Data Driven AI Assisted Green Network Design and Management FULLTEXT01

kth royal institute
of technology
Doctoral Thesis in Information and Communication Technology
Data Driven AI Assisted Green

Network Design and Management
MEYSAM MASOUDI
Stockholm, Sweden 2022

Data Driven AI Assisted Green
Network Design and Management
MEYSAM MASOUDI
Academic Dissertation which, with due permission of the KTH Royal Institute of Technology,
is submitted for public defence for the Degree of Doctor of Philosophy on Friday the 4th
February 2022, at 10:00 a.m. in Sal C, Electrum, Kistagången 16, 164 40, Kista.
Doctoral Thesis in Information and Communication Technology

KTH Royal Institute of Technology
Stockholm, Sweden 2022
© Meysam Masoudi
ISBN 978-91-8040-114-2
TRITA-EECS-AVL-2022:1
Printed by: Universitetsservice US-AB, Sweden 2022

iii
Abstract
The energy consumption of mobile networks is increasing due to an increase in

traffic demands and the number of connected users to the network. To assure the sus-
tainability of mobile networks, energy efficiency must be a key design pillar of the next
generations of mobile networks. In this thesis, we deal with improving the energy effi-
ciency of 5G and beyond networks from two perspectives, i.e., minimizing the energy
consumption of the network, and energy-efficient network architecture design.
In the first part of this thesis we focus on energy-saving methods at the base sta-
tions (BSs) which are the most energy-consuming component of mobile networks. We
obtain a data set from a mobile network operator which contains network load informa-
tion. It is a challenge to use mobile network traffic data to train ML algorithms for sleep
mode management decisions due to the coarse time granularity of data. We propose a
method to regenerate mobile network traffic data taking into account the burstiness of
arrivals. We propose ML-based algorithms to decide when and how deep to put BSs
into sleep. The current literature on using ML in network management lacks of guaran-
teeing any quality of service. To handle this issue, we combine analytical model based
approaches with ML where the former is used for risk analyses in the network. We
define a novel metric to quantify risk of decision making. We design a digital twin that
can mimic the behavior of a real BS with advanced sleep modes to continuously as-
sess the risk and monitor the performance of ML algorithms. Simulation results show
that using proposed methods considerable energy saving is obtained compared to the
baselines at cost of negligible number of delayed users.
In the second part of the thesis, we study and model end-to-end energy consump-
tion and delay of a cloud native network architecture based on virtualized cloud RAN
forming foundations of open RAN. Today large telco players achieved a consensus on
an open RAN architecture based on hybrid C-RAN which is studied in this thesis. Mi-
grating from conventional distributed RAN architectures to the network architectures
based on hybrid C-RAN is challenging in terms of energy consumption and costs. We
model the migration cost, in terms of both OPEX and CAPEX, with economic viability
analyses of a virtualized cloud native architecture considering the future traffic fore-
casts. It is not clear under what conditions C-RAN based architectures are more cost
efficient than D-RAN considering the infrastructure cost of fronthaul and fiber links.
We formulate an integer linear programming (ILP) optimization problem to optimally
design the fronthaul minimizing the migration costs. We solve the problem optimally
using commercial solvers and propose AI-based heuristic algorithm to deal with the
scalability issue of the problem for large problem sizes. Dealing with the trade-off
between network energy consumption and delay is a challenging issue in the network
design and management. In a multi-layer hybrid C-RAN architecture, we formulate an
ILP problem to optimize the delay by storing the popular contents in the edge closer to
the users and to minimize the network energy consumption. Moreover, we investigate
the trade-off between the overall energy consumption and occupied bandwidth in the
network. We demonstrate that intelligent content placement reduces not only delay
but also saves energy by finding a compromise between performance metrics. With
a similar objective of minimizing network energy consumption, we propose a method
for end-to-end network slicing, where logical networks are tailored and customized
for a specific service. As per literature, end-to-end network slicing is optimized for
iv
the first time considering the energy consumption. Most network slicing studies con-
sider only radio access network resources. Intuitively, energy consumption goes down
if more bandwidth resources is allocated to users when RAN segment of the network
is considered. However, with end-to-end energy consumption model, presented in this
thesis, it is demonstrated that increasing bandwidth allocation also increases processing
energy consumption in the cloud and the fronthaul segment of the network. To deal
with this issue, we formulate a non-convex optimization problem to allocate end-to-
end resources to minimize the energy consumption of the network while guaranteeing
the slices’ QoS. We transform the problem to second order cone programming prob-
lem and solve the problem optimally. We show that end-to-end network slicing can
decrease the total energy consumption of the network compared to the radio access
network slicing.
Keywords: 6G, 5G, Energy efficiency, Machine learning, Reinforcement learning,
Network architecture, Sleep modes, Mobile networks.
v
Sammanfattning
Energiförbrukningen i mobilnäten ökar ständigt i takt med ökade trafikvolymer

