density, median
age and movement hours are optional, but the predictions
could also be made without them.
For training, publicly available data about the new cases
and deaths on national level around the world was used from
(1) [17] together with data for cities in Serbia from [18].
D. Linear Optimization Model
There are several optimization models for different resource
exchange strategies implemented within the tool (respirator,
(2) medical personnel, food, disinfection, vehicles and patients
exchange), but only one of them will be described.
Let us consider a network of interconnected cities involved
into resource trading, denoted as C. Travel cost tc[i,j] is
assigned to each link between city[i] and city[j]. Each city[i]
has predicted number of critical patients cp[i] and available
respirator devices r[i]. For each critical patient, a respirator
(3) device is needed. Each city[i] has respirators of price rp[i].
25
 Moreover, each city[i] has amount of budget b[i] that is the ACKNOWLEDGEMENT
maximum amount of money which can be spent. To each
connection between cities, a decision variable x[i,j] is This work has been supported by the Ministry of Education,
assigned to indicate the amount of respirators that will be sent Science and Technological Development of the Republic of
from city[i] to city[j]: Serbia.
x[i , j ] ≥ 0 if trading between cityi and city j will be perfromed , (5)
x[i , j ] = 0 otherwise
REFERENCES
However, there are several resource exchange constraints.
First, the overall amount of available respiratory devices after [1] V. C. Bala, P. Kumar, “Review on COVID-19: Rise of SARS-
exchange is always equal or greater than the number of CoV-2 Pandemic Outbreak”, Borneo Journal of Pharmacy,
critical patients. 3(Special-1), pp. 103-120, 2020.
[2] B. Cruz, M. Dias, “COVID-19: From outbreak to pandemic”,
r[i] +  x[i, j ] − x[ j, i] ≥cp[i], i ∈ C (6)
GSJ Vol. 8 Issue 3, March 2020, pp. 2230-2238.
j∈C
[3] D. Oran, E. Topol, Prevalence of Asymptomatic SARS-CoV-2
Moreover, the total cost of respirator acquisition should not Infection: A Narrative Review, Annals of Internal Medicine,
exceed the available budget: June 2020, pp. 1-7, 2020. https://doi.org/10.7326/M20-3012
[4] G. Spiteri et al., “First cases of coronavirus disease 2019
b[i] ≥  tc[i, j ] + x[ j, i]rp[ j ], i ∈ C (7) (COVID-19) in the WHO European Region, 24 January to 21
j∈C February 2020”, Eurosurveillance, pp. 1-6, 2020.
Finally, the objective function minimizes the overall sum of [5] X. Pan et al., “Potential drugs for the treatment of the novel
costs, considering both the travel cost and respirator price: coronavirus pneumonia (COVID-19) in China”, Virus Research,
Volume 286, 198057, pp. 1-12, 2020.
minmize  tc[i, j ] + x[i, j ]rp[i]
i , j∈C
(8) https://doi.org/10.1016/j.virusres.2020.198057
[6] P. Vanini, “Protection of the Population and the Economy in a
Pandemic” [preprint], pp. 1-21, 2020.
[7] D. M. Prieto et al., “A systematic review to identify areas of
IV. EVALUATION enhancements of pandemic simulation models for operational
use at provincial and local levels”, BMC Public Health, 12(1),
For evaluation, a laptop equipped with Intel i7 7700-HQ pp. 1-13, 2012. https://doi.org/10.1186/1471-2458-12-251
quad-core CPU running at 2.80GHz, GTX1050 GPU with [8] N. Ghaffarzadegan, H. Rahmandad, “Simulation-based
Estimation of the Spread of COVID-19 in Iran”, pp. 1-19, 2020.
2GB VRAM, 16GB of DDR4 RAM and 1TB HDD. In Table
[9] N. Petrović, Đ. Kocić, “Data-driven Framework for Energy-
2, the evaluation results in respirator device exchange scenario Efficient Smart Cities”, Serbian Journal of Electrical
for different number of cities involved are given, considering Engineering, Vol. 17, No. 1, Feb. 2020, pp. 41-63.
the execution time for different processing steps – prediction https://doi.org/10.2298/SJEE2001041P
and optimization (in seconds) and cost reduction (as [10] N. Petrović, S. Koničanin, D. Milić, S. Suljović, S. Panić,
percentage). According to the results, the processing time “GPU-enabled Framework for Modelling, Simulation and
increases with the model size, but it does not exceed one Planning of Mobile Networks in Smart Cities”, ZINC 2020, pp.
second in the executed experiments. On the other side, the 1-6, 2020.
reduction of resource exchange costs depends on the problem [11] G. B. Dantzig and M. N. Thapa, Linear Programming 1 -
Introduction, Springer, 1997.
instance itself and varies from case to case. When it comes to
[12] G. B. Dantzig, A. Orden and P. Wolfe, “The generalized
prediction performance, the relative error was 24%, while simplex method for minimizing a linear form under linear
AUC score was 0.81. inequality restraints”, Pacific Journal of Mathematics, Volume
5, Number 2, 1955, pp. 183-195.
TABLE II [13] R. Lippmann, “An introduction to computing with neural nets”,
EVALUATION RESULTS IEEE ASSP Magazine, vol.4(2), 1987, pp. 4-22.
[14] Y. Bengio, “Learning Deep Architectures for AI”, Foundations
Size Tprediction Toptimization Creduction and Trends in Machine Learning 2(1), 2009, pp. 1-127.
[cities] [s] [s] [%] [15] H. Fathizadeh et al., “Protection and disinfection policies
3 0.29 0.016 31 against SARS-CoV-2 (COVID-19)”, Le Infezioni in Medicina
5 0.31 0.021 28 n. 2, April 2020, pp. 185-191.
[16] N. Petrović, “Approach to Dynamic Adaptivity Simulation in
7 0.38 0.059 42
Fog Computing Scenarios”, TELSIKS 2019, pp. 58-61, 2019.
10 0.51 0.084 63 https://doi.org/10.1109/TELSIKS46999.2019.9002322
[17] Coronavirus Update (Live) [online]. Available on:
https://www.worldometers.info/coronavirus/
V. CONCLUSION [18] Statistika COVID-19 u Srbiji [online]. Available on:
https://covid19.data.gov.rs/
The proposed approach gives promising results, enabling
the timely allocation of resources in a pandemic crisis
situation. However, the tool is still a work in progress and
more data about new COVID-19 cases in Serbia is needed for
better prediction performance.

26
 Tackling the COVID-19 Conspiracies:
The Data-Driven Approach
Nenad Petrović
Abstract – Apart from obvious impact on many aspects of our
everyday lives, the current COVID-19 pandemic has led to
occurrence many conspiracy theories as side-effect, especially via
social media. However, in such confusion among population, the
situation becomes even more difficult to handle, which could
potentially have even more dramatic consequences on global
economy and increase the number of lives lost. In this paper, a
data-driven framework for analysis of the facts related to
COVID-19 disease is proposed. The proposed implementation
leverages semantic knowledge representation, fuzzy reasoning
and correlation analysis.
In this paper, a data-driven approach is adopted to tackle
the conspiracies in context of COVID-19 disease, leveraging
fuzzy reasoning, performed on facts stored within the
semantic knowledge base created as a result of ontology
learning against open scientific literature and making use of
correlation analysis of public data sets for truth value
determination. To the best of author’s knowledge, such
solution has not been publicly presented yet or documented in
published scientific literature.

Keywords – Coronavirus, Correlation analysis, Fuzzy

II. BACKGROUND AND RELATED WORK
reasoning, Ontology, Semantic knowledge representation.
A. Fuzzy Logic
I. INTRODUCTION Fuzzy logic refers to a form of multi-valued logic where the
truth values, apart from 0 and 1 could be any real number
Since the beginning of this year, the pandemic of novel
between 0 and 1. Its goal is to handle the concept of partial
infectious disease COVID-19 has affected almost all aspects
truth, where the truth value may range between completely
of our everyday lives. It was first discovered in China, but
true and completely false. It is similar to human reasoning as
spread quickly to other continents in just few weeks [1-3].
it relies on imprecise information to make decisions.
According to [4], until June 15th, 2020, the total number of
Traditional logic which requires deeper understanding of a
identified cases was 8,039,649, while the disease has taken
system, such as exact equations and precise numeric values.
436,498 lives worldwide. However, due to huge number of
On the other side, fuzzy logic enables modeling complex
cases in some countries, several measures were taken by
systems using a higher level of abstraction originating from
governments in order to reduce to disease spread, such as
human knowledge and experience [12, 13]. In many cases,
social distancing, limiting citizens’ movement within country
control of complex systems by experienced human operators
boarders and abroad, often together with prohibition and
is more successful and effective than automatic methods.
cancellation of huge public events [5]. Therefore, from work
Fuzzy reasoning has been adopted in many complex systems
to personal relations and entertainment, COVID-19 pandemic
that cannot be modelled precisely even under various
has brought many changes to our daily routine, habits and
assumptions and approximations incorporating the human
activities. Despite the fact that the pandemic seems weaker
expert knowledge [13].
nowadays, some of the government measures are still applied
The medical decision-making process is fuzzy in its nature
in some countries due to unstable situation.
[14]. During this process, the doctors need to reason about
However, one of the crucial problems related to COVID-19
linguistic concepts in order to bring a decision about diagnosis
which led to dramatic consequences are the speed of its spread
or make prognosis. However, the conversion from fuzzy
and lack of approved vaccine and medication [6]. for the
nature into outcome often leads to loss of precision [14],
disease. Apart from huge number of lost lives and stagnation
which could have dramatic consequences, when it comes to
of economy as well [7], the situation becomes even more
patient treatment. For that reason, fuzzy logic is a suitable
difficult due to spread and circulation of conspiracy theories
way to provide the doctors with the support needed for
worldwide [8, 9], especially via social networks [10]. One of
handling of linguistic concepts and avoiding such loss of
the most common conspiracy theories is the relation between
precision. Fuzzy logic technologies are applied across many
the deployment of 5G network and COVID-19 spread [8-10],
areas of healthcare, and they have been approved as quite
which even lead to organized vandalism against 5G
effective [14, 15]. For example, in [16], a fuzzy medical
equipment [11].
diagnosis expert system based on quantiles of diagnostic
measures was presented, while the work presented in [17]
Nenad Petrović is with University of Niš, Faculty of Electronic focuses on coronary heart disease risk determination based on
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail: cholesterol measurements.
nenad.petrovic@elfak.ni.ac.rs

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 27

 In this paper, fuzzy reasoning is applied in wider context, in
order to take into account the partial truth of novel facts about
COVID-19 available online within the decision-making
process. However, the proposed approach can be also applied
to support medical decision-making when it comes to
COVID-19 patient assessment and treatment in order to
eliminate any possible conspiracies that can have impact on
treatment.
B. Correlation Analysis
Correlation analysis refers to a statistical method which is
used for discovery of a relationship between two variables and
how strong that relationship is [18]. Correlation coefficient
represents a numerical measure of correlation between the
relative movements of two variables from the observed data
set. Its value is in range between -1.0 and 1.0. Positive value
of correlation means that for positive increase in one variable,
there is also a positive increase in the second variable. On the
other side, negative correlation value indicates that the
variables move in opposite directions - for a positive increase
of one variable, the second variable decreases. In case of
correlation value which is exactly 0, there is no linear
relationship between the observed variables. There are many
types of different correlation coefficients, but the most
commonly used is the Pearson correlation coefficient (denoted
as r) that measures the strength and direction of the linear
relationship between two variables [18]. However, it is not
able to capture nonlinear relationships between two variables
and distinguish between dependent and independent variables.
There are several works that adopt correlation coefficient to
discover relationships between some factors related to
COVID-19 pandemic. In [19], the analysis of the correlation
between confirmed cases of COVID-19 in Spain and several
geographic, meteorological and socioeconomic variables at
the province level was presented. On the other side, in [20], a
correlation between racial demographics and COVID-19 cases
was analyzed.
In this paper, MATLAB’s implementation of Pearson
correlation known as corcoeff1 was used to calculate the value
of correlation between two variables relevant to COVID-19
based on publicly available online data sets. Moreover, the
calculated value is used as truth value for facts stating that
these two observations are related, which is also leveraged
within fuzzy reasoning process relying on semantic
representation of these facts based on ontologies.
C. Semantic Knowledge Representation
Semantic technology enables encoding the meaning of data
separately from the content itself and related applications,
providing the ability to understand data, exchange that
understanding and perform reasoning on top of it. The
formalization of this knowledge representation is made in a
form understandable and suitable for use by both humans and
computers, which is an advantage for usage in reasoning-
1 2
https://www.mathworks.com/help/matlab/ref/corrcoef.html
28
based systems. Within the semantic knowledge bases, the data
is represented with respect to ontologies. An ontology consists
of classes, individuals, attributes and relations. Classes
represent abstract groups, collections or types of objects.
Individuals are instances of these classes. Attributes refer to
properties, characteristics and parameters of classes. Relations
define ways in which classes and individuals are related. For
both the ontologies and facts, the RDF standard language is
used. It is used to define (subject, predicate, object) triplets
that are persisted within the semantic triple stores. On the
other side, SPARQL is used for query execution against the
semantic triple stores. The results are retrieved in order to
support different reasoning mechanisms that enable inference
of new knowledge and facts from the existing knowledge
base.
In [21], ontologies were used for representation of
pandemic simulation scenarios in case of H5N1. On the other
side, in [22], Epidemiology Ontology was introduced to
enable sharing of epidemiology resources within semantic-
enabled platforms. Recently, in [23], an extension of the
Infectious Disease Ontology was introduced, including the
concepts relevant to coronavirus pandemic.
In this paper, semantic knowledge representation is used for
facts about coronavirus pandemic coming from various online
textual sources - open-access scientific paper repositories
(such as PubMed2) and web pages. The transformation of
textual data to semantic representation is performed relying on
ontology learning module built upon the work presented in
[24]. Moreover, an ontology is also used for representation of
fuzzy rules used by the mechanism of reasoning itself, in a
similar way as it was done in [25] to support explainable
artificial intelligence based on fuzzy reasoning in smart cities.
III. IMPLEMENTATION OVERVIEW
In the first step, the user specifies the desired hypothesis
about COVID-19 disease that will be tested using a graphical
user interface in web browser. The tested hypothesis is given
either in form of (condition set->consequence) or pair of
potentially connected concepts that are in relation r(concept1,
concept2). A condition set consists of a set of statements
connected with and - (condition1 ˄ condition2 ˄ …
conditionn). After that, the key terms are extracted from
provided user-defined hypothesis elements (term1 is extracted
from concept1 or term2 is extracted from condition2, for
example) and forwarded as input to the search and retrieval
module. This module has role to execute search queries for
key terms against repositories of publicly available scientific
texts or extract numerical statistic data related to COVID-19
pandemic from web pages. Furthermore, the retrieved raw
data is filtered and stored into a form suitable for further steps.
Moreover, the prepared data is used as input for two different
processing modules. While textual data is forwarded to
ontology learning module to construct a semantic knowledge
base about given terms, the numerical data is leveraged for
correlation analysis by calculating the correlation coefficient
between the data about input pair of concepts.
https://pubmed.ncbi.nlm.nih.gov/
 Public online data
3

Correlation
analysis 6
Search and
1 User interface 2 4 7 Fuzzy reasoning 8
retrieval
Ontology
User 5 Semantic knowledge
learning
representation Results

Fig. 1. Framework overview: 1-User-defined hypothesis 2-Key words for terms from hypothesis 3-Raw textual or numerical data about
key terms 4-Data prepared for processing 5-Representation of facts extracted from text in semantic triplet form 6-Correlation coefficient
values 7-Facts for given terms found in semantic knowledge base 8-Interpretation of results obtained as output of fuzzy reasoning process

The truth value of facts stored within the semantic have truth value above thresholdconspiracy. Finally, the results
knowledge base takes a real number value from range [0, 1] achieved as outcome of the described reasoning process are
and is calculated in case of concept relation testing according interpreted (whether hypothesis is considered as a conspiracy)
to formula: and shown to the user via GUI.
n( t1 ,t2 ) In Fig. 1, an overview of the proposed framework is
+ correlation _ coeff (t1 , t2 ) summarized. Moreover, in Listing I, pseudocode of the
ntexts (1)
Tt1 ,t2 = conspiracy test based on fuzzy reasoning is given.
2
In Eq. (1), n(t1,t2) denotes the number of texts where the Input: hypothesis, thresholdconspiracy
Output: true/false
relation between the terms extracted from concepts given in Steps:
the hypothesis is present; ntexts is the total number of texts 1. Extract key terms from hypothesis;
returned for given terms; correlation_coeff(t1, t2) refers to 2. For each term t in set of key terms
correlation coefficient between the variables relevant to these 3. Retrieve texts from online sources about term t;
4. term facts:=Perform ontology learning about term t based on texts;
terms extracted from public datasets. In case that search 5. Add term facts to the collection of all facts;
module fails to find the relevant datasets for correlation 6. Retrieve data set related to term t if possible;
analysis, than the truth value is not calculated as average, but 7. Discover transitive relations;
only taking into account the percentage of texts where the 8. Calculate truth based on (1);
9. End for each;
relation from the hypothesis is present (n(t1,t2)/ ntexts), instead. 10. conspiracy:=false;
On the other side, in formula for the other type of hypotheses 11. For each fact f in set of facts
considered, t1 represents a condition from the condition set, 12. If (f.truth < thresholdconspiracy)
while t2 refers to the consequence. Moreover, the relevance of 13. conspiracy:=true;
14. End if;
public online data can be taken into account in order to reduce 15. End for each;
the impact of low-quality publications by including a 16. return not(conspiracy);
trustworthiness coefficient of data source where the 17. End
considered fact was collected, denoted as trusti:
N Listing 1. Pseudocode of COVID-19 conspiracy testing algorithm
 trust ⋅ n i ( t1 ,t 2 )
i =1 (2)
ntexts IV. EXPERIMENTS AND EVALUATION
For this value, either impact factor of scientific publication
can be taken into account or some other metric, such as page For evaluation, a laptop equipped with Intel i7 7700-HQ
rank or relevance [26]. Once the truth values for the collected quad-core CPU running at 2.80GHz, equipped with 16GB of
facts used are ready, the fuzzy reasoning process can begin. It DDR4 RAM and 1TB HDD. In Table 1, the processing times
executes a set of SPARQL queries against the semantic of various steps achieved in several experiments of hypothesis
knowledge base to check whether the tested conditions exist testing are given. PubMed papers were used as basis for
and retrieves their truth values. However, after that, there are textual corpus by the ontology learning module, while the
two possible flows, depending on type of hypothesis testing. numerical statistics used for correlation relied mostly on [4],
In case of concept relation testing, if the truth value of facts together with several pages from [27]. The value of
describing the assumed relationship between the concepts c1 thresholdconspiracy was 0.8 in all cases.
and c2 is below given threshold thresholdconspiracy (a real- The presented framework was able to correctly distinguish
valued number between 0 and 1), than the tested hypothesis is between conspiracy (increased investment in 5G increases
considered as a conspiracy. On the other side, when it comes COVID-19 cases) and valid scientific results (death rate is
to the other type of hypotheses in form of condition set- higher among elderly and persons with chronic diseases). As
>consequence, then the hypothesis passes the conspiracy test it can be noticed, the processing time increases with textual
only in case when all the facts that represent the relation corpus size and data set size in tests where correlation analysis
between a conditioni from the condition set and consequence was performed, as expected.

29
 TABLE I [9] M. Uthman et al., “5G Radiation and COVID-19: The Non-
EXPERIMENT RESULTS Existent Connection”, International Journal of Research im
Electronics and Computer Engineering, vol. 8, no. 2, April-June
H Ntexts Tontology Tcorrel Treason Out 2020, pp. 34-38.
[s] [s] [s] [10] W. Ahmed et al., “COVID-19 and the 5G Conspiracy Theory:
increase 10 98 0.13 0.86 false Social Network Analysis of Twitter Data”, J Med Internet Res
(5G, cases) 2020; 22(5):e19458, pp. 1-9, 2020.
[11] Cell Tower Vandals and Re-open Protestors — Why Some
old->deaths 20 147 0.25 0.91 true People Believe in Coronavirus Conspiracies [online]. Available
chronic- 30 289 - 1.01 true on: https://theconversation.com/cell-tower-vandals-and-re-open-
>deaths protestors-why-some-people-believe-in-coronavirus-
conspiracies-138192
[12] S. G. Tzafestas, N. Venetsanopoulos et al., “Fuzzy Reasoning in
V. CONCLUSION AND FUTURE WORK Information, Decision and Control Systems”, Kluwer Academic
Publishers, 1994, https://doi.org/10.1007/978-0-585-34652-6
With respect to the initial results, it can be recognized that [13] C. Carlsson, R. Fullér, “Fuzzy Reasoning in Decision Making
the presented approach has huge potential for adoption within and Optimization. Studies in Fuzziness and Soft Computing”,
various systems and use cases. In future, it is planned to Springer, 2002, https://doi.org/10.1007/978-3-7908-1805-5
integrate it within medical expert systems based on fuzzy [14] G. Gursel, “Fuzzy Logic in Healthcare”, Medical Imaging:
reasoning, such as the one presented in [17], with purpose of Concepts, Methodologies, Tools, and Applications, pp. 1-29,
decisioning improvement, especially when it comes to impact 2017, https://doi.org/10.4018/978-1-5225-0571-6.ch006
[15] K. B. Sundharakumar, S. Dhivya, S. Mohanavalli, R. Vinob
of COVID-19 on patients with chronic heart diseases. Finally, Chander, “Cloud Based Fuzzy Healthcare System”, Procedia
the presented solution will be integrated within the framework Computer Science, vol. 50, 2015, pp. 143-148.
for efficient resource planning in pandemic crisis with goal of [16] H. Choi, K. Han, K. Choi, and J. Y. Ahn, “A Fuzzy Medical
patient risk assessment [28] and leveraged by mobile app [29]. Diagnosis Based on Quantiles of Diagnostic Measures”, Journal
of Intelligent and Fuzzy Systems, vol. 31, no. 6, pp. 3197-3202,
ACKNOWLEDGEMENT 2016.
[17] N. Allahverdi, S. Torun, and I. Saritas, “Design of a Fuzzy
Expert System for Determination of Coronary Heart Disease
This work has been supported by the Ministry of Education, Risk”, CompSysTech’07, pp. 1-8, 2007.
Science and Technological Development of the Republic of [18] S. Senthilnathan, “Usefulness of Correlation Analysis”, SSRN
Serbia. Electronic Journal, pp. 1-9, 2019.
[19] D. Oto-Peralías, “Regional Correlations of COVID-19 in Spain”
REFERENCES [preprint], pp. 1-28, 2020, https://doi.org/10.31219/osf.io/tjdgw
[20] U. V. Mahajan, M. Larkins-Pettigrew, “Racial Demographics
and COVID-19 Confirmed Cases and Deaths: A Correlational
[1] G. Spiteri et al., “First Cases of Coronavirus Disease 2019
Analysis of 2886 US Counties”, Journal of Public Health, pp. 1-
(COVID-19) in the WHO European Region, 24 January to 21
4, 2020, https://doi.org/10.1093/pubmed/fdaa070
February 2020”, Eurosurveillance, pp. 1-6, 2020.
[21] H. Eriksson et al., “Ontology Based Modeling of Pandemic
[2] “Cluster of Pneumonia Cases Caused by a Novel Coronavirus,
Simulation Scenarios”, Studies in health technology and
Wuhan, China” [online], European Centre for Disease
informatics, 129(Pt 1), pp. 755-759, 2007.
Prevention and Control, Stockholm, pp. 1-10, 2020.
[22] C. Pesquita, J. Ferreira, F. Couto, and M. Silva, “The
https://www.ecdc.europa.eu/sites/default/files/documents/Risk%
Epidemiology Ontology: An Ontology for the Semantic
20assessment%20-
Annotation of Epidemiological Resources”, Journal of
%20pneumonia%20Wuhan%20China%2017%20Jan%202020.p
Biomedical Semantics, 2014, 5:4, pp. 1-7.
df
[23] S. M. Babcock, J. Beverley, L. G. Cowell, and B. Smith, “The
[3] B. Cruz, M. Dias, “COVID-19: From Outbreak to Pandemic”,
Infectious Disease Ontology in the Age of COVID-19”
GSJ, vol. 8, no. 3, March 2020, pp. 2230-2238.
[preprint], pp. 1-33, 2020, https://doi.org/10.31219/osf.io/az6u5
[4] Coronavirus Update (Live) [online]. Available on:
[24] N. Petrovic, M. Tosic, “Never-Ending Ontology Learning
https://www.worldometers.info/coronavirus/
Approach Applied to Biomolecular Function Prediction”,
[5] Considerations Relating to Social Distancing Measures in
IcETRAN 2019, pp. 762-767, 2019.
Response to COVID-19 – Second Update [online]. Available:
[25] N. Petrović, M. Tošić, “Explainable Artificial Intelligence and
https://www.ecdc.europa.eu/sites/default/files/documents/covid
Reasoning in Smart Cities”, YuInfo 2020, pp. 1-6, 2020.
19-social-distancing-measuresg-guide-second-update.pdf
[26] H. H. Musa, N. Abdelrhman, “Comparing the Ranking
[6] Y. Song et al., “COVID-19 Treatment: Close to a Cure? – A
Performance of Page Rank Algorithm and Weighted Page Rank
Rapid Review of Pharmacotherapies for the Novel Coronavirus”
Algorithm”, Journal of Computational and Theoretical
[preprint], pp. 1-25, 2020,
Nanoscience, vol. 24, no. 1, pp. 750-753, 2018.
https://doi.org/10.20944/preprints202003.0378.v1
[27] 5G - Statistics & Facts [online]. Available on:
[7] P. Vanini, “Protection of the Population and the Economy in a
https://www.statista.com/topics/3447/5g/
Pandemic” [preprint], pp. 1-21, 2020.
[28] N. Petrović, “Simulation Environment for Optimal Resource
[8] D. Freeman et al., “Coronavirus Conspiracy Beliefs, Mistrust,
Planning During COVID-19 Crisis”, ICEST 2020.
and Compliance with Government Guidelines in England”,
[29] N. Petrović, M. Radenković, and V. Nejković, “Data-Driven
Psychological Medicine, pp. 1–13, 2020.
Mobile Applications Based on AppSheet as Support in COVID-
https://doi.org/10.1017/S0033291720001890
19 Crisis”, IcETRAN 2020, pp. 1-6, 2020.

30
 COMPUTER SYSTEMS
AND INTERNET TECHNOLOGIES II
 Augmented and Virtual Reality Web Applications for
Music Stage Performance
Nenad Petrović
Abstract – In this paper, it is explored how augmented and stage performance by generating 3D worlds in virtual reality
virtual reality web applications can contribute to stage perceived by the audience.
performance in audio-visual arts. As outcome, application case
study is presented: augmented reality loop sampler interface
integrated with virtual reality song visualizer. Moreover, a II. BACKGROUND AND RELATED WORK
modelling tool leveraging metamodel and code generation for
rapid creation of custom augmented/virtual reality applications A. Augmented Reality (AR)
is introduced. According to the achieved results, the proposed
approach significantly speeds up the development of such Augmented reality (AR) enables merging real scenes with
applications. interactive virtual objects, such as 3D geometric models,
images, sounds and texts in a way that these objects overlap
Keywords – Augmented reality (AR), Metamodeling, Media
with physical world [7].
arts, Music, Virtual reality (VR).
Different types of AR-based music instruments and
interfaces for various purposes have been proposed so far. In
I. INTRODUCTION [4], a solution which enables playing electornic music
compositions interactively, remixing and modulating their
Each technological revolution has affected how music is elements by manipulation of simple physical objects was
created, performed and reproduced [1]. For example, in the presented. Furthermore, a more advanced mobile web solution
early 80s, the rise of digital synthesizers, personal computers for music perfromance which enables AR-based interaction
and MIDI standard has changed the way how music was with audience was introduced in [8]. On the other side, in [7],
created forever by enabling high level of automation in music an AR musical game for cognitive and motor rehabilitation
performance and high precision sequencing [2]. Later, in the was shown. It uses printed markers for each tone and patient
mid 90s, home computers have become powerful enough to
has to repeat the given sequence.
handle heavier sound processing tasks within digital audio
workstation (DAW) substituting hundreds of hardware racks,
B. Virtual Reality (VR)
making the sound editing and music production more
comfortable than ever before [3].
Virtual reality (VR) provides a feeling of presence in virtual
However, it was not exception with augmented reality (AR)
environments where user can interact with objects using
and virtual reality (VR) technologies [4, 5] which have
different sensorial channels [4].
emerged in recent years. The main novelty in music art
Several different usage scenarios of VR in music are
introduced by AR consists in using touch and gestures
identified. In [9], a system providing the ability to listen to
together with physical objects to provide innovative music
music and play different instruments in a 3D environment
interfaces and additional layers of expressivity compared to
with real-world panoramic video as background scenes was
traditional music instruments [4]. On the other side, VR
presented. Furthermore, in [10], the authors developed a VR
enables virtual presence in music-related scenarios, making
orchestral application for interactive music experience,
use of novel input methods, such as biofeedback [5].
allowing virtual musical instruments to be repositioned
In this paper, a framework for automated generation of AR
spatially, dynamically and interactively in virtual space.
and VR applications aiming music stage performance based
Finally, in [11], a museum of interactive music instruments in
on a metamodel is presented. It is inspired by metmodelling
virtual reality was shown.
framework for Future Internet infrastructure code generation
presented in [6]. The proposed approach is tested and
C. Web AR and VR Technology
evaluated in two case studies: AR loop sampler interface and
VR song visualizer. The first one enables performing DJ sets
Web is an emerging platform for lightweight augmented
by mixing sound loops, while the second augments the music
and virtual reality services and applications [12]. On the other
side, vendor-specific platforms and native applications have
Nenad Petrović is with University of Niš, Faculty of Electronic several drawbacks, such as limited compatibility, they often
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
require additional hardware accessories required and cost too
nenad.petrovic@elfak.ni.ac.rs
much [12]. For that reason, it was decided to rely on
JavaScript libraries for development of web AR and VR
applications presented in this paper.

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 33

 Fig. 2. Metamodel used for AR/VR application design notation
AR.js1 is an open-source JavaScript library for development
of augmented reality applications which are run in a web
browser. It provides wide compatibility and solid performance
on mobile devices without high hardware requirements,
performing at impressive 60 frames per second, even on older
smartphones [13]. It is quite simple and intuitive, making the
development of full augmented reality applications in just
several lines of HTML and JavaScript code [13]. It includes a
huge set of built-in features and is integrated with other
JavaScript libraries that enable loading 3D objects, position
and location handling, usage of different camera marker
presets and many others.
On the other side, Three.js2 is an open-source, JavaScript-
based library and API which enables rendering rich animated
3D computer graphics inside a web browser. It relies on
WebGL3, but offers a higher-level interface to advanced
capabilities, making construction of complex multimedia
applications much easier, such as loading different types of 3D
models and their animation [14]. On top of that, Three.js
enables easy conversion of traditional 3D scenes into VR
applications4.
III. IMPLEMENTATION
A. Framework Overview
In this paper, a framework for rapid AR/VR application
development (aiming mostly stage performance applications)
is proposed. An illustration of the proposed framework is
given in Fig. 1. User draws a diagram which describes
behaviour of the desired AR/VR application for music stage
performance. For that purpose, a domain-specific visual
notation is used. Once the modelling is done, the constructed
1
https://github.com/AR-js-org/AR.js
2
https://threejs.org/
3
https://www.khronos.org/webgl/wiki/Getting_Started
4 5
https://threejs.org/docs/#manual/en/introduction/How-to-create-VR-content
34
application model is forwarded in JSON format to code
generator. Moreover, the code generator parses the received
JSON application model in order to extract the necessary
information used for code generation. Finally, the code of the
desired AR/VR application is generated as HTML page with
JavaScript code relying on AR.js and/or Three.js.
AR/VR app
1 Code generator3
modelling tool 2
User AR/VR app
Fig. 1. Framework overview: 1-Drawing application model 2-
Application model in JSON format 3-HTML page with JavaScript
code
B. Application Modelling Environment
A domain-specific notation is used for modelling AR/VR
applications, described as a UML metamodel shown in Fig. 2.
As it can be seen, there are two types of application: AR
and VR. Depending on application type the final output will
be either AR.js or Three.js-supported JavaScript code packed
inside HTML page. Each application consists of (input-
>output) mappings. There are several input methods which
are supported: printed markers (QR, barcode or other), spoken
phrases (relying on speech-to-text framework in JavaScript
[15]), keypress events or specific messages received from
server. In case when server messages are used as input
method, the address of that server has to be set. On the other
side, there are possible outputs which occur when specified
input is detected: sound file, 3D object model or text. The
sound can be one-shot or loop, while 3D object and text can
be either static, rotating or include custom animation. Several
formats for both sounds and 3D models are supported.
However, the messages can be also used as output in cases
when message has to be sent when specific input occurs. The
corresponding code for messages leverages AJAX5 which
https://www.w3schools.com/xml/ajax_intro.asp
allows updating the content without reloading the page by or other) is added to the application code. Furthermore, for
sending the requests or receiving page updates in the each of the possible outputs for the considered input method, a
background. code that produces the desired type of output is appended
The visual modelling tool using the described notation inside the event handler for specific input value (for example
based on metamodel is developed within Node-RED6 barcode “1” or keypress of “a”). For sounds and 3D models it
environment, as in [6]. For that purpose, custom nodes were is checked whether loop/rotate parameters are set. Once all the
created. In Fig. 3, a screenshot of the modelling tool is given. mappings present within the model are traversed, the final
The illustrated example shows AR application model that code is stored in a HTML page file.
consists of two mappings for barcode markers, where each of
them produces different looping sound and rotating object as IV. CASE STUDY
output. The model is exported in JSON format.
Edit sound node
Modelling elements sound
path penenadpi/synth In this paper, a case study application combines the
marker
marker
format .wav utilization of both augmented and and virtual reality for music
speech
object loop true
stage performance. It consists of two distinct parts: AR music
sound mixing interface and VR song visualizer. An overview of the
message
sound application is given in Fig. 4.
marker
object

object Loop 2
Fig. 3. Screenshot of AR/VR application modelling environment 1 2A
based on Node-RED
AR app Sound mix Loop 3
C. Code Generation Loop 1 Loop 2 Loop 3 2B

Performer
In Listing 1, pseudocode of the code generation algorithm is Server
given.
Input: JSON application model
Output: HTML page with AR.js/VR.js code 3
Steps:
1. Parse JSON application model;
2. Get all mappings from model; 4 Audience
3. For each mapping in model.mappings VR app
4. Open event handler template for mapping.input;
5. Use mapping.input as condition for handler trigger;
Fig. 4. AR/VR web application for music stage performance: 1-Loop
6. Get possible outputs for given mapping.input; play/stop event 2a-Loop added or removed from mix 2b-Event
7. For each output in mapping.outputs message sent to server 3-VR application code generation 4-Virtual
8. If(mapping.output instance of object) reality scene
9. If(object.animation is rotate)
10. Generate code for loading and rotating 3D model; The AR-based for loop sampling enables live mixing of
11. Else
12. Generate code for loading 3D model;
multiple sound loops. Its target users are music performers. In
13. End if; order to use the AR application, a paper with barcode markers
14. End if; is needed. By pointing camera to the paper, a 3D cube appears
15. … over each of the makers. Each of the cubes has different color
16. If(mapping.output instance of sound)
17. If(sound.animation is loop)
to make the distinction of sound loops easier to the user. A
18. Generate code for loading and looping sound; sound loop is assigned to each of the markers. The sound loop
19. Else is included or excluded from the generated music mix by
20. Generate code for loading 3D model; placing hands over the corresponding barcode marker. When
21. End if;
22. End if;
loop is stopped, the corresponding cube disappears from the
23. If(mapping.output instance of message) view (as shown in Fig. 5).
24. Generate code for AJAX call to send message.value
25. End if;
26. End for each
27. Close event handler template;
28. End for each
29. End
Listing 1. Algorithm for code generation starting from application model

First, the JSON application model exported from modelling

environment is parsed in order to retrieve all the mappings
from application model. After that, for each mapping’s input, Fig. 5. A screenshot of the author playing his song “Relayman” using
a corresponding event handler (marker read, keypress, speech the AR application7

6 7
https://nodered.org/ https://www.youtube.com/watch?v=LL2yGAJ06UQ

35
 Moreover, each time different loop play/stop event is
triggered, a message about the event is sent to the server. The
role of this server is to accept the messages generated by the
AR application and to generate the desired virtual reality
scene for each loop combination. On the other side, the VR
application aims the audience. The VR application code is
delivered by server to them within their web browser, giving
them the impression of presence in a virtual scene which
depends on the music played by the performer. This way, VR
application gives the provides the ability to generate dynamic
song visualization in realtime. An example of VR scene
shown to the audience when synth loop is played is given in
Fig. 6.
Fig. 6. Synth sequence VR scene rendered using Three.js
V. EVALUATION
A laptop equipped with Intel i7 7700-HQ quad-core CPU
running at 2.80GHz, 16GB of DDR4 RAM and 1TB HDD
was used for evaluation. In Table 1, the evaluation results for
case study application are given. Time needed for each of the
steps necessary for case study application creation is
considered and compared to time needed for manual
development of such applications without using the proposed
framework. Both AR and VR parts of the case study
application were considered separately. The first column
denotes application’s name. The second column shows the
time needed for manual creation of application model using
the tool. Moreover, the third column is the time needed to
transform the corresponding application model to functional
AR/VR web application code. Finally, the last column is the
approximate time needed for entirely manual creation of case
study applications, without relying on the framework
proposed in this paper.
TABLE I
EVALUATION RESULTS
App. Model Generation Manual
[s] [s] [h]
AR 339 0.73 5.5
VR 621 1.28 9 Art, Challenges, and Insights”, Proceedings of the IEEE, pp. 1–
VI. CONCLUSION
In this paper, a framework for automated generation of
AR/VR applications aiming music stage performance is given.
The proposed approach dramatically reduces the time needed
for development of such applications – from several hours to
36
just several minutes. Moreover, an example application for
music stage performance leveraging both augmented and
virtual reality was presented. This application enables
affordable and convenient live performance even on distance
(without any additional dedicated hardware), which has
become an important aspect during the recent COVID-19
pandemic crisis due to social distancing which led to
cancellation of music events worldwide.
ACKNOWLEDGEMENT
This work has been supported by the Ministry of Education,
Science and Technological Development of the Republic of
Serbia.
REFERENCES
[1] I. Poupyrev, R. Berry, M. Billinghurst, H. Kato, K. Nakao, L.
Baldwin, and J. Kurumisawa, “Augmented Reality Interface for
Electronic Music Performance”, 9th International Conference
on Human-Computer Interaction, pp. 805-808, 2001.
[2] C. Yavelow, “The Impact of MIDI upon Compositional
Methodology”, ICMC 86, pp. 21-26, 1986.
[3] J. M. Eargle, “An Overview of Digital Audio Workstations”,
Handbook of Recording Engineering, pp. 439-446, 1996.
[4] I. Poupyrev, R. Berry, M. Billinghurst, H. Kato, K. Nakao, L.
Baldwin, and J. Kurumisawa, “Augmented Reality Interface for
Electronic Music Performance”, 9th International Conference
on Human-Computer Interaction, pp. 805-808, 2001.
[5] P. Jolivet, P. Oomen, G. Pribek, “Large Room-scale Virtual
Reality in Audio Visual Arts using Sensors and Biofeedback”,
LINKs serie 3-4, Spatial Sound Institute, pp. 168-179, 2019.
[6] N. Petrovic, M. Tosic, “SMADA-Fog: Semantic model driven
approach to deployment and adaptivity in Fog Computing”,
Simulation Modelling Practice and Theory, 102033, pp. 1-25,
2019. https://doi.org/10.1016/j.simpat.2019.102033
[7] A. G. D. Correa, G. A. de Assis, M. Do Nascimento, I.
Ficheman, R. de D. Lopes, “GenVirtual: An Augmented Reality
Musical Game for Cognitive and Motor Rehabilitation”, 2007
Virtual Rehabilitation, pp. 1-6, 2007.
[8] K. Nishida, A. Yuguchi, K. Jo, P. Modler, M. Noisternig,
“Border: A Live Performance Based on Web AR and a Gesture-
Controlled Virtual Instrument”, International Conference on
New Interfaces for Musical Expression, pp. 43-46, 2019.
[9] X. Lin et al., “Virtual Reality-Based Musical Therapy for
Mental Health Management”, CCWC 2020, pp. 1-5, 2020.
[10] Y. Zhang et al., “Interactive Virtual Reality Orchestral Music”,
SIGGRAPH 2019, pp. 1-2, 2019.
[11] I. A. Zolkifly, L. K. Cheng, Z. A. Sabot, “Virtual Reality
Musical Museum”, Annual International Conferences on
Computer Games, Multimedia and Allied Technology, pp. 29-
34, 2008.
[12] X. Qiao, P. Ren, S. Dustdar, L. Liu, H. Ma, J. Chen, “Web AR:
A Promising Future for Mobile Augmented Reality--State of the
16, 2019.
[13] J. Etienne, Creating Augmented Reality with AR.js and A-
Frame [online]. Available: https://aframe.io/blog/arjs/
[14] J. Dirksen, Learning Three.js: The JavaScript 3D Library for
WebGL, Packt Publishing, 2013.
[15] Converting from Speech to Text with JavaScript [online].
Available: https://tutorialzine.com/2017/08/converting-from-
speech-to-text-with-javascript
 The Application of Search Engine Optimization in
Internet Marketing
Aleksandar D. Andonov
Abstract – In the past few years, more and more visitors to
the internet are approaching webpages through search engines,
instead of utilizing direct links from the other websites. Due to a
large variety of marketing strategies available for the companies,
it has become quite difficult to select a particular strategy that
can enhance the business accordingly. However, SEO is one of
the potential techniques of internet marketing which allows
businesses to show their online presence, by improving their
ranks online and directing the flow of online traffic accordingly.
Thus, this research study aims to investigate how SEO influences
internet marketing in the modern world.
Keywords – Search engine optimization, Internet marketing,
Online traffic.
I. INTRODUCTION
With a number of marketing strategies available for the
businesses, it is quite hard to recognise which particular
strategy is best for the business. Before investing in the
marketing budget, it is important for organisations to
thoroughly research every available marketing strategy in the
market. However, Search Engine Optimisation (SEO) is one
of the potential internet marketing applications, which
enhances organisation's website and its online presence, both
off-page and on-page in order to boost its ranking within
search engines such as Google. SEO is an advanced technique
of making the website easier to categorise and easier to find
for the users. It helps the user to find a particular business
among millions of other companies. Reports suggest that
about 14 billion online searches take place every month.
These figures suggest the strong influence of SEO on internet
marketing. As the website ranks higher in the online search
engines, it will eventually obtain a greater number of clicks
from the users. Hence, the more the clicks website receives,
the greater number of individuals will visualise the website,
thus, increasing the chances of online customers to find the
service or a product of their interest. However, the business
revenues and conversion rate thoroughly increases, as the
website ranks higher in the search engine [1].
The main goal of this article is to thoroughly study the
significance of SEO, and its influence on internet marketing.
To accomplish this goal various research studies will be taken
into consideration that will help to determine the importance
of SEO in the current internet marketing world. However, the
article is organised as follows; Section I provides a brief
Aleksandar D. Andonov is with University of Library Studies and
Information Technologies, Faculty of Information Studies,
Boulevard "Tsarigradsko shose" 119, 1784 Sofia, Bulgaria, E-mail:
a.andonov@unibit.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
introduction of the study, while Section II reflects a thorough
background of SEO and its impact on internet marketing.
Section III provides a comprehensive review of literature from
previous studies, Section IV shows a technique for improving
SEO, and Section V will include the results and discussion in
accordance with the research topic. Lastly, in Section VI some
conclusions are given, while certain recommendations are also
provided accordingly.
A. Research Question
This research article will focus on the below-mentioned
research question throughout the study: How search engine
optimisation influence internet marketing?
II. BACKGROUND
The concept of SEO was first put forward in 1991 when
the first ever website was launched on the internet. In the past,
the term SEO was recognised as SEM (Search Engine
Marketing), as this specific technique was extensively used to
enhance the marketing of business products and services.
Search engines have comprehensively altered the ways
through which online users find information, shop for services
and products, conduct research, connect with others, and
entertain oneself. However, in the current world scenario,
there are a number of search engines on the internet which
ultimately offers businesses an opportunity to optimise their
website and improve its ranking so that the maximum number
of online customers can be served through its platform [2].
III. LITERATURE REVIEW
A. Evolution of SEO in the Marketing World
In the last decade of the 20th century, the landscape of the
search engine was highly competitive. In the beginning, on-
page activities were considered as the only channel to perform
any kind of SEO. This involved ensuring that the web content
is relevant and good enough to attract the audience, and
whether the keywords are repeating regularly throughout the
content present on the meta tags and website. In the past, if
one business had to outrank the other, they would have to
ensure that their number of keywords are higher from their
competitor. However, today such practices are known as
spamming. From 1994 to 1999, a variety of search engines
came into operation including Yahoo, Backrub, Search Engine
Watch, and Goto. But all these search engines that were
utilised for SEO were left behind when Google came into the
37
market [3].
In 2000, Google partnered with Yahoo in order to boost
its search engine, however, this specific decision of Yahoo is
considered its worst move in the search engine history, as
Google was hardly known previously, but when it partnered
with Yahoo, the end result was that the each and every search
result of Yahoo stated: “Powered by Google”. Thus, Yahoo
eventually ended up introducing its biggest competitor in the
online market that is Google. Until this point, SEO involved
ranking of webpages based on basic site structure, domain
names, and on-page content. But at the same time, Google’s
PageRank algorithm and web crawler were considered quite
revolutionary for the retrieval of data. Google considered both
off-page and on-page factors, which means it monitored both
the quality and quantity of external links that directed users to
the website, along with the anchor text utilised in the content.
With two potential updates in algorithm during 2011 and
2012, the SEO of webpages was highly affected, as Google
attempted to remove all the search results from its online
directory in order to reward the best quality websites [4].
However, in the current world scenario, with the
advancement of search engine, accomplishing the highest rank
through SEO is no longer enough to direct the online traffic
towards webpages. In the past few years, the Search Engine
Result Pages (SERP) are heavily cluttered with article listings,
links, videos, images, and advertisements. Thus, it is quite
easy for the top-ranked result to get jumbled up with the
features of other pages. Still, the key phrase and keyword
optimisation is of great importance, but at the same time, the
SEO experts now must consider other offsite components as
well [5].
B. Significance of SEO in Business
The utilisation of SEO by business organisations can put a
huge impact on their online marketing strategy. SEO becomes
more important for those businesses that have a separate
website through which the companies cater the needs of the
consumers. Each website has various needs in terms of SEO,
but it is quite rare that these needs are determined by the
industry or the business. Organic search is considered as the
potential source for attracting the traffic of the business
website, and it is also an important element of the purchasing
funnel which eventually leads the users to carry out the
engagement or conversion. However, being visible as an
adequate platform by search engines, such as Google might
always favour the business in terms of enhanced brand
reputation, brand awareness, and sales. Moreover, SEO plays
an important role to develop a robust foundation for the
businesses by developing an attractive website having
effective and clean user experience which can be easily
discovered through the search engine. At the same time,
developing an authoritative business brand demands
commitment, effort, and patience, while it also relies on
providing quality and valuable service or product that permits
customers to put trust on that specific brand [6].
Similarly, SEO can also put a huge impact on the business
buying cycle. From the perspective of the buyer, this might be
considered as one of the biggest benefits for the business
38
organisation. Relatively, utilising tactics of SEO to spread the
messages regarding ground-breaking services, products, or
good deals, can act as a game-changer for the businesses. It is
important for businesses to ensure that their brands are
properly visible across those online places where the users
needed them the most. Thus, local SEO will play its part by
enhancing the visibility of the brand and allowing potential
consumers to find the relevant answers which will be
thoroughly answered by the businesses. SEO is also quite
beneficial for businesses in terms of cost. Studies suggest that
SEO is relatively much cheaper in comparison to other
marketing tactics, while the outcomes are much significant in
terms of the brand’s bottom line and benefits. By
implementing an SEO strategy, businesses do not pay a cost
for marketing, instead, they make an investment with an aim
to further prosper the business. SEO can be considered as a
long duration strategy, as it can put an effective impact on the
business initially depending on the actions undertaken,
however, these potential actions will put an everlasting impact
that will continue for several years. With the evolution of the
market, it is best for businesses to follow the changes and
trends closely. Even if the website does not implement the
intense recommendations of SEO, by following the basic
steps, the businesses can accomplish and provide a decent
experience to the online users [7].
C. SEO and Internet Marketing
SEO is a technique to alter or develop a webpage in an
optimised manner so that it can become more relevant as
compared to other pages on the internet search engine.
However, the process of SEO can be divided into five major
components including, off-page optimisation, on-page
optimisation, selection and research, indexing the site through
the search engine, and keywords. How the search engine
operates, which search engines are highly preferred by the
online audience, what users search for, and how keywords are
entered within the search engines, are some of the potential
elements which are considered in the strategy of SEO internet
marketing. Searching is one of the potential practices of the
online audience. Several users search different kinds of
information on a mundane basis, and about half of the traffic
on the internet starts with the search engine. Marketing a
website involves enhancing the amount of links. The
significance of SEO can be determined by the fact that it
increases the traffic volume directed to the sites of the
individual through the search engine, thus, making it a reliable
technique for effectively improving exposure and visibility of
the webpage. Also, it can help to enhance customer interaction
and contributes towards improving the overall traffic on the
website [8].
The main aim of the entire SEO process is to accomplish
high ranking in the results of a search engine when a user
enters specific keywords or phrases in it. Apparently, there is
not much difference between internet marketing and SEO, as
both incorporate the same skills and executes the same
functions. Through SEO users get engaged in developing an
online existence. In practice, the consultant of SEO usually
searches for different areas of digital marketing. The
contemporary trends involve optimising a webpage and
perform search engine marketing. For instance, there are
certain elements that take help from SEO in order to select the
potential keywords, such as mobile advertising, behavioural
marketing, contextual marketing, blogging through SEO
content, video content advertising, viral marketing, and social
media marketing. Hence, for this purpose, it is quite important
to have a strong, SEO and internet marketing strategy [9]. The
below Fig. 1 presents how the SEO method is based on
internet marketing reflects.
Fig. 1. SEO as a Subset of SEM
Studies suggest that the impact of branding as evidenced
in the results of the search engine plays a vital role in
enhancing awareness of the products of business vendors. It
has been found out that SEO not only enhances direct traffic
through clicks, but it also promotes profiles of brands within
the setting of online retail. In recent years, the SEO
application in internet marketing has altered dramatically, as
Google search engine is now on the top, carrying out at least
60 percent of the overall web searches. Due to high
competition in the SEO world, the bid prices are driving
through the roof, thus, making it quite impossible for the
limited budget virtual advertisers to obtain top ranking. As a
result, it remains no longer feasible to bid on the highest
searched keyword, and businesses are forced to enhance their
campaign by including multiple search engines and different
keywords. However, in such a kind of environment, SEO can
offer better outcomes on investment, as compared to other
channels of marketing, without any kind of obstacles [10].
Apart from this, studies demonstrates that SEO along with
Social Networking Sites (SNS) can help to create a strong
internet marketing strategy. The significance of marketing
through online channels is a well-proven fact. A business
organisation for the promotion of its services and product
must incorporate three important elements including a well-
defined SEO process, the SNS, and the official website [25].
In recent years, improving webpage ranking through the
utilisation of appropriate marketing strategy has become a
pivotal issue. SNS has thoroughly infiltrated in people’s
mundane lives, as platforms like LinkedIn, Twitter, and
Facebook, provide users an opportunity to share website links,
photos, and videos of products within the cyber environment.
Thus, by combining SNS and SEO (Fig. 2), the business
organisation can ensure to have the best marketing strategy, as
these techniques make sure that the online audiences are
directed to the particular webpage, and are aware of the
product and services at the same time [11].
39
Fig. 2. Internet Marketing Strategy through SNS and SEO
IV. RANDOM INTERNAL LINKING
The SEO based on backlinks most of the time is pointing
to the home page of the website. However, in e-commerce, the
home page is not so important. A visiting customer will be
annoyed when looking for an item and is redirected to the
main page to either repeat the search or navigate through
categories and subcategories. The good SEO optimized
website should implement a technique for improvement of
online marketing and user experience.
There is a so-called SEO “juice” that is flowing through the
home page to the internal pages via links. Search engines are
following the links to discover new pages; therefore, a good
structured linking is very important. The main problem in this
approach is where there are too many categories and
subcategories. The SEO developer is not able to place a direct
link at the home page of all important pages. A user will get
confused and could possibly leave the page. Furthermore,
search engines are not following and indexing approximately
100 links per page. That will make some internal pages
inaccessible for crawlers and could never get indexed even if
they are in the sitemap file. A useful technique that is used in
this case is to have a random link appearing on every page
reload. This is a way to present the internal link to the web
crawler in order to get it indexed. Often revisiting the main
page by crawlers can be set by meta tags, which will result in
several new pages to be discovered, indexed and pushed up.
Using such dynamic links from the home page, makes the
crawler “think” of the new link as something very important.
If this technique is presented after the crawler has visited all
available pages, this will be something new and interesting for
the search engine to index.
V. RESULTS AND DISCUSSION
The evaluation and implementation of SEO in online
marketing demonstrated in the previous section of the article
is discussed in this section.
A. Influence of SEO on Online Marketing
From the results of Ho et al [12] study, it has been found
out that the SEO process can put a strong influence on the
average time of visits, bandwidth variation, ranking, number
of hits, pages, and visits. Although the result of this particular
study suggests that the number of hits, pages, and visits did
not increase significantly, but, on the other hand, the
bandwidth has been enhanced, while the ranking of the
website has also been effectively increased over the time
period due to SEO. As can be analysed from (Fig. 3) that the
ranking of the website which was taken into consideration in
this particular research work was 14th in July 2007, however,
with the passage of time the site ranking was enhanced until it
reached the 2nd ranking in June 2008 [12].
Fig. 3. Monthly Ranking of the Website
Thus, from the above discussed results, it can be said that
SEO can play a crucial in alleviating the business of the
organisation. Studies suggest that in the world of digitisation,
it has become quite significant for the organisations to keep
themselves ahead of their competitors. In order to excel in this
marketing technique, it is compulsory for the businesses to
ensure their visibility online. Since SEO can be extensively
utilised to promote service, product, or the entire brand
through digital devices such as radio, email, and mobile,
hence, companies have a huge opportunity to boost their sales
accordingly [12].
Since the above mentioned results show a potential impact
of SEO on online ranking of the businesses, thus, this
particular result can also be discussed through research studies
which demonstrate that website that is associated with higher
rankings through the search engine, has a better chance to
attract the attention of visitors. It has been found out that users
usually believe upon the results shared by the search engine,
and when the company accomplishes a top position within the
search engine, it directly signals to the users that this
particular website is more effective and can be trusted than
other. The higher the business will be in the search rankings,
the more traffic and clicks the website will generate, hence,
enhancing user experience, and allowing consumers to
become regular buyers [13].
Apart from this, the results of research studies also
demonstrate that SEO acts as a potential source for business
websites in terms of improving its branding by building
credibility and trust within the users. The author of the study
found out that a majority of the online population views those
businesses adequate that are presented on the first page of the
search engine. Thus, it is important for the SEO users to
ensure that their website is subjected to the first page, and it
should keep the users there during return visits in order to
achieve positive reviews and satisfied customers altogether.
SEO can enhance the branding of the website in a similar way
through which it aids to develop credibility and trust because
40
both are focused on search placement of a website on the first
page. When a particular website is regularly placed on the top
of the search of the keyword, this usually suggests that the
business is the potential player in that sector. Through SEO,
the search engines will automatically give a constructive
response to the users regarding its complete portfolio and
what it delivers to them. Thus, it will be the responsibility of
the SEO manager to fulfil the user’s expectations in order to
complete the process of branding [14].
B. SEO and its Effect on Online Marketing Cost
Studies indicate that SEO provides a cost-effective
solution for online marketing. Results suggest that online
marketing through SEO is not a costly strategy, because it
attracts businesses to do permanent investment, without
indulging in the complexity of advertising. Once, the website
is ranked at the top of Google search, the company will be
able to achieve continuous sales. Due to the inbound nature
of SEO, businesses get an opportunity to save potential sums
of money in comparison to outbound tactics such as cold
calling. Although cold-calling can also act as an effective
technique, but it has been found out that the generated cost
through cold-calling is much greater as compared to SEO.
Moreover, since, SEO potentially focuses on those users that
are regularly searching for the services or products, therefore,
the resulting online traffic is much more qualified as
compared to other strategies of marketing, thus, resulting in
cost-effectiveness for the businesses [15].
Since SEO delivers quantifiable and trackable outcomes,
regardless the businesses have a non-e-commerce or e-
commerce website, therefore, the complexity related to ROI
gets effectively minimised. The agencies of SEO are usually
able to monitor nearly every element of their marketing
strategies such as an increase in conversions, traffic, and
rankings. A wide variety of analytics also enhances the
capability of businesses to reach within the granular level and
track the information of demographics and other metrics of
engagement associated with individuals that are interacting
with the website. Through SEO, by monitoring the channel
selected by the online users to complete the process of sales,
the company can be able to introduce further cost-effective
solutions in order to increase the customer base to the
maximum extent possible, which eventually results in an
increase of revenues of the company [16].
From the studies, it has also been found out that in
comparison to PPC, SEO is more cost-effective. The reason
behind this is because SEO involves organic channels to
produce online traffic and accomplish huge sums of money,
although it might take some time, but the outcomes in the long
run are much more authentic as compared to PPC. PPC might
produce rapid traffic, but from a broader perspective, this
online traffic is not sustainable. Paid marketing like PPC only
benefits the company for a short period of time by generating
an average amount of revenues very quickly. Thus, PPC is a
total wastage of both money and time, and therefore,
marketing companies usually adopt SEO with an aim to make
large profits for a longer period [17].
 C. Impact of SEO on Business Performance Through
Online Marketing
The findings from the research studies demonstrate that
SEO can be the imperative solution for the organisations in
order to thoroughly survive in the current competitive
environment of the market. From the results, it was found out
that the users of the internet are most likely to select the
business website from the top 5 results provided on the initial
page of the search engine. Thus, SEO enhances business
performance, by promoting its feature across the social bloc. It
has been found out that users who find a particular website
through Yahoo or Google are more likely to promote it on
Twitter, Facebook, Instagram, and other channels of social
media. Hence, SEO can be quite significant for the running
efficiency of the business website. It provides both direct and
indirect benefits to the website stakeholders. In terms of
direct, SEO ensures to enhance the traffic of the search
engine, while the indirect benefit is that it offers a coherent
checklist (framework) for utilisation before the content is
published on the website. Thus, by improving the rate of
conversion, SEO allows businesses to transform their online
visitors of the website into potential customers. Through this
phenomenon, the best targeted SEO will aid in attracting
genuine online users towards a particular website, which will
have true interest in the selected business. Thus, this will help
to enhance the conversion rate by transforming visitors into
consumers, and eventually increasing the sales of the services
or products. Also, it has been found out that SEO plays an
important role to boost customer engagement altogether.
Aligning the needs of the customer to improve service or
product can be quite crucial for the company. But when
company’s blog or web page is optimised through search
engine, it will find itself in a strong position to interact and
engage with customers [18].
VI. CONCLUSION
SEO is one of the significant applications of internet
marketing that can aid to enhance the company’s website
and its online presence both off-page and on-page, in order
to improve its ranking within the online search engines.
Thus, from the findings of the study, it was evaluated that
the process of SEO can put a strong impact on the website’s
visit time, bandwidth variation, ranking, and the number of
hits. Further, it was investigated that online marketing
through SEO is not an expensive strategy. The current study
also suggests that in comparison to PPC, SEO is more cost-
effective, while it was found out that SEO can act as an
imperative solution for the companies to survive in a
competitive market.
Thus, from the findings of the study, it can be
recommended that businesses, especially start-up businesses
should adopt SEO techniques in order to attract a maximum
number of audiences in a minimum amount of time.
Moreover, it can be recommended that future
researchers should also focus on SEO and SNS collectively
in order to bring awareness among the individuals regarding
the benefits of both these aspects in terms of costs,
41
reliability, accessibility, and durability.
REFERENCES
[1] A.N. Hidayanto, M.S. Adha, M.A. Jiwanggi, and T. Melia, “A
Study of Impact of Search Engine Optimisation to Internet
Marketing Strategy”, International Journal of Services,
Economics and Management, vol. 4, no. 4, pp.298-316, 2012.
[2] A. Heinze, G. Fletcher, and C. Chadwick, “From Search Engine
Optimisation to Search Engine Marketing Management:
Development of a New Area for Information Systems
Research”, UK Academy for Information Systems Conference
Proceedings, 2010.
[3] J. Grappone, G. Couzin, Search Engine Optimization (SEO): An
Hour a Day, John Wiley & Sons, 2011.
[4] J. L. Ledford, Search Engine Optimization Bible, vol. 584, John
Wiley & Sons, 2015.
[5] M. N. A. Khan, A. Mahmood, “A Distinctive Approach to
Obtain Higher Page Rank Through Search Engine
Optimization”, Sādhanā, vol. 43, no. 3, p. 43, 2018.
[6] G. Egri, C. Bayrak, “The Role of Search Engine Optimization
on Keeping the User on the Site”, Procedia Computer
Science, vol. 36, pp. 335-342, 2014.
[7] E. Enge, S. Spencer, J. Stricchiola, and R. Fishkin, The Art of
SEO, O'Reilly Media, Inc., 2012.
[8] N. Yalçın, U. Köse, “What is Search Engine Optimization:
SEO?”, Procedia-Social and Behavioral Sciences, vol. 9, pp.
487-493, 2010.
[9] Educba, Importance of SEO in digital marketing. Available at:
https://www.educba.com/seo-in-digital-marketing/. Accessed
on: 18th April 2020.
[10] J. Jantsch, P. Singleton, SEO for Growth: The Ultimate Guide
for Marketers, Web Designers & Entrepreneurs, SEO for
Growth, 2016.
[11] B. Y. Shih, C. Y. Chen, and Z. S. Chen, “Retracted: An
Empirical Study of an Internet Marketing Strategy for Search
Engine Optimization”, Human Factors and Ergonomics in
Manufacturing & Service Industries, vol. 23, no. 6, pp. 528-540,
2013.
[12] L. H. Ho, M. H. Lu, J. C. Huang, and H. Y. Ho, “The
Application of Search Engine Optimization for Internet
Marketing: An Example of the Motel Websites”, 2010 The 2nd
International Conference on Computer and Automation
Engineering (ICCAE), vol. 1, pp. 380-383, February 2010.
[13] A. Pandey, “Optimising Online Marketing Resources: SEO
Strategy”, International Journal of Technology Marketing, vol.
7, no. 3, pp. 267-277, 2012.
[14] A. Enaceanu, “SEO Techniques for Business
Websites”, Database Systems Journal, vol. 1, pp. 23-26, 2010.
[15] A. K. Kirtiş, F. Karahan, “To Be or Not To Be in Social Media
Arena as the Most Cost-Efficient Marketing Strategy After the
Global Recession”, Procedia-Social and Behavioral
Sciences, 24, pp. 260-268, 2011.
[16] B. Skiera, N. Abou Nabout, “Practice Prize Paper—PROSAD:
A Bidding Decision Support System for Profit Optimizing
Search Engine Advertising”, Marketing Science, vol. 32, no. 2,
pp. 213-220, 2013.
[17] W. T. Kritzinger, M. Weideman, “Search Engine Optimization
and Pay-Per-Click Marketing Strategies”, Journal of
Organizational Computing and Electronic Commerce, vol. 23,
no. 3, pp. 273-286, 2013.
[18] R. Mahajan, “Use of Social Media as a New Investigative Tool
in Marketing Research for Small Business”, International
Journal of e-Education, e-Business, e-Management and e-
Learning, vol. 5, no. 3, p. 129, 2015.
 Improving the Concept of Medication Vending Machine
in the Light of COVID-19 and other Pandemics
Dragan S. Jankovic1, Aleksandar M. Milenkovic2 and Andjelija I. Djordjevic3
Abstract – Pandemics are usually followed by a large number
of infected people, enormous number of critical patients and
unavoidable fatal outcomes. If the disease is transmitted by
contact, one of the most important measures for preventing the
spread of the disease is social distancing. Ensuring social
distancing is one of the most difficult tasks of the governments of
countries during the pandemic. It has been shown in practice
that any measure or way to increase social distance significantly
reduces the number of victims and shortens the duration of the
pandemic. This paper proposes a modification of the existing
concept of a vending machine for distributing drugs (medication
vending machines – MVM), which would enable better social
distancing in the case of pandemics/epidemics, such as the
current COVID-19 pandemic. The proposed modification is
adapted to the healthcare system organization that exists in the
Republic of Serbia and can be implemented in similar health
care systems, such as in the neighboring countries (Western
Balkan countries).
Keywords – Medication vending machine, COVID-19,
Pandemic, Social distancing, Smart city.
I. INTRODUCTION
Vending machines are widely used in many domains
around the world, such as a vending machine for food, drinks,
cigarettes, newspapers, etc. They started a long time ago. The
first vending machine was the coin-operated machine for holy
water dispensing constructed in 17th century by Hero of
Alexandria, Greek mathematician and engineer from
Alexandria [1].
The first modern coin-operated vending machines were
introduced in London, England in the early 1880s, dispensing
postcards. The machine was invented by Percival Everitt in
1883 and soon became a widespread feature at railway
stations and post offices, dispensing envelopes, postcards, and
notepaper. The Sweetmeat Automatic Delivery Company was
founded in 1887 in England as the first company to deal
primarily with the installation and maintenance of vending
machines. In 1893, Stollwerck, a German chocolate
manufacturer, was selling its chocolate in 15,000 vending
machines. It set up separate companies in various territories to
manufacture vending machines to sell not just chocolate, but
cigarettes, matches, chewing gum and soap products [2].
1Dragan S. Jankovic is with University of Nis, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Nis,
Serbia, E-mail: dragan.jankovic@elfak.ni.ac.rs
2Aleksandar M. Milenkovic is with University of Nis, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Nis,
Serbia, E-mail: aleksandar.milenkovic@elfak.ni.ac.rs
3Andjelija I. Djordjevic is with University of Nis, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Nis,
Serbia, E-mail: andjelija.djordjevic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
The first vending machine in the US was built in 1888 by
the Thomas Adams Gum Company [3] selling gum on New
York City train platforms. The idea of adding games to these
machines as a further incentive to buy came in 1897 when the
Pulver Manufacturing Company added small figures, which
would move around whenever somebody bought some gum
from their machines. This idea spawned a whole new type of
mechanical device known as the "trade stimulators".
One of the challenging areas of using a vending machine is
the pharmacy. There are examples of medicine vending
machines (MVM) which are used in the last ten years in
different countries. The drug vending machines have been
used in many countries for the last ten years (US, Canada,
UK, Switzerland, and South Africa).
Despite all current usage of medicine vending machine, we
cannot say that they are widely used and that all potential
which they offer is appropriately and adequately exploited.
There are restrictions to the list of available drugs which is
distributed by means of drug vending machines. Usually in
this way painkillers, oral contraceptives (“morning after pill”),
baby food, etc. are distributed.
There are different types of MVM but more or less they
have the same set of components and supported
functionalities. In [4] is considered one variant of MVM. The
major MVM components are: a scanner to take the input from
a user, a system that includes servo motors for dispensing the
medication, large storage space to store the pills, sensors to
detect the motion of pills, an inventory monitoring system to
keep track of the storage, an industrial standard vertical foam
fill machine to pack the medication separately and a non-
contact laser inkjet printer to print the description which
includes the time at which the medicine must be taken. The
inventory monitoring system also keeps track of the expiry
date of each batch of medicine and sends alert to refill the
storage when the pills run out. It also holds an inbuilt system
to receive money from the user for the drugs that are
dispensed. All these systems are monitored by a central
microprocessor, which is programmed to receive input from
the user via the scanner and to actuate and control all the
necessary components required to dispense the medication
requested by the user.
Different realization of MVM offers a different realization
of some parts of MVM, for example payment system (coins,
banknote, pay card, e-Banking, etc.). Therefore, it is not
surprising that the price of some MVM is only around 1000
euros.
There are different stories of MVM use. In 2014 Arizona
State University has approved InstyMed machines on its
campus. After the drug store at the health services facility in
Arizona State University was closed, school officials set to
42
 Fig. 1. Top 10 prescribed drugs at Health Center Nis
replace it with a vending machine for dispensing prescription
medication [5]. The InstyMed sending machine was available
to students and university employees with approved
prescription which could be used during the next month.
Customers have a voucher with unique ID information and a
24-hour-only code that can transfer over a secure connection
from the prescription doctor to the machine.
A very interesting real-life story in COVID-19 times comes
from Vancouver [6]. The opioid epidemic, which caused the
declaration of public health emergency in Vancouver four
years ago, is made worse by the COVID-19 pandemic. It is a
crisis on top of a crisis. The usual harm-reduction approach, a
designated location where drugs can be taken with oversight
by first aid-trained staff, is problematic. Such sites constitute a
public gathering place and have been closed in order to avoid
public gatherings. The illegal drug supply is getting more
expensive and less predictable. Overdose deaths still happen
often. MySafe, Dispension Industries’ opioid-dispensing
ATM, debuted in Vancouver in December 2019 as part of
Tyndall’s plan to replace risky street drugs with safe,
pharmaceutical-grade opioids and reduce the staggering rate
of overdoses. About a dozen patients use the machine. All
report an improved quality of life and less involvement in the
dangerous street activities that used to pay for their habit.
Previously described cases show that MVM can have a very
good perspective and there are a set of reasons to use them.
But it is the fact that today’s potential of MVM is not well
exploited. There are several reasons. Some of them will be
discussed in the next chapter. It is quite clear that for the
expansion of MVM use some concepts of MVM should be
changed and adapted to a concrete scenario of use.
One modification we proposed in this paper inspired by the
actual COVID-19 pandemic and effort of all countries to
reduce virus SARS-CoV-2 [7] expansion by providing better
social distancing as one of the most efficient activities in the
suppression of the spread of an epidemic.
43
The proposed use of MVM can be considered as one of the
services provided by smart cities, because the smart city
concept includes improvements in health-care domain too [8].
II. REASONS AGAINST USING MVM
There are many obstacles and open-ended issues that have
prevented the mass use of medication vending machines.
Some of them are universal while some are specific to some
country or region, for example different law regulations. We
will number some of them. It is the fact that the number of
medications is huge and that it is not possible to make it
available through the vending machines. So, medication
vending machines can be used for a restricted amount of
medications. In many cases patients ask a pharmacist for some
explanation therefore, a simple MVM cannot cover these
cases and patients should request assistance directly at the
pharmacy, where the staff is available, instead of using MVM.
There are different models of paying for medications. It
differs from country to country. For example, in the Republic
of Serbia for a huge number of medications patients should
pay only part of price or nothing, while for others they should
pay full price. Therefore, MVM should support different
pricing/paying models. For most medications, patients must
have an adequate prescription. So, MVM should be able to
“read” prescriptions and verify a patient’s request. Of course,
there are other practical open questions: where to install
MVM (near a pharmacy or not), the capacity of MVM,
security (especially if there are narcotics), how and when to
add new medications in a machine, etc.
III. COVID-19 AND EPIDEMICS – NEW CHANCE FOR
MEDICATIONS VENDING MACHINES
The emergence and rapid expansion of major epidemics
have an important influence on the daily lives of people
through changing health, economic, working style, social and
political routines. During outbreaks, especially those with a
pandemic character, a set of key activities [9] which are
updated with COVID-19 strategy [10] have been identified
whose strict implementation has an impact on the reduction of
a number of infected people and suppression of the spread of
an epidemic. One of the most important activities is
movement reduction and social distancing [11] (decreasing
the number of possible contacts) especially for vulnerable
groups (elderly people, chronic patients) [12]. It is quite clear
that health facilities and pharmacies are high-risk places in
relation to the possibility of infection with the virus,
especially if the virus is transmitted by droplets i.e. by air such
as SARS-CoV-2 virus. Therefore, the need for people to come
to pharmacies should be reduced, i.e. they should be allowed
to take their medications without entering pharmacies. This
fact again puts the focus on MVM but it is clear that the
existing concept and implementations of vending machines
must undergo significant modification and become part of the
health care system, and not just independent devices that
potentially reduce the cost of dispensing some drugs.
One of the limitations of MVM is their capacity, because it
is not possible to store all existing medicines. It is quite clear
if we have in mind that the number of medications on the
Republic Fund of Health Insurance (RFZO) list in the
Republic of Serbia is higher than 3966 while the capacity of a
machine is less than 200 different kinds of medications.
Therefore, MVM could be modified in direction to provide
only the most frequently prescribed medications. The
drawback of this approach is the fact that usually one visit
results in more than one prescription i.e. more prescribed
medications. That means that patient will get prescribed
medications from MVM only if all prescribed medications are
on the list of most frequently prescribed medications.
Otherwise, the patient again should go to the pharmacy to get
medications which are not on the list of the most frequently
prescribed medications.
Fig. 1 shows top-ten medications prescribed in the Health
Center Nis. We can see that in top-ten medications we have
medications for different diagnoses, but also different
medications used for the same illness (Tensec and Concor are
used for heart illness as well as Binevol which is dominantly
used for hypertension). Practically, that means that a huge
number of medications will not be completely on this list.
So, we must modify MVM in a different way. We propose
a modification of MVM in the way described in the rest of the
paper.
IV. IMPROVED CONCEPT OF MVM
Improved MVM should solve most of the above-mentioned
problems. So, our proposal for improvements follows.
The concept of MVM as a machine for purchasing
medications from a set of medications available through
machine should be changed to machine for obtaining
medications prescribed by general practitioners to the concrete
patient. In this scenario, the patient has a prescription for
some medications given by a general practitioner in some
healthcare institution (HCI) and it is registered in the medical
information system (MIS) of HCI. All prescription data
registered in MIS are sent to the central state repository. This
information should be available to a pharmacy that has an
44
MVM placed outside of the pharmacy. Based on some
identification method (ID, password, QR code at
prescriptions, etc.) the patient gets medications from the
machine (which is previously prepared by a person from the
pharmacy). The activity diagram for this scenario is shown in
Fig. 2. A small drawback of this concept is the time delay
needed until medications become available to the patient from
the time when it notices which of various medications MVM
will use.
Based on proposed concepts MVM should be able to:
1. Receive a patient request – there are two scenarios: patient
sends QR code from the prescription to the concrete
pharmacy which has MVM; a health-care institution is
connected with pharmacies indirectly via RFZO and all
prescriptions are available to pharmacies;
2. To provide a correct price according to each separate
request (the price of a medication is not same for all
patients because some patients have a reduction in price
(different type of health insurance); to provide payment by
coins, banknote (it must return the change) and credit
cards or e-Banking;
3. To provide a required medication within a reasonable time
from the arrival of the request and inform the patient about
it (or to provide medication within a predetermined time,
for example 1 hour);
4. To withdraw medications from MVM if they have not
been taken by the patient within the appropriate time
period (for example 24 hours) and to remove the requests;
5. To note the number of empty slots in machine i.e.
readiness to accept medication for new patients.
Fig. 2. Sequence diagram
Since very often patients ask the pharmacist for instructions
for the use of the drug (each prescription has a part related to
the way of use and daily doses and frequency), a function
which can be added to the MVM, where MVM is printing a
short instruction for the use of the medication at the request of
the patient. It is similar with the interaction with other
medications and contraindications. On the other hand, these
services can be independent of MVM and realized as a part of
the MIS, or as a part of the pharmaceutical information system
(PIS).
The proposed concept of using MVM implies that the
device is located within the pharmacy and is available to
patients outside the pharmacy.
V. POSSIBILITY AND PERSPECTIVE OF USING MVM
IN SERBIA
The proposed concept of MVM can be applied fully in the
Republic of Serbia. The facts supporting this claim are as
follows. The healthcare system of the Republic of Serbia is
mainly based of state healthcare institutions divided into
three-levels: primary (infirmaries which are spread throughout
he whole country), secondary (hospitals which support more
specialties), and tertiary – high specialized hospitals, usually
as a part of clinical centers). Patients can get prescriptions
only from the primary health level institution. On the other
hand, all primary health level institutions have MIS which
send data about all prescriptions to the central system (e-
Recept in RFZO) which is connected with pharmacies (Fig.
3). So, when patients wish to get medications they should go
to any pharmacy, identify themselves (each patient has the
health-care card and unique health-care number (LBO)) and
pay for medications. A special case is provided for chronic
patients who have prescriptions for a 6-month-long therapy
(which is extended in COVID-19 pandemic period to 9
months). So, in this period chronic patients should go to the
general practitioner only once and repeat it after 6 or 9
months. It is an important feature that directly reduces the
number of healthcare institution visits and reduces the chance
that the patient will be infected with the virus during the visit
[11]. Information from pharmacies go to the RFZO
centralized system i.e. for each prescription RFZO has
information when medications are taken by the patient.
Fig. 3. Health-care system in Republic of Serbia
VI. CONCLUSION
This paper proposes a modification of the MVM that should
affect the level of social distancing, which is one of the main
measures in preventing the spread of pandemics such as
COVID-19. The proposed modification can have significant
success in the Republic of Serbia if we have in mind that all
primary health care institutions have adequate MIS connected
with RFZO while RFZO is connected with all state
pharmacies. So, there are real conditions for installing MVM
in pharmacies where patients can get prescribed medicines
without entering pharmacies which will significantly reduce
social contacts inside pharmacies and reduce the spread of the
virus.
The proposed modification of MVM concept is adapted to
the healthcare system organization in the Republic of Serbia,
45
however, it can be used in all countries with a similar
healthcare system, such asex-YU countries.
Besides social distancing there are also some other positive
effects: patients can get medication 24/7, pharmacies can
work more efficiently (inexpensive distribution of medication
and increased number of patients who can be served).
In general, the use of proposed MVM can be considered as
one of highly important services (in health-care domain)
provided by smart cities, because the health of citizens is the
most important goal for all governments.
ACKNOWLEDGEMENT
This work has been supported by the Ministry of Education,
Science and Technological Development of the Republic of
Serbia.
REFERENCES
[1] Heron's Inventions, https://explorable.com/heron-inventions
(accessed May 23, 2020).
[2] L. Oppel, "Drug-Dispensing Machines", BC Medical Journal
vol. 52, No. 1, 2010, https://www.bcmj.org/news/drug-
dispensing-machines (accessed May 23, 2020).
[3] Vending machine, https://www.britannica.com/topic/vending-
machine (accessed May 23, 2020).
[4] A. Brolin, M. Ravisankar, V. Gokulnath, M. Harivishanth,
"Design of automated medicine vending machine using
mechatronics techniques", IOP Conferences Series Materials
Science and Engineering, 402 (1):012044, September 2018.
[5] M. Ackerman, "College Campus Allows Vending Machine That
Will Dole Out Prescription Drugs", https://www.thefix.com
/content/college-campus-allows-vending-machine-will-dole-out-
prescription-drugs (accessed May 23, 2020).
[6] A. Bethune, "Could an Opioid Vending Machine from
Dartmouth make Street Drug Users Safer During COVID-19
Pandemic?", https://www.theguardian.pe.ca/business/regional-
business/could-an-opioid-vending-machine-from-dartmouth-
make-street-drug-addicts-safer-during-covid-19-pandemic-
438458/ (accessed May 23, 2020).
[7] C. C. Lai, T. P. Shih, W. C. Ko, H. J. Tang, and P. R. Hsueh,
"Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-
CoV-2) and Coronavirus Disease-2019 (COVID-19): The
Epidemic and the Challenges", Int. J. Antimicrob. Agents, vol.
55, 2020, doi: 10.1016/j.ijantimicag.2020.105924
[8] A. Visvizi, M. Lytras, "Smart Cities: Issues and Challenges",
Elsevier, June 2019, ISBN: 9780128166390.
[9] WHO, "Infection prevention and control of epidemic- and
pandemic-prone acute respiratory infections in health care",
2014, https://apps.who.int/iris/bitstream/handle/10665/112656/
9789241507134_eng.pdf?sequence=1 (accessed May 23, 2020).
[10] WHO, "COVID-19 Strategy Update", 2020,
https://www.who.int/docs/default-source/coronaviruse/covid-
strategy-update-14april2020.pdf?sfvrsn=29da3ba0_19 (accessed
May 23, 2020).
[11] A. Milenkovic, D. Jankovic, and P. Rajkovic, "Extensions and
Adaptations of Existing Medical Information System in order to
Reduce Social Contacts During COVID-19 Pandemic",
International Journal of Medical Informatics, Elsevier Press,
vol. 141, 2020, doi: 10.1016/j.ijmedinf.2020.104224
[12] P. G. Walker, C. Whittaker, O. Watson et al., "The Global
Impact of COVID-19 and Strategies for Mitigation and
Suppression", Imperial College London, 2020, doi:
10.25561/77735
 Smart Health Services for Epidemic Control
Aldina R. Avdić1, Ulfeta M. Marovac2, Dragan S. Janković3
Abstract – At the time of an epidemic, it is necessary to have
information about the epidemiological situation from verified
sources. The lack of such information can lead to the spread of
news from non-objective sources, irresponsible behavior of
citizens, but also the spread of panic. Epidemic control is one of
the topics in the field of smart health within smart cities. Based
on electronic health records (EHRs), various information on the
epidemiological situation in healthcare institutions can be
extracted. This paper proposes a framework that enables several
smart health services intended for epidemic control and ways to
implement them. Examples of the realization of these services are
given, based on EHRs from Health Center Nis.
Keywords – Epidemic control, Smart health services, e-Health,
Electronic medical reports.
I. INTRODUCTION
Today, information and communication technologies (ICT)
are represented in various spheres of human activity,
including solving everyday problems of urban residents.
Examples of such problems that reduce the quality of life in
cities are traffic jams, inefficient use of resources and the like.
The result of efforts to solve these problems using modern
technologies has led to the creation of a new concept whose
description follows.
Smart city - A city that uses ICT to improve the living
conditions of its citizens. The concept of a smart city, in
addition to IC technologies, uses public resources and social
information in order to increase the quality of life and public
administrative services. Jimenez et al. in [1] defined a smart
city as “a developed urban area that creates sustainable
economic development and high quality of life by excelling in
multiple key areas: economy, mobility, environment, people,
living, and e-government”. Hence, improving services that
contribute to better health of residents is also one of the key
tasks of a smart city, and a prerequisite for solving them are e-
Health and m-Health.
E-Health (electronic health) is innovative and more
efficient provision of health services with the help of modern
technologies and with a very high degree of system
integration, and the possibility of mobility of both doctors and
patients. Examples of e-Health are reflected through the
1Aldina R. Avdić is with State University of Novi Pazar,
Department of Technical Sciences, Vuka Karadžića bb, 36300 Novi
Pazar, Serbia, E-mail: apljaskovic@np.ac.rs
2Ulfeta M. Marovac is with State University of Novi Pazar,
Department of Technical Sciences, Vuka Karadžića bb, 36300 Novi
Pazar, Serbia, E-mail: umarovac@np.ac.rs
3Dragan S. Janković is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: dragan.jankovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
possibility of electronic access to patients' health records,
telemedicine, health knowledge management, mobile health,
etc. In this way, easier and more successful treatment of
patients and reduction of administrative restrictions when
providing medical services are achieved. The advantages of e-
Health are electronic monitoring and recording of patients'
health status, access to this data anywhere and anytime, as
well as fast transfer of information to users via telemedicine
and internet services.
M-Health (mobile health) encompasses a range of health
and medicine services that are realized by combining medical
technologies and mobile and ubiquitous computing.
S-Health (smart health). The definition is similar to the
definition of m-Health, with the emphasis that it is used within
smart cities. In addition, s-Health and m-Health differ in the
source information they use, and the flow of that information.
For m-Health, the source information comes from the user /
patient, while for s-Health, in addition to this information, it
also uses collective data obtained from the infrastructure of
the smart city. As for the difference in data flow, it refers to
the fact that after processing in m-Health, the feedback is
returned to the user, and in s-Health to the user, but also
affects the collective data of the smart city [2].
Epidemic control. A special area within smart health is
dedicated to epidemic control. The existence of electronic data
on the health status of citizens, along with methodologies for
their use and the collective intelligence of the city, improve
the state's competencies in detecting and controlling
epidemics. Citizens' activities and locations can be used to
detect potentially new cases during an epidemic, effectively
identify high-risk sites, and successfully manage an epidemic.
Such methods are also applied in the detection and
organization of other widespread health risks such as, for
example, air pollution [3, 4].
At the time of writing this paper, there is a current
pandemic of a COVID19 disease caused by new type of
corona virus SARS-CoV-2. So, we are witnessing how much
news on this topic are interesting for the public, how much
there is talk of matching real data with those from public
information systems and the like. The existing information
system in Serbia, covid.rs, provides information of the number
of deaths, the number of infected and the number of patients
on a respirator per day [5]. Until 10.6. it was possible to
monitor the epidemiological situation in the cities, but not
after that. On the other hand, the epidemic curve is most often
tracked on sites such as Worldometers and
Endcoronavirus.org, etc. [6, 7].
The motivation for writing this paper stemmed from the
fact that the existing public health services for epidemic
control in Serbia are insufficient and could give citizens more
details about the epidemic than they provide today. Our
contribution is in identifying the relevant functionalities that
46
smart health systems for epidemic control should provide and
the proposal, design and development of a smart health
platform for epidemic control and its most important services.
The platform we propose works in real time, directly takes
data from the medical information system and EHRs and
ensures a large number of details, such as the prevalence of
the virus by age structure, health facilities and an overview of
the number of infected in the appropriate time interval.
The rest of the paper is organized as follows: the second
section describes smart health solutions and services in related
research, the third section describes the services within the
epidemic control platform and describes how they are
implemented, while fourth section gives an overview of their
use. In the end, the conclusion and directions of further
research are given.
II. RELATED WORK AND STATE OF THE ART
Smart health platforms that solve various problems using
information technology have been described in numerous
papers. The paper [8] describes a smart health care model with
a decision support system model approach in public health
service. The paper [9] describes IoT-based (Internet of
Things) smart health services in order to support citizens in
cases of health problems due to air pollution, and [10]
describes a smart health platform that enables numerous
services for interaction with patients.
Papers about the use of information technology in epidemic
prevention and control vary from those aimed at finding a
mathematical model and simulating the spread of the
epidemic [11], to those describing platforms to inform citizens
about the epidemic, and using technology to monitor the
location of patients [12]. An IoT-based epidemic control
platform is described in [13]. The paper [14] proposes the
novel paradigm to control the Chickungunya virus. This
paradigm has advantages such as timely recognition of
infection and reducing the spread of infection, enhancements
in the analysis process, monitoring patients regardless of their
location, benefit for the doctors to work efficiently and for
logical / precise decisions etc. The paper [15] analyzes the
differences in the use of information technology for epidemic
control in China and Western countries. The services available
to our citizens to follow the information about the Covid19
epidemic are mentioned in the previous chapter.
III. EHRS AND THE DATASET
Electronic health reports are made mainly according to the
internal needs of the hospital. They are needed by various
actors such as: medical staff, patients whose health is
documented, clinical research (medical researchers,
pharmacists, epidemiologists, etc.), hospital management to
monitor finances and inventory planning, budget, etc. EHRs
may contain numerical and textual information. Medical data
consist of structured, semi-structured and unstructured data
and therefore require more complex processing involving the
existence of appropriate lexical sources. The structural part
contains the values of certain variables, so that they are the
easiest to process (name, surname, year etc.). The semi-
47
structured part gives descriptive values for some parameters,
but the structure is still known (temperature, pressure,
laboratory analysis etc.). The unstructured part consists of free
text given by the doctor and consisting of symptoms, history,
observations, conclusions etc. Unstructured data contain
linguistically incomplete, informal and non-standard
abbreviations, which complicates computer processing and
analysis. For that, it is necessary to process the data in
advance before the analysis in order to enter them in a
standardized form.
At the time of writing, there is no publicly available corpus
of medical reports in Serbian in electronic form. We used
5000 medical records in the Serbian language from 32 clinics
belonging to the Health Center Nis (DZ Nis), collected by the
information system MEDIS.NET [16]. The corps includes
EHRs from 2018, which contains information about patients
suffering from measles, because there was an epidemic of this
disease in Nis at that time. This corpus is created in
accordance with ethical standards, with de-identification of
patients and medical staff.
Table I gives an example of the medical report we are
processing. In it, we can identify the structural part that
contains the date of the service, the name of the service, the
diagnosis, the diagnostic code, the organizational unit in
which the service was provided and the location of the
service. Also, this report contains an unstructured section
consisting of anamnesis. This part is more complex to process
because it needs to extract the relevant data and convert it to a
standardized format, suitable for further processing.
TABLE I
AN EXAMPLE OF PROCESSED EHR
Date of the service 02-03-18
Name of the service First examination of adults
Anamnesis povisena t 38.5 belicate naslage po ustima,
kasalj, pulmo b.o, hiperemija grla
(en. fever 38.5 whitish deposits in the mouth,
cough, pulmo b.o, hyperemia of the throat)
Diagnosis Morbili – measles
Diagnosis’ code B05
Organizational unit of General medicine
the service
Location of the service Central building
I. THE ARCHITECTURE OF THE PLATFORM
Based on the data available from the set of EHRs, it is
possible to create several services that would provide insight
into the epidemiological situation in real time and be a valid
means of informing citizens. These services are based on
medical data mining of medical reports. They could be
displayed on a public health portal (Fig. 1) and would include
the following use cases:
a) a report on how many people have a specific diagnosis in
the city, on a daily, weekly and monthly basis. Based on this
service, a visitor to the portal would be aware of the epidemic
situation in his city, and whether the number of patients is
increasing or decreasing. This report is created based on an
analysis of the data obtained by entering in the structural field
for the date of the medical service from the EHRs (Fig. 2a).
 • Reduction to the one alphabet, abbreviation
processing and tokenization. As these EHRs are
written in Serbian, the free text can be found in Latin
and Cyrillic. In order to transform the data into a
standard format, in this step the text is translated into
Latin, with special regard to letters with diacritical
marks. After that, the abbreviations are marked, and
then the sentence is divided into tokens.
• Deleting stop words - this step eliminates words that
do not carry meaning.
• Determining negation - only a few negation symbols
Fig. 1. Smart health framework for epidemic control are used in medical reports, so the negation mark is
attached to the close word, to indicate the possible
b) a report on the number of patients with a specific absence of symptoms;
diagnosis, especially by health stations within the city. This • Reduction on the basis - since the Serbian language
would make the citizen aware of how many patients there are has a rich grammar, words can be found in various
in his immediate vicinity. This report is created based on the forms, it is necessary to reduce on the basis. In the
analysis of data obtained by entering in the structural field for absence of a morphological dictionary, as well as for
the location of the medical service from the EHRs (Fig. 2b). faster results, the base may be a prefix of length n or
c) a report on the presence of the disease in different groups stem;
based on age. Based on this, the citizen would be informed • Classification (labeling). After preprocessing the
whether he or one of his relatives belongs to the risk group text, classification can be performed. Classification
affected by the virus. This report is created based on an can be done using machine learning methods,
analysis of the data obtained by entering in the structural field taggers, but also rule based methods, if symptoms are
for the patient's date of birth from the EHRs (Fig 3). marked in the data model.
d) a report on the most common symptoms in patients These steps of normalization of medical reports in Serbian are
diagnosed with the disease. Based on this report, the citizen described in detail in [17]. The symptoms shown in the
would be more informed about the disease and could diagram are labeled with methods shown in the paper [18].
recognize the symptoms and in that case contact a doctor. This
report is created based on the analysis of data obtained by
entering in the non-structural field for leaving doctors' II. THE REALIZATION OF SMART HEALTH
comments, from the EHRs. With adequate processing of this SERVICES
text, the symptoms can be extracted from this filed and can be
shown on the diagram (Fig 4). The epidemic control service shows the situation about the
e) A simple questionnaire on the patient's current health number of patients who have the appropriate diagnosis, how
condition (presence / absence of symptoms) to check whether old they are, where they came for an examination, and what
to visit to the medical station for treatment or not. symptoms they had.
All these services would enable the citizens to be constantly
up to date with the epidemic and take measures to avoid or
treat the disease, and in that way the consequences of the
epidemic will be reduced.
The first three services require data analysis and
visualization, which is not demanding, while the fourth
requires specific word processing for proper symptom
labeling. Analyzing the non-structural part of the EHRs we
used, we came across abbreviations, spelling mistakes,
different word forms and synonyms for the same symptoms.
The anamnesis should be cleared of words that have no
meaning, and words of meaning should be reduced to the
same form. Abbreviations should also be processed and
stored. There are also negative symptoms in the anamnesis, so
the service would not show the true number of patients who
have a symptom if the negation was not considered.
Natural Language Processing (NLP) techniques are
required to extract information in the free text of the EHRs.
The steps necessary for the extraction of information, in our
case of symptoms, for the implementation of the fourth report
are:
Fig. 2. Chart of patients (a) by months and (b) by health facilities

48
 For the realization of this service, it is necessary that a part
of the data stored in the EHRs in hospital systems be publicly
available in order to extract data from them, i.e. the
appropriate knowledge. Fig. 2 shows an overview of the
number of patients by health stations and by months, the Fig.
3 presents patients by age groups, and Fig. 4 shows the
number of patients with the most common symptoms (rash,
cough, fever and pharyngeal hyperemia are extracted).
Fig. 3. Review of patients by age
Fig. 4. Review of most common symptoms
III. CONCLUSION
The realization of health services is an important
component of a smart city, especially in the field of epidemic
control, in order to raise citizens' awareness of protective
measures and to monitor the epidemic in their place and call a
doctor on time. The paper proposes the necessary health
services that would enable citizens to follow the epidemic
situation in real time. These services are based on the existing
information infrastructure, do not require special investments,
and provide the population with a lot of verified information
in the field of epidemic monitoring because they are
centralized. Some of the proposed services have been
implemented and briefly presented in the paper.
ACKNOWLEDGEMENT
This paper is partially supported by the Ministry of
Education, Science and Technological Development of the
Republic of Serbia under projects III44007 and ON 174026.
49
REFERENCES
[1] C. E. Jiménez, F. Falcone, A. Solanas, H. Puyosa, S. Zoughbi
and F. González, “Smart Government: Opportunities and
Challenges in smart cities development”, In Civil and
Environmental Engineering: Concepts, Methodologies, Tools,
and Applications, IGI Global, pp. 1454-1472, 2016.
[2] A. Avdić, D. Janković, “The Importance of e-health in the
Concept of Smart Cities”, In 17th International Symposium
INFOTEH-JAHORINA, pp. 383-387, 2018 (In Serbian).
[3] M. Al-Azzam, M. Alazzam, “Smart City and Smart-Health
Framework, Challenges and Opportunities”, Int. J. Adv.
Comput. Sci. Appl, vol. 10, no. 2, pp. 171-176, 2019.
[4] J. Dutta, F. Gazi, S. Roy, and C. Chowdhury, “AirSense:
Opportunistic Crowd-Sensing Based Air Quality Monitoring
System for Smart City”, In 2016 IEEE SENSORS, IEEE, pp. 1-
3, 2016.
[5] COVID19, https://covid19.rs/ (last visited 20.7.2020)
[6] Worldometers, https://www.worldometers.info/coronavirus/
(last visited 20.7.2020)
[7] End Coronavirus, https://www.endcoronavirus.org/countries
(last visited 20.7.2020)
[8] A. M.H. Pardede, “Framework for Patient Service Queue
System for Decision Support System on Smart Health Care”,
2018.
[9] A. Avdić, E. Kajan, D. Janković, and D. Avdić, “Towards
Context-Aware Smart Healthcare Platform”, International
Journal of Electrical Engineering and Computing, vol. 3, no. 1,
pp. 26-31, 2019.
[10] A. Avdić, U. Marovac, and D. Janković, “Interaktivna
softverska platforma za realizaciju javnih zdravstvenih servisa”,
6. Međunarodna konferencija Upravljanje znanjem i
informatika, pp. 46-52, Kopaonik, 2020 (In Serbian).
[11] J. Wang, X. Wang, and J. Wu, “Inferring Metapopulation
Propagation Network for Intra-City Epidemic Control and
Prevention”, In Proceedings of the 24th ACM SIGKDD
International Conference on Knowledge Discovery & Data
Mining, pp. 830-838, 2018.
[12] G. Pan, G. Qi, W. Zhang, S. Li, Z. Wu and L. T. Yang, “Trace
Analysis and Mining for Smart Cities: Issues, Methods, and
Applications”, IEEE Communications Magazine, vol. 51, no. 6,
pp. 120-126, 2013.
[13] P. D.S. Tavares, E. B. Rodrigues, “IoT-Based Architecture for
Data Analytics of Arboviruses in Smart Cities”, In 2018 IEEE
Symposium on Computers and Communications (ISCC), pp.
952-957, 2018.
[14] S. Rani, S. H. Ahmed, and S. C. Shah, “Smart Health: A Novel
Paradigm to Control the Chickungunya Virus”, IEEE Internet of
Things Journal, vol. 6, no. 2, pp. 1306-1311, 2018.
[15] R. K.R. Kummitha, “Smart Technologies for Fighting
Pandemics: The Techno- and Human-Driven Approaches in
Controlling the Virus Transmission”, Government Information
Quarterly, 101481, 2020.
[16] A. M. Milenkovic, P. Rajkovic, T. Stankovic and D.S. Janković,
“Application of Medical Information System MEDIS.NET in
Professional Learning”, 19th Telecommunications Forum
(TELFOR) Proceedings of Papers, pp. 1474-1477, 2011.
[17] A. Avdić, U. Marovac, and D. Jankovic, “Normalization of
Health Records in the Serbian Language with the Aim of Smart
Health Services Realization”, Facta Universitas, Series
Matematics and Informatics, accepted for publishing, 2020.
[18] A. Avdić, U. Marovac, and D. Jankovic, “Automated Labeling
of Terms in Medical Reports in Serbian”, Turkish Journal of
Electrical Engineering & Computer Sciences, accepted for
publishing, DOI: 10.3906/elk-2002-9, 2020.
 Syntax Analysis of Serbian Language using
Context-free Grammars
Teodora Đorđević1 and Suzana Stojković2
Abstract — Syntax analysis is one of the major phases in
natural language processing. This paper presents a method for
syntax analysis of Serbian language based on the context-free
grammars. First, it describes building a POS tagger for the
Serbian language. Secondly, it defines a context-free grammar
for Serbian language. The syntax analyzer is implemented by
using NLTK tool. Experiments showed that proposed analyzer
creates correct syntax trees for 74% of tested sentences.
Keywords — Natural Language Processing (NLP), POS
tagging, Context-free grammars, Syntax analysis.
I. INTRODUCTION
Natural language processing (NLP) is primarily concerned
with enabling computers to understand and generate text
written in natural language. Typical NLP applications are:
- Information Extraction – is the process of extracting
valuable information from textual documents.
- Machine Translation – deals with applying software in
order to translate text or speech from one language to
another.
- Sentiment analysis – is contextual mining of text which
identifies and extracts subjective information from given
material.
- Text summarization refers to the technique of shortening
long pieces of text. The intention is to create a coherent and
fluent summary having only the main points outlined in the
document.
- Question answering – is concerned with building systems
which automatically answer questions posed by humans in a
natural language.
Syntactical analysis is a very important part in processing
natural languages. If a syntactical analyzer achieves good
results, it enables to create high-quality semantical analyzers,
as the next logical step in Natural Language Processing.
Syntax analysis is usually used in Information Extraction,
Question Answering Systems, Rule-based Machine
Translation, etc.
This paper presents a syntax parser for the Serbian
language developed by using NLTK [1]. In order to develop
the parser, two important problems should be solved:
implementation of a POS tagger and a definition of Context-
free grammar for Serbian language.
1Teodora Đorđević is with the Faculty of Electronic Engineering,
University of Niš, Aleksandra Medvedeva 14, 18000 Niš, Serbia,
E-mail: teodora.djordjevic@elfak.ni.ac.rs
2Suzana Stojković is with the Faculty of Electronic Engineering,
University of Niš, Aleksandra Medvedeva 14, 18000 Niš, Serbia,
E-mail: suzana.stojkovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
This paper is organized in the following way: Section II
explains existing algorithms for syntax analysis of natural
languages. Section III presents a method for Serbian language
POS tagger development. It describes corpora that are used in
this process. The Serbian language context-free grammar is
shown in Section IV. Section V contains some results of the
experiments and Section VI contains summarized results of
the proposed method and discusses a possible future work.
II. RELATED WORK
Syntax analysis represents a procedure in which a sentence
is assigned a certain structure. The idea of teaching a
computer to understand natural language is perhaps as old as
the computer itself. Also, the algorithms designed for syntax
analysis advanced over time. At first, algorithms used context-
free grammars [2] as a set of instructions for parsing.
Although, this seems logical, it is quite difficult to describe an
entire language via a set of rules. One of the biggest problems
with this approach is that no matter how many rules you add
to your grammar it doesn’t improve parsing. The effect of
adding more rules is rather unpredictable parsing caused by
ambiguity of language itself. So the next logical approach was
adding probability. Newer syntax analysis algorithms are
mostly based on statistics [3].
A. Parsing Based on Context-Free Grammars
- The first step in this approach is defining a grammar which
consists of a large amount of rules and after that using some
parsing algorithm. In the later phases, parsing algorithms
for artificial languages can be used. This can be a problem
because these algorithms are designed for languages that
are far less complicated than natural. Because of that, some
algorithms specialized for natural language parsing were
developed.
- When parsing with context-free grammars the searching
method is also very important.
For determining which algorithm should be used for syntax
analysis first we need to consider the search method. There
are two methods: bottom-up and top-down. The task of a
parser is to search through the space of potential syntax trees
so it can find the correct tree for given input. Top-down
parsers search for a syntax tree by trying to build it from the
starting symbol of grammar. The input of a syntax parser is a
sentence. The goal of syntax analysis using top-down search
is to find all the syntax trees whose root is the starting symbol
of the grammar and that those trees include all of the words in
sentence. Although efficient, the main problem with top-down
analysis is generating large amounts of syntax trees. On the
other hand, bottom-up analysis starts with first word in the
50
input array and tries to create a syntax tree from rest of the
words applying grammar rules, one by one. Parsing is
successful if a parser generates the syntax tree with the
starting symbol in its root.
After choosing a search method, we also need to consider
algorithms for searching syntax trees. As a solution to
ambiguity, there is a set of algorithms designed to solve this
problem. They are based on dynamic programming. Dynamic
programming approaches systematically fill in tables of
solutions to sub-problems. When complete, the tables contain
the solution to all the sub-problems needed to solve the
problem as a whole. The algorithms for parsing using this
approach are CKY [4], the Earley algorithm [5] and Chart
parsing [6].
B. Statistical Parsing
Statistical parsing is definitely the modern approach to
syntax analysis and one of its most important features is
solving the ambiguity problem. Sentences can often be
syntactically ambiguous and this is an enormous problem for
parsing. Statistical parsing solves this problem by calculating
the probability for every possible outcome. The result of the
syntactical parsing is then chosen as the syntax tree with the
largest probability. One of the ways to calculate probability
for every syntax tree that is generated is using Probabilistic or
Stochastic Context-free grammar (PCFG) [7]. The only
difference between regular CFG and PCFG is that in PCFG
every rule has its own probability.
Rule in this grammar is:
(1)
where p represents probability that for this non-terminal A is
used this production. Probability for every generated syntax
tree using PCFG is calculated as product of probabilities of all
applied rules.
III. POS TAGGING
Before syntax analysis is started tagging has to be done.
Tagging is very important because syntax rules cannot be
applied to regular words. Syntax rules are built using Part of
speech tags. The Serbian language has ten different parts of
speech. Those are: nouns, adjectives, numbers, pronouns,
verbs, adverbs, prepositions, conjunction, particles and
interjections. The classification represents choosing the right
class for a specific input. As we can see, there are a lot of
possible tags, so tagging is implemented using classification.
Classification is a supervised method of learning which uses a
corpus of words for training. This corpus of words contains
words with a correct tag for every word. After training, there
comes a testing phase. A small part of corpus is left out of
training and used later for testing the accuracy of the created
model.
Classification is implemented using nltk. Natural language
toolkit or nltk is one of the most popular tools for NLP
implemented in Python. The Naïve Bayes Classifier is used
for classification. The Naïve Bayes Classifier considers all of
the word features when deciding which tag will be assigned to
the input. In order to create model of classification we can use
to tag words, first we need to choose a corpus for training.
51
There are no many corpora for Serbian language. Two were
used for classification:
- SrpLemKor [8] – training with this corpus achieved an
accuracy of 92%. The problem with this corpus and model
is the fact that it only supported shallow tagging. It only
provided information whether a word is a noun, a verb, an
adjective… This corpus doesn’t have any information about
the form of the word. If we are only interested in part of
speech, this tagging will be enough. The problem with
syntax analysis is that it requires more information about
words than just part of speech tag. In the Serbian language
it is very important to know case of nouns, adjectives,
pronouns and numbers, for example. Classification is then
implemented with another corpus.
- srWac [9] – web corpus formed from .rs top-level domain.
The first version contains 894 millions of annotated tokens,
and later versions around 600 million. This corpus contains
all the necessary information for every part of speech, but
classification became more complicated due to the large
amount of possible tags. The accuracy achieved with this
corpus as the training set is 91%. It is important to know
that this accuracy is accomplished after reducing some
information from labels that are not important for syntax
analysis. Also, the training set contained around 8,5 million
words from this corpus.
One of the most important parts of this tagging is extracting
features from words. The Naïve Bayes Classifier as one entry
expects features of a word along with its correct label.
Features of the word are chosen to fit the Serbian language.
Some of the features are:
- Word itself
- Whether is the first word in sentence
- Whether is the last word in sentence
- Does it start with capital letter
- Does it all from small letters
- Prefix – length 1
- Prefix – length 2
- Prefix – length 3
- Suffix – length 1
- Suffix – length 2
- Suffix – length 3
- The previous word in the sentence if the word is not first
- The label of the previous word
- The next word in the sentence if the world is not last
- Whether the word contains a hyphen
- Whether it is a number
IV. SYNTAX ANALYSIS
Implementing the syntax analyzer using nltk, needs a
Context-free grammar of the target language. The definition
of the grammar includes:
- A starting symbol of the grammar
- Terminal symbols – strings which contain tags that tagger
recognizes
- Grammar rules that have only one non-terminal symbol on
the left side
Nltk expects that the word that is being analyzed is part of
the grammar. It is very inconvenient to make rules with every
word of one language. First, there are too many words and
also syntax rules cannot be applied to the words of a language.
This is why the terminals of the grammar contain tags instead
of words. Some examples of terminal symbols in the Serbian
language are shown in the Fig. 1.
Fig. 1. Terminal symbols of the grammar
The grammar rules can be divided into several logical parts
[4]:
- Structure of the sentence – The Serbian language is
complicated and one sentence can contain a few simple
sentences. Because of that, the first part of the grammar
is dedicated to sentence structure. The first rule says that
the starting symbol can be expanded as sentences
symbol, sentences symbol can then be expanded as one
sentence or an array of sentences, etc. A single sentence
can contain a subject or sentence members. Sentence
members then can be a predicate and helper words.
- Subject – There are two types of subjects in the Serbian
language – logical and grammatical [10]. There are
multiple rules for subjects according to the grammar
rules. The logical subject is most likely to be a phrase in
genitive, dative or accusative. The grammatical subject
represents the doer of the action and it can only be in
nominative.
- Predicate – can be a verb predicate or a noun predicate.
A verb predicate represents the action that a subject is
doing and rules for this predicate usually include
different combinations of auxiliary verb, main verb,
particles yes/no and self-pronoun. A verb predicate can
be surrounded with different verb additions. Those are
object and adverbial phrases. The object can be direct or
indirect. The direct object is in accusative without
prepositions and represents the sufferer of the action.
The indirect object can be in any other dependent case.
The rules for objects are designed to fit syntax rules for
the Serbian language. Adverbial phrases can describe
place, time, means or cause. There is a set of rules with
verb additions before and after a verb. Also, there are
rules for situations when there is more than one verb
addition. A noun predicate, on the other hand, includes
an auxiliary verb and a noun, adjective or pronoun. The
rules for a noun predicate include different combinations
with auxiliary verb and phrases in a few cases.
52
Fig. 2. The rules for adverbial provision
- Helper words – There are words in the Serbian language
that are not significant for syntax. Those are
conjunctions, particles, abbreviations, punctuation and
residual words (e.g. e-mails). All of these words are
possible expansions for the non-terminal called helper
words.
- Due to large number of rules and limited space all the
grammar rules cannot be shown. Several rules for
adverbial provision are shown in Fig. 2.
To implement a syntax parser using the created grammar, it
is necessary to choose which type of parser from nltk will be
used. This toolkit offers – Recursive descent parser and Shift
reduce parser. The recursive descent parser uses the created
grammar and applies a top-down search method. The
advantage of this parser is that it will always find all possible
syntax trees for a given input. The problem is the fact that this
parser starts from the starting symbol of the grammar and
because of that generates a large amount of syntax trees that
have no connection to a given input. The shift-reduce parser,
on the other hand, starts parsing from a given input and
searches the rules that fit this input. It uses a bottom-up search
method. The problem with this parser in nltk is that it doesn’t
use backtracking. This parser may not find a syntax tree even
if there is one according to the grammar rules. In order to
guarantee that our parser returns a syntax tree if there is one,
the recursive descent parser is used.
When the parser is initialized with a context-free grammar,
it can be used for parsing sentences. First, we need to divide
our input into tokens. After that, those tokens are forwarded to
the POS tagger. When the POS tagger returns tags for a given
input they are forwarded to a syntax parser and the output of
this parser is finally a syntax tree. Example of a syntax tree is
given in Fig 3.
The problem with this parsing is that it can generate more
than one syntax tree for one sentence. Even if often one of
those trees is correct, we must have a deep understanding of
the Serbian grammar to figure out which syntax tree is the
correct result.
Why is the result often more than one syntax tree? The
Serbian language is very complex and there are a lot of
situations where one tag can be covered with more rules.
Some examples of ambiguity are indirect object and adverbial
phrases, indirect object and logical subject...
 V. EVALUATION
The implemented syntax parser is first tested with a large
amount of sentences prepared to fit with the designed
grammar rules and they were all recognized.
Besides these sentences the syntax analyzer was also tested
by using arbitrary text. The text that was chosen is from the
book ,,Programming language Java with solved problems”
[11].
The results are shown in Table I. Sentences with one or
more correct trees represent 74% of the total number of
processed sentences. The remaining sentences (26%) weren’t
recognized by the parser at all, meaning that they did not
result in any syntax tree. There are 0 sentences with only one
generated tree that is wrong.
TABLE I
THE RESULTS ON ARBITRARY TEXT
Number of
Category
sentences in percent
Sentences with one or more
74%
correct tree
Sentences with a wrong tree 0%
Unrecognized sentences 26%
VI. CONCLUSION
In this paper is explained one way of creating a POS tagger
for the Serbian language using available corpora. It is
described how is created context-free grammar for the Serbian
language. Finally, it is presented creating of the syntax parser
using previously mentioned context-free grammar.
From all of the above, we can see that although this parser
can recognize a lot of sentences of the Serbian language
Fig. 3 Example of the generated syntax tree
53
correctly, there is still room for improvement, especially with
ambiguity. One of the main ideas that could solve the issue of
ambiguity is using Probabilistic Context-free grammars to
always determine only one syntax tree.
ACKNOWLEDGEMENT
This work has been supported by the Ministry of Education,
Science and Technological Development of the Republic of
Serbia.
REFERENCES
[1] S. Bird, E. Klein, and E. Loper, Natural Language Processing
with Python, Sebastopol, USA, O'Reilly Media, 2009.
[2] M. Stanković, S. Stojković, and Ž. Tošić, Compiler theory, Niš,
Serbia, Faculty of Electronic Engineering, 2018. (in Serbian)
[3] Jurafsky, and J. Martin, Speech and Language Processing, 2nd
edition, New Jork, USA, Prentice-Hall, 2009.
[4] D. H. Younger, “Recognition and Parsing of Context-Free
Languages in n3”, Information and Control, vol. 10, no. 2, pp.
189-208, 1967.
[5] J. Earley, An Efficient Context-Free Parsing Algorithm,
Doctoral thesis, Carngie-Mellon University, 1968.
[6] R. M. Kaplan, Natural Language Processing, New York,
Algorithmics Press, 1973.
[7] T. L. Booth, "Probabilistic Representation of Formal
Languages", 10th Annual Symposium on Switching and
Automata, pp. 74-81, Waterloo, Canada, 1969.
[8] SrpLemKor, http://www.korpus.matf.bg.ac.rs/SrpLemKor/,
05.06.2020
[9] srWaC – Serbian web corpus, 12.06.2020
http://nlp.ffzg.hr/resources/corpora/srwac/
[10] Ž. Stanojčić, Lj. Popović, Serbian grammar for high school,
Belgrade, Serbia, Zavod za udžbenike, 2011. (in Serbian)
[11] L. Kraus, Programming language Java with solved problems,
Belgrade, Serbia, Academic Mind, 2015. (in Serbian)
 Signaling in 4G/5G with NB-IoT
support in 5G Option 3
Martina Janakieska1, Pero Latkoski1, and Vladimir Atanasovski1
Abstract – This paper describes the signaling use cases in
4G/5G with NB-IoT support that is the new 3GPP radio
standard which “addresses” the requirements of the Inetrent of
Things devices. It provides improved indoor coverage low-flow
devices, low sensitivity, low power consumption, ultra-low cost
and optimized network architecture. NB-IoT is considered to be
the best solution for LPWA technologies.
Keywords – Data over NAS, Deployment scenario, 5G, NB-IoT,
RAT type, Signalling.
I. INTRODUCTION
It is expected that fifth generation of mobile networks will
completely transform the role that telecommunication
technologies have in society. 5G will support a variety of user
scenarios, enhanced mobile broadband, Massive Internet of
Things (mIoT) and Ultra Secure Low Latency. By this way, it
will bring significant changes to both business and our
everyday lives. 5G is considered to be a beginning of a new
era in the industrial revolution. Thanks to the functionality
that this technology brings, we can expect a great
development in all spheres of society. It is expected that 5G
will support IoT and Narrow Band - IoT use cases that will
have deep impact in our everyday life. NB-IoT and LTE-M
are 4G technologies, but also pay a crucial role in 5G system
to support 5G Low-Power Wide-Area Wireless Technology,
LPWA, use cases. According to 3GPP [1], LTE-M and NB-
IoT will be a part of 5G and they are the only 5G technologies
to support 5G LPWA use cases in the foreseeable future.
Furthermore, firmware of the most NB-IoT and LTE-M
devices today can be upgraded to support the 5G NB-IoT and
LTE-M features. Thus, it can be said that NB-IoT and LTE-M
are on the road to 5G. The remainder of this article is
organized as follows. Section 2 gives an overview and key
concepts of NB-IoT technology. Section 3 presents general
5G deployment scenarios. Section 4 provides description of a
signaling in 4G/5G. Finally, the last section conclude the
paper.
II. BASIC CHARACTERISTICS OF NB-IOT
NB-IoT is a cellular Radio Access Technology that is
based on Long Term Evolution technology, introdused by
3GPP in Release 13 [3], for LPWA Networks. The term
1Martina Janakieska, Prof. Dr. Pero Latkoski, Prof. Dr. Vladimir
Atanasovski are with University of Ss. Cyril and Methodius, Faculty
of Electrical Engineering and Information Technology, Karpos 2 bb,
1000 Skopje, Macedonia, E-mails: janakieska.martina@outlook.com;
pero@feit.ukim.edu.mk; vladimir@feit.ukim.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
Narrowband means that NB-IoT devices’s bandwidth is
maximum 200 kHz, which can coexist either in the Global
System for Mobile Communications (GSM) spectrum or by
occupying one of the legacy LTE Physical Resource Blocks as
in-band or as guard-band [2]. There are several inherent
features of NB-IoT technology that make it the best solution
for LPWA. Low power consumption is a prerequisite for
almost 80% of all LPWA use cases [3], ranging from
applications such as smart gauges, smart parking, etc.
Additionally, with the availability of massive connections,
everything around us can be connected smartly. This would
require 50,000 devices per cell, which means that the density
of households per square meter is 1,500 households, and that
each household will have an average of 40 devices [3]. From a
performance standpoint, the NB-IoT guarantees 20+ dB
coverage, approximately 1000x connections, and
approximately 10 years of using only 200 KHz bandwidth [2].
On the other hand, this is a licensed solution that offers
greater security. The most important characteristics of NB-IoT
are: Multitone transmission support, Power Saving Mode
(PSM) and Extended DRX (eDRX), Coverage Enhancement
Method.
A. Control Plane and User Plane Optimization
NB-IoT Support small data transmissions that are based on
Cellular IoT, on both User Plane and Control Plane. This kind
of transission supports small bursts of data efficiently while
guaranteeing the long-range coverage. In this way, NB-IoT
can support more than one data path in CP for the
transmission of user data which is carried by the signaling
messages managed by the Mobility Management Entity
(MME) node in Core Network. The procedures are optimized
to efficiently support the small data transfer, the mandatory
Control Plane Cellular IoT (CP CIoT) Evolved Packet System
(EPS) and optional, User Plane Cellular IoT (UP CIoT) EPS.
CP CIoT optimization encapsulates the packets in Non -
Access Stratum (NAS) by using control plane signaling, thus
reducing the number of control messages when transferring
less data. The use of CP CIoT will optimize energy
consumption, due to the reduction of signalization and the
reduction of "air time". Data transfer takes place in two
directions (uplink and downlink). Under this NB-IoT solution,
no data radio carrier is established, instead packets are sent
through a radio signal.
B. Data over NAS
MME supports small data transmission over NAS, S11-U
and SGi interfaces. This is a part of CIoT EPS optimization.
54
This feature describes the following aspects of small data protocol) [6]. It means that LTE eNodeB decides if uplink
transfer, IP data over NAS and S11-U interface and Non-IP traffic is routed via NR gNB and/or LTE eNB. Option 3a
Data over NAS (DoNAS) and SGi interface [5]. Non-Access differs from the previous one because data from NR is sent
Stratum is a functional layer in UMTS and LTE, between user from/to SGW via S1-U interface, and eNB and gNB only
equipment and core network and is used to manage the exchange control messages (Bearer switching in SGW).
establishment of communication sessions and to maintain Finally, in option 3x, gNB is master NodeB and decides if
continuous communications with the user equipment. Data uplink traffic is routed via NR gNB and/or LTE eNB.
over NAS is a mechanism designed for small data transfer,
over the S1 Application Protocol (S1AP) [6], in 3GPP
systems by using the existing signaling plane for small data
transfer. DoNAS utilizes the Signaling Radio Bearers (SRB)
to carry data from User Equipment (UE) to eNodeB. The user
data is transported throught the MME node, that encapsulates
user data in NAS Packet Data Units (PDUs). This proces
reduces total number of control plane messages. IP and non-IP
data can be sent over Non-Access Stratum. Small data transfer
is even further optimized using Non-IP Data delivery. SCEF –
Service Capability Exposure Function node, over the control Fig. 1. Option 3 deployment (sub)scenarios
plane, is introduced in 5G core network [7]. This node
delivers Non-IP data. Furthermore, SCEF provides an The dashed lines in Figure 1 refers to a Control Plane.
interface for the network services such as Authentication, Thus, benefits of 5G EPS are: Existing EPS system can be
Access Control or discovery. reused; Possibility of “Smooth” migration; CP/UP for
optimized traffic.
In the following Table 1, are presented the advantages and
III. GENERAL 5G DEPLOYMENT SCENARIOS disadvantages of 5G deployment scenarios.