och det växande antalet användare. För att mobilnäten skall kunna vara hållbara, måste
energieffektiviteten vara en viktig designparameter. I den här avhandlingen föreslår vi
hur energieffektiviteten hos 5G-nät och framtida generationers nät kan förbättras ur
två perspektiv, dels genom att minimera de befintliga nätens energiförbrukning och
dels genom en energieffektiv design av framtida nätverksarkitekturer.
I avhandlingen presenteras först en översikt av olika energibesparande funktioner
i mobila nät. Därefter fokuserar vi på metoder för att minska energiförbrukningen i
basstationerna, där den största mängden energi förbrukas. Det finns flera metoder be-
skrivna i litteraturen där man försätter basstationen i olika energisparlägen. Vi föreslår
en trafikberoende metod för att maximera energibesparingen i basstationen. I denna
metod försätts olika delar av basstationen i olika energisparlägen av olika varaktig-
het när basstationen inte är aktiv. Vi definierar ett mått för att beräkna ”risken” för att
basstationen momentant befinner sig i ett energisparläge så att den inte kan betjäna
en användare som vill få access. Denna risk definieras som den fördröjning som upp-
står för användarna. Metodens prestanda utvärderas genom simuleringar baserade på
nätverksdata från verklig trafik från Tele2. Resultaten visar att betydande energibespa-
ringar kan uppnås i nätet, trots en låg risk.
5G-näten kommer att stödja ett brett utbud av tjänster med olika krav. Den nuva-
rande nätverksarkitekturen är dock inte bra på att tillhandahålla heterogena tjänster till
användarna med goda prestanda. För att förbättra dessa möjligheter behöver vi migrera
från konventionella nät till nya arkitekturer. Den nya, flexiblare, arkitekturen bör utfor-
mas effektivt både vad gäller kostnader och energi. Avhandlingens andra del ägnas åt
dessa frågor. Vi undersöker kostnaderna för att gå från konventionella nätarkitekturer
till molnbaserade arkitekturer, s.k C-RAN arkitekturer. Vi föreslår prestandamått för
end-to-end fördröjning och en energiförbrukningsmodell för C-RAN-arkitekturen. Vi
föreslår även s.k. ”edge caching” och nätverksdelning för att förbättra tjänstekvaliteten
(QoS) i näten samtidigt som vi minimerar energiförbrukningen. I avhandlingen visar
vi att metoderna kan spara energi samtidigt som de tillfredsställer användarnas krav på
tjänstekvalitet.
Keywords: 5G, C-RAN, nätverksarkitektur, nätverksdelning, maskininlärning
vi
Acknowledgements
Any long journey is destined to end and so is the PhD journey. During this duration I had
the opportunity to meet and work with lots of nice people and I received numerous helps
from my friends, family, and colleagues. Being at Radio Systems Lab at KTH was full
of experiences where I found very helpful colleagues and friends. I would like to express
my special thanks to Andres, Mats, and Ben for letting me to contribute the educational
activities, being teacher and Lab assistant. I would like to thank Ki Won and Marina for
their constructive comments and discussions in the RS Lab Friday meetings. My special
thanks go to my colleagues, Andres Laya, Amirhossein Ghanbari, Amirhossein Banuazizi,
Yanpeng, Peiliang, Istiak, Haris, Abbasi, Aftab, Forough, Hossein, Mustafa, Milad, Sara,
and Morteza, for their all supports, helps, and nice time we had together and wishing a
nice journey to our new colleagues Ozan, Ziant, Yasaman, Anders, and Mahmoud. Dur-
ing my PhD, I had the opportunity to collaborate with pioneer researchers in my field. I
would like to thank Amin, MohammadGalal, Ozlem for our nice and fruitful discussions,
collaborations, and meetings.
I am sincerely grateful to my supervisors Dr. Cicek Cavdar, and Professor Jens Zander.
They gave me the opportunity to personal and professional skills. We had a wonderful
and long meetings to discuss the novel ideas and research questions. I would like to thank
again Dr. Slimane Ben Slimane for reviewing my PhD thesis, Professor Muhammad A.
Imran for accepting to be opponent, Professor Emil Björnson for chairing the PhD defense
session and Dr. Pål Frenger, Professor Michela Meo, Professor Rolf Stadler, and Professor
Viktoria Fodor for accepting to be in a grading committee. My PhD was supported by two
European Celtic-Next projects, AI4Green and Soogreen. I would like to thank all of our
industry and academic partners for our insightful meetings and discussions with special
thanks to Professor Tijani Chahed from Telecom SudParis, France, and Dr. Azeddine Gati
from Orange Lab, France.
This acknowledgment would be incomplete without thanking my family. My parent,
brother, parent-in-law, and my brothers-in-law. My sincere and heartfelt and endless thanks
go to my wife, Zeinab, for her understanding and continuous supports and motivations
during this journey. Thank you for standing with me to overcome all challenges we faced.
This journey wouldn’t be great without you.
Meysam Masoudi,
Stockholm, January 2022.
Content
Content vii
List of Figures ix
List of Tables xi
List of Acronyms xiii
1 Introduction 1
1.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Challenges of Mobile Networks . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Surge in Traffic Data and Subscribers . . . . . . . . . . . . . . . 3
1.2.2 High Energy Consumption . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 High Total Cost of Ownership . . . . . . . . . . . . . . . . . . . 5
1.2.4 Smart Network Management . . . . . . . . . . . . . . . . . . . . 6
1.2.5 Carbon Footprint and Life Cycle Assessment . . . . . . . . . . . 6
1.3 Literature Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Green Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.2 Network Architecture Design . . . . . . . . . . . . . . . . . . . 10
1.4 Research Questions and Methodologies . . . . . . . . . . . . . . . . . . 17
1.4.1 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.2 Research Methodologies . . . . . . . . . . . . . . . . . . . . . . 20
1.5 Thesis Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 AI Assisted Green Mobile Networks 25