3GPP supports both [4], 5G Core (5GC) network, as well as TABLE I

a new radio access technology named as 5G “New Radio” ADVENTAGED AND DISADVENTAGES OF 5G DEPLOYMENT
(NR), that uses completely new core network with new nodes, SCENARIOS
and 5G Evolved Packet Core (EPC). In the second option is
used the existing LTE (LTE-M) core, with New Radio Access Adventages Disadventages
Network, 5G NR RAN. It is possible to integrate elements of SA Simple to manage Cannot use LTE
different generations in different configuration with 5G [4]: deployments if NR is used
SA (Stand Alone) and NSA (Non-Stand Alone). SA scenario NSA Existing LTE May impact end user
uses only one radio access technology (5G NR or the evolved Deployments experience
LTE Radio Access Network) and the core networks are EPC Existing EPC Cloud support is optional
operated alone. NSA scenario combines NR radio cells and deployments
LTE radio cells using dual-connectivity to provide radio 5GC Cloud native New deployment is
access and the core network may be either Evolved Packet required
Core or 5G Core. There are few options for deploying 5G. In
NSA scenarios, existing EPS nodes should be upgraded to
support 5G functionalities. [4]: IV. EPC SIGNALLING IN 4G/5G USE CASES
• 5G data rates
• 5G subscription control In this section are described few use cases which show how
• Dynamic mobility switching the signaling messages are transmitted 5G test EPS network.
• Network Slicing – Dedicated Core, 5G Gateway All of the tests are executed by taking the traces for test NB-
selection IoT devices (on the corresponding network nodes) and
• RAN volume reporting of secondary Radio Access analyzing them in Wireshark tool. All of the cases refer to
Type (if it’s used) Data Over NAS scenario.
• Separation of Control Plane and User Plane
One of the migration steps, which will be taken into account A. S11-U Support
in this paper is EPS to NSA option #3. It requires the
development of specifications of NR as non-standalone The first use case refers to the signalling messages that are
access, as part of E-UTRAN connected to EPS, not the full 5G transmitted over S11-User Plane (S11-U) interface when the
system. Option #3 has three sub-options, shown in Fig. 1. In UE sends “Create Session Request Message” to the core
option 3, control messages from LTE RAN are sent to MME, Network. i.e. MME node. This node is responsible for
and user data is transmitted via SGW node in core network. managing UE access and mobility, as well as establishing the
LTE RAN communicates with NR via new X2 interface bearer path for UE's. This interface is introduced as a part of
(Stream Control Transmission Protocol – SCTP based the Control Plane CIoT EPS optimization for small data

55
transmission between the MME and S-GW. Moreover, CP
CIoT EPS optimization enables support of efficient transport,
of user data (IP and Non-IP) or SMS messages over control
plane via the MME without triggering data radio bearer
establishment [7]. Thus, Ericsson supports IP based Control
Plane CIoT EPS optimization, Data over NAS, over S11-
U/SGi interface, for IP and Non-IP data as well. This interface
is based on GTP-User Plane (GTP-U) protocol [7]. It is a GTP
version 1 user plane interface and carries small data sent and
received over NAS between the MME and S-GW [7].
Data over S11-U interface supports the following features [7]:
• Allocation and de-allocation of MME S11-U Fully
Qualified Tunnel Endpoint Identifier - FTEID for
S11-U data tunnel termination at MME
• Path management for the S11-U interface in MME
node
• Sending and receiving GTP-U path failure and error
indication on S11-U interface
• S11-U GTPU sessions are recovered after a session
manager restart
• New statistics are introduced to display information
on data packets sent or received over the S11-U per
UE, per PDN and per MME service
• New statistics are introduced to display information
data packet sent/received over S11-U per GTPU
service
IP data over NAS is transmitted over S11-U interface.
Signaling Radio Bearer (SRB) is used for all communication
over the air interface between UE and the eNodeB. The MME
does not send the S1 Initial Context Setup request to the
eNodeB, and the eNodeB does not send RRC re-configuration
to the UE as there is no Initial Context Setup at the eNodeB.
On the other hand, MME sends a NAS Attach Accept to the
UE in S1 Downlink NAS Transport message instead of S1
Initial Context Setup request. The UE sends NAS Attach
Complete in the S1 Uplink NAS Transport message. The
MME establishes the S11-U tunnel during S11 Create Session
Request/Response signalling. S11 Modify Bearer request sent
is not sent to the S-GW after Attach Complete is received
from UE [7]. Create session request is tested with NB-IoT
device on our, previously mention, test network. Our NB-IoT
device sends Create Session Request to MME node through
eNodeB. The results are shown Figure 2. S11-U data tunnel is
established with corresponding Fully Qualified Tunnel End
Point Identifier, IPv4 address and security key, when UE
sends NAS Attach Complete in the S1 Uplink NAS Transport
message. Moreover, the results show RAT-Type NB-IoT.
That means that SGW is configured to support NB-IoT
devices. In this way, MME sends RAT type NB-IoT to SGW,
while requests “Create Session”, instead of E-UTRAN Radio
Access Type.
B. NB-IoT Support – Handling NB-IoT RAT Type
To transfer data from NB-IoT devices, it is important that
the MME node is configured to support NB-IoT functionality.
MME can handle the case where Home Subscriber Server
(HSS) and/or GW not support NB-IoT RAT. Thus, MME can
makes a "bypass" to avoid the impact of SGW and HSS nodes
56
in the network. Due to this functionality, NB-IoT RAT
support is not required by SGW, HSS and other network
nodes, which means that on the Ѕ6a interface, from MME to
HSS, MME will not send NB-IoT RAT type to HSS, as well
as to SGW, through S11 interface.
Fig. 2. Create Session Request Test
However, in the previous scenario, shown in Figure 2, it is
noticed that the SGW node is configured to support NB-IoT,
thus, MME sends RAT Type NB-IoT to SGW in Create
Session Request message. HSS is a database that contains
Authentication Information and Subscriber's Data, where
MME shall take care of mobility of UE/Subscriber [7]. This
interface is used to Authenticate Subscriber, providing
services to subscriber, to store location information of
subscriber sent by MME. In case when the device needs to
update the location, MME sends the Update Location Request
to HSS via S6a interface. Within the message, MME sends the
RAT type to HSS. If HSS is configured to support NB-IoT
RAT type, then MME sends NB-IoT RAT type, otherwise,
MME sends RAT type E-UTRAN. Note that this scenario
doesn’t have impact on NB-IoT data transmission. Our tests
are made with HSS node that does not support NB-IoT. Thus,
when MME sends Update Location Request on S6a interface,
it still sends E-UTRAN RAT type. S6a Interface is diameter
based interface. It can be indicated by the MME node in the
ULR-Flags AVP, Figure 3. If HSS node support NB-IoT RAT
type, then the URL-Flags AVP will show RAT Type NB-IoT.
However, handling of NB-IoT RAT by MME has own
benefits. For example [7]: Enables the possibility to use NB-
IoT as radio for low cost, low complexity devices for low
throughput applications; MME (or combines SGSN-MME)
handles multiple access, now also including RAT NB-IoT;
Possible to deploy RAT NB-IoT without corresponding
support in HSS and/or GWs.
C. Power Saving Mode Test Case
Power Saving Mode facilitates battery life of device as
devices are supposed to be in sleep mode most of the time.
The PSM mode is similar to power-off but the UE is still
registered on the network. Power Saving mode should be
supported by both, MME and UE. Device supporting power
saving mode asks MME for “Active Time” and “Extended
Periodic Tracking Area Update (TAU) Time” value during
attach or RAU/TAU procedure.
Fig. 3. NB-IoT Support on S6a interface
The UE activates PSM by including these two timer values
[7]. The T3324 is the Active Timer which defines the time in
which the UE stays active after idle mode, following the
Attach or TAU procedure. The Exended Periodic TAU Timer,
is T3412. It defines the extended time for an UE to send
periodic TAU. The UE requests for a T3324 Active Timer
value during Attach and TAU procedures. The MME allocates
the T3324 value to the UE. The T3324, determines the
duration in which the device remains reachable for
transaction during the transition from connected to idle mode
[7]. Active timer starts when the device moves from
connected to idle mode. When this active timer expires, the
device moves to Power Saving Mode. Moreover, MME
allows a value of ‘0’ for the T3324 timer. In this case the UE
enters the Power Saving Mode immediately. Our tests are
made with Active Timer set to ‘0’. From Figure 4 we can see
that only T3412 is present and it’s value is 30 min. The T3412
timer is also referred to as the periodic Tracking Area Update
(TAU) timer. Periodic tracking area updating is used to
periodically notify the availability of the device to the
network. The procedure is controlled in the UE. The value of
the T3412 is sent to the UE in the ATTACH ACCEPT
message through S1 interface [7]. Also, can be sent via
TRACKING AREA UPDATE ACCEPT message [7]. The
UE should apply this value in all Tracking Areas of the list of
tracking areas assigned to the UE, until a new value is
received.
V. CONCLUSION
This paper has presented the basic overview of NB-IoT and
it’s characteristics. The idea was to present all (so far)
possible options for the implementation of the 5G network
and which option for the 5G EPC Non Stand Alone is most
appropriate. Further, were present significant signalization
processes in 4G/5G. But we have taken into account the
signalization by the current NB-IoT, because the signalization
is a broad term that covers many details. The basic signaling
57
processes with the existing core network are presented, with
some changes in the network nodes MME, HSS and SGW.
Those nodes are updated to support primary 5G features and
NB-IOT devices. All tests were performed with a test NB-IOT
device, but on a test network that has the same features and
nodes as the existing core network. The packages were
recorded in Wireshark in order to see the signaling messages
that are transmitted over the network through nodes and
interfaces, based on appropriate protocols.
Fig. 4. S1-MME Attach Accept – Timers
ACKNOWLEDGEMENT
This research was supported by the Ericsson
Telecommunications company, Macedonia and Faculty of
Electrical Engineering and Information Technologies. We
thank our colleagues from Ericsson Telecommunications who
provided test lab, insight and expertise that greatly assisted the
research, although they may not agree with all of the
interpretations of the paper.
REFERENCES
[1] https://www.ericsson.com/en/networks/trending/insights-and-
reports/nb-iot-and-lte-m-in-the-context-of-5g-industry-white-
paper
[2] Huawei, “NB-IoT – Enabling New Business Opportunities”,
Huawei Technologies, Bantian, Longgang District, 2015.
[3] P. Salva-Garcia, J. M. Alcaraz-Caledro, Q. Wang, J. B.
Bernabe, and A. Skarmenta, “5G NB-IoT: Efficient Network
Traffic Filtering for Multitenant IoT Cellular Networks”,
University of the West of Scotland, UK, December 10, 2018.
[4] Ericsson Telecommunications, “MakTel NB-IoT Pilot Project –
HLD”, , Macedonia, Rev PA1, January 2019.
[5] S. Tabbane, “5G Networks and 3GPP Release 15”, Nadi, Fiji
Islands, October 23-25, 2019.
[6] Ericsson Learning Services, “LTE L15 Air Interface, Protocols
and Procedures”, 2012, https://www.ericsson.com/assets/
local/portfolio/ericsson-academy/rnam/lte-l15-air-interface-
protocols-and-procedures.pdf
[7] Cisco Systems, Inc, “Ultra IoT C-SGN Administration Guide,
StarOS Release 21.6”, January 1, 2018.
 Supervision and Monitoring System of the Power Line
Poles Using IIoT Technology
Miloš D. Djordjević1, Jana M. Vračar1 and Aleksandra S. Stojković1
Abstract – In this paper, the development of an autonomous
system for the monitoring and supervision of the power line poles
is presented. It is a new approach to the use of Industrial
Internet of Things technologies in a practical and significant way
in the system of smart control systems. The subject of research is
the supervision and monitoring of the angle of slope and
microclimate using Industrial Internet of Things technology and
smart sensor nodes. The system is based on: a group of built-in
sensors (environmental: BME280, BH1750, and ADXL345 3-axis
accelerometer), a microcontroller with a peripheral interface
(PIC) as a core of low levels of the system, a server system and a
wireless Internet using the Global System of Mobile
Telecommunications (GSM) module with General Packet Radio
Service (GPRS) as a communication protocol.
Keywords – Smart control systems, Industrial Internet of
Things, Sensor technology, PIC microcontroller, Monitoring.
I. INTRODUCTION
Industrial Internet of Things (IoT) is focused on a variety of
manufacturing facilities (oil and gas systems, power systems,
etc.), empowering industrial engineering with sensors,
software, and big data analytics to identify problems on time
and create efficient, cost-effective and optimal processes.
Due to the criticality of such infrastructures, the slightest
mistake in the system design leaves consequences primarily to
human safety, and then to the environment and finances.
Therefore, concerning CIoT (Consumer-oriented IoT - CIoT),
IIoT applications are subject to particular limitations and
requirements regarding security, interoperability, scalability,
precision, accuracy, response, reliability. Operation
management in IIoT systems is often more complex than in
CIoT solutions and due to the almost inevitable abundance
and distribution of heterogeneous devices over long
geographical distances. One example of the application of IoT
technology in the industry will be presented in this paper,
based on the power line poles monitoring and monitoring
system. The implemented system monitors the inclination of
the power line poles, to predict the unintended interruption of
electricity supply due to the collapse of the line poles.
1Miloš D. Djordjević is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: milos.djordjevic@elfak.ni.ac.rs
1Jana M. Vračar is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia,
E-mail: jana.vracar@elfak.ni.ac.rs
1Aleksandra S. Stojković is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: aleksandra.stojkovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
II. THEORETICAL BACKGROUND
The number of utility poles (power line poles, power poles,
hydro poles) in the distribution network is large, and the
construction of the power poles is not as good as that of the
carrier poles in the transmission system. In many countries,
wooden poles are prone to rot and decay. Power poles can
bend, warp, or fall completely.
They are endangered by insects and birds, wild animals,
floods, strong winds and inclement weather in the form of
snow and/or rain ice storms. Often these power poles must be
repaired even before they collapse, in case of traffic accidents,
growth of seedlings/trees of inappropriate size near the poles,
as well as collapses in the event of an earthquake. Landslide
are also a particular risk for the poles. Rain and snow can
cause freezing, and asymmetric traction causes the poles to
tilt, indicating a safety risk for power poles.
In [1], the author introduced a street lamp monitoring
system. The system is based on the Raspberry Pi 3. Sensors,
such as an accelerometer that monitors the angle of inclination
of the lamp, a photoresistor (LDR, Light Dependent Resistor)
that serves as a sensor for measuring the intensity of
illumination of the lamp, and a current sensor for monitoring
changes in the current and power of the lamp are used. The
measured parameters are sent to the Cloud (to keep track of
them) and, if necessary, to react and repair the street lamps.
The paper itself pays little attention on how the system works,
specifically how measurements are made and what the period
between the two measurements is. There is no information if
the accelerometer is used, whether it is analog or digital. In
our paper, a 16-bit digital accelerometer is used to obtain an
accurate angle of inclination. Also, the authors used LDR as a
light intensity sensor (which affects on measurement
accuracy). Our system is more precise due to usage of 16-bit
digital light intensity sensor. Within the mentioned paper,
there is no information about which one was used as the
current sensor. In terms of the practical application of the
system, the authors did not provide any results within the
paper to confirm the authenticity of the presented system.
In the second paper [2], the author presented a system for
monitoring the inclination of electric poles, actually made a
comparison of a uniaxial and three-axis accelerometer when
measuring the slope of a power line pole. The system is based
on the BeagleBone Black development environment and the
ADXL335 analog accelerometer, which measures the slope of
the pole. The sensor used in our system, ADXL345 digital
accelerometer, is more precise than analog sensor, made with
MEMS technology.
58
 Fig. 1. Principle of the smart system for supervision and monitoring of the power line poles
An additional advantage, beside those mentioned above, is
lower price range because of microcontrollers used in our
system. Big disadvantage of these systems is need for Wi-Fi
or GSM/GPRS modules on each of the power line pole, which
increases the cost of the system. An advantage in our paper is
the use of far cheaper RF communication modules for sending
information about measured parameters from power poles.
Also, they do not require internet access.
This paper presents an example of a system for supervision
and monitoring of the power line poles using IIoT technology.
The presented system consists of three levels. The first level is
in charge of measuring the parameters of the power line poles.
The second level is used for collecting the measured data and
sending it to the Cloud. The third or highest level of the
system is in charge of monitoring the parameters and
responding. Also, only operators, or people who monitors the
operation of the entire system, have access to this level.
III. WORKING PRINCIPLE OF THE IIOT SMART
SYSTEM
A. Design of the System
Communication between the power line poles (Level 1
measuring nodes) and the control nodes (Level 2 nodes) are
realized by the RF module (RF communication), and can be
briefly described as follows:
The communication is bidirectional (there are two RF
modules (RX and TX)), and therefore the measuring node (at
Level 1) knows when to do the measurement and when to
send data (to avoid unnecessary power consumption) and then
control node (Level 2) sends a message from its RF module
(Fig. 1). Communication speed is 2400 bps and UART
(Universal Asynchronous Receiver/Transmitter) or serial
communication is used.
59
The microcontroller PIC16F628A [3], which represents the
core of the entire Level 1, manages the sensor block, which
serves for measurements and observations. Also, the
communication block is realized using the RF modules. The
sensor part of Level 1 consists of the following sensors:
- BME280 (measures temperature [°C] or [°F],
Relative Humidity [%] and Atmospheric pressure [hPa] or
[mBar]) [4],
- BH1750 (measures light intensity [lx]) (optional) [5],
- ADXL345 (measures slope angle [°]) [6].
The PIC microcontroller constantly "listens" (or wakes up
from sleep mode) when something is received in the RX
buffer. When an adequate message from Level 2 is detected
for the required measurement, the microcontroller sends
messages to the sensors. The message contains a sensor ID
and a request to measure a specific parameter like
temperature, humidity, pressure or slope angle. When a sensor
is "recognized" by an ID based on the control byte (which
parameter is being measured) it performs a measurement, e.g.
BME280 (Temp, Hum, and Press). Subsequently, the
microcontroller must send a request to the sensor to read data
from its memory again. The sensor sends the measured data to
the microcontroller via the I2C bus.
Level 2 is the link between the power line poles and the
control center. The centerpiece of this level is the
PIC18F45K22 [7] microcontroller, which communicates with
Level 1 and Level 3. All communication blocks are using RF
communication because every power line pole (Level l)
consists of the RF module. The second block was
implemented using Wi-Fi or GSM/GPRS modules, to send the
data to the Internet, based on which dispatcher has an insight
into the state of the power line poles. For the microcontroller
to be able to monitor the period between two measurements, it
is necessary to use the RTC (Real Time Clock) module which
calculates the time that was previously set, as well as the
period between the two measurements.
 Fig. 2. Block diagram of intercommunication between system blocks

Level 3 represents the main (control) part of the whole
system. From this level, messages are sent to the control
nodes (at Level 2) when it is necessary to perform a specific
action or to instruct people who are working to repair the
damage caused on the power line poles.
Part of the systems comprising Levels 1 and 2, as well as
the part related to sending data from Level 2 to the Cloud is
based on the M2M communication protocol [8], similar to
some of our previous publications related to the application of
IoT technology [9-10].
B. Usage Scenario
The realization of the smart system for supervision and
monitoring power pole lines can be seen on the blocking
scheme for the whole system shown in Fig. 2. The block
diagram clearly shows how the system is divided into separate
blocks, each of them operates independently, but ultimately
they represent the whole system. On the system startup, the
microcontroller on Level 2 communicates with the RTC
module via the I2C bus, to obtain timing information.
Thereafter, it is examined whether the time has elapsed when
parameter measurement is required. On that occasion, the
microcontroller from Level 2 sends a message to all
measuring nodes at Level 1 a measurement should be made.
For informing all of the nodes a measurement is required, a
message is sent with a unique address, because every node has
a different ID key. The message is sent by an RF module
operating at a frequency of 433 MHz, and the distance ranges
from 10 m to 7.5 km [11], depending on the need, how far the
control node is away from the measuring nodes, or the power
line poles.
When a measurement node at the Level 1 receives a
message intended for it, the measurement of required
parameters is starting. After that, the data is converted to be
sent as a single message. When there is a request for sending
measured parameters, it`s necessary that the body of the
message includes the measuring node address, to keep track of
the exact measuring node or power line pole.
Once a message has been received from a particular
measurement node, it is necessary to parse the message to
extract useful information, which will be sent to the Internet,
so the control center can monitor the results and react as
needed.
When data is sent from Level 1, the microcontroller on
Level 2 observes if the key corresponds to a key defined for a
particular line pole. If it is the same, the data are received.
When the data reception is completed, a message parsing is
followed to extract integer values for each parameter from the
received string. After parsing, the microcontroller writes the
values for each parameter into the variables defined for it.
This concludes the section on RF communication. The next
step is defining the WiFi/GPRS module and sending the data
to the Cloud or database on the server (arrows from Level 2 to
the Cloud) using the MQTT protocol.
Further, the microcontroller checks whether the slope is in a
desirable range (in this case this is set to an angle less than
15°). If the slope angle is greater than 15°, a request for re-
measurement and data transmission is sent every 5 minutes. If
it is evaluated to be in danger of falling, the request is sent to
the team, which is working to repair the damage and fix the
problem.

60
 Fig. 3. Results obtained by implementing the realized system
IV. EXPERIMENTAL RESULTS
The parameters are measured during a period of 24 hours,
from 5th December to 6th December 2019. Based on the
simulation of the power line poles in real conditions with the
resulting slope problems, the results were obtained in
laboratory conditions. Namely, the obtained results (Fig. 3.)
present the functioning of the realized system.
The number of measured values (every 5 minutes) is greater
in the period when there is a higher slope of the power pole
than optimal value (slope angle greater than 15°) until the
problem is resolved. When there is no problem occurred,
parameters are measured in 20 minutes time period.
V. CONCLUSION
This paper describes a reconfigurable smart system for
supervision and monitoring power line poles as part of a
smart city. The presented system is autonomous, which
means that it monitors the atmospheric parameters, where
the parameter of the slope angle is the most important and,
following the requirements of the control center, monitors
and measures the parameters. Based on changes in the slope
angle of the line poles, the adequate measures are being
taken to prevent a possible problem that could result from a
fall of the power line poles.
ACKNOWLEDGMENT
The authors would like to thank the Ministry of Education,
Science and Technological Development, Republic of Serbia,
for financial support.
61
REFERENCES
[1] A. Shaikih, M. Thapar, D. Koli, and H. Rambade, “IOT Base
Smart Electric Pole”, ICECA'18, Conference Proceedings of the
2nd International Conference on Electronics, Communication
and Aerospace Technology, pp. 594-597, India, 2018.
[2] L. De Yong, N. A. Mohamed Radzi, and Y. E. Jalil, "Pole Tilt
Monitoring System Using MEMs Accelerometer and
Beaglebone Black", International Journal of Electrical
Engineering and Computer Science, vol. 11, no. 2, pp. 658-664,
2018.
[3] PIC16F628A:https://web.mit.edu/6.115/www/document/16f628.
pdf, [On-Line], Data Sheet. Accessed: 15.03.2020.
[4] BME280 sensor – BOSCH Sensortec: https://cdn-
shsh.adafruit.com/datasheets/BST-BME280_DS001-10.pdf,
[On-Line], Data Sheet, Accessed: 15.03.2020.
[5] BH1750FVI: Sensor Sensor ICS – Mouser Electronics:
http://rohmfs.rohm.com/en/products/databook/datadatas/ic/sens
or/light/bh1721fvc-e.pdf, [On-Line], Data Sheet, Accessed:
15.03.2020.
[6] ADXL345: https://www.analog.com/media/en/technical-
documentation/data-sheets/ADXL345.pdf, [On-Line], Data
Sheet, Accessed: 15.03.2020.
[7] PIC18F45K22:
http://www.microchip.com/wwwprowwwpr/en/PIC18F45K22,
[On-Line], Data Sheet. Accessed: 01.07.2019.
[8] A. Prijić, Z. Prijić, D. Vučković, and A. Stanimirović, “AADL
Modeling of M2M Terminal”, Microelectronics Conference
(MIEL 2010), 16-19 May 2010, pp. 373-376.
[9] M. Đorđević, D. Danković, “A Smart Weather Station Based on
Sensor Technology”, Facta Universitatis, Series: Electronics and
Energetics, vol. 32, no. 2, pp. 195-210, June 2019.
[10] M. Djordjevic, B. Jovicic, B. Milosavljevic, V. Paunovic, and
D. Dankovic, “Smart House System Based on GSM Mobile and
IoT Technology”, IEEESTEC - 11th Student Projects
Conference, Niš, Serbia, 2018, pp. 29-34. (In Serbian)
[11] Wireless RF module SX1278 - https://cdn-
shop.adafruit.com/product-files/3179/sx1276_77_78_79.pdf,
Data Sheet, Accessed: 15.03.2020.
ENGINEERING EDUCATION
 Low Cost Laboratory Environment for the Use of Optical
Methods for Transmission of Audio Signals
Tsvetan Valkovski1 and Kalin Dimitrov2
Abstract – In this paper we will discuss the possibilities of
transmitting audio signals through optical and optoacoustic
methods. An analysis of existing methods and the advantages of
transmitting audio signals through optical methods will be made.
A model of a Li-Fi system for transmitting audio signals and
building one with a minimum number of electrical elements is
considered. This is in order to implement the system presented in
the laboratory sessions in front of students, the opportunity to
build such systems from them as a practical exercise and the
consideration of the possibilities for transmitting audio signals
through optical methods.
Keywords – Acoustic, Audio, Optical, Optoacoustic, Li Fi.
I. INTRODUCTION
People have been interested in the transmission of audio
signals without delay and with the highest possible quality for
many years. The goal has always been to convey the original
recording to the listener as much as possible - without
additional noise, without "coloring" the sound picture, exactly
what is recorded.
Conventional methods for transmitting audio signals over
short and long distances are well known to us.
If we divide the audio signals into analog and digital, we
could consider different cases.
In the case of analog audio signals, it is accepted to transmit
the audio signal via copper cables and the audio signal itself is
electrical.
With digital audio signals there is a possibility for
transmission on different media - copper, optical, radio, etc. In
this article, digital audio signals will not be of interest because
the goal is to retransmit the original analog audio signal with
as little processing as possible along the transmission path.
II. THEORY
The principle of Li-Fi uses changes in magnitude - the
modulation of light for data transmission. Generally, when the
LED is on, it transmits a little 1 and when it is off, it transmits
a little 0. The frequency changes so fast that it will remain
invisible to the human eye continuing to receive a continuous
white light. Li-Fi only works with a white LED, simply
because the LED turns on and off in astroboscopic way. Fig. 1
1Tsvetan Valkovski is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: tsvalkovski@gmail.com
2Kalin Dimitrov is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: kld@gmail.com
978-1-7281-7142-5/20/$31.00 ©2020 European Union
shows a simple diagram of Li-Fi system model with LEDs. As
shown in the figure, the data from the source is modulated to a
digital base signal and it is sent by modulating the visible light
with LEDs. The most receiving side, the photodiodes are
integrated as a photodetector and the LED communication
circuit of the demodulator, which is capable of converting
light into an electrical signal. The most the signal attenuation
in the Li-Fi system with LED circuit is urgent issue and thus
the electrical signal is amplified before demodulation [1]. The
basic principle of Li-Fi is shown in Fig. 1.
Fig. 1. Block scheme of basic Li-Fi system
The official standardization of Li-Fi was in December
2011, for VLC technology – visible light 802.15.7 standard
completed the project development of the physical layer
(PHY) and the Media Access Control (MAC) [2-3].
If you pay attention to the light source, it can be not only an
LED diode with visible light, but also a laser, lamp, infrared
diode. After reviewing the various LEDs, Laser diodes, and
Tunable lasers with their atomic structure, electron flow, and
the energy level transitions that produce different wavelengths
of light in semiconductor materials, it is concluded that for the
purposes of our study the most suitable is an LED diode with
visible light [4-13].
The features of Li-Fi technology are the capacity,
efficiency, safety and security. Li-Fi provides with greater
bandwidth, which is easy to use and free because it is
unlicensed. Li-Fi provides greater data density compared to
that of standard wireless Wi-Fi technology. The data density
is about 1000 times than other technologies. This is due to less
interference of light than radiofrequency waves. The data
density and bandwidth the output speed is very high. The
system would be cheap as it requires a smaller number of
components. No external power input is required for this
technology and moreover, LED illumination is already
efficient. It eliminates any health hazards caused by
radiofrequency waves. The technology is safe and non-
hazardous because the use of light cannot interfere with any
electronic circuitry. Hacking or data theft is negligible
compared to other wireless technologies since the range of
data transmission is confined to a certain visible area [14-17].
65
 III. OPTICAL METHODS FOR TRANSMISSION OF In the part of the "Audio output" we must have a receiver of
AUDIO SIGNALS an audio signal, which could be anamplified speaker,
multimedia speaker system, phones etc.
After examining the individual blocks, it is time to pay
The article will conduct a study to build a basic Li-Fi
attention to the necessary elemental basis for building the
laboratory environment to demonstrate the transmission of
laboratory environment.
audio signals by optical methods.
The needed elements are:
The experimental part includes a study of the
- The required wiring is standard stereo audio cables and
developments so far, a study of the individual components
3.5 mm jack connectors, known as AUX cable.
needed to build the environment and the selection of
- AnLEDdiode with characteristics 5mm dimensions,
specific, suitable electrical elements and sensors.
6000-15000k nm wave length, 14000-16000 MCD, 3-
From the developments considered so far, the basic
3.2V Voltage and 20mA current.
model of this laboratory environment is presented in
- A resistor with 220 Ohms resistance.
Fig. 2.
- A solar cell panel with characteristics 10x10 cm
dimensions, 6V maximum voltage, 100mA maximum
current.
The possibilities to improve the Li-Fi laboratory
environment are to use a high intensity light source like a
laser, to use a very sensitive solar panel, have a higher audio
input and reduce interference by moving into a dark room.
In further developments, the system could be improved by
increasing the element base in the transmitter and receiver in
order to improve the quality, speed and capabilities of audio
Fig. 2. Block scheme of the basic model of Li-Fi laboratory signals[10].But the basic idea of transmitting audio signals by
environment optical methods remains the same, shown illustratively in
Fig. 4.
In the part of the "Audio input" we must have a source
of audio signal, which could be a smartphone, mp3 player,
laptop, computer, etc.
“Modulation circle LED” is a block which is much
more complex to implement. In it, we must convert the
input audio signal so that we can add power to it, which
will be the carrier signal. At the same time, we should not
degrade the quality of the audio signal, but we should also
strive for a higher power of the carrier signal, which in
turn leads to transmission over longer distances. This is
derived from the laws of light propagation in the free
atmosphere, the higher the power we have, the greater the
distance we will transmit the signal.An easy way to build
this block is through a suitable DC battery and a suitable
resistor, as shown in Fig.3.
Fig. 4. Illustration of proposed Li-Fi System

After building the system, we conducted several

Fig. 3. Wiring diagram of “Modulation circle LED”
experiments with a random audio signal and a fixed one from
a tone generator.
Table 1 and Fig. 5 show the initial experiment with a sine
test tone at 1 kHz. The greatest distance that we made
measurements at is 1 m and the SPL is 37 dB, that is a very
low sound level and we estimated that we would run the
experiment until the sound level dropped by 30 dB.

“Solar panel” block includes standard solar panel for visible TABLE I
light and stereo audio cable such as in the “Modulation circle THE SOUND LEVEL AT DIFFERENT DISTANCES
LED” block. When the solar panel receives a light signal, it
Z,cm 5 10 15 20 25 30
transforms it into an electrical signal and this signal is not the
carrier signal from the battery. This signal is exactly the audio SPL,dB 73,6 73,5 71,9 71,3 71,2 70,5
input signal. Z,cm 35 40 45 50 55 60
SPL,dB 70,1 69,7 68,4 66,2 64,1 58,9

66
 Fig. 5. Graphic representation of the readings for the sound level at Fig. 7. Graphic representation of the readings for the signal level at
different distances different distances

The next experiments we performed were at a frequency of

1kHz emitted by a tongener and measured on an oscilloscope
in order to increase the accuracy of the measurements.
At Fig. 5 we can see that the noise level decreases by less
than 10 dB at half a meter, which for the size of the LED and
the solar panel used is quite a distance, when visible light is
used and taking into account the dispersion and the
polarization.
Fig. 6 and Table 2 show the laboratory setting and the
obtained results at the signal level when changing the
distance.

Fig. 8. Laboratory environment for measuring the signal level at

15 cm and changing the angle of signal reception by changing the
angle of the solar panel cell

TABLE III
THE SIGNAL LEVEL AT 15 CM AND CHANGING THE ANGLE
OF SIGNAL RECEPTION BY CHANGING THE ANGLE OF
THE SOLAR PANEL CELL

Angle, o 0o 15 o 30 o 45 o
U, V 0,34 0,335 0,33 0,32
Angle, o 60 o 75 o 90 o
U, V 0,315 0,285 0.021

Fig. 6. Laboratory environment for measuring the signal level at

different distances

TABLE II
THE SIGNAL LEVEL AT DIFFERENT DISTANCES

Z, cm 15 30 45 60 75 90
U, V 0,34 0,13 0,12 0.10 0.07 0.05

Fig. 9. Graphic representation of the readings for the signal level at

15 cm and changing the angle of signal reception by changing the
angle of the solar panel cell

67
 Fig. 8 and Table 3 show the laboratory setting and the
obtained results at the signal level at 15 cm and changing the
angle of signal reception by changing the angle of the solar
panel cell.The measurements are from 0o to 90o, because after
90o the angle of view is zero and there is practically no way to
receive a signal.
From the results obtained in Tables 2 and 3 and the
graphics in Figs. 7 and 9, the change in the angle of signal
reception does not have such a strong effect as the change in
the distance. Because the drop in the signal level from 0o to
75o is 0.055 volts. While when changing the distance even
from 15 to 30 cm, the drop in the signal level is 0.21 volts.
After the experiments done so far, it can be said that the
laboratory environment is suitable for transmitting audio
signals by optoelectronic methods over relatively short
distances with a wide angle of signal reception.
IV. CONCLUSION
As can be seen from experimental and theoretical data, the
Li-Fi laboratory setting can be performed even at home. On
the financial side, it is a low-budget project, but with the
advantages it offers, it is interesting to consider.The data
transmission experiment by using Li-Fi offers 100 Gbps
maximum speed and it can be up to 240 Gbps theoretically.
In this experiment, we transmit audio signal over Li-Fi to
present the possibilities of this technology to the students at
the Technical University - Sofia.
This laboratory environment could be set for students to
perform in laboratory sessions as part of the practical training
in the discipline "Optoelectronics and Optical
Communications", taught to bachelors at the Faculty of
Telecommunications at the Technical University of Sofia. In
addition, it could be a good basis for upgrading in a bachelor's
thesis by considering various electrical elements, achieved
results, changing characteristics etc.
ACKNOWLEDGEMENTS
We would like to express our sincere gratitude to the
Faculty of Telecommunications at the Technical University of
Sofia for providing the laboratory base.
The scientific research and the results obtained are part of a
contract №202ПД0014-07 for a research project to help
doctoral students: “Using optical methods to collect and
transmit audio signals” of the Technical University of Sofia,
Research Sector.
68
REFERENCES
[1] A. Assabir, J. Elmhamdi, A. Hammouch, and A. Akherraz,
"Application of Li-Fi Technology in the Transmission of the
Sound at the Base of the PWM", 2016 International Conference
on Electrical and Information Technologies (ICEIT), Tangiers,
2016, pp. 260-265, doi: 10.1109/EITech.2016.7519602
[2] C. Gavrincea, J. Baranda, and P. Henarejos, "Rapid Prototyping
of Standard-Compliant Visible Light Communicatios", IEEE
Journal & Magazines, vol. 52, pp. 80-85, July 2014.
[3] IEEE Rev. 802.15.7-2011, "Standard for Local and
Metropolitan Area Networks, Part 15.7: Short-Range Wireless
Optical Communication Using Visible Light", Sept. 2011.
[4] J. B. Carruthers, "Wireless Infrared Communications", Wiley
Encyclopedia of Telecommunications, John Wiley & Sons, Inc.,
2003.
[5] K. Asatani, T. Kimura, "Analyses of LED Nonlinear
Distortions", IEEE Journal of Solid-State Circuits, vol. 13, no.
1, pp. 125-133, February 1978.
[6] N. Theofanous, A. Arapoyianni, "Light-to-Input Nonlinearities
in an R-LED Series Network", IEEE Journal of Quantum
Electronics, vol. 28, no. 1, pp. 34-38, January 1992.
[7] K. Asatani, T. Kimura, "Linearization of LED Nonlinearity by
Predistortions", IEEE Journal of Solid State Circuits, vol. 13,
no. 1, pp. 133-138, February 1978.
[8] P. Petrov, V. Georgieva, "Adaptive Velocity Control for a
Differential Drive Mobile Robot", 2018 20th International
Symposium on Electrical Apparatus and Technologies (SIELA),
Bourgas, 2018, pp. 1-4, doi: 10.1109/SIELA.2018.8447091.
[9] H. Elgala, R. Mesleh, and H. Haas, "Non-Linearity Effects and
Predistortion in Optical OFDM Wireless Transmission Using
LEDs", Intern. Journal of Ultra Wideband Communications and
Systems, vol. 1, pp. 143-150, October 2009.
[10] H. Qian, S. J. Yao, S. Z. Cai, and T. Zhou, "Adaptive Post-
Distortion for Nonlinear LEDs in Visible Light
Communications", IEEE Photonics Journal, vol. 6, pp. 1-8,
2014.
[11] S. H. Ali, S. P, G. R. Embedded, and M. J, "Design and
Evaluation of LiFi Module for Audio Applications", 2018 15th
IEEE India Council International Conference (INDICON),
Coimbatore, India, 2018, pp. 1-5, doi:
10.1109/INDICON45594.2018.8987134
[12] B. Liengme, A Guide to Microsoft Excel 2007 for Scientists and
Engineers, Academic Press, 2008.
[13] G. Popov, K. Raynova, "Diversity in Nature and Technology -
Tool for Increase the Reliability of Systems", 2017 15th
International Conference on Electrical Machines, Drives and
Power Systems, ELMA 2017, pp. 464-466.
[14] G. Popov, "Use of System Diversity as Tool for Identification,
International Conference on High Technology for Sustainable
Development", HiTech 2018, Article number 8566676.
[15] Microsoft Excel 2007 for Scientists and Engineers, Academic
Press, 2008.
[16] L. Johnson, Fiber Optic Active Devices - LEDs, Lasers, and
Detectors, Course, IEEE, 2013.
[17] S. Kulkarni, A. Darekar, and P. Joshi, "A Survey on Li-Fi
Technology", 2016 International Conference on Wireless
Communications, Signal Processing and Networking
(WiSPNET), Chennai, 2016, pp. 1624-1625, doi:
10.1109/WiSPNET.2016.7566413
 IMA-NET: Innovative e-Management Platform for
Academic Network
Galia I. Marinova1 and Krisel L. Tola2
Abstract – In academic networks are priority in higher II. DESIGN AND SOFTWARE REALIZATION OF
education international cooperation. Successful management of
teachers’ and students’ mobilities is one of the main tasks in
IMA-NET PLATFORM
academic networks. Based on the positive experience with the
coordination of CEEPUS Network - CIII-BG-1103-05-2021 A. Design of IMA-NET Platform
Modelling, Simulation and Computer-aided Design in
Engineering and Management, a platform for Innovative e- The design of the IMA-NET platform is presented in Fig. 1.
Management of Academic Network IMA-NET is developed. The Modules developed in IMA-NET are: Database collecting
paper describes in details the software realization of IMA-NET and storing the know-how of teachers in the network (courses,
platform, its modules and its impact on network management. tools, modules); Survey tool to evaluate students’ interest to
different courses offered; Tool to generate context-oriented
Keywords – E-Management, Academic network, Joint
programs, Students’ mobility, CEEPUS, Web based application. Flexible courses with 3 ECTS, composed by 3 modules from
the Database; Certificate generator and assessment tool; Tool
to assist Joint Doctoral program “Thèse en co-tutelle”; Tool
I. INTRODUCTION for Events with Feedback option; Tool for dissemination of
joint achievements and new knowledge created in the
University networks are a trend in higher education network.
development nowadays. CEEPUS program is a good example
of long term academic network implementation. Schuch K.,
based on Central CEEPUS office information, shows in [1]
that 80 CEEPUS academic networks are granted for the
academic year 2019/2020. The CEEPUS Network CIII-BG-
1103-04-1920 Modelling, Simulation and Computer-aided
Design in Engineering and Management [2] is in its 5th year,
it involves 33 partner HEIs from 15 countries from CEEPUS
region [3] and it is considered as a large scale network.
Management of mobilities in a CEEPUS network is a
challenge. The goal is to organize successful CEEPUS events,
short and long term students’ and teachers’ mobilities, using
the grants awarded to the network, in an optimal way. To
achieve this goal, CEEPUS events, individual study and
teaching programs have to be organized early enough to
permit students and teachers from partner institutions to apply
for CEEPUS grants in the deadlines defined by CEEPUS
program and by the National CEEPUS offices in CEEPUS
Fig. 1. Design of IMA-NET platform
partner countries. To face the challenges an Innovative e-
management platform IMA-NET is designed and developed.
The outline of the paper is: In Section II the design of IMA- B. Framework and Approach Implemented for the Software
NET platform and the software realization are considered, in Realization of IMA-NET Platform
Section III the impact of IMA-NET on mobilities is estimated
and Section IV concludes the paper and describes probation The platform IMA-NET supports three different user
planned for the platform. interfaces, respectively:
• Administrator – Coordinator of the CEEPUS
network and Platform support staff;
• Lecturer – Teachers in the CEEPUS network;
• Student – Students in the CEEPUS network.
The administrator is redirected to a control panel interface
1,2
where he/she can view different statistics and trends. The
Galia I. Marinova and Krisel L. Tola are with Technical administrator can create, edit and delete courses, modules and
University-Sofia, Faculty of Telecommunications, Boulevard
topics. Also, when a lecturer or a student signs up, they are
“Kliment Ohridski” 8, 1000 Sofia, Bulgaria, E-mail: gim@tu-
sofia.bg
stored in a pending list and the system administrator is

978-1-7281-7142-5/20/$31.00 ©2020 European Union 69

responsible for accepting or rejecting them. The administrator
has access to all the surveys, including data about the person
who filled them and the topics he selected.
The lecturer can manage topics and modules. The data
entered in the Topics section is used in the survey section.
Each topic belongs to a category. At the end, the survey
shown to the users will be a categorized list of topics fetched
from the lecturers’/administrator’s input.
The lecturer can insert, edit and delete modules. A module
is part of the specific Joint program called Flexible course.
Each course consists of three modules and each module
should have three lessons inside and three practical examples
conform to CEEPUS teacher’s mobility format and
requirements [3]. The web-app offers the possibility to enter a
title for each module and a descriptive text making use of
external text editing libraries.
Students logged in the system are provided a list of flexible
courses they can register themselves in. They are able to view
data around the flexible courses such as modules they need to
follow, place and timesheets.
The information about managers, lecturers, students,
surveys, courses, topics and their connections is saved in an
independent database.
Selecting the right tools to use for developing the web
application for IMA-NET, required a serious study of the
available options. It was opted in for a fully compatible set of
open source programming languages which interact with each
other for the end goal. “Open source” refers to the notion that
the encoded instructions for the programs produced are openly
available for scrutiny by potential users. One of the goals of
open source software development is to enhance the quality of
the software by increasing scrutiny for errors by more
programmers [5].
A combination between HTML, CSS, Javascript,
Codeigniter (PHP framework) and MYSQL [5-7], [9] was
used to accomplish the platform development.
Fig. 2. IMA-NET table structure and relations
70
Codeigniter [8] source framework to build dynamic PHP
applications and the MVC (model-view-controller)
architecture were used for the software design of the platform
IMA-NET.
C. MySQL and Database Structure in IMA-NET
The Database in IMA-NET relies on a relational database
management system (RDBMS) called MySQL. A RDBMS is
a database management system which depends on the
relational model. Database management systems are the
electronic filing containers that help to manage the
information [9]. On this kind of systems, data is organized
into tables compounded by specific fields. Each field contains
the type of data being saved. Some of the main advantages of
relational DBMSs are table’s interconnectivity, structured
queries, scalability, flexibility and security.
Fig. 2 is an illustration of the tables, part of the web
application for IMA-NET, their fields and their relations. All
credentials of the users who can access the system via the
login interface are saved in the table called users. Digging a
bit further we can understand that the primary key of users
table is used as a foreign key in two other tables such as
recovery_req and student_cours. The usage of primary keys
and foreign keys is part of normalization process and this is
essential since it would avoid data repetition in more than one
table. All the tables illustrated in the above schema have gone
through this process.
To connect to this database system, Codeigniter offers a
configuration file called database.php located in
application/config/%workingenvironment% where working
environment can be one of the following folders:
1) Developing
2) Testing
3) Production
 Once opened the file in a text editor all that the user needs
to do is to insert the database credentials inside the
$db['default'] array. All the connection phase is than handled
by Codeigniter. Following the framework rules, all database
actions and business logic should reside inside model folder.
The files inside this folder must extend the model class
inheriting so all properties and methods of that class. Inside
the class we can create than methods which would call the $db
class to execute queries via SQL language according our
needs.
Since there are 3 different user types they are offered a
login interface which shows up when the user would send a
request to the domain. The platform is hosted on the web
address: ima-net.ubt-uni.net. The login page is illustrated in
Fig. 3.
Fig. 3. IMA-NET login page
At this section users can also create an account or recover
their password in case they lost it. The logic behind account
recovery is that the user must insert their e-mail address used
to register initially on the website. Once entered the e-mail
they will receive an activation email giving them the
possibility to set a new password. At the very bottom of the
page you can find a list of partners who made possible the
development of the web application. Each logo contains a link
which will redirect to the corresponding website if clicked.
As stated in the image above, users will have to fill some
formal information such as their name, email, country etc. The
survey continues with a list of topics suggested by the
lecturers registered on the system. Each category has a certain
number of topics and this data is displayed in different tables.
After inserting the correct credentials as an administrator
the user is redirected to the control panel. The panel interface
is the same for all user roles. It differs only from the access
rights each user level has and what he can or can not do within
the system. The first page the user will see as an administrator
after logging in, is the dashboard. This page contains some
useful statistics regarding number of students registered on the
system or selected topics trend gathered from the survey
results. The lecturers tab contains a list of all registered
lecturers appears to the administrator. To display the list a
paginated table is used. This is handy since there may be a lot
of data. Pagination gives the possibility to split these ones into
71
sections of 10. The administrator can interact with this section
by choosing between checking out lecturer details or deleting
their account. The lecturer details page contains info such as
lecturers name, the University he or she comes from and a list
of proposed topics. The topics that a lecturer proposes are than
shown in the surveys topics list.
The whole control panel follows this design principle.
Unlike the most of pages in IMA-NET, the survey section
works as a separate entity of the website. No login to the
system is necessary to have access and fill the survey since it
is supposed that this survey will be filled by persons who are
not necessarily staff or students.
After the users fill the formal information, the survey
continues with a list of topics suggested by the lecturers
registered on the system. Each category has a certain number
of topics and this data is displayed in different tables. To
access the survey a separate link is to follow: http://ima-
net.ubt-uni.net/survey. The web page of the survey is
presented in Fig. 4.
Fig. 4. Web page of the Survey in IMA-NET
III. IMPACT OF IMA-NET PLATFORM ON
MOBILITY MANAGEMENT
IMA-NET platform permits to organize the know-how in
the academic network, to automate and optimize the Flexible
course content and Certificate generation [4] for concrete
CEEPUS events and Joint doctoral program courses, to inform
students and teachers on time about the events and Joint
program courses and to permit them to apply for CEEPUS
grants in the deadlines fixed by CEEPUS program. The
platform IMA-NET performs statistics to analyse the students’
and teachers’ interest and feedback. The dashboard of IMA-
NET is presented in Fig. 5.
 Fig. 5. Dashboard of IMA-NET
As visible from the figure, the topic of Project Management
is currently the one that has attracted the highest interest from
students.
The first experimental implementations showed that IMA-
NET is an excellent instrument to assist the coordination of
the network
IV. CONCLUSION
The web application for the IMA-NET platform presented
in this article will allow CEEPUS members to have an
organized and centralized e-management system for their
summer school sessions. The set of open source programming
languages used to develop this software gives the possibility
to maintain and manipulate gathered data easily significantly
reducing the error margin. Being an online website, one just
needs an internet connection to access it thus eliminating
boundary limits. IMA-NET platform will be probated to
organize Flexible course within CEEPUS, at the Summer
Academia, Summer Schools and International Weeks
incorporated in the annual program of the network. The
effectiveness of the platform will be tested through the
Feedback tool and the Feedback forms generated by IMA-
NET. They will be filled by teachers and students attending
the courses and events and later they will be analyzed.
Management decisions for network will be formulated based
on the feedback results.
ACKNOWLEDGEMENT
The results presented in the paper are partly supported by
the CEEPUS project CIII-BG-1103-03-1920: Modelling,
Simulation and Computer-aided Design in Engineering and
Management.
72
REFERENCES
[1] K. Schuch, "The Uptake of European Programmes in the
CEEPUS Cooperation Area", Center for Social Innovation, pp.
1-24, June 2019
[2] Site of the CEEPUS project CIII-BG-1103-03-1920 (2020):
Modelling, Simulation and Computer-Aided Design in
Engineering and Management, Retrieved February 23, 2020
from: http://ceepusmodcad.ubt-uni.net/
[3] Site of CEEPUS program, Retrieved February 23, 2020 from:
https://www.ceepus.info
[4] G. Marinova, "Impact of University Networking on Students’
Mobility and Motivation", EUNIS 2019, Trondheim, Norway,
5-7 June 2019, https://www.eunis.org/eunis2019/wp-
content/uploads/sites/15/2019/06/S11-3.pdf
[5] D. Delorey, C. Knutson, and C. Giraud-Carrier, "Programming
Language Trends in Open Source Development: An Evaluation
Using Data from all Production Phase Source Forge projects",
Second International Workshop on Public Data about Software,
pp.1-6, June 2007.
[6] G. B. Newby, J. Greenberg, and P. Jones, "Open Source
Software Development and Lotka's Law: Bibliometric Patterns
in Programming", Journal of the American Society for
Information Science and Technology, vol. 54, no. 2, pp. 169-
178, 2003.
[7] W. Cui, L. Huang, L. Liang, and J. Li, "The Research of PHP
Development Framework Based on MVC Pattern", IEEE 2009
Fourth International Conference on Computer Sciences and
Convergence Information Technology, Seoul, Korea, 2009.
[8] A. K. Himawan, "Performance Analysis Framework
Codeigniter and CakePHP in Website Creation", International
Journal of Computer Applications, vol. 94, no. 20, 2014.
[9] R. Nixon, Learning PHP, MySQL, JavaScript, and CSS,
O’REILLY, 2009.
 Expert Tool for Filter Design Program Selection in
Online-CADCOM Platform
Ognyan R. Chikov1 and Galia I. Marinova2
Abstract – The paper describes the software implementation of benchmark example are discussed in Section IV. Conclusions
Expert Tool (ET) for automatic selection of filter design program and implementation in the educational process are drawn in
in the Cloud-based platform Online-CADCOM. It illustrates the Section V.
methodology of development, integration and delivery of
Multicriteria Decision Analysis method used for user-friendly
automatic selection for designing filters through ET. Decision II. DESIGN MODIFICATION IN E-CONTENT
Matrix and MAUT model usage as well as front-end and
backend implementation and integration of the ET are
DEVELOPMENT FOR TOOL SELECTION
described. The platform and the ET are used for education in the
courses of Computer-aided design in Technical University-Sofia, A. Modification Needed in E-content Development to
and for individual students’ projects. Implement the MAUT Additive Aggregation Model Method

Keywords – Expert system, Filter design, Multicriteria decision
analysis, Decision matrix, MAUT model.
I. INTRODUCTION
Online and cloud-based EDA tools are a trend during the
last decade [1]. The Online-CADCOM platform [2], contains
large number of online tools for computer-aided design
(CAD) which are integrated, based on methodology steps. The
tools are first estimated for reliability by tests of their
qualitative and quantitative features and then classified based
on Daniel-Gajsky Y model [3] for design levels abstraction.
Each tool has its own unified Tool Passport template,
containing its features characterization [2,4]. Conform to
Daniel-Gajsky Y design abstraction model, the platform is
structured on categories and subcategories containing tools,
based on their classification and characterization. Analog
filter design is the first automated in analog area, so different
semiconductor designers and/or manufacturers provide their
own online tools [5], often oriented to the devices they
produce. Seven tools for analog filter design are integrated in
Online-CADCOM [6]. The Multicriteria Decision Analysis is
used for automated tool selection for filter design. In the
current paper the approach is implemented in software.
Considerable modifications of the E-content development
process and the methodology for passport tool publication are
introduced in Online-CADCOM in order to integrate the
Expert Tool (ET) in the platform. A benchmark example for
automatic selection of filter design tool from [6] is solved
using the ET.
The paper is organized as follows: The modification in e-
content development is presented in Section II. Online-
CADCOM expert tool database and algorithm implementation
are presented in Section III. The ET implementation and
1,2Ognyan R. Chikov and Galia I. Marinova are with Technical
University-Sofia, Faculty of Telecommunications, Boulevard
“Kliment Ohridski” 8, 1000 Sofia, Bulgaria, E-mail:
gim@tu-sofia.bg
The development and integration of ET consists in solving
2 tasks: extending the tool passport format through adding a
set of mandatory and desired criteria with different weight
coefficients (full list for filters is published in [6]) and
implementation of the Multicriteria design analysis formulas
for performing tool selection. The modifications to develop in
Online-CADCOM for adding the ET are presented in Fig. 1.
Fig. 1. Modifications in Online CADCOM for adding ET
The E-content development for filter design program
selection is oriented to the implementation of MAUT additive
aggregation model method [7]. Additive aggregation model is
based on weighted means and the algorithm works in two
steps based on mandatory and desired criteria from the
specification of the filter to be designed.
The value function F(OPTIONk)is expressed as:
( )=∑ ( ), (1)
where wj is the weight of attribute j and uj(xkj) is the utility
calculated from the value (x) of criterion j in option k.
Two steps follow - Mandatory criteria are defined and for a
given specification of the filter the set of viable options (VO)
is determined. VO is defined as:
=⋂ ∏ ( ), (2)
where Xi, i=1..n - mandatory criteria, σ(Xi=True) - estimate
of selection.

978-1-7281-7142-5/20/$31.00 ©2020 European Union 73

 Mandatory criteria options weights and desired criteria
options weights may vary and may be presented or not in a
different filter design tool. That’s why they are separated in
different steps and implemented separately in the E-content
development. The Passport tool process is updated and for its
creation, a multi-step Wizard component is created, designed
and used. The Wizard component is composed of 5 steps
separated by implementation, but tightly connected by each
other by logic, sequence and requirement. This is needed
because there are different criteria options with different
weight to be implemented.
B. General Workflow for Implementing MAUT Additive
Aggregation Model Method
The ET creation process is built of 5 steps corresponding to
the step workflow bar containing the Options used in MAUT
equation for utility function as follows:
• Tool Basic Information – in the tool Passport;
• Mandatory Criteria Options with weight coefficient (1);
• Desired Criteria Options with weight coefficient (0.5);
• Desired Criteria Options with weight coefficient (0.33);
• Success Screen – This screen is a notification screen that
notifies the user for the successful inclusion and successful
publication in the ET.
The steps are implemented as independent and no one step
can be skipped. The navigation is by the “Next” and “Back”
buttons.
In the first step, the user is filling the well-known attributes
of tool passport from [2]. The Passport and the ET are now
tightly coupled because entering the tool information
automatically defines the information into the ET.
In the third step, the Desired Criteria Options with weight
coefficient (0.5) is shown. Here the options are represented
with a check button. The user can select the desired ones. In
the fourth step, the Desired Criteria Options with weight
coefficient (0.33) are selected. This is the criteria option with
lower coefficient. Mandatory and desired option selection are
shown in Fig. 2. In the last fifth step, the user is notified that
process is completed. From here, the user can be navigated to
dashboard.
Fig. 2. Mandatory and Desired Criteria Options with weight
coefficient 0.33
74
C. Software Implementation of the ET
The software implementation of every step is composed by
so called “component” containing four separated files as
follows:
• Template file which contains the structure of the screen
(HTML Elements). Each component of the screen is built
dynamically, based on the data coming from the database.
There are no static elements, which defines the large
scalability of the platform and easy maintenance. Each
criterion with its options is built dynamically based on
properties described and fetch from the Model in MVC
framework.
• CSS File which contains the cascade sheet styles of the
screen. The major platform styles are exported to a single file
View “styles.css” in the MVC framework, however each step
from the Wizard needs specific tweaks and options to be
implemented. The color of the table header is set as well as
media query tag specified for certain screen resolution.
• Component file – Model Implementation of the template. It
defines which component to dowhat on the screen, and the
user actions provided. Component file is the backbone file of
certain component and in our implementation, a certain step.
The fragment on bellow shows how each component import
specific files that uses and creates its own class that can be
implemented by other classes.
import { Component, OnInit } from '@angular/core';
import { FormGroup, FormBuilder, FormArray, FormControl
} from '@angular/forms';
Component({
selector: 'app-add-tool',
templateUrl: './add-tool.component.html',
styleUrls: ['./add-tool.component.scss']
})
export class AddToolComponent implements OnInit {
• Test File is for automation tests scenarios. Test files are
used for automatic test of implementation of the component.
The connection between the components and the data flow
described above is done by sharing data from parent to child
components. The sharing of the data allows keeping the state
of the data between the steps. The goal of next Section is to
show how the data is saved in the database and how the
Criteria Options are organized in the context of the database
in Online-CADCOM.
III. ONLINE-CADCOM EXPERT TOOL DATABASE
AND ALGORITHM IMPLEMENTATION
A. Database Implementation
The platform database is extended and should be redesigned
to accumulate all the necessary Passport Tool extended
information used in the ET. The whole Passport table is not
updated or changed. A new database tables are added as
described:
• Database Table for Mandatory Criteria Options – This
table contains three rows as showed in Fig. 3.

Fig. 3. Database table for Mandatory Criteria Options
The “frequency_range” row contains the possible options as
described in [6] for “Frequency range / Filter elements”
mandatory criteria. The “Type_of_Filter” row contains the
possible options as described in [6] for “Type of Filter”
mandatory criteria. The “id” row contains the id associated
with the id of the Passport Tool form the “tools” table. This is
the relation between a single tool and mandatory criteria
options with weight coefficient (1).
• Database table for Desired Criteria Options with weight
coefficient (0.5) – This table contains fourteen rows as
showed in Fig. 4.
Fig. 4. Database table for Desired Criteria Options with weight
coefficient (0.5)
The id row is a relation between those criteria options and
the Passport Tool itself.
• Database table for Desired Criteria Options with weight
coefficient (0.33) – This table contains seven rows.
B. Algorithm Implementation
The algorithm of ET is implemented in the platform
backend and works on two stages:
• Receive the query from front-end with the desired
parameter specified by user. All these parameters are
described in the next section.
• Execute the query on the database with these parameters.
• Return the calculated result to the front-end for showing
the best solution for ET program selection.
Before executing any operations on the database, the
backend should be connected to it.
Receiving the query from the front end is made by calling a
separate function called: “performExpertQuerry(params)”
which receives all the parameters from the front-end as
JavaScript Object Notation (JSON) object. This object is in
the following format and contains key-value pairs:
“{ Key: Value }”
In our case the request object is slightly different and
complex. Fragment of it is showed below, as some of the
content is hidden because of the long code list.
{ "mandatory": {
"frequency_range": "1kHz-100kHz" },
75
"optional_coefficient_05": { ... },
"optional_coefficient_033": { ... }
The json object is helping to describe every feature request
from the user and prepopulated in the front-end part of the
platform, where the ET is launched and the query is executed.
The execution of the query to the database is made with
“database queries” containing all the “rules” populated in the
json file. Example query is shown below:
SELECT * FROM mandatory_criteria_options WHERE
rules.mandatory .frequency_range=”1kHz-100kHz”
Then execution is made for all other Desired Criteria
Options. When the results are retrieved, then the calculation of
the coefficient is performed. The code fragment bellow
represents the selection formula. Here the overall weight
coefficients are calculated based on loop cycle and addition of
Mandatory criteria weight. Then, the same loop cycle is made
for Desired criteria weight with coefficients 0.5 and 0.33:
queryResults.forEach(element => {
letoveralCoeff = 0;
switch (element.criteria) {
case mandatory:
overalCoeff += element.weightCoeff;
break;
case optional_05:
overalCoeff += element.weightCoeff;
break;
case optional_033:
overalCoeff += element.weightCoeff;
break;
default:
break;
});
For every retrieved tool that matches a desired criteria
option, a coefficient number is added to the total coefficient
based on the coefficient of the matched option. At the end of
the algorithm, the calculated coefficient for Tools founded in
the platform are sorted and the tool with the highest
coefficient is likely the one that the user will use for designing
the filters based on provided options.
The returned result from the algorithm is array of objects
with JSON format and a fragment is shown below:
[
{"id": "KRGj33nHJsekuDLGXErWuvvAoU42", "value":
1.33},
{"id": "2iyHVhIX7MNpWjMFUMyXjhGDuYA3",
"value": 0.83},
{"id": "4Macg8WlG7OKo56ojenFN2aSz4J3", "value":
0.83}
]
The “id” key is representing the id of the Tool in the
platform, which will be used later to show the Tool on the
screen and the value is the calculated coefficient based on ET.
The next Section shows how the ET is implemented in the
front-end of the platform and a benchmark example is solved.
 IV. ET IMPLEMENTATION AND REAL-LIFE
EXAMPLE
The ET is implemented on the front-end part of the
platform. It can be launched from the main menu of the
platform, which is repositioned and restyled.
The process is designed as workflow wizard steps
container, so the user can provide the valuable information
needed for the ET in small and easy to understand steps. Here
again the workflow has a toolbar with information about all
the needed steps in order to complete the process. They are
marked with number and get coloured when single step
information is provided and completed.
• Basic Category Information Step – on this step the user
can select the category of the tool. In our case, the ET is
implemented for analog filter design and the category is
selected.
The information for this step is provided and the ET
software implementation lets the user go forward to the next
step as makes the “Next” button active. Pressing it will
navigate to the second step.
• Mandatory Filter Criteria – on this step the user must
specify the mandatory filter option for filter design selection.
The checkboxes option implementation allows the user to
select more than one option. For the demonstration of the ET
a benchmark filter design specification from [6] is used. The
options are selected as shown in Fig. 5.
Fig. 5. Selection of mandatory criteria
Fig. 6. Selection of the tools by ET
Once the mandatory options for ET are selected, the process
can continue and the next step from the Wizard is loaded. The
options selected for desired criteria with weight 0.5 are: Flat
Band – Butterwort and Circuit topology – Biquadratic. Then
for Desired Criteria with weight – 0.33 is chosen: Low
76
sensitivity-Monte Carlo. This is the last step from the Wizard
and the user can navigate back to correct his options. Once the
button “Next” is clicked, the ET algorithm is launched and the
needed queries and backend mathematical solutions are
performed. During the calculation the user is seeing the screen
with dynamic loader showing that backend actions are
performed.
The process completes, the loading bar is hidden and the
ET provides to the user the best solution for the desired filter
design options, then additional tools that might also do the
job. The result for the benchmark example is shown in Fig.6.
V. CONCLUSION
The paper described the software realization of an Expert
tool for automatic selection of a program for filter design. The
ET is successfully integrated in the platform Online-
CADCOM. A test example based on benchmark was
performed in order to illustrate the implementation.
The platform Online-CADCOM with the new ET is used in
the educational process in the courses of Computer-aided
design in the Faculty of Telecommunications in Technical
University, Sofia, Bulgaria. The feedback from about 200
Bachelor student using ET is that it helps them to assimilate
better and deeper the filter design process and options.
Further development is connected with the design of other
categories of circuits and systems within the ET and Online-
CADCOM platform.
ACKNOWLEDGEMENT
This study is realized andpartly supported by the CEEPUS
network CIII-BG-1103-04-1920.
REFERENCES
[1] Rik, Top 10 online circuit simulators, Electronics-lab.com, 8,
Dec. 2016, https://www.electronics-lab.com/top-ten-online-
circuit-simulators/
[2] G. Marinova, O. Chikov, “Methodology for Tools Integration in
the Online Assisted Platform for Computer-Aided Design in
Communications”, ICEST 2015, pp. 31-35, 2015.
[3] D. Daniel, D. Gajski, S. Abdi, A. Gerstlauer, and G. Schirner,
Embedded System Design: Modeling, Synthesis and
Verification, Springer, 2009.
[4] G. Marinova, O. Chikov, “E-content Development and Task
Solution Using the Content Management System of Online-
CADCOM”, ICEST 2016, Ohrid, Macedonia, June 28-30, 2016,
pp. 213-216.
[5] B. Rodic, G. Marinova, and O. Chikov, “Algorithms and
Decision Making Methods for Filter Design Tool Selection for a
Given Specification in Online-CADCOM Platform”, ERK 2017,
Slovenian IEEE Section, Portoroz, Slovenia, 25-26 September
2017, pp. 247-251.
[6] G. Marinova, O. Chikov, and B. Rodic, “E-Content and Tool
Selection in the Cloud-based Online-CADCOM Platform for
Computer-Aided Design in Communications”, CONTEL 2019,
Graz, Austria, 3-5 July 2019, pp. 1-4.
[7] J. Butler, D. J. Morrice, and P. W. Mullarkey, “A Multiple
Attribute Utility Theory Approach to Ranking and Selection”, J.
Management Science,vol. 47, no. 6, pp. 800-816, 2001.
Analysis of Pulse Code Modulation in MATLAB / Octave
Environment
Lyubomir Laskov1, Veska Georgieva2 and Kalin Dimitrov3
Abstract – In this paper, we present an approach for
investigating the pulse code modulation for signals with various
shape forms. We have developed an algorithm as well as its
course of implementation in MATLAB and Octave
environments. The software we have implemented makes it
possible to investigate the Pulse Code Modulation of various
analog signals along with the signal parameter influence and
parameters of PCM on the discrete and digital signals. It is
possible to use the presented approach in education, in the field
of engineering, to study this process.
Keywords – Signal processing, Pulse code modulation, Analog
to digital converter, PCM, Computer simulation, MATLAB,
Octave.
I. INTRODUCTION
Pulse code modulation is very common in any digital
device. It is necessary for analog signals, which exist in
nature, to be converted into digital signals, which could be
further processed by the digital devices. After being subjected
to PCM, a continuous signal which has an unlimited number
of possible values, becomes discrete in time and could take a
finite number of values.
The pulse code modulation is studied at the Technical
University of Sofia, at the undergraduate level, within the
"Signals and Systems" seminar, in the following specialties:
"Electronics", "Telecommunications", and "Computer and
Software Engineering". During the seminars, it will be
possible for the students to investigate the transformation
process of analog signal to discrete signal and to digital signal.
Students will explore the influence of various parameters of
PCM and analog signals on the resulted discrete and digital
signals.
We suggest for a program code to be applied in the
environment of MATLAB in order to investigate as many
parameters as possible of the processes performed during
PCM, as well as for signals to be generated, with different
shape forms of a maximum number, and to change the
parameters of the signal as much as possible [1]. It is
preferable to write source code in MATLAB rather than the
1Lyubomir Laskov is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: laskov@mail.com
2Veska Georgieva is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: vesg@tu-sofia.bg
3Kalin Dimitrov is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: kld@tu-sofia.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
MATLAB Simulink because of the supplementary features
giving students the possibility to change signals and the
parameters of the individual processes performed during
PCM, which is not applicable in MATLAB Simulink [2], [3].
The source code writing also results in the expectation that the
students are familiar with the processes taking place during
PCM.
The environment of MATLAB bears a decisive advantage
over other programming languages and software products, due
to its built-in functions [3], [4]. Thus, students would be
assisted considerably in their work. Another reason to choose
MATLAB is that the code written for it, could be started with
small modification on GNU Octave. GNU Octave is an open
source product, supporting Windows, Linux, BSD and Mac
OS. It is very suitable for home usage by students, as it
doesn’t cost anything and could be used on almost any device.
This allows students to use MATLAB not only in the
university, but also to use the same code in Octave at home.
Taking into account the current situation with the Covid 19
pandemic, this is a significant advantage over other paid
software platforms.
II. PROBLEM FORMULATION
The major tasks answered by the study of PCM are
connected with the different kinds of discretization and
quantization, as well as the variety of signal shapes
(sinusoidal, rectangular, triangular, sawtooth, Gaussian, etc.).
The process of Pulse Code Modulation has three basic
steps: discretization, quantization and digitalization.
Discretization is a process of transforming the analog
(continuous) signal to discrete ones. In this step values from
original signals are taken at a specific moment in time and
saved as values of the corresponding samples. The interval
between samples, also known as sample time, is one of the
important parameters of discretization. Its value must meet
Nyquist criteria, which is given in expression 1 [5]:
1
Δt ≤ , (1)
2. f max
where the maximum frequency in the analog signal spectrum
is fmax.
The sampled amplitude values are represented by
quantization as a limited set of levels, which consists in the
conversion of a sample of continuous amplitude into a signal
of discrete time. Quantization could be linear (uniform) or
nonlinear (non-uniform). In uniform quantization, all steps are
equal to each other. In the non-uniform one - the steps are
with different sizes. Another important parameter is the
77
number of levels. In the uniform quantization, step “h” could Start
be found from the number of allowed levels “N” with
expression [5]
S max Setting up
h= , (2) Signal form
Global
N Selection
Parameters
where Smax is the maximum possible value in the analog
signal. The number of allowed levels themselves depends on
the number of bits “n”, which will be used to represent the Setting up Analog
Specific Signal Signals
value of samples in digital sequences. The dependence is the
Parameters Generation
following [6]:
N = 2n . (3)
Setting up
The level of quantized signal differs from the level of
Discretization
sampled signal, as a result of which the noise is added to the Parameters
signal. This noise is known as error in quantization. The
maximum possible value of the error is equal to half of the
biggest step in quantization.
Discretization
The third step is coding, in which the quantized signal
levels are presented in binary form. Most often, symmetric
binary code is used, in which the first bit indicates the polarity
of the signal and the others - the signal level. Setting up
The major tasks that we can solve by computer simulations, Quantization
Parameters
are as follows:
1. Establish a suitable model for signal presentation.
2. Derive algorithms for signal discretization.
3. Derive algorithms for signal quantization. Determination of
Uniform
4. Derive algorithms for signal digitalization. Coefficients of
No Quantization
5. Study waveform and analog signal parameter influence Levels
on the discrete and digital signals.
6. Investigate the discretization and quantization Yes
parameter influence on discrete and digital signals.
7. Present the obtained results graphically. Determination of Determination of
Levels of Levels of
Quantization Quantization
III. BASIC ALGORITHM FOR PULSE CODE
MODULATION
Determination Of
In Fig. 1, we can see the block diagram of the algorithm Quantization
proposed for analyzing the process of pulse code modulation. Signal
The variables and global parameters are fixed at the start.
Following that, choosing the signal shape form is needed, as
well as the setting of specific parameters for the selected type Digitalization
of signals. Signals with different shape forms have different of the Signal
specific parameters. In addition to the signal shape form, it is
necessary to know whether it is periodic or not. Depending on
this, different models of signals will be created. Taking into Graphic Presentation of
the Analog, Discrete and
account all the signal parameters, the students are required to Digital Signals,
devise a mathematical description of the signals studied (the
investigation of the following analog, periodic and aperiodic
signals are to be focused on: rectangular, sinusoidal, End
triangular, Gaussian, and sawtooth).
Once the signal models are created, the parameters of the Fig. 1. Basic algorithm block diagram
discretization must be set. The parameter that will be changed
in order to investigate its influence on the discrete and digital will be used: uniform quantization and non-uniform
signals will be the discretization step. The next step after quantization. The students need to understand the advantages
discretization is to set the parameters of the quantization. In and disadvantages of both types of quantization. Then it is
this step the following parameters will be set: the number of necessary to determine the levels of the quantization, which in
bits per sample, maximum level, minimum level, as well as case of non-uniform quantization require an additional step –
the type of quantization. The following types of quantization the determination of the coefficients of different levels. Тhe

78
next step is to quantize the signal. Then the quantized discrete Fig. 3. Comparing Fig. 2 and Fig. 3, students may notice how
signal needs to be presented as a digital sequence, which is the much smaller the quantization error becomes, when the
result of the pulse code modulation. In the final step of the number of bits presenting each sample increases. The
algorithm it is necessary to show the graphic representation of maximum value of quantization error in Fig. 2 (8 bits per
analog signal, discrete signal, quantized signal, quantization sample) is 2.10-2, while in Fig. 3 (16 bits per sample) it is
error and digital signal. This allows for the study and 5.10-5.
visualization of specific signal parameter and PCM parameter
impact on the digital signal.
The implemented software for investigation of the process
of the periodic and aperiodic signal pulse code modulation,
allows us to devise a graphic representation of the analog
signals, their corresponding discrete, quantized and digital
signals, as well as the quantization error. The implemented
software can run on both Matlab and GNU Octave. The only
difference is that in Octave it is necessary to load a side
library “symbolic”, which is required for the mathematical
description of the analog signal.

IV. EXPERIMENTAL PART

The problems considered are solved by a MATLAB /

Octave environment computer simulation.
Several results from PCM simulations of periodic signals
are shown in the next figures below. In Fig. 2 is shown the
result from the simulation of PCM for sine wave signal in
Matlab environment. During the simulation, it was assumed
that the uniform quantization will be used and each sample
will be represented with 8 bits. The figure shows the analog
signal, the discrete signal, the quantized signal, the digital
signal and the quantization error. Fig. 3. Analog, discrete, quantized and digital signals and error from
The result from the simulation of sine wave signal with quantization for sine wave signal when 16 bits per sample are used
uniform quantization with 16 bits per sample is shown in

Fig. 2. Analog, discrete, quantized and digital signals and error from Fig. 4. Analog, discrete, quantized and digital signals and error from
quantization for sine wave signal when 8 bits per sample are used quantization for uniform quantization of sawtooth pulse train

79
 Fig. 4 shows the result of the simulation of the PCM for the
sawtooth pulse train, when uniform quantization is used and
each sample is represented with 8 bits. The simulation was
done in GNU Octave environment. Fig. 5 shows the analog
signal, the discrete signal, the quantized signal, the digital
signal and the quantization error, for sawtooth pulse train,
with non-uniform quantization with 8 bits per sample. By
comparing Figs. 4 and 5, students can analyze the influence of
the type of quantization on the quantized signal and the
distribution of the quantization error. It can be seen from the
figures that for uniform quantization, the error from
quantization does not depend on the signal level, while for
non-uniform quantization, the error from quantization
increases when the signal level becomes higher.
Fig. 5. Analog, discrete, quantized and digital signals and error from
quantization for non-uniform quantization of sawtooth pulse train
Fig. 6 shows the results of a PCM simulation for a Gaussian
pulse train. During the simulation, it was assumed that
uniform quantization will be used and each sample will be
represented with 8 bits. The simulation was done in
MATLAB Environment. Comparing Fig 4 and Fig. 6, the
students can see differences between natural binary code and
symmetric binary code. The second one is used in Fig. 6, and
as it is seen from it, the first bit is used to show the sign of
signal value. From Fig. 4 it is seen that the first bit changes
depending on signal value, while in Fig 6 it is always the
same.
80
Fig. 6. Analog, discrete, quantized and digital signals and error from
quantization for uniform quantization of Gaussian pulse train
V. CONCLUSION
In this study, we propose an approach for investigating
periodic and aperiodic signal pulse code modulation. We have
developed an algorithm and its course of implementation in
MATLAB and Octave environments. The students can devise
signal models of various types, shape forms and different
parameters. They can also create models of the individual
processes that make up the PCM. Students will be able to
examine the impact of each of the PCM parameters on the
discrete, quantized and digital signals, as well as on the
quantization error. Our next work will be focused on further
developing the proposed approach to conducting web distance
learning exercises.
REFERENCES
[1] W. Y. Yang, T. G. Chang, I. H. Song, Y. S. Cho, J. Heo, W. G.
Jeon, J. W. Lee, J. K. Kim, Signals and Systems with MATLAB,
Springer, 2009.
[2] S. Karris, Signal and Systems with MATLAB; Computing and
Simulink; Modeling, Fifth edition, Orchard Publications, 2012.
[3] V. Georgieva, P. Petrov, Signals and systems - manual for
laboratory exercises, King, 2016 (In Bulgarian).
[4] L. Chaparro, Signal and Systems using MATLAB, Academic
Press (Elsevier), 2011.
[5] G. Nenov, Signals and systems, Novi Znania, Sofia, 2008 (In
Bulgarian).
[6] Е. Ferdinandov, Signals and systems, Siela, Sofia, 1999 (In
Bulgarian).
 Model of Project-Based Learning Platform
Ava Chikurteva1 and Denis Chikurtev2
Abstract – The implementation of the Project Based Learning
(PBL) in education facing a number of problems and difficulties,
prompting teachers to avoid its use and rely on traditional
methods of training. The article examines the role of ICT in PBL,
the benefits of PBL and its main problems in its implementation.
A conceptual model of a PBL platform in education is proposed
to guide and assist the teacher in planning, implementing and
presenting project results. The role of the teacher is leading and
involves preparing, presenting a project, motivating students,
managing the process through continuous communication and
facilitating.
Keywords – Project-based learning, e-learning, Education,
ICT.
I. INTRODUCTION
Project-based learning is defined as “an instructional
method that gives students complex tasks based on
challenging questions or problems that include problem
solving, decision making, research skills. It focuses on issues
that make students come up with basic concepts and principles
in practical terms. Students do their own research on a leading
issue that allows them to develop valuable research skills, as
they participate in the design, problem solving, decision
making and research activities themselves. Students learn
from these experiences, take them into account, and apply
them to the world outside their classroom. A different
teaching technique that encourages the practice of forming
new learning habits, emphasizing creative thinking skills and
allowing students to discover that there are many ways to
solve the problem" [2, 6].
The current project-based learning model was developed by
the Buck Institute for Education in the late 1990s in response
to school reform efforts [14]. Unlike ordinary projects, PBO
requires critical thinking, problem solving and communication
between participants [3]. Students must organize their own
work and manage their own path when working on the
project.
PBL is the interactive method that can most strongly cause
a change in the teaching culture, involving collaboration
between both students and between student and the teacher,
who often acts as a facilitator [4, 8]. The feedback that the
teacher periodically gives is concrete, useful, positive [1].
The advantages the PBL offers over traditional teaching
methods are:
1Ava
Chikurteva is with Institute of Information and
Communication Technologies, Bulgarian Academy of Sciences,
Sofia, Bulgaria, E-mail: avapaleva5@gmail.com
2Denis Chikurtev is with Institute of Information and
Communication Technologies, Bulgarian Academy of Sciences,
Sofia, Bulgaria, E-mail: dchikurteva@gmail.com
978-1-7281-7142-5/20/$31.00 ©2020 European Union
• Students receive adequate preparation for real life.
Project-based learning enables students to relate what
they do in school to the big world around them - to ask
questions and to find a solution themselves, using their
available knowledge and experience. Through project-
based learning, they learn and practice skills that
traditional teaching often neglects - group work,
decision making, process monitoring, in-depth
problem-solving, communication, and feedback. Along
with the new knowledge, students acquire all the skills
required by the rapidly changing global world [9].
• Learning material is better absorbed. Studies show
that, compared to the traditional way of teaching,
students learn the learning material as well or even
better. The advantage of project-based training is that
trainees remember the lessons learned longer and are
able to use the knowledge they have acquired [11].
• Improving students' attitude to learning. According
to the results of studies, project-based learning in
mathematics reduces students' anxiety and creates a
positive attitude towards the subject [10].
• Positive effect on the relationship between students'
achievements and their social background. In
schools that work predominantly on projects, students'
success is not directly dependent on their social status,
while in traditional schools this relationship is clearly
evident.
• Collaboration between students is developing.
Interpersonal conflicts must be seen as an opportunity
for change. Classes with real problems cause students
to develop teamwork and teamwork skills, which are a
prerequisite for successful professional and personal
realization [8]. In addition, during their work, students
learn to overcome various difficulties associated with
more passive classmates, difficulties in coordinating
online projects, meeting deadlines, or using technical
means.
• Increases computer literacy of teachers and students
- ICT is present at every stage of project organization,
implementation and presentation [7, 12]. This unlocks
the potential of information technology in training and
allows PBL to be distance learning.
All these advantages over traditional teaching are only
observed when project-based learning is planned and managed
properly [5, 13].
In order to structure the study model and to formulate its
functions and services, it is necessary to identify and
categorize the problems that teachers face when using the PB
method.
81
 II. PBL PROBLEMS
In PBL, the teacher becomes the facilitator and the students
work under his guidance. This article explores the problems
associated with creating and managing teacher's PB lessons.
The problems that teachers encounter when creating a PB
lesson are the following [11, 13]:
• to identify important issues;
• to structure meaningful tasks;
• to develop knowledge;
• to develop social skills;
• to evaluate what students have learned from the
experience.
The specific challenges faced by teachers at PBL are related
to:
• the selection or design of situations that suggest the
possibility of good projects;
• structuring problems as learning opportunities;
• working with colleagues to develop interdisciplinary
projects;
• dynamic management of the learning and self-learning
process;
• technology integration;
• assisting students in authentic assessment.
There is a great deal of research on PBL, proving the
benefits and effectiveness of the method, ways of its
application, various developments of PB lessons and results of
its application, but there is no comprehensive platform to
integrate this data and provide it to the user in a convenient
and selective form. In order to disseminate and promote the
use of PBL as an integral part of the curriculum, supporting
materials, guidelines, and appropriate conditions for working
with PBL must be created.
III. MODEL OF THE PROJECT-BASED LEARNING
PLATFORM
The research model provides an opportunity for the teacher
to prepare and deliver a project-based lesson, following a
series of steps to create a new project. To create a PB lesson,
it is necessary to follow the technological modular model
presented in Figure 1. Using a modular model allows
consistency in the formation of the individual steps, tasks and
objectives of the lesson. On the other hand, modularity allows
modifying and editing the different parts of the individual
modules for a lesson. This means that the teacher can easily
update or modify existing projects. In this way, teachers can
share models of successfully completed PB lessons that can be
used as a basis for forming good practices for creating
improved future lessons in PBL.
The module manager takes care of organizing the
participating modules in one lesson. It interconnects modules
82
and provides tools and services for working with modules. It
is important to mention that proposed modules are developed
according to the minimal needs for successful PBL
application.
Modules 1 to 5 are the main modules that make up a single
PB lesson. Each of these modules can be removed or edited as
desired by the teacher.
Module 1 - Determination of Expected Results - Working
in a team of teachers to formulate and develop the project
idea; Deciding on the scope of the project; Selection of
indicators for achievement of the result; Determination of
overall results; Working on project design criteria; Creating
an optimal learning environment;
Module 2 - Preparation of leading questions - Design of
questions; Clarification of questions; Generalization of
questions; Development of standard based questions.
Module 3 - Evaluation Planning - Defining the Outcome
and Evaluation Criteria; Compare the product with the result;
Use of headings;
Fig. 1. Modular structure of the PBL system
Module 4 - Project Map - Organization of activities by the
teacher, which creates the environment and provides the
preconditions for successful independent research work of the
students; Directing students' attention to key issues and
focusing on ways to solve the problems as the project
progresses; Collecting resources from students; Forming the
Performance Matrix i.e. developing a plan for the
implementation of the project and its practical
implementation;
Module 5 - Project Management - Understanding the role of
the teacher as a Manager; Sharing goals with students; Using
different methods and tools to solve problems; Use of anchor
points and milestones; Assessment and reflection plan.
Module N - represents any additional module added by the
teacher. The system provides the opportunity to add new or
existing modules both at the end and between already installed
modules.
A. Tools, Functions, Services
A key element in the implementation and application of
PBL is the use of ICT. Thanks to ICT, we can create, manage
and present the content of a project. In order to implement
PBL, we need to research and develop an ICT platform that
enables the creation and management of PB lessons online.
Such a platform should provide a range of tools, functions and
services to facilitate and optimize the work of teachers and
students in creating and delivering a lesson.
The PBL platform should contain a large amount of
materials and data in different fields, subjects and topics.
These materials must be checked for authenticity. Teachers
will be able to add new materials and edit current ones. The
materials to the platform database are: lessons, articles,
lectures, e-books, presentations and videos.
Tools for creating and editing modules have been
developed. These tools are part of the module manager, they
are represented by e specific build-in software functions, part
of the Java script programs. The tools make it easy to edit
texts, assignments, tutorials, and multimedia content.
Evaluation criteria and automated checks of results can be
created and edited.
Functions and services have been developed for the gradual
verification of the completion of the tasks, by setting tags and
a template for the completion of the stages of a project.
Functions have also been developed for connecting different
components from one module to another. In this way,
evaluation criteria or guidelines for the implementation of a
project can be set. Functions and services development are
implemented in the Python scripts, as part of the web server
framework.
Fig. 2. Structure of information on work with the PBL
An important step in working with PBL is to become
familiar with the nature of this type of training and to learn the
basic methods of work. Therefore, specialized information for
teacher training is structured. In this way, they will get into
the matter and become familiar with the opportunities offered
by the developed platform for working with PBL. Figure 2
illustrates developed structure. The main problem in the
structure is "What is PBO". To answer this question
completely, there are several basic categories: basis,
organization, facilitation, application of ICT and result. Each
of these categories contains detailed information organized
into sub-categories represented in the blue rectangles.
83
The main service of the platform under study is the Data
Selection Service, which is presented in Figure 3. When
creating a new lesson according to set criteria: subject, class
and / or keyword, automatically output information about
already completed projects and study materials according to
the assignment of the teacher. This service requires the
introduction of at least one of the specified criteria. Increasing
the criteria increases the specificity of the information
provided by the service. This facilitates the work of the
teacher by limiting the amount of information required by
entering keywords.
By default, the provided information is sorted by class and
subject, where the keywords occur. In addition, the user can
change the ranking criteria according to popularity, rating,
class, subject, date and more.
The information in the database is organized by subjects
and by class. This makes it possible to distinguish the
specifics of the subject, taking into account the age of the
students (by class). When we have opened a subject for a
class, we will have references to basic documents for the
subject of the class - State educational requirements, curricula,
references to other sources of information; lesson units;
development - plan of a real problem for the category;
problem projects that have already been worked on; ideas for
new project-based lessons.
Fig. 3. Organizing information by subject, class, and keyword
IV. TECHNOLOGY DESIGN OF THE PLATFORM
A PBL portal can be implemented in the form of a website
or web application. For this purpose, the latest ICTs in the
field of mobile and web technologies and database
technologies are used.
The model consists of three main blocks: databases,
management system and user interface. Each of these
components plays a key role in the implementation of the
above functionalities of the PBL model.
The database management unit stores all the information
that is needed to create the PB lesson and implementation.
Databases are based on the MySQL technologies and they are
accessed by Python scripts. Into the databases, the
management system stores the already entered training
materials and lesson templates, records the new lessons, as
well as the profile data of each registered user. The sorting
and searching of data is performed by standard database
sorting algorithms.