2.1 BS Sleep Mode Management . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Power Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Traffic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.2 Distribution of Traffic . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.3 Bursty Behavior of Traffic Data . . . . . . . . . . . . . . . . . . 28
2.3.4 Data Generation . . . . . . . . . . . . . . . . . . . . . . . . . . 29
vii
viii CONTENT
2.4 Risk Aware Sleep Mode Management . . . . . . . . . . . . . . . . . . . 29

2.4.1 Risk Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.2 Risk Management . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.3 Digital Twin Model . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.4 Virtual Network Model: A Hidden Markov Process . . . . . . . 32
2.4.5 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 BS Sleep Mode Management Problem . . . . . . . . . . . . . . . . . . . 35
2.5.1 Action Set and Environment . . . . . . . . . . . . . . . . . . . . 35
2.5.2 Reward Function Design . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Selected Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 AI Assisted Network Architecture Design and Management 41

3.1 Cloud RAN Network Architecture Power and Delay Models . . . . . . . 41
3.1.1 Conventional C-RAN Architecture . . . . . . . . . . . . . . . . . 42
3.1.2 Hybrid C-RAN . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.3 Network Delay Model . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.4 Network Power Model . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Migration Costs towards C-RAN . . . . . . . . . . . . . . . . . . . . . . 48
3.2.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2.2 Pool Placement Problem . . . . . . . . . . . . . . . . . . . . . . 50
3.2.3 Solution Methodology . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.4 Selected Results . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Content placement and Functional Split Optimization . . . . . . . . . . . 53
3.3.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Energy Efficient End-to-End Network Slicing . . . . . . . . . . . . . . . 56
3.4.1 Network Architecture for Slicing . . . . . . . . . . . . . . . . . . 57
3.4.2 Slice Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4.3 Minimization of Network Energy Consumption . . . . . . . . . . 60
3.4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Conclusions and Contributions 67

4.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Discussions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . 69
Bibliography 73
List of Figures
1.1 Ericsson traffic forecast by 2026 [1]. . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Ericsson mobile user subscriptions forecast by 2026 [1]. . . . . . . . . . . . 4
1.3 Three main 5G services [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Performance requirements comparison between 4G and 5G [3]. . . . . . . . . 7
1.5 Relationship between methodologies of LCAs for ICT goods, networks and
services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Evolution of RAN (DRAN/C-RAN/H-CRAN) [4]. . . . . . . . . . . . . . . 12
1.7 Functional split model [5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.8 The break down of thesis topics . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1 Data traffic model as interrupted Poisson process [6]. . . . . . . . . . . . . . 29

2.2 Digital twin assisted decision making [6]. . . . . . . . . . . . . . . . . . . . 30
2.3 Digital twin based risk aware sleep mode management [6]. . . . . . . . . . . 31
2.4 Markov model for advanced sleep mode management [6]. . . . . . . . . . . . 32
2.5 Network load generation as input to the network for τ = 0.1 and ζ = 0.5 [6]. 37
2.6 Energy saving vs delay trade-off [6]. . . . . . . . . . . . . . . . . . . . . . . 38
2.7 Normalized daily profile of energy saving [6]. . . . . . . . . . . . . . . . . . 38
2.8 Daily profile of energy consumption. . . . . . . . . . . . . . . . . . . . . . 39
2.9 Performance of risk management algorithm [6]. . . . . . . . . . . . . . . . . 39
3.1 Basic C-RAN architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.2 Two layer CRAN architecture. . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Multi layer C-RAN architecture. . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 An encoder example for 3 RUs, 3 pools, and 5 DUs per pool. RU 1 and 2 are
connected to DU 1 and 2 in Pool 1 respectively, pool 2 serves RU 3 and Pool
3 is inactive. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 TCO of Different Network Architectures [5]. . . . . . . . . . . . . . . . . . 52
3.6 TCO of C-RAN and D-RAN over the years, with TCO minimization, for
greenfield pool deployment, with and without FS [5]. . . . . . . . . . . . . . 52
3.7 Power consumption of the network [7]. . . . . . . . . . . . . . . . . . . . . . 55
3.8 Average delay of the users vs percentage of active users [7]. . . . . . . . . . . 56
3.9 Simple overview of slice manager . . . . . . . . . . . . . . . . . . . . . . . 59
3.10 Computation and communication resources assigned to each slice . . . . . . 59
ix
x LIST OF FIGURES
3.11 Network architecture and slice manager [8]. . . . . . . . . . . . . . . . . . . 61

3.12 Slicing life cycle management with QoS guarantee [8]. . . . . . . . . . . . . 62
3.13 Hourly profile of expected energy consumption and number of users for eMBB
delay sensitive and eMBB delay tolerant slices [8]. . . . . . . . . . . . . . . 64
3.14 Allocated processing resources to each slice and user [8]. . . . . . . . . . . . 65
3.15 Time complexity performance comparison of SBRA and UBRA [8]. . . . . . 65
List of Tables
1 List of frequently used acronyms and definitions . . . . . . . . . . . . . . . . xiv
1.1 Fronthaul capacity requirement for different split points . . . . . . . . . . . 14
2.1 Power and minimum durations from [9, 10] . . . . . . . . . . . . . . . . . . 26
3.1 List of parameters and their definitions. . . . . . . . . . . . . . . . . . . . . 45

3.2 5G frame structure [11]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
xi
List of Acronyms
xiii
xiv LIST OF ACRONYMS
Table 1: List of frequently used acronyms and definitions
Parameter Definition
ASM Advance sleep mode
BS Base station
CaPEX Capital expenditure
CC Central cloud
CoMP Coordinated multipoint
CRAN Cloud radio access network
DRAN Distributed radio access network
DT Digital twin
DU Digital unit
DRAN Distributed RAN
EC Edge cloud
eCPRI Enhanced common public radio interface
eMBB Enhanced mobile broadband
HCRAN Hybrid CRAN
ICT Information and communications Technology
IoT Internet of things
MEC Mobile edge computing
ML Machine Learning
mMTC Massive machine type Communication
OPEX Operational expenditure
PON Passive optical network
QoS Quality of service
RAN Radio access network
RL Reinforcement learning
RDM Risk of decision making
TCO Total costs of ownership
TWDM Time and wavelength division multiplexing
UE User equipment
URLLC Ultra reliable low latency communication
VBS Virtualized BS
VCRAN Virtualized CRAN
Chapter 1
Introduction
The information and communication technology (ICT) industry is rapidly advancing to-
wards 5G and beyond networks with the aim of fulfilling the ever-increasing demands for
higher data rates and offering multitudes of services with diverse quality-of-service (QoS).
These diverse requirements pose a big challenge for the operation and management of 5G
and beyond systems. On one hand, 5G architectures should be capable of satisfying the
demands for different services which require more bandwidth and less latency for certain
types of services. On the other hand, the surge in data traffic and the number of equip-
ment, yields in high investment cost and higher energy consumption [4]. Therefore, there
will be a need to revisit the current architectures and network management methodologies
and fully exploit the capabilities of advanced signal processing techniques and use amount
of data for a proper network planning and operations. Along with the rapid evolution of
network architectures, new challenges such as scalability and network management com-
plexity arise. Artificial intelligence (AI) and machine learning (ML) are promising tools
for handling the complexity of the network management. On the other hand, the conven-
tional approaches can no longer solve efficiently the new raised challenges and deal with
the complexity of the 5G and beyond networks.
After investigating the energy and delay aware resource allocations in the mobile net-
works in [12–25], in this thesis, we deal with the issue of reducing the energy consumption
of future mobile networks and enhance the sustainability of the networks by utilizing AI
and ML based techniques. We highlight the potential opportunities, challenges, and open
problems of energy saving at different segments of the network. In particular, we look into
the energy saving at the base stations, green network architectures, and methods to save
energy in the network while maintaining the offered QoS. In this thesis, we investigate the
open research questions on the energy saving at the base stations (BSs). In this part, we
put effort to save energy at the BSs and meanwhile trying to eliminate the performance
degradation. In this study, we aim at proposing a framework to determine and manage the
risk of using energy saving features at the base stations. This is very crucial in the network,
because usually energy saving comes at the cost of performance degradation. Hence, it is
of great importance to monitor and keep the performance degradation as low as possible.
1
2 CHAPTER 1. INTRODUCTION
For the next step, we look into the end-to-end network architecture and investigate the
open problems and research gaps in green network architecture design. We argue that the
current network architecture is not designed for multitudes and heterogeneous services and
we need to move on to the new architecture. Therefore, we investigate different aspects of
novel network architectures ranging from migration cost to them, to supporting numerous
services, and to tailoring network resources for specific services. In this section of thesis,
we mainly focus on end-to-end network design to 1) evaluate the migration costs to the
new green architectures, and 2) propose joint network resource allocation for improving
the QoS of the users and services while minimizing the energy consumption. The proposed
solutions are evaluated with numerical experiments and the experiments results well prove
the efficiency of the proposed solutions.
The remaining part of this chapter is structured as follows. We provide the background
an motivation of this thesis in Section 1.1. We highlight the challenges and opportunities
of energy saving in the mobile network in Section 1.2. We survey the state of the art and
the research gaps in Section 1.3. We elaborate the research questions and research method-
ologies in Section 1.4. Finally, the thesis contributions are summarized in Section 1.5.
1.1 Background and Motivation