Fig. 4. General Architecture of the PBL platform
The PBL platform management system is made up of
multiple features and objects that are designed to connect the
user interface to the database. The technologies used for
development of the proposed system are Nodejs and Python
Django frameworks. The system includes all the tools and
services for creating and working with lesson modules. This
block executes requests and commands received from the user
interface that are processed according to certain algorithms.
This is how complex automated platform management
processes are implemented. The management system, on the
other hand, sends requests to the database according to
requests received from the user interface or according to the
necessary data for the proper functioning of the working
algorithms or the results of their operation. A module for
providing cyber security and user authorization is also
included in the management system. Depending on whether
the user is a teacher or a student, he or she is granted different
access rights to projects and databases. If a user is not logged
in, then the system only allows access to some information
resources such as - what is PBO, some of the projects already
completed and others. A guest type user is not allowed to
create or edit projects, or to access profile information of other
users.
The user interface provides interaction between the user
and the platform. It provides the necessary menus and buttons
to operate the model. All platform services and tools can be
used through the interface. The interface illustrates the
necessary information requested by the user. This information
can be presented as multimedia or as an electronic textbook.
The connections between the WebUI and the management
system are implemented by Javascript packages. The frontend
visualization and design are developed by HTML and CSS.
V. CONCLUSION
Project-based learning has the huge potential of a method
that develops key competences and skills of the 21st century.
The implementation of the method makes it possible to
increase the efficiency of the education system, implying
intensive, focused and 'smart' use of ICT. The teacher must
have the necessary training to successfully and fully
implement the PBL. The model developed allows the teacher
to apply the method quickly and easily. The model has the
necessary functionality to learn, create and manage project-
based lessons. This model will contribute to the development,
implementation and dissemination of PBL.
For future work it is envisaged development of graphical
web-based user interface and communication services: video,
chat, forum.
84
ACKNOWLEDGEMENT
The paper is supported by the National Scientific Program
"Information and Communication Technologies for a Single
Digital Market in Science, Education and Security
(ICTinSES)", financed by the Ministry of Education and
Science.
REFERENCES
[1] S. Bell, “Project-Based Learning for the 21st Century: Skills for
the Future”, The Clearing House: A Journal of Educational
Strategies, Issues and Ideas, 2010, vol. 83, no. 2, pp. 39-43,
DOI: 10.1080 / 00098650903505415
[2] J. W. Thomas, A Review of Research on Project-Based
Learning, 2000.
[3] P. C. Blumenfeld, E. Soloway, R. W. Marx, J. S. Krajcik, M.
Guzdial, and A. Palincsar, “Motivating Project-Based Learning:
Sustaining the Doing, Supporting the Learning”, Educational
Psychologist, 1991, vol. 26, no. 3-4, pp. 369-398.
[4] J. Larmer, J. R. Mergendoller, “Seven Essentials for Project-
Based Learning”, Educational Leadership, 2010, vol. 68, no. 1,
pp. 34-37.
[5] A. M. Novak, J. S. Krajcik, “A Case Study of Project‐Based
Learning of Middle School Students Exploring Water Quality”,
in The Wiley Handbook of Problem‐Based Learning, 2019, pp.
551-572.
[6] J. S. Krajcik, P. C. Blumenfeld, “Project-Based Learning”, in
The Cambridge Handbook of the Learning Sciences, R. Keith
Sawyer (ed), Cambridge University Press, 2006, pp. 317-334.
[7] V. B. Gómez-Pablos, M. M. del Pozo, and A. G.V. Muñoz-
Repiso, “Project-Based Learning (PBL) through the
Incorporation of Digital Technologies: An Evaluation Based on
the Experience of Serving Teachers”, Computers in Human
Behavior, 2017, vol. 68, pp. 501-512.
[8] C. L. Chiang, H. Lee, “The Effect of Project-Based Learning on
Learning Motivation and Problem-Solving Ability of Vocational
High School Students”, International Journal of Information and
Education Technology, 2016, vol. 6, no. 9, pp. 709-712.
[9] M. E. Beier, M. H. Kim, A. Saterbak, V. Leautaud, S. Bishnoi,
and J. M. Gilberto, “The Effect of Authentic Project‐Based
Learning on Attitudes and Career Aspirations in STEM”,
Journal of Research in Science Teaching, 2019, vol. 56, no. 1,
pp. 3-23.
[10] V. L. Holmes, Y. Hwang, “Exploring the Effects of Project-
Based Learning in Secondary Mathematics Education”, The
Journal of Educational Research, 2016, vol. 109, no. 5, pp. 449-
463.
[11] M. J. Harris, The Challenges of Implementing Project-Based
Learning in Middle Schools , Doctoral dissertation, University
of Pittsburgh, 2015.
[12] D. G. Moursund, Project-Based Learning Using Information
Technology, Eugene, OR: International Society for Technology
in Education, 1999.
[13] Y. R. Totseva, Project – Based Learning – Technology and
Capabilities for Use in Bulgarian Schools, Reports Awarded
with "Best Paper" Crystal Prize, pp. 303-308, 2016.
[14] Buck Institute for Education: http://www.bie.org
 Practical Methods for Teaching of Solar Cell
Characterization
Nikola I. Mitrović1 and Milan D. Stojanović2
Abstract – This paper gives insight into methods for practical
teaching in the field of solar cell characterization. Since machines
for solar cell characterization are mostly situated in some large
factories, it is often hard to give student opportunity to get in
touch with characterization processes. In this paper, the
chamber that can perform basic characterization is designed and
tested. The entire device is built around PIC microcontroller that
is interfaced with multiple sensors and with a display which gives
characterization results to the user. Setting and measuring of
environmental conditions together with measuring solar cell
output parameters as open-circuit voltage or short-circuit
current is enabled. Characterizations of multiple solar cells have
been done, with the goal to confirm solar cells parameters. Main
characteristics of solar cells are then discussed.
Keywords – Solar cell, Characterization, PIC microcontroller,
Education.
I. INTRODUCTION
Solar cells and solar systems are taught in the most of
faculties of electronic engineering worldwide. The main focus
of these studies is mostly structure, basic parameters and
applications of solar systems, but physics and composition of
these systems is also presented. Still, laboratories and
factories that are studying and producing solar cells aren’t that
well spread [1, 2]. Machines needed for these processes are
very expensive, and therefore, can be found mostly in large
factories. Because of that, students often don't have the
possibility to have a direct contact with these technologies and
to deepen their knowledge. In order to compensate this lack of
contact with concrete machines, teachers sometimes resort to
developing some small-size replacements. One example of
this type of machines is a characterization chamber. Since that
it is often case that there aren't any factories using this
machine in the near area, it is almost impossible to achieve
visits of students to those factories [3, 4].
The complete characterization chamber is designed and
tested to describe and simulate characterization process and to
improve practical knowledge of the students. Section II of the
paper explains characterization process and parameters that
should be measured, while section III presents design of
hardware needed for the measuring and assembly of the
chamber. In the section IV examples of using of chamber are
given, and results are discussed.
1Nikola I. Mitrović is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
nikola.i.mitrovic@elfak.ni.ac.rs
2Milan D. Stojanović is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: milan.stojanovic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
II. THEORETICAL BACKGROUND
A characterization of solar cell presents a process of
accurate measuring of basic parameters of a solar cell.
According to these parameters, an appropriate application of
the cell is selected. Solar cells are then combined into panels
and into modules, which represent most used method for
obtaining solar energy. Like for any other measuring, it is
needed to first set some values as standard, so that later
measured results can be compared with the standard ones.
These standard values for measuring are usually acquired in
laboratories, under certain conditions. An industry-wide
standard for indicating solar cells performance was adopted
and marked as STC (Standard Test Conditions) [5].
Output parameters of the solar cell are dependent on the
properties of light that the cell is subjected to and to position
between light source and the cell. Besides that, since solar
cells are mostly made of semiconductors, its parameters shift
with the change in temperature. Under STC, solar cell is
exposed to light irradiance of 1000 W/m2, at the cell
temperature of 25°C and with an Air Мass 1.5 (AM1.5)
spectrum [6]. Sun irradiance presents the power delivered to
the Earth by Sun. It is worth mentioning that an irradiance as
high as STC irradiance isn’t possible to be induced by Sun
everywhere during the day on the planet. Especially is average
irradiance for Europe significantly lower. Air mass spectrum
presents “number of obstacles” that light beam needs to
overcome in order to reach the destination. Since the Sun
beams travel through air, Air Mass spectrum is set to 1.5.
All of the mentioned conditions must be enabled to conduct
appropriate measuring of the cell parameters. Mostly all of the
data on the solar cells and panels given in datasheets is
acquired with measuring under STC. However, datasheet STC
values often tend to be insufficient because irradiance and
temperature change with different parts of the day and with
different seasons. That is why it is preferred to have insight
into values and parameters not only under some constant
conditions, but for the different ranges of different conditions.
Importance of solar cell characterization is essential
because the cells are composing units for solar panels and
solar modules. For the panel to work properly and to provide
maximal power, all of the cells should have as matching
characteristics as possible. Beside electrical characteristics,
physical characteristics, as packaging density or cell shape can
create difference in the output parameters of a solar panel as
well. That is why connection between solar cells is very
important part of the panel designing process [5]. Cells can be
connected in two ways. Since the solar cell is practically a p-n
junction, it has two terminals to be connected. If a p-side of
the solar cell is connected to the p-side of the other cell,
85
connection is parallel, while if a p-side of the solar cell is
connected to the n-side of the other cell, connection is in
series. Both types of connection have its causes, although
mainly, cells are connected serially. With the serial
connection, output voltage of the connection is increased. This
is because output voltage of a single cell is often too small
(0.6 to 0.7 V) to provide appropriate supply for specified
circuit. It is important to note that capabilities of solar panels
are acquired by accurate characterization and appropriate
connection of the solar cells.
III. DESCRIPTION OF THE DESIGNED CIRCUIT
With the goal to give students opportunity to explore
characterization concept in more details, structure of the
designed chamber needs to be divided in multiple parts. As
stated, it is needed to ensure mechanisms to simultaneously
measure temperature and light intensity, and of course to
measure output current and output voltage. The similar
characterization chambers, based on Arduino platform, were
presented in papers [7] and [8]. Basic diagram of the chamber
is given in Fig. 1.
Fig. 1. Basic diagram of the designed characterization chamber
Control part of the chamber is composed out of two printed
circuit boards (PCB). Boards and electrical circuits are
designed using PCB software and later manufactured. One
PCB is designed to provide light that illuminates solar cell
(Light Source Block), and other PCB is designed to perform
control and measuring of the parameters (Control Block).
A. Light Source Block
Basic schematic behind Light Source Block is given in
Fig. 2. This block is composed out of 32 LE diodes that
organized in the groups of 8 diodes. The diodes produce light
of white color and they can be turn on/turn off by transistor
Q1. The concrete model of the transistor is IRF520 [9], which
is chosen because of its superior switching characteristics. A
potentiometer which value is 10 kΩ, mounted on the case of
chamber and connected to the connector J1, is used to regulate
current intensity across the diodes when they turned on. Using
this potentiometer, intensity of light induced by diodes can
change in range from 0 to 2500 lux, by changing current.
86
The benefit of this designed PCB is possibility to solder
diodes that produced light of other colors (red, green, blue,
etc.). It can be used to achieve different spectrum of light and
test the behavior of cells under different wavelengths.
Fig. 2. Schematic of the Light Source Block
Designed and manufactured PCB board for this block is
given in Fig. 3.
Fig. 3. Manufactured PCB for Light Source Block with components
B. Control Block
Control Block is built around microcontroller
PIC18F45K22 [10]. It is a powerful microcontroller that
operates at 5 V and has an internal oscillator. It has 44 pins, so
it can interface a fair amount of other circuits and sensors.
Control Block is divided into environment measurement
subblock and cell parameters subblock.
Environment measurement is comprised of light intensity
measuring which is done using light photo sensor APDS9005
[11] and temperature measuring which is done by using
temperature sensor LM35 [12]. APDS9005 is connected with
resistor, forming a voltage divider. Light intensity determines
output analog current of the APDS9005, that flows through
divider. Voltage from the voltage divider is delivered to the
analog pin of the MCU, where ADC is executed. The
datasheet gives insight into calibration with which light
intensity can be measured. On the other hand, temperature
sensor LM35 does not demand any additional components in
order to operate. It is a three-pin device where one of the pins
is output. Change in the temperature is manifested as a change
in output voltage, which is brought to the other analog pin of
the MCU for ADC.
 Cell parameters subblock is designed to measure main solar
cell properties. Parameters of the greatest interest in solar cell
characterization are short-circuit current (ISC) and open-circuit
voltage (VOC). ISC presents the current between two contacts of
the solar cell when they are short-circuited, while VOC presents
voltage between contacts of the solar cell when they are not
connected (when circuit is open, so there is no current
flowing). However, these values should not be confused with
values measured in some applicable circuits. Basic solar cell
measurements, especially one of ISC, are mostly done under
some resistance, where these parameters are shifted.
Experience has shown that students mostly misinterpret these
values. So, in order to clarify and to properly measure ISC and
VOC, additional circuits must be designed (Fig. 4).
environment. To enable all of this, it is needed to program PIC
MCU. With MCU programming, all values are constantly
being received, calculated and presented to the student or user.
Range of measured ISC is up 15 mA, and up to 8 V for VOC.
For presenting results 128x64 Graphical Liquid Crystal
Display (GLCD) [15] is used. Printed circuit board for all of
the connectors and for all of the described components is
given in Fig. 5. Beside described components, board contains
push buttons, test LEDs, contrast potentiometers, decoupling
capacitors and others.

Fig. 5. Manufactured PCB for Control Block with components

For the end of the design process, it is needed to

encapsulate the board and to provide a suitable box. Final
form of the device is given in Fig. 6.

Fig. 4. Additional circuits for ISC and VOC measurements

Additional two MOSFETs BSS138 [13] are added to the

two measuring circuits to enable accurate measuring.
Transistor Q2 is related to current measuring, while transistor
Q1 is related to voltage measuring. Both transistors have
switching functions in circuits. For the ISC measurement, Fig. 6. Exterior and interior form of the characterization chamber
transistor Q2 is set to on-state to enable short circuit between
two parts of the solar cell, while transistor Q1 is set to off-
state in order to ensure that all of the generated current from IV. CHARACTERIZATION EXAMPLE
the cell will go through the current measuring part. Resistor
R5 is a shunt resistor that is needed for operation of INA213 Characterization can be performed on various solar cells
[14], the circuit used for current measuring. INA213 is an and small panels regardless of their power and size as long as
opamp-based IC which is a current sensing amplifier with the they can fit in the chamber. Example of characterization is
voltage output. Analog output of INA213 is connected to given for the cells showed in Fig. 7. Cells are made of same
another analog pin of the MCU to conduct ADC. Connectors materials, but they differ in sizes.
J12 and J13 are set to give possibility to connect additional For the constant temperature, output parameters of these
resistance for different measuring. For VOC measurement, cells are measured on a range of light intensities. Values of
states of the transistors are inverted. Output of the cell is the output parameters are given constantly on a GLCD, as
connected with the voltage divider, because the output voltage shown in Fig. 8.
of some panels that can be characterized can be greater than
VCC of the MCU that is 5 V. Output of the voltage divider is
forwarded to another analog pin of the MCU, where the value
of VOC is gained with ADC.
All of the analog values received by analog input pins of
the MCU that are used for measuring are, after ADC
processes, further used to determine magnitude of the
parameters. According to the device’s datasheets, it is needed Fig. 7. Characterized solar cells a) Solar cell A (58 mm x 38 mm);
to calibrate and adapt received signals to the changes in b) Solar cell B (53 mm x 18 mm)

87

Fig. 8. Measured values shown in GLCD
According to the literature, with the increase in the light
intensity, linear growth if the ISC and exponentials growth if
VOC is to be expected [2]. Measured results for the set of
different light intensities are given in Fig. 9.
Fig. 9. Measured values for given solar cells
Beside only presenting results, designed chamber can plot
measured parameters on GLCD. For the constant temperature,
chamber measures VOC and ISC for 10 values of light intensity,
ranging from 200 lux to 2000 lux with the step of 200 lux. All
measured values are stored in microcontroller memory and
based on these values, a plot is created. This is very useful for
demonstration of fast characterization and different slopes of
these curves for different PV cells. Plotted results for solar
cell A are given in Fig. 10. In the described way, chamber can
be used to characterize some other photo components as well.
Fig. 10. Demonstration of plotted values on GLCD for PV cell A
Theoretical facts behind solar cell parameters are confirmed
using the chamber. Characterization of different types of the
cells and their capabilities can be further examined. Designed
chamber is easily mobile and adaptable to various courses and
laboratory exercises.
88
V. CONCLUSION
This paper describes design process and testing of the solar
cell characterization chamber that can be used in educational
purposes for teaching of practical methods in this field of
engineering. Using designed device, students can improve
practical skills to complement their theoretical knowledge.
Processes of measuring, testing and data analysis are
emphasized in order to encourage further engagement and to
increase research capabilities.
ACKNOWLEDGEMENT
This work has been supported by the Ministry of Education,
Science and Technological Developmentof the Republic of
Serbia.
REFERENCES
[1] T. C. Kandpal, L. Broman, "Renewable Energy Education: A
Global Status Review", Renew. Sustain. Energy Rev., vol. 34,
pp. 300-324, June 2014.
[2] C. Quinto, L. Ocaña, C. Montes, A. Linares, E. Llarena, B.
González-Diaz, S. González-Pérez, R. Guerrero-Lemus, M.
Friend, and M. Cendagorta, "Design, Start Up and Society
Benefits of a PV Solar Cell Laboratory", 29th European
Photovoltaic Solar Energy Conference and Exhibition, pp.
4113-4118, 2014.
[3] R. Ciriminna, F. Meneguzzo, M. Pecoraino, and M. Pagliaro,
"Rethinking Solar Energy Education on the Dawn of the Solar
Econom", Renew. Sustain. Energy Rev., vol. 63, pp. 13-18,
2016.
[4] G. Nikolov, M. B. Marinov, B. Nikolova, and B. Ganev,
"Characterization of Different Types of Photovoltaic Cells
Using Remote vIrtual System", 2018 Seventh Balkan
Conference on Lighting (Balkan Light), Varna, 2018, pp. 1-4,
doi: 10.1109/BalkanLight.2018.8546992
[5] M. Zeman , K. Jäger, O. Isabella, A. H.M. Smets, and R. A.C.
M.M. van Swaaij, "Solar Energy - Fundamentals, Technology,
and Systems", Copyright of Delft University of Technology,
Delft, Netherlands, 2014.
[6] Elsevier House, Solar Energy Engineering - Processes and
Systems, Amsterdam, The Netherlands, 2013.
[7] W. M. Myint Aung, Y. Win, and N. W. Zaw, "Implementation
of Solar Photovoltaic Data Monitoring System", International
Journal of Science, Engineering and Technology Research
(IJSETR), vol. 7, pp. 591-596, August 2018.
[8] S. A. Jumaat, M. H. Othman, "Solar Energy Measurement
Using Arduino", MATEC Web of Conferences, MUCET 2017,
https://doi.org/10.1051/matecconf/201815001007
[9] IRF520 datasheet. Available at:
https://www.vishay.com/docs/91017.pdf
[10] PIC18F45K22 datasheet. Available at:
https://ww1.microchip.com/downloads/en/DeviceDoc/400G.pdf
[11] APDS9005 datasheet. Available at:
https://docs.broadcom.com/docs/AV02-0080EN
[12] LM35 datasheet. Available at: http://www.ti.com/lit/ds/lm35.pdf
[13] BSS138 datasheet. Available at:
https://www.onsemi.com/pub/Collateral/BSS138-D.PDF
[14] INA213 datasheet. Available at:
http://www.ti.com/product/INA213
[15] GLCD datasheet. Available at:
https://mikroe.com/documents/glcd-128x64-datasheet.pdf
 Servo-system Design for Arduino Due Based
Laboratory Platform
Marija Z. Filipović1, Snežana S. Đorđević2, Vladimir M. Mitić3 and Boban R. Veselić4
Abstract – Realization of a servo-system this is based on
Arduino Due microcontroller is described in this paper. The
developed system was used to create laboratory exercises for
students as well as for experimental purposes. The purpose of
this setup is to demonstrate the difference between systems
working in open and closed control loop, to example different
control strategies and to show how changes in different system
parameters influence the system performance.
Keywords – Arduino Due, Servo-system, Laboratory exercise,
Real-time system, Digital controller.
I. INTRODUCTION
II. SYSTEM DESCRIPTION AND WORKING PRINCIPLE
The purpose of this system as a laboratory exercise for
students is to demonstrate components and working principles
of a servo-system. The students are encouraged to experience
the difference between stable and unstable system, realize
different control strategies, analyze their impact on system
response and learn about different types of responses.
Laboratory platform offers possibility to implement various
controller types in order to notice their performances. Three
types of controllers are currently realized: PI, PID controllers
and a lead compensator.

Servomechanism or servo-system is a type of a closed loop A. System Elements

control system where the controlled variable is mechanical
position or speed. The main characteristic of the The system is composed of the following elements:
servomechanism is fast and precise response of the output Arduino DUE microprocessor [3], CROUZET 82850002dc
variable with respect to the reference input. The name of these motor (24V, 52W) [4], two identical magnetic absolute
systems has been derived from Latin words servus (the one encoders [5], PWM controlled Pololu VNH5019 motor driver
that serves or follows) and mehanism (mechanism). The main [6], a PC power supply and other accompanying active and
requirement in a servo-system design is to have at least one passive components that are necessary for normal system
integrator within the control contour [1]. operation (Fig. 1). Also, a connection to a computer is
The servomechanism working principle is simple: the provided.
output variable is continuously measured and compared with
the reference input via feedback loop. The basic requirement
is minimization of difference between the input and the output
(error signal) in addition to power amplification. This means
that the output variable tracks the input with some accuracy,
where the tracking error is proportional to the system gain [1].
Since the system output can be a mechanical angle, speed
or acceleration, to provide the control task it is necessary to
implement position, speed or current feedback loops.
Servo-systems are used to drive automated processes in
industry, mobile robotics and many other devices used on a
daily basis, such as cruise control systems in a vehicle or
similar [2].
Fig. 1. Block diagram of the designed system

1Marija
Z. Filipović is with the University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: marija.filipovic@elfak.rs
2Snežana S. Đorđević is with the University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: snezanadjordjevic@elfak.rs
3Vladimir M. Mitić is with the University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: vladimir.mitic@elfak.ni.ac.rs
4Boban R. Veselić is with the University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: boban.veselic@elfak.ni.ac.rs
Three voltage levels are required for normal system
operation: 3.3V for Arduino DUE, 5V for encoders and
drivers and 12V for motors. Although Arduino DUE is
incompatible in power supply with other system components,
this particular controller has been chosen because of its
superior processing power, which allows continuous
development and satisfactory performance.
The used encoders in this platform utilize the SSI
communication and are 5V TTL, this is why Logic Leve
Converters were used. For displaying information on the
platform, an LCD was used with four-wire connections, for

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 89

making the wiring much simpler. Also, display menu
navigation is achieved by using an encoder with a push
button.
B. Description of the System Operation
The realized laboratory platform is shown in Fig. 2. In its
simplest form, the system consists of the input knob and the
output knob, where the output knob, driven by the motor,
follows the changes in the input knob. Reference angle is
assigned with the encoder mounted on the knob located on the
right side (1), while the output angle is the motor shaft
position on the left side (2). For easily noticing motor
position, a needle has been attached to the output shaft. The
position itself is measured with the encoder mounted on the
rear side of the motor.
LCD display of 16 by 2 characters resolution is used for a
full control of the system directly on the platform [7]. It shows
the real time values of the reference and the output (2 decimal
digit precision) and also supports setting the referent position
for enabling its precise control with the ability for direct step
response demonstration. Going through the menu the user can
change all the variable parameters of the system including the
controller gains. After the parameter changes are applied, the
experiments can be repeated and the system behavior can be
observed. The goal of this procedure is to experimentally find
the system stability margin and to demonstrate its behavior
under that condition, as well as to show the operation of an
unstable system. In this way, the real physical effects of
different controller parameters are clearly visible.
Fig. 2. Photo of the laboratory platform
The other part of the system is a visual representation of the
various system responses on the PC, where students can
90
observe them, interact with them and record them for further
analysis. This is accomplished using the UART serial
communication interface paired with a custom-made Python
application. By changing the model parameters, it is possible
to record various responses and observe how each parameter
affects the system response.
The designed platform concept allows its operation in
completely stand-alone mode (without PC). Also, the system
can operate using connected PC. The control options over PC
application are shown in Fig. 3.
Fig. 3. PC control interface
III. THE MICROCONTROLLER PROGRAMMING
A. Real Time (RT) Component of the System
Servo systems are intended to work in real-time and
therefore it is necessary that the Arduino DUE program has
both a main loop (loop) and a time-critical loop (RT loop).
During realization, usage of RT loop is crucial, since some
parts of the code must be executed in adequate time instants
i.e.in equidistant time intervals in order to avoid execution
errors. The system must work in RT mode or else it wouldn’t
be able to function at all.
For system to function properly, all calculations must be
done sufficiently fast, i.e. faster than the system discretization
interval. It was experimentally determined that the minimal
discretization interval is equal to 4 ms. In such a short time,
the following must be done: reading of the encoders,
execution of the control loop (speed or position feedback) and
reading of push button encoder for menu navigation. Display
refreshing and communication with PC are not RT critical [8].
B. Parameter Changes and the Display
Beside closing the feedback loop and realization of the
control system itself, implementation of other components
such as display, parameter changes and communication were
also challenging. To enable just taking the control of the
system operation by the PC or the display and releasing it
after that required implementation of a priority scheduler.
It was already mentioned that menu navigation encoder
reading is an RT critical component and therefore its reading,
command interpretation and buffering is done inside the RT
loop. The interpretable commands are: turning clockwise,
turning counter-clockwise and pushing, i.e. enter command.
The main loop executes reading of the buffered commands
and displaying the appropriate content. In case there are no
new commands with in the time of 80 ms, the function for
refreshing the display is executed. This is very important if the
screen displays the value of a tracked variable. This way, the
value will be updated every 80 ms instead of every time the
command changes.
Since the control algorithm is implemented in software, the
transition from positional to speed-tracking system is very
simple. The only thing that needs to be done is to extract the
value of the motor speed which is already being calculated
and apply the appropriate control strategy, while all the other
structures remain the same. This way, the students can
Based on these coefficients, the gain and time constant of
the motor are determined as well as the parameters a and b as:
Km =0.03013, Tm = 0.0146, a=68.49315 and b=2.0637.
To design the digital control system, the Z-transform is
applied with the additional zero-order hold circuit which
approximates D/A convertor transfer function:
{
G ( z ) = Z ⋅ L−1 Gh0 ( s ) ⋅ G ( s ) . } (2)
As the result of the transformation the following is
obtained:
b z ( aT + e − 1) − aTe − e + 1
− aT − aT − aT

observe how the hardware-software coupling influences the G( z) =

system [11]. a2 z 2 − z ( e − aT + 1) + e − aT
(3)
0.000015004 ( z + 0.9178885630498 )
G(z) =
IV. PI CONTROLLER DESIGN ( z − 1)( z − 0.7607 )

As stated earlier, there are currently implemented three For clarity purposes, the following notation in the digital
controllers in the system. Only one controller at a time domain will be adopted:
calculates the control signal and we can choose which one at b0 z + b1
any time during the exercise. All the controllers were G( z) = (4)
z 2 + a1 z + a2
designed using pole placement method [9]. As an example of
the design procedure, PI controller design will be presented. a1 = −1.7607; a2 = 0.7607
b0 = 0.000015004; b1 = 0.009279500
A. Step Response Recording and System Identification
B. Digital PI Controller Design
A special purpose Arduino code was written for
determining the step response of the system. Step response After the system parameters identification, PI controller
assignment is issued via serial communication by applying gains Kp and Ki are determined using the pole placement
50% of the PWM pulse, resetting the time measurement method. Let the closed loop system dynamics be characterized
system and staring the data transfer to the PC. The PC by relative dumping ratio ξ = 0.9 and the natural in damped
receives the data which consists of current timestamp in frequency ωn =12 rad/s.
milliseconds, current position and current speed. The position
vs time us shown Fig. 4, which was used for system The required technical conditions are determined by the
identification [10]. two conjugate complex roots of the characteristic equation in
s-domain, s1,2 whose location are calculated from:

s1,2 = −ζωn ± jωn 1 − ζ 2

s1,2 = −10.8 ± j 5.23 (5)

Design of the digital controller requires that these closed

loop poles are located in the Z-domain using the following
expressions:
1−ζ 2
Fig. 4. Step response of the motor position z1,2 = e−ζωnT ± jωnT ,
z1,2 = 0.95751 ± j 0.020034 . (6)
The transfer function for the given response is represented
as: where T is the discretization interval, which is T=4 ms.
θ (s) km b The generalized form of the desired closed loop system
G ( s) = = = . (1) characteristic equation is given with:
U ( s) s (1 + Tm ) s(s + a)
z 3 + z 2 ( − ( z1 + z2 ) − σ z ) + z ( z1 z2 + σ z ( z1 + z2 ) ) − z1z2σ z = 0 . (7)
It is obvious that the system is astatic with a negatively
dumped response. The steepest tangent is constructed and its σz is the third real pole of the closed loop system.
coefficients calculated as follows: Open loop transfer function with the applied PI controller is
k = 6.026; n = −0.088263 . in the following form:

91
  z 
Wp ( z ) =  K p + K i W ( z ) .
 z −1
Based on this transfer function, it is possible to derive the
characteristic equation of the closed loop system as:
1 + Wp ( z ) = 0 ,
z + z ( a1 −1 + b0 K p + b0 Ki ) + z ( a2 − a1 − b0K p + b1Ki ) − a2 − b1K p = 0 . (10)
3 2
Taking the obtained values of z1,2 into account, the
following holds:
z1 z 2 = 0.917226761; z1 + z2 = 1.91502 .
By comparing the equations (7) and (10), a system of three
equations with three variables (Kp, Ki, z) is obtained:
a1 − 1 + b0 K p + b0 K i + z1 + z2 + σ z = 0 ,
a2 − a1 − b0 K p + b1 K i − z1 z2 − σ z ( z1 + z 2 ) = 0 ,
−a2 − b1 K p + z1 z2σ z = 0 .
By solving these equations, the following values of the
controller parameters, along with the third, real root of the
closed loop system characteristic equation are obtained:
K p = 0.536426177157006; K i = 0.012432547378634;
σ z = 0.837727002982122 .
The designed digital PI controlleris implemented in its
incremental form by the microcontroller, which means that the
current value of the control signal is calculated usingits
previous value, as well as the previous and current position
error signal. The output of the controller is directly
proportional to the PWM pulse width [9].
U ( kT ) = U ( ( k − 1) T ) + Ki ⋅ e ( kT ) + K p ( e ( k ) − e ( k − 1) ) . (12)
Fig. 5. Step response of the positional system
The correct system functionality is verified by observing its
(8) recorded step response. After the controller is implemented by
the microcontroller, the following step response is obtained
(Fig. 5). It is evident from the figure that the desired referent
position is achieved by the designed servo-system. Also, a
larger overshoot than analytically predicted can be noticed.
(9) This overshoot is produced by nonlinear effect of the motor
caused by a dry friction.
V. CONCLUSION
The developed servo-system laboratory platform presented
in this paper, is intended to give students the insight on how
both position and speed servo-system can be implemented
using the same hardware. Differences and transitions between
the two of them can be done by using the appropriate
computer software. Additionally, everything that has been
(11) implemented on a PC can also be implemented on a tablet
computer or server and the communication can optionally be
made available over Wi-Fi connection. Considering that the
most important parts of the system are implemented in
software, the same platform can be used for experimental
research. Advanced topics such as design of different control
strategies as well as further development of the
microcontroller software can be conducted by senior students,
from which they can observe how the design of a real time
system is complex and how much care and attention should be
applied to the smallest of details.
REFERENCES
[1] M. Nakamura, S. Goto, and N. Kyura, Mechatronic Servo
System Control, Saga University, Japan, 2004.
[2] M. Bjekić, Z. Stevanović, Servoregulator Lexium LXM32M,
str.7http://www.empr.ftn.kg.ac.rs/7%20REGULACIJE/Lexium
%20LXM32M.pdf
[3] Arduino Due schematic,
https://www.arduino.cc/en/uploads/Main/arduino-Due-
schematic.pdf
[4] Crouset Motors, model Ø 42 mm 22 to 52 Watts 52 W Part
number 82850002,https://soda.crouzet.com/pn/?i=82850002
[5] AEAT-6010/6012 Magnetic Encoder,
https://docs.broadcom.com/docs/AV02-0188EN
[6] Data sheet Pololu VNH5019 Motor Driver Carrier,
www.https://www.pololu.com/product/1451
[7] LCDWH1602B2-TMI-ET# Data Sheet, lcd-2x16-datasheet,
http://www.winstar.com.tw
[8] K. Ogata, Discrete-time Control Systems, N. Jersey, Prentice-
Hall International, Inc., 1995.
[9] M. B. Naumović, Design of Automatic Control Systems, Faculty
of Electronic Engineering, Niš, 1995.
[10] B. Danković, D. Antić, and Z. Jovanović, Process
Identification, Faculty of Electronic Engineering, Niš, 1996 (in
Serbian).
[11] M. M. Maung, M. M. Latt, and C. M. Nwe, “DC Motor Angular
Position Control using PID Controller with Friction
Compensation”, International Journal of Scientific and Research
Publication, vol. 8, no. 11, November 2018.
92
 Reception of Audio Signals Received from
Different Angles by Optical Methods
Tsvetan Valkovski
Abstract – In this article we will look at the possibilities of
receiving audio signals from different angles through optical and
optoacoustic methods. An analysis of the optical methods for
transmitting information and transmitting audio signals will be
performed.
A model of a laboratory environment for transmitting audio
signals by optical methods, the so-called Li-Fi system, is
considered. The laboratory environment is designed to be
presented to students and to perform experiments in the
laboratory, as well as the opportunity to prepare term papers
related to optoelectronics and acoustics.
Keywords – Acoustics, Optoelectronics, Education, Audio,
Optical, Optoacoustics, Li Fi.
I. INTRODUCTION
In practice, various methods are used to receive and
transmit audio signals. Some of these methods are
optoacoustic, through optical microphones and optical fibers.
They offer extremely large capabilities for collecting all types
of audio signals, their transmission and analysis.
Optoacoustic methods allow us to extract audio signals
using photonic principles for accurate measurement of sound,
vibration, pressure and other key physical characteristics in
places where conventional sensors cannot be used. The use of
fiber optic sensors and sensor systems in microphones,
headphones, noise canceling devices, accelerometers and
pressure sensors provide safe, highly efficient and reliable
service in the most difficult and demanding applications.
Through optical fibers we can transmit audio signals over long
distances, with high speed, minimal delays and without
electromagnetic interference on the track from other electrical
equipment, tracks, etc.
The problem we have focused on is the impossibility and
insufficient efficiency of using conventional methods for
extracting audio signals and the advantages that we could
have using optical methods. These conventional methods of
transmitting audio signals are wired and wireless. The former
have a problem with insufficient device mobility and lower
data rates. With wireless methods, depending on the
technology, there are problems with the reliability of the
connection and the speed of data transmission. The method
for Li-Fi systems is suitable for systems requiring high speed
audio transmission.
Tsvetan Valkovski is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: tsvalkovski@gmail.com
978-1-7281-7142-5/20/$31.00 ©2020 European Union
II. THEORY
Visible light communication (VLC) is a centuries-old
technique that uses visible light to send messages from one
place to another. In the past, communication by flame was an
effective way to transmit signals from one place to another at
a distance. The principle of using marine headlights for ship
signalling is similar. Nowadays, for example, a similar
method can be seen in lights mounted on modern tall
buildings to indicate their presence to aircrafts [1]. Along with
the development of telecommunications, using visible lights
instead of other electromagnetic waves to transmit
information, to attract the attention of scientists by reviewing
a famous experiment by Alexander Graham Bell in 1880
using photons in his experiment [2] is done. In the
experiment, the voice signal was modulated on sunlight and
the information was transmitted remotely from about 200 m.
Efforts to study natural lights and artificial lights for
communication have been going on for decades. In 1979, FR
Gfeler and G. Bapst demonstrated this technical feasibility of
optical wireless communication using infrared light emitting
diodes (LEDs) [3]. Whether infrared or visible light will be
used depends on the source (LED) and the receiver
(photodetector). The photodetector could be a photoresistor, a
solar panel and others.
Visible light communication has many attractive
advantages, which include but are not limited to the following
aspects - wide spectrum, no electromagnetic interference, easy
implementation, low cost, high energy efficiency, health
safety, information security and others.
The basic environment for transmitting audio signals
through visible light consists of six blocks, shown in Fig. 1. In
this article we will focus on the transmission of the signal
through light, i.e. the connection between the LED and the
photodetector.
Fig. 1. Block scheme of basic Li-Fi system
93
 The main losses in Li-Fi systems are observed in the
transitions source-fiber, free space channel (communication
between LED and photodetector). Therefore, the losses must
be minimal.
The losses on the free space channel are mainly of three
types - distance losses, "visibility" losses and angular losses.
The main task in the article is to consider the angular losses
and the problem of obtaining an audio signal from different
angles, because from previous experiments of ours and other
colleagues the influence of the distance in the transmission of
audio signals through visible light is considered.
Different statistical methods can be used to process the data
obtained from the system and estimate the angular losses. One (a)
of them is the so-called "Statistical probability of loss" [4].
For the most part, the angular losses are caused by the angle
of incidence of the light beam on the photodetector, but from
the point of view of diffraction, these angular losses decrease
if the diffraction increases, as this is observed at relatively
short distances [5-7].

III. ANGULAR LOSSES IN LI-FI SYSTEMS

The article will conduct a study on the built Li-Fi laboratory
environment, through which is possible the transmission of
audio signals by optical methods.
The experimental part includes a study of the individual
components, a study of previous developments and the
dependencies through which the angular losses can be studied.
The basic model of the laboratory environment is shown in
Fig. 2.
(b)
Fig. 3. Laboratory environment for measurement of the signal level
at 15 cm and change the angle of signal reception by changing the
angle of the solar panel cell: (a – the laboratory environment includes
Li-Fi system, tone generator and oscilloscope; b – the laboratory
environment includes conveyor to change the angle of the solar
panel)
TABLE I
THE SIGNAL LEVEL WHEN CHANGES THE ANGLE OF SIGNAL RECEPTION
BY CHANGING THE ANGLE OF THE SOLAR PANEL CELL
AT 15 CM DISTANCE

Angle, o 0o 15 o 30 o 45 o
U, V 0,34 0,335 0,33 0,32
Angle, o 60 o 75 o 90 o
U, V 0,315 0,285 0.021

Fig. 2. Block scheme of the Li-Fi laboratory environment

The actual experimental setup is shown in Fig. 3. It

includes a tone generator, Li-Fi system, oscilloscope and
conveyor to change the angle of the solar panel.
Fig. 4 and Table 1 show the results obtained at the signal
level at a distance of 15 cm and change the angle of signal
reception by changing the angle of the photodetector (solar
panel). Measurements were made from 0 to 90 degrees,
because after 90 degrees, the angle of view is zero and there is
practically no possibility to receive a signal.
The results show that the angle of signal reception does not Fig. 4. Graphic representation of the readings for the signal level
strongly affect the reduction of the signal level, as the when changes the angle of signal reception by changing the angle of
decrease in the signal level is 0.005 volts at a change from 0 the solar panel cell at 15 cm distance
to 75 degrees, this is due to the nature of visible light,
diffraction and its uniform distribution throughout the "field After reviewing the results, it can be said that the change in
of vision". the signal level from 0 degrees to 75 degrees is a decrease of
17.46% compared to the initial signal level. Fig. 5 shows the
percentage of signal level losses when changing the angle of
the solar panel [8].

94

Fig. 5. Graphic representation of the signal level losses
After comparing the obtained angular losses in percent with
the standard angular losses for the type of solar panel used, it
can be said that the values obtained by us are relatively close
to the theoretically obtained ones, from the experiments of YS
Khoo, JP Singh, TM Walsh and AG Aberle [9]. They are
obtained by the following formula:
From the measurements and the statistical processing done
so far, it can be said that the angular losses in the Li-Fi system
up to 60 degrees are relatively small, compared to the losses
from changing the distance and those from closing the
visibility. This provides opportunities for transmission and
reception of audio signals with a large "angle of view", which
could be used for various applications in the field of smart
home systems, mixing audio signals, studying the influence of
various factors of the transmission environment.
As it can be seen in Fig. 6, the theoretically obtained graph
and that of the experiment are relatively close to 75 degrees.
Theoretical data and previous research have shown that after
80 degrees, the angular losses become too large and in
practice the signal strength decreases to a maximum.
Fig. 6. Graphic representation of the experimental signal level losses
vs theoretical angle losses
95
After the experiments, statistical processing and
comparison with the theoretical data, it can be seen that the
experimental data are close enough to the theoretical ones and
it can be said that the angular losses in the Li-Fi system will
have an effect after 80 degrees.
Studies on the subject show that a high level of signal
reception, below 5% loss has up to 30 degrees, and in both
directions. This makes a 60 degree deviation from the central
position [10]. Compared to other methods for receiving
signals through free space optics, this method has a relatively
high tolerance, which allows for a high signal level with easy
adjustment.
Given the distance between the LED and the photodetector
(solar panel) and the rotation left and right, this makes a 160
degree angle of "visibility" (d), i.e. the angle from which the
signal can be received. This is a fairly large angle, which
allows us to easily set up the device, transmit signals from a
relatively long distance and the system's resilience to changes
in the environment from external influencing factors [11].
Fig. 7 shows an exemplary diagram of the angle of "visibility"
(d) of a Li-Fi system.
Fig. 7. Block scheme of the angle of “visibility” of Li-Fi system
IV. CONCLUSION
As it can be seen from the experimental part and the results
obtained, the angular losses of up to 60 degrees do not have a
particularly strong effect on the signal strength, as the
movement around these 60 degrees leads to a number of
advantages of the system.
Some of these advantages are: easier setup to receive a
signal from a long distance, the ability to mix different signals
and the stability of the system when changing the angle of any
of the elements (LED or photodetector).
Through this experiment we showed some of the
capabilities of the Li-Fi system for transmission of audio
signals by optical methods, and this task could be assigned to
students in laboratory classes in the discipline
"Optoelectronics and Optical Communications", taught to
bachelors, master’s-additional and master's degree in
equalization training at the Faculty of Telecommunications of
the Technical University in Sofia.
There are opportunities for further experiments on the
subject by adding two or more sources or several separate
receivers, using lenses and other optical instruments and
methods.
 ACKNOWLEDGEMENT
[5]
We would like to express our sincere gratitude to the
Faculty of Telecommunications at Technical University of
Sofia for providing the laboratory base.
The scientific research and the results obtained are part of a
national program "Young scientists and postdoctoral [6]
students", adopted by РМС 577/17.08.2018, for research
project to help young scientists.
[7]
REFERENCES [8]
[1] Z. Wang, Q. Wang, W. Huang, and Z. Xu, “Introduction to [9]
Visible Light Communications”, in Visible Light
Communications: Modulation and Signal Processing, 2018, pp.
1-16, doi: 10.1002/9781119331865.ch1
[2] A. G. Bell, W. G. Adams, Tyndall, and W. H. Preece,
“Discussion on the Photophone and the Conversion of Radiant
Energy into Sound”, J. Soc. Telegraph Eng., vol. 9, no. 34, pp. [10]
375-383, 1880.
[3] F. R. Gfeller, U. Bapst, “Wireless in-House Data
Communication via Diffuse Infrared Radiation”, Proc. IEEE,
vol. 67, no. 11, pp. 1474-1486, November 1979.
[4] A. Karipzhanova, K. Sagindykov, and K. Dimitrov, [11]
“Justirication of the Method and Algorithm of Multidimensional
Parity Control in Distributed Databases of Information
Systems”, 2019 X National Conference with International
96
Participation (ELECTRONICA), Sofia, Bulgaria, 2019, pp. 1-4,
doi: 10.1109/ELECTRONICA.2019.8825600
S. Kulkarni, A. Darekar, and P. Joshi, “A Survey on Li-Fi
Technology”, 2016 International Conference on Wireless
Communications, Signal Processing and Networking
(WiSPNET), Chennai, 2016, pp. 1624-1625, doi:
10.1109/WiSPNET.2016.7566413
C. Gavrincea , J. Baranda, and P. Henarejos, “Rapid Prototyping
of Standard-Compliant Visible Light Communicatios”, IEEE
Journal & Magazines, vol. 52, pp. 80-85, July 2014.
L. Johnson, Fiber Optic Active Devices - LEDs, Lasers, and
Detectors, Course, IEEE, 2013.
Microsoft Excel 2007 for Scientists and Engineers, Academic
Press, 2008.
Y. S. Khoo, J. P. Singh, T. M. Walsh, and A. G. Aberle,
“Comparison of Angular Reflectance Losses Between PV
Modules with Planar and Textured Glass Under Singapore
Outdoor Conditions”, IEEE Journal of Photovoltaics, vol. 4, no.
1, pp. 362-367, January 2014, doi:
10.1109/JPHOTOV.2013.2284544
E. V. Platonova, A. S. Toropov, and A. N. Tulikov, “Simulation
of Energy Input to Solar Panels”, 2019 International Ural
Conference on Electrical Power Engineering (UralCon),
Chelyabinsk, Russia, 2019, pp. 133-137, doi:
10.1109/URALCON.2019.8877633
R. B. Nazmul, “Calculating Optimum Angle for Solar Panels of
Dhaka, Bangladesh for Capturing Maximum Irradiation”, 2017
IEEE International WIE Conference on Electrical and Computer
Engineering (WIECON-ECE), Dehradun, 2017, pp. 25-28, doi:
10.1109/WIECON-ECE.2017.8468880
ENERGY SYSTEMS AND EFFICIENCY I
 Application of Neural Networks for Short Term Load
Forecasting in Power System of North Macedonia
Goran Veljanovski, Metodija Atanasovski, Mitko Kostov, and Pande Popovski
Abstract – Air temperature, type of the day and humidity are
factors that have significant impact on electricity consumption
and power system load on a short term. In the paper, neural
networks are successfully applied for load forecasting on a short
term in a power system of Republic of North Macedonia. The
training of neural network is performed on real input data. The
forecast is obtained on the basis on the trained network and
learned relationship through training process.
Keywords – Neural networks, Load forecast, Power systems.
I. INTRODUCTION
Short term load forecast is very essential for power system
operation in secure and reliable manner. The significance of
electricity demand forecast relies on several reasons:
increased penetration of non–dispatchable, renewable sources,
electricity markets structural changes, optimal unit
commitment and operation of conventional power plants.
Electricity demand forecast depends on several factors:
climate parameters (temperature, humidity, wind speed etc.),
type of the day (working day, Saturday, Sunday, holiday),
demography data, gross domestic product, and economic
standard of the people, applied measures for energy
efficiency, geography and other factors.
Climate and calendar parameters such as temperature, day
type, humidity, and wind speed are one of the factors, which
has significant impact on electricity consumption and power
system load on a short term. Air temperature impact on
system load is especially important from the power system
operational management aspects on short run. This fact is
evident in power system of Republic of North Macedonia due
to high variations of consumption and load in year seasons.
The number of scientific papers on load forecasting has
increased from about 100 before 20 years to more than 1,000
in the last years. Therefore, it is very difficult to make full
overview of the existing methodology, and only some papers
are mentioned here. There is a number of techniques, which
have been used for load forecast, such as: single or multiple
linear or nonlinear regressions, stochastic time series,
exponential smoothing, state space and Kalman filter,
knowledge based approach, neural networks, wavelet
transformations, semi parametric additive model, fuzzy logic
etc.
Goran Veljanovski, Metodija Atanasovski, Mitko Kostov and
Pande Popovski are with University St. Kliment Ohridski - Bitola,
Faculty of Technical Sciences, Makedonska falanga 37, 7000 Bitola,
North Macedonia, E-mail: goran.veljanovski @uklo.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
The first generation methods called analytical methods
includes time series analysis, regression methods [1-3],
similar day method, Wavelet Transform [4-5], etc. Artificial
intelligence methods, also known as the second generation of
the load forecast methods, mainly comprises artificial neural
networks (ANN) [6-8], including deep neural networks [9-10],
random forests, gradient boosting [11], fuzzy logic [12].
Compared with the second generation, the second generation
has become more important due to reduced errors. Some
combination of methods (known as hybrid methods) that
belong to the both generations is also possible [13-14].
Authors in [15] have concluded that all the previous methods
appear to have at least one of the following three limitations:
 Methods work only with a subset of days (electricity
demand forecast is done only for a one day type, e.g.,
working days);
 Narrow time frame is considered in calculations (e.g.,
a few months).
This paper investigates the application of neural networks
for short term load forecasting in power system of North
Macedonia. In the paper, neural networks are successfully
applied load forecasting on a short term in a power system of
Republic of North Macedonia. The training of neural network
is performed on real input data. The forecast is obtained on
the basis on the trained network and learned relationship
through training process. The implementation of the neural
networks is performed by using two independent variables: air
temperature and type of the day. It means the algorithm
calculates suitable power load values around a certain
temperature for a specific type of the day. The hourly data
(8760 per year) for the temperatures and load are used for the
years 2014-2018 as a training dataset, while 2019
(temperatures and load) data are used as a test dataset.
The proposed methodology tries to overcome some of the
previous methods limitations mentioned, above. Namely,
methodology works with all types of days and calculations are
done for a large time frame. Other contributions of the paper
are the use of last seven days for day ahead load forecast and
the fact that for the first time neural networks are applied for
load forecast in power system of North Macedonia using real
input data.
II. METHODS
A. Input Data and Basic Analysis
Historical data on load and temperature was collected for
this study. The case study input data includes hourly power
system load of the Republic of North Macedonia in the years
99
 Load (MWh/h)

Fig. 1. Load duration curve for years 2014−2018 for power system of Fig. 2 Load forecast for the year 2019
Republic of North Macedonia

2014-2019 (8760 per year) [18, 19], as well as corresponding C. Neural Network Training
meteorological information on minimum, average and
maximum temperatures collected from [20] (Fig. 1). After the network has been designed, the next step is to
Figure 2 depicts a daily diagram of power system of Republic train the network. Training an ANN is an iterative process.
of N. Macedonia for the day Jan 08, 2014. Three typical points For training the proposed neural network, Levenberg
(minimum daily load Pmin=836MW, average daily load Marquardt technique is used.
Pavg=1100MW, maximum daily load Pmax=1293MW) are For training the network, the proposed algorithm uses 70%
marked on the diagram and they are used in the analysis from of the average day loads data and temperatures, while 15% are
each daily diagram. An average daily load is an average of all used for the network validation and 15% for testing the
24-hourly loads on a daily diagram. The average daily loads in network. The division of data in the proposed algorithm is
correlation with average daily temperatures are used in this done randomly.
paper for analysis and forecast. In [21, 22] was shown that there
is a strong negative correlation between the power load and the III. RESULTS AND DISCUSSION
air temperature. The determination coefficients for polynomial
regression and sinuses regression of the maximum, average and
After proper training of the network, its generalization
minimum daily load due to the average daily temperature are
performance needs to be validated. New data that is not used
very high, which means that the regression analysis shows high
during training will be fed into the network to see if it can
prediction degree of the daily typical loads from the air
predict these "unseen" data well.
temperature [21-23].
The results obtained from testing the trained ANN on new
data for 365 days of the year 2019 are presented in Fig. 2. The
B. Neural Network Structure and Design graph depicts a plot of ‘actual’ and ‘forecast’ load in MW
values for the day of the year. In addition, for evaluation of
In order to achieve the objective of forecasting electrical the numerical predictions success, several alternative
load, first, multilayer artificial neural networks (ANN) are quantities have been calculated and shown in Table I. The
constructed. The theory of ANN is well known in the quantities used to evaluate the model performance are the
literature [1, 6, 7, 8 and 24]. Multilayer feed forward ANN are following:
used in this research for prediction. The basic element of a Mean-squared error (MSE):
feed forward neural network is the neuron, which is a logical-
mathematical model that seeks to simulate the behavior and
MSE =
( p1 − a1 )2 + ⋅ ⋅ ⋅ + ( pn − an )2 . (1)
functions of a biological neuron [1]. n
The number of input neurons is usually set equal to the
number of input variables. Input variables for this research are Mean-absolute error (MAE):
the following: Day of the week; Temperatures (average
p1 − a1 + ⋅ ⋅ ⋅ + pn − an . (2)
temperature) of the seven previous days; Average loads of the MAE =
seven previous days. For the days of the week, numbers from n
1 to 7 are used and for holidays according to calendar number
8 is used.
Root mean-squared error (RMSE):
This results in 15 neural network input values. The output
of the neural network represents the forecasted load data for
the forecasting day. The model is consisted of one hidden RMSE =
( p1 − a1 )2 + ⋅ ⋅ ⋅ + ( pn − an )2 (3)
layer with 10 hidden neurons per layer. n

100
 TABLE I
MEASURES FOR MODEL EVALUATION
MAE(%) MAE (MW) MSE (MW2) RMSE (MW)
3.04 30.4 1397 37.38
In Eqs. (1)-(3), n represents the total number of instances,
p1,.., pn are the predicted response values, a1,.., an are the
actual response values.
It is evident that there is a high level of accuracy of ANN to
forecast electricity demand.
IV. CONCLUSION
The presented results confirm the applicability and
efficiency of ANN in short-term electricity demand
forecasting. Average daily loads and temperatures are used as
historical data. The forecasted average load for the day ahead
with high level of accuracy, gives the possibility to the system
operator to estimate the necessary amount of energy day
ahead. The methodology presented in the paper uses the last
seven days for day ahead load forecast and for the first time
ANN are applied for load forecast in power system of North
Macedonia using real input data. For future work, experiments
should be extended on real input data for power systems in the
region of South East Europe.
ACKNOWLEDGEMENT
This research is supported by the EU H2020 project
TRINITY (Grant Agreement no. 863874) This paper reflects
only the author’s views and neither the Agency nor the
Commission are responsible for any use that may be made of
the information contained therein.
REFERENCES
[1] W. Charytoniuk, M. S. Chen, and P. Van Olinda,
“Nonparametric Regression Based Short-Term Load
Forecasting”, IEEE Trans. Power Syst., vol. 13, no. 3, pp. 725-
730, Aug. 1998.
[2] S. Rucic, A. Vuckovic, and N. Nikolic, “Weather Sensitive
Method for Short Term Load Forecasting in Electric Power
Utility of Serbia”, IEEE Trans. Power Syst., vol. 18, no. 4, pp.
1581-1586, Nov. 2003.
[3] T. Hong, M. Gui, M. E. Baran, and H. L.Willis, “Modeling and
Forecasting Hourly Electric Load by Multiple Linear
Regression with Interactions”, in Proc. IEEE PES Gen. Meeting,
Jul. 2010, pp. 1-8.
[4] Y. Chen, P. B. Luh, C. Guan, Y. Zhao, L. D. Michel, M. A.
Coolbeth, P. B. Friedland, and S. J. Rourke, “Short-Term Load
Forecasting: Similar Day-Based Wavelet Neural Networks”,
IEEE Trans. Power Syst., vol. 25, no. 1, pp. 322-330, Feb. 2010.
[5] G. Sudheer and A. Suseelatha, “Short Term Load Forecasting
Using Wavelet Transform Combined with Holt_Winters and
Weighted Nearest Neighbor Models”, Int. J. Electr. Power
Energy Syst., vol. 64, pp. 340-346, 2015.
101
[6] W. Charytoniuk and M.-S. Chen, “Very Short-Term Load
Forecasting Using Artificial Neural Networks”, IEEE Trans.
Power Syst., vol. 15, no. 1, pp. 263-268, Feb. 2000.
[7] H. S. Hippert, C. E. Pedreira, and R. C. Souza, “Neural
Networks for Short-Term Load Forecasting: A Review and
Evaluation”, IEEE Trans. Power Syst., vol. 16, no. 1, pp. 44-55,
Feb. 2001.
[8] P. Mandal, T. Senjyu, N. Urasaki, and T. Funabashi, “A Neural
Network Based Several-Hour-Ahead Electric Load Forecasting
Using Similar Days Approach”, Int. J. Electr. Power Energy
Syst., vol. 28, no. 6, pp. 367-373, 2006.
[9] H. Shi, M. Xu, and R. Li, “Deep Learning for Household Load
Forecasting - A Novel Pooling Deep RNN”, IEEE Trans. Smart
Grid, vol. 9, no. 5, pp. 5271-5280, Sep. 2018.
[10] M. Cai, M. Pipattanasomporn, and S. Rahman, “Day-Ahead
Building-Level Load Forecasts Using Deep Learning vs.
Traditional Time-Series Techniques”, Appl. Energy, vol. 236,
pp. 1078-1088, 2019.
[11] S. Papadopoulos and I. Karakatsanis, “Short-term electricity
load forecasting using time series and ensemble learning
methods”, in Proc. IEEE Power Energy Conf. Illinois (PECI),
Feb. 2015, pp. 1-6.
[12] H. H. Çevik and M. Çunka, “Short-Term Load Forecasting
Using Fuzzy Logic and ANFIS”, Neural Comput. Appl., vol.
26, no. 6, pp. 1355-1367, 2015.
[13] P. Lusis, K. R. Khalilpour, L. Andrew, and A. Liebman,
“Short-Term Residential Load Forecasting: Impact of Calendar
Effects and Forecast Granularity”, Appl. Energy, vol. 205, pp.
654-669, 2017.
[14] M. Ghayekhloo, M. B. Menhaj, and M. Ghofrani, “A Hybrid
Short-Term Load Forecasting with a New Data Preprocessing
Framework”, Electr. Power Syst. Res., vol. 119, pp. 138-148,
2015.
[15] M. Tucci, E. Crisostomi, G. Giunta, and M. Raugi, “A Multi-
Objective Method for Short-Term Load Forecasting in
European Countries”, IEEE Trans. Power Syst., vol. 31, no. 5,
pp. 3537-3547, Sep. 2016.
[16] E. A. Feinberg, D. Genethliou, Chapter 12 Load forecasting,
Applied Mathematics for Power Systems, pp.269-282.
[17] I. H. Witten, E. Frank, M. A. Hall, C. J. Pal, Data Mining,
Practical Machine Learning Tools and Techniques, Burlington,
USA, Morgan Kaufmann, 2017.
[18] Dispatching and Operational Reports for Power System Load
Data obtained from MEPSO.
[19] D. Bajs, M. Atanasovski, Longterm Forecast Study of Electrical
Energy and Power Balance and Adequacy Analysis of
Transmission Network of Republic of Macedonia,
Zagreb/Skopje EIHP, 2016.
[20] https://www.wunderground.com/history.
[21] M. Atanasovski, M. Kostov, B. Arapinoski, I. Andreevski,
“Correlation between Power System Load and Air Temperature
in Republic of Macedonia”, Int. Scientific Conf. on Information,
Communication and Energy Systems and Technologies, pp.
213-216, Sozopol, Bulgaria, Jun. 2018.
[22] M. Kostov, M. Atanasovski, G. Janevska, B. Arapinoski,
“Power System Load Forecasting by using Sinuses
Approximation and Wavelet Transform”, Int. Scientific Conf.
on Information, Communication and Energy Systems and
Technologies, pp. 273-276, Ohrid, North Macedonia, Jun. 2019.
[23] S. Vukadinovic, Elements of Probability Theory and
Mathematical Statistics, Belgrade, 1973.
[24] H. B. Demuth, M. H. Beale, O. De Jess, and M. T. Hagan,
Neural Network Design, 2014.
 Fuzzy Multiple Linear Regression for the Short Term
Load Forecast
Andrija Petrušić1 and Aleksandar Janjić2
Abstract – In predicting the hourly demand for energy trading,
the weather forecast error makes the usage of the exact values of
certain predictor variables (such as temperature) very difficult.
Therefore, the fuzzy input variables are often used. This paper
presents a new way of using the fuzzy inputs in linear regression
in forecasting hourly consumption for the day-ahead period.
Unlike the present approach to the fuzzy linear regression, where
the regression coefficients have been fuzzified, in this model the
degrees of membership to the fuzzy sets are used as additional
predictor variables
Keywords – Day-ahead market, Fuzzy regression, Multiple
linear regression, Short term load forecast.
I. INTRODUCTION
For successful decision making in the energy sector, short
term load prediction plays a crucial role in ensuring profitable
and reliable power system planning and operations. The more
accurate the forecast, the more efficient are the operation and
management of an electric utility. These forecasts influence
all stakeholders included in the process of energy supply,
generation, distribution, markets, etc. [1].
According to some authors [2-3], load forecasting can be
divided in 4 categories based on the considered timespan: very
short term (less than a day), short term (one day to two
weeks), medium term (two weeks to three years) and long
term (three years to thirty years or more); whereas according
to others, there are three categories – short term forecasts
which are usually from one hour to one week, medium
forecasts which are usually from a week to a year, and long
term forecasts which are longer than a year [1], [4-5].
These forecasts are equally important for different
operations and can be widely used by all sectors within a
company, e.g. short term load forecasting can be applied in
operations, maintenance, and demand response and energy
market activities [6-7]. Accurate load forecasting helps the
operator to make decisions about unit commitment, decreases
spinning reserve capacity and properly design a generator
maintenance plan. Besides reducing production costs, accurate
load forecast is also essential for the reliability of power
systems. The system operators use the result of load
1Andrija Petrušić is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
andrija.petrusic@elfak.rs
2Aleksandar Janjić is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
aleksandar.janjic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
prediction as the basis of offline network analysis to
determine the vulnerability of the system. If this is the case,
corrective action like the additional energy purchase or load
shedding should be prepared. For the energy trading, the day
ahead load forecast is of crucial importance, because the
inaccurate forecast may cause great financial losses. Some
corrections may be effectuated in the intra-day market, when
the short-term hourly forecast is required.
This paper presents a new way of using the fuzzy inputs in
linear regression in forecasting hourly consumption. Unlike
the present approach to the fuzzy linear regression, where the
regression coefficients are fuzzified, in this model the degrees
of membership to the fuzzy sets are used as additional
predictor variables. This approach allows the input values to
be predefined fuzzy numbers and the output size to remain
crisp, which allows the use of standard statistical tests on
forecast accuracy. It is also possible to use developed
computational methods for linear regression. The method was
tested on real data of a wider consumer area for the period
2014 to 2018, using temperature and period during the month
as fuzzy variables
II. FUZZY ENERGY FORECAST PREDICTORS
A. Fuzzy Variables
In order to determine how and to what extent predictor
variables shape the power supply curve it is necessary to
select the appropriate variables and their types for the model
attributes. Among the meteorological variables, temperature
has the greatest influence on electricity consumption and is
closely correlated with it, so it is often used in electricity
consumption forecasting models. The short term load forecast
is characterized by the large volume and high quality of time
series data. Because of advanced metering technology today,
many utilities are comfortable with the data quality at the
corporate level. However, some data may be still missing or
corrupted.
Furthermore, it is considered that the temperature forecast
is relatively accurate only up to the day ahead for the purposes
of forecasting electricity consumption and this is the main
reason why there are more and more methods that avoid the
use of temperature or any other meteorological attribute.
Another reason is that the correlation between temperature
and electrical consumption varies from data set to data set,
even from year to year, and thus the inclusion of temperature
may have a poor effect on model accuracy [8]. Finally, the
forecast model must be adapted to thermal inertia of the
system.
102

Fig. 1. Fuzzy values of ambient temperatures
Due to these reasons, the temperature should not be treated
as crisp, but rather fuzzy variable. An example of temperature
modelled with a group of triangular fuzzy sets (“around 15”,
“around 20” etc.) is presented in Fig. 1.
Linear models are generally more robust in higher
dimensional space, especially when there is certain amount of
noise in measurements [9].
A. Types of Fuzzy Predictor Variables
The forecast problem can be treated as a problem of one
dependent and multiple independent variables and it is a
suitable situation for analyzing data by the multiple regression
method to help predict future values from past values.
Although the training model may use the fuzzy predictor
variables, the future value of the independent variable may
differ from the predefined set of training variables.
• Case A
In the case when temperature values from the past are
missing or corrupted, they must be estimated with appropriate
fuzzy sets. Any recorded temperature can be transformed to
the N - length vector of membership degrees corresponding to
N temperature fuzzy sets:
FT { x} =  μ1 ( x ) , μ2 ( x ) ,, μ N ( x )  .
The example of transformation with two fuzzy sets A and B
is given in Fig. 2, where the value x1 is transformed to the
vector [μA(x1), μB(x1)].
Fig. 2. Crisp future value
• Case B
If the input predictor variable is a fuzzy set (for example,
hourly temperatures are recorded as triangular fuzzy numbers
‫ܣ‬ሚ, ‫ܤ‬෨ and ‫ܥ‬ሚ ) in Fig. 3, they can be treated as a categorical
variable, also known as factor or qualitative variables that
classify observations into groups. They have a limited number
of different values. In these steps, the categorical variables are
recorded into a set of separate binary variables, and the
problem is solved as standard forecast of the output value
depending on the presence of the fuzzy set (for example the
number ‫ܥ‬ሚ in Fig. 3).
Fig. 3. Same shape of fuzzy variables in the past and future
• Case C
Let the regression model be formed on the basis of the
fuzzy predictor variables A and B. In some cases, the input
values may have a different shape from the predictor variables
on the basis of which the regression model was formed. Let Ai
be a symmetric fuzzy set denoted by (αi, ci)L, with αi as the
center and ci as the spread, with reference function
L ( x ) = max ( 0,1 − x ) and the membership function
determined as in:
μ A ( ai ) = L ( ( ai − α i ) ci ) . (2)
i
Using (1) on intersection points every other fuzzy set (the
fuzzy set C on Fig. 4) is uniquely determined by the
(1) membership degrees to sets A and B.
In traditional fuzzy regression, the deviations between the
observed values and the estimated values are regarded as the
fuzziness of the parameters of the system. The main idea is to
define the possibilistic linear function:
Y = A1 x1 +  + Ap x p . (3)
Fig. 4. Different shape of fuzzy variables in the past and future
103
 More detailed discussions of fuzzy regression can be found
in [10] and [11] where identification of the parameters of the
fuzzy linear regression model is formulated as a linear
programming problem. If the relationship between them is
linear, the case transfers to a multiple linear model. Let Y be a
dependent variable and x1 , x2 ,, x p independent variables –
then a linear model is represented by:
Y = β0 + β1 x1  + β p x p + ε ,
where β0 , β1 ,, β p are unknown coefficient that have to be
estimated, while ε represents the measurement error, or
residual. Let the original predictor variable xn be divided in m
fuzzy numbers. According to (1) it is also possible to form a
regression model that considers these degrees of membership
as predictor variables:
Y = β0 + β1 x1 +  + βn μ1 ( xn ) + β n μ2 ( xn ) +
1 2
+  + β n μm ( xn ) +  + β p x p + ε
m
For the sake of simplicity, a new set of variables Z will be
introduced, consisting of both crisp and segmented fuzzy
predictors, with (p+m-1) dimension:
Z =  x1 , x2 ,, μ1 ( xn ) , μ2 ( x2 ) ,, μm ( xn ) ,, x p  (6)
Since the enlarged set of parameters βi is unknown, it should
be estimated from observations of y and z. Let
bi ( i = 0,1, 2,, p + m − 1) be the estimates in terms of
βi ( i = 0,1, 2,, p + m − 1) . Hence the predicted value of y is:
Yˆ = b0 + b1 z1 +  + bp + m −1 z p + m−1 .
Learning and forecasting procedures for the proposed fuzzy
regression model are as follows.
Step 1) Determine the learning range
The regression model is trained by using three types of
variables (Table I).
Time predictor variables (hour of the day, day in the week
and month) are categorical, while the temperature is presented
by 13 triangular fuzzy numbers, starting from “around minus
20 degrees” to “around 40 degrees” with the spread of 10
degrees.
Step 2) Fuzzification of input data
Depending on types of input data for the temperature
variable, according to Figs. 2 and 3, the extended set Z of
predictor variables is formed.
Step 3) Parameter calculation
The parameter set B is determined, using QR decomposition
as the main fitting algorithm. For robust fitting, M-estimation
can be used to formulate estimating equations and solve them
using the method of iterative reweighted least squares (IRLS).
Step 4) Day-ahead energy forecast
Depending on types of input data for the temperature
variable, the hourly temperature forecast for the next day has
been fuzzified. Other data for the 24 hours period are loaded
and the forecasted energy demand is obtained.
Step 5) Prediction error
Root Mean Square Error (RMSE) as the standard deviation
of the residuals and Mean Absolute Percentage Error (MAPE)
has been calculated.
(4)
TABLE I
TYPE OF PREDICTOR VARIABLE
Predictor Type
Hour categorical
Day categorical
Month categorical
Temperature fuzzy
(5)
Period fuzzy
Load of the same hour of the numerical
previous day
III. CASE STUDY
The fuzzy multiple linear regression was tested on the
actual load power data for the city of Belgrade for the year
2018. The model was trained with data sets of total
consumption for 2014, 2015, 2016 and 2017.
The model is trained with the predictor variables presented
in Table I. The temperature is presented with 13 fuzzy sets,
(7)
starting from “around -20” to “around 40” degrees. The model
is compared with the classical multiple linear regression
model with crisp values of temperature and ordinal day in the
month. Data series for temperature contains only average
hourly values for the whole city. The model is trained with the
MATLAB fitlm function for the total of 35000 records. The
model is tested on actual data for 2018 using the forecasted
temperature day ahead values. Load demand curves for typical
winter (February 5), and summer (August 23) days are given
in Figs. 5 and 6, respectively.
It can be noticed that in winter the electrical load is about
twice as high with marked variations and trend changes from
day to day. The summer is characterized by a more stable
load, with less variation and almost no trend. The change in
trend during the week is reflected in increased loads on the
days after the weekend (Monday and Tuesday), when people
are somewhat more active than in the second half of the week.
Weekend days are characterized by reduced consumption, a
direct consequence of non-working days when people are less
active and industrial production stands still.
It is visible that the fuzzy regression model gives better
results than the classical model. The model is more effective
in extreme weather conditions when predicted values in
winter months show the less deviation from the actual ones.
The comparison of RMSE and MAPE values is given in
Table II.
104
 based on extended set of predictor variables has shown
1300000 better results than classical method. Furthermore, the
1200000 methodology allows to perform statistical tests with already
1100000 reliable and proven methods, for which there is extensive
1000000 computer support. Further research will be focused on
900000 different fuzzy sets modelling
800000
700000 REFERENCES
600000
1 3 5 7 9 11 13 15 17 19 21 23 [1] E. A Feinberg, D. Genethliou, “Load Forecasting”, Applied
Mathematics for Restructured Electric Power Systems, Springer,
TRUE Fuzzy Crisp Boston, MA, 2005, pp. 269-282.
[2] T. Hong, P. Wang, “Fuzzy Interaction Regression for Short
Fig. 5. Predicted and actual values for a winter day Term Load Forecasting”, Fuzzy Optimization and Decision
Making, vol. 13, no. 1, pp. 91-103, 2014.
1100000 [3] A. D. Papalexopoulos, T. Hasterberg, “A Regression based
1000000 Approach to Short Term System Load Forecast”, IEEE Trans.
900000 on Power Systems, vol. 5, no. 4, pp. 1535-1544, 1990.
[4] A. M. Al-Kandari, S. A. Soliman, and M. E. El-Hawary, “Fuzzy
800000
Short-term Electric Load Forecasting”, International Journal of
700000 Electrical Power & Energy Systems, vol. 26, no. 2, pp. 111-122,
600000 2004.
500000 [5] D. Srinivasan, M. A. Lee, “Survey of Hybrid Fuzzy Neural
400000 Approaches to Electric Load Forecasting”, IEEE International
Conference on Systems, Man and Cybernetics, Intelligent
1 3 5 7 9 11 13 15 17 19 21 23
Systems for the 21st Century, Vancouver, BC, 1985, pp. 4004-
TRUE Fuzzy Crisp 4008.
[6] T. Hong, “Short Term Electric Load Forecasting”, Ph.D.
Fig. 6. Predicted and actual values for summer day dissertation, North Carolina State University, Raleigh, USA,
2010.
[7] T. Hong, P. Wang, “On the Impact of Demand Response: Load
TABLE II
COMPARISON OF ERRORS Shedding, Energy Conservation, and Further Implications to
Load Forecasting”, IEEE Power & Energy Society General
RMSE (kWh) MAPE (%) Meeting, 2012, pp. 1-3.
[8] B.-J. Chen, M.-W. Chang, “Load Forecasting using Support
Crisp 42585.7 3.65 Vector Machines: A Study on EUNITE Competition 2001”,
Fuzzy 36915.6 3.26 IEEE Transactions on Power Systems, vol. 19, no. 4, pp. 1821-
1830, 2004.
[9] H. Tanaka, I. Hayashi, and J. Watada, “Possibilistic Linear
Regression Analysis for Fuzzy Data”, European Journal of
IV. CONCLUSION Operational Research, vol. 40, no. 3, pp. 389-396, 1989.
[10] T. Hong, P. Wang, “Fuzzy Interaction Regression for Short
Although the temperature is a desirable attribute in the Term Load Forecasting”, Fuzzy Optimization and Decision
formation of a model for forecasting electrical Making, vol. 13, no. 1, pp. 91-103, 2014.
consumption, the fact that cannot be ignored is that the [11] G. James, D. Witten, T. Hastie, and R. Tibshirani, An
Introduction to Statistical Learning with Applications in R,
precise temperature value for the day ahead is unknown so Springer Texts in Statistics, Springer Science + Business Media,
that the model can only use the forecasted temperature New York, 2015.
value. In this paper, the model with the fuzzy input values
and crisp output variable is developed. The methodology

105
 Power Load Forecast for North Macedonia Using
Machine Learning
Pande Popovski, Mitko Kostov, Metodija Atanasovski and Goran Veljanovski
Abstract – Estimating power load has always been of a great
importance for making power purchase agreements on the
electricity market or maintaining power system stability and
reliability. This paper analyses power load forecasts for North
Macedonia when different combinations of independent
variables are used – air temperature, humidity and wind. Several
types of machine learning models were trained in order to
evaluate their performances and impact on the power load
forecast.
Keywords – Machine learning, Regression, Power load
forecast, Forecast accuracy, Model performance.
I. INTRODUCTION
Estimating power load has always been a practice of key
importance whether it comes to making power purchase
agreements on the electricity market, or maintaining power
system stability and reliability. With the increasing trend of
renewable energy sources over the decades, new questions
arise, one of them being inertia in the power systems. Having
prior knowledge on electrical energy demand by means of
power load forecasting can help to deal with the negative
effects that come with the decrease of power system inertia,
by minimizing the mismatch that can appear between supply
and demand. This can be done through better production
planning, which is greatly determined by the estimated
increase in power demand. Furthermore, power load forecast
can be used when it comes to minimizing power loss and
coping with growing power demand: forecasting would
determine which power units should increase their production
and which generators should be dispatched [1]. Thereby the
accuracy and reliability of an optimal calculated generation
curve is determined by the reliability of the power load data
which is used.
There is a number of factors that play an essential role
when conducting power load forecasting - air temperature,
day of the week, holidays, geographical differences,
demographic information, people's standards of living, etc.
[1].
A very close correlation between air temperature and power
system load is studied for the Republic of North Macedonia,
along with forecast accuracy using sinuses and polynomial
regressions and discrete wavelet transformation [2]. In this
paper, an analysis is made on the effect of air temperature,
Pande Popovski, Mitko Kostov, Metodija Atanasovski and Goran
Veljanovski are with the Faculty of Technical Sciences-Bitola, St.
Kliment Ohridski University, Republic of N. Macedonia, E-mail:
pande.popovski@uklo.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
humidity and wind speed on the power load forecast accuracy:
how a few machine learning models perform when different
combinations of meteorological parameters are taken into
consideration. Therefore, a contribution of the paper is that
several types of machine learning models for the first time
were trained on real data of North Macedonia power system in
order to evaluate and compare their performances.
The paper is organized as follows. After the introduction, a
brief explanation on the used machine learning algorithms is
given. The third section shows the numerical results obtained
from the trained models, while the fourth section concludes
the paper.
II. ALGORITHMS
A. Decision Trees
This type of algorithms obtains its name from the decision
tree (DT) diagram. Decision tree diagram is used to
graphically present a process of making a decision. It shows
the factors that play a role in the decision making, and
conditions that have to be fulfilled, as well as all the possible
solutions. DT algorithms apply the same concept to machine
learning, however a different approach than simple decision
making is used in order to construct a model that behaves like
a DT. These algorithms partition the feature space into a set of
rectangles, and then fit a simple model in each one [3]. The
problem results in constructing the algorithm in a way so it
could decide itself under what criteria to make the partitions.
A number of DT algorithms are common when dealing with
classification and regression problems: ID3 (Iterative
Dichotomiser 3), CART (Classification And Regression
Tree), MARS (Multivariate Adaptive Regression Splines) and
others [3].
There are also a number of methods used to deal with the
drawbacks from using TDs. One of these methods is
Bootstrap Aggregating (commonly called Bagging) [4], more
specifically developed to reduce the negative impact that
variance has on the model performance. In other words, it is
implemented to reduce the sensitivity of a model to different
starting parameters. Bagging method is composed of two
parts: bootstrapping and aggregating [5]. Bootstrapping is a
sampling method, where a sample is chosen out of a set, using
the replacement method. Using sampling, multiple smaller
datasets are formed from the original dataset, and a model is
trained using each one. Then, the predictions made from these
“weaker” models are aggregated to obtain a more accurate
prediction:
106
 1 B
mˆ bag (x ) =  mˆ S (x ) ,
B b =1 b
(1)
C. Gaussian Process Regression
where m̂S is estimator of a model m: XY, for a bootstrap
b

sample Sb, which is obtained from the sample S = {yi, xi} with
i = {1, …, n}.
Decision trees are sensitive to the specific data on which
they are trained. Different subsets of the training data may
result in different resulting decisions, and the predictions may
be completely different. When bagging with decision trees,
one is less concerned about individual trees overfitting the
training data. The only parameter is the number of samples
and consequently the number of trees to include. It can be
selected by continuously increasing the number of trees until
the accuracy starts to stop improving.
B. Support Vector Machines
Support-Vector Machines (SVMs, also called Support-
Vector Networks) are supervised learning models with
associated learning algorithms that analyse data used for
classification and regression analysis [6]. The support-vector
network works in such a way that an input training data
(support vectors) is mapped into a high dimensional feature
space, and in this space a decision surface is constructed with
special properties that ensure high generalization ability of the
network [7]. In other words, the algorithm constructs a
hyperplane, a subspace whose dimension is one less than that
of its ambient space, by using the mapped data from the
training dataset [8]. This hyperplane (or set of hyperplanes)
acts as a sort of a border that divides data points, which is why
SVMs were primarily used for classification problems. Once a
Gaussian processes (GPs) can be thought of as an extension
of multivariate Gaussian distribution. GPs fall in the category
commonly known as Bayesian methods, because of the fact
that they originate from the Bayes' theorem [11].
The Gaussian distribution is very useful in machine
learning because it can conveniently model noise, which is a
very common occurrence in data. Also, they are convenient
for many analytical manipulations. Because of this, Gaussian
distributions are useful for modelling finite collections of real-
valued variables. GPs are the extension of multivariate
Gaussians to infinite-sized collections of real-valued
variables. This extension defines GPs as distributions not just
over random vectors, but also over random functions [11].
Following the assumption that data can be represented as a
sample from a multivariate Gaussian distribution, the
following expression is given [12]:
y  K K*T   , (3)
 y  ~ N  0,  
 *   K* K**  
where tilde (~) denotes "has the probability distribution of", T
indicates matrix transposition. Matrices K, K* and K** contain
the values of a covariance function for all possible
combinations of data points, due to the variance of the
responses differing:
 k (x1 , x1 ) k (x1 , x2 )  k (x1 , xn )
k (x , x ) k (x , x )  k ( x2 , xn ) ,
K = 2 1 2 2

model has been trained, further instances added in the created      

 
space can be classified using the hyperplane. However,  k (xn , x1 ) k (xn , x2 )  k (xn , xn )
modified SVMs can be used for regression analysis as well.
In their most general form, SVMs [9] find a hyperplane in a K * = [k (x* , x1 ) k (x* , x2 )  k ( x* , xn )] ,
space different from that of the input data. It is a hyperplane in
a feature space induced by a kernel K. In other words, because K** = [k (x* , x* )] ,
most of the time a problem does not contain linearly separable
data, it is presented in a higher dimensional space so its data  − ( x − x ' )2 
 
k ( x , x ') = σ e
2
may become linearly separable. This is often called the kernel 2
f
 2 l  ,
trick [10]. This can be seen as a set of functions in a
Reproducing Kernel Hilbert Space (RKHS) [9]. with l and δ - a length parameter and the Kronecker delta
For regression purposes SVMs solve the following function, respectively. The goal is, given n observations y, to
minimization problem: predict y*. The conditional probability p(y*|y) follows a
l Gaussian distribution y:
+ C  y i − f (xi ) ε
2
min f (2)
f K
i =1 (
y * | y ~ N K * K −1 y, K ** − K * K −1 K *T . ) (4)
where C is a "regularization parameter" that controls the
trade-off between empirical error and complexity of the The best estimate for y* is the mean of this distribution:
hypothesis space used. The variables x and y belong to a set of
y * = K * K −1 y , (5)
training data containing l instances. The function f represents
a hyperplane, expressed via the input data x. K represents the and the uncertainty in the estimate is captured in its variance:
kernel that defines the RKHS, and f 2 is the RKHS norm of
var( y* ) = K ** − K * K −1 K *T .
K
the function f. (6)

107
 TABLE I The following three machine learning algorithms generated
PERFORMANCE RESULTS FROM TRAINING DATASET the most accurate models: Ensemble of trees (Bootstrap
Bagged Radial Exponen Aggregation, or Bagged trees), Support Vector Machine
Attributes: Measures (Radial kernel SVM) and Gaussian Process Regression
trees SVM tial GPR
R 0,93 0,91 0,93 (Exponential kernel GPR). The measures used to evaluate
RMSE [MW] 59,29 67,48 59,34 model performance are the following [15]:
Temperature
MSE [MW] 3515,54 4553,07 3521,28
MAE [MW] 45,78 54,69 46,11 Mean-squared error (MSE):
R 0,92 0,91 0,91
Temperature, RMSE [MW] 63,78 66,43 66,21 ( y1* − a1 )2 + ⋅ ⋅ ⋅ + ( y n* − an )2 . (7)
MSE =
Wind MSE [MW] 4068,10 4412,28 4384,16 n
MAE [MW] 50,73 53,37 53,55
R 0,92 0,91 0,91 Mean-absolute error (MAE):
Temperature, RMSE [MW] 63,28 66,72 66,30
y1* − a1 + ⋅ ⋅ ⋅ + yn* − an . (8)
Humidity MSE [MW] 4004,26 4451,58 4396,12 MAE =
MAE [MW] 50,37 53,95 53,82 n
R 0,91 0,91 0,91 Root mean-squared error (RMSE):
Temperature,
RMSE [MW] 66,75 65,67 65,54
Humidity,
MSE [MW] 4456,20 4311,93 4295,65
Wind
MAE [MW] 53,16 52,70 52,72
( y1* − a1 )2 + ⋅ ⋅ ⋅ + ( y n* − an )2 . (9)
RMSE =
n

III. CASE STUDY Correlation coefficient (R):

SYA . (10)
The entirety of the collected data is made up of R=
SY S A
meteorological documentation and power load data. More
specifically the meteorological data consists of the daily where:
averages of the air temperature (expressed in degrees Celsius),
air humidity (expressed in %), and wind speed (expressed in n n n

 ( y − y )(a − a ) (y − y) (a − a )

2 2
mph-miles per hour) for the city of Skopje, North Macedonia i i i* i

throughout the years 2014-2019 [13]. For the same time SYA = i =1 , SY = i =1 , SA = i =1 .
n −1 n −1 n −1
period, the daily averages of the power load in the Republic of
North Macedonia are provided [14].
The data is divided into a training dataset, which consists of In Eqs. (6)-(9), n represents the total number of instances,
the statistical information for the period 2014-2018 (1826 y1*,.., yn* are the predicted response values, a1,.., an are the
instances), and the test dataset with the data for the year 2019 actual response values, and y* and a are their average values.
(365 instances). The performance of the trained models for the
Table I shows the performance results of the trained models
training dataset is evaluated using 10-fold cross-validation.
when the training dataset (years 2014-2018) is tested by 10-
Multiple models were trained using Linear Regression,
fold cross-validation. By observing the acquired numerical
Regression Trees, Support Vector Machines, Gaussian
results from the cross-validation, it can be concluded that the
Process Regression and Ensembles of Trees. Different
performance of the trained models does not change noticeably
combinations of the meteorological factors were used as
by adding humidity and wind speed as attributes alongside air
attributes. In each of the models, the average daily power load
temperature. In fact, slightly worse results are obtained for the
remained as the dependant variable. More specifically, four
models trained with the Bagged trees and Exponential kernel
cases were taken into consideration:
GPR algorithms. A minor improvement in forecast accuracy is
• Attributes: Temperature; Dependant variable: Power load noted for the model trained with Radial kernel SVM.
• Attributes: Temperature and wind; Dependant variable: However, this applies for the training dataset, as different
Power load results might be acquired from the test dataset.
• Attributes: Temperature and humidity; Dependant
variable: Power load
• Attributes: Temperature, humidity and wind; Dependant
variable: Power load.

108
 TABLE II accurate results are achieved when air temperature and
PERFORMANCE RESULTS FROM TEST DATASET humidity are used as independent variables. The models
Bagged Radial Exponen trained with Exponential kernel GPR gain the most accuracy
Attributes Measures from using multiple attributes, while for the models trained
trees SVM t ial GPR
R 0,93 0,95 0,94 with Radial kernel SVM the least improvement is noted.
RMSE [MW] 67,41 56,23 64,49 However, in general the models trained with Radial kernel
Temperature SVM give the most accurate results.
MSE [MW] 4543,66 3162,17 4158,66
MAE [MW] 51,30 41,54 47,80
R 0,94 0,94 0,95 ACKNOWLEDGEMENT
Temperature, RMSE [MW] 59,32 58,97 57,24
Wind MSE [MW] 3518,44 3478,35 3276,04
This research is supported by the EU H2020 project
MAE [MW] 45,14 43,64 41,83
TRINITY (Grant Agreement no. 863874) This paper reflects
R 0,94 0,95 0,95
only the author’s views and neither the Agency nor the
Temperature, RMSE [MW] 58,71 55,36 54,77
humidity
Commission are responsible for any use that may be made of
MSE [MW] 3447,72 3064,59 3000,08
MAE [MW] 43,63 41,44 40,83
the information contained therein.
R 0,94 0,95 0,95
Temperature,
Humidity,
RMSE [MW] 61,69 55,57 55,77 REFERENCES
MSE [MW] 3805,37 3088,48 3109,75
Wind
MAE [MW] 46,84 41,74 41,66 [1] E. A. Feinberg, D. Genethliou, “Chapter 12 Load
Forecasting”, Applied Mathematics for Power Systems,
Performance results obtained for the test dataset (year pp.269-282.
2019) are given in Table II. Results from the test dataset show [2] M. Kostov, M. Atanasovski, G. Janevska, and B.
that when only the air temperature has been chosen as an Arapinovski, "Power System Load Forecasting by using
independent variable, prediction accuracy for the models Sinuses Approximation and Wavelet Transform", ICEST -
Ohrid, North Macedonia, 27-29 June 2019.
trained with Bagged trees and Exponential GPR is worse [3] T. Hastie, R. Tibshirani, and J. Friedman, "The Elements of
compared to the training dataset, whereas there is an Statistical Learning: Data Mining, Inference, and Prediction"
improvement for the model trained with Radial kernel SVM. - Second EDITION, pp. 305-306, Springer, May 2001.
When multiple independent variables are considered, an [4] "Bootstrap Aggregation, Random Forests and Boosted
improvement in forecast accuracy is noted in each model. Trees", https://www.quantstart.com/articles/bootstrap-
Examining the test results further, it can be noticed that aggregation-random-forests-and-boosted-trees.
trained models perform overall better when humidity is taken [5] L. Breiman, ‘‘Bagging Predictors’’, Machine Learning, vol.
as an attribute along with temperature, compared to when 24, no. 2, pp. 123-140, 1996.
wind and temperature are taken as attributes. This leads to the [6] https://en.wikipedia.org/wiki/Support_vector_machine.
[7] C. Cortes, V. Vapnik, “Support-Vector Networks”, Machine
consideration that humidity may be more closely correlated to Learning 20, pp. 273-297, 1995,
power load than wind speed is. When all three attributes are https://doi.org/10.1007/BF00994018.
used to train a model, very slight changes in accuracy can be [8] https://en.wikipedia.org/wiki/Hyperplane.
noticed compared to the case when two attributes are used. [9] T. Evgeniou, M. Pontil, “Support Vector Machines: Theory
and Applications”, chapter from book Machine Learning and
Its Applications: Advanced Lectures, pp 249-257, 2001.
IV. CONCLUSION [10] https://en.wikipedia.org/wiki/Bayes%27_theorem
[11] C. B. Do, H. Lee, "Gaussian Processes", lecture notes,
From several machine learning algorithms used for power November 22, 2008.
load forecast, the following three are presented in the paper: [12] M. Ebden, "Gaussian Processes for Regression: A Quick
Bagged trees, Radial kernel SVM and Exponential kernel Introduction", August 2008.
GPR. Using each of the three algorithms, a number of models [13] https://www.wunderground.com/history.
[14] Dispatching and Operational Reports for Power System Load
are trained where different combinations of meteorological
Data obtained from MEPSO.
factors are used. [15] I. H. Witten, E. Frank, M. A. Hall, C. J. Pal, Data Mining:
From the results it can be concluded that combining the Practical Machine Learning Tools and Techniques,
attributes improves its accuracy. From the four cases, the most Burlington, USA, Morgan Kaufmann, 2017.

109
 K-Nearest Neighbor Regression for Forecasting
Electricity Demand
Metodija Atanasovski, Mitko Kostov, Blagoja Arapinoski, and Mile Spirovski
Abstract – Power system load forecasting plays a vital role in
all aspects of power system planning, operation and control. It is
a basic function for reliable and economical operation of power
systems. This paper analyses the power system load forecast
performed by applying k-nearest neighbour machine learning
model, which is for the first time applied on real data of North
Macedonia power system. The results are compared with
polynomial and sinuses regressions.

Keywords – Power system load, Forecast, Machine learning, K-

nearest neighbour, Correlation, Regression.

I. INTRODUCTION
Electricity load forecasting is a process of projecting future
electric energy demand in order to meet the increasing
demand. Electric load forecasting will control which power
plants should increase their output and which generators
should be dispatched. In terms of the period of load
forecasting, it can be divided into three categories: short-term
load forecasting, medium-term load forecasting and long-term
load forecasting [1].
Short-term load forecasting refers to forecasting electricity
demand several days in advance every hour. An
underestimation of the power load can result in lack of
electricity production. On the other hand, in the case of
overestimation of the load sub-optimal dispatch of power
plants could be scheduled.
The factors that play a key role in forecasting energy
consumption are temperature, type of the day (weekday,
weekend, holiday), geographic differences, people standard,
demographics, etc. The number of scientific articles about
power forecasting is very large. Some of the used techniques
include regression methods [2-4], discrete wavelet transform
[5, 6], neural networks [7-11], fuzzy logic [12]. This paper
studies the forecast of power load consumption by applying k-
nearest neighbour (KNN) machine learning model, which is
for the first time applied on real data of North Macedonia
power system and the results are compared with polynomial
and sinuses regressions. The implementation of the k-nearest
neighbor machine-learning model is performed by using two
independent variables: air temperature and date. It means the
algorithm searches for/calculates suitable power load
candidates around a certain period and temperature. The
hourly data (8760 per year) for the temperatures and load are
Metodija Atanasovski, Mitko Kostov, Blagoja Arapinoski and
Mile Spirovski are with the Faculty of Technical Sciences-Bitola,
University St. Kliment Ohridski, Republic of North Macedonia, E-
mail: metodija.atanasovski@uklo.edu.mk
Fig. 1. Load duration curve for years 2014-2018 for power system of
Republic of N.Macedonia
used for the years 2014-2018 as a training dataset, while 2019
(temperatures and load) data are used as a test dataset. The
effectiveness of the model is evaluated and confirmed by
cross-validation.
The remainder of this paper is structured as follows.
Section II briefly describes the basic statistical analysis of
power system load data and air temperature. The methodology
used in the paper is introduced in Section III. Section IV
elaborates the experimental results and the conclusions are
given in Section V.
II. BASIC ANALYSIS
In this paper, dataset consists of power system load data for
Republic of North Macedonia for the calendar years 2014-2019
on hourly basis (8760 per year) (Fig. 1) [13, 14] and the
matching meteorological data about minimal, average and
maximal air temperatures for the city of Skopje,
North Macedonia [15].
Fig. 2 shows the normalized daily average power load
diagram and normalized daily air temperature diagram from
2014 to 2019, both of which were normalized in the interval
[0 1]. An average load is average of all 24-hourly loads on a
daily diagram. The analysis of the power load data shows that
there is a high variation between hourly loads in the power
system on a year basis. Fig. 2 confirms the strong negative
correlation between the power load and the air temperature

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 110


Loads and Temperatures
Fig. 2. Normalized average daily loads and air temperatures for the
years 2014-2019
which is typical for this region: 1) the electric energy
consumption is large in winter; 2) the positive air temperature
peaks correspond to negative power load peaks and vice
versa.
The regression analyses over power load and temperature
datasets in 2014 and 2015 performed in [16] and [17] examined
the approximations parameters, determination coefficients (R2)
and correlation coefficients [18]. According to [16] and [17], the
determination coefficients for polynomial regression and sinuses
regression of the maximal, average and minimal daily load due
to the average temperature are very high, which means that the
regression analysis shows high prediction degree of the daily
typical loads from the air temperature. The correlation
coefficients for polynomial regression and sinuses regressions
have values in a range between −0.90 to −0.95 what implies very
close negative relation between all the combinations of typical
daily loads and air temperatures.
III. METHODOLOGY
K-nearest neighbour machine learning model [19] is
considered for power system load forecast. The power load as
a dependent variable depends on two independent variables –
average air temperature and date. This means the algorithm
will search for suitable k power load candidates around a
certain period and temperature. After defining the training
dataset (temperatures, dates and power load for the years
2014-2018), the model is tuned by setting an appropriate
parameter for the number of neighbours k, and then it is
trained on the training dataset.
The effectiveness of the model is evaluated by 10-fold and
Leave-one-out cross-validations. Cross-validation is a method
for obtaining reliable estimates of model performance using
only training data. To predict the performance of a model on a
new dataset, it is needed to assess its performance on a dataset
that plays no part in the formation of the model – the test
dataset. By comparing the test performance and training
111
Selecting the
Defining the
independent and
training and Tuning the model
dependendent
testing dataset
variables
Applying the Evaluating the
Training the
model over the model with the
model
testing dataset training dataset
Fig. 3. Phases of the defined machine learning model
performance, overfitting can be avoided. If the model
performs well on the training data, but poorly on the test data,
then it is overfitted. Besides with the correlation coefficient R
and determination coefficient R2, the performance can be
measured in other ways; two of those ways are through the
root-mean-squared error RMSE and mean-absolute error
MAE:
∑
= , (1)
∑ | |
= , (2)
where pi and ai are the predicted and actual values,
respectively, while n is the total number of the test instances.
All the phases of the machine learning model are illustrated
in Fig. 3.
IV. CASE STUDY
This section presents the results obtained by using the
presented methodology. The temperatures, dates and power
load for the years 2014-2018 are used as a training dataset, while
corresponding data for 2019 as a testing dataset. Figure 4
illustrates distributions of the average air temperatures and the
average power loads for the years 2014-2018 through 365 days,
respectively. The blue circles in Figures 4 denotes the
temperatures and power loads in the period 2014-2018,
respectively, while the red stars denote the temperatures and
power loads in the forecast period 2019. Figure 5 illustrates
distribution of the real average power load in the period Mar. 1-
21 2014-2019.
According to the above methodology the average power
system load for the year of 2019 is forecasted on the basis of the
defined training dataset. KNN machine learning model is used
over the two independent variables, average air temperature and
date, and the distance between neighbours is measured by
Euclidean distance function as a commonly used distance
metric. The experiments for this case study show that a suitable
number of neighbours in the model is 30. The minimum and the
maximum of the variable average air temperature data are −15
and +30(C), respectively, while the minimum and maximum of
the variable date are 1 and 365 (the first and the last day in a
year), respectively. Variables measured on different scales
contribute differently to the analysis, which may eventually lead
 Power load [MW]
Temperature [C]
(a)
(b)
Fig. 4. Distribution of average air temperatures and average power
loads for the years 2014-2019
to bias. The variable date (due to the larger range) will outweigh
the variable air temperature, i.e. the variable date will have a
bigger weight in an analysis compared to the variable air
temperature. This means the date will have higher influence on
the calculated distance than the air temperature will do.
Transforming the data to a comparable scale can avoid this
problem. Normalization is a way of standardizing a set of
numbers so each one is between 0 and 1. Hence, both the
variables in this model are normalized in the range [0−1].
The model is evaluated by 10-fold and Leave-one-out
cross-validations through the root-mean-squared error and the
mean-absolute error (30 neighbours used over the 2014−2018
dataset). Results given in Table 1 show that the errors are
smaller when the variables are normalized.
Results of applying the model over the testing dataset (2019,
30 neighbours used) and forecasting the corresponding power
load are given in Table 2 and Figure 6. Table 2 gives an
overview of MAE and RMSE errors and correlation coefficient
112
Fig. 5. Distribution of average power load for the period Mar. 01-21,
2014-2019
TABLE I
PERFORMANCE OF THE MODEL MEASURED ON THE TRAINING DATASET
[MW]
Cross-validation normalized variables non-normalized variables
10-fold 62.2467 66.4218
Leave-one-out 65.8212 62.2133
for power load forecast with different models. Fig. 6 graphically
illustrates the deviation between forecasted and real power load
data. These results show that for power load forecasts, k-
nearest neighbour regression outperforms polynomial and
sinuses regressions.
V. SUMMARY
This paper is the first one using k-nearest neighbour
machine-learning model for investigation the forecast of
power system load in correlation with air temperature and date
on real data of power system of Republic of North Macedonia.
On the basis of statistical analysis, it can be noticed that there
is a close time matching in appearance of power system
maximum load and minimum air temperature. The same time
matching is noticed between power system summer maximum
load and registered maximum temperature in analysed years.
The presented results show that for power load forecasts the
proposed algorithm outperforms polynomial and sinuses
regressions. The effectiveness of the model is evaluated by 10-
fold and Leave-one-out cross-validations. Methodology works
with all types of days and simulation results are given in a
large time frame.
ACKNOWLEDGEMENT
This research is supported by the EU H2020 project
TRINITY (Grant Agreement no. 863874) This paper reflects
only the author’s views and neither the Agency nor the
 TABLE II
COMPARISON OF MAE, RMSE ERRORS AND CORRELATION COEFFICIENT
R OF DIFFERENT MODELS FOR POWER LOAD FORECAST FOR 2019 DATA

Algorithm MAE RMSE R

k-nearest neighbour
38.4046 50.6231 0.9614
(normalized variables)
k-nearest neighbour
39.1885 51.7435 0.9564
(non-normalized var.)
polynomial (order 4) 41.3541 56.4752 0.9488
sinuses (order 4) 40.7253 55.3602 0.9523
sinuses (order 4) + wavelet
41.1788 54.2211 0.9584
transform

Fig. 6. Comparison of power load KNN forecast with normalized

Commission are responsible for any use that may be made of variables and real average power load for the year 2019
the information contained therein.

using Similar Days Approach”, Int. J. Electr. Power Energy

REFERENCES
[1] E. A. Feinberg, D. Genethliou, Chapter 12: Load Forecasting,
Applied Mathematics for Power Systems, pp. 269-282.
[2] W. Charytoniuk, M. S. Chen, and P. Van Olinda,
“Nonparametric Regression Based Short-Term Load
Forecasting”, IEEE Trans. Power Syst., vol. 13, no. 3, pp. 725-
730, Aug. 1998.
[3] S. Rucic, A. Vuckovic, and N. Nikolic, “Weather Sensitive
Method for Short Term Load Forecasting in Electric Power
Utility of Serbia”, IEEE Trans. Power Syst., vol. 18, no. 4, pp.
1581-1586, Nov. 2003.
[4] T. Hong, M. Gui, M. E. Baran, and H. L. Willis, “Modeling and
Forecasting Hourly Electric Load by Multiple Linear
Regression with Interactions”, in Proc. IEEE PES Gen. Meeting,
Jul. 2010, pp. 1-8.
[5] Y. Chen, P. B. Luh, C. Guan, Y. Zhao, L. D. Michel, M. A.
Coolbeth, P. B. Friedland, and S. J. Rourke, “Short-Term Load
Forecasting: Similar Day-Based Wavelet Neural Networks”,
IEEE Trans. Power Syst., vol. 25, no. 1, pp. 322-330, Feb. 2010.
[6] G. Sudheer and A. Suseelatha, “Short Term Load Forecasting
using Wavelet Transform Combined with Holt_Winters and
Weighted Nearest Neighbor Models”, Int. J. Electr. Power
Energy Syst., vol. 64, pp. 340-346, 2015.
[7] W. Charytoniuk, M.-S. Chen, “Very Short-Term Load
Forecasting using Artificial Neural Networks”, IEEE Trans.
Power Syst., vol. 15, no. 1, pp. 263-268, Feb. 2000.
[8] H. S. Hippert, C. E. Pedreira, and R. C. Souza, “Neural
Networks for Short-Term Load Forecasting: A Review and
Evaluation”, IEEE Trans.Power Syst., vol. 16, no. 1, pp. 44-55,
Feb. 2001.
[9] P. Mandal, T. Senjyu, N. Urasaki, and T. Funabashi, “A Neural
Network Based Several-Hour-Ahead Electric Load Forecasting
Syst., vol. 28, no. 6, pp. 367-373, 2006.
[10] H. Shi, M. Xu, and R. Li, “Deep Learning for Household Load
Forecasting - A Novel Pooling Deep RNN”, IEEE Trans. Smart
Grid, vol. 9, no. 5, pp. 5271-5280, Sep. 2018.
[11] M. Cai, M. Pipattanasomporn, and S. Rahman, “Day-ahead
Building-Level Load Forecasts using Deep Learning vs.
Traditional Time-Series Techniques”, Appl. Energy, vol. 236,
pp. 1078-1088, 2019.
[12] Dispatching and Operational Reports for Power System Load
Data obtained from MEPSO.
[13] H. H. Çevik and M. Çunka³, “Short-Term Load Forecasting
using Fuzzy Logic and ANFIS”, Neural Comput. Appl., vol. 26,
no. 6, pp. 1355-1367, 2015.
[14] D. Bajs, M. Atanasovski, Longterm Forecast Study of Electrical
Energy and Power Balance and Adequacy Analysis of
Transmission Network of Republic of Macedonia,
Zagreb/Skopje EIHP, 2016.
[15] https://www.wunderground.com/history
[16] M. Atanasovski, M. Kostov, B. Arapinoski, and I. Andreevski,
Correlation between Power System Load and Air Temperature
in Republic of Macedonia, Int. Scientific Conf. on Information,
Communication and Energy Systems and Technologies, pp.
213-216, Sozopol, Bulgaria, Jun. 2018.
[17] M. Kostov, M. Atanasovski, G. Janevska, and B. Arapinoski,
“Power System Load Forecasting by using Sinuses
Approximation and Wavelet Transform”, Int. Scientific Conf.
on Information, Communication and Energy Systems and
Technologies, pp. 273-276, Ohrid, North Macedonia, Jun. 2019.
[18] S. Vukadinovic, Elements of Probability Theory and
Mathematical Statistics, Belgrade, 1973.
[19] I. H. Witten, E. Frank, M. A. Hall, and C. J. Pal, Data Mining:
Practical Machine Learning Tools and Techniques, Burlington,
USA, Morgan Kaufmann, 2017.