Next generation of mobile networks, 5G and beyond, are required to simultaneously sup-
port multitudes of services categorized into three main classes, namely, 1) enhanced mo-
bile broadband (eMBB), 2) massive machine-type communication (mMTC), and 3) ultra-
reliable low-latency communication (URLLC). The required QoS for each category is ex-
tremely heterogeneous while number of services are growing [26–28]. On the other hand,
it is predicted that the total mobile network traffic will exceed 300EB per month in 2026 [1]
which results in more complicated network management and a surge in energy consump-
tion [4]. Therefore, the possibility and sustainability of supporting such a diverse services
require novel network architectures and enabling technologies that possess a green nature
to well adapt their energy consumption to actual network traffic; for instance, via smart
and agile traffic offloading accompanied by network component sleeping/shutdown, and
even further to enable end-to-end network slicing via a proper dynamic resource allocation
in different hours. In this regard, traffic intermittency and burstiness need to be reflected
on the network availability and on the actual network energy consumption, and the QoS
of users. Also, the allocated computation and communication resources to the offered
services should fairly consider their actual cost of occupancy and/or consumption of the
network infrastructure/resources and service requirements.
Network management schemes for energy efficient resource allocation are extremely
complex and conventional approaches cannot be further extended to future networks due
to their much higher density and complexity [24]. Meanwhile the need for energy saving
continues to vastly grow. Machine learning (ML) is a promising tool that can handle the
complexity of 5G and beyond systems while providing a competitive performance with the
other state of the art tools. Nowadays, AI and its subcategories such as machine learn-
ing and deep learning have been evolving as the most promising solutions to improve the
1.2. CHALLENGES OF MOBILE NETWORKS 3
Figure 1.1: Ericsson traffic forecast by 2026 [1].
performance and robustness of 5G and beyond systems. These solutions are fueled by
the massive amount of data generated in the networks. Implementation of AI becomes
feasible because of the availability of powerful data processing devices and techniques.
These mechanisms allow wireless networks to be both predictive and proactive. In recent
years, AI/ML techniques have been applied in many aspects of 5G design including radio
resource allocation, network management, etc, to improve the energy efficiency of the net-
work, provide higher QoS [29]. However, the impact of applying these solutions on the
overall or end-to-end performance needs more investigation.
1.2 Challenges of Mobile Networks
1.2.1 Surge in Traffic Data and Subscribers

According to the Ericsson traffic forecast [1], the total traffic of network including fixed
wireless access (FWA), and mobile data (2G/3G/4G/5G) will exceed 300EB per month
in 2026. As is illustrated in Fig. 1.1, a major share of mobile traffic belongs to 5G.
Moreover, as is depicted in Fig. 1.2, total number of mobile subscribers as well as 5G
subscribers are increasing over time. This surge in number of users and traffic data, make
the network management a complicated task and hence calls for new network management
schemes [30]. There is a consensus that there will be a three main categories of 5G services
as shown in Fig. 1.3. However, 5G and beyond networks are supporting multitudes of
services under these three main categories, to heterogeneous networks, technologies, and
devices operating in different geographic regions. Therefore, there will be a challenge to
provide dynamic, universal, user-centric, and data-rich wireless services to fulfill the vase
and heterogeneous QoS of services.
Figure 1.2: Ericsson mobile user subscriptions forecast by 2026 [1].
1.2.2 High Energy Consumption
The total electricity consumption of the ICT networks were increased by almost 60 per-
cent from 2007 to 2020 owing to the increasing number of devices, subscribers, services
and network expansion. There are concerns in the industry that 5G will dramatically in-
crease total mobile network energy use if deployed in the same way as 3G and 4G to meet
the increasing traffic demand. In this case, Telecom operators often require to add new
equipment while keeping existing network assets. This method is not sustainable from an
energy cost and environmental perspective. Therefore, continuous improvement in energy
efficiency is essential in order to balance the growth in user numbers and data volumes,
meet current and future traffic demands and simultaneously addressing network energy
consumption and related carbon emissions. In one hand, rolling out the next generation of
mobile networks is inevitable to meet the new service and requirements, on the other hand,
the energy consumption curve is increasing. Therefore, there is a crucial need to make
effort on breaking the energy curve by building, operating and managing networks in more
intelligent and strategic way.
Figure 1.3: Three main 5G services [2].
1.2.3 High Total Cost of Ownership

Mobile operators are willing to provide new services to the users. However, it it is not
energy efficient to provide heterogeneous services over one common architecture and in-
frastructure. Therefore, it is required to update the architecture and move from distributed
RAN to novel architectures that can accommodate higher data rates, lower energy con-
sumption, and satisfying the service requirements. On the other hand, it is very compli-
cated and costly to upgrade the currently deployed architecture to a new architecture. The
total cost of ownership (TCO) is divided into two parts, capital expenditure (CAPEX), and
operational costs (OPEX). With the emerging services and more data to be delivered to
users, the RAN costs, x-Haul costs, core infrastructure costs, energy costs, among other
may dramatically increase, hence, it is crucial to lower the OPEX while satisfying the
service requirements [31–37].
It is expected that the operators face more challenges in terms of CAPEX since the
network expenditure is forecast to rise significantly. Modernizing the network, contending
with new disruptive forces, private mobile networks will challenge the the value change of
the operators. In 5G, new spectrum is available with regulatory stipulations on coverage,
meanwhile operators may need to densify their network to roll-out 5G coexisting with
earlier generations, e.g., 4G and 2G. Therefore, intelligent CAPEX planning in the 5G era
can deliver a good QoS to the users while keeping the network cost as low as possible [38].
For instance, migration from distributed to centralized radio access networks (C-RANs)
can be expensive in terms of capital expenditures due to the initial investment while it
has lower operational expenditures due to pooling baseband processing into the cloud and
reduced power consumption. Then, with the assessment of the the crossover time, the time
it takes to have lower C-RAN TCO than distributed-RAN, one can decide when to migrate
to the new architecture. Therefore, TCO analysis plays a pivotal role and is essential in a
network migration, architecture design, and network planning.
1.2.4 Smart Network Management

The information and communication technology (ICT) industry is rapidly advancing to-
wards 5G and beyond, in order to fulfill the ever-increasing demands for higher data rates
and stringent quality-of-service (QoS). As can be seen from Fig. 1.4, compared to 4G, 5G
networks should be more energy efficient, support more number of connections, provide
higher spectral efficiency, and reduce the end-to-end delay. These diverse requirements
pose a big challenge for the operation and management of 5G and beyond systems. 5G and
beyond systems cannot meet their demands without the integration with artificial intelli-
gence (AI) and machine learning (ML), since the conventional approaches can no longer
solve the new challenges efficiently. Therefore, there will be a need to revisit the actual
methodologies and fully exploit the capabilities of AI/ML for a proper network planning
and operations.
Nowadays, AI and its subcategories such as machine learning and deep learning have
been evolving as the most promising solutions to improve the performance and robustness
of 5G and beyond systems. These solutions are fueled by the massive amount of data
generated in the networks and become feasible because of the availability of powerful data
processing devices and techniques. These mechanisms allow wireless networks to be both
predictive and proactive. In recent years, AI/ML techniques have been applied in many
aspects of 5G design including radio resource allocation, network management, cyber-
security, etc. However, the impact of applying these solutions on the overall or end-to-end
performance still needs more investigation.
1.2.5 Carbon Footprint and Life Cycle Assessment

1.2.5.1 Life cycle management
One way to evaluate the environmental impact of a product or process is life cycle as-
sessment (LCA). LCA is a systematic analytical method in which the potential impact of
ICT including equipment, networks, and service on the environment can be estimated [39].
LCAs have a cradle-to-grave approach where all the stages from raw material to the pro-
vided services and transport and energy supplies are considered. LCA became interna-
tionally standardized by the International Organization for Standardization (ISO) with the
publication of the ISO 14040 [40] series of life cycle assessment standards representing an
important step to consolidate procedures and methods of LCAs. The hand book published
by the The European Commission [41] provide the detailed analysis steps of conducting an
Figure 1.4: Performance requirements comparison between 4G and 5G [3].
LCA. Fig. 1.5 shows three parts of ICT related LCA which each requires its own method-
ologies. Among them, ICT goods are the physical equipment used in the ICT industry. The
other two parts, i.e., ICT networks and ICT services, are not physical entities but logical
structures that are made up by the ICT goods, including hardware and software. On the
other hand, they require building premises, civil works to create cable ways, air condition-
ing, power generators and power storage such as UPS and so on. The environmental impact
assessment of ICT networks reflects the environmental impact of ICT goods employed in
the ICT networks and the environmental impact assessment of ICT services reflects the
environmental impact assessments of ICT goods and ICT networks employed in the ICT
services. Therefore, it is difficult to define the assessment boundary for each part in detail.
However, it is important that the boundaries do not overlap to avoid any double counting
the environmental impact.
The cradle-to-grave approach to embodied energy consumption of the complete pro-

cess of delivering the services to the users. In this thesis, we mainly focus on the energy
consumption of the ICT networks and services. In particular, we focus on minimizing
the energy consumption in the next generation of cellular networks considering the novel
architectures and energy saving features.
Figure 1.5: Relationship between methodologies of LCAs for ICT goods, networks and
services
1.2.5.2 Carbon Footprints
Carbon footprint explains the environmental impact of a product or service over its entire
process based on a LCA method. The world population is producing more than 56 Bil-
lion tons of CO2. The total annual operational carbon emissions of the ICT networks is
estimated to 169 M tonnes CO2 in 2015 [42]. This corresponds to 0.53% of the global
carbon emissions related to energy (about 32G tonnes), or 0.34% of all carbon emissions
(about 50 G tonnes). According to recent studies [43], energy footprint of mobile net-
work operations has reached up to 1.4% of global electricity consumption showing 0.2%
increase compared to 2015. The carbon footprint of mobile networks has 14% share of
that of whole ICT sector. This high global footprint shows the impact of mobile networks
on the environmental challenges and therefore reveals the necessity to drive the research
trends towards the green mobile networks.
On the other hand, mobile network traffic and mobile user subscriptions are growing
exponentially over the years, mainly due to the democratization of smartphones and tablets
and the increase of video content. At the same time, Internet of Things (IoT) related traffic
will gain in importance with the explosion of the number of connected objects. In order to
keep up or improve customers quality of experience (QoE), network operators densify net-
works with additional equipment. This comes with an increased energy consumption. This
results in two main challenges: economic, as operators’ margin is decreasing (OPEX), and
environmental, in a context aiming at reducing greenhouse gas emission. These concerns
spark an increasing interest for optimizing networks’ energy consumption in the research
community, both public and private, academic and industrial.
1.3. LITERATURE STUDY 9
1.2.5.3 Thesis Contribution to LCA and Carbon Foot Print