113
 MATLAB Simulation of
Grid-Connected Photovoltaic System
Pande Popovski1, Natasa Mojsoska2 and Blagoja Stevanoski2
Abstract – In this paper work will be briefly covered the
principle of operation of photovoltaic system, the procedure of
integration of such power plants in the power system, their
impact on the system and the technical norms that they should
meet. Furthermore, the numerical results obtained from the
simulation of a photovoltaic system conducted in Simulink will
be shown and commented on. Along with this, the model with
which the simulation was made, which was made on the basis of
a real photovoltaic system owned by the Technical Faculty -
Bitola, will be discussed.
Keywords – Photovoltaic system, Matlab/Simulink simulation,
Renewable energy sources RES.
I. INTRODUCTION
In the last few decades, especially since the beginning of
the 21st century, the interest in the participation of renewable
energy sources (RES) as producers in the power system has
been progressively increasing. One of the reasons for this is
the continuous increase in consumption, both in developing
countries and in already developed countries. Global warming
has also contributed to this, which has become a major topic
of interest due to its close connection to the use of fossil fuels
[1].
Among other types of renewable sources is the energy
obtained from photovoltaic power plants, which is marked by
the highest increase from the others. In 2002, an average
annual increase of 48% was recorded, and by the end of 2008
the total installed capacity of such power plants in the world
exceeded 16 GW [2]. A report by the European Photovoltaic
Industry Association (EPIA) shows that in 2010 the total
installed capacity of photovoltaic power plants in Europe was
29,777 MW, while in 2019 in the European Union this power
was 130GW. In whole world in 2019 installed capacity of
photovoltaic power energy was 585GW.
A. Standards
The organizations in charge of normalization in
photovoltaic systems are IEC and IEEE - SA.
1Pande Popovski is postgraduate student at Technical Faculty -
Bitola, Makedonska falanga 33, 7000 Bitola, E-mail:
popovski.pande@gmail.com
2Natasa Mojsoska is with University of Sv. Klimenit Ohridski –
Bitola, Technical Faculty - Bitola, Makedonska falanga 33, 7000
Bitola, E-mail: natasa.mojsoska@tfb.uklo.edu.mk
2Blagoja Stavanoski is with University of Sv. Klimenit Ohridski –
Bitola, Technical Faculty - Bitola, Makedonska falanga 33, 7000
Bitola, E-mail: blagoj.stavanovski@uklo.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
The key standards in the field of IEC photovoltaic systems
are:
• IEC 61730 Photovoltaic Safety Module
• IEC 61215 Crystalline silicon terrestrial photovoltaic (PV)
modules - Design qualification and type approval
• IEC 61646 Thin-film terrestrial photovoltaic (PV) modu-
les - Design qualification and type approval
The IEC 61730 standard covers thorough structural
requirements in order to ensure mechanical safety as well as
reliability in the operation of photovoltaic systems during
their lifetime. This standard in combination with the IEC
61215 or IEC 61646 standard practically covers the thorough
construction of photovoltaic modules.
The IEEE - SA standards for photovoltaic systems are:
• IEEE 1262 Recommended Practice for Qualification of
Photovoltaic (PV) Modules
• IEEE 1374 Guide for Terrestrial Photovoltaic Power
System Safety
• IEEE 928 Recommended Criteria for Terrestrial Photo-
voltaic Power Systems.
There are organizations that are committed to developing
and coordinating standards. The most important of these are
the following organizations:
• ISO (International Organization for Standardization) is an
organization composed of 145 national institutions and 193
technical boards, and is the most important organization for
development and standardization.
• IEC (International Electrotechnical Commission) is a
leading organization for the development and adoption of
norms in the field of electrical engineering. It consists of 112
technical boards.
• CEN (Comité Européen de Normalisation) is the
European Normalization Organization, while CENELEC
(European Committee for Electrotechnical Standardization) is
the European Organization for Normation in the field of
electrical engineering.
B. Procedures for Development and Construction of Power
Plants from Renewable Energy Sources
In the Republic of North Macedonia, for the
implementation of projects for renewable energy sources,
more specifically photovoltaic power plants, relevant legal
and regulatory acts have been established. [6] provides
detailed guidelines for investors planning to enter the market
as electrical energy producers from renewable energy sources.
The following steps, procedures, and required documentation
are generally included in this process:
• Establishing a company
• Land use
114
 • Environmental Impact Assessment Study
• Consent for connection to the distribution system
• Building permit
• Gaining temporary status of preferential producer
• Design and construction of the connector
• Approval for use
• License to produce electrical energy
• Registration of the object in the register of objects for
production of electrical energy from renewable energy
sources:
• Awarding the status of preferential producer
• Electrical energy purchase agreement with electrical
energy market operator
II. MATLAB SIMULATION OF GRID CONNECTED
PV SYSTEM
An analysis of the performance of a photovoltaic power
system through simulation using Simulink - a graphically
oriented program tool based on Matlab - will be explained.
The structure of the model to be used for the simulation and
how it works will be explained. The used model is based on
an already existing photovoltaic system that is installed on the
roof at the Technical Faculty - Bitola, and is connected on the
distribution network.
Fig. 1. Model for a photovoltaic plant built in Simulink
An existing database on solar radiation in the area where
the plant is installed will be used for the assessment of the
energy potential of the PV system, as well as for the
consumption of the Technical Faculty - Bitola for two
characteristic days. With the help of the model, the produced
electrical energy will be calculated in such conditions. The
results will be compared with those obtained in the conceptual
design made for the existing power system.
A. Model of the Photovoltaic System
Two PV systems with a power of 10 kW or a total of 20
kW are installed on the roof of the Technical Faculty in
Bitola. The real photovoltaic system for which the model is
made in Simulink is composed of a structure with solar
modules made of polycrystalline silicon with a power of
250W. The modules are grouped into four arrays, where each
array has 20 serial-connected modules. The four arrays via a
DC switchboard with single-pole fuses and surge arresters are
connected to separate inputs from a three-phase inverter.
Compliance of the voltage and current levels of the inverter
115
with the voltages of the series solar modules has been made
taking into account the dependence of the voltage of the solar
modules on the temperature. That temperature range ranges
from -10°C to 70°C. The inverter has a rated power of 20kW,
and has MPPT capability. Furthermore, the inverter with the
main switchboard is connected via a three-pole switch.
Simulink is a graphically oriented tool that can use blocks
to construct and simulate a particular system. The tool
contains a wide library, which also contains a package of
blocks designed to simulate power systems. Using the
required blocks, the model shown in Fig. 1 is constructed.
In addition to the model of the photovoltaic power plant,
oscilloscopes, along with other additional blocks, have been
installed to carry out the necessary measurements. In general,
solar radiation is measured at one of the block inputs for solar
modules, the effective values of the phase voltages and
currents at the output of the power system connected to the
network, as well as the active power at the output of the PV
system. Also, the output power is sampled at certain time
intervals in order to be further used to calculate the produced
electrical energy. The measurements were made with the help
of a measuring block that provides data on the voltages and
currents at the output of the system.
Fig. 2. Instruments for conducting the necessary measurements
In order to measure the effective values of current and
voltage as well as the active power transmitted by the PV
system, it is necessary to divide the voltages and currents from
the measuring block into a module and an argument. The
modules are then divided by √2 to obtain the effective values.
Ultimately, the active power is calculated using the obtained
quantities.
It can be seen from the Fig. 2 images of the instruments and
blocks for the required calculations that the power factor is
not taken into account in the calculation of active power, i.e. it
is adopted to be 1, and the instrument for measuring active
power is used a switching block that misses the active power
signal after a certain time. This is done in order to avoid
inaccuracies in the results due to the occurrence of oscillations
in voltage, current, as well as the phase shift in them at the
moment of connection of the PV system to the network. This
is a transient phenomenon that lasts for a fraction of a second,
but due to the short time of the software simulation, these
phenomena last relatively long and can lead to large errors in
the results.
B. Used Data to Perform the Simulation
A database of two characteristic days was used to perform
the simulation. Specifically, the solar radiation in the area of
the Technical Faculty was measured, as well as its
consumption for one working day (24.01.2020) and one non-
working day (16.02.2020). The measurements were made at a
time interval of 15 minutes.
The data obtained are implemented in the model shown
earlier. Two different models have been used that use the data
from both days.
In the function, the collected data for the whole day are
entered for a time interval of 0.96 seconds in order to shorten
the simulation time. This means that in the simulation one
hour is represented by a time interval of 0.04 seconds, and 15
minutes by a time interval of 0.01 second. This data is
important for the calculation of the produced energy after the
simulation is performed. The resulting solar radiation
diagrams are shown in the figures below Fig. 3 and 4.
Fig. 3. Diagram obtained for solar radiation on 16.02.2020
The resulting algebraic expressions for radiation
dependence and time consumption are implemented in
Simulink through a block that connects a function written in
Matlab to the model in which that block is placed. The inputs
of this block are the coefficients generated by the tool shown
in the pictures before, and also the block that gives the output
time of the simulation at a given moment. The output of the
solar radiation is connected to one of the inputs of the module,
and the consumption block dictates the operation of a three-
phase dynamic load, which represents the TFB.
Fig. 4. Diagram obtained for solar radiation on 24.01.2020
116
III. ANALYSIS RESULTS
After the simulation is released, the oscilloscopes draw the
curves of the production of the photovoltaic system and the
consumption of the faculty. Also, as mentioned in the 0.01
second time interval, the values of the power generated by PV
system in Matlab's desktop are preserved. This data can be
used to calculate the energy produced for both days.
Fig. 5. Production on 24.01.2020
Fig. 6. Production on 16.02.2020
Fig. 5 and 6 shows the power output by the photovoltaic
system for days 24.01 and 16.02.2020 obtained from the
oscilloscope.
The results obtained for the energy produced for the two
days are shown in Tables I and II.
If we compare the quantities of electrical energy produced
by the simulation with the actual read values, we get the
following:
On January 24, 2020, the photovoltaic system produced
71.47 kWh, and according to the simulation, we get produced
energy 67.52 kWh, which is 3.95 kWh, or 5.5% less value
than the real one. On February 16, 2020, the system produced
91.91 kWh, while the simulation results show a value of 86.76
kWh, or 5.15 kWh (5.6%) less than the real value.
In general, the model gives lower values for the produced
power than the real ones. This may be due to the short
simulation time, where transient phenomena occurring as a
problem that result in an error in the results. If this is the case,
the solution would be to increase the simulation time so that
these oscillations have less effect. However, the bad side of
this would be the long duration of the simulation.
 TABLE I TABLE II
SIMULATION RESULTS FOR THE DAY 24.01.2020 SIMULATION RESULTS FOR THE DAY 16.02.2020

Time Production Time Production Time Production Time Production

(kW) (kW) (kW) (kW)
00:15 0.000 12:15 11.342 00:15 0.000 12:15 13.359
00:30 0.000 12:30 11.217 00:30 0.000 12:30 13.368
00:45 0.000 12:45 11.013 00:45 0.000 12:45 13.274
01:00 0.000 13:00 10.713 01:00 0.000 13:00 13.082
01:15 0.000 13:15 10.331 01:15 0.000 13:15 12.772
01:30 0.000 13:30 9.921 01:30 0.000 13:30 12.316
01:45 0.000 13:45 9.498 01:45 0.000 13:45 11.724
02:00 0.000 14:00 9.072 02:00 0.000 14:00 11.079
02:15 0.000 14:15 8.596 02:15 0.000 14:15 10.423
02:30 0.000 14:30 7.948 02:30 0.000 14:30 9.730
02:45 0.000 14:45 7.046 02:45 0.000 14:45 8.989
03:00 0.000 15:00 5.935 03:00 0.000 15:00 8.176
03:15 0.000 15:15 4.733 03:15 0.000 15:15 7.241
03:30 0.000 15:30 3.530 03:30 0.000 15:30 6.180
03:45 0.000 15:45 2.401 03:45 0.000 15:45 5.014
04:00 0.000 16:00 1.422 04:00 0.000 16:00 3.782
04:15 0.000 16:15 0.654 04:15 0.000 16:15 2.575
04:30 0.000 16:30 0.154 04:30 0.000 16:30 1.248
04:45 0.000 16:45 0.000 04:45 0.000 16:45 0.000
05:00 0.000 17:00 0.000 05:00 0.000 17:00 0.000
05:15 0.000 17:15 0.000 05:15 0.000 17:15 0.000
05:30 0.000 17:30 0.000 05:30 0.000 17:30 0.000
05:45 0.000 17:45 0.000 05:45 0.000 17:45 0.000
06:00 0.000 18:00 0.000 06:00 0.000 18:00 0.000
06:15 0.000 18:15 0.000 06:15 0.000 18:15 0.000
06:30 0.000 18:30 0.000 06:30 0.000 18:30 0.000
06:45 0.000 18:45 0.000 06:45 0.000 18:45 0.000
07:00 0.089 19:00 0.000 07:00 0.089 19:00 0.000
07:15 0.106 19:15 0.000 07:15 0.464 19:15 0.000
07:30 0.478 19:30 0.000 07:30 1.709 19:30 0.000
07:45 1.219 19:45 0.000 07:45 3.039 19:45 0.000
08:00 1.977 20:00 0.000 08:00 4.304 20:00 0.000
08:15 3.060 20:15 0.000 08:15 5.699 20:15 0.000
08:30 4.295 20:30 0.000 08:30 7.070 20:30 0.000
08:45 5.568 20:45 0.000 08:45 8.162 20:45 0.000
09:00 6.801 21:00 0.000 09:00 8.846 21:00 0.000
09:15 7.845 21:15 0.000 09:15 9.333 21:15 0.000
09:30 8.561 21:30 0.000 09:30 9.853 21:30 0.000
09:45 8.997 21:45 0.000 09:45 10.443 21:45 0.000
10:00 9.331 22:00 0.000 10:00 11.086 22:00 0.000
10:15 9.651 22:15 0.000 10:15 11.748 22:15 0.000
10:30 9.986 22:30 0.000 10:30 12.349 22:30 0.000
10:45 10.360 22:45 0.000 10:45 12.791 22:45 0.000
11:00 10.784 23:00 0.000 11:00 13.030 23:00 0.000
11:15 11.169 23:15 0.000 11:15 13.111 23:15 0.000
11:30 11.406 23:30 0.000 11:30 13.137 23:30 0.000
11:45 11.464 23:45 0.000 11:45 13.184 23:45 0.000
12:00 11.415 00:00 0.000 12:00 13.273 00:00 0.000

117
 Another cause could be the photovoltaic maximum power
monitoring (MPPT) algorithm. The Perturb and Observe
algorithm works by changing the voltage at the output of the
photovoltaic, which measures the change in output power.
With this method, the time interval controller increases the
voltage as long as the measurements show that the power also
increases. Working in this way can cause oscillations in
voltage and current, and thus in output power. The solution to
this would be to apply another more stable algorithm.
IV. CONCLUSION
With the help of the Simulink software package, it is
possible to create a very precise model of a photovoltaic
power system. Simulink offers the user a wide range of
databases and elements, and gives them complete freedom in
analyzing such a model. For the same, it is possible to obtain
relatively accurate production data, where the accuracy
depends on the input data itself, as well as on how detailed the
plant model is.
From the developed model it can be concluded that such an
approach would be quite useful in the initial stages, i.e. when
a PV power plant is in the design process. In this case, results
would be obtained with satisfactory accuracy, which would
give an idea of how much production capacity the plant will
have. However, when it comes to more detailed analysis, such
accuracy would not be sufficient. An example of this is the
results of the simulation: for the days 24.01 and 16.02 an error
118
was made in the energy production is around 5.5%, which
although for one or two days is a relatively small error.
However, as mentioned, it is possible for the model to be
improved and thus give more accurate results. Improving the
model would be in the direction of how to increase the
simulation time that Simulink will be able to do in a
reasonable amount of time, which will probably improve the
results.
REFERENCES
[1] E. Rakshani, K. Rouzbehi, A. J. Sánchez, A. C. Tobar, and E.
Pouresmaeil, “Integration of Large Scale PV-Based Generation
into Power Systems: A Survey”, 13 April, 2019.
[2] R. Goić, J. Krstulović Opara, I. Penović, D. Jakus, and I.
Zlatunić, “Connection of Large PV Power Plant od
Disstribution Network”, Umag, 16-19 May, 2010.
[3] B. Stevanovski, “Renewable Energy Sources RES”, Subject
lectures, TFB Bitola, 2017.
[4] M. Atanasovski, “New Methods for Distribution of Losses in
Distribution Networks with Dispersed Electricity Production”,
PhD dissertation, FEIT Skopje, 2011.
[5] R. Orlovac, A. Tunjic, “Technical Requirements for Connection
od PV Generators to the Public Low-Voltage Network”,
CIRED, Šibenik, 18-21 May, 2008.
[6] “Guidelines on the Procedures for the Development and
Construction of Power Plants Utilising Renewable Energy
Sources – Photovoltaic Plants”, Project funded by EU, source:
http://www.economy.gov.mk/Upload/Documents/Guidelines_o
n_photovoltaic_plants_B5_final_3.pdf
 Influence of Reduction Factors on High Voltage Cables
on the Transfer of Potentials in the Network
Nikolche Acevski1, Аngela Micevska2, and Аna Manivilovska3
Abstract - The purpose of this paper is to calculate the The self and mutual impedances Zs and Zm on the three cores
reduction factors of 3 single-core HV cables, that is, to see the of the cables are calculated with Carson equations, [5], for
differences in the voltage at two HV 110 kV substations, ρ=300 Ωm. The cables are placed in a PVC pipe, d=0.16 m.
substation Central and transferred to substation South New,
when applying HV cables with various cross-sections placed in ρ 300
Dek = 658 = 658 = 1611.76m , (2)
triangle and plane configurations. f 50

Keywords - Reduction factor, HV cables, Plane, Triangle, D ek

Transfer of potential. Z s = ( 0 . 05 + re + j ⋅ 0 . 1445 ⋅ log )⋅l , (3)
de / 2

I. INTRODUCTION D ek
Z m = ( 0 .05 + j ⋅ 0 .1445 ⋅ log )⋅l . (4)
D
The reduction factor of the HV cable is defined as the ratio
between the residual current (IKV - Je), which continues to the So each cable with an insulated sheath observed along with its
end of the cable, and the error current IKV, (10), (21). It shows return path through the ground is represented by an I-
us how much of the fault current will be injected into the replacement circuit. The mutual impedances between the
grounding conductor at the end of the cable at the grounded phase conductor of phase f=A, B, C and the screens of the
connection at its end. three wires of the cable are calculated according to (5). For
phase A, (6).
Jem
D ek
M ij = 0 . 05 ⋅ l + j ⋅ 0 .1445 ⋅ log ⋅l , (5)
de / 2
p q
Zek v D ek
M 1 A = 0 .05 ⋅ l + j ⋅ 0 . 1445 ⋅ log ⋅l ≡ M s , (6)
Fig. 1. Single-phase equivalent of 3 single-core cables de / 2

When the three single-core cables are laid in a triangular Dek

M 2 A = M 3 A = 0.05 ⋅ l + j ⋅ 0.1445⋅ log ⋅ l ≡ M m . (7)
layout, as in Fig. 2, the followingapplies: D
 Z s Z m Z m Similarly, the remaining own and mutual impedances are
 
Z = Z m Z s Z m (1) calculated Mij = Mji, between the phase conductors of the
Z m Z m Z s  phases (i = B, C) and the three cable screens (j = 1,2,3). For
  Fig. 2, the single-phase equivalent of Fig. 1 [5] will be:
Z + 2⋅Zm
Ze = s , (8)
3
M + 2⋅M m
Je = s ⋅I , (9)
Z s + 2 ⋅ Z m KV

I KV − J e Je Ms + 2 ⋅ Mm
rf = = 1− = 1− . (10)
I KV I KV Zs + 2⋅ Zm
Fig. 2. Triangle configuration
When the strings are placed in a plane, it is valid [5]:
 Z s Z m1 Z m 2 
 
1Nikolce Acevski, Technical Faculty, St. Kliment Ohridski, Z =  Z m1 Z s Z m1  , (11)
Makedonska Falanga 37, Bitola, Republic of North Macedonia, E-  Z m 2 Z m1 Z s 
mail: nikola.acevski@uklo.edu.mk
 
2Angela Micevska, Technical Faculty, postgraduate, E-mail: D ek
micevska.angela@hotmail.com Z s = ( 0 . 05 + re + j ⋅ 0 . 1445 ⋅ log ) ⋅ l , (12)
3Ana Manivilovska, Technical Faculty, postgraduate, E-mail:
de / 2
ana_manivilovska@hotmail.com

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 119

 TABLE I
CATALOG DATA FOR POLYETHYLENE CABLES OF THE TYPE NA2XS(FL)2Y UM=123 KV

D ek Z m2 ⋅ M m1 + Z s ⋅ (M s + M m1 + M m2)
Z m1 = ( 0 .05 + j ⋅ 0 .1445 ⋅ log )⋅l , (13) r f =1−
D Z 2s − 2 ⋅ Z 2m1 + Z s ⋅ Z m2
(M s + 2 ⋅ M m1 + M m2)
D ek + Z m1 ⋅ . (21)
Z m 2 = ( 0 .05 + j ⋅ 0 .1445 ⋅ log )⋅l , (14) Z 2s − 2 ⋅ Z m
2
2D 1 + Z s ⋅ Z m2

 M s M m1 M m 2 
  II. CALCULATION OF REDUCTION FACTORS FOR
M =  M m1 M s M m1  , (15)
DIFFERENT CABLE CROSS SECTIONS
 M m 2 M m1 M s 
 
D ek After the calculations made according to the above
M s = 0 . 05 ⋅ l + j ⋅ 0 . 1445 ⋅ log ⋅l . (16) formulas, the results shown in Table II are obtained. The
de / 2
reduction factor of a bundle of three single-core cables placed
D ek in a triangle is smaller than that in a plane from 3.7% to 12%
M m1 = 0 . 05 ⋅ l + j ⋅ 0 . 1445 ⋅ log ⋅l , (17) due to the obvious symmetry.
D
D ek TABLE II
M m2 = 0 . 05 ⋅ l + j ⋅ 0 . 1445 ⋅ log ⋅l , (18) REDUCTION FACTORS OF HV CABLES
2D
Cross section Triangle Plane
2 ⋅ ( Z s − Z m1 )2
Ze = Zs − , (19) 2000 0.066e -j81.25
0
0.074e-j76.72
0
3 ⋅ Z s − 4 ⋅ Z m1 + Z m 2
0 0
1600 0.082e-j80.88 0.088e-j75.49
Z m2 ⋅ M m1 + Z s ⋅ (M s + M m1 + M m2) 1400 0 0
J e = I KV ⋅ 0.103e-j79.35 0.109e-j75.71
Z s2 − 2 ⋅ Z m
2
1 + Z s ⋅ Z m2 0 0
1200 0.139e-j77.16 0.148e-j74.27
0 0
(M s + 2 ⋅ M m1 + M m2) 1000 0.191e-j74.22 0.198e-j74.77
− I KV ⋅ Z m1 ⋅ (20)
Z 2s − 2 ⋅ Z 2m1 + Z s ⋅ Z m2

120
 - both thorough grounding with grounding resistance
RT1=4.42 Ω and RT2=3.74 Ω,
- the cables tunnel grounding resistance Rkt=6.81 Ω
- the first sections of two cables (their equivalent
impedances)
The values of the groundings are obtained by computer
Fig. 3. Configuration plane
simulation in Matlab. At the places where the phases of the
cables are transposed (nodes 3,4,5,6), groundings are
III. CALCULATION OF SHORT-CIRCUIT VOLTAGE simulated with which the surge arresters are grounded and
CONDITIONS IN SUBSTATION CENTRAL which are predicted to be at R=5 Ω.
The third sections of the two cables are connected to the
Central substation is connected to the South New grounding of the South New substation and the grounding of
substation by two 110 kV cables. The appearance of the the substation which is Z2 = 0.18 Ω. It should be emphasized
circuit is presented in Fig. 4, and in Fig. 5 it’s equivalent that the EVN services have obtained the data for Z1=0.25 Ω
scheme, taking into account the transposition of the phases. and Z2=0.18Ω which are the average values of the
measurements at several locations at these substations. At the
same time, they do not represent grounding resistance of
grounding conductors at substations, but input impedances
that take into account the input impedances of all MV cables
connected to the corresponding substations.
Fig. 5 shows that the network between the Central
substation (node 1) and the South New substation (node 2)is
composed of 6 nodes. After two additional nodes introduces
the transposition of cables to 1/3 of their lengths 2.7 km or
4.11 km respectively, 0.9 km or 1.37 km.It is a non-radial
network, so since the input impedance of the system at node 1
Fig. 4. View of the circuit between Central and South New
substation during a short circuit in Central substation substation Central cannot be calculated by simply summing
the admittances as in other cases in practice. The system can
be solved using the method of independent voltages
3 4
(potentials in the nodes). For this purpose, the matrix of
Zek v1 Zek v1 Zekv1 admits Y with dimensions 6x6 is formed first. Matlab
R R
calculates the impedance matrix Z = Y-1, and then calculates
1
ТС
Централна
the voltage vector in all 6 nodes of the network with the
2 ТС Југ Нова
relation (22)
Ik v − Iem Z1 RT 1 RT 2 Rmz Rk t
Z2 Iem U = Z ⋅I . (22)
5 6
From the power plant operator MEPSO, data were obtained
Zekv2
Zekv2 Zek v2
for three-way and single-circuit short-circuit currents of 110
R R
kV busbars in substation 110/35/10 (20) kV "Central".

TABLE III
Fig. 5. Equivalent circuit scheme between Centraland South New
substation during short circuit in Central substation

According to [4], first the currents that are injected into the
Fig. 6. Appearance of grounding of the Central substation grounding systems of the Central substation and South New
substation are calculated, and then the voltages of the
At 110 kV substation Central are tied, [4], Fig. 6: respective groundings, according to the relation (22).

- the old grounding of the building with an entrance

impedance Z1=0.25 Ω,
- the new mesh grounding with grounding resistance
Rmz=4.508 Ω,

121
 1)
Triangle configuration, cross section 2000 mm2 U CN   247.05 e− j 66.2 
   j 4.45 
 5.309− j2.986

0 + j0 − 0.269+ j1.802 0 + j0 − 0.178+ j1.184 0 + j0 
U JN   2705.45 e 
− 0.178+ j1.184
 0 + j0 6.002− j2.986 0 + j0 − 0.269+ j1.802 0 + j0
 U 3   959.56 e− j12.28 
− 0.269+ j1.802  
Y =
0 + j0 0.737− j3.603 − 0.269+ j1.802 0 + j0 0 + j0

U = = − j 2.96
 0 + j0 − 0.269+ j1.802 − 0.269+ j1.802 0.737− j3.603 0 + j0 0 + j0   U 4  1798.14 e 
− 0.178+ j1.184

0 + j0 0 + j0 0 + j0 0.555− j2.386 − 0.178+ j1.184

 U   945.14 e − j16.29 
 0 + j0 − 0.178+ j1.184 0 + j0 0 + j0 − 0.178+ j1.184 0.555− j2.386   5
  
 U 6  1772.27 e − j 5.54 
0.180 + j 0.034 0.010 − j 0.028 0.121+ j 0.002 0.064 − j 0.016 0.118 − j 0.002 0.061− j0.020
0.010 − j 0.028 0.160 + j 0.026 0.057 − j0.018 0.108 − j0.004 0.055 − j 0.021 0.105 − j 0.006 4)

0.121+ j 0.002 0.057 − j0.018 0.181+ j 0.336 0.126 + j 0.111 0.093 − j 0.018 0.072 − j 0.023
Configuration plane, cross section 1600 mm2
Z= 
0.064 − j 0.016 0.108 − j 0.001 0.126 + j 0.151 0.174 + j0.342 0.072 − j0.023 0.087 − j 0.019  5.409− j3.098 0 + j0 − 0.332+ j1.871 0 + j0 − 0.214+ j1.227 0 + j0 
0.118 − j 0.002 0.055 − j 0.021 0.093 − j0.018 0.072 − j0.023 0.236 + 0.492 0.160 + j 0.219  
   0 + j0 6.102− j3.098 0 + j0 − 0.332+ j1.871 0 + j0 − 0.214+ j1.227
0.061 − j 0.020 0.105 − j 0.006 0.072 − j0.023 0.087 − j0.019 0.160 + j 0.219 0.230 + j0.491 − 0.332+ j1.871 0 + j0 0.865− j3.743 − 0.332+ j1.871 0 + j0 0 + j0 
Y = 
 0 + j0 − 0.332+ j1.871 − 0.332+ j1.871 0.865− j.3.743 0 + j0 0 + j0 
U CN   294.6 e − j 65  − 0.214+ j1.227 0 + j0 0 + j0 0 + j0 0.628− j2.454 − 0.214+ j1.227
   j 5.66 
 
U JN   2708.7 e
 0 + j0 − 0.214+ j1.227 0 + j0 0 + j0 −0.214+ j1.227 0.628− j2.454

 U 3   969 e − j13.01  0.178 + j0.034 0.011− j 0.028 0.120 + j0.003 0.064 − j0.016 0.118 − j0.003 0.062 − j0.020
U = = − j 2.32
  0.011− 0.028
j 0.159 + j 0.026 0.058 − j 0.018 0.100 − j0.001 0.056 − j 0.021 0.105 − j0.006
 U 4  1803.2 e  
0.120 + j0.003 0.058 − j0.018 0.185 + j 0.319 0.127 + j 0.143 0.094 − j0.017 0.073 − j0.023
 U   937.9 e − j16.83  Z = 
0.064 − j0.016 0.107 − j0.001 0.127 + j0.143 0.178 + j 0.318 0.073 − j 0.023 0.087 − j 0.019
 5
   0.118 − j 0.003 0.056 − j0.021 0.094 − j0.017 0.073 − j0.023 0.244 + j0.474 0.164 + j 0.210
 U 6  1756.2 e − j 4.62   
0.062 − j0.020 0.105 − j0.006 0.073 − j 0.023 0.087 − j 0.019 0.164 + j0.210 0.238 + j0.473

2)
Configuration plane, cross section 2000 mm2
 5.351− j3.12 0 + j0 − 0.295+ j1.884 0 + j0 − 0.194+ j1.235 0 + j0  U CN   262.1e − j 53.03 
     j 4.48 
U JN   2670.6 e
0 + j0 6.044− j3.12 0 + j0 − 0.295+ j1.884 0 + j0 − 0.194+ j1.235

− 0.295+ j1.184 0 + j0 0.789− j3.759 − 0.295+ j1.884 0 + j0 0 + j0 

Y =   U 3   972.5 e − j11.57 
 0 + j0 − 0.295+ j1.884 − 0.295+ j1.884 0.789− j3.769 0 + j0 0 + j0  U = = − j 2.73

− 0.194+ j1.235

0 + j0 0 + j0 0 + j0 0.588− j2.471 − 0.194+ j1.235
  U 4   1878 e 
 0 + j0 − 0.194+ j1.235 0 + j0 0 + j0 − 0.194+ j1.235 0.588− j2.471  U   958.4 e − j15.36 
0.178+ j0.034 0.111− j0.029 0.121+ j0.003 0.064− j0.016 0.119− j0.002 0.062− j0.020  5
  
0.011− j0.029
 0.159 + j0.027 0.058− j0.018 0.107 − j0.001 0.056 − j0.021 0.106 − j0.006  U 6  1763.7 e− j 5.18
0.120 + j0.003 0.058 − j0.018 0.178+ j0.321 0.124 + j0.144 0.094 − j0.017 0.073− j0.023
Z= 
0.064 + j0.016 0.107 − j0.001 0.124 + j0.144 0.171+ j0.319 0.073− j0.023 0.088 − j0.019
0.119 − j0.002

0.056 − j0.021 0.094 − j0.017 0.073− j0.023 0.236 + j0.475 0.161+ j0.212
 IV. CONCLUSION
0.062 − j0.020 0.106 − j0.006 0.073− j0.023 0.088 − j0.019 0.161+ j0.212 0.230 + j0.474

1. The reduction factor of a bundle of three single-core

U CN   299.2 e − j 59.49  cables placed in a triangle is smaller than that in a plane
U   j 5.4  from 3.7% to 12% due to the obvious symmetry.
 JN   2680.5 e 
 U 3   976.5 e − j12.54  2. On the one hand, smaller reduction factors contribute to
U = = − j 2.28
 reducing the voltage in the Central substation and thus
 U 4  1793.7 e 
relieving its grounding, but on the other hand they
 U   962.8 e− j16.4 
 5   contribute to increasing the transferred potential in the
 U 6   1768 e − j 4.77  South New substation and increasing the risks of
excessive touch and step voltages in and around the South
3)
Triangle configuration, cross section 1600 mm2 New substation.
 5.355− j2.971 0 + j0 − 0.297+ j1.792 0 + j0 − 0.195+ j1.178 0 + j0  3. When choosing a HV cable, it is necessary to choose the
 
 0 + j0 6.048− j2.97 0 + j0 − 0.297+ j1.792 0 + j0 − 0.195+ j1.178 optimal configuration in terms of meeting the criteria for
− 0.297+ j1.792 0 + j0 0.794− j3.585 − 0.297+ j1.792 0 + j0 0 + j0  safety in and around both substations.
Y = 
 0 + j0 − 0.297+ j1.792 − 0.297+ j1.792 0.794− j3.585 0 + j0 0 + j0 
− 0.195+ j1.178 0 + j0 0 + j0 0 + j0 0.59− j2.356 − 0.195+ j1.178
 
 0 + j0 − 0.195+ j1.178 0 + j0 0 + j0 − 0.195+ j1.178 0.59− j2.356 
REFERENCES
0.179 + j0.033 0.010 − j0.028 0.121 + j0.002 0.064 − j0.016 0.119 − j0.003 0.062 − j0.020
 
0.010 − j 0.028 0.160 + j0.026 0.057 − j0.018 0.107 − j0.002 0.055 − j0.021 0.105 − j 0.006 [1] HD637 S1: Power Instalations Exceeding 1 kV AC, 1999.
0.121 + j0.002 0.057 − j0.018 0.186 + j0.334 0.129 + j0.150 0.094 − j0.018 0.073 − j0.024
Z =  [2] J. Nahman, "Neutral Grounding in Power Distribution
0.064 − j0.016 0.107 − j 0.002 0.129 + j0.150 0.179 + j0.333 0.073 − j0.024 0.087 − j0.020
0.119 − j0.003 0.055 − j 0.021 0.094 − j0.018 0.073 − j0.024 0.249 + j0.494 0.168 + j0.220
Networks", Naucna kniga, Belgrade, 1980 (in Serbian).
  [3] J. Nahman, V. Mijailović, "High Voltage Power Stations",
0.062 − j0.020 0.105 − j0.006 0.073 − j0.024 0.087 − j0.020 0.168 + j0.220 0.243 + j0.493
Academic Mind, Belgrade, 2005 (in Serbian).
[4] New substation 110/35/10 (20) kV Central, book 5, grounding,
Basic project, TIMEL, Nikolce Acevski, 2019.
[5] R. Ackovski, M. Todorovski, Groundings and Grounding
Systems in Power Networks, Skopje, 2017.

122
 Evaluation of Reliability of Power Substations
Considering the Uncertainty of the Distributed Generation
Atanas Iliev1, Dimitar Dimitrov2 and Natasha Dimishkovska3
Abstract – This paper presents an upgrade of an existing fuzzy
logic method for evaluation of the reliability of power substations
with implemented distributed generation, considering
uncertainty and imprecision of the input data. The method is
presented on a 110/35 kV/kV substation and gives satisfactory
results from an engineering point of view and provides a high
level of precision and model simplicity.
Keywords – Reliability, Power substation, Fuzzy logic,
Uncertain data, Distributed generation.
I. INTRODUCTION
As the fossil fuels are approaching their expiration date,
introducing the Renewable Energy Sources (RES) to the
power system has been proven to be an acceptable alternative
solution. Using the RES for local power production can
benefit the environment, the community and one county’s
economy.
However, there are many technical challenges the
Distributed Energy Sources (DERs) are facing. The most
important is providing an uninterrupted power supply to the
customers. Considering that the distributive networks are
more exposed to outages, the implementation of DERs may
have a huge impact on their reliability [1]. Therefore, the
reliability of the power substation where the DERs are
connected has to be properly evaluated in the designing
process. Power substations are power facilities that are part of
the power system and their main function is to distribute the
electrical energy and/or transform it to certain voltage levels
[2]. Power substations are supposed to perform reliably in a
way that provides quality electrical energy supply, with lowest
exploitation costs.
The estimation of the reliability parameters of the
components in the power substations contains high level of
insecurity and uncertainty, which are still given statistical
character. The statistical analysis starts with calculation of the
mean value of the reliability parameters of the components
and further this data are used for estimation of the reliability
1Atanas Iliev is with Faculty of Electrical Engineering and
Information Technologies, Rugjer Boshkovikj 18, Skopje 1000,
North Macedonia,E-mail: ailiev@feit.ukim.edu.mk
2Dimitar Dimitrov is with Faculty of Electrical Engineering and
Information Technologies, Rugjer Boshkovikj 18, Skopje 1000,
North Macedonia, E-mail: ddimitar@feit.ukim.edu.mk
3Natasha Dimishkovska is with Faculty of Electrical Engineering
and Information Technologies, Rugjer Boshkovikj 18, Skopje 1000,
North Macedonia, E-mail: dimishkovskan@feit.ukim.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
parameters of a certain single line diagram of power
substation [3]. However, it is more appropriate to assume that
the unavailability of the components and the uncertainty
belong to a certain range, rather than assuming that they have
fixed value [1]. Additionally, the dependency on weather
conditions makes the stand-alone DERs systems quite
unreliable, and their implementation to the utility grid, has a
major impact on the frequency and power balancing.
Therefore, in this paper, the fuzzy logic theory is used for
estimation of the reliability of single line diagrams of power
substations with implemented DERs. This method, compared
to the conventional methods, offers wider range of scenario
analysis that can cause failure of the power substation. The
main characteristic is that the values of the unavailability of
the substation, the duration of energy not supplied, as well as
the energy not supplied due to component failure in the
substation are also fuzzy numbers, defined with proper
membership function.
II. MAIN CONCEPTS OF THE FUZZY LOGIC
THEORY
Fuzzy set theory and the fuzzy logic represent branches of
the science and computer engineering, which belong to the
artificial intelligence field. Unlike the conventional Boolean
logic, where each element belongs to the given set or not, in
the fuzzy set theory, each element has its membership degree.
The membership degree varies between 0 (the element does
not belong to the set) and 1 (the element completely belong to
the set). Therefore, fuzzy logic is also known as multi-value
logic and represents an expansion of the classical bi-value
Boolean logic [4].
The fuzzy set A in the universe of discourse U defines as
set of ordered pairs:
= ( , ( ))| ∈ (1)
where, ( ) represents the membership function of the fuzzy
set A. The membership function defines the degree of
membership of the variable x to the fuzzy set A.
It is obvious that the fuzzy set definition is an expansion of
the classical set theory. Respectively, if the range of the
membership function expands, the uncertainty and the
indeterminacy increase. Otherwise, if the range of the
membership function reduces, the uncertainty and the
indeterminacy decrease, and the fuzzy set turns into
conventional set.
The design of the fuzzy sets depends on two factors:
identifying the state space and the corresponding membership
function. The specification of the membership function is
123
subjective, which means that there is diversity of solution to
one problem, since different expert take different approach.
The reason behind that are the individual differences of
human perspective to the world and abstract concepts.
If a real number contains certain level of uncertainty and
unreliability, or the analysed variable’s value can be estimated
in a certain range [ ; ], it is more suitable to model an
interval number, with membership function defined as:
1 ≤ ≤
~( )=
0 < ⋀ >
Fuzzy numbers are fuzzy subset that is normal and convex.
Fuzzy numbers are expansion of the ordinary real numbers
∈ ℜ and interval numbers. They are used for modeling
unreliable data and information with convex membership
function defined with values in the range [0,1] (for instance,
the value is approximately a). In Fig. 1, a fuzzy number A and
its alpha-cuts and ′ are shown, defined with proper interval
number. Membership functions construction depends on two
factors:
- Identification of state variables and state space, and;
- Determining the proper function of belonging, which
is subjective, meaning that there are many different
solutions of one problem
A. Uncertainty Modelling
Uncertainty defines as a feature of the real systems, caused
by unpredicted events, which affects their reliability of proper
working. Uncertainty can be caused by many reasons, mostly
regarding the lack of information and knowledge to run the
systems [3]. Fuzzy logic describes uncertainty with
probability of happening with numbers between 0 and 1.
Additionally, fuzzy logic provides analysis of uncertain events
even if the probability is in a certain interval.
If the value of a variable is in a certain range, for instance
[ ; ], the membership function of the variable is defined
as:
1 ≤ ≤ Assuming that the weather conditions of a certain location
( )=
0 < ∧ >
That represents a fuzzy triangular number. Generally the
fuzzy triangular numbers are represented with two end points
and , and a peak point . If there is an - cut, then the
fuzzy interval is presented as in Eq. (5). Fuzzy numbers can
be treated as regular fuzzy numbers, regarding the
mathematical operations applied.
=[ ; ; ]
( ) ( ) ( )
=[ ; ; ] (5)
Graphical representation of fuzzy triangular numbers is
shown in Fig. 1.
( )
−
(2)
0
1 1 2 2 3
Fig. 1. Triangular fuzzy number
III. THE PROPOSED APPROACH
The DERs depend highly on the weather conditions, and
because of that, their power generation varies at any analysed
moment. Therefore, analysing the total power generation can
be a problem if the analysis is limited to a certain power
capacity. Using fuzzy numbers can expend the analysis to a
defined range of possible power generation, and using a
triangular fuzzy numbers and membership functions for
reliability evaluation of power substations provides simpler
energy loss analysis [5].
DERs penetration into the power system can cause
destabilisation of the frequency, damages to the installed
equipment and decrease of reliability of supply. In this paper,
the fuzzy numbers are used for estimation of reliability of
distributive substations with implemented DER.
In [6] a method for reliability estimation of based on fuzzy
logic substations is presented. The paper justifies the usage of
fuzzy logic for reliability estimation of power systems.
Additionally, in [7] a fuzzy logic method is used for optimal
distributed generation allocation in radial distribution system.
In this paper the fuzzy logic method is applied for reliability
estimation of more complex substations with DERs
connected.
(3)
are known, the power production of the DER can be easily
calculated according the expressions for power production for
different RES. The reliability of supply of the whole
substation can be calculated using the expressions for serial
and parallel connections of the components.
If K denotes the number of components connected in a
series, and if ( , ( )) describes the unavailability, then
the equivalent unavailability of a serial union of components
(4) for a given - cut defines as [3]:
( )
=∑
( )
−∑
( )
∙∑
( )
+⋯≅
( )
∑ (6)
The unavailability can be represented as triangular fuzzy
number = [ ; ; ] as well. In that case, the
equivalent unavailability will also be a fuzzy triangular
number Fig. 2.
=[ ; ; ] = [∑ ;∑ ;∑ ]. (7)
124
 TABLE I
1 2
⋯ COMPONENTS’ UNAVAILABILITY RATES
Component Voltage Unavailability
Fig. 2. Serial connection of the components (kV)
Circuit breaker 110 [2.85; 3.00; 3.10] ∙ 10
1 Disconnector 110 [3.01; 3.05; 3.25] ∙ 10
Bus_HV 110 [5.18; 5.53; 5.75] ∙ 10
2
Circuit breaker 35 [4.21; 4.60; 4.90] ∙ 10
Disconnector 35 [3.06; 3.20; 3.33] ∙ 10
⋮ Bus_MV 35 [4.80; 4.97; 5.3] ∙ 10
Power transformer 110/35 [4.50; 4.67; 4.75] ∙ 10
DER 110 [3.58; 3.70; 3.85] ∙ 10
Fig. 3. Parallel connection of the components

3
7
1 2 A B
4 6

8
5

Fig. 4. Combination of the serial and parallel connection of the C

components 110 kV

If the components are connected in parallel, as shown in

Fig. 3, then the equivalent unavailability of the system is
defined in the following manner:
( ) ( )
=∏ . (8)
And when the unavailability of the components is given as D E F
a fuzzy set interval, then the equivalent unavailability is not a
50 MW
fuzzy number, but a fuzzy polynomial. 50 MW 50 MW

( ) ( ) ( )
= , =
= [∏ [ +( − ) ]; ∏ [ +( − ) ]]. (9)
Usually the components in the system are connected both in
series and in parallel connection (Fig. 4). Then, the equivalent G
unavailability of the systems is calculated with the 35 kV
combination of Eqs. (6)-(9).

Fig. 5. Single line diagram of 110/35 kV/kV substation

IV. TEST EXAMPLE
The test example analyses a 110/35 kV power substation
with 50 MW DER connected to the 110 kV bus. The DER
may represent a big RES power plant (wind farm, solar power
plant) or a smaller power plant with diesel generator or a
batteryas a back-up. The single line diagram of the test
example is given in Fig. 5. The unavailability rates of the
components are given in Table I.
It is assumed that the DER can satisfy half of the load, and
that it can provide power in case of power lines’ outage.If one
of the transformer branches is in outage, the load will be
satisfied from the distributed generator connected to the 110
kV bus. But, if both of the transformer branches are in outage,
or there is a fault on the 35 kV bus, then the power supply will
be interrupted.
Since the DERs power production varies and it is time-
dependant, the reliability of distributed generator cannot be
represented with a single value [8]. The unavailability rate of
the DER, which is considered to be a RES, includes the
metrological data for the analysed location, as well as the
probability of failure of the generator. The fuzzy
representation of DER’s power production is shown in Fig. 6.
The probability of one branch failure is calculated by using
the expression (7). For more accurate and easier calculation,
the substation is divided into parts. Each part represents a
branch with the appropriate switching equipment.
The probabilities of failure of each part of the substation,
computed using the expressions (2)-(4), are given in Table .

125
Membership
function
a1 a2 a3
Time
Fig. 6. Fuzzy triangular number representation of DER’s power
generation
TABLE II
UNAVAILABILITY OF SUBSTATION’S DIFFERENT PARTS
Part Unavilability
A [3,15; 3,31; 3,43] ∙ 10
B [3,15; 3,31; 3,43] ∙ 10
C [2,04; 2,12; 2,21] ∙ 10
D [5,21; 5,38; 5,55] ∙ 10
E [5,21; 5,38; 5,55] ∙ 10
F [3,61; 3,73; 3,88] ∙ 10
G [1,09; 1,14; 1,20] ∙ 10
The probability of failure of the whole substation is
calculated using Eqs. (6)-(9). The unavailability of the whole
substations is:
= [3,13; 3,26; 3,40] ∙ 10 (10)
The results indicate a highly reliable system, with a low
probability of failure. The Energy Not Supplied (ENS) due to
substation failure with DERs implemented is a product of the [6] A. Iliev, M. Zlatanoski, “Reliability Evaluation of Electric
probability of failure of the substation and the energy
supplied.
= ∙ 100 ∙ 8760 =
= [27,42; 28,56; 29,80]MWh (11)
The unavailability of the whole substations without DER [8] B. M. Radhakrishnan, D. Srinivasan, and R. Mehta, “Fuzzy-
included, and the ENS in that case are:
= [3,16; 3,29; 3,43] ∙ 10 (12)
= [27,66; 28,81; 30,07]MWh (13)
The results indicate a slight increase in reliability with
DERs implementation, even without storage system.
V. CONCLUSION
The power system needs to maintain flexibility in order to
be able to respond quickly to the shifts in power load and
power supply. That is especially important when there are
DERs incorporated in the distributive networks.
Although DERs improve reliability of power supply of the
standard power systems, their dependency on the weather
conditions may have an impact on the frequency stability.
Therefore, it is important to estimate the probability of failure
of the equipment installed in the power substations in the
design process.
This paper shows that the fuzzy logic method can be used
for reliability estimation of power substations with DERs
implemented.
REFERENCES
[1] M. B. Ndawula, S. Z. Djokic, and I. Hernando-Gil, ”Reliability
Enhancement in Power Networks under Uncertainty from
Distributed Energy Resources”, Energies, 2019.
[2] Working Group A3.06. Final Report of the 2004-2007
International Enquiry on Reliability of High Voltage Equipment.
CIGRE, 2012.
[3] J. Nahman, Fuzzy Logic Based Network Reliability Evaluation,
Belgrade, Republic of Yugoslavia, Faculty of Electrical
Engineering, 1997.
[4] H.-J. Zimmermann, Fuzzy Set Theory-and its Applications,
Fourth Edition, New York: Springer, Science+Business Media,
2001.
[5] Y. Zhang, J. Wang, and Z. Li, Uncertainty Modeling of
Distributed Energy Resources: Techniques and Challenges,
Springer, 2019.
Substations Considering Uncertainty of the Input Data”,
Conference Proceedings of the Fourth MakoCigre Conference,
Paper No. B3-01, Ohrid, 26-29 September, 2004.
[7] A. Metia, S. Ghosh, “Fuzzy Based DG Allocation for Loss
Minimization in a Radial Distribution System”, Balkan Journal of
Electrical &Compoter Engineering, pp. 115-123, 2015.
Based Multi-Agent System for Distributed Energy Management
in Smart Grid”, International Journal of Uncertainty, Fuzziness
and Knowledge-Based Systems, pp. 781-803, 2016.
126
 Reduction of Energy and Power Losses in Distribution
Network Using Energy Storage Systems
Nikola Krstić
Abstract – Energy storage systems (ESS) are most often used
as a backup, or additional power supply in times of power
shortage, but today, in addition, their role in regulating power
flows is becoming more frequent. Therefore, this paper discusses
the possibility of reducing energy and power losses and
improving the efficiency of the distribution network using ESS.
Optimal charge and discharge power of ESS, as well as their
location in the distribution network will be determined by
metaheuristic optimization methods of particle swarm
optimization (PSO) and genetic algorithm (GA). Standard
IEEE18 distribution network was used as a test network, and the
obtained results of computer simulations done in Matlab
software package, presented at the end of the paper, were used
for verification.
Keywords – Energy storage system, Particle swarm
optimization, Genetic algorithm, Distribution network losses.
I. INTRODUCTION
One of the bigest problems of power system and
distribution network within, is the univen load diagram.
Adequate regulation of power consumption can be used to
align the load diagram, which results in the release of part of
network capacity and the reduction of network losses.
Regulation of power consumption and thus the active power
flow in the distribution network can be done, among other
things, by using ESS, that, depending on the mode of
operation, can be viewed as generators or as consumers.
General idea in this paper is to charge ESS in time of low
consumption, when it will act as a consumer, and to do their
discharge in time of high consumption, when it will act as a
distributed generator. This would somewhat equalize the
power across the network branches over time and relieve
certain branches, that would to some extent reduce losses. In
order for this to be possible, it is very important to determine
the optimal location of ESS [1-3] and its power over time,
whichwill be the primary goal of this paper. Level to which
power and energy losses in distributed network can be
reduced byconnectiong ESSwill largely depend on the
efficiency and type of policy of charge/discharge process,
maximum power of the ESS, as well as its total energy
capacity [4-7]. Also, in this case, the shape of load diagram
plays a crucial role in the reduction of losses in distribution
network. In particular, the paper will consider two types of
load diagrams, connection of one, twoandthree ESS,
wherefour different efficiency levels and four different
N. Krstić is with the Department of Power Engineering, Faculty of
Electronic Engineering, University of Niš, Aleksandra Medvedeva
14, 18000 Niš, Serbia, E-mail: nikola.krstic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
policies of charge/discharge process of ESS will be observed.
Metaheuristic optimization methods of PSO and GA [8] will
be used to find optimal location and operating power of ESS.
Steady state will be determined by backward/forward sweep
method [9] for the calculation of power flows in radial
distribution networks. Load in the network will be of the
constant power type, because that is the most critical type of
load in terms of maintaining an acceptable voltage profile in
the event of change in power flows in the network.Special
contribution of this paper can be seen in measuring the impact
of efficiency of charge/discharge process, state of charge
policies of ESS and the shape of load diagram onthe losses in
distribution network. To measure how many losses can be
theoreticaly reduced using ESS, this paper does not consider
the cost of instalation and size of ESS in finding the optimal
solution, as can most often be seen in literature. Also,
proposed methodology is simple and determines the locations
and operating powers of ESS in the same step, in relation to
the two-step procedure most often used.
II. OPTIMIZATION PROBLEM DEFINITION
Optimal location and power of charge and discharge of
ESS, is nonlinear optimization problem with constraints.
Control variables in this optimization problem are the
locationsof ESS in the network (network node) and their
powers during operation. Constraints of the control variables
are given by Eqs. (1) and (2):
< < , (1)
where are: - power of ESS, - maximum charging
power of ESS, - maximum discharging power of ESS.
∈ , ,... , (2)
where are: - network node index in which ESS is placed,
, ,... - set of network node indexes in which
placement of ESS is possible.
In this optimization problem, dependent variables are node
voltages, currents in network branches and the state of charge
of ESS. Constraints of these variables are given by Eqs. (3),
(4) and (5):
< < , (3)
is the network node voltage,
< , (4)
is the network branch current,
127
 < < , (5)
where SOC represents the state of charge of ESS.
In this paper, load and ESS powers are constant in one-hour
time period (ΔT), based on what, state of charge of ESS at the
endof i-th hour in a day is determined by Eq. (7) in the case of
charging, and by Eq. (8) in the case of discharging:
= ,
= + ⋅ ⋅
Based on Eq. (12), it can be concluded that if load diagrams
are with the same mean power (same energy consumption),
the one with less variance and smaller differences in load
power (flatter load diagram) will have lower energy losses.
These theoretical considerations can be illustrated by a
simple example of a uniformly loaded power line which
power changes in time.In this case, the power distributions
(6) along the power line during the maximum and minimum load
in the case of connection of ESS and without it, are given in
, (7) Figs. 1 and 2, respectively.

= + ⋅ ⋅ , (8)

where are: - state of charge of ESS at a begining,

- state of charge of ESS at the end of the i-th hour,
- power of ESS, - total energy capacity of ESS,
- efficiency of the charge/discharge process.
Based on Eqs. (7) and (8) it is clear that the charging power
of ESS is greater, and the discharging power is less than zero.
In addition to the constraints in the form of inequality, there
are also constraints in the form of equality, which applay to Fig. 1. Power distribution during the maximum load
power balance equations and determine the steady state. These
constraints can be expressed as:

= + , (9)

where are: - generated active power (comes from network,

or ESS that work in generation mode), - active power
consumed (used by allloads, counting ESS in charging mode),
- losses of active power.
= + , (10) Fig. 2. Power distribution during the minimum load
where are: - generated reactive power, - reactive power
consumed by loads, - losses of reactive power. In this Assuming that the maximum load is twice the minimum,
paper, connected ESS are presented as constant power loads, with the same duration times, and charging and discharging
whose power can be negative (generator mode) or positive processes are without losses, equalizing the first derivative of
(consumer mode). energy losses over location and power of ESS to zero, the
Two-parameter criterion function given by Eq. (11), that optimal location and power of ESS are obtained:
takes into account not only thetotal losses of energy, in a
certain period of time, but also the amount of maximum = , (13)
power losses in the network branches, will be used.
= . (14)
= ⋅ + ⋅∑ , (11)
Locating and using ESS acording tothe optimal values from
where are: - criterion function that needs to be minimized,
Eqs. (13) and (14), energy losses can be reduced by 8.89%, in
- energy losses in the distribution network, -
relation to the case without ESS.
maximum power losses in the i-th network branch, , -
weight coefficients.
IV. SIMULATION RESULTS AND DISCUSSION
III. THEORETICAL BASIS OF THE PROPOSED In order to check the efficiency and verification of the
SOLUTION mentioned optimization methods in finding the optimal
solution for the problem of reducing the losses in distribution
Based on characteristic parameters of the load diagram, network using ESS, computer simulations were performed in
energy losses can be calculated from Eq. (12) [10]: the Matlab software package.
= ⋅( + )⋅ , (12) Standard IEEE18 distribution network shown in Fig. 3, was
used asa test network. Paper considers two types of load
where are: - total time duration of power losses, - diagrams, both with the same average load power
average load power, - variance of load power, - (Pm=2679 kW), but with different load factors (m1 =0.788,
proportionality coefficient between the power losses and the m2=0.669), that are shown on the Figs. 4 and 5, respectively.
square of load power.

128

Fig. 3. IEEE18 distribution network
Fig. 4. First load diagram
Fig. 5. Second load diagram
Time dependence of active and reactive load power of each
node is the same, and follows the time dependence of the
consumption given on the load diagram.
Maximum charge/discharge power and total energy
capacity of ESS are selected in such a way that they do not
represent a limiting factor for achieving the optimal solution
in any of the observed cases.
As mentioned, paper will consider four different values
forthe efficiency of charge/discharge process of ESS, that
are: = 0.7, = 0.8, = 0.95 and = 1.
Four different policies related to the state of charge, will
also be consider. According to the first, state of charge of ESS
remains the same at the end as it was at the beginning of the
day, in the second it decreases by 0.1, in third by 0.2 (ESS
discharges), while in the fourth it increases by 0.1 (ESS
charges). Total energy capacity of all connected ESS in every
case is 7000 kWh.
Values of acceleration coefficients in PSO, are both equal
to 0.8, while the coefficient of inertia decreases linearly drom
1 to 0.25 depending on the ordinal number of the iteration.
Selection in GA was implemented so that only the best four
129
individuals were allowed to cross, with the probability linearly
decreasing from the first to the fourth. To obtained the
location of ESS, arithmetic crossing was used, while their
powers were inherited from one parent for the first half of the
day, and from the other for the second half. Mutation is
implemented, so that, with a probability of 10% it allows a
random change in a certain range for both the power and
location of ESS.
Values forweight coefficients of energy and power losses
are determined in a way thatenergy losses have twice the
priority of the sum of maximum power losses in the network
( =0.05 $/kWh, =0.3 $/kW).
Table I contains the results related to the operation of the
distribution network without ESS.These results include, the
total energy losses during the day, sum ofmaximum power
losses in the network branches and the value of the criterion
function (cost function) for both load diagrams.
TABLE I
REPRESENTATIVESIZES OF NETWORK OPERATION WITHOUT ESS
Diagram 1 2
Wloss(kWh) 3520.797 3983.784
ΣPlossmax(kW) 238.648 350.007
a1⋅Wloss ($) 176.04 199.19
a2⋅ΣPlosmax($) 71.59 105.00
Criterion function ($) 247.63 304.19
Connection of one, two and three ESS is considered and
implemented, butit tourned out that for this test network the
best results are obtained in the case of two connected ESS,
and they will be shown.
Figs. 6 and 7 show the total power of connected ESS and
their average state of charge during the day, for the first and
second load diagram, respectively. In this case, the state of
charge at the beginning and end of the day is the same, with
the efficiency of the charge/discharge process of η=0.95.
Fig. 6. Power and state of charge of ESS (first load diagram)
In Figs. 6 and 7 it can be seen that power of ESS in a large
extent followed the patern in which it was positive (charging
mode of ESS) when load was below average and negative
when load was above average (discharging mode of ESS).This
applies to both optimization methods, but in the case of GA it
is more pronounced, which can be seen in somewhat greater
operating powers of ESS and higher oscillations in the state of
charge of ESS compared to the those in the case of PSO.

Fig. 7. Power and state of chargeof ESS (second load diagram)
Table II shows the values of representative sizes of criterion
function and the locations of ESS, for the case of two
connected ESS, obtained using both optimization methods.
successfully solved. Based on the obtained results, it can be
concluded that GA proved to be more successful in finding a
solution to a given optimization problem than PSO. This
applies to both energy losses and maximum power losses
reduction. Better results were achieved in the case of
connection of more than one ESS and with more uneven load
diagram. Although applied on the passive distribution
network, methodology used is universal and can be also
applied in the presence of distributed generation, if their
generation powers like load powers are known in advance.
Disadvantage of used methodology in high dependency of its
efficiency on the accuracy of the forecast of injected powers,
could be greatly overcome by an additional degree of control
that would generate adequate management actions of ESS in
case of deviations of actual and forecasted values, which
could be a topis for further research.

TABLE II REFERENCES
REPRESENTATIVE SIZES OF NETWORK OPERATION WITH TWO ESS
[1] R. Jin, J. Song, J. Liu, W. Li, and C. Lu, “Location and
diagram 1 2
Capacity Optimization of Distributed Energy Storage Systems
Optim. method PSO GA PSO GA in Peak-Shaving”, Energies, January 2020, vol. 13, no. 3, 513
Locations 1, 6 1, 5 6,21 6, 23 DOI: 10.3390/en13030513
Wloss(kWh) 3498.1 3476.3 3818.5 3713.2 [2] H. Zhu, F. Tian, Y. Li, K. Deng, Y. Feng, and X. Tang,
ΣPlossmax(kW) 217.32 195.09 290.74 226.25 “Distributed Energy Storage Optimization Location Method
a1⋅Wloss ($) 174.90 173.81 190.92 185.66 Based on Improved Voltage Distribution”, ICAESEE, 2019,
IOP Conf. Series: Earth and Environmental Science, vol. 446
a2⋅ΣPlosmax($) 65.20 58.53 87.22 67.87
042072, 2020.
Crit. function($) 240.10 232.34 278.14 253.53 [3] C. Kanti Das, O. Bass, G. Kothapalli, T. Mahmoud, and D.
Habibi, “Optimal Placement of Distributed Energy Storage
Although they are not shown in table, the results indicate a Systems in Distribution Networks Using Artificial Bee Colony
large dependence of energy losses and maximum power losses Algorithm”, Applied Energy, October 2018.
reduction on the efficiency of the charge/discharge process od [4] T. Mahmoud, C. K. Das, “An Optimal Allocation and Sizing
ESS. Value of the criterion function (cost) goes from $244.89 Strategy of Distributed Energy Storage Systems to Improve
to $221.77, for the first, and from $274.15 to $244.41, for the Performance of Distribution Networks”, Journal of Energy
second load diagram when the efficiency changes from 0.7 to Storage, vol. 26, December 2019, DOI:
10.1016/j.est.2019.100847
1.This goes to the point, that for efficiency below 0.8 it is no
[5] M. Nick, M. Hohmann, R. Cherkaoui, and M. Paolone,
longer possible to achieve a reduction in energy lossses “Optimal Location and Sizing of Distributed Storage Systems
compared to the mode without ESS.Also, network losses in Active Distribution Networks”, IEEE Grenoble Conference,
reduction has shown to be very sensitive to increasing or 2013, DOI: 10.1109/ptc.2013.6652514
decreasing the state of charge of ESS at end of the day [6] M. Stecca, L. Remirez-Elizondo, T. Soeiro, P. Bauer, and P.
compared to those at the beginning. Value of the criterion Palensky, “A Comprehensive Review of the Integration of
function (cost) ranges from $239.33 to $217.59 for the first Battery Energy Storage Systems into Distribution Networks”,
and from $267.69 to $240.35 for the second load diagram, at IEEE Open Journal of the Industrial Electronic Society, March
10 % recharge and 20% discharge relative to the full state of 2020, pp. 46-65, DOI: 10.1109/OJIES.2020.2981832
[7] X. Tang, K. Deng, Q. Wu, and Y. Feng, “Optimal Location and
charge. Locations of ESS in the distribution network were
Capacity of the Distributed Energy Storage System in a
most often nodes in the middle of the first and second major Distributed Network”, IEEE Access, January 2020, vol. 8, pp.
branch in the network (nodes 5, 6, 21, 23). Power of 15576-15585, DOI: 10.1109/ACCESS.2020.2967435
connected ESS followed theoretical predictions, it was [8] J. Radosavljević, Metaheuristic Optimization in Power
positive (load mode) during low loads, and negative Engineering, London, Institution of Engineering and
(generator mode) during high loads. Technology, 2018.
[9] J. A. Michline, S. Ganesh, “Power Flow Analysis for Radial
V. CONCLUSION Distribution System Using Backward/Forward Sweep Method”,
International Journal of Electrical and Computer Engineering,
vol. 8, no. 10, pp. 1628-1632, 2014.
In this paper, the optimization problem of optimal power [10] N. Rajaković, D. Tasić, and G. Savanović, Distribution and
and location of ESS in the standard IEEE18 distribution Industrial Networks, Belgrade, Academic Mind, 2004, pp. 152-
network, in order to reduce energy and power losses, is 164 (in Serbian).

130
ENERGY SYSTEMS AND EFFICIENCY II
 Design and Regulation of the HF Inverter for Contactless
Power Transfer Systems
Nikolay Madzharov1 and Desislav Iliev2
Abstract – Resonant inverters with width regulation of the duration t0, i.e. its width regulation is realized by control pulse
power supply have been used many years and have been applied dephasing. The wave-forms of the inverter alternative current
in the High Frequency (HF) inverters in different applications. (AC) are shown in Fig. 2: f) natural commutation mode
Since they use the HF transistors (MOSFETs, IGBTs, SiC FETs (NCM); g) slave commutation mode (SCM). The working
and GaN) in their circuit the resonant inverters receive a new
sequence and the formation of AC in the two modes are
development, from an applied point of view. Their utilization in
regulable and contactless systems is perspective. The circuit explained by the marked symbols of the together working
realization of the Resonant Inverters is a half or full bridge with couple devices on the wave-forms Fig. 2.
free-wheeling diodes and specific switching algorithm of the
control system, which are natural or slave commutations. A
procedure for design and analysis of these inverters is proposed,
characterized by simplification, natural purity and reliability.
The procedure is a part of computerized educative-scientific
program. All kinds of resonant inverters are included in the case
of powering contactless power transfer systems (CPTS).

Keywords – Contactless power transfer systems, High

frequency inverter, Specific switching algorithm, Resonant
inverter.
Fig.1. HF resonant inverter powering contactless power transfer
system
I. INTRODUCTION

There are a number of principles for controlling the power

of HF (High Frequency) inverters [1-4]. One of them and
probably the oldest, is the method of unclear width regulation
of the DC input voltage in the inverters [1-2], [4-5]. It is based
on the control frequency half period divided into two intervals
- 0÷λ, λ÷π. The equivalent DC voltage is 0 or a part of its
rated value in the first part and in the second it is equal to the
total supply voltage. The usage of this principle gives definite
regulating possibilities of the inverters. At the same time some
important disadvantage are found. The most important of
them is the inverter insufficient commutation stability and
increased losses when the load has highly variable parameters,
such as contactless transmitters.
With the usage of the advanced HF transistors the principle
of width regulation of the power supply is applied in bridge
inverter with free-wheeling diodes (Fig. 1) [2], [4-5]. The
regulation power supply is made easier here by the dephasing
of control pulses of the together working devices as it is
shown in Fig. 2 a)-d). When there is a control delay t0 of the
switch VT3 toward the switch VT2 and of the switch VT4
toward the switch VT1 then to the inverter AC circuit
(between the points a and b) the transformed voltage Eab is
fed, as it is shown Fig. 2 e). It is with cut heard front with

1Nikolay Madzharov is with Technical University of Gabrovo,

BG, Faculty of Electrical Engineering, H. Dimitar Str. 4,

5300Gabrovo, Bulgaria, E-mail: madjarov@tugab.bg Fig. 2. The HF inverter basic wave-forms a)-d) control pulses wave-
2Desislav Iliev is with Technical University of Gabrovo, BG, forms; e) wave-forms of the transformed voltage; f) current wave-
Faculty of Electrical Engineering, H. Dimitar Str. 4, 5300 Gabrovo, forms with natural commutation mode (NCM); g) current wave-
Bulgaria, E-mail: deso_1975@abv.bg forms with slave commutation mode (SCM)

978-1-7281-7142-5/20/$31.00 ©2020 European Union 133

 In order for the treated inverters to work normally, here it is
mandatory that the current of the respective couple devices
have a resonant character with fixed frequency ωn of the
natural oscillations. It must be underlined that in base of the
procedure for analysis and design suggested here is the
determination of that frequency ωn and the values of the
inverter circuit elements to guarantee it.
II. EXPOSITION
In order to show the complete procedure of the analysis and
design, some main parameters will be presented here, because
they are important parts of its construction.
A. Equivalent Circuit of the Inverter and Contactless Power
Transfer Systems.
For analysis and design of whole system it is equally
correct to use a parallel or serial equivalent circuit (EC) [6-8].
Of course, this refers to Contactless Power Transfer System
(CPTS) with well-defined parameters (geometric dimensions,
air gap, electrical parameters and etc.) [8-10]. It should be
emphasized that both cases compromise with the accuracy of
the description of electrical magnetic processes, but the
resulting inaccuracy is acceptable. Nevertheless, the question
remains whether the two ECs are equal? The theoretical
results of the mode, quantities and values of the natural
oscillation frequency and the elements don't coincide with the
practical ones in parallel EC. There is no difference when
using the EC series [9]. The investigations show that these
problems will disappear completely if the value of the
inductance of CPTS - Lc is calculated in the EC series, instead
of the one calculated for the parallel EC. The conclusion made
in [2,5,11] is that in the same inverter alternating current
circuit, only the obtained value of the CPTS inductance Lc in
the EC series provide the set operating mode and the inverter
parameters. This means that only the calculations made by the
EC series correctly represent the processes occurring in the
inverter. In accordance with the above, the EC of the inverter
between two switching transistors should be considered as a
series circuit Lc-Ce-Re, independently of the type of
alternating current circuit. A supply voltage Ea is applied to
the equivalent circuit (Fig. 3).
Fig. 3. Equivalent circuit in the operating range of the transistors
B. The Phase Angle δ in the AC Circuit - The Angle of
Deviation of the Alternative Voltage Ug and the First
Harmonic of the AC Circuit of the Inverter
It is accepted that the analysis and design of the inverter can
be done separately for each individual case. However, this
complicates their research and comparative evaluation and
134
does not give clear notion of the general parameters of the
inverter, determining its actual properties. It has been proven
that the phase angle δ of the AC circuit is the main parameter
for these inverters regardless of their circuit, operating mode
and CPTS parameters. The same value and waveforms of the
alternative voltage Ug are obtained at a constant angle δ and a
constant value of the inductance Lc is obtained. This confirms
the connection between the parameters of alternating current
and direct current, regardless of the differences in the circuit
and the algorithm.
C. Natural Oscillation Frequency ωn in Inverter
The correct answer of the question – what is the meaning of
“natural oscillation frequency” can be given on the base of the
accurate analysis of the inverter. Because of the different
natural oscillation frequencies, the characteristics equations
and their roots can be calculated.
In the more common case the natural oscillation,
summarizing geometrically independence on their number,
amplitudes, phases and commutation frequency of the devices,
determine the consequential oscillation. The characteristics
and evaluation for this general oscillation in [2,5] are made for
the input current of the inverter.
The frequency of natural oscillation of the inverter ωn can
be determined by its waveforms located in the half-cycle of
the control frequency. The ratio between this ωn and the
commutation frequency ω (ωn/ω) can be found and can be
used as a basis for characterizing the different modes of the
inverter, as well as for determining its general and specific
circuit parameters. For this ratio ωn/ω the main correlations
are correct [2]:
- for the advisable mode; (1)
for the unsuitable mode. (2)
Therefore, the frequency ωn must be determined by the
wave-form of the input current during the operating period of
the switching devices. It does not depend on the inverter
circuit and the number of integral differential equations
describing the electrical magnetic processes in the inverter.
D. Calculation Procedure of RI (Resonant Inverter) width
Adjustment
In accordance with the forms in Fig. 2 e) and f) for the
natural oscillation half period Tn/2 and the frequency ratio
ωn/ω the equations are obtained:
 ;
NCM , (3)
 2 2 ;
SCM . (4)
 The half-period Tn/2 and the ratio ωn/ω can be determined if Table I is made with the following input data: P = 1000W;
the values of τ1 and τ2 are given depending on the known f = 20kHz; E = 300V; cosφLOAD = 0.25. The computer
deviation from the control pulses, for example simulations are made using LTSPICE software. The wave-
τ1=0.5t0/2≈ 0.25t0 and τ1=0.5t0. forms obtained by simulations are shown in Fig. 4 - AC
According to the harmonic analysis variant used, the alternative inverter output current for NCM and in Fig. 5 - the
alternative voltage of the inverter is assumed to be a pure sine same quantity for SCM. The values of the HF inverter
wave. The AC is not perfect, i.e. the current pulse is received elements and some characterizing quantities for the work of
with its real waveform. In accordance with these assumptions, the inverter are obtained by calculations and shown in
it is easy to prove that the inverter current is described by the Table I.
equation:
TABLE I
THE INVERTER ELEMENTS VALUES AND SOME QUANTITIES
(5) C, Lc, Vgm, Io,
Elements RI,Ω L1,μH
nF μH V A
and the amplitude of the alternative voltage is:
644 3,3
, (6) SCM 12,9 398 201 578
600 3,7
644 3,3
where and φ1 - phase of the first harmonic of NCM 13 401 207 450
the inverter AC. 600 3,8
In [2,4] for the border operation mode of the inverter an
equation is worked out: When determining the intervals τ1 and τ2, in practice, the
absolute reliability of the quantities is obtained from the
0,406 0,165 0,188 . (7) calculations (up line, Table I) and from a computer simulation
The design is made when the fundamental equation kept: (under line, Table I). Accepted with minimal or no repetitions.
When changing the input data of the inverter, the contactless
; 1,2 1,5 . (8) power supply system and the load parameters, similar data
and experimental results are obtained.
As a result, the following sequence of the calculations can
be proposed:
1) For the relevant mode next is determined: half period
Tn/2 and frequency ratio ωn/ω from Eq. (3) or (4);
2) Angle δ from Eq. (8);
3) Angle φ1 from Eq. (7);
4) Voltage UGm from Eq. (6).
5) The known parameters of power, frequency and load
determine the capacitance C of the primary coil CPTS. The
condition here is to guarantee the calculated angle δ.
6) The equivalent real Re and reactive (capacitance) Xce
resistances of the tank load are determined in the series
equivalent circuit (Fig. 3). Fig. 4. AC alternative inverter output current for NCM
7) The value of the inductance Lc is determined by the
known formula for the natural oscillation frequency of the
equivalent LC–CE–RE circuit (Fig. 3) of the inverter in the
operating range of SD:
8) The currents of transistors and free-wheeling diodes and
average value of the input current of the inverter I0 are
calculated.
The criterion for the accuracy of the calculations is the
equation between the current I0 and its theoretical value
determined by: I0 T = P/E.
Fig. 5. AC alternative inverter output current for SCM
III. COMPUTER SIMULATIONS
The reliable and efficient operation of the prototype unit
The described procedure can be used with small used in contactless power supply was ensured by exhaustive
modifications for the design of all high frequency AC circuit theoretical study and compliance with the requirements for
and the matching method. The calculating example shown in control pulses and element using SiC MOSFET or GaN.

135
 IV. CONCLUSIONS
An integrated classification of the RI (Resonant Inverters)
switching mode is proposed using the ratio ωn/ω between the
frequency of natural oscillation of the inverter ωn and the
control frequency ω.
The procedure for analysis and design of the resonant
inverter, where it is mainly ensured that the natural frequency
has approximate accuracy, is presented. This has been
confirmed by computer simulations and real experiments.
It has been proven that RI with free-wheeling diodes
operates in the "quasi-boundary" mode, and the difference of
RI without such diodes is the frequency ratio of ωn/ω>1 and
ωn/ω< 1.
For RI analysis and design with supply voltage width
regulation, it is sufficient to define the circuit elements to
create the necessary conditions for the oscillating process. It
has a fixed frequency ωn and a frequency ratio ωn/ω during the
operating period of switching devices.
ACKNOWLEDGEMENT
The carried out research is realized in the frame of the OP
“Scienceand Education for Smart Growth 2014–2020”,
Project CoC “Smart Mechatronic, Eco- and Energy Saving
Systems And Technologies“, №BG05M2OP001-1.002-0023.
REFERENCES
[1] G. Kraev, N. Hinov, and D. Arnaudov, "Multiphase DC-DC
Converter with Improved Characteristics for Charging
Supercapacitors and Capacitors with Large Capacitance",
Annual Journal of Electronics, V6, B1, pp.128-131, 2012.
136
[2] T. S. Todorov, N. D. Madzharov, D. T. Alexiev, and P. T.
Ivanov, Autonomous Inverters, Gabrovo, 1996.
[3] I. Nedyalkov, D. Arnaudov, and N. Hinov,"Modelling of a Bi-
Directional Converter from a Power Supplying System with
Application", PCIM Europe 2018; International Exhibition and
Conference for Power Electronics, Intelligent Motion,
Renewable Energy and Energy Management.
[4] A. Vuchev, T. Grigorova, and Y. Madankov, "Static
Characteristics of a Phase-Shift Controlled Series Resonant DC-
DC Converter", XXVII International Scientific Conference
Electronics – ET2018, 13-15 September 2018, Sozopol,
Bulgaria, pp. 253-256.
[5] G. Kraev, N. Hinov, N. Gradinarov, and D. Arnaudov, "Study
of Parallel Single Transistor ZVS Resonant Inverter with Direct
Settle of Steady Mode", International Exhibition
and Conference for Power Electronics, Intelligent Motion,
Renewable Energy and Energy Management, PCIM Europe
2013, Nuremberg, Germany.
[6] G. Covic, J. Boys, "Inductive Power Transfer", Proceedings of
the IEEE, vol. 101, no. 6, pp. 1276-1289, June 2013.
[7] D. Arnaudov, N. Madzharov, and N. Hinov, "Rotary Wireless
Inductive Transmitter Powered by ZVS Resonant Convertor",
Elektronika Ir Elektrotechnika, vol. 25, no. 4, 2019.
[8] H. Daskalova, "On the Choice of Wireless Data Transmission
Systems", Compendium Modeling and Management of
Information Processes, Sofia, Bulgaria, 2009, pp. 18-29.
[9] V. Rankovska, "Using Wireless Interfaces in a Smart Home
Model", ICEST 2016 Conference Proceedings, Ohrid,
Macedonia, pp. 325-328.
[10] A. Van den Bossche, V. Valchev, D. Van de Sype, and L.
Vandenbossche, "Ferrite Losses of Cores with Square Wave
Voltage and DC Bias", Journal of Applied Physics, vol. 99, no.
8, 15 April, 2006.
[11] W. Choi, W. Ho, X. Liu, and S. Hui, "Bidirectional
Communication Techniques for Wireless Battery Charging
Systems and Portable Consumer Electronics", IEEE Conf.
Applied Power Electron., Palm Springs, CA, 2010, pp. 2251-
2257.
 Analysis of Electromagnetic Characteristics
of Asynchronous Induction Machine
Igor Sterjovski, Blagoja Arapinoski, Metodija Atanasovski, Vesna Ceselkovska
Abstract – Electromagnetic field analysis is important for the TABLE I
design and operation of induction machines. The electromagnetic GENERAL DATA FOR INDUCTION MACHINE
field of waves contains information about the position of the
General data Value
stator, rotor and the mechanical parameters of IM. The aim of
Rated Output Power (kW) 0.550
this research is to show the type of electromagnetic analysis of a
three-phase asynchronous generator with a wound rotor through Rated Power Factor 1
a concrete example. The example is contained in the ANSYS Frequency (Hz) 50
Maxwell Finite Element Analysis software tool. The results of the Rated Voltage (V) 380
analysis including their graphical representations are presented Rated Speed (rpm) 1500
in the paper. Operating Temperature (C) 75
Number of poles (p) 4
Keywords – Аsynchronous induction machine, ANSYS Winding Connection Y
Maxwell, Electromagnetic analysis, Finite element method.

I. INTRODUCTION

Although synchronous generators with permanent magnets

occupy a significant place in wind power conversion,
asynchronous wound rotor generators have an even more
significant place than previous generators. If an asynchronous
generator is used with a wound rotor then the ability to control
the generator through the rotor circuit is opened. The main
advantage of an asynchronous generator with a wound rotor is
the possibility to obtain it from output power greater than the
nominal, without overheating the machine. These generators
have the ability to produce maximum power over a wide
range of speed variations. Asynchronous generators with a
wound rotor together with asynchronous cage generators rotor
find special application in wind turbines whose power is
within the frame of MW.
For the purpose of the research, the asynchronous wound
rotor generator is chosen from the software’s library presented
in Fig. 1. The value of the rated output power and the
frequency are fixed, therefore the input data for the analysis
are power factor, voltage and speed. The nominal data of the
induction machine is shown in the Table I.
This example shows how electromagnetic analysis can be
performed in a relatively fast and accurate manner in a 3D
transient mode. The commissioning of the asynchronous
generator powered by a wind turbine has been analyzed. The
results for magnetic induction, flux, torque and current values
in stator windings for analyzed time of 75 ms are given.
Igor Sterjovski, Blagoja Arapinoski and Metodija Atanasovski
are with the Faculty of Technical Sciences-Bitola, St. Kliment
Ohridski University, Republic of Macedonia, E-mail:
sterjovskiigor@gmail.com; blagoja.arapinoski@tfb.uklo.edu.mk;
metodija.atanasovski@tfb.uklo.edu.mk;
vesna.ceselkoska@uklo.edu.mk
Fig. 1. Geometry of asynchronous wound rotor generator
The paper is organized as follows. After the introduction,
the basic definitions of 3D transient (time domain) are given
in Section II. Section III gives data overview and presents the
electromagnetic field analysis. Section IV concludes the
paper.
II. ANSOFT MAXWELL’S FORMULATION FOR
3D ELECTROMAGNETIC TRANSIENT ANALYSIS
Maxwell's equations in the moving and the stationary part
include fixed coordinate system and particular convention for
simpler formulation where motion is involved [1]. For the
rotational type of motion, cylindrical coordinate system is
used. This coordinate system has z-axis aligned with the
actual rotation axis. Thus, the motion term is eliminated for
the translational type of motion [1].
The formulation used by the Maxwell transient module is
strongly coupled with the finite element part and also coupled
with the mechanical elements [1]. In this case, the
electromagnetic force/torque is calculated using the virtual
work approach.
The following three Maxwell's equations are relevant for
transient (low frequency) applications [1]:

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 137

 ∇xH = σ ( E ) , (1) Therefore, for the 3D transient solver, many quantities will
be automatically calculated and displayed as a 2D graph (a
∂( E ) function of time): voltage (current), flux linkage, back emf.
∇xE = , (2)
∂t Other global variables can also be calculated by the 3D
transient solver and displayed as 2D graphs, such as power
∇xB = 0 , (3) loss, core loss, stranding loss, mechanical power such as
where H represents magnetic field, E represents electric field force/torque, speed and displacement [1].
and B stands for the magnetic induction.
The following two equations directly result from the above III. 3D ELECTROMAGNETIC TRANSIENT ANALYSIS
equations: OF АSYNCHRONOUS INDUCTION MACHINE
1 ∂B
∇x ∇xH + = 0, (4)
σ ∂t

∇xB = 0 . (5)

The final result is a formulation where vector fields are

represented by first order edge elements and scalar fields are
represented by second order nodal unknowns.
Field equations are coupled with circuit equations for both
solid and stranded conductors because, in the case of applied
voltage supplies, the currents are unknown. For the case of
voltage driven solid conductors, the following equation is
used to account for theohmic drop across the i-th conductor Fig. 2. Mesh of finite elements of asynchronous machine with
loop [1]: wound rotor in 3D geometry

VRi =  J
RC ( i )
0i ( E + v x B) dR ,
where J0i represents the current density. The current density
J0i corresponds to 1A of net current in loop i and vanishes
outside loop i. The induced voltage can be derived from the
following equation [1]:
ei = −  H i ⋅ BdR .
RC ( i )
Here, integration is performed on the entire conductor area.
Stranded conductors are considered not to induce eddy
currents therefore they are placed in non-conductive areas.
This means that in order to calculate the ohmic voltage drop,
we cannot use the same process as a solid conductor. Instead,
we use lumped parameters to represent the DC resistance of
the windings. We obtained the induced voltage due to the total
flux linkage in a similar way to the solid conductor. In both
cases, an external inductor and capacitor can be added. For
time discretization, use the backward time stepping scheme
[1]:
 dx 
t + Δt
(6) As mentioned above, the tool used in designing and
obtaining the results in this research is method. Finite element
method is a powerful method that allows difficult to solve
complex differential equations that describe Ansoft - Maxwell
software that has the well-known finite element electrical
machines in such analysis to be reduced to a large number of
easily solvable equations, and the processing speed of modern
computers allows them to be solved relatively quickly. The
(7)
applied software allows entering the geometry of the problem
under consideration, both in terms of dimensions and in terms
of the construction materials of individual engine elements
[2-3], [5]. Depending on the requirements for the accuracy of
the calculations themselves, the same software offers the
option to perform the calculations in a plane in 2D, or in 3D
space. In the research all simulation analyzes are made in
three dimensions.
All complex equations are solved by the finite element
method. The software tool creates a mesh of finite elements
that for the corresponding example is shown in Fig. 2 [2].
The accuracy of the result depends on the mesh density. Of
course, this is given by the user and depending on the goal we
want to achieve, the mesh density is adjusted. Thus, after

=
{ x } − { x}
t + Δt t

. (8)
 
 dt  Δt
For the non-linearity allowed in 3D transient applications,
the classic Newton-Raphson algorithm is used. The transient
solver in Maxwell supports coil terminals and winding
definitions. Therefore, the number of coil turns can be
specified in the model, which is necessary for calculating
global quantities with high engineering value such as flux
linkage and coil back emf [1].
creating the mesh of final elements, the solutions are obtained.
The next step after drawing the mesh of finite elements is
the decisive part in which the value of the magnetic vector
potential is calculated at each point of the considered model
according to the following expression given below [4-5]:
А = { Аx , Аy , Аz } . (9)
Determining the distribution of the magnetic field gives
values for the magnetic vector potential that move within the
corresponding limits. When the magnetic vector potential is

138
determined, the densities of the equivalent potential magnetic
force lines are defined, which connect points with equal
potential. The results obtained from the specific solutions can
be displayed in tabular or graphic format.
The distribution of the magnetic induction vector in all
parts of the motor is related to the value of the magnetic
vector potential with the following equation [4-5]:
B = rotA . (10)
All of this is shown in Fig. 3, from which it can be seen to
what extent the magnetic induction moves in the
asynchronous induction machine with a wound rotor that is
subject to analysis. The analyzed part refers to when the
machine is powered by a wind turbine. The generator speed is
1659.95 rpm, the time is 0 seconds, and the position of the
rotor in relation to the old position is 7.5 degrees.
If we analyze it well, from Fig. 3 it can be concluded that
magnetic induction has the highest values in the rotor and
Fig. 3. Distribution of magnetic induction in 3D domain of
asynchronous machine in generator mode driven by wind turbine
Fig. 4. Values of the currents in the stator in the considered period
139
stator slots and in the air gap, because here is the smallest area
through which the magnetic force lines close.
An analysis of the torque can also be performed. For that
purpose, i.e. to determine this characteristic it is necessary to
set the time of analysis of the transient mode. In the
example,it is analyzed up to a time of 75 ms, because it is
approximately the time required to clearly see the stabilization
of the generator. This is shown in Fig. 4, where it is clear that
at the beginning the currents are several times higher, and
after 50 ms, the stabilization process starts and the current in
the stator of the generator gets its nominal value.
Fig. 5 shows a graphic of the torque from the start of the
asynchronous machine up to 75 ms. The picture shows that the
moment has a negative value, which explains that the
asynchronous induction machine is operating in generator
mode.
From the characteristic of Fig. 5 it can be noticed that the
value of the torque is the largest around the 6th ms from the
start of the generator, and that value is a critical moment.
Because then the critical slip has the greatest value and is
proportional to the active resistance of the rotor winding, and
completely depends on impedance of rotor winding. From the
characteristic of Fig. 5 it can be seen that in a time of 50 ms
the so-called stationary mode begin, at that moment the torque
generated by the induction machine is balanced with the static
resistance moment of the drive machine.
Fig. 4 shows that when starting the generator, the initial
values of the currents are significantly higher and behave
unstable until the torque is equal to the static moment of the
drive machine in the case of the wind turbine. Due to the
small value of the rated power of the machine that is 550 W,
the initial values of the currents do not have large values the
ratio is around 2.5In. The rated current of the asynchronous
generator with a wound rotor is 0.85 A. Fig. 6 shows the
distribution of the flux at an arbitrarily selected point in the air
gap, for a different position of the rotor circuit in relation to
the stationary magnetic axis of the stator. The torque in the
asynchronous machine is created because of the rotating
magnetic field with flux Φ and the current in the winding of
the rotor I2 . If the machine is connected to a constant voltage
 Fig. 5. Characteristic of the rotating torque in function of time
Fig. 6. Distribution of the flux in the air gap at an arbitrarily selected point in function of time
network, the flux can be considered constant, and this can be
confirmed from the diagram in Fig. 6 because it is clear that
the flux is moving in a very narrow range. From Fig. 7 it can
be concluded that the torque depends proportionally to the
value of the speed.
Fig. 7. Characteristic of the torque in function of speed
IV. CONCLUSION
The methodology of using the finite element method to
model the three-dimensional magnetic field proposed in this
paper is the basic concept of determining the magnetic field
distribution in the three-phase asynchronous induction
motor. It can be concluded from an analysis that the three-
dimensional calculation of the magnetic field can provide
us with enough data to calculate the electromagnetic
characteristics of the three-phase asynchronous induction
140
motor with wound rotor and analyze its performance
through them. In addition, the data can be used to calculate
and plot the electromechanical characteristics of the motor,
or to optimize the analysis of the motor to find more
effective and improved motors.
REFERENCES
[1] ANSYS/ANSOFT Maxwell 2D, 3D User’s Guide V.14
[2] B. Arapinoski, M. P. Radevska, V. Ceselkoska, and M. Cundev,
"Modeling of Three Dimensional Magnetic Field in Three-
Phase Induction Motor with Double Squirrel Cage", TEM
Journal, vol. 2, no. 1, 2013, pp. 80-82.
[3] B. Arapinoski, M. Cundev, M. P. Radevska, and V. Ceselkoska,
"Electromagnetic Analysis of Three-Phase Asynchronous Motor
with Double Cage Rotor with Finite Element Method in 3D",
Еighth Counseling of the Macedonian National Committee in
Cigre, Ohrid, 2013.
[4] B. Arapinoski, M. P. Radevska, M. Atanasovski, and M.
Spirovski, "Mathematical Model and Calculation of Operating
Characteristics of Three – Phase Asynchronous Motor with
Double Squirrel Cage", 53rd International Scientific Conference
on Information, Communication and Energy Systems and
Technologies, Sozopol, Bulgaria, June 2018, pp. 301-304.
[5] B. Arapinoski, M. P. Radevska, V. Ceselkoska, M. Spirovski,
and M. Atanasovski, "Calculation of parameters of Three-Phase
Induction Motor with Double Squirel Cage with FEM 3D", 51st
International Scientific Conference on Information,
Communication and Energy Systems and Technologies, Ohrid,
Republic of Macedonia, 2016, pp. 355-358.
 Comparison of Methods for Control of
Linear Induction Motors
Bojidar G. Markov1 and Diana G. Markova2
Abstract – This article makes a comparative analysis between
the vector control, and the method of direct electromagnetic
force control of linear induction motors. The dynamic models of
linear induction motors applying the two control methods are
presented. For verification of the effective control of linear
induction motors comparative simulations are made.
Keywords – Linear induction motor, Direct force control, Field
orientated control, Dynamic model.
I. INTRODUCTION
Asynchronous electric drives are the most common type of
electric drive in the industry. For optimal results, different
control methods for induction motors are used. In the end of
the 20th century and the beginning of the 21st century, various
methods for control of machines with alternative currents
have been developed.
In practice, there are two general types for control of
induction motors - open and closed control systems.
The modern open control systems are built on the principle
of U/f speed control, also known scalar control. The closed
control systems are built on the principle of vector control or
direct control of the electromagnetic moment force.
The vector control methods are controlling the flux of the
rotor, and electromagnetic torque of the motor. The
measurement of controlled quantities can be direct and
performed by measuring controlled quantities, or indirectly,
by additional calculations [1].
An alternative method of controlling an induction motor is
the so-called direct torque control. A scheme for direct torque
control is shown in [2]. This method of control significantly
reduces the controlled parameters [3]. In the method for direct
control of the electromagnetic moment, hysteresis zones of
change of the magnetic flux and the moment of the motor
shaft are introduced. If the stator current coupling or torque
goes out of the set hysteresis zones, a signal is sent to the
inverter, which, by changing its output voltage, brings the flux
reduces the controlled parameters [3]. In the method for direct
control of the electromagnetic torque, the hysteresis zones
change the magnetic flux and the torque of the motor. If the
stator current, or torque goes out of the set hysteresis zones,
the signal is sent to the inverter, which, by changing its output
voltage, brings the flux coupling and the electromagnetic
torque to the desired values.
Lately in different sectors of industry, as well as in
everyday life, linear motors are used ever more frequently.
Control of these motors is carried out through methods for
control of alternative current motors. It is therefore reasonable
to research and compare the different contemporary methods
for control of linear induction motors.
The purpose of this article is to create and make the
simulation verification control of linear induction motor with
systems for vector controls and direct force control of the
electromagnetic force.
II. DYNAMIC MODEL OF LINEAR INDUCTION
MOTOR
When we compose the dynamic model of a linear induction
motor, it is necessary to make the following assumptions:
Only the values of the inductor can be measured;
- The influence of end effects is ignored;
- The additional friction forces of the secondary element
are not taken into account;
- The power supply of the inductor of the linear induction
motor is three-phase, symmetrical and balanced;
- The magnetic system of the linear induction motor is not
saturated.
The dynamic model of a linear induction motor is
performed in a d-q system. In [4] and [5] a d-q electric model
(Fig. 1) of a linear induction motor is shown, without taking
account the end effects.
Rs ωeλsq Lls L lr (ωe − ω)λrq
+

coupling and the electromagnetic moment to the desired ird

isd
values. Usd Lm Rr
An alternative method of controlling an induction motor is
the so-called direct torque control. A scheme for direct torque
−
control is shown in [2]. This method of control significantly
Rs ωeλsq Lls L lr (ωe − ω)λrq
+
1Bojidar G. Markov, University of Food Technologies, Technical ird
isd
Department, 26 Maritca Blvd., 4002 Plovdiv, Bulgaria, E-mail: Usq Lm Rr
bojmarkov@abv.bg
2Diana G. Markova, IT Academy, 14 Valko Shopov Str., 4023 −
Fig. 1. Dynamic model of a linear induction motor in d,q system
South Industry zone, Plovdiv, Bulgaria, E-mail: didkata1@abv.bg

978-1-7281-7142-5/20/$31.00©2020 European Union 141

 In Fig. 1 we have the following designations: The electromagnetic force in the vector control is:
Rs - the active resistance of the inductor;
2 π Lm P
Rr - the active resistance of the slider; Fe =
3 τ Lr 2
(
λrd isq . ) (4)
Lm - the magnetizing inductance;
Lls =Ls -Lm - the leakage inductance in the inductor, where The equation of the electric drive of the LIM is:
Ls - inductance of the inductor; 1 dvr
Fe = + FL , (5)
Llr =Lr -Lm - the leakage inductance of the slider, where M dt
Lr - the inductance of the slider; where: FL - the external resistance force; M - mass of the
isd , isq - the inductor currents in d , q axes; motor; P - pole pairs.
u sd , usq - the induction voltages in d , q axes
ω = ( π / τ ) ⋅ vr - angular speed of the slider, where vr are III. DIRECT FORCE CONTROL
linear speed of the slider; τ is the pole pitch; ωe = 2τ f1 -
The scheme according to which the direct force control is to
angular speed of the magnetic field of the inductor, where f1 be carried out is shown in Fig. 2 [6].
is the frequency the power supply of the inductor; λsd , λsq -
fluxes of the inductor in d , q quantities; λrd , λrq - fluxes of λ *s
the slider on d , q quantities; and λm air gap flux. SWITCH
Fe* λ
The dynamic equations describing the electromagnetic TABLE
processes of the LIM in d,q consist of the substitute schemes Fe
shown on fig.1 are as follows:
FLUX AND FORCE
disq 1  R 1−σ  π Lπ L ESTIMATOR
= usq −  s + isq − veisd − m vλrd + m λrq;
dt σLs σLs σTr  τ σLs Lr h σLs LrTr
disd 1  R 1−σ  π Lπ L Fig. 2. Schematic diagram of DFC
= usd −  s + isd + veisq + m vr λrq + m λrd ;(1)
dt σLs σ
 s L σ Tr  τ σ L L
s r h σLs LrTr

dλ rq Lm π 1 Flux and Force Estimator

= i sq − (v e − v r )λ rd − λ rq ; The values for estimating the space vectors of voltages and
dt Tr τ Tr currents in the inductor in the d-q coordinate system for DFC
dλ rd Lm π 1 are as follows:
= i sd + (v e − v r )λ rq − λ rd ,
dt Tr τ Tr 
e  di e
U ds = Rs ise + Ls ds ,
where: Tr =Lr / Rr and σ = 1 − L2m / Ls Lr . dt

The electromagnetic force is expressed by: e e diqse
U qs = Rs iqs + Ls , (6)
2 π Lm P dt
Fe =
3 τ Lr 2
(
λrd isq − λrq isd .) (2)
 
di e diqse
where Ls ds and Ls is the variation of the inductor
The basis of vector control is to keep the flux vector of the dt dt

slider constant, i.e. λr = λrd , and λrq = 0 . In this case the fluxes in d,q. The estimation fluxes for the inductor are
obtained form equations (6), as follows:
dynamic equations of the LIM are:
e e 
λds =  (U ds − Rs ise ) dt ,
di sq 1  R 1−σ  π Lπ
=u sq −  s + isq − v e i sd − m vλrd e e  (7)
dt σLs σ
 sL σT r  τ σL s Lr h λqs =  (U qs − Rs ise ) dt .
disd 1  R 1−σ  π Lm
= u sd −  s + isd + ve isq + λ rd ; Eqs. (6) and (7) show that the fluxes of the inductor can be
dt σLs σ
 sL σT r 
τ σL L T (3)
s r r
estimated by measuring the currents in the inductor along the
Lm π d-q axes.
0= isq − (ve − vr )λrd ; When the values of the stator fluxes and currents are
Tr τ known, the value of the electromagnetic force can be
dλrd Lm 1 calculated by the formula:
= isd − λrd .
dt Tr Tr
Fee =
2τ 2
(
3π P e e e e
iqs λds − ids λqs . ) (8)

142
 Switch Table and Controllers IV. VECTOR CONTROL
Тo obtain the reference current in the inductor, the vector (FOC-FIELD ORIENTATED CONTROL)
modulation of a three-arm inverter, consisting of 6 IGBT
transistors shall be applied. The switching position of the
The FOC in the linear induction motor is realized
IGBT transistors is shown in Fig. 3. In Fig. 3 it is evident that
analogously to the FOC of a rotary induction motor according
there are eight switching positions. Two of them have zero
to the scheme shown in Fig. 4 [7].
voltage and therefore zero current.

ω ref i qs* u ds ua
ub
λ *dr ωr ids* uqs
uc
λ dr

i qs
i ds

1
s

ω sl ωsl ωr π
τ
v

Fig. 3. Inverter output voltage

The presence of voltage at the output of the inverter is

marked with 1, and its absence with 0. If the combinations of Fig. 4. Schematic diagram of LIM with FOC
on and off are marked with S A , S B , and SC , the vector of the
In Fig. 4 the designation ωsl , is the slip frequency, with
inductor voltage can be expressed as follows:
value:
  j
2π
j
4π  1 iqs
2
U s = U dc  S A + S B e 3 + SC e 3 . (9) ωsl = . (10)
3   Tr ids
 
Analyzing the six sectors of the voltage and fluxes shown V. SIMULATION AND VERIFICATION
in Fig. 3, the limits of change of the angles between the
sectors are estimated by the following limits:
For verification of the proposed models for LIM control
−30 < θs1 < 30 ; 30 < θs2 < 90 ; 90 < θs3 < 150 ; systems, LIM with the following technical data is used:
150 < θs4 < 210 ; 210 < θ s1 < 270 ; 270 < θs1 < 330 . U H = 380 V; I H = 400A; f H = 50 Hz; Rs = 0,016 Ω;
The stator fluxes angle in addition to the linear Rr = 0,09 Ω; Lr = 3,2 mH; Ls = 3,5 mH; Lm = 11mH;
electromagnetic force hysteresis has the optimal value of the P = 2; τ = 0,015m; M = 165kg.
state vector when directly controlling the electromagnetic
force of a linear induction motor. Table I shows the possible As a result of the simulation the following results have been
commutations of the vectors of voltage, fluxes and obtained, Figs. 5 and 6.
electromagnetic force in a linear induction motor.
For stator currents
TABLE I DTC: Starting current 1267 A, transient current 538 A at a
POSSIBLE COMMUTATION OF THE VECTORS OF VOLTAGE AND load of 792 kg, Fig. 5. Duration of transient current time
FLUXES 0.05 second, current 400 A. The first peak current is
Position Sector instantaneous at 0.001 second.
λs Fe 1 2 3 4 5 6
FOC: Starting current 1359 A, transient current 543 A at a
θs1 θ s 2 θ s 3 θ s 4 θ s5 θ s6 load of 792 kg, Fig. 6. Transient current time of 0.05 seconds,
set current 400 A. It is noteworthy that the peak current at
0 1 110 010 011 001 101 100 FOC occurs with a delay of 0.05 seconds compared to the
0 111 000 111 000 111 000 DTC.
-1 011 001 101 100 110 010
1 1 100 110 010 011 001 101
0 000 111 000 111 000 111
1 001 101 100 110 010 011

143
144

1400
Stator Current DFC
It can be seen that in FOC the processing of the speed
1200 reference is more accurate than in DFC, where there is a
1000
deviation from the reference, Fig. 6.
800

600

VI. CONCLUSIONS
Is [A]

400

200

-200
From an energy point of view, the DFC method has better
-400 indicators, which can be seen from the obtained results of the
-600
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
measured stator current.
t [s]
From the point of view of slider movement accuracy, the
Stator Current FOC
1400 method of LIM control via FOC gives better indicators.
1200
The research shows that in the system in direct control of
1000
the electromagnetic force, the speed of the slider follows the
800

600
set speed to the system, and the electromagnetic force follows
X: 0.6648
the set load of the electric drive.
Is [A]

Y: 398.8

400

200 Both models for LIM control are equally applicable in

0 terms of the technical degree of complexity.
-200
In the system in direct control of the electromagnetic force,
-400
the speed of the slider follows the set speed to the system, and
-600
0 0.1 0.2 0.3 0.4 0.5
t [s]
0.6 0.7 0.8 0.9 1
the electromagnetic force follows the set load of the electric
Fig. 5. Stator current drive. Both models for LIM control are equally applicable in
terms of the technical degree of complexity.
Speed DFC
1.6

1.4
REFERENCES
1.2

1
[1] S. N. Manias, "Power Electronic and Motor Drive Systems",
0.8
Elsever Inc., 2017.
v [m/s]

0.6
[2] R. P. Aguilera, P. Acuna, G. Konstantinou, S. Vazquez, and J. I.
0.4
Leon, "Control of Power Electronic Converters and Systems",
0.2
Academic press, Elsevier Inc., 2018.
0
[3] S. Gdaim, M. Abdelatiff, and M. Mimouni, "Direct Torque
-0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Control of Induction Machine Based on Intelligent Techniques",
t [s]
International Journal of Computer Applications, vol. 10, no. 8,
Speed FOC
1.6 November 2010.
1.4 [4] E. F. da Silva, E. B. dos Santos, P. C.M. Machado, and M. A.A.
1.2
de Oliveira, "Dynamic Model for Linear Induction Motor", IEEE
International Conference on Industrial Technology, 2003.
1
[5] M. N. Raju, M. S. Rani, "Mathematical Modeling of Linear
v [m/s]

0.8
Induction Motor", International Journal of Engineering &
0.6 Technology, vol. 7 (4.24), pp. 111-114, November 2018.
0.4 [6] M. Cirrincione, M. Pucci, “An MRAS-Based Sensorless High
0.2
Performance Induction Motor Drive with a Predictive Adaptive
Model”, IEEE Transaction on Industrial Electronics, vol. 52, no.
0
0 0.2 0.4 0.6 0.8 1
t [s]
1.2 1.4 1.6 1.8 2
2, pp. 532-551, April 2005.
Fig. 6. Slider speed [7] A. Yousef, A. M. Samir, "Review of Field Orientated Control of
Induction Motor", International Journal for Research in
Emerging Science and Technology, vol. 2, no. 7, pp. 5-16, July
2015.

144
 The Cost of Indecision in Energy Efficiency. A Cost of
Opportunity Analysis for an Industrial Consumer
Ioan-Sevastian Bitir Istrate1, Cristian Gheorghiu2 and Miruna Castravete3
Abstract – Organisations are not always taking the In the 2012 and 2016 Energy Audits, the ESCO identified
recommendations of ISO 50001 group of standards into account and proposed for implementation six EPIAs, summarily
when establishing their internal operational Guidelines and, by presented in Table I.
doing so, they end up neglecting some rather significant Energy
Performance Improvement Actions (EPIAs). In this paper, the
TABLE I
Cost of Indecision generated by a poor Energy Management
PREVIOUS EPIAS IMPLEMENTATION HISTORY
Procedure will be analysed for an industrial consumer. The Cost
of Opportunity Analysis will reveal the financial losses generated Identified Implemented
by not implementing the already identified EPIAs as soon as EPIA
in in
financially possible. Replacement of CFL lighting with
2012 2012-2016
LED lighting
Keywords – Energy efficiency, Energy performance indicators, Energy Measurement System
2012 2014-2018
Energy performance improvement actions, Energy services, development
Energy management. Heat Recovery from the Brick
2012 N/A
Treatment Oven
Water Pumps redesign and
2012 N/A
I. INTRODUCTION Variable Speed Drive Installation
Variable Speed Drive installation
2016 2016-2017
on the ventilation system
Energy efficiency and, implicitly, the environmental impact Refurbishing of the Brick
reduction became a top priority for the European Union, thus 2016 2018
Treatment Oven insulation
engaging every energy sector in a continuous effort to Heat Recovery from the Brick
2016 N/A
improve its specific Energy Performance Indicators (EPIs). Treatment Oven
Water Pumps redesign and
Furthermore, as energy prices continue to rise, the 2016 N/A
Variable Speed Drive Installation
improvement of the aforementioned EPIs can and will lead to
an improvement of the overall profitability of the
As it can be observed, the Organization did not implement
Organization, regardless of the energy sector that it operates
the Heat Recovery EPIA or the Water Pumping System EPIA
in. The importance of developing an Energy Management
as of 2020, as it considered the Capital Expenditures
System, in accordance with ISO 50001 group of standards
(CAPEX) and the Simple Payback Period to be too high.
guidelines should thus be a priority.
The Top Management of the Organization did not approve
In accordance with the en-force legislation [1]-[3], every
the expenditures throughout the analysis period of 12 years.
Organization with an annual equivalent energy consumption
The internal regulations of the Organization state that the
of over 1,000 t.o.e. (tons of oil equivalent) must undergo an
Simple Payback Period required of an EPIA must be lower
Energy Auditing process by an authorized Energy Services
than 1.5 years in order to be deemed viable from an
Company (ESCO) every four years.
economical point of view.
In this paper, the Case Study of an industrial consumer
This paper will continue to analyze the Cost of Indecision
which operates in the bricks manufacturing sub-sector will be
for these two EPIAs by an Opportunity Cost Analysis, in
presented. The analyzed Organization underwent the previous
order to evaluate the real economic impact of the Top
Energy Auditing Process in 2012 and 2016. The main goal of
Management decision making policy. In Sections II and III a
the proposed methodology is to reduce the “efficiency gap”,
brief description of the Water Pumping system redesigning
as defined in [4] and to increase the implementation rate of
EPIA, respectively the Heat Recovery EPIA will be presented.
EPIAs in the industrial sector, in which energy efficiency isn’t
In Section IV the authors will present the Opportunity Cost
usually a top priority [5].
Analysis. Section V will summarize conclusions of the paper.