In particular, this thesis is built around the requirement of reducing the energy consumption
of the networks including services and brings together both novel techniques such as Ma-
chine Learning and conventional techniques such as optimization theory to investigate on
not only the potential energy saving of the networks but also alleviate the carbon footprint
of mobile networks in ICT sector and globally. In particular, with the thyme of energy con-
sumption reduction, the contributions of the thesis is classified into two main categories,
namely, AI assisted green network management and AI assisted green network design. In
each of these categories, we seek to investigate the potential energy savings for the future
of mobile networks considering the services QoS requirements.
Considering service-based energy saving, each section of this thesis investigate specific
segment of the network to shed light on the potential focus areas not only to save energy but
also to save costs. In fact, it may not be possible that operators can make use of all proposed
solutions simultaneously, but it provides various methods that can be chosen according to
the situation. The proposed methods are divided into four groups: 1) Global energy saving
methods, 2) Energy saving methods at the BSs, 3) Service-based resource allocation for
network slicing and content placement. These energy savings are obtained by various
techniques and methods varies from the energy saving in one segment of the network to
end-to-end resource allocation to save energy. One should note that these methods are for
different cases (full traffic, low peak, average, etc), have different maturities (some can be
applied today, other are promising but still need further studies) and not all these methods
can be added and used simultaneously although some of them can.
1.3 Literature Study
1.3.1 Green Networks

Green networks has been intensively studied during the last decade considering both fixed
[44–46] and wireless networks [47–49]. Most of the studies in both fields focus on putting
different components of the network in sleep, reducing the active elements of the network
or power control challenging different network performance metrics, such as reliability,
throughput or delay [47–50]. Studies show that a considerable share of the overall net-
work energy consumption take place within BSs [51]. In order to maintain the operational
sustainability of such networks, it is crucial to devise efficient means for BS energy mini-
mization while avoiding any adverse impact on the offered QoS to served users, e.g., delay,
for data services [25, 52]. Recently in [10], it is shown that BS transceiver chain compo-
nents can be divided according to their (de)activation time scale into 4 groups. Each group
of components (with similar transition times) can be simultaneously put to sleep for a cer-
tain time duration, defined as sleep modes (SMs). Implementing SM can save BS energy
at the cost of bearing some service interruption. This interruption, e.g., delay/dropping,
may have more severe impact during daytime when network traffic congestion usually
increases. Therefore, meticulous SM implementation is required to mitigate QoS degra-
dation. The authors in [53] study cell DTX in heterogeneous networks and investigate the
impact of BS density and traffic load on potential energy savings. The study in [54] aims at
load-adaptive SM management considering deep, long-term SM subject to QoS guarantee.
In [55], the ON-OFF state of 4G BS is evaluated, quantifying the impact of BS sleep on
the QoS, e.g., dropping/delay. In [56], the authors evaluate the ON-OFF state of mm-wave
small cells, considering cellular traffic data from Tokyo. In [57], the authors propose a
stochastic model to perform dynamic tuning of the BSs configuration using ASMs.
Recently, machine learning (ML) techniques have attracted the great interests to be
used in the management of BS sleep modes (SMs) operations. However, implementing
such SMs can save BS energy at the cost of incurring serving delay to users. Thus, it is
crucial to devise novel yet efficient schemes to utilize ASMs without any adverse impact
on the offered QoS to the served users. Prior works report promising energy saving at
BSs utilizing ML in ASMs [58–63]. The authors in [58–60] investigate 4 SMs with pre-
defined order of (de)activation of SMs. They decide on duration of each SM. In [58] a
heuristic algorithm is proposed to implement the ASMs. They investigate the impact of the
synchronization signaling periodicity on the energy saving performance of ASMs. In [59],
the authors propose a dynamic algorithm based on Q-learning for SM selection. In [60],
they propose a traffic-aware strategy to build a codebook for mapping the traffic load to the
possible actions using Q-learning algorithm. In these studies, the authors assume a pre-
defined order of the decisions which may not result in optimal energy saving. The authors
in [61], relaxes the pre-defined order of SMs and propose a traffic-aware Q-learning based
algorithm for SM selection. However, they train their network based on fixed traffic pattern.
Since control signaling can limit the energy saving of ASMs, the study in [62] proposes
a control/data plane separation in 5G BSs which allows to the implement deeper SMs,
e.g, SM4, or longer SM duration. This separation can also be leveraged in other network
settings such as co-coverage scenarios when basic coverage cells may carry all periodic
control signaling, leaving capacity-enhancing ones with higher ability to sleep and save
energy. In [63] we propose to differentiate between the load levels and we consider High
and Low load levels as part of our states. However, this study is based on tabular methods
which 1) cannot capture long/short term time dependencies and 2) it is limited to two load
levels.
1.3.2 Network Architecture Design