1Ioan-Sevastian Bitir Istrate is with University POLITEHNICA of II. WATER PUMPING SYSTEM EPIA
Bucharest, Faculty of Power Engineering, Splaiul Independenței 313,
060042 Bucharest, Romania, E-mail: ioanistrate2005@yahoo.com
2Cristian Gheorghiu is with University POLITEHNICA of A. Reference Case
Bucharest, Faculty of Power Engineering, Splaiul Independenței 313,
060042 Bucharest, Romania, E-mail: cristian.gheorghiu@upb.ro The Water Pumping System (WPS), briefly described in
3Miruna Castravete is with University POLITEHNICA of
Table II, is made up of thirteen electrically driven pumps of
Bucharest, Faculty of Power Engineering, Splaiul Independenței 313,
various rated voltages and powers.
060042 Bucharest, Romania, E-mail: mirunacastravete@gmail.com

978-1-7281-7142-5/20/$31.00 ©2020 European Union 145

 Due to a decrease in the water consumption as a result of Pump with a smaller, more efficient one equipped with a
various technological process improvements, the WPS VSD, two sets of calculations were done.
operates at a maximum of 71% of it’s designed, rated load. Firstly, the contribution of the VSD (∆ ) was taken into
As it can be observed from Table II, the Significant Energy account by considering an electricity consumption reduction
Use in the WPS is represented by the MV 403 cold water of 25% of the proposed scenario:
pumps, which have a rated voltage of 6 kV and a rated power 1 ℎ
of 315 kW. Out of the five MV 403 pumps only one is ∆ = , ∙ ∙ ∙( , − , ) (1)
currently in use, and working at an average load of 30% with
an actual average flow of 500 m3/h. 1 ℎ
∆ = 200 ∙ 1,440 ∙ ∙ (1 − 0.7)
TABLE II
WATER PUMPING SYSTEM ℎ ℎ
∆ = 92,407 = 92.41
Electrically Total In Average
Rated
Driven Quantity Installed Operatio load
Power Secondly, the energy economy (∆ ) obtainable by
Pump Power n
[-] [kW] [pcs.] [kW] [pcs.] [%] replacing the old electrically driven pump with a new,
(4)=(2)*(3 (5) (6) efficient, electrically driven pump equipped with a VSD was
(1) (2) (3)
)
determined by applying the following equation:
MV 403 315 5 1,575 1 30
MV 253 110 1 110 1 80
1 1 ℎ
∆ = , ∙ ∙ − , ∙ ∙ ∙ , (2)
VDF 300 75 3 225 1 75
MV 403 110 2 220 1 90
MV 403 90 2 180 1 80 1 1 ℎ
TOTAL 13 2,310 5 71 ∆ = 315 ∙ 1,440 ∙ − 200 ∙ 1,440 ∙ ∙ 0.7
0.84 0.935

B. Proposed EPIA ℎ ℎ
∆ = 324,385 = 324.39
The EPIA proposed in the 2012 Energy Audit consisted in From a financial point of view, the CAPEX required to
replacing the MV 403 cold water pump with a smaller, more implement the Water Pumping System EPIA is presented in
efficient MV253x3 cold water pump with a rated power of Table IV.
200 kW and a rated voltage of 0.4 kV, powered by a three-
phase asynchronous motor equipped with a Variable Speed TABLE IV
Drive (VSD). CAPEX OF THE WATER PUMPING SYSTEM EPIA
As proven by [6]-[8], the installation of a VSD can lead to a
Cost Value M.U.
reduction of the electricity consumption of up to 38,9% at a MEC-MV253x3 Acquisition Cost 46,200 EUR
rotation speed reduction of only 20%. For this case, based on Toshiba VFAS3-4200 KPC VSD
9,400 EUR
the average load of the electrically driven pump and the VSD Acquisition Cost
capabilities, an electricity consumption reduction of 25% will Hydraulic System modification procedure 4,500 EUR
be considered. Removal of the existing electrically driven
2,400 EUR
pump
The main technical and operational inputs used in the Installation of new electrically driven pump 2,600 EUR
energy economy forecast are presented in Table III. Additional costs required to power the new
4,300 EUR
pump
TABLE III CAPEX 72,400 EUR
MAIN TECHNICAL AND OPERATIONAL INPUTS USED IN THE ENERGY
ECONOMY FORECAST – WATER PUMPING SYSTEM EPIA The Operational and Maintenance Costs (OPEX) are
Input Notation Value M.U. considered constant in both scenarios, thus it can be ignored
Average technological process -
8,160 h/year from the technic and economic analysis presented in Section
functioning time IV.
Average Cold-Water Pump functioning
1,440 h/year
time
Actual Rated Power , 315 kW III. HEAT RECOVERY EPIA
Actual Average Load 30 %
Rated Motor Efficiency 84 %
A. Reference Case
Proposed Scenario Rated Power , 200 kW
Proposed Scenario Average Load , 70 %
Proposed Scenario Rated Motor The Heat Treatment Oven (HTO) is the Significant Energy
93.5 % Use when it comes to Natural Gas consumption, being
Efficiency
Total Installed Power Reduction 115 kW responsible for up to 51% of the entire equivalent energy
VSD energy usage reduction 25 % consumption of the Organization.
At the time of the Energy Audit, in 2012, the Waste Gasses
In order to properly quantify the electricity consumption were not used at their full energy potential and were simply
reduction obtainable by replacing the existing Cold-Water cooled down from their average temperature of 827 degrees
Celsius to about 400 degrees Celsius and then dispersed into

146
the atmosphere. The average annual operation time of the ℎ
= 0.9 ∙ 2,000 ∙ 1.61 = 2,898
HTO is THTO=2,000 hours/year and it is fueled by natural gas
with LHV=42 MJ/ . The average characteristics of the
HTO exhaust gasses are presented in Table V. Considering an average efficiency of the oven of 95%, the
annual fuel saving ( ) will be:
TABLE V 2,989 ℎ
EXHAUST GASSES CHARACTERISTICS = = = 3,050 (7)
0.95
Characteristic Notation Value M.U.
Waste Gas Temperature = 827 ˚C From a financial point of view, the CAPEX required to
Excess air coefficient 1.2 - implement the Heat Recovery System EPIA is presented in
9.54 % Table VI.
Volumetric Composition 19.03 %
71.43 % TABLE VI
Combustion air temperature 25.7 ˚C CAPEX OF THE HEAT RECOVERY SYSTEM EPIA
Maximal fuel consumption 427
Cost Value M.U.
B. Proposed EPIA Piping 42,000 EUR
Heat Exchanger 93,000 EUR
Electrically Driven Fan used for Air Intake 12,000 EUR
The EPIA proposed in the 2012 and 2016 Energy Audits
Installation of the HRS 16,000 EUR
consisted in implementing a Heat Recovery System (HRS) in CAPEX 163,000 EUR
order to efficiently use the exhaust gasses temperature to
optimize the Natural Gas consumption by preheating the The Operational and Maintenance Costs (OPEX) are
combustion air (see Fig. 1) [9]. estimated at approximatively 12,900 EUR/year and are
generated by the periodic cleaning and maintenance procedure
of the pipe system and the electricity consumption of the air
intake fan.

IV. COST OF OPPORTUNITY ANALYSIS

A. Input Data and General Hypothesis

Fig. 1. Simplified HRS flow diagram
The main criterions used in the technic and economic
analysis of the EPIAs were the Net Present Value – NPV (8),
In order to evaluate the energy consumption reduction
the Internal Rate of Return – IRR (9), the Simple Payback
obtainable by implementing this EPIA, the specific heat of the
Period (10), determined by considering a variable annual net
exhausted gasses was determined by using the following
income and the Benefit – Cost Analysis – BCA (11).
equation:
−
= ∙ , = 1.506 (3) = − [ ] (8)
∙ ∙ (1 + )
The volume of exhausted gasses is considered for an excess
where ttst is the analysis time-frame, in years, which is
air coefficient of = 1,2 and an absolute humidity of =
considered to be 12 years in this case [11], It is the yearly
10 , by applying the equation: income in the tth year, in EUR/year, Ct are the yearly
,
expenditures in the tth year, in EUR/year, a is the discount
=1+ =1+ ∙ ∙ (1 + 0.00161 ∙ ) (4) rate, in %/year and IC is the investment cost, in EUR.
,
, −
= 12.72 = = 0[ ] (9)
, (1 + )
The energy content of the exhausted gasses ( ) is where the CAPEX can be included in the yearly expenditures
determined considering the maximum air temperature after the as a depreciation cost.
heat recovery process:
= [ ] (10)
= ∙ ∙ ∙ − = 1.61 [ ] (5) ∑

Considering a HRS efficiency of 90% and an average

= [−] (11)
annual operation time of 2,000 hours/year, the estimated
energy economy ( ) will be: In order to quantify the yearly income that the Organization
missed out on, a historical evolution of the electricity and the
ℎ
= ∙ ∙ (6) natural gas prices, as average national values, is presented in
Table VII.

147
 TABLE VII proper periodical Energy Audits and a continuous
ELECTRICITY AND NATURAL GAS PRICES EVOLUTION quantification of the cost of energy reduction [12].Further
Year Electricity Price Natural Gas Price research is needed to evaluate how the Cost of Indecision can
[-] [EUR/MWh] [EUR/MWh] be used as a hierarchy setting criterion when comparing
2012 92.9 20.25 multiple EPIAs.
2013 94.4 23.4
2014 92.9 23.9
2015 96.1 23.85
2016 84.4 22.85
2017 77.9 30.2
2018 85.1 30.35
2019 95.7 33
2020 108 34.5
2021 110.2 37.54
2022 112.43 40.84
2023 114.,72 44.29

Fig. 2. Cumulative Cashflows for the EPIA implementation scenario

As it can be observed, both prices increased from 2012 until
now, with electricity prices increasing with an average of
16.24% from 2012 to 2020 and natural gas prices increasing
with an average of 70.37% from 2012 to 2020. Based on the
average price growth rate of 2.03%/year for electricity and
8.80%/year for natural gas, the prices for 2021, 2022 and 2023
were forecasted.
B. Technic and Economic Analysis Results
As it can be observed in Table VIII, both EPIAs would’ve
led to extremely attractive Economic Indicators. However,
due to the indecision of the Top Management of the
Organization, those Indicators were not achieved. In order to
evaluate the accurate amount of capital that the Organization
lost as a result of its inaction, the annual income generated by
depositing the not-invested capital is considered.
TABLE VIII
TECHNIC AND ECONOMIC ANALYSIS RESULTS
Economic
WPS EPIA HRS EPIA
Indicator
NPV [EUR] 148,238 366,801
IRR [%/year] 41 38
SPP [years] 2.30 1.72
BCA [-] 2.05 1.45
V. CONCLUSION
The total CAPEX of the unimplemented EPIAs amount to
232,400 EUR. Considering an average interest rate of
2,5%/year for a 12 years deposit, the total NPV obtained by
not investing in the EPIAs is approximatively 312,553 EUR.
This means that after nine years the Organization lost 153,305
EUR by not implementing the two analyzed EPIAs (Fig. 2).
In conclusion, the Cost of Indecision – “Inaction Loss” –
can be used as a next-level incentive for Organizations,
Energy Services Companies and Energy Management
Personnel when analyzing EPIA which lead to a negative set
of financial indicators, in order to make them more attractive
to the decision makers. This paper proved that presenting the
Cost of Indecision to the top management of an industrial
company can lead to a significant increase in the approval and
implementation of EPIAs, especially when combined with
(blue) and the reference scenario (orange)
REFERENCES
[1] The European Parliament and the Council, “Directive (EU)
2018/844 Amending Directive 2010/31/EU on the Energy
Performance of Buildings and Directive 2012/27/EU on Energy
Efficiency,” PE/4/2018/REV/1, Official Journal of the European
Union, L156/75, 19.06.2018.
[2] The Romanian Parliament, “The Energy Efficiency Law no.
121/2014,” Official Gazette, Part I, No. 574 from 01.08.2014.
[3] The Romanian Parliament, “Law no. 160/2016 for modifying
and updating the Energy Efficiency Law no. 121/2014”, Official
Gazette, Part I, No. 562 from 26.07.2016.
[4] A. Hasanbeigi, L. Price, H. Lu and W. Lan, “Analysis of
Energy-Efficiency Opportunities for the Cement Industry in
Shandong Province, China: A Case Study of 16 Cement Plants”,
Energy, vol. 35, no. 8, pp. 3461-3473, 2010.
[5] J. Harris, P. Bostrom and R. B. Lung, ”Crossing the Valley of
Death: Policy Options to advance the Uptake of Energy-
Efficient Emerging Technologies in US Industry”, Industrial
Energy Technology Conference, pp. 17-19, 2011.
[6] P. Bhase, M. Lathkar, ”Energy Conservation using VFD”, 2015
International Conference on Energy Systems and Applications
(ICESA), DOI: 10.1109/ICESA.2015.7503406, 2015.
[7] E. Morris, D. Armitage, ”A Guide to Standard Medium Voltage
Variable Speed Drives, Part 2”, pp. 7-13, 2012.
[8] R. Paes, ”An Overview of Medium Voltage AC Adjustable
Speed Drives and IEEE Std. 1566 – Standard for Performance
of Adjustable Speed AC Drives Rated 375 kW and Larger”,
Joint Power Engineering Society – Industrial Society Technical
Seminar, IEEE Southern Alberta Chapter 1-78, 2011.
[9] E. Cagno, D. Accordini and A. Trianni, ”A Framework to
Characterize Factors Affecting the Adoption of Energy
Efficiency Measures within Electric Motors Systems”, 10th
International Conference on Applied Energy (ICAE), Energy
Procedia, vol. 158, pp. 3352-3357, 2018.
[10] B. Delpech, B. P. Axcell and H. Jouhara, “A Review on Waste
Heat Recovery from Exhaust in the Ceramics Industry”, E3S
Web of Conferences 22:00034, 2017, DOI:
10.1051/e3sconf/20172200034
[11] The Romanian Government, “Decision no. 2139/30.10.2004 for
approving the normal period of use for capital goods,” Official
Gazette, Part I, No. 46/13.01.2005.
[12] K. Abubakar, A. Haneef, J. G. Ringim and A. Musa, “An
Approach to Energy Saving and Cost of Energy Reduction
using an Improvement Efficient Technology”, Open Journal of
Energy Efficiency, vol. 4, pp. 61-68, 2015.

148
SIGNAL PROCESSING I
 The Effects of Coefficient Quantization
on New CIC FIR Filters
Biljana P. Stošić
Abstract — Novel selective, highly effective and low complexity
CIC (Cascaded-Integrator-Comb) FIR (Finite Impulse Response)
filters, for applications in communication systems, are described
here. The paper proposes the use of classical CIC filters as
functions in generating new filters, by spreading the delays in the
CIC filter comb stages. To verify the frequency response
characteristic improvements of the suggested CIC FIR filters,
comparative study of the frequency responses of the generated
filters with classical CIC filters is provided.
Quantization has a great influence on digital signal processing
applications. The effects of coefficient quantization as one type of
quantization that influences a filter characteristic are
investigated here.
Keywords — Digital filters, CIC filters, FIR filters, Linear
phase characteristic, Coefficient quantization.
I. INTRODUCTION
In 1981, Hogenauer presented with hardware-efficient
linear phase FIR (Finite Impulse Response) filter class so
called CIC (Cascaded-Integrator-Comb) filters [1]. They are
very efficient in terms of computational complexity due to
their properties such as no required multipliers and no
required memory for the storage of filter coefficients. A CIC
filter that uses only additions/subtractions reduces the
computational demands.
It is well known that the classical CIC filters themselves
suffers from a passband droop which is unacceptable in many
applications. The improvement of the CIC filter frequency
response attracted much attention during the last several years.
Hence, it is of great interest to get a flat passband using a
compensation filter [2-5].
CIC filter is what the name indicates: a cascade of simple
integrator and comb filter stages. A new selective low-pass
CIC filters based on the classical CIC filters and obtained by
spreading the delays in the comb stages, are designed by
author and presented in [6-12]. Compact forms of new
functions, their characteristics in the frequency domain as well
as the performance achievements over the classical CIC filters
and different solutions presented earlier, are described.
Digital filters are very important type of filters in many
digital signal processing (DSP) applications. Quantization has
a great influence in DSP. Coefficients quantization represents
a type of quantization that influences a frequency response of
digital filter. Hardware implementation of FIR filters requires
Biljana P. Stošić is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
biljana.stosic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE 151
quantization of filter coefficients to the word size of registers.
The goal is to obtained quantized frequency characteristic that
has more similarity to frequency characteristic of infinite
precision FIR filter.
The novelty presented in this paper are novel selective,
highly effective and low complexity CIC FIR filters, designed
as several cascaded non-identical CIC filter section. They are
given in a closed form providing the filter coefficients
analytically. Some characteristics of them are shown here to
demonstrate their usefulness in communication systems. After
designing a digital filter, the effects of the coefficient
quantization on its frequency characteristic have to be
examined and studied. If the quantized filter does not meet the
target specifications, the designer needs to redesign the filter.
This paper analyzes the effects of coefficient quantization in a
new designed CIC FIR filter. In particular, it will be study
how magnitude response of the filter is affected by coefficient
quantization.
This work is unscrewed into two parts: a filter design and a
coefficient quantization. The design procedure of the novel
CIC FIR filters is discussed in Sections II to IV. Then, Section
V is dedicated to verification of the designed filter. Section VI
gives a review of coefficient quantization’s effects on filter
frequency characteristic. The summary is given in Section VII.
II. MULTIPLIERLESS LINEAR-PHASE CIC FILTERS
The normalized CIC FIR filter function of one section (i.e.
CIC section) in z -domain is defined as a cascade of simple
integrator and comb filter stages
1− z−N
H (N , z) =
1
N
( )
⋅ 1 + z −1 + z −2 +  + z −( N −1) =
N ⋅ (1 − z −1 )
(1)
The filter function of a comb is H c ( N , z ) = (1 − z − N ) , and its
time domain representation is yc (n) = x(n) − x(n − N ) .
Function H i ( z ) = 1 / (1 − z −1 ) represents the filter function of
an integrator and yi (n) = x(n) + y (n − 1) is its difference
equation in time domain.
The classical CIC FIR filter function of normalized
amplitude response characteristic is defined as
 1 − z−N 
K
H (N , K , z) =   , (2)
 N ⋅ (1 − z −1 ) 
 
where N is the decimation factor, and K is the number of
sections [1].
III. NON-RECURSIVE FORMS OF DESIGNED CIC FIR between the stopband attenuation and the parameter L . For
FILTERS chosen parameter N , the minimum attenuation increase
rapidly by increasing it. Data show that the new filters have
The new designed CIC FIR filters having normalized much better performances especially for filter order from 5 to
response can be written in non-recursive form 10.
60
L L=3
  1 N −3   1 N −2 −r   50 L=2

Stopband attenuation
   N −1 

improvement [dB]
H ( N , K , L, z ) =  z −r  ⋅  z  40
 N − 2 r =0   r =0

  (3) 30

L 20
 1 N −1   1 N −r   1 N +1 −r  
z −r
  N +1    N +2 
10
⋅  ⋅ z ⋅ z  L=1
N 

0
r =0   r =0   r =0  5 10 15 20
where N and L are freely selected integer parameters, and N
parameter K takes value of K = 5L . This function represents Fig. 1. Improvements of the stopband attenuation of new CIC FIR
five cascade-connected non-identic CIC sections ( H(N − 2, z) , filters over classical CIC filters

H(N −1, z) , H(N, z) , H(N +1, z) , and H(N + 2, z) of the form The frequency characteristics of suggested filters are
given by Eq. (1)) which are repeated L times. calculated in MATLAB by evaluating the function at the
In general, the non-recursive implementation is
sample points defined by setting z = e jω , where ω = 2π ⋅ f
( N −1)⋅ K
1
H ( N , K , L, z ) =
H 0 ( N , K , L)
⋅  h( N , r ) ⋅ z − r , (4) reprents angular frequency in rad/s. By use of Euler’s identity,
the frequency characteristic is obtained
r =0
where H 0 ( N , K , L) presents the normalized constant for the L
 ( N − 2)ω ( N − 1)ω Nω 
unit magnitude response at f = 0 , and h ( N , r ) represents the jω  sin 2
sin
2
sin
2 
H ( N , K , L, e ) =  ⋅ ⋅  ⋅
coefficients of the impulse response. Vector h( N , K , L) has a  ( N − 2) ⋅ sin ω ω ω 
( N − 1) ⋅ sin N ⋅ sin
total of M = ( N − 1) ⋅ K + 1 impulse response coefficients  2 2 2 
L
satisfying h ( N , r ) = h ( N , M − r ) relation.  ( N + 1)ω ( N + 2)ω  K ( N −1)ω
 sin 2
sin
2  −j
As examples of non-recursive implementation, filter ⋅ ⋅  ⋅ e 2
functions calculated for even and odd values of parameter N  ( N + 1) ⋅ sin ω ( N + 2) ⋅ sin ω 
(e.g. 4 and 5) and K = 5 (calculated for L = 1 ), are observed.  2 2
For classical CIC filter functions H ( N , K , z ) , satisfying Eq. and K = 5L . (9)
(4), the normalized constants are 1024 for N = 4 , and 3125 The normalized amplitude characteristics of the suggested
for N = 5 . The impulse response coefficient vectors are given filters are defined as the magnitudes of the complex filter
in Eqs. (5) and (6), respectively: frequency responses H ( N , K , L, e jω ) . Then, the magnitude
hCIC (4,5) = {1,5,15,35, 65,101,135,155, characteristics represent the absolute values of them.
(5) The linear phase characteristic of new filter functions is
155,135,101, 65,35,15,5,1}
φ ( N , K , L, ω) = − ( N − 1) ⋅ K ⋅ ω / 2 , K = 5L . (10)
hCIC (5,5) = {1,5,15,35, 70,121,185, 255,320,365, The constant group delay characteristic is
(6)
381,365,320, 255,185,121, 70,35,15,5,1} τ ( N , K , L, ω) = − dφ ( N , K , L, ω) / dω = ( N − 1) ⋅ K / 2 ,
For new CIC filter functions, H ( N , K , L, z ) , the impulse K = 5L . (11)
response coefficient vectors are given in Eqs. (7) and (8),
respectively: IV. DESIGNED EXAMPLE AND COMPARISON
h(4,5,1) = {1,5,14, 29, 49, 71,90,101,101,90, 71, 49, 29,14,5,1} (7)
In this part, an example is given in order to show
h(5,5,1) = {1,5,15,34, 64,105,154, 205, 250, 281, 292, effectiveness of the frequency response of the designed
(8)
281, 250, 205,154,105,64,34,15,5,1} selective CIC filters in comparison to the classical CIC filters.
In order to fairly compare these filters in Fig. 2, the
The constants H 0 ( N , K , L) are 720 for N = 4 , and 2520 for normalized magnitudes are calculated for the same group
N =5. delay and the same number of delay elements. The designed
As stated before, one can see the coefficient symmetry of filters have an equal number of cascaded sections K .
proposed FIR filter functions from Eqs. (7) and (8). Also, the However, the classical CIC filters have an identic form in all
calculated constants and impulse response coefficients are cascades, and the new designed filters have cascade of non-
smaller compared to those calculated for classical CIC filters. identic CIC filter sections. The novel CIC filter functions
The stopband attenuation improvements of new filters over show higher selectivity and higher insertion losses in
the classical CIC filters, calculated for several values of stopband.
parameter L are depicted in Fig. 1. Notice close relation

152
 filters are smaller than 1. Because of that, B + 1 -bit format
which reserve 1 bit for sign, 1 bit for integer part and B − 1
bits for the fractional part has been chosen for coefficient
representation.
As examples of affecting the magnitude response of a filter
by coefficient quantization, classical CIC filters and new CIC
FIR low-pass filters with different orders are designed and
different number of bits are considered. Normalized curves are
pictured in Figs. 4-7.

Fig. 2. Comparison of normalized magnitudes in dB:

N = 4 and K = 5 .

The zero locations and their multiplicities are shown in Fig.

3, for case of classical and proposed filters and chosen
parameters N = 6 and K = 10 . CIC filters have N − 1
different zeros with multiplicity K, zr = e j⋅2π ⋅r / N ,
r = 1, 2,..., N − 1 . The number of zeros is ( N − 1) ⋅ K . All zeros
lied on the unit circle, Fig. 3a. New filters have zeros more
evenly distributed as well as their multiplicities reduced, Fig. Fig. 4. Classical CIC filter curves in dB: N = 4
3b.

(a) (b)
Fig. 3. z -plane zero locations with multiplicities: (a) classical filter
and (b) new filter; case N = 6 and K = 10 .

V. COEFFICIENT QUANTIZATION EFFECTS

Fig. 5. Classical CIC filter curves in dB: N = 5
A design of digital filters involves finding the coefficients
of the filter. For computing the coefficients of digital filter,
infinite precision arithmetic is used (the number of digits is
only limited by available memory of the system).
In DSP system, the number of bits in designing a filter is
limited by word length of the register used to store them
(registers have a fixed number of bits). Quantization methods
like rounding and truncating are used to quantize the filter
coefficients to the word size of the register [13], [14].
As it is known, a common filter design tool like MATLAB,
can run the design algorithm and return the filter coefficients.
The designed coefficients have to be converted or quantized to
a fixed-point representation [13]. A signed fixed-point
representation with B + 1 total bits including sign is used here.
Notice that filter coefficients of the novel designed CIC FIR
Fig. 6. New CIC filter curves in dB: N = 4

153

Fig. 7. New CIC filter curvess in dB: N = 5
It can be concluded that the quantization of filter
coefficients creates a deviation in the frequency response of
the filter as seen in Figs. 4-7. Note that if the wordlength B is
not large enough, there will be undesirable effects. In
summary, after coefficient quantization, a filter having a
frequency response diverge from the frequency response of a
filter with unquantized coefficients is obtained. The
quantization effect is more visible and more significant in the
stopband area.
VIII. CONCLUDING REMARKS
In this paper, novel selective CIC filter functions are
designed based on cascade connection of five successive CIC
filter sections duplicated an integer number of times. All
designed filters have zeros more evenly distributed and
located on the unit circle.
Closed-forms for the designed filter coefficient computation
are presented. The calculated coefficients have integer values
smaller than those of classical CIC filters which are important
during the hardware realization because of the finite
wordlength effects. In practice, it is reasonable to do
coefficient quantization of a single filter and measure the
effect on its frequency response. Implementing a higher-order
filter requires a higher number of bits for coefficient
representation.
A classical CIC filter which is a very often used in many
DSP applications can be replaced most directly by suggested
CIC filters which still remains the most effective FIR filter
characteristics but have better frequency characteristics.
ACKNOWLEDGMENT
The Ministry of Education, Science and Technological
Development of the Republic of Serbia has supported this
research.
154
REFERENCES
[1] E. B. Hogenauer, “An Economical Class of Digital Filters for
Decimation and Interpolation”, IEEE Transactions on
Acoustics, Speech and Signal Processing, April 1981, pp. 155-
162.
[2] G. Jovanović Doleček, S. K. Mitra, “A Simple Method for the
Compensation of CIC Decimation Filter”, Electronics Letters,
2008, vol. 44, no. 19, pp. 1162-1163.
[3] J. O. Coleman, “Chebyshev Stopbands for CIC decimation
Filters and CIC-Implemented Array Tapers in 1D and 2D”,
IEEE Trans. on Circuits and Systems I, 2012, vol. 59, no. 12,
pp. 2956-2968.
[4] G. Jovanović Doleček, A. Fernandez-Vazquez, “Novel Droop-
Compensated Comb Decimation Filter with Improved Alias
Rejections”, Intern. Journal of Electronics and Communications
(AEU), May 2013, vol. 67, no. 5, pp. 387-396.
[5] G. Jovanović Doleček, M. Laddomada, “An Improved Class of
Multiplerless Decimation Filters: Analysis and Design”, Digital
Signal Processing, 2013, vol. 23, no. 5, pp. 1773-1782.
[6] B. P. Stošić, V. D. Pavlović, “On Design of a Novel Class of
Selective CIC FIR Filter Functions with Improved Response”,
Intern. Journal of Electronics and Communications (AEU),
August 2014, vol. 68, no. 8, pp. 720-729.
[7] B. P. Stošić, V. D. Pavlović, “Design of Selective CIC Filter
Functions”, Intern. Journal of Electronics and Communications
(AEU), December 2014, vol. 68, no. 12, pp. 1231-1233.
[8] B. P. Stošić, D. N. Milić, and V. D. Pavlović, “New CIC Filter
Architecture: Design, Parametric Analysis and Some
Comparisons”, IETE Journal of Research/Institution of
Electronics and Telecommunication Engineers, Taylor &
Francis, vol. 61, no. 3, May-June 2015, pp. 244-250.
[9] B. P. Stošić, V. D. Pavlović, “Using Cascaded Non-Identical
CIC Sections to Improve Insertion Loss”, Journal of Circuits,
Systems, and Computers, World Scientific, vol. 24, no. 6, July
2015, 19 pages.
[10] B. P. Stošić, D. N. Milić, and V. D. Pavlović, "Design and
Properties of New CIC FIR Filter Functions", Proceedings of
the Romanian Academy, Series A: Mathematics, Physics,
Technical Sciences, Information Science, vol. 16, no. 4, 2015,
pp. 539-546.
[11] B. P. Stošić, V. D. Pavlović, "Design of New Selective CIC
Filter Functions with Passband-Droop Compensation",
Electronics Letters, 2016, vol. 52, no. 2, January 2016, pp. 115-
117.
[12] B. P. Stošić, V. D. Pavlović, “Closed-form Design of New Class
of Selective CIC FIR Filter Functions“, LII International
Scientific Conference on Information, Communication and
Energy Systems and Technologies - ICEST 2017, Serbia, Niš,
June 28-30, 2017, pp. 31-35.
[13] A. V. Oppenheim, R. W. Schafer, Discrete-Time Signal
Processing. Vol. 3, Prentice Hall Englewood Cliffs, NJ, 2010.
[14] B. P. Lathi, R. A. Green, Essentials of Digital Signal
Processing, Cambridge University Press, 2014.
 Improvement of Energy Efficiency of PAM Constellation
by Applying Optimal Companding Quantization in
Constellation Designing
Slobodan A. Vlajkov1, Aleksandra Ž. Jovanović2 and Zoran H. Perić3
Abstract – In this paper improvement of energy efficiency for
pulse amplitude modulation (PAM) constellation in additive
white Gaussian noise (AWGN) channel is considered.
Conventional PAM constellation with equidistant and
equiprobable constellation points does not achieve satisfactory
energy efficiency. Considering this fact, it is clear there is a need
to design new PAM constellations with goal to increase energy
efficiency compared to conventional PAM. In order to meet
proposed requirement we design PAM constellation by
establishing an analogy with scalar quantization. In particular,
we perform geometric and probabilistic constellation shaping by
establishing an analogy with piecewise uniform scalar
quantization obtained through linearization of optimal
compression characteristic.
Keywords – PAM constellation, Energy efficiency, Optimal
companding quantization, Gaussian source.
I. INTRODUCTION
One of the actual problems in digital communication is
improvement of constellation energy efficiency [1-5]. The
pulse amplitude modulation (PAM) constellation with
equidistant and equiprobable constellation points does not
achieve satisfactory energy efficiency. On the other side,
energy efficient PAM constellation with non-uniform
Gaussian-like distribution of constellation points is too
complex. In order to define constellation having higher energy
efficient than conventional PAM constellation and lower
complexity than PAM constellation with non-uniform
probabilities of constellation points, in this paper we design
PAM constellations by employing both the geometrical and
probabilistic constellation shaping. In particular, we establish
an analogy between constellation points and representation
levels of scalar quantization to perform geometrical and
probabilistic constellation shaping.
In order to reduce complexity caused by totally different
probabilities of constellation points, we get idea to divide a set
of constellation points into smaller subsets within which
1Slobodan A. Vlajkov is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: vlajkov.slobodan@gmail.com
2Aleksandra Z. Jovanović is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: aleksandra.jovanovic@elfak.ni.ac.rs
3Zoran H. Perić is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
zoran.peric@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
constellation points have the same probability, while
probabilities of constellation points from different subset are
different. These requirements we meet by establishing
analogy between constellation points and representation levels
of piecewise uniform scalar quantization. Also, since the
maximization of energy efficiency comes from Gaussian
distribution of constellation point probabilities, we suppose
that amplitudes being quantized have Gaussian probability
density function (pdf). Application of quantization theory in
constellation design has been also exploited in [6-10]. Unlike
those analyzes, in analysis performed in this paper we paid
special attention to quantizer support limit optimization with
goal to increase energy efficiency of constellation designed by
establishing analogy with quantization.
II. SYMBOL ERROR PROBABILITY OF PIECEWISE
UNIFORM PAM CONSTELLATION
Design of PAM constellation is based on definition of two
basic elements, the amplitudes and probabilities of
constellation points. In this paper, we are interested in
designing PAM constellation consisting of subsets, whereby
the probabilities of constellation points within one subset is
the same, while the probabilities of the points from different
subsets are different. Similar rule holds for amplitudes of
constellation points. Thus, distance between adjacent
constellation points within subset is constant, while distances
between adjacent constellation points from different subsets
are different. In hence, we introduce two indexes i, j. First
index i is used to define subset of constellation points, while
second index j is used to define a certain point in subset. Let
the numbers of constellation points in subsets be equal M/L,
where M represents the total number of constellation points
and L the number of subsets. Now, we can define
constellation by means of the ordered pairs of values
{ai,j, Pi = P(ai,j)}, i = 1,...,L, j = −M/(2L)+1,...,0,1,...,M/(2L),
where ai,j is the amplitude of constellation point and P(ai,j) is
the probability of occurrence of constellation point. Due to
PAM constellation symmetry, in further investigating we
consider constellation points from positive part of real axis
specified with indexes i = 1,...,L, j = 1,...,M/(2L). One can
emphasize that there are two special cases of PAM
constellation defined above. For L = 1, constellation
represents constellation with equiprobable and equidistant
constellation points. For L = M/2, constellation is consisted of
non-equprobable and non-equidistant constellation points.
155
 The average power per symbol for piecewise-uniform III. PAM CONSTELLATION DESIGN INSPIRED WITH
constellation is: OPTIMAL COMPANDING SCALAR QUANTIZATION
L M (2 L )

i =1 j =1
( )
a 2 = 2  P a i , j ai2, j , (1) We design PAM constellation by establishing analogy
between PAM constellation and scalar quantization, precisely,
while signal to noise ratio per bit (SNR) in additive white between constellation points and representation levels of
Gaussian noise (AWGN) channel is [1]: piecewise linear companding scalar quantization of
amplitudes having Gaussian pdf with zero mean value and
SNR = 10 log 10
a2
. (2) unit variance [9], [10]:
2
σ log 2 M
n
1  x2 
p (x ) = exp − , − ∞ < x < ∞ . (8)
In channel signal is degraded by adding AWGN sample n, 2π  2
with zero mean value and variance σn2 = N0/2, to useful signal
sample. Then, the sample at the receiver input is r = ai,j + n, We assume that piecewise linear companding scalar
causing that the received sample r has the conditional quantization is obtained by performing linearization of
Gaussian pdf, with the same variance as noise but with mean optimal compression characteristic for Gaussian source. Let
value equal to ai,j.[1]: define piecewise linear compression characteristic on interval
c(x): [−xmax, xmax] →[−xmax, xmax], where xmax is already
( ) 1  r − a i , j ( )2
 mentioned quantizer support limit.
p r ai, j = exp − , − ∞ < r < ∞ . (3) Quantization symmetry gives us opportunity to consider
2π σ n  2σ 2n 
only positive arguments of compression characteristic. If we
As MAP criteria is superior compared to ML criteria [1], [11], declare L as the number of segments of the piecewise linear
we adopt decision rule according to MAP rule: compression characteristic, and ti, i = 1, 2,...,L+1, t1 = 0 and
tL+1 = +∞ as the segment thresholds for positive range of
value, then the i-th segment is defined on [−ti+1, −ti)∪ [ti, ti+1).
( ) ( )
D (r ) = ai , j , if p ai , j r ≥ p a p ,q r , p = 1, L, q = 1,
M
2L
. (4) The linearization of compression characteristic is done in
manner that the linearized characteristic and nonlinear
Due to the symmetry of PAM constellation, we consider compression characteristic have equal values at the segment
thresholds of decision regions noted as: mi,j, i = 1,...,L, thresholds. Since the numbers of representation levels within
j = 1,...,M/(2L)+1, where m1,1 = 0, mL,N/(2L)+1 = +∞, segments are equal, we can define segment thresholds as
mi,N/(2L)+1 = mi+1,1 and from Eq. (4): follows:

a i , j + a i , j −1 xmax
mi , j = , i = 1, , L, j = 2, ,
M c (ti ) = (i − 1) , i = 1,  , L + 1 . (9)
2 2L L

ai ,1 + ai −1,M (2 L ) σ n2 P By solving Eq. (9) per ti we derive an expression for segment

mi ,1 = + ln i −1 , i = 2,, L . (5) thresholds:
2 ai ,1 − ai −1,M (2 L ) Pi
 x 
That led us to following formulation for decision rule: ti = c −1  (i − 1) max , i = 1,  , L + 1 , (10)
 L 
D(r) = ai,j, if mi,j ≤ r < mi,j+1. Now, we have all the necessary
elements to estimate constellation energy efficiency by and after that the expression for the segment width in the
deriving formula for symbol error probability. To do that, we positive part wdi = ti+1 − ti. By taking into consideration that
utilize general expression for error probability [1], [6-11]:
the number of levels within segments is M/L, the
representation levels can be determined by:
L M (2 L )
( [
Pe = 2  P r ∉ mi , j , mi , j +1 y i , j P ai , j
i =1 j =1
) )( )
 1 M
ai , j = t i +  j −  d i , i = 1,, L, j = 1,, (11)
 2  2 L
L M (2 L )
( )
mi , j +1

i =1 j =1
( )
= 1 − 2  P a i , j  p r ai , j dr , (6)
mi , j
where:
wherein we substitute pdf with Eq. (3) to derive expression for wd i 2L
symbol error probability: di = = (ti +1 − ti ), i = 1,, L (12)
M (2 L ) M

L M (2 L )   a i , j − mi , j  m − ai, j 
( )
Pe =   P ai , j erfc

 + erfc i , j +1
 
 , . (7)

Owing to established analogy between representation levels of
i =1 j =1   2σ n   2 σ n  piecewise linear companding scalar quantization and
constellation points of piecewise uniform PAM constellation,
In Eq. (7), erfc( ) represents complementary error function [1]: we actually determine the amplitudes of constellation points
by means of Eqs. (11) and (12). Now, we have to define the
{ }
+∞
erfc(z ) = 2 π  exp − t 2 dt . second key parameter of constellation, the probability of
z

156
occurrence of constellation points. We achieve this by
defining probability of representation levels. We firstly define
the probability that the amplitude with Gaussian pdf belongs
to a certain segment [12]:
t i +1
P (t i ≤ x < t i +1 ) =  p (x )dx =
ti

1  t   t 
= erfc i  − erfc i +1 , i = 1, , L (13)
2   2  2 

After that, by following analogy and assumption of equal

probabilities of constellation points in the subset, we
determine the probabilities of constellation points

P (t i ≤ x < ti +1 ) L   t  t 
( )
P ai , j =
M (2 L )
= erfc i  − erfc i +1
M 
 .
 (14) Fig. 1. For M = 16, SNR = 18 dB and for different numbers of
 2  2  segments, symbol error probability in function of xmax

Now, we specify optimal compression characteristic for TABLE I

amplitudes having Gaussian pdf with zero mean value and OPTIMAL VALUES OF XMAX FOR 16-PAM, 32-PAM, 64-PAM AND
unit variance [12]: DIFFERENT NUMBERS OF SEGMENTS
 x 
1 − erfc  xmaxopt
 6
c(x ) = x max   sgn (x ) (15) M L=2 L=4 L=8 L=16 L=32
 x max  16 2.7 3.7 5 × ×
1 − erfc 

 6  32 2 3.7 5 6 ×
64 2 3.7 5 6 7
Then, linearization of this optimal compression characteristic
we perform by determining segment thresholds from Eq. (10): We also investigate 32-PAM constellation (see Fig. 2) in
the same manner as we investigate 16-PAM. From Fig. 2 we
 1   x  
t i = 6 erfc −1 1 − (i − 1)1 − erfc max  , can conclude that the relations among constellations having
 L   6   (16) L=2, 4, 8, 16 and referent conventional 32-PAM are similar to
i = 1,  , L + 1 above explained observations for 16-PAM. Constellation with
16 subsets has the lowest error probability and thus the
highest energy efficiency, but the complexity of constellation
IV. NUMERICAL RESULTS is also the highest. Thereby the achieved gain in terms of
energy efficiency over constellation with L=8 is far smaller
In this section we describe and explain the obtained than the achieved gains between constellations having L=4
numerical results. We start elaboration with Fig. 1 where for and L=2, L=8 and L=4.
16-PAM constellation and different numbers of subsets L we
present symbol error probability dependence on xmax. We also
present this dependence for referent constellation,
conventional 16-PAM constellation. We can notice that in
comparison with conventional PAM, for xmax values around 1,
all proposed constellations have lower energy efficiency
viewed as achieved symbol error probability for a given SNR.
On the other hand, for values of xmax between 2 and 3
constellation with L=2 has higher energy efficiency than the
conventional PAM, but lower than efficiency of constellations
with L equals to 4 and 8. For xmax values greater than 4,
constellation with L=2 has very poor performances comparing
to other presented constellations in Fig. 1. PAM with L=4 has
wider range of xmax, 2 ≤ xmax ≤,7, wherein error probability is
lower than that of referent constellation. For xmax greater than
Fig. 2. For M = 32, SNR = 25 dB and for different numbers of
2, the most complex constellation having L=8 has the highest
segments, symbol error probability in function of xmax
energy efficiency in comparing with all other constellations.
In Table I we summarize the optimal values of xmax noted in
We perform additional analysis and conclude the optimal
Fig. 1. We also find that the same optimal values for xmax are
values for xmax do not depend on SNR. In Table I, we tabulate
obtained for other values of SNR as well.
the obtained optimal values of xmax for 32-PAM.

157
 The results obtained for 64-PAM are presented in Fig. 3.
The Pe dependence on xmax has the similar behaviour as in
Figs. 1 and 2. As we expet 64-PAM with L=32 has the lowest
value of Pe, but very closely to Pe value for L=16. In order
words, for L equals 32 the achieved gain in terms of symbol
error probability over constellation having L=16 is the
smallest, wherby the constellation complexity is the highest.
Optimal values of xmax for 64-PAM constellation are also
presented in Table I. It is interesting that in Table I, for L
equals 4, 8 or 16 the values of xmaxopt are the same for all M.
This points out that xmaxopt depends on the number of subsets,
but does not depend on the total number of constellation
points.
Fig. 3. For M = 64, SNR = 30 dB and for different numbers of
segments, symbol error probability in function of xmax
For different 16-PAM constelations and corresponding
optimal values of xmax, the symbol error probability in
function of SNR is presented in Fig. 4. From Fig. 4 we can
conclude that apart from 16-PAM having L = 2, constellations
having L=4 and L = 8 are superior in terms of energy
efficiency over convetional 16-PAM. The energy efficiency of
16-PAM having L = 2 can be improved by adjusting number
of constellation points within subsets as it was done in [10].
Fig. 4. Symbol error probability versus SNR for different
16-PAM constellations
V. CONCLUSION
In order to achieve better energy efficiency of PAM
constellation than it is the efficiency of conventional PAM
constellation and smaller constellation complexity than it is
158
the complexity of constellation having non-equiprobable
constellation points, in this paper we design piecewise
uniform PAM constellations in terms of distance between
adjacent constellation points and the probabilities of
constellation points. One issue of our research is to determine
optimal value for xmax subject to symbol error probability, i.e
subject to energy efficiency. We demonstrate that optimal
values of xmax depend on the number of subsets L, but do not
depend on the signal to noise ratio SNR and the total number
of constellation points M. The obtained results also show that
as the number of subsets increases, the symbol error
probability decreases. On the other hand, by increasing the
number of subsets the constellation complexity grows. Beside
this, the gain in energy efficiency obtained by doubling L
decreases as L increases. These findings point out that we
have to be very careful with selection of PAM constellation
parameters.
ACKNOWLEDGEMENT
This research was supported by the Science Fund of the
Republic of Serbia, 6527104, AI- Com-in-AI.
REFERENCES
[1] J. G. Proakis, M. Salehi, Digital Communication, McGraw-Hill,
2008.
[2] L. Jiang et al, “A Novel Multi-level Constellation Compression
Modulation for GFDM-PON”, IEEE Photonics Journal, vol. 11,
no. 2, pp. 1-11, Art no. 7901611, 2019.
[3] X. Wu et al, “A Probabilistic Shaping Method Based on
Intrinsic Bit-level Labeling and Asymmetric Polygon
Modulation for Optical Interconnects”, Optics Communications,
vol. 444, pp. 68-73, 2019.
[4] Z. Qu, I. Djordjevic, and J. Anderson, “Two-dimensional
Constellation Shaping in Fiber-Optic Communications”,
Applied Sciences, vol. 9, no. 9, Art no. 1889, 2019.
[5] J. Cho, P. Winzer, “Probabilistic Constellation Shaping for
Optical Fiber Communications”, Journal of Lightwave
Technology, vol. 37, no. 6, pp. 1590-1607, 2019.
[6] Z. H. Perić, “Nonlinear Transformation of One-dimensional
Constellation Points in order to Error Probability Decreasing”,
Facta Universitatis, Series: Electronics and Energetics, vol. 11,
no. 3, pp. 291-299, 1998.
[7] Z. H. Perić, S. M. Bogosavljević, “Performance of Nonuniform
PAM Constellations for Gaussian Channel”, Electronics and
Electrical Engineering, vol. 53, no. 3, pp. 27-30, 2004.
[8] Z. H. Perić, N. Milošević, A. Ž. Jovanović, and Z. Nikolić,
“Design of Piecewise Uniform PAM Constellation”, SAUM
2014, Conference Proceedings, pp. 109-111, Niš, Serbia, 2014.
[9] Z. H. Perić, A. Ž. Jovanović, and S. A. Vlajkov, “Comparative
Analysis of Various PAM Constellations”, SAUM 2016,
Conference Proceedings, pp. 27-30, Niš, Serbia, 2016.
[10] S. A. Vlajkov, A. Ž. Jovanović, and Z. H. Perić, “Approach in
Companding-Quantization-Inspired PAM Constellation
Design”, IET Communications, vol. 12, no. 18, 2018.
[11] S. A. Vlajkov, A. Ž. Jovanović, and Z. H. Perić, “Comparative
Analysis of ML and MAP Detectors for PAM Constellation in
AWGN Channel”, Facta Universitatis, Series: Electronics and
Energetics, vol. 33, no. 2, pp. 217-226, 2020.
[12] N. S. Jayant, P. Noll, Digital Coding of Waveforms, Englewood
Cliffs, NJ: Prentice-Hall, 1984.
 The Influence of Compression Parameter μ on the Energy
Efficiency of PAM Constellation Based on μ-law
Companding Quantization
Slobodan A. Vlajkov1, Aleksandra Ž. Jovanović2 and Zoran H. Perić3
Abstract – In literature, design of pulse amplitude modulation
(PAM) constellation by establishing an analogy between
constellation points and representation levels of piecewise
uniform scalar companding quantizer has been considered with
the aim to outperform the energy efficiency of the conventional
PAM constellation with equidistant and equiprobable
constellation points and also with the aim to achieve complexity
lower than the complexity of energy efficient PAM constellation
with non- equidistant and Gaussian-like probability distributed
constellation points. In this paper a focus is placed on application
of μ-law companding quantization in PAM constellation
designing. It is noted that the existence of compression
parameter μ provides more sophistic geometric and probabilistic
constellation shaping. In particular, a thorough analysis of
influence of compression parameter μ on the energy efficiency of
PAM constellation is performed with the aim to specify μ that
provides desirable energy efficiency.
Keywords – PAM constellation, Symbol error probability,
Energy efficiency, μ-law companding quantization.
I. INTRODUCTION
Although pulse amplitude modulation (PAM) is the earliest
kind of modulation, it still has the wide area of usage.
Namely, PAM is used in Ethernet communication, in photo-
biology, in transmission over optical networks [1-3].
However, communication systems increasingly demand better
spectral and energy efficiency of transmission techniques.
This motivates us to direct our research on upgrading and
improving PAM constellation.
PAM constellation with equidistant and equiprobable
constellation points does not satisfy energy efficiency
requirements. On the other side, PAM constellation with non-
equidistant and Gaussian-like probability distributed
constellation points is energy efficient but too complex. There
is a need for PAM constellation that outperforms non-uniform
PAM constellation in terms of complexity and conventional
PAM constellation in terms of energy efficiency. In order to
meet these requirements, in [4] we presented PAM
constellation design inspired by piecewise uniform optimal
1Slobodan A. Vlajkov is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: vlajkov.slobodan@gmail.com
2Aleksandra Z. Jovanović is with University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: aleksandra.jovanovic@elfak.ni.ac.rs
3Zoran H. Perić is with University of Niš, Faculty of Electronic
Engineering, Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mail:
zoran.peric@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
companding quantization. In this paper, with the same aim,
we investigate PAM constellation design based on
establishing an analogy between constellation points and
representation levels of piecewise uniform μ-law companding
quantization. Similar as in [4], we perform geometric and
probabilistic shaping of PAM constellation by using method
from quantization designing to arrange constellation points.
The methods from quantization designing were also used in
constellation designing in numerous papers [5-10], where it
was shown that quantization designing methods can be
efficient tools in constellation designing.
In comparison to PAM constellation designing in [4], PAM
constellation design based on μ-law companding quantization
offers more freedom in constellation designing. Actually, the
existence of compression parameter μ provides more sophistic
geometric and probabilistic constellation shaping. Thus in this
paper our research is aimed to determine how compression
parameter μ influences performances of PAM constellation.
The constellation performances we measure through
constellation complexity and energy efficiency i.e. error
probability dependence on signal to noise ratio in additive
white Gaussian noise (AWGN) channel.
II. PAM CONSTELLATION BASED ON PIECEWISE
UNIFORM μ-LAW COMPANDING QUANTIZATION
Design of our PAM constellation is realized by dividing set
of constellation points into subsets where within one subset
distances between adjacent constellation points are same,
while they are different for different subsets. The similar rule
is adopted for probabilities of occurrence of constellation
points. Within one subset the probabilities are the same, while
the probabilities of constellation points from different subsets
are different. Then, the indexes i and j are introduced to
denote subset (index i) and to denote constellation point
within certain subset (index j). In accordance with this, the
constellation point is defined with {ai,j, Pi = P(ai,j)}, i = 1,...,L,
j = −M/(2L)+1,...,0,1,...,M/(2L), where ai,j represents the
amplitude or coordinate of constellation point in one
dimensional space, Pi = P(ai,j) is the probability of occurrence
of constellation point, M is total number of constellation
points and L is total number of subsets. Beside this, equal
number of constellation points within subsets M/L is assumed.
The existence of symmetry in distribution of constellation
points enable us to focus on the determination of constellation
points on positive part of real axis, i = 1,...,L, j = 1,...,M/(2L)
[4].
159
 Now, we can start PAM constellation designing by For symmetric pricewise-uniform PAM constellation, the
establishing analogy between constellation points of PAM average power per symbol is:
constellation and representation levels of piecewise uniform
L M (2 L )
companding scalar quantization. As the main gain in energy
efficiency comes from Maxwell-Boltzmann distribution,
a 2 = 2
i =1
 P (a )a
j =1
i, j
2
i, j . (5)
which is a Gaussian distribution sampled at discrete
amplitudes across a finite amplitude range [11], we presume Then, signal to noise ratio per bit (SNR) in additive white
that amplitudes at the input of scalar companding quantizer Gaussian noise (AWGN) channel is [13]:
used in constellation designing have Gaussian probability
density function (pdf) with zero mean value and unit variance 2a 2
[4-10]. SNR = 10 log10 , (6)
N 0 log2 M
The piecewise uniform companding quantization used in
constellation designing we obtain by performing linearization where N0/2 is AWGN power spectral density.
of μ-law compression characteristic [12]:
Since maximum a posteriori probability (MAP) criteria is
xmax  x  recognized as the best criteria for error probability
c(x ) = ln1 + μ  sgn (x ), x ≤ xmax . (1)
ln (1 + μ )  xmax  minimization [13-15], we consider MAP detection of signal
transmitted through AWGN channel. According to derivation
In Eq. (1) μ is compression parameter and xmax is quantizer conducted in [4], thresholds for decision regions on positive
support limit. If we declare L as the number of segments of
part of real axis are: mi,j, i = 1,...,L, j = 1,...,M/(2L)+1, where
the piecewise linear compression characteristic, and ti, i = 1,
m1,1 = 0, mL,N/(2L)+1 = +∞, mi,N/(2L)+1 = mi+1,1 and
2,...,L+1 as the segment thresholds for positive range of value,
then the i-th segment is defined on [−ti+1, −ti)∪ [ti, ti+1). The ai , j + ai , j −1 M
linearization of compression characteristic is performed in mi , j = , i = 1,  , L, j = 2,  ,
2 2L
such way that nonlinear and linearized compression (7)
characteristics are matched in segment thresholds [4], [8]. ai ,1 + ai −1,M (2 L ) N0 P
mi ,1 = + ln i −1 , i = 2,  , L
Since the numbers of representation levels within segments 2 (
2 ai ,1 − ai −1,M (2 L ) Pi )
are equal, the segment thresholds are:
With Eqs. (3), (4) and (7) all needed elements for estimating
 x  x  1
 symbol error probability of PAM constellation in AWGN
ti = c  (i − 1) max  = max (1 + μ )
−1 (i −1)
L − 1, i = 1,  , L + 1 . (2)
 L  μ  channel are specified. Then, by referring to expressions
derived in [4], the following expression for symbol error
Then, the representation levels uniformly distributed within probability can be formulated:
segments can be determined by:
L M (2 L ) a   − mi , j   mi , j +1 − ai , j 
ai , j =
2 L  1  3 
 j − ti +1 −  j − ti  , i = 1,  , L, j = 1,  ,
M Pe =   P(a )erfc i, j
i, j

N 0 
+ erfc
 N0
  , (8)

M  2 2  2L
i =1 j =1
   

(3)
III. NUMERICAL RESULTS
Based on established analogy between representation levels
and constellation points we actually determinate amplitudes of
In this section, we analyze the influence of compression
constellation points by specifying representation levels of
parameter μ on performances of PAM constellations designed
piecewise uniform companding quantizer with Eqs. (3) and
by establishing analogy with linearized μ-law companding
(2). The second key parameter of constellation, the
quantization. We start our analysis with Fig. 1 where for 16-
probabilities of occurrence of constellation points we
PAM constellation (M = 16) and different numbers of subsets
determine by specifying the probability of representation
L and different values of μ we present symbol error
levels. The probability that the amplitude with Gaussian pdf
probability dependence on xmax. We also present this
belongs to a certain segment is covered in [4], [8]. By
dependence for referent constellation, conventional 16-PAM
following analogy and assumption of equal probabilities of
constellation. From Fig. 1 we can conclude that for all L all
constellation points within the subset, we determine the
PAM constellations having μ values in range from 50 to 255
probabilities of constellation points:
have very poor performances comparing to conventional PAM
constellation. For xmax values ranging from 1 to 4, 16-PAM
P(ti ≤ x < ti +1 ) L   ti   t 
( )
P ai , j =
M (2 L )
= erfc
M 
 − erfc i +1  , (4) constellations having μ=5 also underperforms referent PAM
 2  2  constellation. However, on xmax interval 4≤ xmax ≤10, symbol
error probability decreases as value of xmax increases, whereby
where erfc( ) represents complementary error function constellation performances considerably exceed the
{ }
+∞
erfc(z ) = 2 π  exp − t 2 dt . performances of conventional 16-PAM constellation.
z

160

Fig. 1. For M = 16, μ = 5, 50, 255, SNR = 20 dB and for different
numbers of subsets L, symbol error probability in function of xmax
We continue our analysis by considering behaviour of
symbol error probability in function of μ for a given xmax and
different L (see Fig. 2). From Fig. 2 we can notice that in
comparison with conventional PAM, for xmax = 4.2 our
constellations having μ values lower than 5 are superior in
terms of symbol error probability. It is important to point out
that constellation with μ = 1 achieves the best results among
others. From Fig. 2 we can also notice that the symbol error
probability decreases with the increase of number of subsets,
whereby the rate of gain decreases with the increase of L.
Since the increase of number of subsets increases
constellation complexity, the compromise should be made
when the choice of L is in question.
Fig. 2. For M = 16, xmax = 4.2, SNR = 20 dB and for different
numbers of subsets L, symbol error probability in function of μ
In Fig. 1 the Pe dependence on xmax for 16-PAM
constellations having μ = 5 and different number of subsets
shows that gain in decrease of symbol error probability
161
considerably increases with xmax. The disadvantage which
follows with xmax augmentation is very low values of
probabilities of constellation points for such PAM
constellations. In order to constraint probability of occurrence
of constellation point on value relevant for transmission
through the channel, the appropriate xmax value should be
determined.
In Fig. 3 we show the Pe dependence on SNR for 16-PAM
constellations specified in Table I. As we have previously
stated, from Fig. 3 is evident that the symbol error probability
and thus the energy efficiency can be improved by decreasing
μ and by increasing L. Also, we can see that PAM
constellations having μ = 1 outperform conventional PAM as
well as all other constellations in whole range of SNR and for
all L values. In order to quantify the achieved gain in energy
efficiency, in Table I, for conventional and our PAM
constellations, we tabulate the required SNR values for a
given symbol error probability. It is seen that to achieve the
symbol error probability equal to 10-6, 16-PAM constellation
having μ = 1 and L = 8 requests the lowest SNR value equal to
19.8 dB, while conventional PAM constellation needs higher
SNR value (24 dB). In order word, we can state that in
comparison to conventional PAM constellation our PAM
constellation specified with μ = 1 and L = 8 achieves gain in
energy efficiency equal to ΔSNR = 24 dB − 19.8 dB = 4.2 dB
for symbol error probability 10-6. This means that by
employing our PAM constellation specified with μ = 1 and
L = 8 the considerably energy saving can be achieved. More
precisely, if the desired level of symbol error probability
is 10-6, the useful signal power can be reduced by more than
double, i.e. by 104.2/10 = 2.63 times.
Fig. 3. For M = 16 and different μ and L, symbol error probability
in function of SNR
From Fig. 3 and Table I we can also draw the following
conclusion. For a given μ, it can be noted that for ΔSNR
obtained by doubling L holds: ΔSNRL = 1−ΔSNRL = 2 >
ΔSNRL = 2−ΔSNRL = 4 > ΔSNRL = 4−ΔSNRL = 8. For example, for
μ=1 holds: ΔSNRL = 1−ΔSNRL = 2 = 3 dB,
ΔSNRL = 2−ΔSNRL = 4 = 0.9, ΔSNRL = 4−ΔSNRL = 8 = 0.3dB.
Actually, benefit achieved by doubling L decreases with
increasing L. Besides this, we should have in mind that the
increase of L leads to more complex constellation. All these
points out that it should be careful with increasing L. Finally,
we compare PAM constellation proposed in this paper with
PAM constellation from [4]. The best 16-PAM constellation
from [4] achives symbol error probability equal to 10-6 for
SNR[4] = 22 dB, which means that the gain achieved in energy
efiiciency ΔSNR[4] = 24 dB − 22 dB = 2 dB is smaller than the
gain achieved with PAM constellation based on piecewise
uniform μ-law companding quantization having μ = 1 and
L = 8. The achieved gain with PAM constellation specified
with μ = 1 and L = 8 is for ΔSNR − ΔSNR[4] = 4.2−2 = 2.2 dB
greater.
TABLE I
FOR A GIVEN PE THE REQUIRED SNR OF 16-PAM CONSTELLATIONS
SPECIFIED FOR DIFFERENT μ AND DIFFERENT L (SNR* IS THE
REQUIRED SNR OF CONVENTIONAL 16-PAM CONSTELLATION,
ΔSNR = SNR* − SNR)
μ L xmax Pe SNR [dB] SNR* [dB] ΔSNR [dB]
2 21 3 ICEST 2020, Niš, Serbia, 2020.
1 4 20.1 3.9
8 19.8 4.2
2 23.4 0.6
3.7 10-6 24
3 4 22.8 1.2
8 22.7 1.3
2 25.1 −1.1
5 4 24.6 −0.6
8 24.6 −0.6
IV. CONCLUSION
In order to outperform conventional PAM constellation in
terms of energy efficiency and to achieve lower constellation
complexity than it is complexity of constellation with non-
equidistant and non-equiprobable constellation points, we
proposed method for designing piecewise uniform PAM
constellation based on piecewise uniform μ-law companding
quantization. The main issue of this paper is to determinate μ
value that provides desirable energy efficiency. We
demonstrate that the PAM constellations with values for μ
greater than 5 are not suitable solutions. On the other hand,
PAM constellation with values for μ smaller than 5
outperforms conventional PAM constellation in terms of
energy efficiency. In comparison to conventional PAM
constellation, the energy efficiency gain is up to 4.2 dB for the
symbol error probability equal to 10-6. This means that it is
possible to reduce the signal power by more than half
simultaneously retaining symbol error probability. If we
constraint constellation complexity on two subsets, the energy
efficiency gain is also considerable and amounts up to 3dB.
162
ACKNOWLEDGEMENT
This research was supported by the Science Fund of the
Republic of Serbia, 6527104, AI- Com-in-AI.
REFERENCES
[1] R. van der Linden, N. C. Tran, E. Tangdiongga, and T. Koonen,
“Optimization of Flexible Non-uniform Multilevel PAM for
Maximizing the Aggregated Capacity in PON Deployments”,
Journal of Lightwave Technology, vol. 36, no. 12, pp. 2328-
2336, 2018.
[2] I. N. Osahon, M. Safari, and W. O. Popoola, “10-Gb/s
Transmission over 10-m SI-POF with M -PAM and Multilayer
Perceptron Equalizer”, IEEE Photonics Technology Letters, vol.
30, no. 10, pp. 911-914, 2018.
[3] Y. Hsu, et all, “2.6 Tbit/s On-chip Optical Interconnect
Supporting Mode-division-multiplexing and PAM-4 Signal”,
IEEE Photonics Technology Letters, vol. 30, no. 11, pp. 1052-
1055, 2018.
[4] S. A. Vlajkov, A. Ž. Jovanović, and Z. H. Perić, “Improvement
of Energy Efficiency of PAM Constellation by Applying
Optimal Companding Quantization in Constellation Designing”,
[5] Z. H. Perić, “Nonlinear Transformation of One-dimensional
Constellation Points in order to Error Probability Decreasing”,
Facta Universitatis, Series: Electronics and Energetics, vol. 11,
no. 3, pp. 291-299, 1998.
[6] Z. H. Perić, S. M. Bogosavljević, “Performance of Nonuniform
PAM Constellations for Gaussian Channel”, Electronics and
Electrical Engineering, vol. 52, no. 3, pp. 27-30, 2004.
[7] Z. H. Perić, A. Ž. Jovanović, S. A. Vlajkov, “Comparative
Analysis of Various PAM Constellations”, SAUM 2016,
Conference Proceedings, pp. 27-30, Niš, Serbia, 2016.
[8] S. A. Vlajkov, A. Ž. Jovanović, and Z. H. Perić, “Approach in
Companding-quantization-inspired PAM Constellation Design”,
IET Communications, vol. 12, no. 18, pp. 2305-2314, 2018.
[9] I. B. Đorđević, A. Ž. Jovanović, M. Cvijetić, and Z. H. Perić,
“Multidimensional Vector Quantization-based Signal
Constellation Design Enabling Beyond 1 Pb/s Serial Optical
Transport Networks”, IEEE Photonics Journal, vol. 5, no. 4,
2013.
[10] I. B. Đorđević, A. Ž. Jovanović, Z. H. Perić, and T. Wang,
“Multidimensional Optical Transport Based on Optimized
Vector-quantization-inspired Signal Constellation Design”,
IEEE Transactions on Communications, vol. 62, no. 9, pp.
3262-3273, 2014.
[11] F. R. Kschischang, S. Pasupathy, “Optimal Nonuniform
Signaling for Gaussian Channels”, IEEE Transaction on
Information Theory, vol. 39, no. 3, pp. 913–929, 1993.
[12] N. S. Jayant, P. Noll, Digital Coding of Waveforms, Englewood
Cliffs, NJ: Prentice-Hall, 1984.
[13] J. G. Proakis, M. Salehi, Digital Communications, McGraw-
Hill, 2008.
[14] S. A. Vlajkov, A. Ž. Jovanović, and Z. H. Perić, “Comparative
Analysis of ML and MAP Detectors for PAM Constellation in
AWGN Channel”, Facta Universitatis, Series: Electronics and
Energetics, vol. 33, no. 2, pp. 217-226, 2020.
[15] A. Alvarado, E. Agrell, and F. Brannstrom, “Asymptotic
Comparison of ML and MAP Detectors for Multidimensional
Constellations”, IEEE Transactions on Information Theory, vol.
64, no. 2, pp. 1231-1240, 2018.
 How Initialization is Related to Deep Neural Networks
Generalization Capability: Experimental Study
Ljubinka Sandjakoska1 and Frosina Stojanovska2
Abstract – Deep learning has lately received significant
research effort, considering its capabilities followed by the
advantages in many domains. The focus of this paper is on
improving the generalization ability, which is a key for successful
implementation of deep neural networks. In order to answer the
question how initialization is related to the deep neural networks'
generalization capability, an experimental study is done. The
study consists of several experiments, which belong to the data-
independent initialization techniques. Using different
approaches, the change of generalization gap is presented.
Keywords – Deep learning, Initialization, Regularization,
Generalization.
I. INTRODUCTION
Deep learning has lately received significant research
effort, since its capabilities, followed by its application
advantages in many domains. The concept of deep learning
realized in neural networks, is implemented in a variety of
domains, such as: image [1-3], audio [7-8] and video
processing [9-10], natural language processing [4], [6],
automatic machine translation [5], [11], fraud detection [12-
13], in healthcare - personalized medicine, advanced
diagnosis, drug discovery [14-19], advertising [20] and
finance [21], social media [22-23], cybersecurity [24-26], etc.
The paradigm of deep learning usually is defined as
hierarchical learning, where complex nonlinear functions are
used. Learning distributed and hierarchical representation of
distinctive data features is allowed by the ability of knowledge
extraction from row data. The in-depth processing does not
indicate the data size, but the model (neural network) size, in
terms of the number of hidden layers of neurons - processing
units. Theoretically, deep learning follows connectionism
philosophy. According to that, the individual biological
neuron or the individual feature in the machine learning model
is not intelligent by itself, but a large population of neurons or
features acting together can exhibit intelligent behavior.
The intelligent behavior of deep neural networks relies on
the sequential operations on input data, that are realized in the
hidden layers of neurons. Standard machine learning
operations: initialization, activation, training, testing, error
calculation and back-propagation are included, followed by
more advanced such as recursive application of convolution
1Ljubinka Sandjakoska is with Faculty of Computer Science and
Engineering, UIST “St. Paul the Apostle”, Ohrid, Republic of North
Macedonia, E-mail: ljubinka.gjergjeska@uist.edu.mk
2Frosina Stojanovska is with Faculty of Computer Science and
Engineering, UKIM, Skopje, Republic of North Macedonia, E-mail:
frosina.stojanovska@finki.ukim.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
and pooling, full connection operations, batch normalization,
etc. The chain of transformations of inputs to outputs forms a
credit assignment path (CAP) depth. If the CAP is bigger than
2 than we have deep processing which allows deep learning.
A very important aspect of deep learning is that the layers of
features are not designed by human engineers; they are
learned from data using a general-purpose learning procedure.
The key concept is the depth of the model, which allows the
model to learn data representations with multiple levels of
abstraction.
The remainder of this paper is organized as follows. Section
II presents the theoretical insights of initialization and
regularization, with its state-of-the-art methods. Experimental
design of this study is presented in Section III, where the
description of the dataset, the experimental setup and
evaluation measures are included. The obtained results from
the experimentation and discussion are in the next section. In
the end, the concluding remarks are given.
II. THEORETICAL INSIGHTS
A. Initialization
Initialization of Deep Neural Networks (DNNs) is a starting
point of the learning process. The mechanism which
determines the basic network parameters’ behavior is known
as initialization. Initialization is based on creating a pattern for
assigning values for the basic deep networks parameters –
weights and biases. The learning process and the numerical
stability of the network depend on the values of these
parameters. Hence, initialization gives the difference between
the network that converge in reasonable time and the network
whose loss function does not change even with hundreds of
thousands of iterations. Finding an appropriate range for the
weights protects the deep network from training with too
small or too high values. Both situations cause a problem, and
the signal is not allowed to reach the deep layers. If the
weights are too small, then the signal decreases as it passes
through each layer of the network, until it is too small to be
useful. This means that the network will learn slowly. If the
weights in the network start too high, then the signal grows as
it passes through each layer until it is too large to be useful for
the divergence of the network.
Generally, the initialization strategies can be group into
initialization with zeros or other constant, and initialization
with random numbers. First, by eliminating weights the deep
model is reduced to a linear model since the derivative of the
loss function is the same for all weights in each subsequent
iteration. It is important to note that, unlike weights, if the
163
biases are 0, there will be no problem and the values for each
neuron will be different, but the symmetry will also be
disturbed. Initialization of random numbers leads to the two
previously mentioned problems of too low or high weights
known as the vanishing gradients of the activation function
and the exploding gradients. These, in turn, are often
associated with deep network optimization. Although the
differences between the optimization and regularization are
out of the scope of this paper, we will emphasize that
initialization is one of its junctions’ points.
Depending on the amount of information derived from
available data, weight initialization techniques are divided
into three main groups: Data Independent, Data Dependent,
and Pre-Training approaches. According to this classification,
for this paper, we work with data independent techniques.
Namely, we do not make use of any training data to determine
the initial parameter values of a network, but we work by
sampling numbers from different distributions and using such
numbers as the initial values.
In this experimental study are used the following
initialization strategies: Uniform, Normal, Lecun Uniform and
Lecun Normal [31], [32], Xavier Uniform and Xavier Normal
[33], Kaming Uniform and Kaming Normal [34]. The uniform
method is an initialization technique where the weights are
derived from uniform distribution with -0.05 and 0.05 as
bounds. The normal strategy sets the weights from normal
distribution with mean and standard deviation. The Lecun,
Xavier and Kaiming methods use a scale factor defined as:
= , (1)
= , (2)
( )/
= , (3)
where fanin is the number of input units and fanout is the
number of output units in the weight tensor.
B. Generalization Ability of the Network
The regularization of DNNs is one of the key concepts that
determine the success in terms of its’ effectiveness and
efficiency. Regularization techniques arise as solutions to the
problem of overfitting. This problem occurs in cases such as:
over parameterization of the DNN or small training set.
Overfitting results in decreasing the generalization ability
of the network. If the DNN can give good performance not
only on training data but also on new, previously unseen data,
then we can say that the DNN has good generalization ability.
Improving the generalization capability can be done using
different regularization strategies. According to Kukavckas’ et
al. taxonomy [30], the regularization strategy can be via: data,
the network architecture, the error function, the regularization
term and via optimization. In the research scope of this paper
the regularization is done implicitly, by working with the data.
The main idea is to see what will be the effect of initialization
on error function and loss function, without using some
explicit regularization method. The error function
expresses consistency of outputs with targets. Classical
approaches of regularization strategies work with
regularization term or adding a regularizer into the
164
loss function. The regularization term is independent of
the targets and does encoding of other properties to
provide assumptions about the mapping other than the
consistency of outputs with targets.
III. EXPERIMENTAL DESIGN
To answer the question How the initialization is related to
Deep Neural Networks Generalization capability experimental
study is done. The study consists of several experiments, that
belong to the data-independent initialization techniques. For
the purpose of this research we have tested DNNs initialized
by: Uniform, Normal, Lecun Uniform and Lecun Normal
[31], [32], Xavier Uniform and Xavier Normal [33], Kaming
Uniform and Kaming Normal [37].
The main task of this experimental study is a quantum
chemistry problem - prediction of potential energy surfaces
experiments. For that purpose, SchNet model is applied.
SchNet is a convolutional network with continuous
convolutional filters designed to model the quantum
interactions in molecules [27]. The network architecture is
consisting of input module that includes molecular descriptors
- vector of nuclear charges and an inter-atomic distance
matrix, which forms 128 embedding, for every feature. The
number of features is 128. The input module is followed by 6
interaction modules. The interactions depend on the current
atom representation, and according that - convolutional filters
are created. Actually this block consists of continuous-filter
convolution with filter-generating network with dimension
64x. The shifted softplus activation, defined as ln(0.5ex + 0.5),
is also part of this module. Specifically, this architecture is not
using weight sharing across multiple interaction blocks, but
residual connection. The total number of trained weighs is
447488. The output module is formed by neurons for the
property U0. This deep learning model is used in the
experiments for testing the importance of different
initialization techniques for the learning process and
generalization capability. The problem is to predict the total
energy of a molecule U0, which represents the sum of atomic
energy contributions of the atoms in the molecule [28].
A. Dataset Benchmark
The benchmark dataset is the QM9 dataset [29]. This
dataset contains 133,885 organic molecules with chemical
potentials and energies. The total energy of the molecules is
defined as the property U0. The dataset is randomly separated
into a train set with 70% of the molecules (94,000 molecules),
validation set with 15% (20,000 molecules) and test set with
the remaining 15% (19,885 molecules).
B. Experimental Setup and Evaluation Measures
Training and testing of the model was done with the Google
Colab, the product by Google Research, specifically with
Python 3 Google Compute Engine (Tesla P4 GPU, Intel Xeon
CPU 2.00GHz, 13GB memory). Each experiment was
executed for 50 epochs, which lasted about 2 hours. The
SchNet model was trained with 0.0001 learning rate, and an TABLE I
exponential learning rate decay with ratio 0.96. As optimizer MAE AND RMSE OF THE ENERGY U0 PREDICTIONS IN KCAL/MOL FOR
ADAM is used. The loss function for training is the mean THE TRAIN SET (TRAIN), VALIDATIONS SET (VAL) AND TEST SET (TEST)
WITH ALL OF THE INITIALIZATION TECHNIQUES
squared error (MSE) in kcal/mol. The batch size is 100, the
both beta arguments are 0.5, and the cut off function is cosine.
To evaluate the accuracy of predicted total energy of the
model, mean absolute error (MAE) and root mean squared
error (RMSE) in kcal/mol were used.

IV. RESULTS AND DISCUSSION

The results from the experiments show that the
initialization technique can be critical for the model’s ultimate
performance. Table I holds the values of MSE and RMSE
values in kcal/mol for testing of each model on the train set,
validation set and test set. From these results, it is notable that
the generalization of the models is preserved for all models,
i.e. the validation set, and test set have values that do not
differ for big difference rate from the performance of the train
set. The uniform and normal distributions with the Lecun and
Xavier settings stand out as better initialization techniques.
Figure 1 depicts the impact on the learning process of the
models with different initializations, by showing the training
loss and validation loss curves, as well as the curve of MAE
values. From these curves, it is notable that the Kaiming and
Kaiming normal techniques are the worst initializations. For
these models, it takes more time to reach the standard values
set by the other models. Lecun and Xavier have a similar
learning process with both uniform and normal variants.
The standard normal and uniform initializations have an
average performance compared to the other initializations. In
the epochs close to 50, the differences become very small, this
drives the conclusion that the number of epochs also depends
on the initialization technique. Hence, if the model is trained
for only 20 epochs, then the Xavier and Lecun initializations
terminate with a model with more prominent generalization
and learning capability than the other methods. Here should
be mentioned that, our aim is to obtain loss less than 0.1 for
not more than the half of the total number of epochs, or 25.

V. CONCLUSION

Using experimental study, we try to depict the influence of

the initialization techniques on regularization. The main
contribution of this paper is the experimentation in atomistic
simulations. Best to our knowledge, a lot of research effort in
developing DNNs is put in domains such as computer vision
or text processing but is missing developing deep learning
methods specified for chemo-informatics. Usually in this kind
of applications, general purpose procedures for initialization
and regularization are used. We believe that more domain-
specialized deep learning approaches will give better results.

REFERENCES
[1] A. Karpathy, L. Fei-Fei, “Deep Visual-Semantic Alignments for
Generating Image Descriptions”, Proceedings of the IEEE Fig. 1. a) Train loss, b) Validation loss,
c) RMSE for predicting Energy U0 in kcal/mol

165
 Conference on Computer Vision and Pattern Recognition, pp.
3128-3137, 2015.
[2] J. Donahue, L. A. Hendricks, S. Guadarrama, M. Rohrbach,
S. Venugopalan, K. Saenko, and T. Darrell, “Long-Term
Recurrent Convolutional Networks for Visual Recognition and
Description”, Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 2625-2634, 2015.
[3] J. Hwang, Y. Zhou, “Image Colorization with Deep
Convolutional Neural Networks”,Stanford University, Tech.
Rep. 2016.
[4] T. Young, D. Hazarika, S. Poria, and E. Cambria, “Recent
Trends in Deep Learning Based Natural Language
Processing”, IEEE Computational Intelligence Magazine, vol.
13, no. 3, pp. 55-75, 2018.
[5] J. Zhang, C. Zong, “Deep Neural Networks in Machine
Translation: An Overview”, IEEE Intelligent Systems, no. 5, pp.
16-25, 2015.
[6] A. R. Sharma, P. Kaushik, “Literature Survey of Statistical,
Deep and Reinforcement Learning in Natural Language
Processing”, Proceedings of IEEE International Conference on
Computing, Communication and Automation, pp. 350-354,
2017.
[7] H. Purwins, B. Li, T. Virtanen, J. Schlüter, S. Y. Chang, and T.
Sainath, “Deep Learning for Audio Signal Processing”, IEEE
Journal of Selected Topics in Signal Processing, vol. 13, no. 2,
pp. 206-219, 2019.
[8] B. Chen, W. Luo, and D. Luo, “Identification of Audio
Processing Operations Based on Convolutional Neural
Network”, Proceedings of the 6th ACM Workshop on
Information Hiding and Multimedia Security, pp. 73-77, 2018.
[9] W. Zhang, L. Xu, Z. Li, Q Lu, and Y. Liu, “A Deep-Intelligence
Framework for Online Video Processing”, IEEE Software, vol.
33, no. 2, pp. 44-51, 2016.
[10] S. Zhang, X. Pan, Y. Cui, X. Zhao, and L. Liu, “Learning
Affective Video Features for Facial Expression Recognition via
Hybrid Deep Learning”, IEEE Access, vol. 7, pp. 32297-32304,
2019.
[11] M. R. Costa-Jussà, A. Allauzen, L. Barrault, K. Cho, and H.
Schwenk, “Introduction to the Special Issue on Deep Learning
Approaches for Machine Translation”, Computer Speech &
Language, vol. 46, pp. 367-373, 2017.
[12] X. Zhang, Y. Han, W. Xu, and Q. Wang, “HOBA: A Novel
Feature Engineering Methodology for Credit Card Fraud
Detection with a Deep Learning Architecture”, Information
Sciences, 2019.
[13] A. Roy, J. Sun, R. Mahoney, L. Alonzi, S. Adams, and P.
Beling, “Deep Learning Detecting Fraud in Credit Card
Transactions”, Proceedings of Systems and Information
Engineering Design IEEE Symposium (SIEDS), pp. 129-134,
2018.
[14] R. Fakoor, F. Ladhak, A. Nazi, and M. Huber, “Using Deep
Learning to Enhance Cancer Diagnosis and
Classification”, Proceedings of the International ACM
Conference on Machine Learning, vol. 28, New York, USA,
2013.
[15] A. Mansoor, J.J. Cerrolaza, R. Idrees, E. Biggs, M. A. Alsharid,
R. A. Avery, and M. G. Linguraru, “Deep Learning Guided
Partitioned Shape Model for Anterior Visual Pathway
Segmentation”, IEEE Transactions on Medical Imaging, vol. 35,
no. 8, pp.1856-1865, 2016.
[16] D. Nie, H. Zhang, E. Adeli, L. Liu, and D. Shen, “3D Deep
Learning for Multi-Modal Imaging-Guided Survival Time
Prediction of Brain Tumor Patients”, Proceedings of
International Conference on Medical Image Computing and
Computer-Assisted Intervention, Springer, Cham, pp. 212-220,
2016.
166
[17] C. Angermueller, H. J. Lee, W. Reik, and O. Stegle, “DeepCpG:
Accurate Prediction of Single-Cell DNA Methylation States
Using Deep Learning”, Genome Biology, vol. 18, no. 1, pp.
1-13, 2017.
[18] B. Ramsundar, S. Kearnes, P. Riley, D. Webster, D. Konerding,
and V. Pande, “Massively Multitask Networks for Drug
Discovery”, arXiv preprint arXiv:1502.02072, 2015.
[19] D. Quang, Y. Chen, and X. Xie, “DANN: A Deep Learning
Approach for Annotating the Pathogenicity of Genetic
Variants”, Bioinformatics, vol. 31, no. 5, pp. 761-763, 2015.
[20] H. Zhang, Y. Ji, W. Huang, and L. Liu, “Sitcom-Star-Based
Clothing Retrieval for Video Advertising: A Deep Learning
Framework”, Neural Computing and Applications, vol. 31, no.
11, pp. 7361-7380, 2019.
[21] J. B. Heaton, N. G. Polson, and J. H. Witte, “Deep Learning in
Finance”, arXiv preprint arXiv:1602.06561, 2016.
[22] Z. Zhang, Q. He, J. Gao, and M. Ni, “A Deep Learning
Approach for Detecting Traffic Accidents from Social Media
Data”, Transportation Research Part C: Emerging Technologies,
vol. 86, pp. 580-596, 2018.
[23] M. K. Hayat, A. Daud, A. A. Alshdadi, A. Banjar, R. A. Abbasi,
Y. Bao, and H. Dawood, “Towards Deep Learning Prospects:
Insights for Social Media Analytics”, IEEE Access, vol. 7, pp.
36958-36979, 2019.
[24] S. Mahdavifar, A. A. Ghorbani, “Application of Deep Learning
to Cybersecurity: A Survey”, Neurocomputing, vol. 347, pp.
149-176, 2019.
[25] P. Viswanath, “A Study on Deep Learning in Cyber Security”,
International Journal of Engineering Science, vol. 9, no. 4, pp.
21482-21483, 2019.
[26] A. A. Diro, N. Chilamkurti “Distributed Attack Detection
Scheme Using Deep Learning Approach for Internet of Things”,
Future Generation Computer Systems, vol. 82, pp. 761-768,
2018.
[27] K. Schütt, P. J. Kindermans, H. E. S. Felix, S. Chmiela, A.
Tkatchenko, and K. R. Müller, “Schnet: A Continuous-Filter
Convolutional Neural Network for Modeling Quantum
Interactions”, Advances in Neural Information Processing
Systems, pp. 991-1001, California, USA, 2017.
[28] K. T. Schütt, F. Arbabzadah, S. Chmiela, K. R. Müller, and A.
Tkatchenko, “Quantum-Chemical Insights from Deep Tensor
Neural Networks”, Nature Communications, vol. 8, no. 1, pp. 1-
8, 2017.
[29] R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. Von
Lilienfeld, “Quantum Chemistry Structures and Properties of
134 Kilo Molecules”, Scientific Data, vol. 1, no. 1, pp.1-7,
2014.
[30] J. Kukačka, V. Golkov and D. Cremers, “Regularization for
Deep Learning: A Taxonomy”, arXiv preprint
arXiv:1710.10686, 2017.
[31] Y. A. LeCun, L. Bottou, G. B. Orr, and K. R. Müller, “Efficient
Backprop”, Neural Networks: Tricks of the Trade, pp. 9-48,
Berlin, Heidelberg, 2012.
[32] G. Klambauer, T. Unterthiner, A. Mayr, and S. Hochreiter,
“Self-Normalizing Neural Networks”, Advances in Neural
Information Processing Systems, pp. 971-980, California, USA,
2017.
[33] X. Glorot, Y. Bengio, “Understanding the Difficulty of Training
Deep Feedforward Neural Networks”, Proceedings of the
Thirteenth International Conference on Artificial Intelligence
and Statistics, pp. 249-256, Sardinia, Italy, 2010.
[34] K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into
Rectifiers: Surpassing Human-Level Performance on Imagenet
Classification”, Proceedings of the IEEE International
Conference on Computer Vision, pp. 1026-1034, Santiago,
Chile, 2015.
 Using Neural Networks Application for the Font
Recognition Task Solution
Oleksii Bychkov1, Kateryna Merkulova2, Georgi Dimitrov3, Yelyzaveta Zhabska4,
Iva Kostadinova5, Pepa Petrova6, Pavel Petrov7, Iglika Getova8 and Galina Panayotova9
Abstract – The purpose of the paper is research and
development of font visual classification system, that will allow
users to work with fonts more effectively, specifically with the
use of the required font fetching on pre-screening, employment
of the wide range of filters and font management tools
application.
There were reviewed basic fonts classification approaches with
the definition of pivot advantages and disadvantages. After
assessing these approaches, it was decided to deploy neural
network with the font symbols images as an input.
As a result, neural networks driven computer fonts
classification system, that allows to classify fonts by slope, basic
style and weight according to their graphical representation, was
developed and integrated into the font organizer. High
classification accuracy of the developed neural network model
was confirmed by control tests. Percentage of correctly classified
fonts during the system testing reaches 91% in determining
weight, 92% in determining basic style, 96% in determining
slope. The developed system can be applied to tasks of fonts
classification solution as supplement for digital documents
structure determination, and as system of font selection for
design creation.
Keywords – Convolutional neural network, Machine learning,
Image classification, Font classification.
I. INTRODUCTION
The computer font is a file that contains description of
alphabetical, digital, reserved or pseudo-graphic characters
1Oleksii Bychkov is with Taras Shevchenko National University
of Kyiv, Kyiv, Ukraine, E-mail: bos.knu@gmail.com
2Kateryna Merkulova is with Taras Shevchenko National
University of Kyiv, Kyiv, Ukraine. E-mail: kate.don11@gmail.com
3Georgi Dimitrov is with University of Library Studies and
Information Technologies, Sofia, Bulgaria, E-mail:
geo.p.dimitrov@gmail.com
4Yelyzaveta Zhabska is with Taras Shevchenko National
set, designed for displaying these characters (the text in
particular) by the operating system or the program. Every font
has its slope, weight (100-900), and one of the basic
typefaces: Serif, Sans-Serif, Decorative, Script. There is an
enormous number of fonts, but the standard of their
classification still wasn’t developed. Fonts with a common
subfamily name often considerably distinct from each other
by slope as much as by weight. In cases when development of
all weight types is not needed, the authors of the font families
create what is known as “shifts” - when actual font weight
corresponds to another weight. Also, author often names the
font as Regular after developing the single font in the sub-
family regardless of the font type. The lack of the
classification standard causes the problem of font search and
selection.
II. PROBLEM DEFINITION AND SOLUTION
METHODS
Classification task consists of 3 sub-problems: recognition
of “basic” slope, classification by weight, classification by
style.
It is possible to apply a simple approach to determine the
font’s weight: there can be calculated the number of black
pixels in the image and font can be assigned to the closest
class by this sum (the average value).
However, in the process of analyzing the fonts it was found
that font weight more depends on image edges rather than
number of black pixels. Those are outlined fonts with inside
hollow edges. This issue can be resolved using the Edge
Detection methods. An attempt to apply them was made. The
result of symbol image processing with the use of
convolutional matrix is shown in Fig. 1.

University of Kyiv, Kyiv, Ukraine, E-mail: y.zhabska@gmail.com

5Iva Kostadinova is with University of Library Studies and

Information Technologies, Sofia, Bulgaria, E-mail:

i.kostadinova@unibit.bg
6Pepa Petrova is with University of Library Studies and

Information Technologies, Sofia, Bulgaria, E-mail:

p.petrova@unitbit.bg
7Pavel Petrov is with University of Economics, Varna, Bulgaria, Fig. 1. The result of symbol image processing with the use of
Edge Detection and convolutional matrix methods
E-mail: petrov@ue-varna.bg
8Iglika Getova is with University of Library Studies and

Information Technologies, Sofia, Bulgaria, E-mail:

i.getova@unibit.bg Fig. 2. Variations of an “a” symbol
9Galina Panayotova is with University of Library Studies and

Information Technologies, Sofia, Bulgaria, E-mail:

g.panayotova@unibit.gb

978-1-7281-7142-5/20/$31.00 ©2020 IEEE 167

 Used methods are not sufficient since they don’t consider
the width of the symbol, as the study results show. Some fonts
have combined edges that are difficult to interpret as well, so
these fonts can be classified only based on patterns of human
symbol perception.
The most difficult of the sub-problems is the task of style
classification. Because of the reason, that the vast number of
different fonts are complicated to describe by basic distinct
features, the font can be classified to a certain style (Fig. 2).
It was decided to apply machine learning methods,
specifically neural networks, considering the tasks complexity
and the problem of the unassisted various font classes features
description. Neural networks are the most frequently exploited
image classification problems algorithm, because it
demonstrates the best pattern recognition and classification
results among other algorithms [1], [2].
After analyzing of how various font styles design differs, it
was decided to use Convolutional Neural Network. This
network provides an opportunity to recognize fonts features
irrespective of position and deformation – it selects important
classification patterns, such as feature maps and filters, taking
into account correct supervisor responses [3], [4].
For the reason of the wide range of symbol variations by
size and design, the use of typical Multi Layer Perceptron
(MLP) network would lead difficultness in high accuracy
classification. Also, it requires big neurons number
construction, and, as a result, power efficiency.
Slope classification task, that is the case of font belonging
to Italic type, firstly was tried to decide with the application of
MLP network, that solved the problem with the high accuracy.
Nevertheless, under continued data increasing with more
complex Script style fonts, network was reducing its accuracy.
It was resolved with the including the CNN network as well
(Fig. 3).
formula: f(x)=max(0,x). In the research [7], [8] described that
ReLU activation function can increase the stochastic descent
gradient convergence speed by factor 6 in comparison to
hyperbolic tangent and sigmoid function.
Softmax function was used for the last layer, that is
confidence of belonging to one of the classes depends on
confidence in other classes. It can be described by the formula
with the following:
=∑ , (1)
where zi is i-th neuron input value; yi is output i-th neuron
value.
The function categorical_crossentropy was used for the
error determination. It is used for Multi Class classification,
when the only one class can belong to the font. The following
formula defines this function:
, = −∑ , (2)
where p and q are values of expected and received results.
Weight update was performed with the adaptive learning
rate method, that is Adadelta algorithm. It considers weight
update and gradient values history. During the training on
simplified data sample, it demonstrates faster convergence in
comparison with the regular SGD. In Fig. 4 neural network
model is presented.
Fig. 4. Neural network model for style classification
TABLE I
IMAGE OPERATIONS

Operation Allowability of applying

Allowable for all sub-tasks, except

Rotate
for slope determination
Fig. 3. Loss function values during MLP network training before and Allowable for all sub-tasks, except
after Script type fonts inclusion Scale
for weight determination

III. EXPERIMENTAL RESEARCH RESULTS AND Horizontal flip Allowable for all sub-tasks
ANALYSIS Vertical flip Allowable for all sub-tasks
Designing and training of the network were performed with Vertical and Horizontal
the use of Keras Python library. It was decided to apply the Allowable for all sub-tasks
flip at the same time
sequential layer combinations method: Conv > Conv >
Shift, Wrap Non-allowable for all sub-tasks
MaxPooling; it is inspired by “VGG NET” architecture, that is
one of the most accurate networks for the present [5], [6]. Size
and number of kernels were defined for every task Font classification by slope was performed with the use of
individually with priority given to classification accuracy, that “IW” symbols, and font classification by weight and style –
must consist over 90%. It was established, during the testing with the use of “RoWGa”. These symbols appropriately
stage, that slope classification with the 5x5 sized Conv layers represent the typical features of corresponding classes, as was
8 kernels and 3x3 sized 16 kernels for classification by weight established during analytical research. Corresponding image
and style demonstrates the highest results. sizes are 90х36 and 350х36.
For all layers, except for the last one, ReLU activation During the research, the basic image operations were
function was used. This function can be defined by following studied and their applying possibilities for data augmentation
were determined [9], [10].

168
 Rotate and Scale operations are applicable for several data TABLE III
samples, however the usage of these operations assumes SAMPLES DATA AMOUNT AND THE NETWORK TRAINING RESULT
image palette enlargement to prevent the loss of fonts
features. This will cause the required power supply increasing, Neural network
Style Confidence
that is not appropriate for tasks with limited computer activation
resources.
In the process of testing there were prepared and classified
the following number of fonts: 826 fonts were classified by Script 0.54
style, 664 – by weight, 1128 – by slope. Image generation and
execution of operations for image number augmentation and
data balancing provided the following results: 12754 images Decorative 0.46
were classified by style with the use of 3 mirroring and letter
reordering types; 12074 images were classified by weight the
same way; 4628 images were classified by slope with the use Serif or Sans-
0.0
of size modification and letter reordering. serif
Neural network training example is depicted in Fig. 5.
Figure demonstrates that network was trained correctly, TABLE IV
validation loss value and training loss value are close, that WIREONE SANS-SERIF FONT ANALYZING WITHOUT GUIDED
means retraining effect wasn’t revealed. Also, the stochastic BACKPROPAGATION
gradient descent appliance can be seen on the graph. Because
of validation loss curve, that roll down in the figure, it can be Confid
concluded that configuration is advantageous. Table 2 Style Neural network activation
ence
contains the Loss and Accuracy functions values for every
data sample.
Serif 0.02

Sans-serif 0.44

Decorative 0.22

Fig. 5. Loss and Accuracy functions graph during the training of Script 0.33
classification by style

TABLE II TABLE V
SAMPLES DATA AMOUNT AND THE NETWORK TRAINING RESULT OVERPASS THIN ITALIC FONT ANALYZING WITHOUT GUIDED
BACKPROPAGATION
Traini
Test
Trainin ng loss_t acc_t loss_te acc_t
g type image
image
rain rain st est Neural network
s Style Confidence
s activation
By style 12274 480 0.09 0.924 0.12 0.909
By
11594 480 0.1 0.937 0.13 0.91
weight Italic 0.85
By
4168 460 0.08 0.946 0.1 0.938
slope

During the analytical research of the neural network

principle of association establishment between the classes, the
Gradient-weighted Class Activation Mapping visualization
method was used [11-13]. It presents the heat map, that
depicts areas that attract neural network attention and intensity
of the attention.
Fig. 6. Activation heat map of Decorative style recognition class
Fig. 7. Activation heat map of Serif style recognition class
Non-italic 0.15
The neural network attention focus on unusual FasterOne
font lines demonstrated in Fig. 6. It means that the font can be
classified as Decorative style. In Fig. 7 is depicted activation
in serif location areas that is typical for Serif class.
Used method can provide activation areas of each class.
Table 3 contains the examples of controversial points, when
the neural network wasn’t confident of the class the font
belongs to.

169
 The AccuracyChecker module, that can be used for the external devices" and National Science Program "Information
assessment of the correctly classified fonts percentage was and Communication Technologies for Unified Digital Market
developed to check the classification accuracy. For each in Science, Education and Security". Part of the research
classification by style, weight and slope 100 fonts were equipment was donated by a company FADATA.
prepared. With the use of neural network module performed
the font classification and comparation of the classification REFERENCES
with the correct answer. After that percentage of the right
answers were counted and displayed. Module performing [1] M. Matusugu, K. Mori, Y. Mitari, and Y. Kaneda, “Subject
results is presented in Table 3.5. Network training metric Independent Facial Expression Recognition with Robust Face
comparation with performed accuracy check is presented in Detection Using a Convolutional Neural Network”, Neural
Fig. 8. Networks, 2013, vol. 16, no. 5-6, pp. 555-559.
[2] Y. Zhu, T. Tan, and Y. Wang, “Font Recognition Based on
TABLE VI Global Texture Analysis”, IEEE Transactions on Pattern
CLASSIFICATION ACCURACY CHECK RESULTS Analysis and Machine Intelligence, 2001, vol. 23, no. 10, pp.
Class Accur 1192-1200.
Classification Test fonts [3] C. Сhapman, “Understanding the Nuances of Typeface
es acy
Classification”, Retrieved from https://www.toptal.com/
By slope 100 2 96%
designers/typography/typeface-classification
By weight 100 9 91% [4] D. Soni, “Dealing with Imbalanced Classes in Machine”,
By style 100 4 92% Retrieved from https://towardsdatascience.com/dealing-with-
imbalanced-classes-in-machine-learning-d43d6fa19d2
[5] D. Tao, X. Lin, L. Jin, and X. Li, “Principal Component 2-D
loss_train acc_train loss_test Long Short-Term Memory for Font Recognition on Single
Chinese Characters”, IEEE Transactions on Cybernetics, 2015,
acc_test Accuracy vol. 46, no. 3, pp. 756-765.
0.937

0.946

0.938
0.924

0.909

[6] Z. R. Königsberg, “A Mixed Lyapunov-Max-Plus Algebra

0.96
0.92

0.91
0.91
0.13
0.12

Approach to the Stability Problem for Discrete Event

0.09

0.08
0.1

0.1

Dynamical Systems Modeled with Timed Petri Nets”, Neural

BY STYLE BY WEIGHT BY SLOPE
Fig. 8. Networks training metric and classification accuracy
comparation
IV. DISCUSSION AND CONCLUSION
The main problem of fonts classification consists in features
indistinctness and strong variation that complicates the
process of classification. In this paper the basic fonts
classification approaches of fonts classification were reviewed
with the definition of pivot advantages and disadvantages.
After assessing these approaches, it was decided to deploy
neural network with the font symbols images as an input.
Neural network selects patterns, such as features maps, filters,
that it can be considered in classification process based on the
correct answers of the supervisor.
Neural networks driven computer fonts classification
system, that allows to classify fonts by slope, basic style and
weight according to their graphical representation, was
developed. Percentage of correctly classified fonts during the
system testing reaches 91% in determining weight, 92% in
determining basic style, 96% in determining slope. The
developed system can be applied to tasks of fonts
classification solution as supplement for digital documents
structure determination, and as system of font selection for
design creation.
It can be concluded, based on results, that neural networks
successfully solved the fonts classification problem.
ACKNOWLEDGEMENT
Parallel & Scientific Computations, 2011, vol. 19, no. 1-2, pp.
35-50.
[7] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet
Classification with Deep Convolutional Neural Networks”,
Retrieved from http://papers.nips.cc/paper/4824-imagenet-
classification-with-deep-convolutional-neural-networks.pdf.
[8] O. Bychkov, K. Merkulova, and Y. Zhabska, “Software
Application for Biometrical Person’s Identification by Portrait
Photograph Based on Wavelet Transform”, 2019 IEEE
International Conference on Advanced Trends in Information
Theory (ATIT), pp. 253-256.
[9] C. Tensmeyer, D. Saunders, and T. Martinez, “Convolutional
Neural Networks for Font Classification”, 2017 14th IAPR
International Conference on Document Analysis and
Recognition (ICDAR), 9-15 November 2017, DOI:
10.1109/ICDAR.2017.164
[10] J.-P. Aumasson, Ç. Çalık, W. Meier, O. Özen, R. C.-W. Phan,
and K. Varıcı, “Improved Cryptanalysis of Skein”, International
Conference on the Theory and Application of Cryptology and
Information Security ASIACRYPT, December 6-10, 2009,
Tokyo, Japan, pp. 542-559.
[11] X. Glorot, Y. Bengio, “Understanding the Difficulty of Training
Deep Feedforward Neural Networks”, Proceedings of the
Thirteenth International Conference on Artificial Intelligence
and Statistics (AISTATS) 2010, Chia Laguna Resort, Sardinia,
Italy, vol. 9 of JMLR: W&CP 9, pp. 249-256.
[12] S. Ruder, “An Overview of Gradient Descent Optimization
Algorithms”, Retrieved from http://ruder.io/optimizing-
gradient-descent/index.html
[13] N. Samadiani, H. Hassanpour, “A Neural Network-Based
Approach for Recognizing Multi-Font Printed English
Characters”, Journal of Electrical Systems and Information
Technology, vol. 2, no. 2, September 2015, pp. 207-218,
https://doi.org/10.1016/j.jesit.2015.06.003

This work is supported by the PPNIP-2020-03/09.03.2020

"Extraction, processing and analysis of parametric data from

170
SIGNAL PROCESSING II
 Adaptive 3D Interpolation of Images
Rumen P. Mironov
Abstract - In this paper a scheme of adaptive 3D interpolator,
suitable with the structure of neighborhood picture elements in
local area 2x2x2 is developed. The type of the interpolation
changes adaptively depending on the presence or absence of
contours in the 3D area of the studied fragment, which improves
the visual quality. On this base an algorithm for adaptive 3D
interpolation of halftone images, worked in Matlab environment
for operating system Windows 7 is developed. The given
experimental results from working of developed interpolator in
different directions, based on a set of real test video sequences
are obtained.
Keywords - Digital image processing, 3D adaptive image
sequences interpolation, Tensor processing.
I. INTRODUCTION
The basic methods for 3D interpolation of halftone image
sequences can be separated on non-adaptive (nearest, tri-
linear, 3D cubic, spline or tri-cubic interpolation) and adaptive
interpolations [1-4]. There is also a large set of methods for
interpolating of video sequences that use in one form or
another the motion vector between the individual key frames
[5-6].The basic idea of all of them is 2D interpolation of the
two key frames of the sequence and then restoration of the
intermediate frames.
With the increasing of interpolation order of the images the
non-adaptive methods decreases the sharp of the transitions.
At the same time, as the interpolation order increases, false
contours appear in the homogeneous areas of the images. [3],
[5]. To reduce them different adaptive methods for 3D
interpolation of image sequences are developed [5-6] and etc.
Based on developed in [7] method for 2D interpolation of
halftone images the new adaptive 3D interpolation method is
developed. The adaptation is performed on the base of
analysis of local characteristics of brightness of 8
neighborhood picture elements in two reference frames. On
the other hand, these points represent a tensor of rank 3. The
interpolation order is changed adaptively (zero or bilinear) in
the local area of the reference frames and after then the points
of the intermediate frames are calculated in the same way. In
this paper the new adaptive 3D interpolation method for
image slices is presented and estimation of mean-square errors
and signal to noise ratios for the developed method and
trilinear interpolation are described.
Rumen P. Mironov is with the Faculty of Telecommunications,
Technical University of Sofia, Boulv. Kl. Ohridsky 8, 1000 Sofia,
Bulgaria, E-mail: rmironov@tu-sofia.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
II. MATHEMATICAL DESCRIPTION
The input reference frames A and B of size MxN from
m-level halftone image sequences and the output interpolated
frames A* and B* of size pMxqN can be represented by the
matrices as described in Eq. (1) and Eq. (2):
A MxN = {a(i, j ) / i = 0,M − 1; j = 0, N − 1 },
(1)
B MxN = {b(i, j ) / i = 0,M − 1; j = 0, N − 1 },
A *pMxqN = {a* (k, l) / k = 0, pM − 1; l = 0, q N − 1} ,
(2)
B *pMxqN = {b* (k, l) / k = 0, pM − 1; l = 0, q N − 1} ,
where p and q are the interpolation coefficients in horizontal
and vertical directions respectively [7].
The interpolated image elements a*(k,l) and b*(k,l) in these
frames can be presented as linear convolution of the 4
neighbor elements a(i,j) and b(i,j):
1 1
a * (k, l) =   w m,n (r, t)a(i + m, j + n ),
m =0 n =0
1 1
, (3)
b (k, l) =   w m,n (r, t)b(i + m, j + n ),
*
m =0 n =0
where: r = 0, p ; t = 0, q . The interpolation coefficients given
in the following expression:
w m,n (r, t) = F.Zrm,n (r, t) + F.Bl m,n (r, t) , (4)
are dependent from the logical function F which give the type
of the interpolation (zero or bi-linear). The dependence of the
logical function F from the kind of brightness transitions in
current local area is given in [7].
The 3D interpolation can by represented as a generalized
3D tensor of rank 3 with size 2x2x2 using the scheme shown
on Fig. 1.
The First Step is calculation of interpolated points a*(k,l)
and b*(k,l) in referenced frames by the adaptive 2D
interpolation algorithm.
The Second Step is calculation of point c*(k,l) in the frame
C, using the points a*(k,l) and b*(k,l) with the same adaptive
interpolation algorithm in 1D direction.
The presented algorithm is compared with well known
trilinear interpolation algorithm [4], which is realized by
following three steps:
First Step: Calculates the points a00, a01 in frame A and b00,
b01 in frame B by bilinear interpolation in horizontal direction.
Second Step: Calculates the points c00, c01 in frame C by
bilinear interpolation in depth direction.
Third Step: Calculate the point c(k,l) in frame C by bilinear
interpolation in vertical direction.
173
 b(i,j) b00 b(i,j+1)

c00

a(i,j+1)
a(i,j)
a00 b*(k,l)
B

c*(k,l)
C 12TH
A a*(k,l)
b(i+1,j)
b01
b(i+1,j+1)
c00

a(i+1,j) a01 a(i+1,j+1)

Fig. 1. Generalized tensor presentation of 3D interpolation

III. EXPERIMENTAL RESULTS

On the base of presented method an algorithm for adaptive 13TH

3D interpolation of halftone image sequences, worked in
Matlab environment for operating system Windows 7 is
developed.
The test movie “XYLOPHONE.MPG” with size 320x240
and 256 brightness levels is used for the experiments.
In Fig. 2 four frames from this movie – the 12th, 13th, 14th
and 15th frames with size 320x240 and 256 brightness levels
are presented. The 12th and 15th frames are referenced and 13th
and 14th are calculated.
In Fig. 3 the 12th and 15th frames with size 160x120 and
256 brightness levels after 3D decimation with factor ½: ½:
⅓are presented. The 13th and 14th fames are removed. 14TH
In Fig. 4 the same reconstructed frames – 12th, 13th, 14th and
15th frames with size 320x240 and 256 brightness levels after
trilinear interpolation algorithm with factor 2:2:3 are
presented. In Fig. 5 are presented the same reconstructed
frames – 12th, 13th, 14th and 15th frames with size 320x240 and
256 brightness levels after adaptive 3D interpolation
algorithm with factor 2:2:3.
For quality measurement the following metrics ware used -
Mean Square Error (MSE), Normalized Mean Square Error
(NMSE), Signal to Noise Ratio (SNR) and Peak Signal to
Noise Ratio (PSNR), equations are given below. The MSE,
NMSE, SNR and PSNR gives the objective representation of 15TH
the restored image quality [8]. Fig. 2. Input four frames (12th, 13th, 14th, 15th) of movie
1 M −1N −1 “XYLOPHONE.MPG”
MSE =   ( X i , j − X̂ i , j ) 2 ;
M × N i = 0 j =0
M −1N −1 ( X i , j − X̂ i , j ) 2
NMSE =  ;
i = 0 j =0 ( X i , j )2 (5)
1
SNR = 10 lg , dB;
NMSE
255 2 Fig. 3. The basic two frames (12th and 15th) of movie
PSNR = 10lq , dB . “XYLOPHONE” after 3D decimation by factor ½: ½: ⅓
MSE

174
 12th 12th

13th 13th

14th 14th

15th
Fig. 4. Reconstructed frames of movie “XYLOPHONE” after
trilinear interpolation by factor 2:2:3
In Table I the results obtained from calculation of MSE,
NMSE, SNR and PSNR for 12th, 13th, 14th and 15th frames for
trilinear interpolation and the adaptive 3D interpolation
algorithms are shown.
15th
Fig. 5. Reconstructed frames of movie “XYLOPHONE” after
adaptive 3D interpolation by factor 2:2:3
In comparing with other high order 3D interpolations – 3D
spline, tri-cubic and approaches with using neural networks
[3-6], the presented algorithm owns lower calculation
complexity and better visual quality (see Figs. 2, 4 and 5).

175
 TABLE I video sequences but for reconstruction of 3D graphic objects
in medicine, digital image processing and visualization
Frame 12 Frame 13 Frame 14 Frame 15
systems as: digital photography, videoconference systems,
Trilinear Interpolation
security systems and etc.
MSE 142.4325 128.3140 129.5764 141.5571
NMSE 0.0094 0.0085 0.0085 0.0092
SNR ,dB 20.2865 20.6958 20.6912 20.3555 ACKNOWLEDGEMENT
PSNR, 26.5947 27.0481 27.0055 26.6215
dB
The author thanks the National Research Fund of the
Adaptive 3D Interpolation Bulgarian Ministry of Education and Science for the financial
MSE 96.4109 86.2590 87.3387 96.0127 support by the contract KP-06-H27/16.
NMSE 0.0063 0.0057 0.0057 0.0062
SNR ,dB 21.9814 22.4205 22.4044 22.0416
PSNR, 28.2895 28.7728 28.7187 28.3075 REFERENCES
dB
[1] D. V. Pham, V. C. Phan, "Proposing Storage Structures and
Interpolation Algorithms of 3D Spatial Data", Proc. on 4th EAI
IV. CONCLUSION International Conference on Nature of Computation and
Communication, ICTCC 2018, pp 81-97, Context-Aware
On the base of presented experimental results for error Systems and Applications, and Nature of Computation and
estimation of adaptive 3D interpolation on halftone image Communication, Eds. Phan Cong Vinh, Vangalur Alagar,
Springer, 2018.
sequences the following conclusions can be made:
[2] Z. Jia, Y.-J. Chang, T.-H. Lin, and T. Chen, "Dense
- the given results from comparing of trilinear and Interpolation of 3D Points based on Surface and Color", Proc.
adaptive 3D interpolations shows increasing of SNR on 18th IEEE International Conference on Image Processing,
and PSNR with about 2 dB; pp. 885-888, 2011.
- the developed 3D interpolation filter doesn't include [3] B. Sun, S. Zhou, "Study on the 3D Interpolation Models Used
phase distortions in the processing images; in Color Conversion", IACSIT International Journal of
- the use of optimal thresholds for selection of Engineering and Technology, vol. 4, no. 1, February 2012, pp.
homogenous and contour blocks lead to decreasing of 57-60.
mean-square error, normalized mean-square error and [4] C. Q.T. Thanh, N. T. Hai, "Trilinear Interpolation Algorithm
for Reconstruction of 3D MRI Brain Image", American Journal
increasing of signal to noise ratio and peak signal to
of Signal Processing, vol. 7, no. 1, pp. 1-11, 2017.
noise ratio with about 7-10%; [5] A. P. Mahmoudzadeh, N. H. Kashou, "Evaluation of
- to minimize mean-square error in horizontal and Interpolation Effects on Upsampling and Accuracy of Cost
vertical directions for p and q must be used variables in Functions-Based Optimized Automatic Image Registration",
the space from 2 to 5; International Journal of Biomedical Imaging, vol. 2013, Article
- the maximal effective interpolation for the local ID 395915, 19 pages, Hindawi Publishing Corporation, 2013.
characteristic of images can be achieved by using [6] S. Niklaus, L. Mai, and F. Liu, "Video Frame Interpolation via
coefficients p, q between 2 and 4. Adaptive Convolution", arXiv:1703.07514, 2017.
[7] R. Mironov, Local Adaptive Interpolation of Halftone Images,
The developed method leads to significant reduction of New Approaches in Intelligent Image Processing, Ed. R.
image distortions in the reference frames and in the Kountchev, Publ. by WSEAS Press., Chapter 13, pp. 195-212,
intermediate frames of the interpolated sequences and can also 2013.
be generalized to higher dimensions and can replace the much [8] W. Pratt, Digital Image Processing, Wiley Interscience, New
more complex multidimensional tensor representations. The York, 2007.
described interpolation algorithm can be used not only for

176
 Noise Adaptive Wiener Filtering of Images
Lora Petkova1 and Ivo Draganov2
Abstract – In this paper we propose a technique for application
of the Wiener filter over images with unknown level of present
Gaussian noise. The variance of the noise is predicted by
exploring the intensity distribution of relatively homogenous
areas from the image and then passed to the Wiener filter as an
input argument prior to applying it in contrast to its traditional
implementation. Also, extensive study has been made on the
influence of the filter mask size over the quality of processed
images given this approach. Promising results are gathered with
comparison to the original implementation of the Wiener filter.
Keywords – Image denoising, Gaussian noise, Wiener filter,
Adaptive filtration.
I. INTRODUCTION
Image filtering is important process [1] as a preliminary
stage in various detection, recognition and analysis systems,
which take visual input. Camera sensors, optics, compression
algorithms and the surrounding environment all affect
resulting images with various types of noises at different
levels. These noises directly render influence over the
accuracy of the algorithms used for further processing of the
images. One of the popular and widely used filters against
image noise is the Wiener filter [2].
Larhmiri [3] propose denoising system for medical images,
in particular such produced by magnetic resonance imaging
(MRI), X-rays and computed tomography, employing Wiener
filtering. It is applied in iterative fashion with stop criterion
based on energy decrease. Improved quality of the images is
reported with comparison to other algorithms like wavelet
packet, nonlocal Euclidean means and the original Wiener
implementation.
Hasan and El-Sakka [4] propose the structural similarity
index (SSIM) as a measure for optimization of the work of the
Wiener filter which in turn is the core of the Block matching
and 3D filtering (BM3D). Additionally, 3D thresholding has
been introduced which further improves the efficiency of the
algorithm beyond the capabilities of the BM3D alone.
Suresh et al. [5] propose two-dimensional Wiener filter
with finite impulse response which incorporated the adaptive
cuckoo search. It was specifically designed for multispectral
satellite images in which noise with normal distribution could
be detected with variances of different level. Convergence
studies reveal the algorithm to be stable and comparable in
performance to 10 other popular algorithms.
1Lora
Petkova is with Technical University of Sofia, Faculty of
Telecommunications, 8 Kliment Ohridski Blvd., 1756 Sofia,
Bulgaria, E-mail: lora.petkoovaa@gmail.com
2Ivo Draganov is with Technical University of Sofia, Faculty of
Telecommunications, 8 Kliment Ohridski Blvd., 1756 Sofia,
Bulgaria, E-mail: idraganov@tu-sofia.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
Anam et al. [6] combine Wiener filter with edge detection
and dilation in order to filter computed tomography images
using wire phantoms. They make it possible to estimate the
inherent noise from a multislice scanner. All three parameters
– the kernel size of the filter, the edge detector activation level
and the dimensions of the morphological operator were varied
in order to achieve optimal results of the filtering which prove
better than the base implementation of the Wiener filter.
All discussed techniques lead to significant increase of the
noise removal from digital images but they either demand
considerable computational and memory resource or are too
specific to the type of filtered images. In this study we
propose a much simpler technique of applying the Wiener
filter with the approximated noise variance as input argument
estimated from input image with AWGN. Additionally, we
test various filter masks in size and try to find a dependence of
the output image quality to the noise variance. As a result a
flexible scheme for the selection of the mask dimensions is
proposed. In Section II follows a description of the proposed
algorithm with experimental results in Section III and finally a
conclusion is made in Section IV.
II. PROPOSED ALGORITHM
A. Original Wiener Filter Implementation
Wiener filter [7] works over digital images suppressing
effectively Additive White Gaussian Noise (AWGN)
according to the following expression:
(, )= + ( ( , )− ), (1)
where I is the intensity of a pixel locating in the i-th row and
the j-th column of the input image, J – the intensity of a pixel
from the output image, and σ2 – variance of the present in the
input image noise. The average brightness or also the local
mean from the mask κ of the filter with size P x Q pixels is
denoted with m and it equals:
= ∑, ( , ). (2)
∈
The variance is found from:
= ∑, ( , )− . (3)
∈
If σ2 is not known, then the filter takes the mean of local
variances for all positions of the mask. This is significant
limitation in the processing principle of the Wiener filter both
from computational point of view and with regards of the
quality of filtered images. Below, we propose a way to
overcome this limitation.
177
B. Approximation of the Noise Variance dimensions 256x256 pixels, all in 8-bit grayscale form the
image database. They were consecutively and homogenously
According to the approach of Liu et al. [8] the noise filled three times with Additive White Gaussian Noise with
variance from unknown image could be approximated in the zero mean and variance of 0.001, 0.01 and 0.1.
following way. Considering the noise as additive, for a given At first, the variance of present noise was approximated.
square area of the image the following equation in vector form Results are shown in Table I.
holds [8]:
( , ) = ( , ) + ( , ), (4)
where I is the noisy image area with its centre located at the
(i0, j0)-th pixel, I0 – original area content and n – only the
noise over the same zone. Along arbitrary direction
represented by its unit vector v, it then follows that [8]:
( )= ( )+ , (5)
where var{.} means variance estimation. Minimal variance
could be found from set of tested directions using the
Principal Component Analysis (PCA) [8] corresponding to the
direction vmin which is also the eigenvector of the minimal
eigenvalue of the covariance matrix [8]:
= ∑ ( − )( − ) , (6)
where N is the number of analysed areas over horizontal and
vertical direction. From (6) it could easily be derived that [8]:
( )= + , (7)
where λmin is the minimal eigenvalue of the related covariance
matrix. Applying (7) over relatively homogenous part of the
Fig. 1. Proposed algorithm
image will return λmin(KI0) ≈ 0 and then the approximated
variance will be [8]:
It seems that the noise variance is most accurately estimated
= ( ). (8) for mid-levels around 0.01 and approximation error increases
The selection of flat enough areas from the noisy image for lower and higher levels. Variance of about 0.1 and bigger
based on intensity change [8] is done by finding the is not typical for most practical cases in daily-based mass user
covariance matrix of the gradient map [9] from each tested photography. At these magnitudes enhancement of the
zone. Then, the eigenvectors and eigenvalues of it are incorporated algorithm needs to be developed. Approximation
calculated and the maximal eigenvalue selected. If it goes time is found as 0.4067 sec on average for all images.
under a certain threshold, found by iterative testing, the area is The influence of the original Wiener filter mask size on the
considered flat and the approximate noise variance is quality of processed images at the same 3 noise variances is
estimated. given in Fig. 2 for PSNR and in Fig. 3 for SSIM.

C. Proposed Algorithm Flowchart TABLE I

NOISE VARIANCE APPROXIMATION ACCURACY
Once the approximated noise variance is found the Wiener
filter (1)-(3) could be applied with it but still its second Original
0.0010 0.0100 0.1000
parameter – the mask size could take and arbitrary integer variance
value, greater than 2. Within our study we test wide range of Approximated
0.0011 0.0098 0.0701
mask sizes w using the approximated noise variance and try to variance
find optimal wopt depending on it. Exemplary values are given Relative error,
10.00 2.00 29.90
in Section III – Experimental results. The complete set of %
steps of the proposed here algorithm is presented by the
flowchart in Fig. 1. Most efficient the original Wiener filter tends to be for
images with noise variance 0.001 where the highest values
emerge for both PSNR and SSIM. In particular, the filtration
III. EXPERIMENTAL RESULTS with a 3x3 mask is the most successful, as we observe how by
reducing the size of the mask, the filtration efficiency
The testing platform consists of IBM PC compatible decreases, the same is observed for SSIM, where the
Lenovo computer with 64-bit Intel Celeron CPU B830 difference between the best and the least performing
operating at 1.80 GHz, 4 GB of RAM, and OS Windows 10 realization is only 0.023.
Pro. All tests were performed in Matlab R2014A For Gaussian noise, with increasing variance, the filtration
environment. Six, non-compressed, popular images, namely efficiency decreases. At σ2 = 0.01 the best result is observed
Boat, Barbara, Clown, Lena, Peppers, and House with

178
for PSNR when applying a mask 5x5 - 26.22 dB and the worst
for 11x11 - 24.8 dB. For SSIM the greatest efficiency is for a
mask 7x7 - 0.735.
At σ2 = 0.1 there is a peak for the medium-sized mask of
9x9 in PSNR, and for SSIM the results are equally satisfactory
for a mask from 7x7 to 11x11 with a value for all of them
equal to 0.436.

Fig. 4. Resulting PSNR from both filters

The average filtering time of the proposed filter is less and

works better for all masks with an exception of the case for
5x5 pixels.
Fig. 2. Achieved PSNR by Wiener filter at various mask sizes

The average PSNR after processing all test images,

containing noise with σ2 = 0.1, 0.01 and 0.001, with the
Wiener filter (W) and the proposed (P) one when mask size
varies between 3 and 11 is given in Fig. 4. Peak value appears,
most like spikes, for w = 3 being 30.79 dB for the W and
33.02 dB - for the P at σ2 = 0.001. These are the global
maximums. Another spikes could be observed for w = 5 at σ2
= 0.01, and w = 7 at σ2 = 0.1 where both filters produce almost
identical resulting images with PSNR around 25.8 dB and 20
dB, respectively.

Fig. 5. Preservation level of the image structure after filtering

We can also conclude that compared to the Wiener filter for

which there is a PSNR varying 0.07 to 0.4 dB higher than the
proposed implementation, they come to be almost equally
effective. In SSIM, minimal differences of 0.005 to 0.05 in
favour of the Wiener filter are again observed. On the other
hand, the execution time of the proposed filter is less for all
mask sizes.
Visual comparison between the resulting image of Lena
obtained from the original Wiener filter and the proposed here
Fig. 3. Mask size influence over SSIM for Wiener filter
implementation could be seen in Fig. 7. Edges in the Wiener
filtered image are somewhat more deteriorated and the overall
The increase in filtering efficiency with respect to SSIM
presence of artefacts in homogenous areas is higher. In the
(Fig. 5) is most pronounced at 5x5 pixels with a difference of
image processed with the proposed realization the same zones
just over 0.1. For the other sizes of the mask the difference
are smoother and look more natural.
decreases to 0.5. The average difference in execution time
Based on average quality results for all tested images
(Fig. 6) of the filtration in the proposed variant with the most
experimenting with wide range of mask sizes and at various
effective noise suppression is about 1 ms. From the obtained
variances of induced noise, the following flowchart for
results we can estimate that the filter is most effective with a
selecting optimal parameters could be proposed (Fig. 8).
5x5 mask.

179
 Fig. 6. Processing time at various noise levels and mask size
a) Original b) Noisy
c) Wiener filtered d) Proposed filter result
Fig. 7. Image of Lena at various processing stages
IV. CONCLUSION
The presented algorithm for digital image adaptive filtration
suppress additive white Gaussian noise in natural images
leading to PSNR up to 33 dB. Its efficiency in terms of overall
(PSNR) and structural (SSIM) preservation of the original
image content is higher than that of the Wiener filter on which
it is based. The mask size of the proposed implementation
could be adapted to various noise levels in stepwise fashion.
Image category classification based on content and source
origin with their naturally induced noises could also be
employed in further development of the proposed filter.
Additional testing over larger sets of images would extend the
scheme for adaptive selection of the mask size.
180
Fig. 8. Adaptive selection of proposed filter mask size
REFERENCES
[1] R. Gonzalez, R. Woods, Digital Image Processing, 4 ed.,
Pearson, 2017.
[2] W. Pratt, Introduction to Digital Image Processing, CRC Press,
2013.
[3] S. Lahmiri, “An Iterative Denoising System based on Wiener
Filtering with Application to Biomedical Images”, Optics &
Laser Technology, vol. 90, pp. 128-132, 2017.
[4] M. Hasan, M. El-Sakka, “Improved BM3D image Denoising
using SSIM-optimized Wiener Filter”, EURASIP Journal on
Image and Video Processing, Paper No. 25, 2018.
[5] S. Suresh, S. Lal, C. Chen, and T. Celik, “Multispectral Satellite
Image Denoising via Adaptive Cuckoo Search-based Wiener
wilter”, IEEE Transactions on Geoscience and Remote
Sensing, vol. 56, no. 8, pp. 4334-4345, 2018.
[6] C. Anam, T. Fujibuchi, T. Toyoda, N. Sato, F. Haryanto, R.
Widita, F. Haryanto, R. Widita, I. Arifand, and G. Dougherty,
“An Investigation of a CT Noise Reduction using a Modified of
Wiener Filtering-edge Detection”, Journal of Physics:
Conference Series, IOP Publishing, vol. 1217, no. 1, p. 012022,
May 2019.
[7] J. Lim, Two-Dimensional Signal and Image Processing,
Englewood Cliffs, NJ, Prentice Hall, 1990.
[8] X. Liu, M. Tanaka, and M. Okutomi, “Noise Level Estimation
using Weak Textured Patches of a Single Noisy Image”, 19th
IEEE International Conference on Image Processing, pp. 665-
668, Orlando, FL, USA, Sept. 30 – Oct. 3, 2012.
[9] X. Zhu, P. Milanfar, “Automatic Parameter Selection for
Denoising Algorithms using a No-reference Measure of Image
Content”, IEEE Transactions on Image Processing, vol. 19, no.
12, pp. 3116-3132, 2010.
 Along-Track and Cross-Track Noise Analysis of Minimal
Curvature on Mars Orbiter Laser Altimeter Data
Nicole Christoff1 and Agata Manolova2
Abstract – The planetary surfaces are explored by ground
samples or by indirect observations such as images and
topographic data. This article examines the topography data
from the Mars orbiter laser altimeter (MOLA). The analysis is
based on the noise variance of the minimal curvature. A
histogram approach to assessing longitudinal and transverse
lines laying on geographical features (such as craters) in
altimeter data is suggested.

Keywords – Histogram analyse, Along-track, Cross-track,

Altimeter data.

I. INTRODUCTION
The Mars surface is explored from numerous space Fig. 1.Representation of the maximum curvature k1 and the
missions, such as NASA’s Mars Global Surveyor (MGS) [1]. minimum curvature k2 lying on a surface
The information for the surface relief is acquired with the
altimeter on the board on the spacecraft. Typically, a band at
the surface along the trajectory of the satellite’s orbit is III. METHOD
recorded [2]. They are two types of traces: along-tracks and
cross-tracks. The first ones follow the longitudinal direction Consider the plane, defined by , , (see Fig. 1
and the second ones follow the orthogonal direction along the (courtesy Polette et al. [6])), a triangulated surface to an
trajectory of the satellite (the transverse direction) [3]. intersection point of two osculating circles, with a normal
On MOLA data, Jarmołowski and Łukasiak [4] focuses on and tangent vector . The maximum curvature (k1) and the
the assumption of the presence of white noise in the data. In minimum curvature (k2), where , are shown in red
two perpendicular directions they performed maximum and blue. The main directions are orthogonal and associated
likelihood (ML) estimation on the covariance model. It is used with the values of k1 and k2, with:
to identified the correlation lengths and the result matrix is
1
fitted into the empirical estimation function. k1 = , (1)
r1
II. DATASET 1
k2 = . (2)
For this research, we use MOLA data [5] on a scale of 128 r2
pixels per degree, where the resolution is ∼500 m per pixel.
There is another method for calculating curvatures and
The sample represents the equidistant cylindrical projection
principal directions. The normal curvature , for all directions
and has the following geographic coordinates: Longitude:
at any point on the surface, is defined by the ratio:
88.87◦W ∼ 61.26◦W Latitude: 14.85◦S ∼ 25.69◦S and cover an
area of 1 050 006 km2. II
The input data is a point cloud stored in a polygon file k= , (3)
I
format. For the discrete curvature computation and analyses,
in which is based the following article, we create a 3D mesh where and are respectively the first and second
from the point cloud. fundamental forms. The differential vector is the tangent
to the surface in a direction defined by and . Mynatt et
al. [7] describe the first fundamental form (I), according to
1Nicole Christoff is with Technical University of Sofia, Faculty of Lipschultz [8], as the inner product of with itself:
Telecommunications, blvd. Kliment Ohridski 8, 1796 Sofia, ∙ ∙
Bulgaria, E-mail: nicole.christoff@tu-sofia.bg
2Agata Manolova is with Technical University of Sofia, Faculty of ∙ 2 ∙ ∙ (4)
Telecommunications, blvd. Kliment Ohridski 8, 1796 Sofia, 2
Bulgaria, E-mail: amanolova@tu-sofia.bg

978-1-7281-7142-5/20/$31.00 ©2020 European Union 181

 Fig. 2.Results on MOLA sample: b) representation of k2, the minimal curvature lying on a surface a)

For the calculation of angles, distances and surfaces on the For local characterization of the surface (i.e. elliptical,
surface, Mynatt et al. [7] uses the first fundamental parabolic or hyperbolic) Mynatt et al. [7] uses the second
coefficients ( , and ). From the partial derivatives, the fundamental coefficients ( , and ). Likewise, the matrix
matrix of the first fundamental form is obtained: is characterized by:
E F
I = . (5) . (7)
F G
The second fundamental form quantifies the spatial velocity From the first and second fundamental forms, we can
of change of in each direction, determined by at a point calculate the Weingarten endomorphism :
on the surface, taking the internal product of the two
differential vectors: , (8)
where is the inverse of the matrix I.
∙
It should be noted that the eigenvalues of L define the
∙
magnitudes of the principal curvatures at the given point and
∙ 2 ∙ ∙ (6) the eigenvectors of L characterize the orientations of the
∙ principal curvatures in the plane.
2

182
 An assessment at a point on the surface, about the shape
and the orientation, can be succeeded by the Gaussian
curvature and the mean curvature . They are affected by
the minimum curvature and the maximum curvature by
the following relations:
, (9)
. . (10)
The mesh sample used in our research is presented in Fig. 2
(theupperpicture). We computed the differential descriptor the
minimal curvature k2 on it (see Fig. 2 the lower one). In
general, curvatures attempts information on the concavity or
convexity of a surface along its principal directions [9]. This
information is characterised by the sign of H. In case of H< 0,
the convex areas are shown and presented in blue. The
concave areas are yellow (H> 0). The relatively flat areas are
green (H ≈ 0 and K ≈ 0). We proposed k2 as a good descriptor
of the Mars topography [10].
Using different distributions of k2 and their respective
covariance matrix, we describe how the assessment of
transverse and longitudinal lines changes.
We denote both types of lines, by the quantile of the
distribution of values, represented by and (the solutions
of Eq.(11)), which lies outside the interval , (Rees
[11]):
, μ ∗ . (11)
The variables
, are calculated by the percentage of
outliers % and the product of the mean value µ and the
standard deviation σ. The outlier values are considered in the
intervals ]−∞, ] and [ ,+∞[.
One sample of a histogram (see Fig. 3) gives the density of
the distribution of the k2 data. In Fig. 3 (on the left) is the 2D
representation of the histogram and (on the left), a bivariate
histogram is created, showing the distribution of the k2 values
of the sample from Fig. 2 (the lower one). It represents a right
skewed distribution, with µ = −0.0124, σ = 0.0142, the median
is −0.0043 and the mode −1.1848.
Fig. 3.A histogram of k2, the minimal curvature lying on a surface of Mars
183
IV. RESULTS AND DISCUSSION
The proposed approach of assessing was tested with Matlab
2017.
We choose a small area of our MOLA sample, which
contains one huge impact crater in the middle and several
small ones (see Fig. 2 small panel A). One of the challenges is
the detection of the small craters (with diameter ≤ 5 km). We
can observe on the computed minimal curvature , lying on
the same area (see Fig. 2 small panel B).
The results of four cases of outliers are shown in Fig. 4.
Outliers lying in the interval ]−∞, ] contain the craters. That
is the significant information about the topology that we want
to keep (Fig. 4 1a), 2a), 3a) and 4a)). The outliers in the
interval [ ,+∞[ represent actually the along-track and cross-
track information (Fig. 4 1b), 2b), 3b) and 4b)). The solutions
of Eq. 11, when p = 1% represent all values that enter into the
set of values marked in blue (Fig. 4 1a) and 1b)). The rest of
the solutions and their respective colours are as follow:
p = 5% in orange (Fig. 4 2a) and 2b)), p = 10% in purple
(Fig. 4 3a) and 3b)) and p = 20% in green (Fig. 4 4a) and 4b)).
As we can observe, increasing the percentage of values in
the positive range, the noise becomes more visible, while the
craters are clearer when the percentage is smaller. We must
keep in mind that the more we try to separate the noise, the
topology itself.
V. CONCLUSION
The present analysis aims to examine the distribution of
longitudinal and transverse lines added over different
geographical features (such as craters). That will help the
delimitation of added points in the topology for their further
removal and reconstruction of the mesh. This is crucial in the
detection of craters based on a minimal curvature approach, as
adding additional, incorrect information to the topography
would mislead the recognition system and, accordingly, the
number of false positives results, concerning small craters,
would be very high.
 Fig. 4. Visualization of the quantile of the distribution for values lying outside intervals (a) ]−∞,
ACKNOWLEDGEMENT
This work was supported by the National Program ”Young
Scientists and Postdoctoral Students”, PMC No 577,
17.08.2018, of Ministry of Education and Science of Bulgaria.
REFERENCES
[1] A. Albee, R. Arvidson, F. Palluconi, and T. Thorpe, “Overview
of the Mars Global Surveyor Mission”, Journal of Geophysical
Research: Planets, vol. 106, no. E10, pp. 23291-23316, 2001.
[2] M. C. Malin, K. S. Edgett, B. A. Cantor, M. A. Caplinger, G. E.
Danielson, E. H. Jensen, M. A. Ravine, J. L. Sandoval, and K.
D. Supulver, “An Overview of the 1985-2006 Mars Orbiter
Camera Science Investigation”, International Journal of Mars
Science and Exploration, vol. 5, pp. 1-60, 2010.
[3] J. van Zyl, “Introduction to the Physics of Remote Sensing:
Imaging geometries,” 2006. [Online].Available:
http://www.its.caltech.edu/ee157/lecture
note/Imaging%20geometries.pdf
[4] W. Jarmołowski, J. Łukasiak, “A Study on Along-Track and
Cross-Track Noise of Altimetry Data by Maximum Likelihood:
Mars Orbiter Laser Altimetry (Mola) Example”, Artificial
Satellites, vol. 50, pp. 143-155, 2015.
[5] M. T. Zuber, D. E. Smith, S. C. Solomon, D. O. Muhleman, J.
W. Head, J. B. Garvin, J. B. Abshire, and J. L. Bufton, “The
Mars Observer laser altimeter investigation”, Journal of
184
] and (b) [ ,+∞[
Geophysical Research: Planets, vol. 97, no. E5, pp. 7781-7797,
1992.
[6] A. Polette, J. Meunier, and J.-L. Mari, “Shape-Curvature-
Graph”: Towards a New Model of Representation for the
Description of 3D Meshes”, in Augmented Reality, Virtual
Reality, and Computer Graphics - 4th International Conference,
AVR 2017, Ugento, Italy, June 12-15, 2017, Proceedings, Part
II, In Lecture Notes in Computer Science, LNCS 10325,
Springer, 2017, pp. 369-384.
[7] I. Mynatt, S. Bergbauer, and D. D. Pollard, “Using Differential
Geometry to Describe 3-D Folds”, Journal of Structural
Geology, vol. 29, no. 7, pp. 1256-1266, 2007.[Online].
Available: http://www.sciencedirect.com/science/article/pii/
S0191814107000193
[8] M. M. Lipschultz, Theory and Problems of Differential
Geometry,Schaum’s Outline,New York: McGraw-Hill, 1969.
[9] N. Christoff, A. Manolova, L. Jorda, and J.-L. Mari, “Feature
Extraction and Automatic Detection of Martian Impact Craters
from 3D Meshes”, 2017 13th International Conference on
Advanced Technologies, Systems and Services in
Telecommunications (TELSIKS), 2017, pp. 211-214.
[10] N. Christoff, A. Manolova, L. Jorda, S. Viseur, S. Bouley, and
J.-L. Mari, “Level-Set Based Algorithm for Automatic Feature
Extraction on 3D Meshes: Application to Crater Detection on
Mars”, Computer Vision and Graphics, Cham: Springer
International Publishing, 2018, pp. 103-114.
[11] D. G. Rees, “Foundations of Statistics,” CRC Press, pp. 244-
249, 1987.
 Analysis for Watermark in Medical Image using
Watermarking with Wavelet Transform and DCT
Stoyan Kushlev1 and Rumen P. Mironov2
Abstract – An analysis for increasing the transparency of
digital watermarking of medical images, using discrete Wavelet
transform and DCT is presented. The analysis ensures
watermarking with high resistant of the watermark, to
commonly used types of malicious image processing. The results
for the analysis of the transparency on the watermark image is
based on the signal to noise ration and mean-square error after
watermarking the image.
Keywords – Medical Image Watermarking, Discrete Wavelet
transform, Discrete Cosine transform.
I. INTRODUCTION
Advances in telecommunications and computer science
have introduced drastic changes to the modern health care
organization, including services as: Picture Archiving and
Communications System (PACS), information management
systems in hospitals which forms the information technology
infrastructure for a hospital based on the DICOM (Digital
Imaging and Communication in Medicine) standard, medical
imaging facilities. Such services allow for the medical
personnel, as well as for the patients to have access to:
transmission and interpretation of medical images for
diagnostic purposes, remote access to personal medical
information [1-3].
Digital watermarking has showed that it have properties for
enhancing security measures for various multimedia
applications. Applications of the digital watermarking for
medical imaging are discussed in [4]. The applicability of
watermarking for security applications is studied in [2].
The new medical systems required medical images to have
unauthorized modifications protection and copyright
protection. The unauthorized modifications protection, as
discussed in [5], is of protection from unrestricted
modification of the information. Copyright protection, as
discussed in [4], is of protection from disseminated medical
information over Internet..
Watermarking systems are characterized whit the
watermark invisibility, security of the watermark, robustness
of the watermark and the ability for reversible watermarking.
The importance of each is determined on the application and
how it is used [6-7].
As discussed in [5], [8] the watermarking techniques can be
1Stoyan Kushlev is with the Faculty of Telecommunications,
Technical University of Sofia, Boul. Kl. Ohridsky, 8, Sofia 1000,
Bulgaria, E-mail: skushlev@mail.bg
2Rumen P. Mironov is with the Faculty of Telecommunications,
Technical University of Sofia, Boul. Kl. Ohridsky 8, Sofia 1000,
Bulgaria, E-mail: rmironov@tu-sofia.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union
classified based on domains, in which the watermarking is
done, the watermark preprocessing, the transparency on the
watermark. Watermark robustness can be simulated using
unauthorized attacks directed to the integrity of the
watermark. Such attacks are unauthorized removal, adding or
detection of watermark. The removal and adding of
watermarks are active attacks while the detections of
watermarks are passive attacks.
An outline of the medical image watermarking field that
uses various techniques to embed watermark data and utilize
various functions to detect tampered regions is given below in
the paper [9].
In the present work an analysis of the watermark for digital
watermarking of medical images using Wavelet transform and
DCT is described.
The presented watermarks characteristics, together with a
DWT and DCT watermarking algorithm, ensures high
resistance of the watermark to various types of malicious
attacks. The obtained experimental results, for some simulated
attacks over the test medical images are, made on the base of
mean-squared error and signal to noise ratio of the
reconstructed images. The robustness of the watermark
against some attacks are tested with the post processing of
watermarked images by adding of Salt and Pepper noise,
Gaussian noise, filtration whit median filters and average
filters.
II. MATHEMATICAL DESCRIPTION
The algorithms for embedding and extracting the
watermark using DWT and DCT, on which is based the
developed watermark preprocessing, are described in details
in article [5]. In the current work we are analyzing how the
image used for the watermark is affecting the result.
Embedding of the watermark is based on the following
formula [5]:
, 1, 2, … , , 1
where Ck are the coefficients of the transformed insertion
block, Vk are the coefficients of the transformed watermark, k
is the consecutive number of the coefficients and a is
parameter which determines the depth of the watermarking.
This shows that the transparency of the watermark after
embedding is based on two factors: the values of the
watermark itself (Vk), and the depth of the watermarking
coefficient (a).
The watermark is obtained using the following formula [5]:
⁄ , 1, 2, … , , 2
185
where Ck' are the coefficients of the transformed insertion
block of the watermarked image, Ck are the coefficients of the
transformed correspondent block of the original image, Vk' are
the coefficients of the transformed watermark. The extracted
watermark is in discrete cosine space. Over Vk' we are
applying inverse discrete cosine transform to receive the
image used for watermarking.

III. EXPERIMENTAL RESULTS

For analyzing the connection between the image used as the
watermark and the transparency on the watermarked image,
three test images are used. The images have size 512x512
pixels, 256 gray levels and are shown in Fig. 1a, b, c.
Fig. 1c. Input X-ray test image “Right rib cage”

The images used as watermarks are the letter “K” with and
without histogram correction, based on the image used for
watermarking. Those image are shown in Fig. 2a, b.

Fig. 2a. Watermark without histogram correction

Fig. 1a. Input X-ray test image “Spine”

Fig. 2b. Watermark with histogram correction

Based on the described effect of the watermark in Section

II, algorithms for embedding and extracting of watermarks
have been developed and have been simulated with MATLAB
programs.
The test images show in Fig. 1 are transformed by 3 level
Discrete Wavelet transform. The transformed block with the
maximum entropy is used for embedding the input watermark
[5].
The robustness on the watermark is tested against some
popular attacks as:
• Adding of Gaussian noise with mean 0 and variance
Fig. 1b. Input X-ray test image “Left rib cage” 0.01;
• Adding of Salt and Pepper noise;
• Median filter filtration with kernel size 3x3;
• Filtration of Gaussian noisy image with average filter;
• Filtration of Salt and Pepper noisy image with median
filter.
Efficiency of the watermark is estimated with the
following metrics [5], [10-11]:

186
 • The peek signal to noise ratio (PSNR) estimates the
transparency of the embedded watermark to the human TABLE II
eyes; Test Images “Spine” “Left rib “Right rib
• The normalize cross-correlation (NC) is used for cage” cage”
determining how close the extracted watermark is Watermarked test image
SNR, dB 34.01 14.72 35.02
compared to the original; PSNR, dB 59.95 39.27 59.80
• The mean square error (MSE) and the normalized mean MSE 6.58E-02 7.68 E+00 6.81E-02
square error (NMSE) estimate the change between the NMSE 3.97E-04 3.37E-02 3.15E-04
original and the watermark images. NMSE, % 3.97E-02 3.37E+00 3.15E-04
The obtained results for the test images when using the NC 0.98 0.97 0.99
NC, % 98.5 97.1 99.23
image from figure Fig. 2a as a watermark are summarized in Watermarked test image with added of Gaussian noise
Table I [5]. The results when using the Watermark with SNR, dB 4.67 4.66 4.68
histogram correction (Fig. 2b) are summarized in Table II. PSNR, dB 30.61 29.21 29.46
MSE 5.65E+01 7.79E+01 7.36E+01
TABLE I NMSE 3.41E-01 3.42E-01 3.40E-01
Test Images “Spine” “Left rib “Right rib NMSE, % 3.41E+01 3.42E+01 3.40E+01
cage” cage” NC 0.8451 0.8865 0.8253
Watermarked test image NC, % 84.51 88.65 82.53
SNR, dB 19.86 18.16 18.71 Watermarked test image with added of Salt and Pepper noise
PSNR, dB 45.8 42.72 43.49 SNR, dB 15.99 12.42 16.01
MSE 1.71E+00 3.48E+00 2.91E+00 PSNR, dB 41.93 36.97 40.79
NMSE 0.1E-01 0.15E-01 0.1E-01 MSE 4.17E+00 1.31E+01 5.42E+00
NMSE, % 0.1E+01 0.15E+01 0.1E+01 NMSE 2.52E-02 5.73E-02 2.51E-02
NC 0.57 0.53 0.55 NMSE, % 2.52E+00 5.73E+00 2.51E+00
NC, % 57 53 55 NC 0.85 0.88 0.83
Watermarked test image with added of Gaussian noise NC, % 85.49 87.58 82.57
SNR, dB 4.66 4.67 4.69 Watermarked test image with median filtration
PSNR, dB 30.6 29.22 29.47 SNR, dB 20.14 14.22 19.20
MSE 5.67E+01 7.78E+01 7.35E+01 PSNR, dB 46.08 38.77 43.98
NMSE 3.4E+01 3.4E-01 3.4E-01 MSE 1.61E+00 8.63E+00 2.60E+00
NMSE, % 3.4E+01 3.4E+01 3.4E+01 NMSE 9.70E-03 3.78E-02 1.20E-02
NC 0.34 0.46 0.42 NMSE, % 9.70E-01 3.78E+00 1.20E+00
NC, % 34 46 42 NC 0.93 0.92 0.92
Watermarked test image with added of Salt and Pepper noise NC, % 93.42 91.94 91.92
SNR, dB 15.54 14.02 14.27 Watermarked test image with added of Salt and Pepper noise and
PSNR, dB 40.48 38.57 39.05 filtration with median filter
MSE 0.58E+01 0.9E+01 0.8E+01 SNR, dB 10.14 8.09 8.64
NMSE 3.5E-02 4E-02 3.8E-02 PSNR, dB 36.08 32.64 33.42
NMSE, % 3.5E+00 4E+00 3.8E+00 MSE 1.60E+01 3.54E+01 2.96E+01
NC 0.36 0.41 0.39 NMSE 9.68E-02 1.55E-01 1.37E-01
NC, % 36 41 39 NMSE, % 9.68E+00 1.55E+01 1.37E+01
Watermarked test image with median filtration NC 0.85 0.86 0.83
SNR, dB 16.13 17.4 17.99 NC, % 85.05 86.48 83.12
PSNR, dB 45.07 41.95 42.77 Watermarked test image with added of Gaussian noise and filtering with
MSE 0.2E+01 0.41E+01 0.34E+01 average filter
NMSE 0.12E-01 0.18E-01 0.16E-01 SNR, dB 19.38 14.00 18.60
NMSE, % 0.12E+01 0.18E+01 0.16E+01 PSNR, dB 45.32 38.55 43.38
NC 0.57 0.52 0.55 MSE 1.91E+00 9.08E+00 2.98E+00
NC, % 57.0 52.0 55.0 NMSE 1.15E-02 3.98E-02 1.38E-02
Watermarked test image with added of Salt and Pepper noise and NMSE, % 1.15E+00 3.98E+00 1.38E+00
filtration with median filter NC 0.93 0.93 0.84
SNR, dB 18.55 16.99 17.6 NC, % 92.85 92.50 84.19
PSNR, dB 44.49 41.54 42.38
MSE 0.23E+01 0.46E+01 0.38E+01
NMSE 0.1E-01 0.2E-01 0.17E-01
IV. CONCLUSION
NMSE, % 0.1E+01 0.2E+01 0.17E+01
NC 0.56 0.52 0.55 An analysis for increasing the transparency of digital
NC, % 56 52 55 watermarked medical images, using Discrete Wavelet
Watermarked test image with added of Gaussian noise and filtering with
transform and Discrete Cosine transform, is presented. The
average filter
SNR, dB 8.86 8.6 8.6 analysis result are based on the signal to noise ratio, mean-
PSNR, dB 34.8 33.15 33.38 squared error and normalized cross-correlation, of the
MSE 2.15E+01 3.15E+01 2.99E+01 reconstructed images between the original and the
NMSE 1.3E-01 1.4E-01 1.4E-01 reconstructed images, when adding malicious attacks.
NMSE, % 1.3E+01 1.4E+01 1.4E+01 Compared to the result in [5], when is used watermark without
NC 0.34 0.43 0.41
NC, % 34 43 41 histogram correction, we see that the signal to nose ratio is
higher when using the watermark with histogram correction,

187
while achieving high robustness for the watermark. This will
allow to a great extent to verify the reliability of the medical
data, transmitted and recorded as images, while reducing
artifacts from the watermarking. All this leads to the
conclusion that, the presented analysis can be used for
developing algorithm for histogram correction, based on the
input image.
V. REFERENCES
[1] P. Koushik, G. Ghosh, and M. Bhattacharya, “Biomedical
Image Watermarking in Wavelet Domain for Data Integrity
Using Bit Majority Algorithm and Multiple Copies of Hidden
Information”, American Journal of Biomedical Engineering,
2012, vol. 2, no. 2, pp. 29-37.
[2] H. Nyeem, W. Boles, and C. Boyd, “A Review of Medical
Image Watermarking Requirements for Teleradiology”, Journal
Digital Imaging, 2013, vol. 26, pp. 326-343.
[3] L. Siau-Chuin, J. M. Zain, “Reversible Medical Image
Watermarking for Tamper Detection and Recovery”, 3rd IEEE
International Conference on Computer Science and Information
Technology, 2010, vol. 5, pp. 417-420.
[4] G. Coatrieux, H. Maitre, B. Sankur, Y. Rolland, and R.
Collorec, “Relevance of Watermarking in Medical Imaging”,
Proceeding of IEEE EMBS International Conference on
188
Information Technology Applications in Biomedicine, 2000,
pp. 250-255.
[5] R. Mironov, St. Kushlev, “Medical Images Watermarking using
Wavelet Transform and DCT”, LIII International Scientific
Conference on Information, Communication and Energy
Systems and Technologies (ICEST 2018), June 28-30 2018,
Sozopol, Bulgaria, pp. 119-122.
[6] W. K. Pratt, Digital Image Processing, 4th Ed., John Wiley &
Sons. Inc., Hoboken, New Jersey, 2007.
[7] R. C. Gonzalez, R. E. Woods, Digital Image Processing, Third
Ed., Pearson Education Inc., 2008.
[8] N. Singh, M. Jain, and S. Sharma, “A Survey of Digital
Watermarking Techniques”, International Journal of Modern
Communication Technologies & Research, August 2013.
[9] Mahasweta, J. Joshi et al. “Watermarking in DCT-DWT
Domain”, International Journal of Computer Science and
Information Technologies (IJCSIT), vol. 2, no. 2, 2011, pp.
717-720.
[10] I. Cox, M. Miller, J. Bloom, J. Fridrich, and T. Kalker, Digital
Watermarking and Steganography, 2nd Edition, Elsevier,
Burlington, 2007.
[11] A. Ustubioglu, G. Ulutas. “A New Medical Image
Watermarking Technique with Finer Tamper Localization”.
Journal of Digital Imaging. Springer International Publishing,
2017, pp. 1-17.
MICROWAVES, ANTENNAS AND
PROPAGATION
 On the Gas Sensing Properties of Microwave Transducers
Giovanni Gugliandolo1, Mariangela Latino2, Giuseppe Campobello3, Zlatica Marinkovic4,
Giovanni Crupi5and Nicola Donato3
Abstract – In this work it is reported about the development
and characterization of a microwave interdigital capacitor (IDC).
The IDC structure was realized in microstrip line technology and
its surface was functionalized by screen printing a titanium-
dioxide (TiO2) layer on the electrodes and the sensing properties
towards O2 have been evaluated. Electrical characterization was
carried out in a frequency range spanning from 1.5 GHz to 3.5
GHz. The sensor response was evaluated at different binary gas
mixtures O2/N2by means of an Agilent 8753ES Vector Network
Analyzer (VNA).
Keywords – Microwave characterization, Microstrip resonator,
S-parameter measurements, Oxygen sensor.
I. INTRODUCTION
In recent years, microwave technique provided a new route
for the development of gas sensors. Traditional gas sensors
are based on resistive, capacitive and amperometric effects
[1], [2] and, most of the time, they operate at high temperature
(400°C). Despite conventional gas sensors, device based on
microwave transducers are attracting considerable attention
because of their interesting properties in terms of power
consumption, response time and operating temperature [1],
[3]–[7]. A low-cost electronic interface can be developed for
such sensors with an accurate resonance frequency estimation
[8]–[10]. Moreover, they are fully compatible with the
wireless technology so that they can be easily integrated into
antennas [11]–[14].
The working principle of a microwave transducer is based
on the propagation and reflection of an electromagnetic wave
into a sensing material. When exposed to the gas target, the
sensing layer interacts with the gas molecules modifying its
dielectric properties and its characteristic electrical response.
As a consequence, the velocity of the electromagnetic wave is
affected. Beside the sensing material, that acts as an active
1Giovanni Gugliandolo is with Istituto Nazionale di Ricerca
Metrologica, Strada delle Cacce, Torino, Italy, and with Politecnico
di Torino, Corso Duca degli Abruzzi, 24, Torino, Italy, E-mail:
g.gugliandolo@inrim.it
2Mariangela Latino is with the MIFT Department, University of
Messina, Italy, E-mail: mlatino@unime.it
3Giuseppe Campobello and Nicola Donato are with the
Department of Engineering, University of Messina, Contrada di Dio,
S. Agata, Messina, Italy, E-mails: gcampobello@unime.it,
ndonato@unime.it
4Zlatica Marinkovic is with University of Niš, Faculty of
Electronic Engineering, Aleksandra medvedeva 14, 18000 Niš,
Serbia, E-mail: zlatica.marinkovic@elfak.ni.ac.rs
5Giovanni Crupi is with the BIOMORF Department, University of
Messina, Via Consolare Valeria, 98100 Messina, Italy, E-mail:
crupig@unime.it
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
layer, a microwave gas sensor needs of a propagative structure
for the electromagnetic wave. Several geometries have been
explored as propagative structures in gas sensing, in most
cases, the microstrip technology is employed. For instance, in
[11]–[13] a sensitive layer was combined with a rectangular
patch antenna and the sensing performance towards a gas
target has been evaluated. The combination of a sensing layer
with an antenna allows development of wireless sensor nodes
[15]. Moreover, in [1], [4], [16]–[21] different geometries of
microwave resonators have been explored for gas sensing
with interesting results in terms of response time and
measurement accuracy. Among the various structures
proposed over the years, in [5], [22]interdigital capacitors
(IDCs)covered by a sensing layer have been employed as gas
sensors.
In this paper, an IDC, used as propagative structure, is
covered by a thick layer of titania (TiO2), that acts as a
sensitive material. The developed prototype is used as probe
for oxygen (O2) sensing. It is fabricated using the microstrip
technology and its electrical response towards different
concentrations of O2 is measured in the frequency range
spanning from 1.5 GHz up to 3.5 GHz. According to the
literature, the TiO2 metal oxide is suitable for this purpose
since the O2 molecules absorbed on its surface can be easily
desorbed even at room temperature [23]. This reduces the
sensor power consumption and makes the measurement
process more practical. Furthermore, the nanostructured TiO2
employed allows increasing the number of O2 molecules
adsorbed on the surface, thereby improving the device
sensitivity.
The main goal of this work is to employ an IDC as
microwave transducer for O2 monitoring at room temperature
and at concentrations between 0% and 100% v/v. A possible
application for this device is in oxygen therapy, where the O2
concentration is usually greater than 85%. The device could
be also useful in spacecraft oxygen level detection, in these
cases the O2 concentration is above 60% [24].
This paper is organized as follows. In Section II, a
description of the fabricated prototype is reported and details
about the experimental setup are given. In Section III the
sensor response at different O2 concentrations is evaluated in
terms of the input admittance parameter (Y11), focusing on the
frequency range from 2 GHz to 2.5 GHz. Finally, conclusions
are givenin Section IV.
II. EXPERIMENTS
The device presented in this work is a two-port IDC
employed as a microwave transducer for oxygen sensing. It is
fabricated on a 1.6-mm FR4 substrate by using the LPKF
Protomat S103 PCB milling machine. The substrate is
191
 200
characterized by a dielectric constant εr = 4.2 and a loss
tangent tanδ = 0.015. Two microstrip lines with an input
100
impedance of 50 Ωare connected to two SMA connectors that
act as input and output ports for the microwave device. A

Real part (Ω)

0
picture of the device is reported in Fig. 1. The IDC is made by
six fingers with a length of 15 mm, width of 0.36 mm and a
-100
spacing of 0.36 mm. The board dimensions are relatively
small (20 mm x 30 mm x 1.6mm).A thick layer of TiO2 is a) Re(Z11) @ 0% O2
-200 Re(Z21) @ 0% O2
deposited by screen printing at the top of the IDC structure as Re(Z22) @ 0% O2
sensitive layer. After deposition, the sample is placed inside -300
Re(Z12) @ 0% O2

oven at 60°C for two hours in order to reduce the humidity 1500 2000 2500 3000 3500
content. The employed TiO2 is composed of 56.5% anatase, Frequency (MHz)
19.8% rutile and 23.6% brookite. Further information about
100
synthesis and characterization of this sensing material is
reported in [25]. 50

Imaginary part (Ω)

0

-50

-100 Im(Z11) @ 0% O2
b) Im(Z21) @ 0% O2
-150 Im(Z22) @ 0% O2
a) b) Im(Z12) @ 0% O2
Fig. 1. (a) Geometry of the developed prototype (W = 20 mm, -200
1500 2000 2500 3000 3500
L = 30 mm, Ws = 4 mm, Lf = 15 mm, Wf = S = 0.36 mm), Frequency (MHz)
(b) Fabricated IDC structure after the sensing material deposition
Fig. 2. Comparison between the real (a) and imaginary part (b)
The device is tested upon a controlled atmosphere by means of the Z-parameters
1.0
of a fully automated gas control system. This apparatus is
equipped with certified gas bottles and mass flow controllers 0.5
in order to set the preferred gas mixture inside the testing
chamber. The control system allows the operator to select the 0.0
Real part (S)

gas mixture, set the concentration value for each gas and
control/monitor the gas flow. The device under test (DUT) is -0.5

placed inside a Polytetrafluoroethylene (PTFE) test chamber Re(Y11) @ 0% O2

-1.0
for the electrical characterization. Because of the relatively Re(Y22) @ 0% O2

small chamber volume (5 cm3), it is possible to set or purge a) Re(Y11) @ 100% O2

-1.5 Re(Y22) @ 100% O2
the gas content in short time. The fabricated prototype, placed
inside the testing chamber, is connected to the Agilent 2000 2100 2200 2300 2400 2500
Frequency (MHz)
8753ES VNA for test. Before acquisition, the VNA is
1.0
calibrated using the Short-Open-Load-Through (SOLT)
standards. The microwave response, in terms of the scattering 0.5
(S-) parameters, is recorded for each O2 concentration set-
Imaginary part (S)

point. Measurements performed on the DUT are reported in 0.0

the next section.
-0.5

III. RESULTS -1.0

Im(Y11) @ 0% O2
Im(Y22) @ 0% O2
b) Im(Y11) @ 100% O2
The TiO2-IDC device is tested inside the test chamber with -1.5 Im(Y22) @ 100% O2
a binary gas mixture composed by oxygen and nitrogen (N2).
2000 2100 2200 2300 2400 2500
The O2 concentration is varied from 0% (pure N2 inside
Frequency (MHz)
chamber) to 100% (no N2 inside chamber). The gas flow is set
to 100 cm3/min and this value is maintained constant for the Fig. 3. Comparison between the real (a) and imaginary part (b)
whole measurement session. The DUT behavior at O2 of Y11 and Y22 parameters
concentration variation is studied in terms of admittance (Y-) As illustrated in Fig. 2, the measured data show that the
parameters, which are achieved from the measured S- impedance (Z-) parameters of the DUT are quite close to the
parameters. Measurements are carried out considering the condition Z11 = Z22 = − Z12 = − Z21 and, analogously, the Y-
frequency range from 1.5 GHz to 3.5 GHz. However, since a
parameters are quite close to the condition
substantial variation on the admittance parameters is recorded
between 2.0 GHz and 2.5 GHz, this band is considered in the Y11 = Y22 = Y12 = Y21 . This is valid for any O2 concentration
study. value, however, for brevity, the Y-parameters are plotted only

192
for the lowest and highest concentration of O2: 0% and 100% In Fig. 5, the real and imaginary parts of Y11 are plotted at
of O2 (see Figs. 3 and 4). Therefore, the attention has been different O2 concentrations. As shown, the O2 concentration
restricted to the analysis of the frequency-dependent behavior changes are transduced by the microwave device into
of Y11. variations of Y11. Peaks and dips in the spectrum are shifted in
1.0 frequency and their amplitude is affected as well. In this work,
the frequency shifts in Y11 are used as markers for O2
0.5 detection.
At certain frequencies, the imaginary part of Y11 becomes
0.0
null and the parameter is purely real. At these frequencies the
Real part (S)

device resonates and it is possible to relate the change in the

-0.5
resonant frequency to a change in the O2 concentration. In the
-1.0
Re(Y21) @ 0% O2
investigated bandwidth, three resonant frequencies can be
Re(Y12) @ 0% O2
a) Re(Y21) @ 100% O2
found: f1 = 2.21 GHz, f2 = 2.33 GHz, f3 = 2.38 GHz (at 0%
-1.5 Re(Y12) @ 100% O2 O2).
In Fig. 6the three resonant frequencies are reported as a
2000 2100 2200 2300 2400 2500
function of O2. The frequency dependence on O2
Frequency (MHz)
1.0
concentration seems to be almost linear for all the resonances.
The sensitivity (S) of each resonant frequency to changes in
0.5 the concentration of O2is evaluated as:

f rx − f r 0
Imaginary part (S)

0.0
S= , (1)
-0.5
%O2

-1.0
Im(Y21) @ 0% O2 Where frx is the resonance frequency at the generic %O2
Im(Y12) @ 0% O2
oxygen concentration and fr0 is the resonant frequency when
b) Im(Y21) @ 100% O2
-1.5 Im(Y12) @ 100% O2
there is no oxygen inside the test chamber.
The sensitivity is plotted in Fig. 7. As regards f1 and f3, it
2000 2100 2200 2300 2400 2500
varies from -2 MHz/% to 56 kHz/%. On the other hand, the
Frequency (MHz)
sensitivity for f2 is almost constant and equal to -330 kHz/%.
Fig. 4. Comparison between the real (a) and imaginary part (b)
2400
of Y21 and Y12 parameters
0.5
2350
Frequency (MHz)

0.0
2300
0% O2
f1
Re[Y11] (S)

-0.5 5% O2
f2
10% O2 2250 f3
15% O2
-1.0 20% O2
40% O2 a) 2200
70% O2 0 20 40 60 80 100
-1.5
100% O2
O2 concentration (%)
2000 2100 2200 2300 2400 2500
Fig. 6. Resonant frequencies vs. oxygen concentration: frequencies at
Frequency (MHz) which the Y11 parameter becomes purely real
1.0
0.5
0.5
0.0
0.0
0% O2
Sensitivity (MHz / %)
Im[Y11] (S)

5% O2 -0.5
-0.5
10% O2
15% O2 -1.0
-1.0
20% O2
40% O2 -1.5
-1.5 70% O2 b) f1
f2
100% O2 -2.0
-2.0 f3
2000 2100 2200 2300 2400 2500
-2.5
Frequency (MHz) 0 20 40 60 80 100
Fig. 5. Frequency dependence of the real (a) and imaginary part (b) O2 concentration (%)
of Y11 at different oxygen concentrations. At three resonant Fig. 7. Sensitivity of each resonant frequency. For f1 and f3 the
frequencies, the imaginary part is null and the Y11 parameter sensitivity is high for O2 concentrations less than 20%. The
becomes purely real sensitivity of f2 is almost constant in the entire range explored

193
 IV. CONCLUSION
An interdigitated capacitor was designed and realized in
FR4 substrate, the interdigitated fingers were covered by a
titania layer in order to test the sensing properties towards
oxygen. The device was exposed in a controlled atmosphere
with an oxygen concentration range spanning from 0% to
100%. The sensing performance was evaluated by measuring
S-parameters and converting as Y-ones. The attention has
been focused on the shift of the resonant frequencies in the
input admittance parameter. The reported data can be
considering a promising solution in order to develop oxygen
low power sensing devices.
REFERENCES
[1] A. Bogner, C. Steiner, S. Walter, J. Kita, G. Hagen, and R.
Moos, “Planar Microstrip Ring Resonators for Microwave-
Based Gas Sensing: Design Aspects and Initial Transducers for
Humidity and Ammonia Sensing”, Sens. Switz., 2017, doi:
10.3390/s17102422
[2] E. Comini, G. Faglia, and G. Sberveglieri, Eds., Solid State Gas
Sensing. Boston, MA, Springer US, 2009.
[3] G. Barochi, J. Rossignol, and M. Bouvet, “Development of
Microwave Gas Sensors”, Sens. Actuators B Chem., vol. 157,
no. 2, pp. 374-379, Oct. 2011, doi: 10.1016/j.snb.2011.04.059
[4] J. Rossignol et al., “Microwave-Based Gas Sensor with
Phthalocyanine Film at Room Temperature”, Sens. Actuators B
Chem., vol. 189, pp. 213-216, Dec. 2013, doi:
10.1016/j.snb.2013.03.092
[5] A. Rydosz, E. Maciak, K. Wincza, and S. Gruszczynski,
“Microwave-Based Sensors with Phthalocyanine Films for
Acetone, Ethanol and Methanol Detection”, Sens. Actuators B
Chem., vol. 237, pp. 876-886, Dec. 2016, doi:
10.1016/j.snb.2016.06.168
[6] B.-H. Kim et al., “A Gas Sensor Using Double Split-Ring
Resonator Coated with Conducting Polymer at Microwave
Frequencies”, IEEE SENSORS 2014, Nov. 2014, pp. 1815-
1818, doi: 10.1109/ICSENS.2014.6985379
[7] G. Gugliandolo, D. Aloisio, S. G. Leonardi, G. Campobello, and
N. Donato, “Resonant Devices and Gas Sensing: from Low
Frequencies to Microwave Range”, 2019 14th International
Conference on Advanced Technologies, Systems and Services
in Telecommunications (TELSIKS), Oct. 2019, pp. 21-28, doi:
10.1109/TELSIKS46999.2019.9002368
[8] G. Campobello, G. Cannatà, N. Donato, A. Famulari, and S.
Serrano, “A Novel Low-Complex and Low-Memory Method for
Accurate Single-Tone Frequency Estimation”, 4th International
Symposium on Communications, Control and Signal
Processing, ISCCSP 2010, Limassol, Cyprus, 3-5 March 2010,
Article number 5463459.
[9] G. Campobello, G. Cannatà, N. Donato, M. Galeano, and S.
Serrano, “An Accurate and Simple Frequency Estimation
Method for SensorApplications”, D’Amico A., Di Natale C.,
Mosiello L., Zappa G. (eds) Sensors and Microsystems. Lecture
Notes in Electrical Engineering, vol. 109. Springer, Boston,
MA, 2012.
[10] G. Campobello, A. Segreto, and N. Donato, “A New Frequency
Estimation Algorithm for IoT Applications and Low-Cost
Instrumentation”, 2020 IEEE International Instrumentation and
Measurement Technology Conference (I2MTC), Dubrovnik,
194
Croatia, 2020, pp. 1-5, doi:
10.1109/I2MTC43012.2020.9128364
[11] B. Wu, X. Zhang, B. Huang, Y. Zhao, C. Cheng, and H. Chen,
“High-Performance Wireless Ammonia Gas Sensors Based on
Reduced Graphene Oxide and Nano-Silver Ink Hybrid Material
Loaded on a Patch Antenna”, Sensors, vol. 17, no. 9, p. 2070,
Sep. 2017, doi: 10.3390/s17092070
[12] H. Lee et al., “Carbon-Nanotube Loaded Antenna-Based
Ammonia Gas Sensor”, IEEE Trans. Microw. Theory Techn.,
vol. 59, no. 10, pp. 2665-2673, Oct. 2011, doi:
10.1109/TMTT.2011.2164093
[13] G. Gugliandolo, K. Naishadham, N. Donato, G. Neri, and V.
Fernicola, “Sensor-Integrated Aperture Coupled Patch
Antenna”, 2019 IEEE International Symposium on
Measurements & Networking (M&N), Jul 2019, pp. 1-5, doi:
10.1109/IWMN.2019.8805023
[14] H. Torun, F. Cagri Top, G. Dundar, and A. D. Yalcinkaya, “An
antenna-coupled split-ring resonator for biosensing,” J. Appl.
Phys., vol. 116, no. 12, p. 124701, Sep. 2014, doi:
10.1063/1.4896261
[15] G. Campobello, M. Castano, A. Fucile, and A. Segreto, WEVA:
A Complete Solution for Industrial Internet of Things. Ad-hoc,
Mobile, and Wireless Networks, ADHOC-NOW 2017, Lecture
Notes in Computer Science, vol. 10517, Springer.
[16] S. Chopra, S. Natarajan, and A. M. Rao, “Gas Sensing Using
Carbon Nanotube-Based Resonator”, Proceedings of IEEE
Sensors, pp. 399-402, doi: 10.1109/ICSENS.2004.1426183,
2004.
[17] D. Aloisio, N. Donato, “Development of Gas Sensors on
Microstrip Disk Resonators”, Procedia Eng., vol. 87, pp. 1083-
1086, 2014, doi: 10.1016/j.proeng.2014.11.351
[18] G. Bailly, A. Harrabi, J. Rossignol, M. Michel, D. Stuerga, and
P. Pribetich, “Microstrip Spiral Resonator for Microwave-Based
Gas Sensing”, IEEE Sens. Lett., vol. 1, no. 4, pp. 1-4, Aug.
2017, doi: 10.1109/LSENS.2017.2716413
[19] G. Bailly et al., “Influence of the Design in Microwave-based
Gas Sensors: Ammonia Detection with Titania Nanoparticles”,
Procedia Eng., vol. 168, pp. 264-267, 2016, doi:
10.1016/j.proeng.2016.11.185
[20] G. Gugliandolo, A. Mirzaei, M. Bonyani, G. Neri, M. Latino,
and N. Donato, “Electrical Characterization of Microstrip
Resonators Based on Nanostructured Sensing Materials”,
Sensors and Microsystems, LNEE, vol. 457, 2018, pp. 29-34.
[21] S. Walter, A. Bogner, G. Hagen, and R. Moos, “Novel Radio-
Frequency-Based Gas Sensor with Integrated Heater”, J. Sens.
Sens. Syst., vol. 8, no. 1, pp. 49-56, Jan. 2019, doi:
10.5194/jsss-8-49-2019
[22] G. Bailly, A. Harrabi, J. Rossignol, D. Stuerga, and P. Pribetich,
“Microwave Gas Sensing with a Microstrip Interdigital
Capacitor: Detection of NH3 with TiO2 Nanoparticles”, Sens.
Actuators B Chem., vol. 236, pp. 554-564, Nov. 2016, doi:
10.1016/j.snb.2016.06.048
[23] H. Wang et al., “A Room Temperature Oxygen Gas Sensor
Based on Hierarchical TiO2”, 2016 IEEE International
Conference on Manipulation, Manufacturing and Measurement
on the Nanoscale (3M-NANO), Jul. 2016, pp. 199-202, doi:
10.1109/3M-NANO.2016.7824936
[24] E. M. Cortright, Apollo Expeditions to the Moon, Washington,
DC: U.S. Government Printing Office, 1975.
[25] D. M. Tobaldi, S. G. Leonardi, R. C. Pullar, M. P. Seabra, G.
Neri, and J. A. Labrincha, “Sensing Properties and
Photochromismof Ag–TiO 2 Nano-Heterostructures”, J. Mater.
Chem. A, vol. 4, no. 24, pp. 9600-9613, 2016, doi:
10.1039/C6TA03760G
 Artificial Neural Network Modeling of Interdigital
Capacitor Sensor for Oxygen Detection
Zlatica Marinković, Giovanni Gugliandolo2, Mariangela Latino3, Giuseppe Campobello4,
Giovanni Crupi5 and Nicola Donato4
Abstract – Microwave interdigital capacitors (IDCs) realized in
standard microstrip technology whose electrodes are covered by
a titanium-dioxide (TiO2) layer can be used for sensing oxygen,
as their RF behaviour change when exposed to different oxygen
concentrations. In order to model the variation of the frequency
dependent behaviour of an IDC with oxygen concentration, in
this paper, artificial neural networks (ANNs) are exploited. The
modelling results match very well the corresponding measured
values over the whole range of oxygen concentration from 0 to
100%.
Keywords – Admittance (Y-) parameters, Artificial neural
networks, Interdigital capacitors, Oxygen sensing.
I. INTRODUCTION
Artificial neural networks (ANNs) have been efficiently
applied to build behavioral models of the RF characteristics of
different microwave devices, such as microwave transistors,
RF MEMS switches, microstrip antennas, etc. [1]-[6]. The
models include the dependence on frequency and on different
parameters such as device dimensions or operating conditions,
e.g. temperature or bias conditions. The models can be
developed from the measured data or from the data simulated
in accurate electromagnetic (EM) simulators. In any case,
once developed an ANN based model provide fast and
accurate response for any input value from the input
parameter space, without need for additional measurements or
simulations. Therefore, ANN based models are suitable for
efficient analysis of the device behavior with the changes of
input parameters or for optimization of the device behavior.
In this paper the neural based admittance (Y-) parameter
modeling is applied to an interdigital capacitor (IDC) aimed
for gas sensing. Namely, if covered by appropriate gas sensing
1Zlatica Marinković is with the University of Niš, Faculty of
Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš,
Serbia, E-mail: zlatica.marinkovic@elfak.ni.ac.rs
2Giovanni Gugliandolo is with Istituto Nazionale di Ricerca
Metrologica, Strada delle Cacce, Torino, Italy, and with Politecnico
di Torino, Corso Duca degli Abruzzi, 24, Torino, Italy, E-mail:
g.gugliandolo@inrim.it
3Mariangela Latino is with the MIFT Department, University of
Messina, Italy, E-mail: mlatino@unime.it
4Giuseppe Campobello and Nicola Donato are with the
Department of Engineering, University of Messina, Contrada di Dio,
S. Agata, Messina, Italy, E-mails: gcampobello@unime.it,
ndonato@unime.it
5Giovanni Crupi is with the BIOMORF Department, University of
Messina, Via Consolare Valeria, 98100 Messina, Italy, E-mail:
crupig@unime.it
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
layers, the IDC characteristics change when exposed to
different concentration of the gas to be detected [6]-[10]. This
is one of the approaches followed to develop microwave gas
sensors, a second one consists in the design of resonators,
equipped with sensing layers able to change the resonance
frequency value when exposed to gas targets [9]-[10].
Therefore, it is useful to have a model of the IDC which
includes the dependence of its frequency-dependent
characteristics on the gas concentration. In the particular case,
the studied IDC, realized in microstrip technology, is covered
with a tick layer of titanium-dioxide (TiO2), which is able to
absorb the oxygen (O2) molecules, thereby making the IDC
suitable for O2 sensing applications [5]. Therefore, the
dependence of the device characteristics in terms of
admittance (Y-) parameters versus O2 concentration and
frequency are modeled by means of ANNs.
The paper is structured as follows. In Section II, a
description of the proposed neural model is given. Section III
contains details of the modeled device and the developed
model, as well as the comparison of the obtained results with
measured reference values. The concluding remarks are given
in Section IV.
II. PROPOSED ANN MODEL
As mentioned above in Introduction, the IDC characteristics
modeled for sensing purposes are the admittance (Y-)
parameters. The measured Y-parameters of the interdigital
capacitor considered in this work roughly satisfy the condition
Y11 = Y22 = Y21 = Y12, as discussed in [11]. Therefore, it is
enough to model only one of the four measured Y-parameters,
let it be Y11. Having in mind that the characteristics vary when
the oxygen concentration varies from the reference
concentration and the frequency behavior under any
concentration value follows a pretty much similar trend, to
ensure the modeling accuracy, the proposed model accounts
for the values of Y11 at a chosen reference oxygen
concentration, i.e., the model is realized by applying the so-
called prior knowledge input (PKI) neural modeling approach
[1].
Namely, PKI ANN models have, besides original inputs,
additional inputs correlated with the output parameters. The
proposed model is shown in Fig. 1.
195

Fig. 1. Proposed ANN model for the studied sensor
The model is composed of two ANNs. The first one,
ANN1, is trained to model the frequency dependence of the
real and imaginary parts of Y11 at the reference oxygen
concentration. The main ANN, ANN2, relates the real and
imaginary parts of Y11 with the oxygen concentration and
frequency, having the information about Y11 at the considered
frequency at the reference concentration.
Therefore, both ANNs have two outputs, corresponding to
the real and imaginary parts of Y11. ANN1 have one input
corresponding to the frequency, whereas ANN2 have four
inputs corresponding to the oxygen concentration, frequency
and Y11 parameter at the reference oxygen concentration.
The ANNs used in this work are multilayered perceptron
networks (MLP), having neurons grouped into layers: input
layer, output layer and one or more hidden layers [1]. The
number of neurons in input layer is equal to the number of
input parameters and the number of output layer is equal to
the number of output parameters. The number of hidden
neurons is determined in the ANN training phase. The
neurons are characterized by a transfer function, which is the
unity function for the input layers, sigmoid function for
hidden layers and linear for the output layer. The neurons
from one layer are connected to all neurons from the
following layer and each connection is weighted. During the
training the ANN parameters, i.e., connection weights and
thresholds of the neuron transfer functions, are adjusted to
obtain the ANN response matching the target values with the
defined accuracy. In this work, The ANNs were trained by
applying the Levenberg-Marquardt training algorithm [1],
which belongs to the backpropagation type of training
algorithms. For both ANNs, during the training the ANNs
with different number of hidden neurons in two hidden layers
were trained with the aim to find an ANN structure providing
the best accuracy. The ANNs with two hidden layers were
chosen as two-hidden layer ANNs have better generalization
ability than one-hidden layer ANNs. Generalization ability is
a measure of the capability of an ANN to give accurate output
also for the input values not used for the ANN training. The
accuracy of the trained ANNs was estimated through average
test error, worst case error (WTE) and correlation coefficient
r, defined as in [1]. After the ANNs are trained, the model can
be used for fast estimation of the IDC characteristics for
different values of oxygen concentration without need for any
additional measurement.
III. RESULTS AND DISCUSSION
A picture of the modeled IDC after the sensing material
deposition is shown in Fig. 2 [6].
196
Fig. 2. Photo of the investigated IDC structure after the sensing
material deposition
The data used for the ANN model development were
obtained by measuring the device scattering (S-) parameters
for different oxygen concentrations from 0% (no presence of
oxygen) and 100% (only oxygen present) in the frequency
range from 1.5 GHz to 2.925 GHz (step of 1.25 MHz) using
the measurement setup described in [11]. The S-parameters
were converted into the Y-parameters, which are further used
for the ANN model training and verification. The
measurements were performed for the following values of
oxygen concentrations: 0%, 5%, 10%, 15%, 20%, 40%, 70%,
and 100%.
Having in mind a non-uniform sampling of the oxygen
concentration levels, the concentration of 70% was chosen as
the reference one, as it is in the part of the input space where
the values are sampled with a bigger step. All available data
referring to the concentration of 70% (1141 samples) were
used to train the ANN1, giving the frequency dependence of
Y11 for the reference concentration. The best ANN has six
neurons in the first hidden layer and four neurons in the
second hidden layer. The real and imaginary part simulated by
the model (lines) compared with the reference values are
shown in Fig. 3.
0.8
Re(Y 11 ) - reference values
0.6 Re(Y 11 ) - ANN model
Im(Y 11 ) - reference values
0.4 Im(Y 11 ) - ANN model
0.2
0
-0.2
-0.4
-0.6
-0.8
1.5 2 2.5 3
f [GHz]
Fig. 3. Y11 parameter at 70% of the O2 concentration: (symbols)
measurements and (lines) ANN simulations
ANN2 was trained using the data referring to the following
values of concentrations: 0%, 10%, 20%, 40%, 70%, and
100%, whereas the data referring to 5% and 15% were used
for the ANN test purpose. To build the training set, for each
frequency appearing in the training set, the Y11 parameter real
and imaginary parts referring to the concentration of 70%
were calculated by using ANN1. As there were 1141 samples
available per a concentration, the training set consisted of
6846 samples, and the test set consisted of 2282 samples. The
best ANN has 15 neurons in the first hidden layer and 12
neurons in the second hidden layer. This ANN predicts well
the Y11 parameter with good accuracy not only for the oxygen
concentrations used for the training purpose, as well as for the
test values of the concentration. For illustration, in Fig. 4, a
comparison of the simulated values with the target reference
values for two values of the oxygen concentration is shown.
The values referring to these concentrations (40% and 100%)
belong to the training set. A good agreement of the simulated
and reference values indicate that the ANN learnt the training
data very well. The ANN achieved also satisfactory
generalization accuracy, i.e., it gives accurate responses also
for the values not used in the training set. This is illustrated in
Fig. 5 where the simulated and reference values for the
0.6
Re(Y 11 ) - reference values
Re(Y 11 ) - ANN model
0.4 Im(Y 11 ) - reference values
Im(Y 11 ) - ANN model
0.2
concentrations of 5% and 15% not used for the ANN training
are reported.
The developed model can be used further to analyze the
frequency behaviour of Y11 at an arbitrary value of oxygen
concentration. To illustrate this, in Fig. 6, the real and
imaginary parts of Y11 generated by the developed model for
the oxygen concentrations for 0% to 100% with the step of
10% were shown in the frequency range from 2 GHz to
2.5 GHz.
As it can be noticed from Figs. 4 and 5, in some cases there
are certain deviations from the reference values for the
frequencies where the peaks and deeps appear. However, as in
this type of sensors the resonat frequency shifts are more of
interest for sensing purpose, these deviations in the areas of
peaks and deeps can be considered as acceptable.
0.5
Re(Y 11 ) - reference values
0.4 Re(Y 11 ) - ANN model
Im(Y 11 ) - reference values
0.3 Im(Y 11 ) - ANN model
0.2

0.1
Y11 [S]

0
0

-0.2 -0.1

-0.2
-0.4
-0.3

-0.6 -0.4
1.5 2 2.5 3 1.5 2 2.5 3
f [GHz] f [GHz]

(a) (a)

0.8 0.5
Re(Y 11 ) - reference values
0.6 0.4 Re(Y 11 ) - ANN model
Im(Y 11 ) - reference values
0.4 0.3 Im(Y 11 ) - ANN model
0.2 0.2
0
0.1
Y11 [S]

-0.2
0
-0.4
-0.1
-0.6
Re(Y 11 ) - reference values
-0.2
-0.8 Re(Y 11 ) - ANN model
Im(Y 11 ) - reference values -0.3
-1
Im(Y 11 ) - ANN model
-1.2 -0.4
1.5 2 2.5 3 1.5 2 2.5 3
f [GHz] f [GHz]

(b) (b)

Fig. 4. Y11 at two training values of the O2 concentration (symbols – Fig. 5. Y11 at two test values of the O2 concentration (symbols –
measurements, lines - ANN simulations): (a) 40%, (b) 100% measurements, lines - ANN simulations): (a) 5%, (b) 15%

197
 0.4 in the present case) as the additional ANN inputs. The
developed model exhibs good learning and genralization
0.2 accuracy. Such a model can be further used for the analysis of
the shift in the resonant frequencies i, which can be used as
0 markers for the oxygen detection.
0% O2
-0.2 10% O
2 ACKNOWLEDGEMENT
20% O2
-0.4 30% O2
40% O2
The work presented in this paper has been supported by the
-0.6 50% O Ministry for Education, Science and Technological
2
60% O2 Development of Serbia.
-0.8 70% O2
80% O2
-1 90% O2
REFERENCES
100% O2
-1.2 [1] Q. J. Zhang and K. C. Gupta, Neural Networks for RF and
2 2.1 2.2 2.3 2.4 2.5 Microwave Design, Boston, MA: Artech House, 2000.
f [GHz] [2] H. Kabir, L. Zhang, M. Yu, P. Aaen, J. Wood, and Q. J. Zhang,
“Smart Modeling of Microwave Devices”, IEEE Microw. Mag.,
(a) vol. 11, no. 3, pp. 105-118, May 2010.
0.8 [3] Z. Marinković, G. Crupi, A. Caddemi, V. Marković, and D. M.
M.-P. Schreurs, “A Review on the Artificial Neural Network
0.6 Applications for Small-Signal Modeling of Microwave FETs,”
Int. J. Numer. Model. Electron. Netw. Dev. Field, vol. 33, no. 3,
0.4 May/June 2020, e2668.
0.2
[4] Z. Marinković, T. Kim, V. Marković, M. Milijić, O. Pronić-
0% O2 Rančić, T. Ćirić, L. Vietzorreck, “Artificial Neural Network
0 10% O2 Based Design of RF MEMS Capacitive Shunt Switches”,
20% O
2
Applied Computational Electromagnetics Society Journal, vol.
-0.2 30% O2 31 no. 7, pp. 756-764, July 2016.
40% O2 [5] J. Jin, F. Feng, W. Na et al. “Recent Advances in Neural
-0.4
50% O2 Network Based Inverse Modeling Techniques for Microwave
-0.6 60% O2 Applications”, Int. J. Numer. Model. Electron. Netw. Dev.
70% O2 Field, e2732, 2020.
-0.8 80% O2 [6] G. Gugliandolo, D. Aloisio, S. G. Leonardi, G. Campobello, and
90% O2 N. Donato, “Resonant Devices and Gas Sensing: from Low
-1
100% O2 Frequencies to Microwave Range”, 14th International
-1.2 Conference on Advanced Technologies, Systems and Services
2 2.1 2.2 2.3 2.4 2.5 in Telecommunications (TELSIKS 2019), Oct. 2019, pp. 21-28.
f [GHz] [7] G. Bailly, A. Harrabi, J. Rossignol, D. Stuerga, and P. Pribetich,
(b) “Microwave Gas Sensing with a Microstrip InterDigital
Capacitor: Detection of NH3 with TiO2 nanoparticles”, Sensors
Fig. 6. Y11 for the oxygen concentrations from 0% to 100%: and Actuators, B: Chemical, 236, pp. 554-564, 2016.
(a) real parts, (b) imaginary parts [8] A. Rydosz, E. Maciak, K. Wincza, and S. Gruszczynski,
“Microwave-Based Sensors with Phthalocyanine Films for
Namely, with a change of the oxygen concentration the Acetone, Ethanol and Methanol Detection”, Sensors and
resonant frequency (when Y11 becomes real, i.e., having the Actuators B: Chemical, 237, pp. 876-886, 2016.
imaginary part equal to zero) is shifted. As the model gives [9] D. Aloisio and N. Donato, “Development of Gas Sensors on
very good accuracy in the areas where the imaginary part of Microstrip Disk Resonators”, Procedia Eng., vol. 87, pp. 1083-
Y11 crosses the zero, the model can be efficiently used for the 1086, 2014, doi: 10.1016/j.proeng.2014.11.351
analysis of the device as an oxygen sensor. [10] G. Gugliandolo, A. Mirzaei, M. Bonyani, G. Neri, M. Latino,
and N. Donato, “Electrical Characterization of Microstrip
Resonators Based on Nanostructured Sensing Materials”,
IV. CONCLUSION Sensors and Microsystems, LNEE, vol. 457, 2018, pp. 29-34.
[11] G. Gugliandolo, M. Latino, G. Campobello, Z. Marinkovic, G.
Crupi, and N. Donato, “On the Gas Sensing Properties of
In this paper an ANN based model for modeling of the Microwave Transducers”, 55th Scientific Conference on
admittance parameters of an interdigital capacitor versus Information, Communication and Energy Systems and
frequency and oxygen concentration was proposed. As for the Technologies – ICEST 2020, September 10-20, 2020, Niš,
considered device all four Y-parameters are roughly equal, it Serbia.
was enough to develop a model for only one parameter, Y11.
The PKI approach was applied, having the values of the real
and imaginary parts of Y11 at a reference concentration (70%

198
 Analysis of Dual Band PA Nonlinear Performances by
Application of Digital Linearisation Model
Aleksandar Atanasković1, Aleksandra Đorić1, Nataša Maleš-Ilić1
Abstract – A dual band power amplifier designed to sustain
high efficiency at two diverse working frequencies, 2.1 GHz and
5 GHz, is linearised by the technique that uses baseband signal
processing in the digital domain. Generation of the dual band PA
matching circuits consist of two individual steps. Firstly,
transformation of the selected complex impedances at the two
operating frequencies into the real impedance identical for the
both observed frequencies is performed. Secondly, synthesis of
the dual band filter in order to provide matching of the real
impedance with the termination impedances is carried out. The
linearisation of the dual band amplifier is accomplished by the
2nd order nonlinear baseband signals, which modulate the carrier
2nd harmonic. The generated signals in this manner are then
induced at the gate and drain of the dual band PA transistor.
The assessment of the linearisation influence was performed for
the two baseband signals spaced in frequency by diversified
intervals up to 80 MHz for output levels-Pout from 25 dBm to
35 dBm for both considered frequencies.
Keywords – Dual band PA, dual band filter, digitally processed
baseband signals, 2nd harmonic, IM products.
I. INTRODUCTION
Recently, in the modern communication systems, necessity
of implementation of electronic equipment that can
concurrently operate over several frequency bands is rapidly
growing. Consequently, deployment of multiband PA has
become an attractive research theme in the area of wireless
microwave systems. The literature includes various reports for
different multistandard PA solutions [1-3].
In the former study phases, the linearisation technique that
utilizes the digital processing of the baseband signals was
deployed in order to reduce the PA nonlinearity. The adequate
linearisation signals of the 2nd order, that modulate the 2h (2nd
harmonic) of the fundamental carrier, were formed by this
approach and the modulated signals at the 2h are then fed to
the amplifier transistor gate and drain. In accordance with the
analysis carried out in [4-10], due to the 2nd order transistor
nonlinearity, the interference between the modulated 2h and
the basic useful signals creates the additional IM products of
the 3rd order. These additional IM products may lessen the
primary nonlinear products resulting from transistor
nonlinearity by suitable selection of the phases and amplitudes
of the linearisation signals.
Aleksandar Atanaskovic, Aleksandra Đorić and Nataša Maleš-
Ilić are with University of Niš, Faculty of Electronic Engineering,
aleksandar.atanaskovic;
Aleksandra Medvedeva aleksandra.djoric;
14, 18000 Niš, Serbia, E-mails:
aleksandar.atanaskovic@elfak.ni.ac.rs;
natasa.males.ilic)@elfak.ni.ac.rs.
aleksandra.djoric@elfak.ni.ac.rs; natasa.males.ilic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
The estimation of the proposed technique effects was
performed through the simulation analysis and experimentally
on several single stage PA [4-6] and two-way Doherty PA
topologies [7-9], as well for the mixer [10].
In the research of this paper, we extend the application of
the formerly developed digital linearisation model to the
designed PA which operates at two different frequency bands
with satisfying accomplishment of the main operating
requirements, such as efficiency, gain and linearity.
Hitherto, the modeled dual band PA was successfully
linearised by the formerly deployed analogue linearisation
approach [11] that uses the 2nd order distortions of the
fundamental signals (IM2) around the 2h frequencies.
In this paper, for the first time, the digital linearisation
approach that uses the digitally processed baseband signals
was applied on the dual band PA designed to operate at two
divers frequencies, f1 (2.1 GHz) and f2 (5 GHz). Basic steps of
the dual band MN designing method [12-15] are shortly
described in this paper. The effect of the linearisation
application is inspected by simulation for the QAM baseband
source, which spectrum has two frequency components
spaced by different values starting from 20 MHz and going to
80 MHz, as also for the different Pout levels 25 dBm, 30 dBm
and 35 dBm.
II. APPLICATION OF THE DIGITAL LINEARISATION
MODEL ON THE DUAL BAND PA
The applied linearisation model uses the baseband signal
transformation DSP system (Fig. 1) to generate adequate in-
phase IIM2 and quadrature-phase QIM2 components of the 2nd
order nonlinear signals. The IIM2 and QIM2 signals are properly
tuned in magnitude by coefficients a{i2h|o2h} and phase by
θ{i2h|o2h} and modulate the fundamental carrier 2h. The
modulated linearisation signals shaped in this way are then
conducted into an IQ modulator with a central frequencies of
the fundamental carrier 2h. The signals are then lead at the
input and at output of the dual band PA transistor by the BPF
with central frequencies 2f1 and 2f2. Indices i2h and o2h mark
the signals that are supplied at the amplifier transistor gate and
drain, respectively.
199
 Fig. 1. Scheme of dual band PA linerisation model which uses the digital processing of the baseband signals
A. Dual Band PA Model
The tool used for the synthesis of the PA source and load
dual band MN (matching network) was software ADS
(Advanced Design System).
The dual band PA was modeled in configuration with
distributed elements with aim to perform concurrently at two
broadly spaced frequencies, f1 = 2.1 GHz and f2 = 5 GHz, and
the Cree transistor model CGH40025F GaN was used in the
dual band PA design.
The dual band matching was realized through two major
stages [11-15]. In the first stage, selection of the optimum
impedances at the two operative frequencies, 2.1 GHz and
5 GHz, was made based on the ADS load-pull/source-pull
method with the aim to accomplish high efficiency at both
observed frequencies. The obtained matching source and load
normalized impedances in the case of the 2.1 GHz are
Zsf1 = (0.124 + j0.096) Ω and Zlf1 = (0.188 + j0.114) Ω,
respectively, while for the 5 GHz they are
Zsf2 = (0.174 – j0.086) Ω and Zlf2 = (0.166 – j0.088) Ω. The
dual band PA is powered by VD = 28 V, IDQ = 350 mA. These
selected complex impedances at two working frequencies f1
and f2 are converted into the real impedance (R) that has the
same value for the both frequencies. In the second stage, the
dual band filter theory was used to convert previously gained
real impedance R to 50 Ω terminations at two targeted
frequencies.
B. Results
The evaluation of the applied linearisation model on the
modeled dual band PA was executed through the ADS QAM
source baseband test. The frequency separation of the
baseband signal pairs goes from 20 MHz to 80 MHz in regard
200
to the central frequencies 2.1 GHz or 5 GHz. These signals are
conducted simultaneously at the PA input for several Pout
levels of the fundamental signals: 25 dBm, 30 dBm and
35 dBm.
The IM products of the 3rd (IM3) and 5th (IM5) order are
represented in the Fig. 2 to Fig. 5, for both considered
frequency bands, with and without application of the
linearisation model for various Pout levels and diverse
frequency span between the tones.
It should be pointed out that represented results for all
analyzed Pout were obtained in case when the optimization of
the linearisation parameters (phase and amplitude of the
baseband linearisation signals) was carried out for 40 MHz
input signal differences regarding frequency.
In the lower band, around the 2.1 GHz frequencies, the
highest suppressing of the IM3 products is achieved for the
20 MHz signals separation for all observed Pout levels. In the
case of the 25 dBm Pout, satisfactory reduction of the IM3
products was accomplished until the 50 MHz tones spacing.
The IM3 levels are lessened 16 dB for the 20 MHz span and
with the increase of the frequency separation between the
tones their reduction decrease to 3 dB at the 80 MHz
separation in the case of the 30 dBm Pout. Also, the IM3
levels are decreased by 15 dB for the 35 dBm Pout and the
20 MHz signals spacing, whereas at the 80 MHz spacing their
reduction level falls to 9 dB.
Very symmetrical suppression grade is achieved, regarding
the IM5 products, for the 25 dBm and 30 dBm Pout. The IM5
levels descending around 3 dB for the 25 dBm Pout for all
considered tone spans while for the 30 dBm Pout their
reduction values goes from 15 dB for the 20 MHz span to
12 dB for the 80 MHz signal separation. For the highest
considered Pout of 35 dBm, the IM5 levels descend by 10 dB
in the case of the 20 MHz spacing until 2 dB at the 80 MHz
tones separation.
 In the higher band, around the 5 GHz frequencies, when the signals are formed in the baseband by using DSP and
Pout is 25 dBm, the IM3 products are descended by 12 dB for appropriate hardware, which energy consumption is more
the 20 MHz tones separation whereas their suppression value optimal than the linearisation circuit devices needed in the
for the 80 MHz drops to 1 dB. For the 30 dBm Pout, the IM3 analogue method.
levels decline for 15 dB for the 20 MHz spacing until the
value of 10 dB at the 80 MHz interval between the tones. In
the case of the Pout 35 dBm, the reduction level of the IM3
distortions drops from 9 dB at 20 MHz spacing until 3 dB for
the 80 MHz tones shifting.
The IM5 products reduction level for the 25 dBm Pout is
around 8 dB and for the 30 dBm Pout is around 12 dB for all
observed frequency intervals while for the 35 dBm Pout
satisfactory suppression is accomplished only for the smallest
tone spacing. a) b)

a) b) c)
Fig. 3. 5th order IM products of the dual band amplifier at 2.1 GHz
before and after linearisation: a) Pout =25 dBm, b) Pout =30 dBm,
c) Pout =35 dBm

c)
rd
Fig. 2. 3 order IM products of the dual band amplifier at 2.1 GHz
before and after linearisation: a) Pout =25 dBm, b) Pout =30 dBm,
c) Pout =35 dBm a) b)

Furthermore, the linearisation results of the dual band PA

achieved in this paper by application of the digital
linearisation model are compared with the acquired results
with application of the analogue linearisation approach
executed previously [11]. Observing the analogue
linearisation method results of dual band PA presented in [11]
we can spot firstly divers IM3 and IM5 signal power levels
before linearisation in comparison with the digital
linearisation method. Explanation for this was found in
different analysis performed in ADS during simulation of the c)
analogue and digital approaches, in the first case it was Fig. 4.3rdorder IM products of the dual band amplifier at 5 GHz
Harmonic Balance and in the digital we used Data flow. before and after linearisation: a) Pout =25 dBm, b) Pout =30 dBm,
Secondly, even though the IM3 products were lessened by c) Pout =35 dBm
greater rate in the analogue than in the digital linearisation, Also, the digital linearisation model was applied on the dual
especially for 2.1 GHz signal frequency, the IM5 products band PA for both targeted frequencies simultaneously, and
were lower approximately the same. results gave a certain number of additional mixed cross IM
Moreover, the digital linearisation model applied in this distortions between the each two signals from different bands.
paper does not require additional sources of 2h signal for The research showed the same results as for the case when the
linearisation and amplifiers in the linearisation circuits as it is analogue linearisation method is performed, that these mixed
in the analogue approach, [11]. The baseband linearisation IM products cannot be descended simultaneously. In the

201
future study, focus will be directed on discovering a solution
for concurrent reduction of all IM distortions which spectrum
contains. One of the ideas is to include circuits at the PA
transistor gate and drain that should short the signals at the
difference frequency components between the two operational
bands of the PA.
a) b)
c)
Fig. 5. 5th order IM products of the dual band amplifier at 5 GHz
before and after linearisation: a) Pout =25 dBm, b) Pout =30 dBm,
c) Pout =35 dBm
III. CONCLUSION
This paper offers results of the digital linearisation
procedure applied on the formerly designed dual band PA.
The suggested linerisation model uses the 2nd order signals
that modulate the 2h of the fundamental carrier for the 3rd and
5th order IM products reduction. The dual band PA was
designed to operate concurrently at two frequency bands
around 2.1 GHz and 5 GHz. The dual band PA linearisation
effects were demonstrated through the QAM baseband test at
both working bands. The digital linearisation model attains
satisfactory results concerning the repression of the IM
products of the 3rd and 5th order in case when it is applied to
each band separately. It should stress that the digital
linearisation performed in this paper achieves good
linearisation results with the linearisation circuit topology that
is less demanding regarding the power supply and complexity
in comparison with the analogue linearisation method
demonstrated in our earlier papers.
ACKNOWLEDGMENT
This work has been supported by the Ministry of Education,
Science and Technological Development of the Republic of
Serbia.
202
REFERENCES
[1] K. Rawat, F. M. Ghannouchi, “Dual-Band Matching Technique
based on Dual-Characteristic Impedance Transformers for Dual-
Band Power Amplifiers Design”, IET Microw. Antennas &
Propag., vol. 5, no. 14, pp. 1720-1729, 2011.
[2] Y. Ding, Y. X. Guo and F. L. Liu, “High-Efficiency Concurrent
Dual-Band Class-F and Inverse Class-F Power Amplifier”,
Electronics Letters, vol. 47, no. 15, pp. 847-849, July 2011.
[3] P. Wright, J. Lees, J. Benedikt, P. J. Tasker and S. C. Cripps, “A
Methodology for Realizing High Efficiency Class-J in a Linear
and Broadband PA”, IEEE Trans. Microw. Theory Tech., vol.
57, no. 12, pp. 3196-3204, December 2009.
[4] A. Atanasković, N. Males-Ilić, K. Blau, A. Đorić and B.
Milovanović, “A Novel Linearization Technique based on
Modified Baseband Signal that Modulates Carrier Second
Harmonic”, TELSIKS 2013, Conference Proceedings, pp. 192-
195, Niš, Serbia, 16-19 October, 2013.
[5] A. Atanasković, N. Males-Ilić, K. Blau, A. Đorić and B.
Milovanović, “RF PA Linearization using Modified Baseband
Signal that Modulates Carrier Second Harmonic”, Microwave
Review, vol. 19, no.2, pp. 119-124, December 2013.
[6] A. Đorić, N. Maleš-Ilić and A. Atanasković, “RF PA
Linearization by Signals Modified in Baseband Digital
Domain”, Facta Universitatis, Series: Electronics and
Energetics, vol. 30, no. 2, pp. 209-221, June 2017.
[7] A. Đorić, N. Maleš-Ilić, A. Atanasković and V. Marković,
“Linearization of Broadband Doherty Amplifier by Baseband
Signal that Modulates Second Harmonic”, IEEE EUROCON
2017, Conference Proceedings, pp. 206-211, Ohrid, Macedonia,
6-8 July, 2017.
[8] A. Đorić, A. Atanasković, N. Maleš-Ilić and M. Živanović,
“Linearisation of RF PA by Even-Order Nonlinear Baseband
Signal Processed in Digital Domain”, International Journal of
Electronics, Published online: vol. 106, Issue 12, pp. 1904-
1918, 27 June 2019.
[9] A. Đorić, A. Atanasković, B. Alorda, and N. Maleš-Ilić,
“Linearization of Doherty Amplifier by Injection of Digitally
Processed Baseband Signals at the Output of the Main and
Auxiliary Cell”, TELSIKS 2019, Conference Proceedings, pp.
339-342, Niš, Serbia, 23-25 October, 2019.
[10] A. Đorić, N. Maleš-Ilić, A. Atanasković and B. Milovanović,
“Mixer Linearization by Modified Baseband Signals”, Sinteza
2016, Conference Proceedings, pp. 299-303, Belgrade, Serbia,
22 April 2016.
[11] A. Atanasković, A. Đorić, B. Alorda, N. Maleš-Ilić, “Dual Band
Power Amplifier Linearization”, TELSIKS 2019, Conference
Proceedings, pp. 37-40, Niš, Serbia, 23-25 October 2019.
[12] X. Fu, D. T. Bespalko and S. Boumaiza, “Novel Dual-Band
Matching Network for Effective Design of Concurrent Dual-
Band Power Amplifiers”, IEEE Trans on Circuits and Systems
I-Regular paper, vol. 61, no. 1, pp. 293-301, January 2014.
[13] X. Fu, Dual-band Power Amplifier for Wireless Communication
Base Stations, Master Thesis, Waterloo, Ontario, Canada, 2012.
[14] X. Fu, D. T. Bespalko and S. Boumaiza, “Novel Dual-Band
Matching Network Topology and its Application for the Design
of Dual-Band Class J Power Amplifiers”, IEEE MTT-S
International Microwave Symposium Digest, Conference
Proceedings, Montreal, 2012.
[15] H.-Y. A. Yim, K.-K. M. Cheng, “Novel Dual-Band Planar
Resonator and Admittance Inverter for Filter Design and
Applications”, IEEE MTT-S Int. Microwave Symp. Dig.,
Conference Proceedings, October 2005.
 Design of Asymmetrical Slot Antenna Array in
Corner Reflector
Marija Milijic1 and Branka Jokanovic2
Abstract – In this letter, a design procedure of the K-band slot
antenna array in corner reflector is described. The array consists
of eight slots as radiating elements with parallel-series feeding by
coplanar waveguide (CPW) transmission line. The uniform
rectangular slots stand asymmetrically in relation to the CPW
feeding line. The apex of the corner reflector is at the distance
around 1.4λ0 in respect to the center frequency fc=25.875 GHz
from the antenna array. The corner reflector increases antenna
gain and improves control of radiation pattern that is desirable
for many modern wireless applications. The proposed antenna
array features gain of 21.6 dBi and side lobe suppression of
14.2 dB at the center frequency, while it keeps satisfied radiation
properties in a wide frequency range between 24.8 GHz and
27.1 GHz.
Keywords – Antenna array, Corner reflector, Asymmetrical
slot antenna, CPW-fed antenna, Parallel-Series feeding, CPW T-
junction.
I. INTRODUCTION
With the rapid growth in wireless communications across
the world, the demands for fast transmission speeds and great
broadband capacities have risen exponentially to enable easy
access the internet, IoT and M2M applications [1]. In order to
cater the high data transmission rate, the spectrum efficiency
and smarter and reliable service for connection lots of people
and objects, the millimeter wave frequency region is being
proposed as promising solution to the bandwidth and capacity
issues [1]. The International Telecommunication Union (ITU)
has announced that 5G wireless communication systems can
operate in millimeter-wave bands where the most promising
frequency ranges are 24.25 to 27.5 GHz, 37–40.5 GHz,
66–76 GHz [2]. Meantime, the Federal Communications
Commission (FCC) has estimated the spectrum of around
11 GHz above 24 GHz for the latest fifth generation wireless
technology systems in the United States [3].
However, investigation and development of millimeter-
band antennas have been started to meet the needs of modern
5G wireless communication system. Firstly, atmospheric
attenuation and path losses become critically higher at these
frequencies [4-6]. Furthermore, the received signal power
weakens at millimeter-wave frequencies as compared to lower
frequency bands as a result of smaller effective aperture size
1Marija Milijic is with the Faculty of Electronic Engineering, University of
Niš, 18000 Nis, Serbia, E-mail: marija.milijic@elfak.ni.ac.rs
2Branka Jokanovic is with the Institute of Physics, University of Belgrade,
Pregrevica 118, 11080 Pregrevica, Serbia, and The Academy of Engineering
Sciences of Serbia, Kraljice Marije 16, Belgrade, Serbia, E-mail:
brankaj@ipb.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
of single element antenna. Therefore, to improve and optimize
the antenna performance in the millimeter-wave frequencies
communication systems, dealing with a problem of mentioned
losses, the use of high gain antenna arrays should be
fundamental.
Lately, CPW-fed slot antennas have been popular and used
as radar sensors and for mobile communication applications
[7-9]. This type of antennas is characterized by its wideband
operation, compact size, planar structure and cheap and
uncomplicated fabrication. Still, since the radiator and the
ground plane are etched onto one side of the substrate, the
energy will be radiated bidirectionally, which results in a low
antenna gain. The antenna should radiate only in a half the
atmosphere to satisfy the requirements of wide-band
applications, where high-gain and unidirectional radiation are
crucial requirements.
This paper presents a centrally offset fed antenna array of
identical rectangular slots positioned asymmetrically relative
to the CPW line. In our papers [10-11] we demonstrated
advantages of using asymmetrical slots in the design of
centrally and serially fed antenna arrays. Here we proposed
the linear slot antenna array placed in a corner reflector in
order to increase the gain and side lobe suppression. Presented
simulated results show very satisfying radiation properties in a
wide frequency range making proposed antenna as good
choice for 5G and radar applications. Asymmetrical slot
antenna array with corner reflector exhibits 5 dBi greater gain,
0.55 GHz wider frequency range with appropriate radiation
pattern and almost two times wider impedance bandwidth for
the criterion S11<-10 dB than the version with parallel
reflector plate.
II. THE ARRAY OF CPW-FED ASYMMETRICAL
SLOTS IN CORNER REFLECTOR
The antenna array of eight identical asymmetrical slots,
CPW feeding line and CPW T-junction on the dielectric
substrate having relative permittivity of εr=2.54 and height of
0.508 mm is depicted in Fig. 1. The array is placed parallel to
the apex of corner reflector with an angle α=90° at the
distance h=16 mm away, that is around 1.4λ0 (at the center
frequency fc=25.875 GHz). The center frequency
fc=25.875 GHz is determined as the median of range
24.25–27.5 GHz, proposed by ITU [2]. The array is formed of
two equal four-slot subarrays set at the different distances
from the CPW T-junction. To ensure in-phase feeding of both
subarrays, the CPW T-junction is moved off the symmetry
203
 Fig. 1. Asymmetrical slot antenna array in corner reflector

a)

a) b)
Fig. 2. Feeding of antenna subarrays by:
a) 2 generators b) CPW T-junction

axes for λg/4, where λg=9 mm is CPW line wavelength at the

center frequency fc [10]. The 60 Ω CPW line (featuring the
strip w1=0.9 mm and gap g1=0.1 mm) is used to enable
feeding for both sub - arrays dividing power into two 120 Ω
CPW feed lines featuring the strip w2=0.3 mm and gap
g2=0.375 mm for feeding every four slots sub - array.
The antenna substrate has the geometry of 77 mm x
34.2 mm x 0.508 mm. The substrate is placed between two
metal panels, set under α=90° angle, that form a corner
reflector. The dimensions of each reflector panel are 77 mm x
45 mm. CPW feeding line penetrates the connection of two
reflector panels through a hole with the sufficient holes’
b)
diameter (2 mm) to minimize the undesirable influence of the
metallic panels. Each slot has a rectangular shape with width a
and arms of different length b1 and b2 resulting in entire slot
length l=b1+b2+w2+2g2. Also, the mutual distance between
neighboring slots is d=9 mm. The commercial software tool
WIPL-D [12] is used to obtain the simulated results of
proposed antenna array.
The process of antenna design has two steps. Firstly, both
antenna subarrays are fed by two independent generators in
the beginning of their CPW feeding lines (Fig. 2a) to obtain
optimized slots dimensions at the center frequency fc.
Afterwards, the subarrays of slots with optimal dimensions are
connected via CPW T-junction (Fig. 2b). The Fig. 3. shows
the simulated radiation patterns through two steps of antenna
optimization for the different slots width a values ranging
from a∈[0.5 mm, 1.5 mm]. The length of arms b1 and b2 are
adjusted in order to have center frequency at fc=25.875 GHz
while total slot length l is about 9 mm. It can be seen that c)
when antenna subarrays are fed by two generators (dash Fig. 3. Radiation pattern in xOz plane (φ=0°) of the slot array fed by
lines), the gain is above 21.5 dBi: 21.6 dBi for a=0.5 mm, two generators (dashed) and by CPW T-junction (solid) at 25.875
GHz for different slot widths a: (a) a=0.5mm, (b) a=1mm, (c)
a=1.5mm. Antenna array is in the corner reflector with angle α=90°

204
 a) the xOz plane (φ=0°)

Fig. 4. S11 parameter versus frequency of the eight asymmetrical slot

array fed by CPW T-junction for different slots width a. S11 is
normalized to the impedance of CPW feeding line (60 Ω)

22.2 dBi for a=1 mm and 22.6 dBi for a=1.5 mm. Side lobe
suppression is 14.3 dB for a=0.5 mm, 13.7 dB for a=1 mm
and 15.2 dB for a=1.5 mm. Furthermore, the both subarrays
are joined through the CPW T – junction (solid lines)
resulting in a bit reduced gain: 19.7 dBi for a=0.5 mm, 20.2
dBi for a=1 mm and 21.6 dBi for a=1.5 mm. Similarly, side
lobe suppression has also decreased and it is 12.3 dB for
a=0.5 mm, 13.2 dB for a=1 mm and 14.2 dB for a=1.5 mm.
The decrease in gain and side lobe suppression when the array
is excited by CPW T – junction with reference to separated
feeding for two sub-arrays, can be a consequence of the b) the yOz plane (φ=90°)
reflected power from the neighboring sub – array through the
CPW T – junction. Although, the least deviation between Fig. 5. Radiation patterns of the slot array fed by CPW T-junction
simulated radiation patterns for two different way of feeding in the corner reflector at 24.8 GHz, 25.875 GHz and 27.1 GHz
is when slot has width a=1.5 mm.
Moreover, the S11 parameters of considered antenna array TABLE I
fed by CPW T-junction for different slots width a values from MAIN CHARACTERISTICS OF THE A SLOT ARRAY FED BY CPW T-
JUNCTION IN THE CORNER REFLECTOR AT THE CENTER FREQUENCY
range from a∈[0.5 mm, 1.5 mm] are presented in Fig. 4. The
fc=25.875 GHZ AND AT THE EDGE FREQUENCIES flef=24.8 GHZ AND
reflection coefficient S11 is normalized to the impedance of
fuef=27.1 GHZ
CPW feeding line (60 Ω) where generator is attached. It is less
Frequency [GHz] Feeding Gain [dBi] SLS [dB]
than -10 dB for the frequency range between 24.5-26.3 GHz
2 generators 22.6 15.2
(dash line) for the slot width a=0.5 mm, 24.4-26.6 GHz (dot fc=25.875
CPW –T junction 21.6 14.2
line) for slot width a=1 mm and 24.8-27.1 GHz (solid line) for
2 generators 21.95 10.6
slot width a=1.5 mm. Based on the presented simulated results flef=24.8
CPW –T junction 19.7 10
in Figs. 3-4, it can be concluded that array of eight
2 generators 22.5 11
asymmetrical slots (with width a=1.5 mm) positioned in the fuef=27.1
CPW –T junction 19.7 10.3
corner reflector with an angle α=90° has the best simulated
results considering gain, side lobe suppression and S11
center frequency. Furthermore, it is observed in the yOz plane
parameters. Its maximum gain, when CPW T-junction is used
Fig. 5b that antenna gain stays almost unchanged at the
for feeding, is 21.6i dB at the center frequency
(center and edge frequencies while the side lobe suppression
fc=25.875 GHz. If maximum allowed gain deviation is set to
drops to 8-9 dB. Table I gives the comparison between the
be 2 dB, the frequencies between 24.8 GHz to 27.1 GHz
main parameters of the asymmetrical slot array in the corner
satisfy this criterion because gain is 19.7 dBi (Fig. 5) at the
reflector through the design steps at center and edge
specified edge frequencies. Side lobe suppression is 10-10.5
frequencies. 3D radiation pattern at the center frequency fc, for
dB at the edge frequencies, which is 3.5-4 dB less than at the
both the xOz and yOz plane is displayed in Fig. 6.

205

a) the xOz plane (φ=0°)
b) the yOz plane (φ=90°)
Fig. 6. 3D Radiation patterns of the asymmetrical slots array fed
by CPW T-junction in corner reflector at 25.875 GHz
Considering all simulated radiation patterns and the results
shown in Figs. 3-6, it can conclude that the proposed antenna
array exhibits operating frequency range from 24.8 GHz to
27.1 GHz that is about 9% of the central frequency.
In respect to previous results [10] of eight asymmetrical slot
array with parallel reflector plane, the proposed antenna array
with corner reflector with the same feeding structure of CPW
T-junction, has 5 dBi greater gain, 0.55 GHz wider frequency
range with appropriate radiation pattern and almost two times
bigger impedance range with satisfied criterion S11<-10 dB.
Side lobe suppression is the same for both antenna arrays
(13-14 dB) as it is expected for the uniform power
distribution.
Due to strict requirements to be met by the 5G network,
which relate to the high gain and wide operating bandwidth,
the obtained results show that CPW-fed asymmetrical slot
array in corner reflector is a very promising solution.
III. CONCLUSION
Design of millimeter-band antennas is probably the most
challenging feature in fifth generation communications. CPW
fed antennas have been regarded as the most optimal choice in
206
mobile phones and smart devices because of their planar
structures and inexpensive fabrication of feeding lines and
radiating elements on the same side of substrate. Unlike the
microstrip antennas with a backside ground plane, the CPW
fed slot antennas need the reflector to establish a radiation in
half of the atmosphere. There are previous published results of
CPW fed asymmetrical slot array with parallel reflector plane
that feature acceptable gain and bandwidth characteristics.
Although, in order to compensate losses that cause the
received signal power reduction, the mm-wave
communication systems are required high gain antennas.
The proposed CPW-fed asymmetrical slot array is with
corner reflector of 90° angle that significantly improve its gain
and bandwidth characteristics. Its gain is 21.6 dBi at the
center frequency of 25.875 GHz while it is around 20 dBi in
the range from 24.8 GHz to 27.1 GHz. Although, the corner
reflector antenna is 3D structure, it is still light weight and
simple for fabrication which together with greater gain and
increased impedance bandwidth makes it very suitable for
mm-wave applications.
REFERENCES
[1] V. Milosevic, B. Jokanovic, O. Boric-Lubecke, and V. M. Lubecke,
“Key Microwave and Millimeter Wave Technologies for 5G
Radio”, in Powering the Internet of Things with 5G Networks, V.
Mohanan, R. Budiatru, I. Aldmour, Eds. IGI Global, July 2017,
DOI: 10.4018/978-1-5225-2799-2.
[2] W. Hong, K. Baek, and S. Ko, “Millimeter-Wave 5G Antennas for
Smartphones: Overview and Experimental Demonstration”, IEEE
Trans. Antennas Propag., December 2017, vol. 65, no. 12, pp. 6250-
6261.
[3] M. J. Marcus, “5G and "IMT for 2020 and Beyond"“, IEEE
Wireless Commun., Aug. 2015, vol. 22, no. 4, pp. 2-3.
[4] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, “Key Elements to Enable
Millimeter Wave Communications for 5G Wireless Systems”, IEEE
Wireless Commun, vol. 21, no. 6, pp. 136-143, Dec. 2014.
[5] T. S. Rappaport et al., “Millimeter-wave mobile Communications
for 5G Cellular: It will Work!”, IEEE Access, pp. 335-349, 2013.
[6] Q. Zhao, J. Li, “Rain Attenuation in Millimeter Wave Ranges”, 7th
Int. Symp. Antennas Propag. & EM Theory, pp. 1-4, Oct. 2006.
[7] M. E. de Cos Gómez, H. Fernández Álvarez, C. García González,
B. Puerto Valcarce, J. Olenick, and F. Las-Heras, "Ultra-Thin
Compact Flexible Antenna for IoT Applications”, 2019 13th Eur.
Conf. Antennas Propag. (EuCAP), Krakow, Poland, 2019, pp. 1-4.
[8] A. I. Afifi, D. M. Elsheakh, A. B. Abdel-Rahman, A. Allam, and S.
M. Ahmed, “Dual Broadband Coplanar Waveguide-Fed Slot
Antenna for 5G Applications”, 2019 13th Eur. Conf. Antennas
Propag. (EuCAP), Krakow, Poland, 2019, pp. 1-3.
[9] A. Zaidi, A. Baghdad, W. A. Awan, Halima, A. Ballouk, and A.
Badri, “CPW Fed Wide to Dual Band Frequency Reconfigurable
Antenna for 5G Applications”, 2019 Int. Conf. Wireless Tech.,
Emb. and Intel. Sys. (WITS), Fez, Morocco, 2019, pp. 1-3.
[10] M. Milijić, B. Jokanović, “Design of CPW-fed Asymmetrical Slot
Array for K Band Applications”, 54rd Int. Scientific Conference
ICEST 2019, pp. 19-22, Ohrid, North Macedonia, June 27-29.
[11] M. Milijić, B. Jokanović, “Uniplanar Series Fed Antenna Array
with Asymmetrical Slots for K-band Applications“, 14th Int. Conf.
TELSIKS 2019, pp. 144-147, Niš, Serbia, 23-25 October.
[12] WIPL-D Pro v15.0, WIPL-D Team
 Radar Antenna Array Designing for
Unmanned Aerial Vehicles (UAV)
Ivaylo Nachev1 and Peter Z. Petkov2
Abstract – In this paper, an application of a simple phased the air gap between ground plane and antenna is Δ=3mm. The
antenna array intended to increase the range of a radar module operating frequency is chosen to be ƒс=10.525GHz (a
is presented. The radar with improved range can find an frequency preferred in the production of small size radar
application in control of unmanned vehicles – mainly drones, but modules).
also cars and robots for civil and special purposes. The first part
Taking into account the abovementioned design method [4],
of this paper describes design of a patch antenna with operating
frequency of 10.5 GHz (Amateur radio X-band). Comparison of the antenna dimensions are calculated and presented in Table
the simulation results with measured parameters of fabricated I. In Fig. 2 are photographs of the fabricated prototypes. Fig. 3
antenna is made. The second part of the paper describes shows the construction of the suspended substrate and the air
simulation results for how these models could be multiplied and gap.
incorporated into a series or corporate feed array. The delay
lines technique used for beam steering, along with for beam
management algorithm are presented.

Keywords – Autonomous vehicles, Drones, Radar, Antenna,

Phased antenna array, Beam scanning, Delay line phase shifting.

I. INTRODUCTION
Fig. 1. The HB-100 radar module
In the last years, cost-effective radar systems and modules
find their wide application in many end-user products. TABLE I
Deployment of new radar modules can be combined with DIMENSION OF PATCH ANTENNA
patch antennas and antennas arrays. In this way, the radar
range and accuracy will increase. Given the low price of patch W L w MS Line Feed Line
antennas and cheap manufacturing process, these modules are [mm] [mm] [mm] [mm] [mm]
suitable for large–scale production. The antenna arrays with a Calculated 12.2 12.96 5.3 12.4 6.2
scanning antenna beam finds special application in radar Adjusted 13.1 11.00 5.3 12.4 6.2
systems because the beam can crawl the space in front of the after
antenna. In this way, different objects ahead of the vehicle can simulation
be detected. This can integrate for vital application in
difference autonomous vehicles – cars [1], drones [2] and
robots [3]. There are many published developments with
various complexity and manufacturing cost. The present paper
examines the application of one of the most inexpensive ones.

II. PATCH ANTENNA

The patch antenna as radiating elements is used, as
dimensions are determined by a standard design method, well Fig. 2. Fabricated patches for antenna array
explained in [4]. The employed dielectric substrate for this
design is Rogers RO4003 [5]. In order to improve the
bandwidth and efficiency of antennas, a design approach
called “suspended substrate” is chosen, where the thickness of

1Ivaylo Nachev is PhD student with the Faculty of

Telecommunications at Technical University of Sofia, 8 Kl. Ohridski
Bvld., Sofia 1000, Bulgaria E-mail: ivaylonachev@yahoo.com
2Peter Z. Petkov is Associate Professor with the Faculty of

Telecommunications at Technical University of Sofia, 8 Kl. Ohridski Fig. 3. The realization suspended substrate
Bvld., Sofia 1000, Bulgaria E-mail: pjpetkov@tu-sofia.bg

978-1-7281-7142-5/20/$31.00 ©2020 European Union 207

 In Fig. 4 are given S11 parameters of the simulated and
manufactured patch antenna. The figures are quite similar,
therefore regardless the handmade process imperfection, the
prototyped model is quite functional. The difference in the
bandwidth between the simulated and the fabricated antennas
is mainly due to imperfections of the air gap, constructed with
simple spacers. In Fig. 5 the measured data of the radiation
pattern of the patch for E and H plane is displayed.
The feed network includes implementation of Butler matrix
for realization of beamforming. The Butler matrix consists N
input with N output port. The design of Butler matrix
constructed by combined of hybrid coupler, phase shifter and
crossover [8]. The values of the typical Butler matrix are
presented in Table II.
TABLE II
VALUES OF BUTLER MATRIX

Patch element I II III IV

port I II III IV
electrical length 45o 90 o 135 o 190 o
extensions
electrical length 135 o 0o 225 o 90 o
extensions
electrical length 190 o 135 o 90 o 45 o
extensions
electrical length 90 o 225 o 0o 135 o
extensions

TABLE III
Fig. 4. S parameters of simulated and fabricated patches RELATION BETWEEN BEAM STEERING AND PATCHES PHASES

Beam Patch 1 Patch 2 Patch 3 Patch 4

sreeting
0o 0o 0o 0o 0o
-10o 0o 45o 90o 180o
+10 180o 90o 45o 0o

Below are described the scan angles of the beamforming.

The original goal is beam steering of +-10deg, therefore the
abovementioned Butler matrix can satisfy this requirement. In
Fig. 5. Radiation pattern of patch antenna (measured) Fig. 7 is shown realization of delay line, fooled Eq. (1)
, (1)
III. THE SCANNING RADAR ANTENNA where θ is the increased electrical length of microstrip line
and L is the length of the line without extension.
This method of designing patch antennas array can be
considered as an advanced version of the series feed antenna
array methods. It is important to keep distance [0.4λ ÷ 0.7λ]
between the centres of radiating elements. In Fig. 6 is shown
the simulated antenna array model. The feed networks for this
kind of antenna include microstrip lines, quarter wave
transformers and hybrid T-junctions [6-7].

Fig. 7. Delay line realization

In Fig. 8 a graphical relation between patches, antenna feed

Fig. 6. Simulated model of antenna array
network and phase shifter delay lines is shown. Building such
a system requires input power to be divided in four equal
parts. Beam steering is accomplished by switching between
microstrip lines with different lengths, different models RF
switches or PIN - to do this, it is necessary to use two diodes
connected in series to increase the coupling factors due to the
high frequency. This leads to required beam scanning (Table
IV) [9-12].

208
 Fig. 11. Radiation Pattern in scanning modes
Fig. 8. The corporate feed network model on H plane (Azimuth)

IV. SIMULATION RESULTS TABLE IV

COMPARISON OF GAINS ON BEAM DIFFERENT STATE

The patches and antenna array are simulation with Gain [dB]
openEMS (open source electromagnetic simulator). In Fig. 9
is displayed the return loss of the array elements and S11 on Patch array 0o 15.3
the input of antenna system. Fig. 10 shows azimuth and
Patch array -10o 14.95
elevation radiation pattern on normal mode. Fig. 11 shows the
beam in different scanning cases – using phases from Table Patch array + 10o 14.89
III and implementation of beamforming using 2-nd and 4-th
row of Butler matrix. Table IV presents a comparison of gains Patch array -24o 12.85
in beam`s different state. Patch array + 24o 12.55

From Fig. 11 and Table IV is evident that the level of side

lobes in 24O case is too high. Which would lead to incorrect
information in scanning mode. The advantage of using the 2-
nd and 4-th order of Butler matrix is bigger scanning angles
than using 1-st and 3-rd order of the same matrix or
implementation from Table III. The disadvantages in this case
are creation of grating lobes with high gain, which makes this
method inapplicable for radar applications like in the
presented case. With a more complex feed network and
employment of advanced scanning software, it is possible to
make a radar antenna with all four states presented in Fig. 11.
Antennas with similar radiation patterns may find application
Fig. 9. S11 parameters
MIMO [13] and 5G [14] systems where multi beams
propagation is a desirable t.

V. SAMPLE ALGORITHM FOR BEAM FORMING

CONTROL CODE AND CONTROL
OF AUTONOMOUS VEHICLES

As addition to the beam steering antenna array, authors

propose a simple algorithm for programmable controllers to
perform scanning for targets. Fig. 12 shows an algorithm
block diagram of the code for beam forming control. The
Fig. 10. Radiation Pattern 0 deg work mode
algorithm is fundamental in the movement of Autonomous
Vehicles when moving along preset coordinates without
human intervention. The main thing in the proposed algorithm
is that the scanning beam detects an object in front of it. When
detecting an object in front of one of the states of the radiation
diagram, the moving structure moves in the opposite direction
in order to avoid a collision.

209

Fig. 12. Algorithm block diagram of the code for radar scanning
process
VI. CONCLUSION
Simulated, designed and analyzed patch antenna integrated
into an electronic scanning antenna array presented in this
article. The used methodology can be used for design of a
patch antenna arrays for different radar applications. Designed
antennas are suitable for modifying and increase scopes of
some existing radar modules operating at that frequency. An
example of such radar is a HB100 radar module [15].
Advantage of design antenna in this paper under the stock
radar antenna is a beam scanning. The radar antenna array
disadvantages are the high cost, complicated feed network and
limited beam tilt. Modified radars with antenna array can be
implemented in different autonomous vehicles like small
robots, drones and various radar systems for civil and
specialized application.
210
ACKNOWLEDGEMENT
This work is supported by the National Scientific Program
"Young Scientists and Postdoctoral Students" (2018-2020)
and is approved by the Council of Ministers of the Republic of
Bulgaria № 577 / 17.08.2018 (Annex № 11).
REFERENCES
[1] S. Chatterjee, "A 77 GHz BCB Based High Performance
Antenna Array for Autonomous Vehicle Radars", 2018,
Electronic Theses and Dissertations,
https://scholar.uwindsor.ca/etd/7505
[2] S. Tonetti et al., "Distributed Control of Antenna Array with
Formation of UAVs", IFAC Proceedings Volumes, vol. 44, no.
1, 2011, pp. 7848-7853.
[3] Y. Kanayama et al., "A Stable Tracking Control Method for an
Autonomous Mobile Robot", IEEE International Conference on
Robotics and Automation Proceedings, 1990.
[4] B. A. Constantine, Antenna Theory: Analysis and Design, 3rd
Edition.
[5] Rogers RO4003 specification: https://www.mirrorsemi.com/pdf/
Rogers_RO4003C_Data_Sheet.pdf
[6] J. R. James, P. S. Hall, Handbook of Microstrip Antennas, 1993.
[7] M. Gachev, Microstrip Schemes.
[8] C.-C. Chang, R.-H. Lee, and T.-Y. Shih, "Design of a Beam
Switching/Steering Butler Matrix for Phased Array System",
IEEE Transactions on Antennas and Propagation, vol. 58, no. 2,
2009, pp. 367-374.
[9] I. Iliev, S. Kolev, and S. Manev, "Reduction of the Frequency
Offset Influence in OFDM using an Optimal PRS Coding and
MLSD", Internat. Journal of Research in Computer and
Communication Technology, vol. 3, no. 1.
[10] E. Carrasco, M. Barba, and J. A. Encinar, "X-band Reflectarray
Antenna with Switching-Beam Using PIN Diodes and Gathered
Elements", IEEE Transactions on Antennas and Propagation,
vol. 60, no. 12, 2012, pp. 5700-5708.
[11] A. Akhtar, H. M. Alahi, M. Sehnan, "Simulation of Phased
Arrays with Rectangular Microstrip Patches on Photonic Crystal
Substrates", 2012.10.12
[12] J. Ehmouda, Z. Briqech, and A. Amer, "Steered Microstrip
Phased Array Antennas", World Academy of Science,
Engineering and Technology, vol. 49, 2009, pp. 321-323.
[13] I. Iliev, I. Georgiev, and M. Budzevski. "Concatenated “MMSE-
Sequential Search”, Algorithm for Multi User Detection in
SDMA Uplink", ICEST 2011, vol. 3, 2011, pp. 591-594.
[14] M. S. Sharawi, M. Ikram, and A. Shamim, "A Two Concentric
Slot Loop Based Connected Array MIMO Antenna System for
4G/5G Terminals", IEEE Transactions on Antennas and
Propagation, vol. 65, no. 12, 2017, pp. 6679-6686.
[15] HB 100 radar module datasheet
https://www.limpkin.fr/public/HB100/HB100_Microwave_Sens
or_Module_Datasheet.pdf
 Design of well-matched Microwave Slot Antenna on a
Flat Metal Grounded Plate using Neural Model
Zoran Stanković1, Nebojša Dončov1, Biljana Stošić1, Maja Sarevska2,
Ivan Milovanović3
Abstract – A neural model of microwave slot antenna on a flat
metal grounded plate based on MultiLayer Perceptron network
(MLP) is described here. The presented model is a part of the
software package called PMWA ANN Design which is developed
for design of a well-matched planar microwave antenna. This
software allows for fast estimation of the flat metal grounded
plate and slot dimensions for desired central frequency of the
microwave slot antenna to make the antenna well-matched to the
feed line. The current version of this software allows for design
of antenna matched to 50 Ω feed line in S and C frequency
bands.
Keywords – Slot antenna, Neural networks, Neural model,
Antenna optimization.
I. INTRODUCTION
Thanks to being small, relatively simple and cheap for
production, with compact architecture that is adjustable on the
housing of small communication devices, small
electromagnetic coupling with human body and ability to
work in multiple frequency bands, printed antennas are more
and more used in modern communications [1-2].
Classical approach in printed antenna design is use of
detailed electromagnetic (EM) model gained from EM
simulators based on one of the following methods: MoM
(Method of Moments) [3-4], FEM (Finite Element Method)
[3], SSPE (Split-Step Parabolic Equation) [4], FDTD (Finite
Difference Time-Domain) [3-4] and TLM (Transmission Line
Matrix) [3,5]. These methods are based on numerical solving
of Maxwell’s equations in frequency or time domain. Because
of that EM simulators require strong hardware platform and
their simulations may take very long time depending on
antenna geometry complexity and available hardware
resources. These disadvantages of EM simulators may
seriously slow down the antenna development especially in
the case when in the design process there are very large
number of changes of physical parameter values and
1Zoran Stanković, Nebojša Dončov and Biljana Stošić are with
University of Niš, Faculty of Electronic Engineering, Aleksandra
Medvedeva 14, 18000 Niš, Serbia, E-
mails:zoran.stankovic@elfak.ni.ac.rs, nebojsa.doncov@elfak.ni.ac.rs,
biljana.stosic@elfak.ni.ac.rs
2Maja Sarevska is with AUE-FON University Skopje, Str. Kiro
Gligorov b.b. 1000 Skopje, North Macedonia, E-mail:
maja.sarevska@fon.edu.mk
3Ivan Milovanović is with Singidunum University, Educational
Center Niš, Danijelova 32, 11000 Belgrade, Serbia, E-mail:
imilovanovic@singidunum.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
observation of these changes influence on EM antenna
characteristics. An alternative approach in design of printed
antenna represents the usage of ANN (Artificial Neural
Network) based models [6-7]. Antenna neural model can
perform much faster than the model developed in some of EM
simulators, but at the same time it has accuracy that is close to
that of EM simulators which motivates increased usage of
ANN for realization of the efficient models of printed
antennas [8-12]. This is a result of very fast propagation of
signal through neural network from its input to output, as well
as fast signal processing in the network’s neurons.
In this paper, a neural model of slot antenna on a flat metal
grounded plate, as well as software for estimation of this
antenna dimensions for working frequency of well-matched
antenna to feed line are presented. This type of printed
antenna possess all previously presented properties of printed
antennas and for that reason large attention is paid for its
usage in RFID systems [13-15].
This paper represents modification and upgraded results of
the software for neural modeling of the planar bowtie antenna
presented in [16]. In that software, a neural model of slot
antenna presented here is added with a goal to further develop
software tool for neural modeling of different types of printed
antennas in microwave band.
II. MICROWAVE SLOT ANTENNA: EM MODEL
The architecture of rectangle slot antenna that is realized on
a flat rectangle metal plate is presented in Fig. 1.The basic
geomtrical parameters that have an impact on the
electromagnetic properties of the antenna are length of metal
plate (l), width of the metal plate (w), slot length (ls), slot
width (ws) and displacement of feeding point relative to the
geometrical slot center (f). In the case when the antenna is
realized on a plate with large dimensions (l, w >> ls, ws) then
by using the Babinet's principle [1] the input impedance of
slot antenna is determined by the relation
η2 , (1)
zS =
4zD
W here η represents an intrinsic impedance of free space, and
zD an input impedance of complementary planar dipole in dual
electromagnetic system (dual dipole) (Fig. 2). In similar way,
using the Babinet's principle and the dual electromagnetic
system, the components of electrical and magnetic fields of
the slot antenna in a far field are determined if we know the
211

Fig. 1. Architecture of the microwave slot antenna on a flat metal
grounded plate
Fig. 2. Slotantenna analysis using a dualplanar dipole based
onBabinet's principle
components of electrical and magnetic fields of dual planar
dipole
EθS = H θD
EϕS = H ϕD
E (2)
H θS = − θ2D
η
H ϕS = − 2
EϕD
η
However in the practical cases of the slot antenna usage where
it is built in the compact housing of a communication device,
the antenna should be of small dimensions. Also the metal
plate used for the slot antenna realization has reduced
dimensions that are comparable with slot antenna dimensions.
In this case, the analyses of input impedance and
electromagnetic radiation of the slot antenna are much more
complex and Eqs. (1) and (2) cannot be applied directly, that
is the reason why we use EM simulators. In our case, the EM
model of slot antenna is realized by use of functions given in
Antenna Toolbox of MATLAB [17-18]. Antenna Toolbox
uses MoM to calculate antenna properties (like input
impedance), surface properties (like current and charge
distribution), and field properties (like near-field and far-field
radiation pattern) [17-18]. Here, a case of slot realized on
square plate with dimensions a = l = w, where a is 50% longer
than the slot length (a = 1.5⋅ls), is observed. Also, the standard
antenna excitation where the feeding point is on half-length of
slot length (f = 0) is considered. Based on this the variable
212
parameters of EM model are slot length (ls) and slot width
(ws). EM model of the realized slot antenna is used for
generation of training and testing samples for the neural
model and also for presentation of EM characteristics of the
antenna for values of geometrical parameters that are derived
with optimization using the neural model.
III. MICROWAVE SLOT ANTENNA:NEURAL MODEL
Architecture of the neural model of the microwave slot
antenna on a flat metal grounded plate (MLP_SL) that is
developed using the Multilayer Perceptron (MLP) network [6-
7] is depicted in Fig. 3. Unlike the EM model, instead of the
slot width, the neural model takes the relative ratio of the slot
width and slot length represented as k = ws/ls. The model has
the main purpose to map the space of physical parameters of
the slot antenna (slot length ls and relative ratio of the slot
width and slot length k) to a space of the antenna operating
characteristics,such as the resonant frequency fr and the
minimum s11 parameter value (value of s11 parameters at
antenna resonant frequency - s11min), at the given feed line
impedance (zf =const)
[ f r , s11 min ] = f MLP (l s , k ) . (3)
In matrix representation, the neural model has the vectors of
input variables x=[ls k]T, and output variables y= [fr s11min].
For the model realization, MLP neural network, described by
function y=y(x,W,B), is used, where y is a network processing
function, W = {w1, w2,…wH+1} is a set of connection
weighting matrices and B = {b1, b2,…bH+1} is a set of bias
matrices (H is the total number of hidden layers in the
network). Based on these definitions, the MLP_SL neural
model is described as:
[ fr s11min ] = y([l s k ]T ,W , B) . (4)
Output of MLP layers can be described by function:
y l = F ( w l y l −1 + b l ) l = 1, 2,.., H + 1 , (5)
where the vector yl-1 represents the output of (l-1)th layer, wl
represents a connection weight matrix between (l-1)th and lth
layer neurons, and bl is a vector containing biases of lth layer
neurons. The activation function of neurons in hidden layers is
the hyperbolic tangent sigmoid transfer function:
e u − e −u . (6)
F (u ) =
eu + e − u
The activation function of neurons in the output layer is linear
which indicates the MLP network output in the following
form:
y = w H +1y H + b H +1 , (7)
where wH+1 is the connection weighing matrix between
neurons in Hth hidden and output layers.
 Fig. 3. The neural model of the microwave slot antenna on a flat
metal grounded plate

The MLP model described above is notated as

MLPH-N1-…-Ni-…-NH where H is the total number of hidden
layers, and Ni is the number of neurons in the ith hidden layer.
For example, notation MLP2-12-10 denotes the MLP model
whose neural network has 2 hidden layers with 12 and 10
neurons in layers, respectively. In total, the neural network
has 4 layers (input, output and two hidden layers).
MATLAB environment is used for implementation of
MLP_SL neural model. Training and test sample sets for
training and testing the network are generated by EM model
Fig. 4. Scattering diagrams of MLP2-13-13 neural model on the test
of the microwave slot antenna presented in Section II. These
set for frand s11min outputs
data sets are given in one of the following formats {(xt,yt)} or
{([lstkt]T,[frt st11min])}, with xt representing the vector of input
trained. During the network testing, average test error (ATE)
combinations of the variables lst and kt, and yt the vector
for both network outputs and correlation coefficient (rPPM) so-
containing desired outputs of the neural network frtandst11min
called testing metrics are observed in order to discover a
for given input. The superscript t indicates that these samples
neural network with good generalization properties. This leads
belong to training and testing sets. The input parameter ranges
to highly accurate neural model for determination of the
defined as l [m]∈[0.025, 0.1] and k∈[0.05, 0.12] are used to
resonant frequency and the minimum s11 value. Table I
train neural network. The chosen parameter ranges provide the
contains the results for several MLP_SL neural networks
resonant frequency in the range frt∈[2.16, 9.12] GHz covering showing the best test statistics.
the S and C microwave bands.
By using microwave slot antenna EM model for the case of TABLE I
impedance zf = 50 Ω and the uniform distribution of samples, TESTING RESULTS OF MLP_SL NETWORKS WITH
THE BEST TEST STATISTICS
the following training and testing sets are generated:
MLP_SL network ATE(fr) [%] ATE(s11min) [%] rPPM
([l st k t ]T , [ f rt , s11
t
min ]) |
 (8) MLP2-13-13 1.50 6.30 0.9438
t  MLP2-12-5 1.48 6.44 0.9435
l s ∈ [0.025 : s1 : 0.1], k t ∈ [0.05 : s2 : 0.12] MLP2-16-11 1.41 6.54 0.9434
MLP2-9-7 1.41 6.26 0.9424
where the parameter values s1 = 0.005 and s2 = 0.01 are used MLP2-11-4 1.57 7.22 0.9422
to generate the training set, as well as values s1 = 0.007 and MLP2-15-11 1.35 5.97 0.9418
s2 = 0.013 are used for the test set. As a result, the training and
test sets with 127 and 66 samples, respectively, are obtained. For the realization of the MLP_SL neural model, the
During the model training, Levenberg–Marquardt training MLP2-13-13 neural network, that has the highest value of
method, with a given maximum root mean square error on test correlation coefficient (rPPM), is chosen. The scattering
samples of 10-4,is used. In order to define the best possible diagrams of MLP2-13-13 neural model on the test set for fr
MLP_SL model, a number of MLP neural networks with two and s11min outputs are shown in Fig. 4. Similarly like in the
hidden layers and different number of neurons in them, is model presented in [16] the output of neural model fr has very

213
small scattering relatively to the referent values. In the are: Δfr = 5 MHz, lsmin = 2.5 cm, lsmax= 10 cm, dl = 0.1 mm,
scattering diagram for s11min output it can be noticed an area of kmin =0.05, kmax= 0.12, dk=0.001.
moderate scattering (SA1) with s11min ≥ -20 dB and an area of
larger scattering (SA2) with s11min < -20 dB where the neural
model shows smaller accuracy.

IV. SOFTWARE PMWA ANN DESIGN

Software PMWA ANN Design represents modification and
expansion of BT ANN Design ZM software presented in [16].
Part of this software that relates to slot antenna design consists
of slot antenna EM model represented in Section II and
MLP_SL model described in Section III. In the first step of
this tool MLP_SL model is used in order to determine
geometrical parameters (antenna length and width) for desired
antenna resonant frequency that will give minimal value of
s11parameter, or will provide best matching to the chosen feed
line. In the next step, the EM model of slot antenna allows to
estimate the dependence of the s11 parameter versus the
frequency, bandwidth, gain, and radiation pattern in order to
(a)
check whether the proposed antenna architecture meets the
design specifications.
The algorithm presented in the first step is taken from [16]
and is:
1. Combinations of the input variables defined with the set
X = {(ls, k), ls= lsmin:dl:lsmax, k= k min:dk:k max} are applied at
the input of MLP_SL neural model. Then for these
combinations the output sets YSA1 = {(fr, s11min)} for
s11min≥ -20 dB and YSA2 = {(fr, s11min)} for s11min< -20 dB are
generated. In this notation lsmin, dl, lsmax and kmin, dk, kmax
present minimal value, step and maximal value for slot
length (l) and slot width/length ratio (k),respectively.
2. In the output set YSA1, the couple-winner (fr(w1),s11min(w1))
whose value for s11min is minimal in all set Y is found, while
the resonant frequency fr(w1) meets the requirement
|fr(t) - fr(w1)| ≤Δfr, where fr(t) is desired resonant frequency and
Δfr is largest permitted deviation from it.
(b)
3. In the output set YSA2, the couple-winner (fr(w2), s11min(w2))
whose value for s11min is closest to the mean value to that
parameter in the set YSA2 is found, while the resonant
frequency fr(w2) meets the requirement |fr(t) - fr(w2)| ≤ Δfr. If
YSA2 is empty set then the coupler-winner is defined as
(fr(w2), s11min(w2)) = (0,0).
4. If s11min(w2)<s11min(w1) is limited, the couple-winner is
(fr(w), s11min(w)) = (fr(w2), s11min(w2)), in contrary it is
(fr(w), s11min(w)) = (fr(w1), s11min(w1)). For optimal values of l
and k the values that match couple-winner (fr(w), s11min(w)) (c) (d)
are gotten. Fig. 5. PMWA ANN Design outputs for 2.45 GHz slot antenna:
(a) Optimized ls, ws and a parameters, (b) s11 parameter vs.
V. EXAMPLESOF ANTENNA DESIGN USING frequency, resonant frequency, antenna bandwidth, (c) directivity in
the azimuthal plane (-4.84 dBi) and (d) in elevation plane (6.37 dBi)
SOFTWARE PMWA ANN DESIGN

Different slot antennas are designed by use of PMWA ANN

Design software: one for RFID application in 2.45 GHz ISM
band and one that may be used in ultra-wideband real-time
location systems (UWB-RTLS) in band [7-9] GHz with
8 GHz central frequency (Figs. 5 and 6). The parameter values

214

(a)
(b)
(c) (d)
Fig. 6.PMWA ANN Designoutputs for 8 GHz slot antenna:
(a) Optimized ls, wsandaparameters, (b) s11 parameter vs. frequency,
resonant frequency, antenna bandwidth, (c) directivity in the
azimuthal plane (-4.33 dBi) and (d) in elevation plane (5.81 dBi)
VI. CONCLUSION
Using neural models, which avoid complex numerical
Maxwell’scalculations of equations, the PMWA ANN Design
software facilitates and speeds up the design of microwave
antennas. Further development of this software will be aimed
at expanding the library of its models with new types of
antennas.
215
ACKNOWLEDGEMENT
The paper is supported by the Ministry of Education,
Science and Technological Development of Serbia.
REFERENCES
[1] C. A. Balanis, Antenna Theory: Analysis and Design, Wiley,
New York, 2005.
[2] R. Waterhouse, Printed Antennas for Wireless Communications,
John Wiley & Sons, Ltd, 2007.
[3] M. N.O. Sadiku, Numerical Techniques in Electromagnetics,
2nd edition, CRC Press LLC, 2001.
[4] L. Sevgi, Electromagnetic Modeling and Simulation, First
Edition, The Institute of Electrical and Electronics Engineers,
John Wiley & Sons, Inc., 2014.
[5] N. Doncov, B. Milovanovic, and Z. Stankovic, ”Extension of
Compact TLM Air-vent Model on Rectangular and Hexagonal
Apertures”, Applied Computational Electromagnetics Society
(ACES) Journal, vol. 26, no. 1, pp.64-72, 2011.
[6] S. Haykin, Neural Networks, New York, IEEE, 1994.
[7] Q. J. Zhang, K. C. Gupta, Neural Networks for RF and
Microwave, Artech House, 2000.
[8] M. Milijić, Z. Stanković and B. Milovanović, “Efficient Model
of Slotted Patch Antenna based on Neural Networks”,2009 9th
International Conf. on Telecommunication in Modern Satellite,
Cable, and Broadcasting Services, Niš, 2009, pp. 384-387.
[9] M. Pandit, T. Bose, “Application of Neural Network Model for
Designing Circular Monopole Antenna”, International
Symposium on Devices MEMS, Intelligent Systems &
Communication (ISDMISC) - Proceedings published by
International Journal of Computer Applications (IJCA), April
2011, Gangtok, Sikkim, India, pp. 18-21.
[10] M. Chetioui, A. Boudkhil, N. Benabdallah and N. Benahmed,
“Design and Optimization of SIW Patch Antenna for Ku Band
Applications using ANN Algorithms”,2018 4th International
Conference on Optimization and Applications
(ICOA),Mohammedia, 2018, pp. 1-4.
[11] U. Ozkayaa, L.Seyfia, “Dimension Optimization of Microstrip
Patch Antennain X/Ku Band via Artificial Neural Network”,
Procedia - Social and Behavioral Sciences, 2015, vol. 195, pp.
2520-2526.
[12] R. Ghayoula, A. Gharsallah, N. Fadlallah and M. Rammal,
“Artificial Neural Networks ANN Approach for the Synthesis of
Patch Antennas”, 2007 IEEE Radar Conference, Boston, MA,
2007, pp. 285-290,DOI: 10.1109/RADAR.2007.374229
[13] Y. Zhou, “A Novel Slot Antenna for UHF RFID Tag”, IET 3rd
International Conference on Wireless, Mobile and Multimedia
Networks (ICWMNN 2010), Beijing, 2010, pp. 254-257, DOI:
10.1049/cp.2010.0664
[14] M. Ramírez, J. Parrón, “Slot Antenna for RFID Readers”, 2013
7th European Conference on Antennas and Propagation
(EuCAP), Gothenburg, 2013, pp. 1726-1729.
[15] W. Abdelrahim, Q. Feng, ”A Compact Wideband Slot Antenna
for Universal UHF RFID Reader”, Progress In Electromagnetics
Research Letters, vol. 73, pp. 137-144, 2018.
[16] M. Sarevska, Z. Stankovic, N. Doncov, I. Milovanovic and B.
Milovanovic, “Design of well-matched UHF Planar Bowtie
Dipole Antenna using Neural Model”, 2019 14th International
Conference on Advanced Technologies, Systems and Services
in Telecommunications (TELSIKS), Nis, Serbia, 2019, pp. 331-
334, doi: 10.1109/TELSIKS46999.2019.9002319
[17] https://www.mathworks.com
[18] http://ece.wpi.edu/mom/
 An Automatic System for Antenna Radiation Pattern
Measurement
Ilia Iliev1 and Ivaylo Nachev2
Abstract – In this paper is presented automatic antenna
pattern measurement system. The system provides the
opportunity to measure different types of antennas in different
frequency ranges. The measured pattern can be presented in
Cartesian, polar or 3D spherical coordinate system.
Keywords – Antenna measurement, 3D radiation pattern,
Automatic system, Measurement algorithm, DSS BPSK.
I. INTRODUCTION
With the developing of the wireless technologies, it is
required widespread use of antennas and, at the same time,
inexpensive and efficient methods for measuring their
parameters. The market offers various automatic measurement
systems - from fully professional to hobby applications. The
professional systems are usually part of a measurement
equipment of an anechoic chamber [1-4] and they are very
expensive. The semi-professionals systems are cheaper but
with limited features and performance [5-6], i.e. work
frequency, weight of the measured antenna, only 2D pattern
diagram and etc.
In this work an automatic system for antenna radiation
pattern measurement is demonstrated. The main features are:
low cost, easy to assemble with wide availability components,
possibility to measure 3D radiation pattern diagram with
sufficient accuracy. This article presents different segments
and applications of the proposed system. The radiation
diagram measurements are performed with a small angle step
using a method of the both antennas. The measured antenna is
mounted on 3D rotated stand. An algorithm has been
developed to synchronize the rotation control of the stand and
measurement process, which further increases the accuracy of
the measurement.
II. AUTOMATIC MEASUREMENT SYSTEM
The designed system automatically measures antenna’s 3D
radiation pattern. The measurement is based on the method of
both antennas. The receiving and transmitting parts of the
staging are made with innovative devices and modules used in
radio communication equipment. Fig. 1 shows block diagram
1Ilia Iliev, Professor at the Faculty of Telecommunications at the
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: igiliev@tu-sofia.com
2Ivaylo Nachev is a PhD student at the Faculty of
Telecommunications at the Technical University of Sofia, 8 Kl.
Ohridski Blvd, Sofia 1000, Bulgaria, E-mail:
ivaylonachev@yahoo.com
978-1-7281-7142-5/20/$31.00 ©2020 European Union
of a classical set. The distance between both antennas must be
L > (2D)/λ, which is larger than near field antenna areas for
the operation frequency [7-9].
Fig. 1. Block diagram of the both antennas RP measurement
method
The block diagram of the automatic system transmitting
part is shown in Fig. 2. The transmitting part in this work
consists SDR Adalm Pluto [10] and it is possible to use
another type of transmitter. This SDR can emit radio signal in
frequency range between 325MHz÷3800MHz. The
transmitted signal bandwidth is wide enough to perform the
measurement and can be achieved by an appropriate
modulation method. Here Direct Spread Spectrum technique
and BPSK is applied. The transmission of a suitable signal is
carried out with specialized software included in the Adalm
Pluto package. For optimal system performance the SDRs
software is loaded on USB flash drive and the power supply
can be accomplished by using a power bank. The SDR
transmitter is connected to an antenna with predefined
parameters.
ANTTX
SDR transmitter DSS, BPSK
Power
supply
Program in USB flash drive
Fig. 2. Block diagram of an automatic system transmitting part
The receiving part of the measurement settings is more
complex – Fig. 3. In the design production, this is the
measuring side of the antenna with unknown parameters. The
blocks are:
• Band Pass Filter – for filtering the received signal from
measured antenna. This is in order to mitigate the
interferences and to increase the sensitivity of the
receiver. The band pass must be approximately equal to
216
 the bandwidth of the transmitting signal. For the and VADC is ESP32 ADC measured value, Pdbm is calculated
multifunctionality of the system, there has to be multiple power in [dBm]. Fig. 4 shows the transfer function in
filters for different frequency bands; graphical manner.
• RF low noise amplifier – increases the sensitivity of the
measurement. This is necessary when measured antennas
are with low gain. However, this part can be neglected;
• AD8318 RF logarithmic detector [11] measures the
power of the received signal from the antenna. The
module must be used in a shielded box in order to
eliminate side effects. The RF logarithmic detector and

Pout [dBm]
transmitter frequency ranges determine the operational
frequencies of the automatic measurement system;
• Lora WiFi (ver2) [12]. The build-in ESP32 is used for
the analog to digital conversion (ADC) of the signal
from AD8318 RF logarithmic detector and data post
processing. This module exchanges data with the PC via
USB interface;
• PC with software for synchronization between the
measurement block and the mechanical rotating units, as
well as for data processing. The software processes the
measured results in an appropriate format;
• Arduino uno – controls the stepper motors. The module Fig. 4. AD8318 RF logarithmic detector transfer function
communicates with PC via USB port and uses special
written driver based on CNC protocol;
Begin
Nema 17 Stepper
Motor Y axis
Arduino A4988 Stepper (elevation)
Stand
UNO Motor Driver Setup of the initial variables:
Nema 17 Stepper
Motor X axis
- scan step and time delay;
USB
(azimuth) - min and max Azimuth angles AZ_scan;
Power - min and max Elevation angles EL_scan;
PC ANTrx
supply
USB

AD8318 RF
Setting the initial position of the motors
Lora RF LNA
Logarithmic BPF
WIFI(V2) Module
Detector
Measurement:
for angle az in (AZ_scan);
Fig. 3. Block diagram of Rx part of the automatic system for angle el in (EL_scan);
set motors position on (az,el);
To simplify the receiving part, instead of the logarithmic Power measurement Pdbm(az,el);
detector and ESP32, a SDR (with the appropriate software) in Time delay;
reception mode can be used. end { for angle el};
In this work the measured receiving power is obtained and end { for angle az};
calculated by approximated pre-determined AD8318 RF
logarithmic detector transfer function. A signal with a certain
voltage depends on the input power is received from the Save data, Plot pattern diagrams,
logarithmic detector. It is measured by a 12-bit ADC of Go to motors to start position
ESP32. For practical purposes and ESP32 RAM limits the
logarithmic detector transmission function can be determined End
by measurement and after an approximation process can be
included in the software. In Eq. (1) is given such polynomial Fig. 5. Block diagram of automatic system software algorithm
approximated characteristic:
The TX RF signal has a constant envelope and this allows
= −38.3 + 256 − 669 + (1) the ADC measurement rate to be 6000 samples per second
+849 − 562 + 150, (sps). One measured value is obtained after averaging 3000
ADC samples. The speed at which a measurement is
where = /1000, performed is 2 values per second.

217
 To minimize inaccuracy, the measured values are
processed with a 5th order low pass linear interpolation FIR
filter and it is programmed in ESP32. The filter sample
frequency is also 2 sps.
The measuring unit is placed on a stand that allows
rotation in X and Y planes. The rotations during the
measurement are performed with the help of stepper motors
NEMA 17.
The synchronization between the rotation of the motors
and the measurement process is done by PC software. A block
diagram of automatic system software algorithm is presented
in Fig. 5. The angles and steps for azimuth and elevation are
set in advance. In the measurement process, all angles values
are setup sequentially. After rotating to given position, the
system measures the received power. In this way, the Fig. 7. S11 parameters
synchronization between measurement and rotation is
achieved.
IV. MEASURED RESULTS
The angular velocity of movement depends on the
measuring speed and the angle step. The stepper motor control
Measured patterns are presented in different coordinate
algorithm provides the ability to limit the speed and
systems. For comparison, the Figs. 8 and 9 give measured and
acceleration of rotation. This is necessary so as not to destroy
simulated radiation pattern of the patch antenna in polar and
the mechanical part or mechanically damage the measured
Cartesian coordinate system. Fig. 10 shows radiation pattern
antenna.
in 3D spherical coordinate system. The transmitting antenna is
The measured information can be presented in Cartesian,
TP-Link ANT2409A [15].
polar or 3D spherical coordinate system.

III. TEST PATCH ANTENNA

To test the functionality of the proposed measurement
system (Fig. 6b) a simple patch antenna is designed for
operating frequency =2.4 GHz. The radiated element of the
designed patch has circular form and a subtended substrate
method has been chosen [13]. For additional simplification the
substrate has =1 (dielectric constant of air) and height
h=5 mm - this simplifies the fabrication of the antenna. The
diameter of the antenna in this case [14] is:

= . (2)

The fabricated antenna is presented in Fig. 6a.

The reflection S11 parameter is simulated with openEMS
software. The simulated and the measured results are shown
in Fig. 7.
Fig. 8. Measured RP in polar coordinate system

From the results can be concluded that the proposed

measuring system has the necessary efficiency and accuracy -
measured results are similar to the simulated ones. The stand’s
metal construction and the influence of the reflected waves are
a major disadvantage in the automatic measurement system.
This is one of the reasons for the difference between the
simulated and measured results. To eliminate reflected waves,
the stand can be fabricated of plastic or wave-absorbent
materials.

Fig. 6. (a) Tested patch Fig. 6 (b) Automatic

antenna measurement systems

218

Fig. 9. Measured RP in Cartesian coordinate system
Another disadvantage is the incorrect positioning of Rx part
and the possibility of entering an incorrect value from the
measurement. In the software data processing with FIR low
pass filter is used for correcting this shortcoming.
The autonomic radiation pattern measurement system has
advantages over a classical method of the both antennas. This
system has the possibility to present radiation pattern in 3D
spherical coordinate system. The error from the rotation of the
antenna is equal to the motor’s step error.
Fig. 10. Measured RP in 3D spherical coordinate system
V. CONCLUSION
The discussed system is relatively cheap and easy to
assemble, with accuracy close to professional measurement
systems. For transmitting and receiving blocks described in
the article can be used SDR integrated into anechoic chamber.
The components of the proposed system can be applied to
219
other radio communication measurement equipments. The
mechanical stand with a control device, stepper motors and
receiver measurement part is also suitable for a radio location
or navigation applications.
ACKNOWLEDGEMENT
This work is supported by the National Scientific Program
"Young Scientists and Postdoctoral Students" (2018-2020)
and is approved by the Council of Ministers of the Republic of
Bulgaria № 577 / 17.08.2018 (Annex № 11).
REFERENCES
[1] Antenna Tripods, Masts and positIoners - EUTmanipulators –
Horn positioners (c2001-2016),
http://www.ramayes.com/tripods_masts_positioners.htm
[2] Sunol Sciences Corporate Website (2015),
http://www.sunolsciences.com
[3] 2120cr multi-axis positioner (maps) all info (2014),
http://www.ets-lindgren.com/all/2120CR
[4] 8-Axis Robotic Antenna Measurement System,
https://www.nsi-mi.com
[5] Wave and Antenna Training Kit WATS-2002 – manual
available on http://www.manntel.com/edue/p2_2.html
[6] J. T. Foley, V. Omelianov, S. Koziel, and A. Bekasiewicz,
“Low-cost Antenna Positioning System Designed with
Axiomatic Design“, MATEC Web of Conferences, vol. 127,
article number 01015, 2017, The 11th International Conference
on Axiomatic Design (ICAD 2017), DOI:
10.1051/matecconf/201712701015
[7] R. C. Johnson, H. Jasik, Antenna Engineering Handbook, New
York, McGraw-Hill Book Company, 1984.
[8] T. Fukasawa et al., “Extended S-parameter Method Including
Radiation Pattern Measurements of an Antenna”, IEEE
Transactions on Antennas and Propagation, 2012, vol. 60, no.
12, pp. 5645-5653.
[9] Z. Živković, A. Šarolić, Measurements of Antenna Parameters
in GTEM Cell, 2010.
[10] Adalm Pluto datasheet available on
https://www.analog.com/media/en/news-marketing-
collateral/product-highlight/ADALM-PLUTO-Product-
Highlight.pdf
[11] AD8318 RF Logarithmic Detector datasheet
https://www.analog.com/media/en/technical-
documentation/data-sheets/AD8318.pdf
[12] Lora WiFi v2 datasheet available on
https://heltec.org/project/wifi-lora-32/
[13] N.-W. Liu, L. Zhu, W.-W. Choi, “Low-Profile Wide-
Beamwidth Circularly-Polarised Patch Antenna on a Suspended
Substrate”, IET Microwaves, Antennas & Propagation, 2016,
vol. 10, no. 8, pp. 885-890.
[14] C. A. Balanis, Antenna Theory: Analysis and Design, John
Wiley & Sons, 2016.
[15] TP-Link ANT2409A datasheet available on https://images-
eu.ssl-images-amazon.com/images/I/9196zJReTVS.pdf
 Inverse Reading of ITU-R P.1546 Propagation Curves for
the Broadcast Transmitter EM Field Level
Milos Radojković1, Olivera Pronić-Rančić2 and Zlatica Marinković2
Abstract – In this paper, a method for inverse reading of
curves for the expected level of electromagnetic field originating
from broadcast transmitters, given in Recommendation ITU-R
P.1546, is proposed. The proposed method is based on the
application of artificial neural networks and enables efficient
determination of distances for given field levels for the
frequencies up to 4000 MHz. Accuracy of the method is
illustrated by an example of curves related to the frequency
3050 MHz and several values of height h1 not used for the neural
model development.
Keywords – Broadcast transmitter, Electromagnetic field level,
ITU-R P.1546.
I. INTRODUCTION
Broadcast communications are point-to-area
communications and their application is very common
nowadays (broadcasting of radio and television signals,
broadcasting of signals from mobile communication base
stations, etc.). One of the most important steps in design a
broadcasting system is determining the coverage area. The
coverage area represents the zone in which the offered service
has the appropriate signal strength, necessary for the provision
of quality services and is determined by the maximum
distances from the transmitter for the given electromagnetic
(EM) field strength level.
The coverage area is determined in several directions in
azimuth (usually, 36 directions i.e., every 10° or 120
directions, i.e., every 3°, depending on the transmitted power).
There are different methods for determining the maximum
distance at which the appropriate field level is achieved, but
the statistical method defined in ITU-R P.1546
recommendation is the most frequently used method [1].
According to the recommendation, the determination of the
EM field level is based on interpolations of the values given
by the empirically derived propagation curves as the function
of distance, transmitting antenna height, h1, frequency and
percentage of time. As such, Recommendation ITU-R P.1546
requires a large number of interpolation calculations for
transmitter coverage range determination.
The aim of this paper is to present a method for inverse
reading of propagation curves from ITU-R P.1546
recommendation using artificial neural networks (ANNs), as
1Miloš Radojković is with Energoprojekt Entel, Bulevar Mihajla
Pupina 12, Belgrade 11070, Serbia, E-mail:
mradojkovic@ep-entel.com
2Zlatica Marinković and Olivera Pronić-Rančić are with the
University of Niš, Faculty of Electronic Engineering, Aleksandra
Medvedeva 14, 18000 Niš, Serbia, E-mails:
zlatica.marinkovic@elfak.ni.ac.rs, olivera.pronic@elfak.ni.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
an extension of the method described in [2]. Namely, the
ANNs have ability to learn the relationship between the input
and output data and the learned relationship can be
subsequently used to generate an accurate response for the
new input values. ANNs have been widely used in the field of
RF and microwaves [3]-[10]. They have been applied for the
automated reading of the empirical propagation curves from
the ITU recommendation, [5], [7]. As mentioned, in [2] the
ANNs were applied for the inverse reading of the propagation
curves for the frequency of 600 MHz. In this paper, the
approach is extended to be valid for the entire frequency range
from 100 to 4000 MHz as defined in Recommendation ITU-R
P.1546.
The paper is organized as follows: in Section II a brief
introduction to ITU-R P.1546 Recommendation is given. The
proposed ANN method is explained in Section III. Section IV
contains the most illustrative results generated by using the
developed ANN model for inverse reading of the ITU-R
P.1546 EM field level curves. Finally, the concluding remarks
are given is Section V.
II. ITU-R P.1546 RECOMMENDATION
ITU-R P.1546 Recommendation is based on a family of
curves representing the field strength radiated by a transmitter
having 1kW effective radiated power (e.r.p). The curves are
given for frequencies of 100, 600 and 2000 MHz for
transmitter antenna heights (h1) of 10, 20, 37.5, 75, 150, 300,
600 and 1200 m. The curves are given for the propagation of
radio waves over land for 50%, 10% and 1% of the time and
50% of the locations in the area 500 m by 500 m as well as for
50%, 10% and 1% of the times for 50% of the locations for
radio propagation waves above cold and warm sea surfaces.
Fig. 1 shows the expected field level for a radio wave of
2000 MHz frequency propagated over land, for 50% of the
time. According to ITU-R P.1546 Recommendation, each
curve represents a signal strength in dB(µV/m) for distances
up to 1000 km and discrete values of height h1.
For the propagation of radio waves over the sea, the height
h1 represents the height of the antenna above sea level. For the
propagation of radio waves over land, h1 depends on the
available information about terrain when propagation
predictions are being made.
In cases when detailed information about terrain is not
available, h1 is equal to the antenna height if distance d is less
or equal of 3 km. If d is greater than 3 km and less than
15 km, h1 is calculated according to below equation:
h1 = ha + (heff − ha )(d − 3) / 12 . (1)
220

Fig. 1. Curves for expected field strength level at 2000 MHz for
propagation over land according to ITU-R P.1546 Rec. [1]
where ha is antenna height above the ground level and heff is
antenna height over the average level of the ground between
distances of 3 to 15 km.
In cases when detailed information about terrain is
available, h1 is equal to the antenna height over the average
level of the ground between distances of 0.2d to d. If d is
greater than 15 km, h1 is equal to the heff.
The curves in Recommendation ITU-R P.1546 are given for
discrete values of frequency, parameter h1 and distance d.
Accordingly, if at least one of the above three parameters
differs from the values in the tables of ITU-R P.1546
Recommendation, it is necessary to perform interpolations for
the specific required values.
If interpolation is performed for h1 according to ITU-R
Recommendation P.1546, the following equation is used:
log(h1 / hinf )
Eo = Einf + ( E sup − Einf ) , (2)
log(hsup /hinf )
where Einf is the signal level for the required distance read
from the curve of the nearest height less than h1 (hinf) while
Esup is level of the field for the required distance read from the
curve of the nearest height greater than h1 (hsup). If h1 is
greater than 1200 m the hinf is 600 m, while hsup is 1200 m.
If interpolation is performed for distance d, where d can be
any distance in the range 1 km to 1000 km from transmitter,
the following expression is used:
log(d / dinf )
Eo = Einf + ( Esup − Einf ) , (3)
log(d /d )
sup inf
221
where Einf is signal level read from the h1 curve for the nearest
distance less than distance d (dinf) while Esup is signal level is
read from the h1 curve for the nearest distance greater than
distance d (dsup).
If the frequency of the system differs from one of the
discrete values of the frequencies given in ITU-R P.1546
recommendation, the interpolation is performed according to
the following equation:
log( f / f inf )
Eo = Einf + ( E sup − Einf ) , (4)
log( f sup / f inf )
where Einf is signal level read from the h1 curve for distance d
for the nearest frequency less than frequency f (finf), while Esup
is signal level read form the h1 curve for distance d for the
nearest frequency greater than frequency f (fsup). If f is less
than 600 MHz finf is equal to 100 MHz while in all other cases
f is equal to the 2000 MHz.
Recommendation ITU-R P.1546 can be used for systems
operating in the range 30 MHz to 4000 MHz and for systems
where h1 does not exceed 3000 m.
Radio frequency band is a resource whose distribution is
defined by multiple different official documents and is
divided among a large number of users. Most of the
broadcasting systems contain operating parameters that do not
match the discrete values of ITU-R P.1546 Recommendation,
therefore in design of a broadcasting system, a large number
of interpolations have to be performed according to
expressions (2), (3) and (4) to determine the EM field level in
the coverage area. As already mentioned, one of the most
important steps in the design of broadcasting systems is to
determine the coverage range of the system, i.e.,
determination of the maximum distance for the requested
field level, which creates a need for inverse reading of the EM
field level curves. As mentioned in Introduction, one of the
efficient methods for inverse reading of the curves from the
ITU-R P.1546 Recommendation is based on artificial neural
networks [2]. In the next section an extension of the method,
with extended validity against the frequency is proposed.
III. PROPOSED METHOD
The proposed method is based on the use of multi-layered
artificial neural networks for inverse curve reading. A multi-
layered neural network consists of basic units, neurons,
grouped in layers; input layer, one or more hidden layers and
the output layer. The number of neurons in the input layer
depends on the number of input parameters, while the number
of neurons in the output layer corresponds to the number of
output parameters. The number of hidden layers and the
number of neurons in each hidden layer is determined during
network training. The neural network learns the dependence
between input-output parameters through optimization
procedures, which optimize the network parameters (weights
of connections between neurons and thresholds of activation
functions of neurons), in order to make the network response
as close as possible to the desired dependence between input-
output parameters.
 special care was taken to avoid network over-learning for the
inputs different than the training data to assure that the
network has good generalization. Generalization was checked
by testing the network for input values that were not used
during learning. The test of the ANN model accuracy is done
Fig. 2. ANN model proposed in [2] calculating average test error (ATE - Average Test Error),
maximum error (WCE - Worst Case Error) and correlation
coefficient [3]. Moreover, as an additional check whether the
network exhibits over-learning was done visually, by
checking the network response plotted versus different input
values other than those used in network training.
After testing the developed models, the network with three
Fig. 3. ANN model proposed in this paper hidden layers with 10, 12 and 14 neurons in the first, second
and third hidden layers, respectively, was chosen as the best
Fig. 2 shows the method proposed in [2]. It is an ANN
model for determining the distance for a given level of EM
trained to calculate the distance for the given EM field level
field.
and height h1.
To illustrate the training accuracy, Fig. 4 shows the
The ANN is trained with the values of the EM field level
reference curves for the frequency 1300 MHz (one of the 39
corresponding to one chosen values of frequency. In [2]
frequencies used during the training) and the heights of 10,
results for frequency of 600 MHz were given.
20, 37.5, 75, 110, 150, 200, 250, 300, 400, 500, 600, 750, 900,
In order to avoid the development of a separate neural
1050 and 1200 m and curves generated by the neural network
network for each frequency defined by the ITU-R P.1546
for the same input parameters. It can be observed that the
Recommendation, or for an operating frequency of the
reference values generated by the neural model follow very
transmitter the model is extended to enable calculation in
well the reference values, which confirms good performance
entire frequency range without using individual interpolation
and good modeling accuracy.
calculations, as shown in Fig. 3.
As can be seen from Fig. 3, the training is performed with
three input parameters, frequency, f, height, h1, requested EM
field level, E and the logarithmic values of the distance, log d,
giving the relation log(d) = f (f, h1, E). Therefore, ANN
consists of three neurons in the input layer and one neuron in
the output layer. It should be mentioned that logarithmic
representation of the distance gave better results than the
linear representation.

IV. RESULTS

The first ANNs trained with the data given in the

Recommendation (discrete values of the frequencies 100
MHz, 600 MHz and 2000 MHz and the discrete values of the
heights 10, 20, 37.5, 75, 150, 300, 600 and 1200 m) did not
provide satisfactory accuracy of the model for any value of
frequency and height. Therefore, a larger training set was
formed, with frequencies from 100 to 4000 MHz with a step
of 100 MHz and heights of 10, 20, 37.5, 75, 110, 150, 200,
250, 300, 400, 500, 600, 750, 900, 1050 and 1200 m. The Fig. 4. Comparison of distance values generated by using the
preparation of the data used for the training was performed developed neural model and the reference values for the training
using expressions (2) and (4). During the development of the frequency 1300 MHz
most accurate model, several networks with one input layer,
two and three hidden layers and with one output layer were Fig. 5 shows the test results for the values different from
tested. Different combinations of the number of neurons in those used during training. From Fig. 5 it can be noticed that
each hidden layer were used, in order to obtain the ANN the values generated by neural network model are very close
giving the most accurate results. As input neurons have a to the reference values. To illustrate better the modeling
buffer function, their function is a unit linear function; accuracy for smaller distances, in Fig. 6 the distance values up
neurons in the hidden layers have a sigmoid transfer function to 20 km were given for the same input parameters as in Fig.
and neurons in the output layer have a linear transfer function. 5. Values from Fig. 5 are shown in higher resolution for
The Levenberg-Maquardt algorithm [3] was applied for distances up to 20 km. In addition, the error calculation results
network training. ANN training and testing was performed in for mentioned test set referring to 3050 MHz are presented in
the Matlab software package. Model accuracy was tested by Table I.
comparing network response and reference values. Also, a

222
 As it is shown in Table I the maximum average error is
slightly greater than 0.1% while the maximum error is slightly
greater than 1.5%. The correlation coefficient is very close to
one, which also confirms the high accuracy of the modeling.

Fig. 5. Comparison of distance values generated by the developed
neural model and the test values not used for training (height h1 of
28, 65, 90, 130, 350, 550 and 700 m and frequency of 3050 MHz)
V. CONCLUSION
Recommendation ITU-R P.1546 gives the EM field level
for discrete values of frequency and antena height. Therefore
to determine the field strength level or the distance for the
given field strength level it would be necessary to perform a
series of interpolations, the caluculations of the coverage area
complex. The method proposed in this paper provides the
ability to instantly calculate the distance by finding the
response of the neural network for any value of the frequency
and antenna height. As each neural network can be described
using a set of equations based on linear and exponential
functions, the calculation of distances using a developed
neural network is simple and can be performed in any
software environment.
The presented method of determining the distance for a
given field level of a broadcast transmitter can be used as an
efficient tool for determining the coverage area of the
transmitter using the statistical method defined by
Recommendation ITU-R P.1546. The aim of future research is
to improve the methods for application in special cases, when
h1 is less than 10 m, as the field strength level for the
mentioned values is not given in the curves.

REFERENCES
[1] P.1546-6, „Method for Point-to-Area Predictions for Terrestrial
Services in the Frequency range 30 MHz to 3 000 MHz“, ITU,
08/2019.
[2] M. Radojković, Z. Marinković, and A. Atanasković, „Inverse
Reading of the ITU-R P.1546 Curves for the EM Field Level
Originating from Broadcast Transmitters”, 63th International
ETRAN Conference, Palić, Serbia, June 11-14, 2018, pp. 334-
337 (in Serbian).
[3] Q. J. Zhang, K. C. Gupta, Neural Networks for RF and
Microwave Design, Boston, MA, Artech House, 2000.
[4] C. Christodoulou, M. Georgiopoulos, Applications of Neural
Networks in Electromagnetics, Artech House, Inc. Norwood,
Fig. 6. Comparison of distance values generated by the developed MA, USA, 2000.
neural model and the test values not used for training (height h1 of [5] Z. Stanković, N. Vasić, “Neural Approach in Modeling of the
28, 65, 90, 130, 350, 550 and 700 m and frequency of 3050 MHz for Propagation Curves from Recommendation ITU-R P.370-7”,
the distances up to 20 km) XLVII ETRAN Conference, Herceg Novi, Serbia and
Montenegro, June 8-13, 2003, pp. 269-272.
TABLE I [6] J. E. Rayas-Sanchez, “EM-based Optimization of Microwave
TEST STATISTICS FOR TRAINING AND TEST DATA Circuits Using Artificial Neural Networks: The State-of-the-
Art”, IEEE Trans. Microw. Theory Tech., vol. 52, no. 1, pp.
ATE [%] WCE [%] r 420-435, Jan. 2004.
training set 0.1083460 1.55836 0.999981 [7] B. Milovanović, Z. Stanković, M. Ostojić, and M. Milijić,
“Prediction of Electromagnetic Field Levels Using a Neural
test set, h1=28m 0.0670648 0.281798 0.999994 Model Based on ITU-R P.1546 Recommendation“, 51st ETRAN
test set, h1=65m 0.1021850 0.305793 0.999998 Conference, Herceg Novi - Igalo, Montenegro, June 4-8, 2007,
test set, h1=90m 0.0606628 0.291294 0.999998 pp. MT1.3.1-4 (in Serbian).
test set, h1=130m 0.0564776 0.224658 0.999997 [8] Z. Marinković, G. Crupi, A. Caddemi, V. Marković, and D.
test set, h1=350m 0.0751861 0.32886 0.999995 M.M.‐P. Schreurs, “A Review on the Artificial Neural Network
test set, h1=550m 0.0792733 0.321617 0.999992 Applications for Small‐Signal Modeling of Microwave FETs”,
Intern. Journal of Numerical Modeling: Electronic Networks,
test set, h1=700m 0.0736781 0.353862 0.999992 Devices and Fields, vol. 33, no. 3, e2668, May/June 2020.

223
 GPU-Enabled Simulation of a Combined Macrodiversity
System in Gamma-Shadowed Nakagami-m Fading
Channel
Nenad Petrović, Dejan Milić, Suad Suljović, Samir Koničanin, Selena Vasić
Abstract – We propose the combined diversity receiver
consisting of macro-diversity (MD) selection combining (SC)
receiver and two micro-diversity (µD) maximum ratio combining
(MRC) receivers, as a solution to the impediments a mobile
signal undergoes in a cellular mobile radio Gamma-shadowed
channel with a short-term Nakagami-m fading. The closed-form
expressions for the moment generating function (MGF) for both
µD MRC receivers’ output signal and macro-diversity SC
receiver’s output signal envelopes are derived. By varying
Nakagami-m short term fading severity parameter m and the
order of Gamma distribution c we have considered and discussed
the change in system performances and graphically presented
numerical results. Moreover, we integrate the obtained formulas
within the simulation environment for optimal network planning
leveraging general-purpose computation on GPU for faster
execution.
Keywords – Gamma shadowing, Nakagami-m fading, GPGPU,
Maximal Ratio Combining (MRC), Selection Combining (SC),
Moment Generating Function (MGF), GPU-enabled simulation.
I. INTRODUCTION
In a Gamma-shadowed, Nakagami-m fading wireless
channel, on its way between a transmitter and a receiver, the
signal is spread out into multiple components that travel along
different paths. This causes a substantial signal degrading
effect commonly known as fading. Both short- and long-term
fading degrade channel performances and limit capacity of a
wireless communication system. The short-term fading
(multipath fading) causes a signal envelope variation. A long-
term fading can be modelled with Gamma distribution or
long-normal distribution. When long-term fading is described
using Gamma distribution, the expressions for outage
probability (OP) and bit error probability (BER) are in closed
form. Another degrading effect, known as shadowing causes
power variation in a signal envelope. All these effects should
be considered when it comes to impact on performance.
To mitigate the effects of both long-term and short-term
fading on the performances in cellular mobile radio systems
Nenad Petrović, Dejan Milić, Suad Suljović and Samir Koničanin
are with University of Niš, Faculty of Electronic Engineering,
Aleksandra Medvedeva 14, 18000 Niš, Serbia, E-mails:
nenad.petrovic@elfak.ni.ac.rs, dejan.milic@elfak.ni.ac.rs,
suadsara@gmail.com, samirk@telekom.rs
Selena Vasić is with the Faculty of Information Technology,
Metropolitan University, Tadeuša Košćuška 63,11000 Belgrade,
Serbia, E-mail: selena.vasic@metropolitan.ac.rs
978-1-7281-7142-5/20/$31.00 ©2020 IEEE
[1], a macro-diversity (MD) receiver is utilized. MD receiver
has been proved to improve the performances as it is almost
impossible that independent signals in a diversity system all
experience deep fading at the same time. µD systems can
reduce the impact of fast fading, while the MD system reduces
shadowing effects. In the scientific testing of signal
transmission so far, using diversion techniques that reduce the
impact of fading and increase channel capacity, MRC and
EGC combining techniques, followed by SC and SSC
techniques have shown the best results [2]. Considering the
complexity of receiver tuning with MRC and EGC combiners,
in practice a more realistic solution is to use SC combiners,
that picks the branch with the greatest signal-to-noise ratio
(SNR), especially when the correlation between channels is
minimal [3]. For MRC combiners, if the noise power is the
same at all the inputs, then the square of the signal at the
output will be equal to the sum of the squares of input signals.
The distribution of the signal at the input to the MRC receiver,
as well as at the output of the MRC receiver, increases the
mean power in direct proportion to the number of inputs [4].
In wireless channels where there is no optical visibility
between transmitter and receiver (no LOS-line of sight), and
the signal is affected by fading, there are no conditions for
formation of a dominant component. Then, Nakagami-m
distribution is successfully used to describe the signal
propagation [5]. Nakagami-m distribution is generalized
distribution characterized by two non-negative parameters: m
that controls the shape of distribution and Ω which represents
the mean power (it affects the spread) [6]. Larger values of m
indicate less-degrading fading conditions. When m approaches
to infinity, Nakagami-m fading channel becomes no-fading
channel. When parameter m equals to one, Nakagami-m
distribution is reduced to a Rayleigh distribution. If m is 0.5,
Nakagami-m distribution becomes a single-sided Gaussian
distribution representing the most severe fading environment
that can be described by Nakagami-m distribution [7].
The analysis of average fade duration (AFD) and average
level crossing rate (LCR) of MD system with two µD
receivers in Gamma-shadowed Rician fading channel has
been conducted [8]. Similar statistical analysis of MD system
with two µD MRC receivers in a channel undergoing Gamma
long-term fading and Nakagami-m short-term fading has been
performed in [9]. Another work applies moments-based
approach to performance analysis of an L-branch redetections
EGC combiner over independent or correlated Nakagami-m
fading channels [10]. MGF is successfully approximated with
the aid of Padé approximants theory. Based on MGF-
approach, the first order statistics such as average output SNR,
224
the average symbol-error probability (SEP) and the outage MGF of yi in case of = 2 branches at the output of the µD
probability (OP) are accurately estimated. MRC receiver will be:
Based on the encouraging results reported in terms of
positive effects on wireless system performances when MD M yi (s ) = e − yi s = e − ( yi1 + yi 2 )s = e − yi1s e − yi 2 s =
system is applied, in this paper we will consider the combined
receiver with one MD SC and two µD MRC receivers in a
4m 2mΩi Γ 2 (m + 1 / 2)
Gamma-shadowed Nakagami-m multipath fading channel. = M y1 (s )M y 2 (s ) = . (4)
MGF of the signals at the output of µD MRC receivers will be Γ 2 (m)(m + sΩi )2m +1
calculated and the results used to further evaluate the MGF of
the signal at the output of MD SC receiver. The analysis of In a Gamma-shadowed channel, the random variables Ω1
MGF is needed to calculate the first-order statistics of the and Ω2are Gamma-distributed [13]:
considered MD system such as outage probability (OP) and −(Ω1+Ω2 ) / Ω0  1− ρ 2 
e   ∞ ρ 2i
bit error probability (BEP). To the best of authors’ knowledge
such a combined receiver has not been analyzed in related
pΩ1Ω 2 (Ω1Ω 2 ) =
Γ(c )
⋅ 
i =0 i!
technical literature. The analytical and simulation results of
this study can be applied in design and performance
evaluation of MD wireless communication systems.
(Ω1Ω 2 )i +c−1
⋅ ; Ω1 ≥ 0, Ω 2 ≥ 0. (5)
Γ(i + c )Ω 02i +2c (1 − ρ 2 )
2i +c

II. MOMENT GENERATING FUNCTION

where c is order of Gamma distribution, Ω0 denotes the mean
value of Ω1 and Ω2 while ρ represents the correlation
Fig. 1 presents the model of the proposed MD system.
coefficient.
At the output of the MD SC receiver the moment-
generating function (MGF) of the signal is given by [4]:
∞ Ω1 ∞
M y (s ) = dΩ1 dΩ 2 M y1 (s ) pΩ1Ω 2 (Ω1Ω 2 ) + dΩ 2
  
0 0 0

Ω2 ∞ Ω1

 dΩ1M y2 (s ) = 2 dΩ1 dΩ 2 M y1 (s ) pΩ1Ω 2 (Ω1Ω 2 ). (6)

 
0 0 0

By substituting terms (4) and (5) into expression (6), we

Fig. 1. MD system with MD SC receiver and two µD MRC receivers
obtain the expression for MGF:
Signal envelopes xij have Nakagami-m distribution [11]:
8m2mΓ 2 (m + 1/ 2) ∞ ρ 2i1
M y (s ) =  ⋅
( )
p x x ij =
2  m  2 m−1 − Ω
  xij e
m m
i
xij2
; x ij ≥ 0, i, j = 1,2. (1)
Γ(c)Γ 2 (m) i1=0 i !Γ(i + c )Ω2i1+2c 1 − ρ 2
1 1 0 ( ) 2i1+c

ij
Γ (m )  Ω 
Ω1 Ω2
− −
where Γ(m) is gamma function, m is Nakagami-m short-term ∞
Ω1i1+c  Ω
Ω0  1− ρ 2  1 Ω0  1− ρ 2 


 
  dΩ i1+c −1  
fading severity parameter and Ω i = xij2 is a power of desired
⋅ dΩ1 e 2Ω 2 e (7)
0 (m + sΩ1 ) 2 m+1
0
signal, respectively. After transformation yij=xij2, PDF of the
signal xij becomes: The integral I1 in the expression (7) is:
m 1 − m y
( ) Γ(2m)  Ωm 
m− ij Ω1
p yij yij = yij 2 e Ω i ; yij ≥ 0, i, j = 1,2. (2) − Ω 2 / Ω 0 1− ρ 2 
  I1 =  dΩ2Ωi21 + c −1e   =
0
MGF of random variable yij>0 is [12]:
e −Ω / Ω (1−ρ )
∞ 2 i1 + i2 + c
( ) Ω1
∞
M yij (s ) = e
− y ij s − y ij s 1 0

=  dyij e p yij yij = =  . (8)

0 (i1 + c ) i2 = 0 (Ω (1 − ρ )) (i
0
2 i2
1 + c + 1)i 2

2m m Ω1i / 2
= Γ(m + 1 / 2). (3)
Γ(m)(m + sΩ i )m+1 / 2 After inserting the expression (8) in the expression (7), we
can rewrite the MGF in the following form:

225
 8Γ 2 (m + 1 / 2 ) ∞ ∞
ρ 2i1 Ω 0− ( 2i1 +i2 + 2c ) performance deteriorates. As a result of the combined effects
M y (s ) =
mΓ(c )Γ 2 (m )
 ( + c )(i1 + c )(i1 + c + 1)i2
⋅ of the observed parameters, it is noted that the correlation
i1 = 0 i2 = 0 i1 ! Γ i1 coefficient ρ exhibits higher influence on MGF for higher
values of Nakagami-m severity parameter m and higher values
− 2Ω1 / Ω0 1− ρ 2  of Gamma long-term fading severity parameter.
∞
1 Ω12i1 +i2 + 2c +1−1e  
⋅
(1 − ρ )
2 2i1 +i2 + c  dΩ1 (1 + (s / m)Ω1 )2m+1
. (9)
0 10
m=1, c=1, ρ=0.2, Ω0=1
m=2, c=1, ρ=0.2, Ω0=1
Using the following identity [14; 3.383]: m=3, c=1, ρ=0.2, Ω0=1
1
m=1, c=2, ρ=0.2, Ω0=1
∞ q −1 − px
x e  p
dx = a Γ(q )Ψ q, q + 1 − v, 
m=1, c=3, ρ=0.2, Ω0=1
 (1 + ax )v
−q
(10) m=1, c=1, ρ=0.4, Ω0=1
0  a 0,1
m=1, c=1, ρ=0.6, Ω0=1

My[s]
m=1, c=1, ρ=0.8, Ω0=1
In the above expression the factor ψ (s, d, t) refers to a 0,01
confluent hypergeometric function. Hence, the last integral in
expression (9) can be denoted by I2 and written in the
1E-3
following form:
2Ω1
−
∞
Ω12i1 + i2 + 2c +1−1 Ω 0  1− ρ 2  1E-4

I2 =  dΩ1
( (
1 + s / m Ω1 2m +1) )
e   = -5 0 5
s[dB] 10 15 20

2 i1 + i 2 + 2 c +1
Fig. 2. MGF of MD SC receiver output signal, when m, c and ρ
m changed
=  Γ(2i1 + i2 + 2c + 1) ⋅
s
IV. MODELLING AND SIMULATION ENVIRONMENT
 2m 
⋅ Ψ  2i1 + i2 + 2c + 1,2i1 + i2 + 2c − 2m + 1, . (11)
 sΩ 0 1 − ρ 2 ( ) 
 Using the calculations of fading effect and our previous
work presented in [15] and [16], the simulation environment
By substituting the (11) in the expression (9) we can obtain is built that use the general-purpose GPU (GPGPU)
the formula for MGF of the MD SC receiver’s output signal: programming paradigm. This enables substantionally faster
calculations of huge data streams in non-graphic applications
8Γ 2 (m + 1 / 2) ∞ ∞ ρ 2i1 (m / Ω 0 )2i1 + i 2 + 2c on GPU. It has been approved as an effective approach to
M y (s ) = 
Γ(c )Γ 2 (m ) i1 = 0i 2 = 0 i1!Γ(i1 + c )(i1 + c ) performance improvement in fading channel simulations [17].
First, user draws a model of base station and antenna
configuration using the graphical modelling environment with
Γ(2i1 + i2 + 2c + 1)
⋅ ⋅ domain-specific notation introduced in [15]. Then, via
(i1 + c + 1)i2 (1 − ρ 2 )
2i1 + i 2 + c 2i + i + 2 c +1
s 1 2 graphical interface, the parameters about both the network
configuration related to transmitter, interference, channel,
  receiver and service consumers can be set. The modelling
2m
⋅ Ψ  2i1 + i2 + 2c + 1,2i1 + i2 + 2c − 2m + 1,  (12) environment is run in web browser and implemented relying
 sΩ 0 1 − ρ 2 ( ) 
 on JavaScript, Three.js, HTML and CSS. In Fig. 3, a
screenshot of mobile network modelling and simulation
III. NUMERICAL RESULTS environment aiming smart cities run in web browser is given.
Fig. 2 depicts the MGF derived in (12) for different values
of parameters m, c and ρ. It represents the way the MGF of
MD SC receiver output changes in terms of instantaneous
signal-to-noise ratio s for several values of Nakagami-m short-
term fading severity parameter m, Gamma long-term fading
correlation coefficient ρ and average signal power Ω.
It can be observed that the increase in fading parameter m
and parameter c causes the decrease of MGF, which has a
positive effect on the system performances. Another effect is
that higher values of the fading parameter m affect the MGF
more intensively. Also, if the correlation coefficient ρ drops to
zero, the outage probability of the system decreases, causing
therefore the improvement of the system performances. On
Fig. 3. Mobile network modelling and simulation environment run in
the other hand, if the correlation is very strong and approaches web browser: autonomous cars (consumers) and antenna (service
to one, the outage probability is higher and the system provider)

226
 After that, the model parameters are sent to the backend
which is responsible for calculation of quality of service
(QoS)-related parameters considering the effect of fading and
leveraging CUDA1-enabled C code for execution on GPU. A
kernel for MGF calculation in CUDA-enabled C program is
shown in Fig. 4. It provided around 25 times faster execution
than our previous work from [15] and 38 times speed-up in
case of similar system with three µD MRC receivers [18].
__global__ void calculate(float* x, float* y)
{
int tid = threadIdx.x + blockIdx.x * blockDim.x;
while (tid < N)
{ IEEE Press, New York, USA, 1994.
y[tid] = my(x[tid]);
tid += blockDim.x * gridDim.x;
} IEEE Transactions on Wireless Communications, vol. 11, no. 9,
} pp. 3316-3325, 2012.
Fig. 4. CUDA kernel for GPU-enabled MGF calculation
The calculated values, together with user-defined model
parameters are used to construct input for linear optimization
process with respect to linear program from [15]. Finally, the
process of linear optimization is performed resulting with
optimal base station configuration. Moreover, the obtained
optimization results can be further transformed to software-
defined radio or base station hardware-specific commands for
adaptive improvement of QoS in real scenarios. In Fig. 5, an
overview of the modelling and simulation environment for
network planning is given.
Modelling Fading Linear
1
tool 2
calculation 3
optimization
User
Fig. 5. Overview of web-based tool for modelling and simulation:
1- Drawing a diagram; 2-Model; 3-Calculated QoSvalues;
4-Optimization results; 5-Hardware-specific commands
V. CONCLUSION AND FUTURE WORK
The analyses presented above refers tothe scenario of a
cellular mobile radio system where the signal undergoes
Gamma long-term fading and Nakagami-m short-term fading.
Combined MD system with one MDSC receiver and two µD
MRC receivers has been proposed to minimize deteriorating
effects in the channel. The higher values of Nakagami
parameter m induce the decrease of MGF and therefore the
enhancement of the bit error probability. The value of
correlation coefficient is inversely related to system
performances. The higher the ρ the worse system
performance. The derived formulas are integrated within the
softwaresimulation for purpose of QoS parameter calculation
aiming optimal network resource planning. GPGPU approach
is used to speed up the calculations of fading effect and has
proven to be a powerful tool in this kind of analysis.
ACKNOWLEDGEMENT
This work has been supported by the Ministry of Education,
Science and Technological Development of the Republic of
Serbia.
1
https://developer.nvidia.com/cuda-zone
REFERENCES
[1] M. K. Simon, M. S. Alouini, Digital Communication over
Fading Channels, Wiley: New York, 2000.
[2] G. L. Stüber, Principles of Mobile Communication: 2nd ed.,
Technology & Engineering, Kluwer Academic Publisher, 2000.
[3] E. A. Neasmith, N. C. Beaulieu, "New Results in Selection
Diversity", IEEE Transaction on Communications, vol. 46, pp.
695-704, 1998.
[4] S.Panic et al.,Fading and Interference Mitigation in Wireless
Communications, CRC Press: USA, 2013.
[5] C. W. Jakes, Microwave Mobile Communications, Piscataway,
[6] N. Wang, X. Song, and J. Cheng, "Generalized Method of
Moments Estimation of the Nakagami-m fading Parameter",
[7] P. Shankar, Fading and Shadowing in Wireless Systems, New
York: Springer, 2012.
[8] N. Sekulovic et al.,"Second Order Statistics of System with N-
Branch Micro Diversity Operating over Gamma Shadowed
Rician Fading Channel", International Journal of
Communication systems, vol. 27, no. 2, pp. 390-400, 2014.
[9] B. Jaksic et al.,"Level Crossing Rate of Macro diversity System
in the Presence of Multipath Fading and Shadowing", Radio
engineering, vol. 24, no. 1, pp. 185-191, 2015.
[10] I. Temelkovski et al., "Moments of Macrodiversity SC Receiver
with two Microdiversity EGC Receivers over Gamma
Shadowed Rician Multipath Fading Channel", TEM Journal,
vol. 4, no 3, pp. 292-296, 2015.
[11] M. Nakagami, "The m-distribution, a General Formula of
Intensity Distribution of Rapid Fading", Statistical Methods in
Command
4
generator
5
Radio Wave Propagation, Oxford, U.K., Pergamon, 1960.
Optimal network
[12] J. Malhotra, "Performance Analysis of Diversity Combining
configuration Multichannel Receivers in Generic-Gamma Fading Channels",
Tamkang Journal of Science and Engineering, vol. 14, no. 4, pp.
333-340, 2011.
[13] S. Suljović et al.,"Level Crossing Rate of Macro-diversity
System in the Presence of Fading and Co-channel Interference",
Revue Roumaine des Sciences Techniques-Serie El. et
Energetique, vol. 64, no. 1, pp. 63-68, 2019.
[14] I. Gradshteyn, I. Ryzhik, Tables of Integrals, Series, and
products. Academic Press, New York 1994.
[15] N. Petrović, S. Koničanin, D. Milić, S. Suljović, and S. Panić,
"GPU-enabled Framework for Modelling, Simulation and
Planning of Mobile Networks in Smart Cities", ZINC 2020, pp.
1-6, 2020, https://doi.org/10.1109/ZINC50678.2020.9161773
[16] D. Milić, S. Suljović, N. Petrović, S. Koničanin, and S. Panić,
"Software Environment for Performance of Relay Signal by DF
Technique Influenced by k-µ Fading", INFOTEH 2020, pp. 1-4,
2020, https://doi.org/10.1109/INFOTEH48170.2020.9066304
[17] A. F. Abdelrazek et al., "A Novel Architecture using NVIDIA
CUDA Tospeed up Simulation of Multi-Path Fast Fading
Channels", pp. 1-5, 2009.
[18] D. Krstić et al., "Moment Generating Function of
Macrodiversity System with Three Microdiversity MRC
Receivers in Gamma Shadowed Nakagami-m Fading Channel",
CSNDSP 2016, pp. 1-6, 2016,
https://doi.org/10.1109/CSNDSP.2016.7573984
227
 MEASUREMENT,
CONTROL SYSTEMS AND ICT
 Measurement and Analysis of PM Particles
Emitted by Automotive Brakes
Emil Iontchev1, Rosen Miletiev2, Rumen Yordanov3 and Ilian Damyanov4
Abstract – All vehicles emit fine particles into the air during
the braking due to the tyres and brakes friction. The brakes may
be the most significant particulate matter (PM) contributor,
especiallyin places with a high traffic density and braking
frequency. As the PM particles are associated with different
respiratory symptoms even sometimes with a premature death it
is important to analysed the PM emitted by the automotive
brakes and their distribution from PM0.3 to PM10. The current
paper discusses the measurement and analysis of the PM particle
sizes from 0.3µm to 10µm during the slow down process, their
time and size distribution according to the brake intensity.
Keywords – Brakes, Fine particles, Particle distribution.
while the emission of fine and ultrafine particles are still
poorly studied.
The current paper discusses the measurement and the
analysis of the brake particle emissions during the slow down
processes with diameter from 0.3µm to 10µm. In the same
time a temperature monitoring of the rotation disk is
accomplished to measure the thermal heating using the
thermal imaging estimation method [9]. The analysis shows
the particle concentration of all three PM categories and
particle count @ 0.1L air for six different particle sizes – from
0.3µm to 10µm.

II. PM MEASUREMENT METHOD

I. INTRODUCTION
The particulate matter emission measurements are
The particulate matters (PM) are divided into three implemented on the front disc brakesof passenger vehicleas
categories– coarse (diameter of between 2.5µm and 10µm), the braking force is greater than the rear wheels.As the brake
fine (0.1 to 2.5µm in diameter) and ultrafine (<0.1µm in particles are evolved during the braking process, the particle
diameter) particles [1]. The particles over 10µm in diameter movement direction is expected to lieinto the coarse size
are efficiently filtered in the nose while the particles below fraction [8], the PM sensor input is positioned under the brake
<0.1µm in diameter directly pass to the blood steam through shoes to collect the maximum portion of the emitted particles
the alveoli. The particles less than 2.5μm in diameter are (Fig. 1). The tire is also removed as the air ventilation during
associated with respiratory problems while the particulate the disk rotation scatters the brake wear particles.
matter particles with diameter between 2.5 and 10 µm are
responsible for asthma symptoms, eye and lung problems and
premature death [2]. At presentthe real-time PM counting
methods are based on light scattering or absorption processes.
The most of the commercial low-cost PM counting sensors
have the typical accuracy from 10 to 15% [3-5] and may
separate the above mentioned PM groups. The comparison
study of some PM sensor performance was established at [6].
The vehicles are one of the major source of PM particles
due to the tyres and brakes friction [7]. A recent review shows
that brake-wear emissions are more than50% by mass of total
non-exhaust traffic-related PM10 emissions and up to 21% by
mass to total traffic-related PM10 emissions [8]. But most of
the studies concern the particles with diameter over 2.5µm

1Emil Iontchev is with the Higher School of Transport “T.

Kableshkov” 158 Geo Milev Street,Sofia 1574, Bulgaria, E-mail:
e_iontchev@yahoo.com
2RosenMiletiev is with the Faculty of Telecommunications at

Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,

Bulgaria,E-mail: miletiev@tu-sofia.bg
3Rumen Yordanov is with the Faculty of Telecommunications at

Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,

Bulgaria,E-mail: rsyordanov@tyahoo.com
4Ilian Damyanov is with the Faculty of Telecommunications at

Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,

Bulgaria,E-mail: idamyanov@tu-sofia.bg Fig. 1. Particulate measurement method and assembly

978-1-7281-7142-5/20/$31.00 ©2020 European Union 

 The particulate matter are measured using a low-cost sensor
type CP-15-A4-CG, which is commercial grade laser-based
sensor with digital outputs (UART and PWM) (Fig. 2). The
measurement range is defined as 0.3-10μm, TTL 3.3V level
outputs, 3s response time and PM2.5 consistency of ±10μg/m3
or ±10% [10]. The declared degree of the approximation to
the professional instrument is equal to 98%.

Fig. 2. PM sensor type CP-15-A4-CG

The PM sensor is connected to USB port via TTL to USB

converter and the sensor is powered from USB port as the
maximum sensor power consumption not exceeded 100mA.
The sensor measurement packets are captured, recorded and
analyzed by PM measurement tool (Fig. 3) specially designed
tool at MATLAB environment on virtual COM port, which is
created by the TTL/USB converter.
Each measurement data are saved as HEX values (RAW
format) and stored in a TXT file (decoded format) using CSV
format:
• Time (seconds from start recording)
• PM1.0 (µg/m3)
• PM2.5 (µg/ m3)
• PM10 (µg/ m3)
• particles 0.3µm @ 0.1L air
• particles 0.5µm @ 0.1L air
• particles 1.0µm @ 0.1L air
• particles 2.5µm @ 0.1L air
• particles 5.0µm @ 0.1L air
• particles 10µm @ 0.1L air
The measurement process is initiated by selecting the real
or virtual COM port from “Serial port:” menu and open it by
clicking on OPEN button. As the connections is established,
the recording process is started when the START button is
pressed the system shows the recording time in “Scan time”
field. The recording process is finished when the STOP button
is pressed and the PM graphics may be generated by clicking
on “Data analysis” button. During the measurement the real-
time values of the PM concentrations and particle numbers are
shown in the corresponding field in the tool.
As the recording process is accomplished the disk thermal
image is made by FLIR camera installed laterally to the
rotation disk.
.
Fig. 3. Measurement and recording tool
III. EXPERIMENTAL RESULTS AND ANALYSIS
The PM measurements are accomplished using different
situations:
• Test 0 - no brakes
• Test 1 – speed 40km/h, 3-th gear, no blocking
brakes, low load slow down
• Test 2 – speed 40km/h, 3-th gear, blocking + no
blocking brakes
• Test 3 – speed 30km/h, continuous brake, slow
stop
• Test 4 – speed 50km/h, high load slow down
• Test 5 – speed 60km/h, extremely high load slow
down
The first test (Test 0 – no brakes) is performed to measure
the PM levels in the test environment and compare the levels
with the other experimental results. The measured PM
concentrations are as follows PM1.0 = 34 (µg/m3),
PM2.5 = 38(µg/m3), PM10 = 53 (µg/m3), while the particle
numbers @ 0.1L air are equal to 5500 (0.3µM), 1925
(0.5µM), 165 (1.0µM), 13 (2.5µM), 1 (5.0µM), 2 (10µM).
Fig. 4 represents the PM concentrations during the tests
from Test 1 (case a) to Test 5 (case e), while Figure 5
represents the particle count @ 0.1L air during the tests from
Test 1 (case a) to Test 5 (case e).


 (a)
(a)

(b)
(b)

(c) (c)

(d)
(d)

(e)
(e)
Fig. 4. PM concentration for Test 1÷5
Fig. 5. Particle count @0.1L air for Test 1÷5


 The PM measurements shows that during the slow down
process the particle concentrations rise significantly from
50µg/m3 (Test 0 levels) to approximately 1500µg/m3 and
remain at the constant levels during the slow down duration
(Fig. 4c) for all particle sizes. The PM10 particles are
dominant but PM1.0 and PM2.5 concentrations exceeded 15-
20 times the Test 0 levels.
The particle count analysis shows that the particle numbers
are also changed dramatically during the slow down process.
The number of the smallest size particles (below 0.5µm)
exceeded the sensor measuring range of 65535 particles (the
particle number is sent as a 16bit value) in all cases especially
during the continuous time braking (Fig. 5c).
Fig. 6. FLIR thermal image of the rotating disk
The FLIR thermal image of the rotation disk at the end of
the experiments shows significant rise of the disk surface
temperature to 216oC due to the friction. Therefore the
complete particle analysis also requires to take into account
not only the distribution of the particles according to their size
but also the high process temperature and possible chemical
compounds used in the rotation disk and brake shoes
production process as the metal and fiber particles may have
different influence on the human health.
IV. CONCLUSION
The experiment results represent that the braking process
emits significant number of particulate matters of all sizes
which are associated with asthma and respiratory symptoms.

Therefore the emitted PM particles, particularly in urban
environments and major cities, may cause several lung
problems and their emissions have to be reduced significantly
by adding the particle collecting system to the automotive
construction.
ACKNOWLEDGEMENT
The authors would like to thank the Faculty of
Telecommunications at the Technical University of Sofia for
the financial support.
REFERENCES
[1] J. Thongplang, “Particulate Matters: Why Monitor PM10 and
PM2.5?”, Last Updated: Dec 19, 2018.
[2] M. Morais-Almeida, N. Santos, A. M. Pereira, M. Branco-
Ferreira, C. Nunes, J. Bousquet, and J. A. Fonseca, “Prevalence
and Classification of Rhinitis in Preschool Children in Portugal:
a Nationwide Study”, Allergy, vol. 68, no. 10, 1278-1288, 2013.
[3] P. Kumar, L. Morawska, C. Martani et al., “The Rise of Low-
Cost Sensing for Managing Air Pollution in Cities”,
Environment International, vol. 75, pp. 199-205, 2015.
[4] E. G. Snyder, T. H. Watkins, P. A. Solomon et al., “The
Changing Paradigm of Air Pollution Monitoring”,
Environmental Science & Technology, vol. 47, no. 20, pp.
11369-11377, 2013.
[5] N. Castell, F. R. Dauge, P. Schneider et al., “Can Commercial
Low-Cost Sensor Platforms Contribute to Air Quality
Monitoring and Exposure Estimates?”, Environment
International, vol. 99, pp. 293-302, 2017.
[6] M. Badura, P. Batog, A. Drzeniecka-Osiadacz, and P. Modzel,
”Optical Particulate Matter Sensors in PM2.5 Measurements in
Atmospheric Air”, 10th Conference on Interdisciplinary
Problems in Environmental Protection and Engineering EKO-
DOK, vol. 44, 2018.
[7] A. Thorpe, R. M. Harrison, “Sources and Properties of Non-
Exhaust Particulate Matter from Road Traffic: A Review”, Sci.
Total Environ., vol. 400, Issues 1-3, August 2008, pp. 270-282.
[8] T. Grigoratos, G. Martini, “Brake Wear Particle Emissions: A
Review”, Environ Sci Pollut Res, vol. 22, 2015, pp. 2491-2504.
[9] S. Pleshkova, A. Bekiarski, and K. Peeva, “Testing Thermal
Images Characteristics for Thermal Images Quality Estimation”,
Latest Trends on Systems – Volumes I & II, Proceedings of the
18th International Conference on Systems, (part of CSCC '14),
Santorini Island, Greece, July 17-21, 2014, pp. 251-256.
[10] http://yeetcen.youhaovip.com/products/p000006 - A4 CG Laser
PM2.5 Sensor description, accessed 10.2019
 Investigation of the Method of RMS Measurement Based
on Moving Averaging
Anna A. Kostina1, Plamen M. Tzvetkov 2 and Andrey N. Serov3
Abstract – Currently, the measurement method of root mean
square value (RMS), based on samples accumulation, is the most
popular. The advantages of this method include ease of
implementation, high accuracy and additional opportunities to
reduce the methodological component of the total measurement
error. Studies show that increasing measurement time also
contributes to a reduction in error. However, in a number of
measurement and automatic control problems, one of the main
requirements is to reduce the measurement time. This problem
can be solved by a method based on moving averaging, that
includes the advantages of the “classical” method, but is largely
devoid of its shortcomings. An algorithm of RMS measurement
for implementation on the basis of recursive and non-recursive
moving average filters is proposed. The influence of the
sinusoidal signal parameters on the RMS measurement error is
analyzed. An analytical expression is obtained that allows to
estimate the RMS measurement error from the signal
parameters and the number of samples. Approaches are
proposed for reducing the RMS measurement error when
applying the measurement method under consideration. It is
shown that the application of post-filtration of RMS
measurement results can reduce the final RMS measurement
error. The veracity of the analytical expressions obtained in the
paper is confirmed by the results of simulation at check points.
Keywords – Root mean square, Simulation, Digital filtration,
Measurement accuracy, Frequency deviation.
I. INTRODUCTION
The root mean square (RMS) of a periodic signal is the
parameter, which is determined by the expression [1]:
TA
1 The investigated method has several advantages. First of
x
2
X id = (t ) dt , (1)
TA 0
where TA – total observation (measurement) time, which is
selected proportional (multiple) of the real value of the input
signal period; x(t) – input periodical signal.
1Anna A. Kostina is with National Research University “Moscow
Power Engineering Institute”, Diagnostic Information Technologies
Department, Krasnokazarmennaya st. 14, 111250 Moscow, Russia,
E-mail: Kostina.AA.mpei@yandex.ru
2 Plamen M. Tzvetkov is with Technical University of Sofia,
Department of Electrical Measurements at Faculty of Automatics,
Bulgaria, Sofia 1000, 8 Kliment Ohridski blvd, E-mail:
Tzvetkov@tu-soia.bg
3 Andrey N. Serov is with National Research University “Moscow
Power Engineering Institute”, Diagnostic Information Technologies
Department, Krasnokazarmennaya st. 14, 111250 Moscow, Russia,
E-mail: SerovAN@mpei.ru
978-1-7281-7142-5/20/$31.00 ©2020 IEEE 
For a polyharmonic input signal,
M
x (t ) = X
i =1
i 2 sin(ω i t + α i ), (2)
expression (1) can be represented as follows [1]:
M
X id = X
i =1
2
i , (3)
where Xi – RMS of the i-th spectral component of the input
signal; ωi – angular frequency of the i-th spectral component;
αi – initial phase of the i-th spectral component; M – number
of considered spectral components.
For digital measuring devices, the instantaneous values of
the input signal come in the form of discrete samples by using
an analog-to-digital converter (ADC). For this reason, the
direct implementation of the definition of RMS (see formula
(1)) is impossible. There are a large number of digital methods
for RMS measurement, which are based on the analysis of the
signal in the time [2]-[8] or frequency domain [9]-[10]. Some
digital RMS measurement methods are designed to measure
sinusoidal input signals. These methods include the three-
sample method [8] and the method based on the application of
the second derivative of the signal [2]. The application of
these methods in conditions of nonsinusoidal signals results to
significant measurement errors. Currently, the operation
algorithms of most digital measuring instruments for RMS
measurement are based on the application of a samples
accumulation method (ASS) of the input signal [7].
all, this method is applicable for measurement sinusoidal and
polyharmonic signals. The small number of required
arithmetic operations ensures the simplicity of its practical
implementation. Due to the additional measurement of the
signal frequency [11]-[16] and the adjustment of the number
of accumulated samples (or the initial phase of the signal), it
is possible to significantly reduce the methodological error.
One of the parameters that influence the measurement error is
the measurement time. With increasing measurement time, the
methodological error tends to decrease. However, in a number
of problems of measurement and control theory, measurement
time is a critical parameter, which imposes serious limitations
on this measurement method.
Thus, the problem arises of finding high accuracy RMS
measurement methods, that requires a short measurement time
for the case of the real electrical networks. In the paper, a
samples accumulation method is analyzed. This method
satisfies the properties, which are describe above.
 II. THE RMS SAMPLES ACCUMULATION
MEASUREMENT METHOD

A. General Description

This method is based on the application of a zero-order

polynomial as a restoring function and the subsequent
implementation of the RMS determination for the
reconstructed signal (see expression (1)). Moreover, RMS can
be obtained by using the following formula [5]-[6]:

N −1
1
X=
N  x [i ] ,
i =0
2
(4)

where N – number of averaged samples; x[i] – signal samples.
The number of averaged samples is related to the
measurement time by the following relation:
N period.
TA = NTS = = mT ,
fS
where m – number of averaged input signal periods; T – input
period value; TS – sampling time; fS – sampling frequency.
Fig. 1. The dependence of the RMS measurement error (moving
averaging method) on the input signal frequency
where Tnom – nominal or measured value of the input signal
(5) The second approach to reducing the error is to ensure that
the lower equation of condition (7) is satisfied. If the initial
phase of the signal satisfies the condition:
α = 0.25π − 0.5( N − 1)ωTS ± 0.5πk , (10)

B. Measurement Error then the RMS measurement error for the case of a sinusoidal
signal is equal to zero.
In [5]-[6], a relation was obtained for calculation the RMS As in the previous case, since the frequency of the signal is
measurement error for the case of a sinusoidal input signal: unstable, and the initial phase of the signal is discrete (due to
the discrete nature of the signal samples), condition (10)
cos(( N − 1)ωTS + 2α ) sin(ωNTS ) cannot be fulfilled absolutely precisely. This results to the
δX = − , (6)
2 N sin(ωTS ) RMS measurement error.

where ω – input signal angular frequency (true value); α –

initial phase of the input signal.
The dependence of the error on the input signal frequency is
shown in Fig. 1. The figure shows that there are singular
points where the dependence takes on a zero.
Special points where the error takes zero values are
achieved when the following conditions are met:
 sin(ωNTS ) = 0,
 (7)
cos(( N − 1)ωTS + 2α ) = 0.
The fulfillment of the upper condition of expression (7) can
be ensured by adjusting the number of averaged samples:
k kT T A
N= = = , (8)
fTS TS TS
where k – integer value (in expression (8) is accepted k = m).
Since the number of processed samples is a discrete, and
the frequency of the input signal is unstable in time, relation
(8) is not absolutely accurate. This results to the appearance of
a methodological error. In practice, the following formula can
be used to determine the number of measured samples:
III. THE RMS MOVING AVERAGING
MEASUREMENT METHOD
A. General Description
As already described above, one of the drawbacks of the
samples accumulation method is a large measurement time.
Since N samples are required to obtain a measurement result,
the measurement time is equal to NTS = TA. That is, the
sampling time of the RMS measurement result is TA. The
update rate of the RMS measurement result can be increased.
It is achieved if for the calculating of the next measurement
result, the signal samples, which are used for the previous
RMS result. The sampling time of the RMS measurement
results is equal to the sampling time of the input signal if the
RMS measurement is performed for the N-1 previous samples
of the input signal relative to the current n-th sample. In this
case, the formula for RMS calculating takes the following
form:
N −1
1
X [n] =
N  x [n − i] .
i =0
2
(11)

N = round (TA ⋅ f S ) = round (m ⋅ Tnom ⋅ f S ), (9)


 As for the previous method, the number of averaged
samples can be obtained by expression (5). Expression (11)
can be implemented on the basis of a moving average non-
recursive filter of the N-th order. The RMS measurement
transducer can also be implemented by applying a recursive
structure:
x 2 [ n] x 2 [n − N ]
X [ n] = + X [ n − 1] − . (12)
N N
B. Measurement Error
The error of the n-th measured RMS sample of the
considered method can be calculated by using expression (6)
(provided that the initial phase of expression (6) is corrected,
the input signal is sinusoidal):
cos(( N − 1)ωTS + 2(α + nωTS )) sin(ωNTS )
δ X (n) = − , (13)
2 N sin(ωTS )
where α – initial phase of the input sinusoidal signal at the
first sample for the total measurement time.
Fig. 2 shows the time dependence of the RMS measurement
error with the application of expression (11). It can be seen
from Fig. 2 and expression (13) that the measurement error in
time domain characterized by a sinusoidal behavior with a
frequency equal to the real value of the signal frequency ω.
To reduce the measurement error, as well as for the samples
accumulation method, the following approaches can be
applied:
– increase of the total measurement time;
– adjustment of the number of averaged samples according
to the dependence (9);
– adjustment of the initial phase of the signal according to
the relation (10);
– application of additional low-pass filtration.
As can be seen from dependence (13), an increase in the
measurement time results to a decrease of the RMS relative
measurement error. For large measurement time values, the
maximum RMS error can be represented as:
Fig. 2. The dependence of the RMS measurement error for the
method of moving averaging on the number of measurement result

1
δ X ( n) ≤ − . (14)
2 N sin(ωTS )
However, an increase in the measurement time results to a
deterioration in the dynamic characteristics of the method.
Since the result of the RMS measurement with the use of this
method is the average value of the RMS for the considerate
measurement interval, changes in the RMS of the input signal
will more slowly manifest themselves on the measurement
result. For this reason, the application of this approach in the
considerate measurement method is ineffective, since it
neutralizes the advantages of this method compared to the
samples accumulation measurement method.
Using the adjustment of the number of samples allows to
reduce the value of the factor sin(ωNTS) in expression (13).
The finite error of frequency measurement and the discrete
nature of the number of averaged samples (parameter N)
interfere with ensuring the equality of this factor to zero. In
the case of using this approach, the relative RMS
measurement error can be represented as:
 1 
δ f + ω cos(( N − 1)ωTS + 2(α + nωTS ))
 2 N
δ X (n) = − . (15)
2 f S sin(ω TS )
where δf – frequency relative measurement error.
The application of adjustment of the initial phase for this
measurement method is not justified, since such adjustment
violates the measurement algorithm represented by
expressions (11) and (12).
The application of low-pass filtration of the RMS
measurement results makes it possible to reduce the influence
of a time-variable factor of expression (13) –
cos((N–1)ωTS+2(α+nωTS)). In this case, the applied low-pass
filter should be designed to suppress the non-dc spectral
components with the frequency of the input signal (frequency
of the variable component of expression (13)). In addition, the
constant (DC) component of the RMS measurement result
(useful signal) should be transmitted to the filter output
unchanged. The RMS relative measurement error for this
approach takes the form:
H (ω ) cos(( N − 1)ωTS + 2(α + nωTS ) + ϕ (ω ))
δ X (n) = − , (16)
2 N sin(ωTS ) sin −1 (ωNTS )
where H(ω) – magnitude response of the applied low-pass
filter for the frequency (true value) of the input signal ω; φ(ω)
– phase response of the applied low-pass filter for the
frequency (true value) of the input signal ω.
IV. SIMULATION MODELING
Simulation mathematical modeling was performed in
Simulink software package. By Simulink a simulation model
of the RMS measurement transducer was developed based on
the method of moving averaging of samples (MA) and based
on the samples accumulation measurement method (ASS). At
the output of the measurement converter, a digital IIR filter is
 TABLE 1. REFERENCES
SIMULATION RESULTS FOR RMS MEASUREMENT TRANSDUCERS OF
DIFFERENT TYPES (ERROR IS REPRESENTED IN PERCENT)
[1] A. E. Emanuel “Powers in nonsinusoidal situations - a review of
RMS MC Frequency relative deviation, δf , % definitions and physical meaning,” IEEE Trans. on Power
type -5.0 -1.0 0.0 1.0 5.0 Delivery, vol.: 5, issue: 3, 1990, pp. 1377–1389.
ASS 2.6 0.51 0.0 0.50 2.4 [2] A.M. Zayezdny; Y. Adler; I. Druckmann “Short time
measurement of frequency and amplitude in the presence of
MA 2.6 0.51 0.0 0.50 2.4
noise,” IEEE Transactions on Instrumentation and
MA+BF 1.8 10-2 9.2 10-4 8.0 10-5 8.8 10-4 1.5 10-2 Measurement, vol.: 41, issue: 3, 1992.
MA+CF 8.7 10-2 0.10 0.10 0.10 9.0 10-2 [3] G. E. Mog; E. P. Ribeiro. “One cycle AC RMS calculations for
ASS+SA 0.11 5.0 10 -3
0.0 5.0 10 -3
0.13 power quality monitoring under frequency deviation,” 11th
International Conference on Harmonics and Quality of Power,
MA+SA 0.11 5.0 10-3 0.0 5.0 10-3 0.13
2004, pp. 700-705.
MA+BF+SA 1.6 10-4 8.0 10-5 8.0 10-5 8.2 10-5 1.6 10-4 [4] Predrag B. Petrovic “Root-mean-square measurement of
MA+CF+SA 0.10 0.10 0.10 0.10 0.10 periodic, band-limited signals”, IEEE International
Instrumentation and Measurement Technology Conference
applied, constructed according to the analog prototype of 4-th Proceedings, 2012, pp. 323-327.
[5] Fan Wang; M. H. J. Bollen “Frequency-response characteristics
order Butterworth filter (BF) or 3-rd order Chebyshev filter
and error estimation in RMS measurement,” IEEE Transactions
(CF). The developed model provides for the possibility of on Power Delivery, vol.: 19, issue: 4, 2004, pp. 1569-1578.
additional adjustment of the number of averaged samples [6] Andrey N. Serov, Nikolay A. Serov, Petr K. Makarychev.
(SA) according to the frequency measurement result. The "Comparative Analysis of the RMS Measurement Methods
simulation results are shown in Table 1. When performing the Based On the Averaging of the Squares of Samples", 2020 30th
simulation, a sinusoidal signal with a unit amplitude and a International Conference Radioelektronika
zero initial phase was applied. The sampling frequency is (RADIOELEKTRONIKA), 2020.
equal to 10 kHz, the measurement time of one frame is equal [7] A. Serov, A. Shatokhin, A. Novitskiy, D. Westermann
to 0.02 seconds. The total simulation time is 2 seconds. When “Investigation of the method of RMS measuring based on the
digital filtration of the square of samples,” 2018 18th
designing the output post-filters, the following characteristics
International Conference on Harmonics and Quality of Power
was applied: passband limit – 10 Hz; stopband limit is 45 Hz; (ICHQP), 2018, pp. 1-6.
maximum passband deviation – 0.1 %; minimum stopband [8] V.S. Melentiev; V.I. Batishchev; A.N. Kamyshnikova; D.V.
rejection is 0.01. Rudakov “An improvement in the methods used for the
measurement of the integrated characteristics of harmonic
signals,” Measurement Techniques, vol.: 54, no.: 4, 2011, pp.:
V. CONCLUSION 407-411.
[9] S.N. Mikhalin, V.M. Gevorkyan “The problems of digital
As a result of the research done, the following results were processing of signals in the system of automated power quality
obtained: control and accounting quantity of electricity (ASQAE),” MPEI
1. The method of moving averaging of samples allows to Vestnik, no.: 1, 2005, pp. 86-92.
get the RMS measurement result in less time compared to the [10] Andrzej Konstanty Muciek “A Method for Precise RMS
"classical approach". Measurements of Periodic Signals by Reconstruction Technique
With Correction,” IEEE Transactions on Instrumentation
2. An analytical expressions (13)-(16) has been obtained to
and Measurement, vol.: 56, issue: 2, pp. 513-516.
take into account the measurement error of the proposed [11] B. Boashash “Estimating and interpreting the instantaneous
method. The application of these expressions makes it frequency of a signal. I. Fundamentals,” Proceedings of the
possible to calculate the maximum value of the RMS IEEE, vol.: 80, issue: 4, 1992.
measurement error for the given parameters of the input signal [12] B. Boashash “Estimating and interpreting the instantaneous
and RMS measurement transducer. In addition, the inverse frequency of a signal. II. Algorithms and applications,”
problem can be solved: for a given maximum value of the Proceedings of the IEEE, vol.: 80, issue: 4, 1992.
RMS measurement error, the maximum deviation of the signal [13] M.M. Begovic, P.M. Djuric, S. Dunlap and A.G. Phadke,
frequency can be estimated (which is equal to the maximum “Frequency tracking in power networks in the presence of
harmonics,” IEEE Transactions on Power Delivery, vol.:
frequency measurement error in the case of adjusting the
8, issue: 2, 1993, pp. 480-486.
number of samples). [14] A. Serov “Frequency estimation methods for stationary signals,”
3. The approaches to reduce the RMS measurement error by 2017 International Conference on Industrial Engineering,
the proposed method are developed. The approaches consist in Applications and Manufacturing (ICIEAM), 2017, pp. 1-6.
adjusting the number of averaged input signal samples and [15] D. Agrez, “Power system frequency estimation in the shortened
post-filtration of the RMS measurement results. It was shown measurement time,” Proceedings of the 18th IEEE
that adjusting the number of averaged samples and low-pass Instrumentation and Measurement Technology Conference,
post-filtration can significantly (up to 103 times) reduce the vol.: 2, 2001, pp. 1094-1098.
RMS measurement error. [16] P. Petrovic, S. Marjanovic and M.R. Stevanovic, “New
algorithm for measuring 50160 Hz AC values based on the
4. By application of Simulink, a simulation model of RMS
usage of slow A/D converters,” IEEE Transactions on
measurement transducer is developed. Instrumentation and Measurement, vol.: 49, issue: 1, 2000, pp.
166-171.


 Monitoring of Air Pollution in Skopje in order to
Determine Possible Causes
Vladimir Mijakovski1 and Monika Lutovska2
Abstract – Brief review of few major air pollutants of Skopje is
presented in this article. State automatic monitoring system for
ambient air quality measures up to 7 air pollutants is presented.
It consists of network of 17 fixed and 1 mobile monitoring
stations installed throughout the country. Performed analysis of
pollutants’ inventory reveals that suspended particles (PM10 and
PM2.5) are the most critical air pollutants in Skopje in the last
couple of years.
Keywords – Monitoring, Causes, Air pollution, Skopje.
I. INTRODUCTION
The use of oil and oil derivatives, traffic, increasing
industrialization and intensification of production inevitably
lead to an increase in the concentration of harmful substances
in the air. Although in traces, their presence has been shown
to have a significant detrimental effect on human health,
biosphere and other material goods. There are substances in
the air, such as: nitrogen oxides (NOx), sulphur dioxide (SO2),
ammonia (NH3), carbon monoxide, particulate matters, heavy
metals, hard degradable organic compounds, etc., that have a
direct or indirect negative impact if their presence is in
concentrations higher than normal.
Nitrogen monoxide (NO) and nitrogen dioxide (NO2), or
general NОx, are the most important nitrogen oxides, that
occur as an air pollutant. NO2 reactive gas is formed by
oxidation of nitrogen monoxide (NO). A major source of NO
and NO2 is high temperature combustion process (processes
that are carried out in motor vehicles and power plants). The
amount of NOx is increased in winter due to intensive use of
fossil fuels. The largest amounts of nitrogen oxide emissions
in Republic of North Macedonia are emitted during the
production of electricity and thermal energy (48%), due to the
dominant electricity production with coal as a fuel. Traffic
emissions, also, have large share (38%) of total emissions of
this pollutant.
Sulphur dioxide, SO2, is a colourless, toxic gas with sharp
and irritant odour, and acidic properties. Its toxic impact on
humans occurs at mass concentration in the air of about 6
mg/m3 or more, causing serious health problems. Today,
sulphur dioxide, SO2, is considered as one of the main
pollutants in the atmosphere from anthropogenic sources, and
thus intensive measures are taken to reduce its emissions.
1Vladimir Mijakovski is with the University “St. Kliment
Ohridski”, Faculty of Technical Sciences- Bitola 7000, N.
Macedonia, E-mail: vladimir.mijakovski@tfb.uklo.edu.mk
2Monika Lutovska is with the University “Mother Theresa”,
Faculty of Technical Sciences - Skopje1000, N. Macedonia, E-mail:
monika.lutovska@unt.edu.mk
978-1-7281-7142-5/20/$31.00 ©2020 IEEE 
Considering that electricity production is the main source for
SO2 emissions, its emissions trends vary and depend highly on
coal consumption. There are many industries that emit
significant quantities of SO2 in the ambient air: the oil
industry emits SO2 or H2S during refining of petroleum
products, electricity production plants that use coal with high
concentrations of sulphur, sulphide smelters, installations for
production of paper and pulp. In 2014, the estimated national
SOx emissions amounted to 83.141 kilotons, where a key and
dominant source with 92%, are the combustion processes
from fuels (coal and oil) in power generation.
Suspended particles or particular matter (PM), as one of the
most common pollutants in the air is a mixture of solid and
liquid particles, suspended in the air with wide range of sizes
and chemical structure. They can originate from natural
sources (sea salt, desert areas dust, wildfires, volcanic ash
pollen) and anthropogenic sources (industrial processes,
combustion of fossil and biofuels - vehicles, power plants and
households, transport and waste incineration). The most
significant share in particles emission (PM2.5 – 51%, PM10 –
35%), according to the conducted inventory of suspended
particles, belongs to domestic heating, especially due to
incomplete burning wood in old stoves. Recent studies, [1]
have shown that key sources of suspended particles emission
are heat and electricity production processes (PM2.5 -25%,
PM10 - 22%), as well as production processes especially in
the area of metallurgy. The reduction in emissions in the
recent years, is due to lower consumption of firewood in
winter period.
II. EMISSION MEASUREMENTS, LIMIT VALUES AND
STANDARDS IN USE
Republic of North Macedonia has an emission inventory
reporting obligation, [2], towards the Convention on
transboundary air pollution and its eight protocols as well as
the European Environmental Agency (EEA). In accordance
with the Convention and its protocols, our country is required
to annually report data on emissions of the main air pollutants
covered by the Convention. According to EU legislation, limit
values of air pollution levels are the strictest rules. In a case of
exceeding the limit values, measures to reduce the levels of
pollutant concentrations must be defined and enforced by the
authorities. In addition, for the protection of human health and
vegetation, critical levels have been defined, Table I.
In the Republic of North Macedonia, the monitoring of
ambient air quality is performed by the Ministry of
Environment and Physical Planning, which manages the State
Automatic Air Quality System, as well as the Institute of
 TABLE I
LIMIT VALUES OF AIR POLLUTANTS
Allowed
Averaging Limit number of
Pollutant
period values exceedances in
a year
Sulphur dioxide, 1 hour 350 μg/m3 24
SO2 24 hours 125 μg/m3 3
Nitrogen dioxide 1 hour 200 μg/m3 18
NO2 1 year 40 μg/m3 - Fig. 2. Total national SO2 emissions for the period 1990-2015
PM10 suspended 24 hours 50 μg/m3 35
particles 1 year 40 μg/m3 - In the period from 2013 to 2014 emissions remained quite
stable (-1%), and after 2014 remained at the same level due to
Public Health (IPH) with the Centres for Public Health in the fact that no major changes were carried out. In 2016, SO2
Skopje and Veles.Additionally, the air quality monitoring is emissions in Skopje are also the highest in the month of
performed by certain installations that have an obligation in December. Maximal hourly concentrations of SO2 reaching up
accordance with the requirements of the IPPC permit. to 52 μg/m3, were registered at monitoring stations in
The Ministry of Environment and Physical Planning Miladinovci at the end of the second half of December. Fig. 3
manages the State Automatic Monitoring System for Ambient shows the hourly SO2 concentrations for Skopje
Air Quality, which consists of 17 fixed and 1 mobile agglomeration region in December 2016 [4, 5].
monitoring station throughout the country, located on
following places: 5 measuring stations in Skopje, 2 measuring
stations in Bitola, and one measuring station in Veles,
Municipality of Ilinden, Kichevo, Kumanovo, Kochani,
Tetovo, Kavadarci, Gostivar, Strumica and in the village of
Lazaropole, Fig. 1 [3].

Fig. 3. Hourly concentrations of SO2 for Skopje agglomeration

region in December 2016

Fig. 1. State automatic ambient air quality monitoring system
Automatic air quality monitoring stations monitor the
following pollutants: SO2, NO2, CO, ozone (O3), PM10,
PM2.5, Benzene, toluene, ethylbenzene, and xylenes (as
BTEX).
Within this paper, a comparison of the emissions of SO2
and PM10 at the level of the entire country and especially for
Skopje was made.
SO2 emissions
Total national SO2 emissions amounted to 110 kt in 1990.
In 2014, emissions decreased for 24% due to decreasing
emissions from public electricity and heat production, and
amounted to total of 83 kt. In the years 2009 and 2011
emission peaked compared to 2010, due to higher
consumption of coal by the major thermal power plant REK
Bitola. Shorter operation periods of the second by installed
capacity thermal power plant REK Oslomej in the period 2012
– 2013 decreased coal consumption up to 60%, thus reducing
SO2 emissions.
PM10 emissions
Main emission sources of PM10 in Macedonia for 2014 were
residential heating with a share of 36% in total PM10
emissions, industrial processes and product use - ferroalloys
production with 33%, energy sector with 20% and agriculture
with 7%. Fugitive emissions and waste are minor sources of
PM10 emissions. Total national PM10 emissions in 1990
amounted to 48 kt, and in 2014 emissions decreased by 32%
reaching 33 kt. Declining emissions from industrial processes
are the main reason for this decrease. Low emissions for the
years 2001, 2002 and 2009 compared to the other years are
related to low emissions from ferroalloys production due to
shorter operating hours and produced quantity of ferrosilicon
(80-90% decrease compared to 2014). Emissions reduction of
24% was recorded from 2013 to 2014, as a result of a decline
in ferroalloys production and decrease in biomass
consumption from residential heating. Total national PM10
emissions for the period 1990-2015 are shown in Fig. 4.
Measurements in Skopje showed that the average annual
concentration of PM10 was 72,82 μg/m3 in the municipality of
Centar and 57,06 μg/m3 in the municipality of Karposh. It is
noted that the highest measured concentrations of PM10 in
Skopje were recorded in winter, [4, 5].



Fig. 4. Total national PM10 emissions for the period 1990-2015
Fig. 5. Hourly concentrations of PM10 for Skopje agglomeration
region in December 2016
TABLE II
REVIEW OF МКС STANDARDS FOR MEASUREMENT OF POLLUTING
SUBSTANCES IN THE AIR
Substa Measuring method
nce
МКС EN 14212:2013 - Ambient air - Standard
SO2 method for the measurement of the concentration of
sulphur dioxide by ultraviolet fluorescence
МКС EN 14211:2013 - Ambient air - Standard
NO,
method for the measurement of the concentration of
NO2,
NOx
nitrogen dioxide and nitrogen monoxide by
chemiluminescence
МКС EN 12341:2014 - Ambient air - Standard
gravimetric measurement method for the
PM10
determination of the PM10 or PM2.5 mass
concentration of suspended particulate matter
МКС EN 12341:2014 - Ambient air - Standard
gravimetric measurement method for the
PM2.5
determination of the PM10 or PM2.5 mass
concentration of suspended particulate matter
МКС EN 14626:2013 - Ambient air - Standard
method for the measurement of the concentration of
CO
carbon monoxide by non-dispersive infrared
spectroscopy
МКС EN 14625:2013 - Ambient air - Standard
O3 method for the measurement of the concentration of
ozone by ultraviolet photometry
МКС EN 14662-3:2016 - Ambient air - Standard
BTEX
method for the measurement of benzene
concentrations - Part 3: Automated pumped sampling
with in situ gas chromatography

According to measurements for 2016 in Skopje, highest
PM10 emissions were observed in the month of December.
Enormous hourly concentrations of PM10 that reached over
1000 μg/m3, are registered at monitoring stations Rectorate
and Lisiche, during the second half of December. Fig. 5 shows
the hourly concentrations of PM10 for December 2016.
A detailed description of the methods for measuring air
pollutants as well as quality control of measurements are
given in the European CEN standards, which are accepted in
the Republic of North Macedonia. The following table shows
the MKC EN standards for measuring the concentrations of
pollutants in the air, Table II.
III. STATE AUTOMATIC MONITORING SYSTEM FOR
AMBIENT AIR QUALITY (SAMSAAQ)
SAMSAAQ consists of: Air quality calibration laboratory,
state air quality monitoring network and air quality data
management system [3].
Air quality calibration laboratory maintains the traceability
of the reference gas standards towards conducted
measurements through calibration of the measuring
instruments; Participates in inter laboratory comparison
exercises organized by EC/Joint Research Center and the
WHO; Approves measurement systems (methods, equipment,
networks and laboratories); Ensures the accuracy of
measurements; Maintains accreditation according to ISO/IEC
17025 for the reference methods and Keeps good contacts and
passes the information from the EC, AQUILA, EC-working
groups to the personnel at the local networks.
As mentioned earlier in the article, the air quality
monitoring system consists of 17 automatic monitoring
stations, 6 low volume samplers and 4 high volume samplers.
Each automatic measuring station measures environmental
and meteorological parameters with installed data acquisition
set. The following measurement instruments are used:
For the measurement of SO2:
• TEI 43C SO2 analyzer, from manufacturer Thermo
Environmental Instruments;
• TI 100SO2 analyzer, from manufacturer Teledyne API;
• ML2050 SO2analyzer, from manufacturer Casella
monitor.
For the measurement of NO, NO2 and NOx:
• TEI 42C NO-NO2-NOx analyzer, from Thermo
Environmental Instruments;
• T200 NO-NO2-NOx analyzer, manufactured by
Teledyne API;
• ML2041 NO-NO2-NOx analyzer, from manufacturer
Casella monitor.
For the measurement of CO:
• TEI 48C CO analyzer, from manufacturer Thermo
Environmental Instruments;
• T300 CO analyzer, manufactured by Teledyne API;
• ML 2030 CO analyzer, manufactured by Casella
monitor.
 For the measurement of O3:
• TEI 49C O3 analyzer, from manufacturer Thermo
Environmental Instruments;
• T400 O3 analyzer, manufactured by Teledyne API;
• ML 2010 O3 analyzer, manufactured by Casella
monitor.
For the measurement of PM10/PM2.5:
• FH 62 I-R, PM2.5 and PM10 Automated Particulate
Monitor analyzer, manufactured by Thermo ESM
Andersen;
• Sharp 5030, PM2.5 and PM10 Automated Particulate
Monitor Analyzer, by Thermo Fisher Scientific;
• BAM-1020 Continuous Particulate Monitor Analyzer,
from Met One Instruments, Inc.;
• Grimm 180, PM10 / PM2.5 Environmental Dust
Monitor Analyzer, by Grimm Aerosol Technik;
• MicroPNS LVS 16 Sequential Low Volume Sampler
sequential low volume sampler, from manufacturer
Umwelttechnik MCZ GmbH;
• Low Volume Sampler LVS 3.1 low volume sampler,
from ComdeDerenda;
• LVS3 / MVS6 low volume sampler, from
manufacturer Sven LeckelIngenieurbüro GmbH.
The schematics of the air quality data management system
is shown in Fig. 6.
Fig. 6. Schematic diagram of the State Automatic Monitoring System
for Air Quality
AIRVIRO is a web-based system for Air Quality
Management developed by SMHI and Apertum IT AB. It is
the core of the SAMSAAQ. AIRVIRO is an integrated system
for time series data handling, emission inventories and
dispersion modelling. It is a client / server web integrated
system for Air Quality Management, including time series
data handling, emission inventories and dispersion modelling.
Validation and verification of the measured data in the
system is done on a daily basis. Based on the validated data, a
daily, monthly and annual report is prepared and published on
the web portal for air quality: http://air.moepp.gov.mk. The
current situation with the ambient air quality in the Republic
of North Macedonia can be monitored in real time on the
same website.
By collecting, processing and utilizing meteorological
data, AIRVIRO has possibility to display wind rose needed
for dispersion modeling.

IV. MEASURES FOR REDUCTION OF POLLUTANTS
Having in mind that the Earth is becoming an urban planet
and more than 60% of the global population now lives in
cities, it is expected that most people will suffer the most
severe consequences from pollution in urban areas. In the city
of Skopje lives approximately one third of the total population
of Republic of Macedonia and the influx of people from the
region create substantial daily migrations. Some activities
have been undertaken to implement medium and long-term
measures outlined in the National Plan for the ambient air
protection, [6]. To improve air quality at the level of Skopje
region agglomeration, draft versions of improvement plans are
prepared during 2015 and in the first half of 2016. Increasing
the share of renewable energy sources in total energy
consumption has led to reduction of pollutants emission. In
this regard, activities listed in adjustment permits with
operational plans of heat production installations are
implemented. A study for the gasification of the Skopje region
was also prepared. At local level, gasification network is
extended with introducing gasification facilities in housing
and schools in Skopje. Most of the measures are implemented
in the capital because of the highest traffic frequency. Out of
this reason, from 1 January 2015, import of vehicles with at
least Euro 3 standard is permitted. In the municipality Centar,
a measure for speed limit in certain streets even up to 30 km/h
is conducted. Namely, a special traffic regime (using the ring
road) for trucks whose final destination is not the city of
Skopje, is introduced during the winter period. Also, as an
experience taken from world capitals such as London, Vienna,
Klagenfurt, Linz, Helsinki, Brunek, Stuttgart, Stockholm,
Gothenburg and al., application of CaMg (CH3COO)4
(calcium magnesium acetate) on the roads in the city during
the winter season has started. It is used to control and reduce
concentrations of PM10 and PM2.5 suspended particles.
V. CONCLUSION
According to Eurostat data, urban areas in the EU are
responsible for 80% of energy consumption [6]. Given the
current performed inventory of pollutants in individual
sectors/activities in Skopje, it is evident that electricity and
heat production, burning firewood for households heating,
transport and industrial processes have the greatest
contribution to the air pollution. Namely, suspended particles
are the most critical pollutant in all previous years. Thus,
exceedances above the limit values of suspended particles of
size up to 10 μm repeatedly higher than the daily average limit
value at all measurement locations especially in winter period
are recorded. The national government, together with local
and regional authorities, shares responsibility for combating
global warming and enormous greenhouse gas emissions,
through efficient energy consumption and the use of
renewable energy sources [6]. In this regards, certain
measures for reduction of pollutants emission have already
been introduced in the last few years, [7]. Although the effects
from implementation of these measures are difficult to be
analysed since their utilization lasted for several days during
the entire winter period, it is especially important that their
implementation is carried out by cooperation among all the
relevant factors between the city and the municipalities.
Therefore, it seems that the primary activity to which the most
attention should be paid is to raise public awareness,
education and change the behaviour/habits of citizens. In
addition, increasing of green areas, modernizing and
improving efficiency in energy infrastructure, increasing the
use of renewable in electricity production, increasing the use
of natural gas [8], as well as the development of a better
traffic management plan, are actions that should be the
primary goals on the agenda for tackling air pollution [4].
REFERENCES
[1] Report of Processed Data on Environmental Quality in 2015,
Ministry of Environment and Physical Planning of Republic of
Macedonia.

[2] Republic of Macedonia/Informative Inventory Report 1990 –
2015 (under the Convention on Long Range Transboundary Air
Pollution-CLRTAP), Ministry of Environment and Physical
Planning of the Republic of Macedonia, Skopje, 2017.
[3] Website: http://airquality.moepp.gov.mk/.
[4] Strategy on Climate Changes - Resistant Skopje 2017,
Department of environmental protection and the nature of the
city of Skopje
[5] Website: http://www.skopjecentar.ekoinformatika.mk/
[6] Sustainable Energy Action Plan for development of the City of
Skopje, April 2011
[7] National Plan for the Protection of the Ambient Air in the
Republic of Macedonia for the Period 2013 – 2018, Government
of the Republic of Macedonia, Skopje, December 2012
[8] Strategy for Energy Development in the Republic of Macedonia
until 2030, Ministry of economy, 2010.
 Load Disturbance Rejection in Hydro Turbine Control
Based on Wiener Filtering
Teofana Puleva1 and Elena Haralanova2
Abstract – In this paper the task of load disturbance rejection modeled as a stochastic process [1] with an exponential
in a hydro generator speed and power control is considered. The distribution described by the PDF function
load change is described as a stochastic process. It is shown that
the problem of random load disturbance rejection can be pv ( x ) = μ e − μ τ . (1)
formulated as an optimal controller design in sense of minimum
mean-square error (Wiener filter design). Controller structure The reciprocal value of µ has a sense of a mean value of
based on dynamic compensation as well as on direct design with time intervals between successive load rate changes. The
rational transfer function by taking into account the plant non- covariance and the spectral density function of this stochastic
minimal phase characteristics is obtained.
process are
Keywords – Load rejection, Water turbine speed control, 2σ v2 μ
Wiener filter design. Kv (τ ) = σ 2 e
−μ τ
, Sv (ω ) = , (2)
ω2 + μ 2
I. INTRODUCTION where σ v2 is the load standard deviation. The load spectral
density function can be presented by the following expression
The problem of random load disturbance rejection in a
hydro generator operation in power limited energy system can 2σ v2 μ
be based on the stochastic approach to the control system. The Sv (ω ) = .
task of load disturbance rejection is formulated as an optimal ω 2 (ω 2 + μ 2 )
controller design problem in sense of minimum mean-square The turbine-generator system dynamic behavior in presence
error (Wiener filter design) [1-2]. In [3] a robust technique of load disturbance can be presented by the equation
based on weighted multichannel Wiener filter for speech
processing is considered. It takes into account the speech y ( s ) = Wuy ( s )u ( s ) + Wvy ( s )v( s ) . (3)
distortion due to signal model errors explicitly in its design
criterion. The investigations in this paper can be considered as The transfer functions with respect to the control signal u and
a continuation of previous research works related to multiple to the load disturbance v in (3) are:
model adaptive controller design based on Wiener filtering
[4]. The efforts are focused on the determination of the k (−Tw s + 1)
simplified single controller structure by more precise Wuy ( s ) = , Wvy (s) = kv , (4)
(Tm s + 1)(0.5Tw s + 1) Tm s + 1
modelling of the stochastic load variation. This approach can
be applied to other control design problems describing the
where Tw is the water column time constant; Tm - mechanical
local network operation of renewable energy sources.
time constant; k and kv - proportional coefficients with
II. CONTROLLER DESIGN BASED ON WIENER respect to the control signal and to the disturbance
FILTER respectively.
In Fig. 1 the block diagram of the turbine speed control
system in presence of load disturbance and measurement
The dynamic behavior of a system in which a hydro turbine
noise is shown. The system output is compared to the desired
unit is connected to a power limited energy system is
reference transformation W0 ( s ) .
investigated. The main function of the hydro turbine governor
in this operation mode is to compensate the unpredictable load
variation in order to keep the frequency on its nominal value. Wvy ( s )
The load rate change can be described as a stochastic process
that takes random values in particular time intervals. The time
Wc (s ) Wuy ( s)
intervals between successive load rate changes can be

Teofana Puleva is with Technical University of Sofia, Faculty of

Automatics, Kl. Ohridski St. 8, 1000 Sofia, Bulgaria, E-mail:
tpuleva@tu-sofia.bg W0 (s )
Elena Haralanova is with Technical University of Sofia, Faculty of
Automatics, Kl. Ohridski St. 8, 1000 Sofia, Bulgaria, E-mail: Fig. 1. Load disturbance in speed control of the hydro generator
tpuleva@tu-sofia.bg local network operation

978-1-7281-7142-5/20/$31.00 ©2020 European Union 

 The spectral density function of the random signal s is where Sη = N 2 , Ds = 2Tl kv2σ v2 .
For a given spectral density function of the random signal s,
2 2Tl kv2σ v2
Ss (ω ) = Wvy ( jω ) Sv (ω ) = 2 , (5) the optimal system transfer function is
ω (1 + Tm2ω 2 )(1 + Tl 2ω 2 )
c1 s 2 + c2 s + c3
Wopt ( s) = , (11)
where Tl = μ .−1
b0 s + b1 s 2 + b2 s + b3
3

After block diagram algebra in Fig. 1 the topology of the

diagram shown in Fig. 2 can be obtained. The controller will c1 = b1 − N (Tm + Tl ) , c2 = b2 − N , c3 = kσ v 2Tl . (12)
be designed for small perturbations around the equilibrium
point r=0. If the optimal system transfer function is already determined
W = Wopt , the controller transfer function can be obtained:
WWvyvy( s( )s )
W
Wc ( s) = , (13)
Wuy ( s )
Wpl (1 − W )
Wvy ( s ) Wc ( s )
where W pl is the plant transfer function ( Wpl = Wuy ). The
turbine speed and power controller design takes into account
Fig. 2. Modified block diagram
special features of the plant related to its non-minimal phase
characteristic. The Eq. (13) can not be applied to controller
The error spectral density function from the general Wiener
design when the plant exhibits non-minimal phase behaviour.
problem is
In this case another controller structure and design algorithm
2
Sε = 1 − W Ss + W Sη ,
2
(6) can be developed. One approach for solving this problem is to
apply direct design method based on rational transfer
where W is the closed loop transfer function. Therefore the functions description by taking into account plant non-
problem of random load disturbance rejection in a hydro minimal phase characteristics [8-9] when the desired closed
generator operation in power limited energy system can be loop trasfer function is already obtained. The control signal u
formulated as an optimal controller design problem in sense of is formed by the following polynomial expression:
minimum mean-square error (Wiener filter design). The linear
Ru = Sr − Ty , (14)
least-squares filter performs optimal separation of the
measured signal and additive noise, based on their spectral where R, S и T are polinomials of s satisfying the
characteristics. The optimal control system transfer function corresponding degree requirements, r and y are the reference
in sense of minimum mean-square error is [2, 5-7]. and the output signal respectively. The polinomials R, S and T
1  W0 S s  can be determined by solving the Diophantine equation
Wopt ( s ) = , ϕϕ = S s + Sη , Ssη = 0 , (7)
ϕ  ϕ  + ( AR + BT ) y ( s ) = BSr ( s ) , (15)
where operations of factorization and separation have to be where the polynomials B and A are the plant transfer function
carried out in order to obtain a causal solution. In case numerator and denominator, and the ratio y ( s ) / r ( s ) is the
W0 ( s ) = 1 and in presence of measurement noise modeled as a
desired transfer function Wopt . In presence of non-minimal
white noise, the separation in Eq. (7) may be avoided. Hence
the Wiener filter design can be carried out by the formula phase characteristics, the numerator is presented as a product
of two components B( s ) = B + ( s ) B − ( s ) , where the non-
1
Wopt ( s ) =
ϕ
(ϕ − )
Sη , Sη = N 2 . (8) minimal phase term is B − ( s ) . The optimal transfer function
in case of functional constraints is [1-2, 7]
The factorization of Eq. (8) leads to the following
expression: E W0 ES s 
Wopt ( s) =   , (16)
Eϕ  Eϕ  +
b s3 + b1 s 2 + b2 s + b3
ϕ= 0 . (9)
s(Tl s + 1)(Tm s + 1) E = −Tw s + 1 , E = Tw s + 1 , (17)
The coefficients bi , i = 0,3 are calculated from the where the non-minimal fraction E is analytical function in the
following relationships: left half-plane and E is analytical in the right half-plane.
The factorization of the input signal spectral density
2 2 2 2
(
b0 = NTmTl , b1 = N Tm + Tl + 2 NTmTl b2 , ) (10) function is ϕϕ = S s + Sη , where ϕ is given by Eq. (9).
The separation determines the causal term which can be
b22 = N 2 + 2 Ds , b3 = kσ v 2Tl , written as:


 W0 ESs  m1 s 2 + m2 s + m3
  = , (18) 10
Bode Diagram

 Eϕ  + s(Tl s + 1)(Tm s + 1) 0

Magnitude (dB)
-10

m1 = d 0TmTl + d1Tl + d 2Tm , m2 = d 0 (Tm + Tl ) + d1 + d 2 , m3 = d 0 ,

-20

-30

where the coefficients d i , i = 0,3 depend on the plant and the -40

-50

load parameters 45

 1  0

Phase (deg)
jω ESs  jω +  ES s
Tm 
d0 = lim , d1 = lim 
-45
,
jω = 0 Eϕ jω =−
1 Eϕ -90
Tm 10
-4 -3
10 10
-2
10
-1
10
0
10
1
10
2

Frequency (rad/sec)
 1
 jω +  ES s (a)
T
d 2 = lim  l  . Bode Diagram

jω =−
1 Eϕ 10

Tl 0

The optimal transfer function with functional constraints is

Magnitude (dB)
-10

-20

(m1 s 2 + m2 s + m3 )(−Tw s + 1)
-30

Wopt ( s) = . (19) -40

(b0 s 3 + b1 s 2 + b2 s + b3 )(Tw s + 1) -50

45

If the transfer function of the optimal system is already 0

Phase (deg)
determined W = Wopt , the controller transfer function can be -45

obtained in accordance with Eq. (13) and can be described by -90

10
-4
10
-3
10
-2
10
-1
10
0
10
1 2
10

the following general expression Frequency (rad/sec)

(b)
p s 3 + p1 s 2 + p2 s + p3 Fig. 3. Optimal system frequency response under (a) variable load
Wc ( s) = 0 . (20) variance; (b) variable load time constant
s(Tl s + 1)(Tm s + 1)
Bode Diagram
10

In general the controller coefficients pi , i = 0,3 depend on 0

Magnitude (dB)

the variable load characteristics descibed by load variance σ v2

-10

-20

and load time constant Tl . -30

-40

-50
90

III. EXPERIMENTAL RESULTS 0

Phase (deg)

-90

The hydro turbine model is described by the transfer -180

function Eq. (4) with k = 1.25 , Tm = 12.8 s and Tw = 1.18 s . -270

10
-4
10
-3
10
-2
10
-1
10
0
10
1
10
2

Frequency (rad/sec)
The shaping filter parameters are: load time constant (a)
Tl = 30 s and coefficient k v = 1.25 . The variance fluctuation
Bode Diagram
interval is σ v2 = [0.1,1.1] and the load time constant interval is 10

0
Tl = [20, 65] s .
Magnitude (dB)

-10

The optimal transfer functions are determined in both cases: -20

without functional constraints defined by Eq. (11) and by -30

taking into account non-minimal phase plant characteristics -40

90
(Eq. 19). The experiment consists of random varying load
conditions. The idea is to present the system performance in 0
Phase (deg)

the whole operational range. Although, hydro turbine -90

generators typically work long time intervals under almost -180

constant load conditions, such sudden changes in the load -270

10
-4
10
-3
10
-2
10
-1
10
0
10
1 2
10

conditions can accrue. The frequency responses under Frequency (rad/sec)

variable load characteristics are obtained. In Figs. 3 and 4 the (b)

optimal system frequency responses under variable load Fig. 4. Optimal system frequency response with functional
variance and variable load time constant are shown. constraints under (a) variable load variance; (b) variable load time
constant
It can be noticed the optimal system frequency response
ensures a zero steady-state error in the low frequency range up
The load characteristics variation reflects on the natural
to 0.1 rad/s that corresponds to the load signal frequency
frequency in a relatively small range less than half of decade.
band.


 Load variation tracking Frequency deviation
25 50.5
Bode Diagram
100
20 Load 50.4
80 System output
15
Magnitude (dB) 50.3
60
10

50.2

O u tp u t [ % ]
40 5

F re q u e n c y [ H z ]
20 0 50.1

0 -5
45 50

-10
0 49.9
Phase (deg)

-15

-45 49.8
-20

-90
-25 49.7
0 200 400 600 800 1000 1200 1400 1600 1800 0 200 400 600 800 1000 1200
Time [s] Time [s]
-135
-4 -3 -2 -1 0 1 2
10 10 10 10 10 10 10
Frequency (rad/sec) Fig. 6. Load variation tracking Fig. 7. Frequency deviation
(a)

120
Bode Diagram IV. CONCLUSION
100
The problem of random load disturbance rejection in a
Magnitude (dB)

80

60 hydro generator operation in power limited energy system is

40
formulated as an optimal controller design problem in sense of
minimum mean-square error (Wiener filter design). The
20

0
45

0
turbine speed and power controller design in case of
functional constraints is investigated. The controller structure
Phase (deg)

-45

-90
based on dynamic compensation as well as on direct design
-135

-180
with rational transfer function is determined. Simulations with
10
-4 -3
10
-2
10 10
-1

Frequency (rad/sec)
(b)
Fig. 5. Controller frequency response under (a) variable load variance
(b) variable load time constant
The both control algorithms use the optimal system
transfer function (Wiener filter) that requires a factorization of
the spectral density function. The direct design with rational
transfer function method avoids the separation Eq. (18) that
makes easily to design the control algorithm.
In Table I some results from the design procedure based
on minimum mean square error method (Wiener filter design)
are shown. It can be noticed the mean square error increases
significantly in case of functional constraints. The controller
structure based on dynamic compensation as well as on direct
design with rational transfer function method is determined.
The both controllers have the same order and an integral term.
TABLE I
OPTIMAL TRANSFER FUNCTION
Without functional constraints
Dv = 2 , Tl = 30 s
0.72s 2 + 0.34s + 0.07
Wopt ( s) =
3
s + 0.84s + 0.35s + 0.072
min mean-square error 0.72 10 −2
In case without functional constraints the controller structure
following the design procedures given by Eqs. (13) and (14) is:
7.37 s 3 + 15.07 s 2 + 4.74s + 0.62
Wc ( s ) =
s 3 + 3.04s 2 + 0.1s
R ( s ) = 0.1s 3 + 1.13s 2 + 1.39 s , S ( s ) = 1.72 s 2 + 4.5s + 0.86
T ( s ) = 4.37 s 3 + 10.81s 2 + 4.57 s + 0.86
The simulation results shown in Fig. 6 show a very good
system performance in tracking the load signal and keeping the
frequency about its nominal value in Fig. 7.
10
0
10
1
10
2
proposed algorithms are carried out in order to explore the
systems performance. This approach can be applied to other
control design problems describing the renewable energy
sources (RES) operation in electric power system where the
both generated power from RES as well as the load have a
random nature. In case of stochastic characteristics varying in
a wide range an appropriate adaptation procedure applied to
the controller parameters can be carried out [10].
ACKNOWLEDGEMENT
The authors would like to thank the Research and
Development Sector at the Technical University of Sofia for
the financial support.
REFERENCES
[1] A. Pervozwanski, Automatic Control Theory, Nauka, Moskva,
1996 (in Russian).
[2] R. Gray, An Introduction to Statistical Signal Processing,
Cambridge University Press, 2004.
[3] S. Doclo, A. Spriel, J. Wouters and M. Moonen, “Frequency-
With functional constraints Domain Criterion for the Speed Distortion Weighted
Multichannel Wiener Filter for Robust Noise Reduction”,
Speech Communication, vol. 49, 2007, pp. 636-656.
−1.70s3 + 0.84s2 + 0.42s + 0.07
Wopt (s) = [4] T. Puleva, “Multiple Model Wiener Filter for Load Disturbance
1.18s4 +1.99s3 +1.24s2 + 0.43s + 0.07
Rejection in Hydro Generator Local Network Operation”,
min mean-square error 3.5 10 −2 Proceedings of the Technical University of Sofia, vol. 58, Book
1, pp. 33-43, 2008.
[5] R. Brown, P. Hwang, Introduction to Random Signal and
Applied Kalman Filtering, John Wiley & Sons, Inc., 1992.
[6] P. Diniz, Adaptive Filtering: Algorithms and Practical
Implementation, Springer, 2008.
[7] M. Kolos, I. Kolos. Methods of Linear Optimal Filtration,
Lomonosov Moskow State University Press, 2000.
[8] K. Astrom, B. Wittenmark, Computer Controlled Systems:
Theory and Design, 3th edition, Prentice Hall, 1997.
[9] M. Lund, Pole Placement Design, Toolbox for MATLAB.
[10] R. Horowitz, B. Li, and J.W. McCormick, “Wiener-filter-based
Minimum Variance Self-tuning Regulation”, Automatica, vol.
34, no. 5, pp. 531-545, 1998.


Multi-Objective LMI based Robust State-feedback Design
of a Hydro Generator Speed Controller
Andrey Yonchev1 and Teofana Puleva2
Abstract – This paper describes the dynamic behaviour of a
hydro generator speed control in case of parametric uncertainty.
The investigations are focused on the multi-objective LMI based
H 2 / H ∞ state-feedback controllers design. The system
performance is investigated for both conventional lead lag
controller and PI state space controller. A modern design
technique based on LMI H 2 / H ∞ controller synthesis is
considered. The simulation results obtained from the both
conventional and proposed controllersare given. They show the
advantages of the designed robust control system.
Keywords – Hydro generator, Speed control, LMI based
robust control.
I. INTRODUCTION
performance specifications have to be transformed in terms of
LMIs. The efficiency of LMI method remains important due
to some reasons. First it is applicable to many systems without
restrictions on infinite or pure imaginary invariant zeros.
Additionally LMI based approach is valuable and remarkable
due to the existence of effective convex optimization
algorithms [4] and software [5] plus the MATLAB package
Yalmip and SeDuMi solver.
II. HYDRO GENERATOR LINEARIZED MODEL
The linearized model of a hydro turbine dynamics in per
units form (p.u.) can be presented by the equations:
ΔU = 0.5ΔH + ΔG , (1)

The linearized hydro turbine generator model parameters Δ Pm = ΔH + ΔU , (2)

vary significantly with the unpredictable load variations. This
fact leads to the adaptive and robust control algorithms. They d ΔU
Tw = −ΔH , (3)
are appropriate for control systems characterized with dt
significant dynamics variation depending on real time
measured variables [1]. Malik et al. have presented a self- where Tw is the water column time constant at any load. After
tuning controller tracking the plant parameters change in order some algebraic manipulations in Eqs. (1)-(3), the transfer
to provide an optimal performance over the wide operating function between the turbine opening position and the turbine
range. The simulation study includes both fixed plant and mechanical power can be obtained:
variable plant parameters that depend on the gate position
varying from the load conditions. Jiang considers the analysis −Tw s + 1
ΔPm = ΔG , (4)
and design of a hydraulic turbine governor using optimal 0.5Tw s + 1
robust control theory [2]. The advantage of this approach is
that the designed governor will guarantee the stability and The linearized model transfer function depends on the time
performance of the speed control for the entire turbine. Recent constant Tw which can be expressed as
works of robust control application are presented in [3]. Eker
considered a single-input multi-output (SIMO) governor LU 0
Tw = , (5)
design based on nonlinear plant model and polynomial H ∞ ag H 0
optimization. The system performance based on the multi-
loop cascade robust controller in the presence of load where L is length of the penstock [m] and ag is the
disturbances is investigated in the time and frequency acceleration due to gravity [ m / s 2 ] . The water column time
domains. The robust governor ensures that the overall system
remains asymptotically stable for all norm bounded constant at rated power TW is
uncertainties and permanent system oscillations.
LQr
The goal of this paper is to derive stabilizing multi- TW = , (6)
objective H 2 / H ∞ state-feedback controllers. Stability and ag AH r

1Andrey Yonchev is with the Technical University of Sofia,
Faculty of Automatics, Kliment Ohridski 8, 1000 Sofia, Bulgaria,
E-mail: ayonchev@tu-sofia.bg
2Teofana Puleva is with the Technical University of Sofia, Faculty
of Automatics, Kliment Ohridski 8, 1000 Sofia, Bulgaria, E-mail:
tpuleva@tu-sofia.bg
where H r and Qr are the hydraulic head and the turbine flow
rate at rated power and A is the pipe area [ m 2 ] . It can be seen
the water column dynamics exhibits a non-stationary behavior
depending on the variable water flow at wide range load
variation [6-7].

978-1-7281-7142-5/20/$31.00 ©2020 European Union 

 Hydro Generator (HG) Speed and Power Control The input signal u = [uc w]T implies the control signal and
The frequency of the power system is dependent on the
the disturbance, and the output is y = [ Δω ΔG ]T . The
active power balance. The load change reflects
instantaneously as a change in the electrical torque output Te matrices in Eq. (12) have the following form:
of the generator. This causes an interaction between the  Kd 1   1
mechanical torque Tm and the electrical torque Te which − 0   0 Tm 
results in speed variation as determined by the equation of  Tm Tm   
 2 2 2  2 
motion: A= 0 − +  , B = − 0 , (13)
 Tw Tw Ts   Ts 
1 
Δω r = (ΔPm − ΔPe ) , (7) 1   1 
Tm s  0 0 −   0
 Ts   Ts 
where ΔPm and ΔPe are the mechanical and the electrical
1 0 0  0 0
power deviations from their initial values in p.u., Tm is the C=  , D= .
mechanical starting time. The frequency dependent 0 0 1  0 0
characteristic of a composite load may be expressed as To ensure the integral control action the model (12) is
extended by extra integral state:
ΔPe = ΔPL + K d Δωr , (8)
xi (t ) = Δω ref − Δω (t )
where ΔPL is a non-frequency sensitive load change, and K d
The augmented state space description is:
is a load-damping constant in the frequency sensitive load.
In accordance with the principles of the primary frequency x = Ax (t ) + Buc (t ) + ΓΔωref , (14)
control the generators operating in parallel with the electric
power system have to share the load when a sudden change in
system load occurs. This is accomplished by the governor  A 0  B1  0 
A= 1  , B =   , Γ =   , C = C 1 0 , (15)
speed-load characteristic  −C 0 0 1 
1
ΔP = − Δω , (9)  2 1
T

Rp C1 = [1 0 0] , B1 = 0 −  ,
 Ts Ts 
where R p is the speed regulation (droop factor). The T
and x =  xT xi  .
permanent droop determines the speed regulation under steady
state conditions. The transient droop compensation is given by The control signal can be expressed as:

1 + sTR u (t ) = − K c x (t ) + K i xi (t ) + u w (t ) = − Kx + u w (t ) , (16)
Wc ( s ) = , (10)
1 + ( RT / R p ) sTR
where K = [ Kc − Ki ] , Ki is the integral term gain, K c is
where TR is the reset time and RT is the temporally droop the proportional term gain and uw (t ) is the control signal term
factor. For stable operation the optimal values of the which is used to compensate the load change. To obtain the PI
controller parameters are given by [6-7]. The block diagram of state space controller parameter K the LQR or pole placement
the conventional turbine governor is shown in Fig. 1. design technique can be used.
Tw
RT = [2.3 − (Tw − 1)0.15]
, TR = [5 − (Tw − 1)0.5]Tw . (11)
Tm III. POLYTOPIC MODEL OF A HYDRO GENERATOR
The state space description of the linearized hydro
generator model is The considered hydro-generator system can be described
using the following polytopic model:
x (t ) = Ax(t ) + Bu(t )
, (12) x = A(t ) x + B1 (t ) w + B2 (t )u
y(t ) = Cx(t )
z∞ = C1 (t ) x + D11 (t ) w + D12 (t )u , (17)
here x = [ Δω ΔPm ΔG ]T is the state vector (see Fig.1). z2 = C2 (t ) x + D22 (t )u
where the state-space matrices range in a polytope is
1
Rp
 A(t ) B1 (t ) B2 (t )    Ak B1k B2 k  
 C (t ) D (t ) D (t )  ∈ Co :   C D11k

D12 k   ,
Δωr  1 11 12    1k
1 − sTw 1
 C (18)
D22 k  
Load Ref 1
Ts s + 1
Wc ( s )
1 + 0.5Tws Tm s + D C2 (t ) 0 D22 (t )   2k 0
Turbine ΔPL Rotor k = 1, K
Governor
Fig. 1. Block diagram of hydro turbine governor system

and with w is denoted the disturbance signal. Such polytopic IV. MULTI-OBJECTIVE LMI BASED ROBUST
models can be used to describe plants with uncertain and/or STATE-FEEDBACK DESIGN
time-varying parameters [8]. The system matrix

 A(t ) B1 (t ) B2 (t )  In this paper we will perform multi-model H 2 / H ∞ state-

S (t ) =  C1 (t ) D11 (t ) D12 (t )  feedback design with pole placement constraints using the
LMI technique. We consider the situation that the state vector
C2 (t ) 0 D22 (t ) 
can be measured. Let T∞ ( s ) and T2 ( s ) be the closed-loop
varies within a fixed polytope of matrices, transfer functions from w to z∞ and z 2 , respectively. We aim
to design a state-feedback law u = Kx in order to
 k k

S (t ) ∈ Co {S1 ,  , S k } :=   α i Si : α i ≥ 0,  α i = 1 , (19) • keep the H ∞ norm of T∞ under a given value γ 0 > 0 ;
 i =1 i =1 
• keep the H 2 norm of T2 (LQG) under a given value
where S1 , , Sk are called vertex systems: ν 0 > 0 [9].
In order to perform multi-model H 2 / H ∞ state-feedback
 A1 B11 B21   Ak B1k B2 k 
design with pole placement constraints using the LMI
S1 (t ) =  C11 D111 D121  , Sk =  C1k
 D11k D12 k  . (20) technique we seek an unique quadratic Lyapunov function
C21 0 D221  C2 k 0 D22 k  X = X 2 = X ∞ that makes the desired control objectives in the
polytope and the presented LMI system has to be solved [10].
Taking all the above in mind, S (t ) represents a convex Minimize αγ 2 + β Trace(Q ) over Y, X, Q and γ 2 subject to
combination of system matrices S1 , , Sk . The
the
nonnegative elements 1 , , ak are denoted as polytopic
a  Ak X + XAkT + B2 k Y + Y T B2Tk B1k XC1Tk + YD12T k 
 
coordinates of S. For the hydro-generator system described by  B1Tk −I D11T k  < 0 , (21)
Eqs. (12)-(13), we have the following four vertex systems:  C1k X + D12 k Y D11k −γ I 2 
 
 − K d / Tml 1 / Tml 0   0 1 / Tml 
A1 =  0 −2 / Twl 2 / Twl + 2 / Ts  , B1 =  −2 / Ts 0   Q C2 k X + D22 k Y 
 XC T + YDT X >0
 0 0 −1 / Ts   1 / Ts 0   2k 22 k 
Trace(Q) < ν 02 .
 − K d / Tml 1 / Tml 0   0 1 / Tml 
γ <γ
2 2
A2 =  0 −2 / Twu 2 / Twu + 2 / Ts  , B2 =  −2 / Ts
 0  0

 0 0 −1 / Ts   1 / Ts 0  Denote the best solution by ( X * , Y * , Q * , γ * ) , the obtained

state-feedback gain is expressed by K * = Y * ( X * ) −1 . This gain
 − K d / Tmu 1 / Tmu 0   0 1 / Tmu  satisfies the worst case quality specification.

A3 =  0 −2 / Twl  
2 / Twl + 2 / Ts  , B3 =  −2 / Ts 0  T∞ ( s ) ∞
≤γ* T2 ( s ) 2 ≤ Trace(Q* ) .
 0 0 −1 / Ts   1 / Ts 0 

 − K d / Tmu 1 / Tmu 0   0 1 / Tmu  V. SIMULATION RESULTS

A4 =  0 −2 / Twu  
2 / Twu + 2 / Ts  , B4 =  −2 / Ts 0 
   1 / Ts
The control system performance is investigated by
0 0 −1 / Ts 0 
comparing the control system based on the conventional
controller described by (10), state space PI controller (14)-
1 0 0  0 0
C=  , D= , (15) with the designed system based on the investigated multi-
 0 0 1 0 0 objective LMI based robust controller. The following plant
where the hydro generator rotor time constant Tm and the parameters are assumed: servo system time constant
water column time constant Tw take values in the interval: Ts = 0.4 S; water column time constant Tw = 3 S; rotor time
11 ≤ Tm ≤ 13 , 2 ≤ Tw ≤ 4.5 , K d = 1 and the servo system time constant Tm = 12 S, load-damping constant K d = 1 ; speed-
constant is Ts = 0.4 s .The notation Tl and Tu refers the lower power droop factor R p = 0.04 . A conventional lead-lag
and the upper bound of the interval for corresponding time controller in accordance with Eqs. (10)-(11) is designed. In
constants. Fig. 2 the turbine generator speed under variable water
We use Robust Control Toolbox of Matlab® to define the column time constant in the range of [2, 4.5] is shown. It can
polytopic model (17) of the hydro generator system and to be seen the overshoot increases significantly. The load
perform the multi-objective robust state-feedback design. disturbance of 20% in t=100 s involves an overshoot increase
from 3.9% to 7.8% when the water column time constant
increases from the nominal value of 3s to 4.5s.


 A state-space controller augmented with an integral state is robust H ∞ LMI based controller ensures H ∞ performance
explored. The matrices of the performance index are: γ=7.99e-004. The simulation results obtained with H 2 LMI
 I4 O(4,1)  based controller are similar to those determined by the H ∞
Q= , R = 10
O(4,1) 10  controller.

The integral gain and the state space controller are: VI. CONCLUSION
Ki = 1, K p = [ 6.25 0.45 1.23] .
The simulation results obtained with H 2 LMI based This paper considered the dynamic behaviour of a hydro
generator speed control in presence of parametric uncertainty.
controller are similar to those determined by the H ∞ The polytopic model with two parameters was obtained. The
controller. investigations were focused on the multi-objective LMI based
Turbine - Generator speed [p.u.] H 2 / H ∞ state-feedback controllers design. The system
2

Tw=2
performance was investigated for both conventional lead-lag
1.5
Tw=2.5
Tw=3
Tw=3.5
controller and PI state space controller. The time domain
Tw=4
Tw=4.5 specifications of these systems are very sensitive to the
1 parametric uncertainty. A modern design technique based on
LMI H 2 / H ∞ controller design was considered. The results
0.5

based on simulation for the two conventional and proposed

0
controllers were presented. They point the efficiency of the
designed robust control system. Using this approach the time
-0.5
0 20 40 60 80 100 120 140 160 180 200
domain specifications satisfy the technological requirements
of the hydro generator speed control.
Time [s]

Fig. 2. System response with conventional controller and variable

water column time constant
1.4
Turbine generator speed [p.u.] ACKNOWLEDGEMENT
1.2

1 The authors would like to thank the Research and

0.8 Development Sector at the Technical University of Sofia for
0.6
Tw=2 s the financial support.
0.4 Tw=2.5 s
Tw=3 s

REFERENCES
0.2 Tw=3.5 s
Tw=4 s
0 Tw=4.5 s

-0.2

-0.4
[1] O. Malik, Y. Zeng, “Design of a Robust Adaptive Controller for
0 20 40 60 80 100
Time [s]
120 140 160 180 200
a Water Turbine Governing System”, IEEE Trans. on EC, vol.
Fig. 3. System response with state space PI controller and variable 10, no. 2, pp. 354-359, 1995.
water column time constant [2] J. Jiang, “Design an Optimal Robust Governor for Hydraulic
Turbine Generating Units”, IEEE Trans. on EC, vol. 10, no. 1,
In Fig. 4 we show a set of time responses of the four pp. 188-194, 1995.
polytopic systems with the H ∞ LMI based robust optimal [3] I. Eker, “Governors for Hydro Turbine Speed Control in Power
Generation: a SIMO Robust Design Approach”, Energy
controller. The time domain specifications satisfy the Conversion and Management, vol. 45, Issues 13-14, pp. 2207-
technological requirements of the HG speed control. 2221, 2004.
1.5
Polytopic systems - time responses
[4] Y. Nestorov, P. Gahinet, Interior-Point Polynomial Algorithms
in Convex Programming, SIAM, Philadelphia, PA, 1994.
1 [5] P. Gahinet, A. Nemirovski, A. Laub, and M. Chilali, LMI
states 3 Control Toolbox for Use with MATLAB, The MathWorks, Inc.,
0.5 states 2
2000.
states 1
[6] IEEE Working Group, “Hydraulic Turbine and Turbine Control
0
Models for System Dynamic Studies”, Transactions on Power
System, vol. 7, no. 1, pp. 167-179, 1992.
-0.5
[7] P. Kundur, Power System Stability and Control, McGraw-Hill,
-1
Inc., 1994.
[8] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear
-1.5 Matrix Inequalities in Systems and Control Theory, SIAM
0 10 20 30 40 50 60 70 80 90
time [s] books, Philadelphia, 1994.
[9] P. Khargonekar, M. A. Rotea, “Mixed H2/H∞ Control: a Convex
Fig. 4. All polytopic systems responses with H ∞ controller
Approach”, IEEE Trans. on AC, 39, pp. 824-837, 1991.
Note that the polytopic systems have two parameters, which [10] A. Yonchev, M. Konstantinov, and P. Petkov, Linear Matrix
vary in the considered intervals. The corresponding time Inequalities in Control Theory, Demetra, Sofia, 2005, (in
responses vary in a relatively small range in comparison to the Bulgarian).
corresponding responses shown in Figs. 2 and 3. The designed


Impact of Microclimate Changes and Noise to Workers as
a part of an Implemented Health and Safety System
Ivo Kuzmanov1 and Roberto Pasic2
Abstract – Impact of microclimate changes to workers has
always been a special health and safety issue to each business
entity who obligatory fulfil the national laws regarding health
and safety. Especially if the microclimate changes are related to
worker circulations among processes where they have to do
business activities in cold and hot chambers as a part of a daily
routine several times a day. On the other hand if the daily
routine has production activities in a room where the noise is a
special problem as well then we have a special situation. These
two main concerns are “must do” problems to solve by
themselves, but combined with a situation where some accidents
happened, we certainly have a situation to show and to deal with.
This paper presents only a small part from the research (real
activity in real business entity) and some of the equipment and
measures (especially made for the purpose) used to deal with the
same one.
Keywords – Health and Safety, Production system,
Measurement, Negative impact to workers, Noise.
I. INTRODUCTION
The paper presents only a part from an extensive research,
as still ongoing applicative activity, with several key activities
such as: creating a measurement unit (specially made and
even improved one) [1-8] as one of the equipments used
during the processes, measurements on different parameters
(temperature, humidity, noise) into production capacity in
Republic North Macedonia, but also dealing with the occurred
problems one by one in short notice and finally continuous
monitoring on defined measurement spots. In addition to
results published in [9], in this research we are presenting only
the situation at first – specially created unit and some of the
additional measurements regarding temperature and humidity
[9], and we are focusing on some problems which occurred
during the process. Also we are presenting some of the
solutions made which are in a stage of so called real
applicative testing mode or approved mode (used on a daily
base) into the business entity previously mentioned. Also, we
are presenting some additional problems on the spot which
occurred during the processes of dealing with the first defined
problem.
1Ivo Kuzmanov is with the Faculty of Technical Science at the
University St. Kliment Ohridski Bitola - UKLO, Bitola 7000,
Republic Noрth Macedonia, e-mail: ivo.kuzmanov@tfb.uklo.edu.mk
2Roberto Pasic is with the Faculty of Technical Science at the
University Kliment Ohridski Bitola - UKLO, Bitola 7000, Republic
North Macedonia
978-1-7281-7142-5/20/$31.00 ©2020 IEEE 
But first of all some small things about the entity – it’s a
production company with several capacities (work stations) in
different parts of Republic North Macedonia, and considering
the year fluctuation of workers, with more than 100 up to 300
employees. The focus during the first contact with the
company was on several key problems regarding health and
safety issues on different micro locations and different spots
into the capacities. But one of the main problems was the
health and safety issue regarding the workers which during
day activities had the mentioned problems – transition from
cold to hot rooms and vice versa.
From all of the presented problems by the company, the one
presented here is the most interesting because starts with one
problem – circulation from cold to hot chambers and then to
normal work conditions several times a day (even in one stage
in every 15 minutes).
This problem during the process of observation evolved
with multiple problems regarding health and safety. At first
the real problem was the circulation of workers who were
exposed to so called ‘temperature shocks’ from -3°C or in
some cases even -5°C to normal conditions (19 - 23°C) and
then to extremely hot conditions where temperatures were
even +45°C. Also one of the problems was the circulation
directly from hot to cold or otherwise several times during the
daily period of work 8 and even more working hours per day.
This problem occurred and concerned 8-12 people at any
time, or in some cases 15-25 people during shifts. Other thing
that was a real concern was the HTZ equipment (HTZ is a
short notice used in Macedonia, given for health and safety
equipment) given to the workers and the conditions that they
work in – wet floors, wet surfaces, lot of equipment into the
room and finally lot and lot of electricity wires (although all
of them were in a good condition, even some of them brand
new instalation).
At that stage, that was the starting point why we created
also a portable unit (2 units) presented in detail in the previous
publication [9], so we could conduct multiple measurements
in several occasions and even during the hole day. As starting
point even we conducted a real calibration with equipment
which is quite expensive, so we could get precise
measurement data from a portable unit. That was the maybe
the only way to get a real picture during various
measurements in multiple days (even is some cases 8 hours
per day), about what is happening and things could be seen
from a different angle.
But during the processes and monitoring of the whole thing,
other problems occurred: noise at one of the spots with a daily
activity with 4-8 people at spot (with circulation into hot and
cold chambers of the same ones), and a thing that was not said
at first (documents were there) – previous injuries and
multiple and common illnesses. Especially injuries and the TABLE I
noise (86-89 dB) was a new thing at that stage, even for the PRESENTING SMALL PART FROM A 60 MINUTE MEASUREMENT
company. And finally the paper presents only a small part IN ONE DAY

from the processes of monitoring, processes of creating Temperature (°C) Humidity (%RH)
solutions on spot (by fulfilling legal obligations) and finally 1 17.2 47
the process of monitoring the created solutions in practice and
2 16.3 51.2
setting them as a daily routine (must do thing).
3 16.7 51.3
4 23.1 56.3
II. THE MEASUREMENT UNIT AND THE RESULTS 5 25.7 57.8
6 10.0 61.1
The measurement unit used as a portable one was in-house 7 6.9 70.1
created [1-8] and tested in various different cases. We are 8 3.2 67.8
using 8 additional pieces of the same one – not only here, but 9 1.8 48.5
also in different locations. During the processes of comparison 10 -2.0 54.7
to other equipment we have concluded that the measurement 11 -2.1 57.7
unit brought us real and secure results. 12 -2.8 68.0
The same one during the process of creation till now, was 13 7.8 51.8
compared to several other equipments (more than 7 14 9.5 51.2
producers) even in real measurements on spot and gave us 15 13.5 51.6
results with a slight variation (in some cases ±0.1 to 16 28.7 52.3
maximum variation of ±0.2).The equipment is in detail 17 34.5 50.8
presented into [9], using previous knowledge from [1-8], and
18 36.9 47.8
real knowledge, but here inside of the same one given in Figs.
19 41.3 46.7
1 and 2 is presented. Additionally, the measurements
20 16.7 51.8
(temperature and humidity) from the same one – measured in
different time and data relative to [9] – are given in Table I.
The data presented into Table I, are in one day in a time
frame since 9.05 till 10.00 o’clock by local time. At this stage
we only get a segment from the whole picture in what kind of
conditions workers do their everyday routine activities. That
day, the outside measurements (as key factor for comparison)
during the processes of inside measurements were 10.1 till
15.3 Celsius degrees and humidity of 40.7% at the starting
point at 9.00 o’clock. That was a starting point which brought
us to another problem at that stage. The problem was that in
Republic North Macedonia although there is a health and
safety law and multiple regulations regarding the same one,
there are only some recommendations considering the
temperatures and humidity (not obligatory - only
recommendations). And considering all of the
Fig.1.Sensor node structure – Arduino NANO recommendations, they brought us to data about the
temperature in work conditions for winter from 18 to 21
Celsius degree and for summer from 22 till 25 Celsius
degrees, once more as a recommendation.
At this stage, we have workers in hot and cold chambers
with variations of temperatures. So considering the data and
the regulations, we started a survey and considering the facts
available on web, we found several information’s regarding
this thing from Balkan countries and some even from EU, but
the most real ones were the Slovenian regulations and the
Serbian ones [10-11].
According to the same ones, the Serbian regulations and the
Slovenian ones, regarding preventive measures for health and
safe work were quite good ones and quite similar. During the
process of research we found tabulated results into the spot
6.4 from the Serbian regulations [11] that were one of our
leads in which way should we go. That was a good lead
because the given results show even the outside temperature
Fig. 2. Temperature and humiditysensor – inside the box and the defined way of working. Regarding the outside


conditions (about + 15 Celsius degrees) the humidity
maximum value should be 75 (we were even better) and the
temperature at maximum + 28 Celsius degrees.
Another thing which was interesting was the situation that
we used quite a large period of time reconsidering and re-
reading the Macedonian law and regulations and we found
that the work conditions should not harm or in any case
influence the workers (still not regulated the temperatures or
humidity with numbers). During this process we had also
some communication with health and safety inspectors as
well.
Seeing the real situation and being on spot, as previously
said that this is only a part of extensive research we found two
other problems. The first one, were the documents about
previous injuries (business entity secret – so this will not be a
thing to mention into this or other papers) and illness (colds,
frequent flues, sneezing as a frequent thing among workers on
spot etc.). But the second one considering previous
measurement and actions to deal with the problem
(temperature shocks) was a real surprise. It was the noise in a
room where most of the time spends several workers at any
time of the day. It even seems weird because it was spotted
after some inventory change and maintainer activities. This
was unexpected because all of the activities were done to
reduce even more the noise and to create a better production
process. We got the same notice from the workers as well –
the noise was larger since the new equipment was installed
and since preventive maintain activities on the actual
equipment were done.
Having in mind the previously mentioned things, a series of
measurements of the noise were conducted using Testo
equipment (quite professional one) in order to get a better
overview of the main problem that occurred.
Table II presents only a small part of the measurements in a
room of 200 meters with more than 15 machines inside. But,
we focus on only 5 of them.
TABLE II
PRESENTING SMALL PART FROM NOISE MEASUREMENTS
NOISE (dB)
1 80
2 79.9
3 86.8
4 87.1
5 89.0
6 88.7
7 78
8 78.1
9 76.7
10 76.5
Table II presents 10 measurements on 5 different spots,
taking the principle 2 measurements on spot. Especially
surprising where the data conducted regarding spot 2
(measurements 3 and 4) and 3 (measurements 5 and 6) which
had some different measurement data after some changes to
the equipment and after maintainer activities. That was maybe
the bigger surprise, but that will be discussed in the future
work (we had conducted a whole monitoring what happened

using FMEA). And FMEA is abbreviation for Failure Mode
Effects Analysis.
It’s quite important to notice that the Macedonian
regulations [12] regarding noise – defined in special
regulation, defines the low and upper limits of noise which are
from 80 – 87 dB. Considering this fact we had a situation with
the spots really on upper limit or even more, and a different
problem to solve. Finally, we are still solving problems in the
process that occurs. Also, the handy transportation of weight
is a problem which has to be solved.
III. THE SOLUTIONS AND FUTURE STEPS
Seeing the previous presented, some steps were made and
some still are in a move so we could get a better environment
for the workers. Several conducted activities made it possible
to create better conditions for workers. First regarding the
situation with the HTZ equipment – although there was quite
a good one when we came the same one was not appropriate
for the situation. So dealing whit the fact that there were some
comes and goes from workers to hot or cold chambers, first of
all we solve the thing by using a line in which most of the
products are transported not by human hand. But for the
processes where we have to have a human hand we created a
cycle activity in which each worker is on spot only using the
cold chamber but no more than 15 minutes a day with a
special HTZ equipment for cold conditions on it. Hot
chambers are not used by humans at this stage – the transport
is on track. This is specially solved also by the real focus from
the management. So, as previously said the HTZ equipment
(updated version of the same one with some significant
efficiency to health and safety but also with some financial
savings for the company) now is only appropriate for the
room where the noise was in some parts. The situation with
the noise is also a thing that is solved with some changes but it
will be presented in the future work. Now, we are only dealing
with a new thing in two spots where a hand weight is
transported, but on small distances. We would like to solve it
as well.
Further, Table III presents the momentarily measured
values of the noise, from which one could get a real picture
what happened. The problem solution will be presented in
some other publication.
TABLE III
PRESENTING SMALL PART FROM NOISE MEASUREMENTS AT THE
MOMENT ON THE SAME SPOTS AS PREVIOUS
NOISE (dB)
1 73.2
2 73.1
3 78.8
4 78.7
5 83.8
6 83.8
7 77.9
8 78.0
9 76.4
10 76.5
 IV. CONCLUSION
Finally, we could conclude that this paper presents not only
previously created measurement unit, shown in detail in [9]
and based on Arduino NANO MCU, but also whole problem
on spot in a real business entity. The same one presents the
starting point – daily routine in humidity and hot and cold
chambers, but also the process of monitoring in which several
new problems occurred. Occurred problems as new ones were
the noise after the maintainer activity and the new equipment
installed which brought us to additional measurements and
additional activities, and also the illnesses and injuries which
were a part of that process as well. So from one problem we
got several different ones. Some of the activities and
measurements done are presented into the paper. The
important thing to mention is that the created unit is used in
more than 8 different spots at this stage, even in different
business entity at the moment so we could get a real data, but
we are also considering to improve the same one with adding
a noise sensor to the same one (small one – we are in a search
for appropriate and good one at the moment).
Finally seeing the data aimed and the amount of the same
ones at this stage we could say that some future publications
will be made as well.
At the end we are also considering to create a sensor node
with web based application for online collection of data in real
time as soon as we solve the ad – in to the measurement units
with a noise sensor, but with some financial support from
sponsors.

REFERENCES
[1] M. McRoberts, Beginning Arduino, Technology in Action,
APress.
[2] C. Amariei, Arduino Development Cookbook, Packt Publishing,
2015.
[3] S. Hussain, N. Schofield, and A.W. Matin, “Design of a Web-
based Application for Wireless Sensor Networks”, The 17th
International Workshop on Database and Expert Systems
Applications, pp. 319-326, 2006.
[4] Arduino Nano, https://store.arduino.cc/arduino-nano.
[5] Sensirion SHT71,
[6] https://www.sensirion.com/en/environmental-sensors/humidity-
sensors/pintype-digital-humidity-sensors/
[7] MicroSD Card Spec.
https://www.sdcard.org/developers/overview/index.html
[8] S. M. Ferdoush, A Low-cost Wireless Sensor Network System
using Raspberry Pi and Arduino for Environmental Monitoring
Applications, Master of Science Thesis, University of North
Texas, 2014.
[9] Ivo Kuzmanov, Roberto Pasic, “Real Impact of the Microclima
Changes to Workers who Circulate from Hot to Cold Chambers
– A Part from an Extensive Research into Industrial Entity”,
TEMEL International Journal, vol. 3, issue 2, pp. 1-5, December
2019.
[10] https://www.informiran.si/portal.aspx?content=temperatura-na-
delovnem-mestu&showMenu=1&showRightFrame=1
[11] https://www.paragraf.rs/propisi/pravilnik_o_preventivnim_mera
ma_za_bezbedan_i_zdrav_rad_na_radnom_mestu.html
[12] http://www.mtsp.gov.mk/WBStorage/Files/pravilnik_bucava.
pdf
 Hardware Implementation and Comparison of CRA and
TRA when Trying to Recover the AES-128 Key
Ivaylo Vladimirov1, Desislava Nikolova2 and Zornica Terneva3
Abstract – In this scientific paper, a comparison of two side-
channel radio attacks is made: CRA – Correlation Radio Attack
and TRA – Template Radio Attack. Both attacks are tested on
mixed-signal chips that include both the radio transceiver and
the digital logic on the same integrated circuit. In systems like
this, the analog transmitter unintentionally leaks sensitive data,
related to the cryptographic hardware components and the
software being run on the CPU, by broadcasting it on a shifted
frequency. The basic conception is that information from the
electromagnetic leakage is gathered and then exploited by
analyzing it via the side-channel attacks in order to crack the
AES-128 algorithm. The system needed for the experiment
consists of: a PCA10040 chip by Nordic Semiconductor, an
USRP n210 by СР-Ettus Research and a Linux-based computer.
Keywords – Side-channel attacks – Correlation Radio Attack,
Template Radio Attack, AES-128, Bluethooth, PCA10040,
USRP.
I. INTRODUCTION
The mixed-signal chips, also known as Radio Frequency
Integrated Circuits (RFICs), have become increasingly
popular thanks to the drive for ever smaller and cheaper
components in microelectronics. They are used in the
construction of widespread wireless communication systems,
such as mobile phone, tablet, Bluetooth, Wi-Fi and IoT
devices. Structurally their analog and digital circuitry reside
on the same piece of a die. This physical proximity naturally
leads to noise issues and novel side-channel attacks that can
break the implemented cryptography algorithm [1-2].
The BLE Bluetooth protocol is a wireless standard used for
exchanging data between different devices over short
distances using radio waves in the 2.4 GHz to 2.48 GHz band.
It is used in a combination with AES-128 in these kinds of
circuits for secured package based device-to-device
communication [3].
The Advanced Encryption Standard (AES) is a
specification for the encryption of electronic data. It is based
on a design principle known as a substitution–permutation
1Ivaylo Vladimirov is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: ivaylo.h.vladimirov@gmail.com
2Desislava Nikolova is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: desislava.v.nikolova@gmail.com
3Zornitza Terneva is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: z.terneva@abv.bg
978-1-7281-7142-5/20/$31.00 ©2020 European Union 
network, and is efficient in both software and hardware
implementations. The key size used for an AES cipher
specifies the number of transformation rounds that convert the
input, called the plain-text, into the final output, called the
cipher-text. The numbers of rounds are as follows: 10 rounds
for 128-bit keys; 12 rounds for 192-bit keys; 14 rounds for
256-bit keys [4].
The two types of side-channel power analysis techniques
utilized in the completion of a full key recovery attack against
AES-128 compared in this paper are:
The Correlation Radio Attack (CRA) is a model based
analysis. It is generally assumed that the data leakage through
the power side-channel depends on the energy required to flip
the bits from one state to the next at a given time. When used
in combination with the Hamming weight model it can be
applied to provide a probability value for the likelihood of a
given AES-128 sub-key to be the one used in the encryption
[5].
The Template Radio Attack (TRA) exploits the power
consumption measured during the operation, expressed as a
discrete function of time to reconstruct the sequence of
operations during the secret computation and derive
information about the AES-128 key. It proceeds in two
phases: training (Consisting in measuring the electrical
activity of a clone system during numerous encipherments and
creating “templates” based on the collected information called
“traces”) and attacking (Consisting of collecting a few traces
and matching them with the existent templates) [6].
II. CONCEPT
The key observation of this research is to make a
comparison between and implement in hardware the two
power analysis techniques. They are based on the idea that
noise coupling propagates sensitive information from the
digital domain to the radio transmission chain, which
broadcasts it at a much larger distance than normal EM leaks
[1].
Fig. 1. Steps that lead to the radio transmission of the leak
 Information related to the digital processor’s activity leaks
into other parts of the chip. Components of that leak are likely
to reach the transmisive analog portion, in particular the radio
transmission chain. They contain frequencies in the range of
the radio’s baseband signal. The leaked information then
couples with the baseband signal via the mixer, amplifier or
the Voltage Controlled Oscillator (VCO) that is part of the
carrier-frequency synthesizer. In any of the cases the result is
unintended amplitude/frequency modulation of the carrier [7].
The capacitive coupling with the VCO leads to an
amplitude modulation of its output signal. Then the
parasitically modulated carrier is amplified by the power
amplifier and radiated by the analog antenna over a potentially
very long distance. These steps also shown in Fig. 1 are as
follows: the generation of digital noise, its frequency
characteristics, its propagation into other parts of the circuit,
and finally its emanation via the radio [8].
III. EXPERIMENTAL SETUP AND EXECUTION
The general setup idea is that one side (See Fig. 2 B-left
side) there is a computer controlling the RFI circuit, which
repeatedly transmits data via Bluetooth BLE using the same
AES-128 key, on the other (See Fig. 2 B-right side) there is
another system, functioning as a software defined radio
(SDR), tuned at the specific frequency of the leak. The SDR
system consists of a USRP device and a PC connected via an
Ethernet cable and stores the traces, used for the performance
of the two power analysis, on the hard drive of the computer.
Fig. 2 shows both the preliminary idea and the realized setup.
Fig. 2. The experimental setup: A) The physically realized one,
B) The preliminary scheme
In order to use the template attack the attacker needs to
have a copy of the whole system, which is needed for the
creation of the templates used for its execution. This step may
take days even weeks, because for generation of one high
quality template a 20 to 30 thousand traces need to be
collected. This is equivalent to two to three hours of
collection.
After a good set of templates is acquired, the experiment
continues with a real time key recovery. First we collect about
3000 traces (about 12 minutes) using the setup shown in
Fig. 2. Then using only PC containing all required software
the two attack techniques are run. An example and full
description of the attacks and how they run can be found in
[9].

After the performance of an attack a table containing the
numerical results is displayed in the Linux terminal window,
an example result can be seen in Fig. 3. The first two lines
show the bytes of the best guess and of the known key. The
PGE is the Partial Guessing Entropy - the "distance" of the
best guess from the known byte (0 when the guess is correct).
At the end the number of correct bytes is shown.
Fig. 3. The terminal window results after performing an attack
IV. COMPARISON AND CONCLUSION
This leakage used and described in this scientific paper is
not due to the design error of an individual vendor, but to a
fundamental difficulty in designing mixed-signal chips.
When making a comparison between the two attacks we
can see that the CRA is a simpler technique needing less
preparation but using more traces collected at the moment to
return a better result. Its computational speed is fairly good
depending on the power of the computer and the number of
traces provided (the lesser traces the fasted the attack, but also
the lower the chance for a full recovery of the key). It also
uses all given traces before making a decision on which is the
best guess for a certain sub-key, which additionally slows the
attack. The full execution took about 11 minutes.
In comparison, the TRA’s algorithm is more complex and
needs more computation power. It requires a lot of preparation
– a copy of the system that will be attacked and its excessive
pre evaluation. Its computational speed is very fast and makes
a decision on which is the best guessed sub-key after a certain
value for the probability is reached. The full execution took
about 1 minute and 30 seconds, which is 10 times faster and
needed a only a little over 100 traces to recover the full key,
which is 30 times the traces less than collected.
ACKNOWLEDGEMENT
The research described in this paper is partially realized
within the Erasmus+ mobility program with the help of the
professors at SUPELEC University in Rennes, France.
The implementation of the codes for the two side-channel
power analysis techniques used in this paper are described in a
scientific document: “Software implementation of CRA and
TRA to recover the AES-128 key using side-channel signals
with Python3” [9] that will also be presented at the ICEST
2020 conference.
REFERENCES
[1] G. Camurati, S. Poeplau, M. Muench, T. Hayes, and A.
Francillon, “Screaming Channels: When Electromagnetic Side
Channels Meet Radio Transceivers”, CCS'18: Proceedings of
the 2018 ACM SIGSAC Conference on Computer and
Communications Security, January 2018, pp. 163-177.
[2] A. Afzali-Kusha, M. Nagata, N. K. Verghese, and D. J. Allstot,
”Substrate Noise Coupling in SoC Design: Modeling,
Avoidance, and Validation“, Proceedings of the IEEE, vol. 94,
no. 12, pp. 2109-2138, Dec. 2006.
[3] I. Hibbert, "Bluetooth.4 - Bluetooth Protocol Family", Itronix
Inc, 2006.
[4] D. J. Bernstein, “Cache-Timing Attacks on AES”, Report, 2005.
[5] V. K. Rai, B. V. Reddy, S. Tripathy, and J. Mathew,
“Correlation Power Analysis and Effective Defense Approach
on Light Encryption Device Block Cipher”, Security and
Privacy, vol. 2, no. 5, September/October 2019.

[6] M. A. El Aabid, S. Guilley, and P. Hoogvorst, “Template
Attacks with a Power Model“, 2007.
[7] S. Bronckers, G. Van der Plas, G. Vandersteen, and Y. Rolain,
Substrate Noise Coupling in Analog/RF Circuits, 2009.
[8] H. Li, A. T. Markettos, and S. Moore, “Security Evaluation
Against Electromagnetic Analysis at Design Time”, Tenth IEEE
International High-Level Design Validation and Test Workshop,
Napa Valley, CA, 2005, pp. 211-218.
[9] D. Nikolova, I. Vladimirov, and Z. Terneva, “Software
Implementation of CRA and TRA to Recover the AES-128 Key
Using Side-Channel Signals with Python3”, ICEST 2020.
 Software Implementation of CRA and TRA to
Recover the AES-128 Key using
Side-Channel Signals with Python3
Desislava Nikolova1, Ivaylo Vladimirov2 and Zornica Terneva3
Abstract – In this scientific paper is presented a software
implementation of two side-channel radio attacks: CRA –
Correlation Radio Attack and TRA – Template Radio Attack.
The main goal is to recover the AES-128 key from an unknown
system. Both attacks are developed in Python3 using a Linux-
based computer. The algorithm uses Probability Density
Functions and the Hamming-weight model. The results are
calculated by the Partial Guessing Entropy. The software is run
on the CPU.
Keywords – Side-channel attacks – Correlation Radio Attack,
Template Radio Attack, AES-128, Python3, Hamming-weight
model, Partial Guessing Entropy.
In this paper we are presenting a few python libraries,
which are going to be used in the software implementation.
The name of the package click comes from Command Line
Interface Creation Kit. It gives the opportunity the command
line interfaces to be written with as little code as possible in
order to function correctly.
Numpy, usually referred as np is the most common library
for dealing with arrays and computing them fast in python.
Pyplot from matplotlib is used mainly for visualization of the
results. The package os in this approach is used only for
finding the correct path and managing the files in the
directory.

I. INTRODUCTION

Every type of system has sound, heat, light or other

electromagnetic radiation while operating. They can be
correlated with what is happening inside the system. These
physical sources can be used as side channels if they leak out
information (Fig. 1). Side channel attacks are used to derive
the key of a cryptosystem and use its weakness in the physical
implementation [1].
In this paper, we are going to analyze two types of attacks -
TRA and CRA. Template radio analysis relies on a
multivariate model of the side-channel traces. Here we present
a practical and efficient implementation. CRA correlation
radio analysis is used to identify the parameters of the leakage
model. Fig. 1. Steps that lead to the radio transmission of the leak
All systems used in these experiments are using AES-128
(Advanced Encryption Standard). AES is a specification for The function multivariate_normal from scipy.stats is used
the encryption of different electronic systems and its data. for the normalization process in CRA. Pickle implements
One of the leading programming languages in data science binary protocols for serializing and de-serializing a structure
is Python. Although it might be a little bit slow, it gives a lot just like our configuration. The library itertools is used
of solutions to the most common problems. The best because it implements a number of iterator building blocks
advantage of the language is the wide variety of libraries and similar to Haskell. And finally binascii contains methods to
already developed tools in it. convert from binary to various ASCII-encoded binary
representations.

1Desislava
II. MATH MODEL OF THE ATTACKS
Nikolova is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
A. Correlation Radio Attack (CRA)
Bulgaria. E-mail: desislava.v.nikolova@gmail.com
2Ivaylo Vladimirov is with the Faculty of Telecommunications at

Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria, E-mail: ivaylo.h.vladimirov@gmail.com
3Zornitza Terneva is with the Faculty of Telecommunications at
Technical University of Sofia, 8 Kl. Ohridski Blvd, Sofia 1000,
Bulgaria. E-mail: z.terneva@abv.bg
The first step of the CRA is to generate an intermediate
value during the cryptographic operation, encryption or
decryption. The selection function must be chosen. This is a
function of a variable data value and part of the key. In the
equation below I is the intermediate value. The variable data

978-1-7281-7142-5/20/$31.00 ©2020 European Union 

sample which is generally plain text or cipher text is d and k is | | | |
parameters: , - the average value of the signal, noise
a part of the key which will be called the subkey from this
distribution also known as the covariance matrix of noise.
point on [2]:
They are calculated using statistical formulas (4), (5) and (6)
on all collected data for a given key. The final equation is (7).
I = f(d, k) (1)
| |
μ = ∑ χ (t) (4)
Then measurements must be done while the device is doing | |
the cryptographic operation. χ (t) − μ
In addition, the intermediate values for each of the used Θ i, : = ⋮ ⋮ ⋮ (5)
variable data samples and the power consumption for those | |
intermediate values must be calculated using a power model χ (t) − μ
such as the Hamming Distance Model. This calculation must | |
Ω = Θ Θ (6)
be repeated for all possible values of the subkey.
These calculated data values are the hypothetical power
consumption values.
The Hamming distance between D and R, also described by K ∗ = argmax exp − χ ∗ (t) −
| |
H(D ⊕ R), encloses the Hamming weight model which ( )
assumes that R = 0. The independent variable b encloses
offsets, time dependent components and noise and a is a scalar
gain between the Hamming distance and W the power | | | | | |
consumed (2): μ . Ω . χ ∗ (t) − μ (7)

W = aH(D ⊕ R) + b (2)

The last step is a comparison between the real and the III. PYTHON MODEL OF THE ATTACKS
theoretical power consumption for each subkey in order to
find the one that is most likely to be used in encryption, based A. Correlation Radio Attack (CRA)
on the highest correlation coefficient [3].
The variance dependencies are: σ = a σ + σ . Тhe Correlation Radio Analysis runs a standard correlation
correlation factor between the Hamming distance and the attack against a data set, trying to recover the key used for the
measured power to assess the linear model-fitting rate is . observed AES operations. The attack works by correlating the
The covariance between both random variables is normalized amplitude-modulated signal of the screaming channel with the
by the product of their standard deviations. The dependences power consumption of the SubBytes step in the first round of
can be observed on equation (3). AES, using a Hamming-weight model.
Firstly we load the known key, the number of traces and the
( , ) √ number of points. Then we create lists to store the information
ρ = = = = (3)
for the best guess, the PGE and the stored_cpas from maxcpa.
The following steps are for each and every one of the key
bytes:
B. Template Radio Attack (TRA)
The output of the CRA and its maximum value are creates
as empty lists.
Many models can be inferred using a template attack. The
For every guessed subkey, its value is printed. Then we
first step is to make a calculation about the distance between
initialize arrays and variables to zero. We create a hypothesis
the template probability density functions (PDFs). A template
using the HW model. The mean of the hypothesis and the
attack will give better results if the PDFs are far one from
mean of all points in the trace are performed for every subkey.
each other. As first order, it is proportional to the number of
And for every one trace of the guessed subkey, we are
commutations of the gates. The energy dissipated by a portion
calculating the difference between the hypothesis and the
of the hardware is thus measured by a Hamming distance.
difference in the traces. After this we calculate the output of
However, the best way to distinguish the templates is to build
the CRA and print the maximum value.
them in such a way that their dissipations are as different from
We find and print the PGE for every key byte in order to
each other as can be. As the dissipation is correlated to the
compare them easily.
Hamming distance, the best mapping is the Hamming distance
At the end we present the final results for the best achieved
between two successive values of the keys [4].
guess for every key and the PGE.
On average, it requires a lot of preparation and retrieval of
many power traces, but the advantage of the SRA is that it is
B. Template Radio Attack (TRA) - Learning
then necessary to collect very few traces of the attacked
system to completely break the encrypted communication. It
The data set should have a considerable size in order to
is even possible to extract the key from only one trace. The
allow the construction of an accurate model. In general, the
construction of a template consists in calculating the pair of


more data is used for template creation the less is needed to
apply the template.
Creating the template is the so-called training phase
consisting of measuring the electrical activity of a clone of the
system that will be attacked during a large number of
encryptions executed with K number of different keys on
random data sets. To perform it, the attacker must have access
to and full control over a copy of the system that will be
attacked.
This is necessary because you need to create models -
"templates" for each possible key that can be used. In practice,
it requires a lot of preparation and retrieval of many power
"traces", but the advantage of the TRA is that it is then
necessary to collect very few "traces" of the attacked system
to completely break the encrypted communication. It is
possible to extract the key from only one "trace".
In the code for every byte in the key we are using a
template S-Box from the S-Box, which is used as a lookup
table in AES. Then we sort the traces by the Hamming-weigh
(HW) model.
Make blank lists - one for each Hamming weight and fill
them up with the traces. Then we have to check if there is at
least a trace for each HW. They are switched to numpy arrays,
where we find the average values and the sum of their
differences.
The next step is to locate the points of interest, which is
done by finding the maximum sum of their differences. In
order to finish the training phase, we create the mean matrix
and the covariance matrix for each HW.
We display the points of interest and the two matrixes for
better understanding of the results.
C. Template Radio Attack (TRA) - Attacking
Template Radio Analysis applies the already prepared
template to a new system with the intention to attack the key
in a new data set. The template’s directory must be the
location of a previously created template with compatible
settings as same trace length.
The real attack starts first with creating the lists for the
scores, the best guess and the Partial Guessing Entropy (PGE)
this saves a byte which indicates the distance between real and
best guess. The closer it is to 0, the closer it is to the real
value. It can have a value of 255 if it is completely wrong
For each and every one of the bytes in the key, we are doing
the following steps:
In order to de-serialize the data stream of the point of
interest, covariance matrix and mean matrix, we use the
loads() option of the library pickle. To simplify and speed up
the calculations, the number of reports is reduced, leaving
only the points of interest.
Principal Component Analysis (PCA) is used for their
selection. This is a statistical approach allowing the
identification of patterns/trends/patterns in the data. In it, with
the help of a covariance matrix, only the most significant
values are selected. Creating these templates requires taking
measurements in about 20 hours, while collecting attack data,
with less than 1,000 traces, takes only 15 minutes. We create a

ring buffer for keeping track of the last N best guesses using a
window with start index 0 and 10 dimensions
The result of the total average values of the signal is the
covariance matrix.
For each trace, we take the key points and put them in a
new matrix. In addition, after this, each and every key is being
tested. A Hamming-weight (HW) model is made of the S-
Box. The HW is normalized and is transformed into a
probability density function. The logarithm of this one is
stored in a list together will all of the keys for a given trace.
The last list is sorted and sored into the main matrix. If a
proper key is found, we display the subkey and the number of
traces in which it was found.
The best guess for every key and the PGE are displayed at
the end.
IV. RESULTS
A. Running an Attack Command
In order to run the two analysis in the Linux terminal there
are a few parameters that you need to know. The most
important for the attacks are the number of traces --num-traces
(the maximum available if you pass 0), and the attack window
defined by --start-point and --end-point (or the entire trace if
you do not specify them) [5].
In order to choose the window one should plot (--plot) the
traces. Use the two shown plots to manually identify the first
round of AES, at the output of the S-Box. The attack works
better this way, because we eliminate a lot of points which
slow down the processing and add noise. The tool will first
show all the traces aligned on the same plot (careful if there
are many of them) and then the Sum of Absolute Differences
(SOAD) - which should be higher where the traces differ [6].
$ sc-attack --plot --data-path=../traces/mbedtls_1m_home –
name =mbed_villa_500 --start-point=0 --end-point=0 --num-
traces=1000 cra
Fig. 2. Example command
B. Running a template creation command
First, we use the template set to create a template. We first
create a folder where we can store the template. Again, we
could set the window, but we do not need it. The template
attack first uses the SOAD to detect the points of interest that
is the point of the trace that "leak". They are simply the peaks
of the SOAD. One can configure how many peaks you want
with --num-pois and the minimum distance between adjacent
peaks with --poi-spacing. Note that the number of traces is
very large, so it is better to use the plot option on a smaller
subset.
$ sc-attack --data-path=../traces/tinyaes_anechoic_10m --
start-point=0 --end-point=0 --num-traces=129661 tra_create
tra_templates/example_template --num-pois=10 --poi-
spacing=1
Fig. 3. Example command
 C. Numerical Results
The numerical results that appear after using the attack
command consist of a table with five lines. The first two lines
show the bytes of the best guess and of the known key,
respectively. The PGE is the Partial Guessing Entropy, which
is the "distance" of our best guess from the known byte, and
therefore it is 0 when the guess is correct. Our tool ranks the
guesses, with the most likely at 0 and the least likely at 255,
and the PGE is the rank of the known key.
Fig. 4. The result given in the terminal
D. Graphic Results
Fig. 5. Plot of all used traces aligned
The –plot tool of the sc-attack command will first show all
the traces aligned (see Fig. 5) on the same plot and then the
Sum of Absolute Differences (see Fig. 6) (which should be
higher where the traces differ, e.g. because of the key).
Fig. 6. Plot of the Sum of Absolute Differences graph

V. CONCLUSION
In this paper we presented an approach for software
implementation of two side-channels attacks with the main
goal of recovering the AES-128 key from an unknown
system. CRA and TRA can be used in the field of security
depending on the given model and its encryptions.
Although the opportunities for those systems are a lot, the
programming realization is fairly similar and can be optimized
in order to meet the required results. In our approach we are
using the Probability Density Functions and the Hamming-
weight model to calculate the Partial Guessing Entropy. The
results are plotted in order to be easily understandable and
analyzed. Because of its simplicity and variety of built-in
functions Python3 was used as the main programming
language.
The future of the field of security systems is not clear but
can have different directions depending on our focus and
research.
ACKNOWLEDGEMENT
The research described in this paper is partially realized
within the Erasmus+ mobility program with the help of the
profesors at SUPELEC University in Rennes, France.
The hardware implementation of the codes for the two side-
channel power analysis techniques used in this paper are
described in a scientific document: “Hardware
implementation and comparison of CRA and TRA when
trying to recover the AES-128 key” [7] that will also be
presented at the ICEST 2020 conference.
REFERENCES
[1] G. Camurati, S. Poeplau, M. Muench, T. Hayes, and A.
Francillon, “Screaming Channels: When Electromagnetic Side
Channels Meet Radio Transceivers”, CCS'18: Proceedings of
the 2018 ACM SIGSAC Conference on Computer and
Communications Security, January 2018, pp. 163-177.
[2] V. K. Rai, B. V. Reddy, S. Tripathy, and J. Mathew,
“Correlation Power Analysis and Effective Defense Approach
on Light Encryption Device Block Cipher”, Security and
Privacy, vol. 2, no. 5, September/October 2019.
[3] E. Brier, C. Clavier, and F. Olivier, “Correlation Power
Analysis with a Leakage Model“, Cryptographic Hardware and
Embedded Systems - CHES 2004: 6th International Workshop
Cambridge, MA, USA, August 11-13, 2004.
[4] M. A. El Aabid, S. Guilley, and P. Hoogvorst, “Template
Attacks with a Power Model“, 2007.
[5] H. Li, A. T. Markettos, and S. Moore, “Security Evaluation
Against Electromagnetic Analysis at Design Time”, Tenth IEEE
International High-Level Design Validation and Test Workshop,
Napa Valley, CA, 2005, pp. 211-218.
[6] G. Camurati, “Screaming_channels”,
https://github.com/eurecom-s3/screaming_channels;
[7] I. Vladimirov, D. Nikolova, and Z. Terneva, “Hardware
Implementation and Comparison of CRA and TRA when Trying
to Recover the AES-128 Key”, 2020.
 AUTHOR INDEX

A H
Acevski, Nikolce 119 Haralanova, Elena 244
Andonov, Aleksandar 37
Arapinoski, Blagoja 110, 137 I
Atanasković, Aleksandar 199 Iliev, Atanas 123
Atanasovski, Metodija 99, 106, 110, 137 Iliev, Desislav 133
Atanasovski, Vladimir 54 Iliev, Ilia 211
Avdić, Aldina 46 Iontchev, Emil 231

B J
Bitir Istrate, Ioan-Sevastian 145 Janakieska, Martina 54
Bychkov, Oleksii 167 Janjić, Aleksandar 102
Janković, Dragan 42, 46
C Jokanović, Branka 203
Campobello, Giuseppe 191, 195 Jovanović, Aleksandra 155, 159
Castravete, Miruna 145
Ceselkovska, Vesna 137 K
Chikov, Ognyan 73 Koničanin, Samir 224
Chikurtev, Denis 81 Kostadinova, Iva 167
Chikurteva, Ava 81 Kostina, Anna 235
Christoff, Nicole 181 Kostov, Mitko 99, 106, 110
Crupi, Giovanni 191, 195 Krstić, Nikola 127
Kushlev, Stoyan 185
D Kuzmanov, Ivo 252
Damyanov, Ilian 231
Dimishkovska, Natasha 123 L
Dimitrijević, Aleksandar 19 Laskov, Lyubomir 77
Dimitrov, Dimitar 123 Latino, Mariangela 191, 195
Dimitrov, George 167 Latkoski, Pero 54
Dimitrov, Kalin 65, 77 Lutovska, Monica 239
Djordjević, Andjelija 42
Djordjević, Miloš 58 M
Djordjević, Snežana 89 Madzharov, Nikolay 133
Djordjević, Teodora 50 Maleš-Ilić, Nataša 199
Djorić, Aleksandra 199 Malinović, Nemanja 11
Donato, Nicola 191, 195 Manivilovska, Ana 119
Dončov, Nebojša 211 Manolova, Agata 181
Draganov, Ivo 177 Marinković, Zlatica 191, 195, 220
Marinova, Galia 69, 73
F Markov, Bojidar 141
Filipović, Marija 89 Markova, Diana 141
Marovac, Ulfeta 46
G Merkulova, Kateryna 167
Georgieva, Veska 77 Micevska, Angela 119
Getova, Iglika 167
Gheorghiu, Cristian 145
Goranov, Goran 3
Gugliandolo, Giovanni 191, 195
Mijakovski, Vladimir 239 S
Milenković, Aleksandar 42 Sandjakoska, Ljubinka 163
Miletiev, Rosen 231 Sarevska, Maja 211
Milić, Dejan 224 Serov, Andrey 235
Milijić, Marija 203 Spirovski, Mile 110
Milovanović, Ivan 211 Stanković, Zoran 211
Mironov, Rumen 173, 185 Sterjovski, Igor 137
Mitić, Vladimir 89 Stevanovski, Blagoja 114
Mitrović, Nikola 85 Stojanović, Milan 85
Mojsoska, Natasa 114 Stojanovska, Frosina 163
Stojković, Aleksandra 58
N Stojković, Suzana 50
Nachev, Ivaylo 207, 216 Stošić, Biljana 151, 211
Nikolova, Desislava 256, 259 Strobl, Peter 19
Suljović, Suad 224
P
Panayotova, Galina 167 T
Pašić, Roberto 252 Terneva, Zornitsa 256, 259
Perić, Zoran 155, 159 Tola, Krisel 69
Petkov, Peter 207 Tzvetkov, Plamen 235
Petkova, Lora 177
Petković, Ivan 15 V
Petrov, Pavel 167 Valkovski, Tsvetan 65, 93
Petrova, Pepa 167 Vasić, Selena 224
Petrović, Nenad 23, 27, 33, 224 Veljanovski, Goran 99, 106
Petrušić, Andrija 102 Veselić, Boban 89
Popovski, Pande 99, 106, 114 Vladimirov, Ivaylo 256, 259
Predić, Bratislav 11 Vlajkov, Slobodan 155, 159
Pronić Rančić, Olivera 220 Vračar, Jana 58
Puleva, Teofana 244, 248
Y
R Yonchev, Andrey 248
Radojković, Miloš 220 Yordanov, Rumen 231
Radmanović, Miloš 7
Rajković, Petar 15 Z
Rankovska, Valentina 3 Zhabska, Yelyzaveta 167
Roganović, Miloš 11