On one hand, energy consumption of mobile networks is increasing, on the other hand
emerging number of services are going to be offered in the network. Current distributed
network architecture is not capable of supporting heterogeneous services efficiently in
terms of energy consumption and capacity. Therefore, novel architectures must be de-
signed considering challenges from several perspectives, including but not limited to en-
ergy consumption, cost, operation and maintenance, delay, handovers, among others. Net-
work architectures with centralization and virtualization of resources are promising solu-
tions for such a network with vast number of services and increasing energy consumption.
Significant research efforts have been recently focused on C-RAN technology. The main
idea of C-RAN is, for the sake of cost and energy consumption reduction, to centralize
all functionalities pertaining to digital baseband computation and cooling at certain cluster
centers, only leaving behind RF and analog processing at the cell sites. On top of the afore-
mentioned benefits, this centralization can also be an enabler for coordinated multipoint
(CoMP) technology, allowing for more rapid inter-cell coordination on the millisecond
scale, hence better resource utilization and EE, especially for cell-edge users.
Despite appealing features of C-RAN architectures, there are disadvantages to be ad-
dressed in the literature. When all processing units are centralized, the capacity of the links
between radio units and center cloud may not be enough at some hours of the day. The
medium between radio units and center cloud must be scalable and low cost, meanwhile
have enough capacity to support transformation of raw I/Q signals to the center cloud. Fur-
thermore, an extra transportation delay is introduced in the network in which may make it
challenging to support delay sensitive applications. C-RAN architecture has been evolving
during past years to support multiple services requirements while maintaining the energy
consumption as low as possible. One way of improving the C-RAN architecture is to
add another layer of processing enabling the distribution of processing tasks between the
center cloud and the new layer called edge cloud. We call this flexible semi-centralized
architecture herein as hybrid C-RAN (H-CRAN) as proposed in SooGREEN. H-CRAN
is a multi-layer architecture where physical layer network function splits are optimized to
minimize power and midhaul bandwidth jointly. H-CRAN leverages the previous C-RAN
structure with functional splitting in a three-layer architecture to share the processing tasks
between CC and EC. In Figure 1.6, we depict the evolution of distributed RAN (DRAN) to
H-CRAN. In DRAN, each radio unit is equipped with digital unit and they are connected
to the core network via backhaul link. In C-RAN the idea is to pool all digital units and
RUs are connected to this pool via fronthaul link. In H-CRAN, another layer of processing
is added between RUs and the DU pool (central cloud), where the processing functions can
be distributed between processing layers. In the context of realizing a practical H-CRAN,
several research efforts have been exerted. H-CRAN consists of three layers namely, cell
layer, EC layer, and CC layer. The cell layer includes RUs that are being densified, each
serving several user equipment (UEs). The RUs are connected to the ECs. In fact, an EC
mainly acts as an aggregation point where the data of a group of RUs is collected at this
point. The fronthaul between the RUs and ECs is assumed to be mmWave links. Further-
more, the ECs transmit the aggregated data to the CC via midhaul. In this architecture,
the CC and ECs are equipped with DUs which are able to perform function processing of
requested content. Therefore, these DUs can serve any connected RUs by sharing their
computational resources. For instance, in upstream, traffic from cells can be partially pro-
cessed at the EC so that bandwidth requirement can be relaxed for midhaul links, then the
remaining processing will be conducted at the CC. However, the EC is usually less energy
efficient than the CC, because the number of DUs at the CC is larger than that in each EC.
Hence, sharing infrastructure equipment offers a multiplexing gain that results in higher
energy saving at the CC. The tradeoff becomes whether to save midhaul bandwidth with
improved delay performance by distributing functions at the ECs or to gain from power
saving which is an intrinsic feature of centralizing all functions at CC. Since in H-CRAN
we have a processing layer closer to the users, delay sensitive services can be served at the
EC, while delay-tolerant applications are served at the CC to benefit from energy saving
features of the cloud. We further distribute the processing functions to tackle the problem
Figure 1.6: Evolution of RAN (DRAN/C-RAN/H-CRAN) [4].
of capacity limitation in the fronthaul considering the delay constraints.
1.3.2.1 Functional Split

Although the C-RAN idea is a promising architecture, several challenges remain to be
tackled to practically apply it side-by-side with the arising technologies. Compared to its
preceding DRAN architecture, full centralization of physical-layer processing, including
channel coding, modulation, and Fast Fourier Transform (FFT), can significantly upscale
the fronthaul signal bandwidth in several orders of magnitude, especially as the number of
serving antennas and the number of spatial streams grow larger. That is, such an effect can
become much more limiting when technologies like massive MIMO are employed. This
challenge also exists in uplink and downlink directions alike. For instance, the fronthaul
bandwidth also scales up when higher signal resolution (finer quantization) is required in
uplink multi-user detection scenarios, like in CoMP joint reception. Therefore, more con-
venient, and potentially flexible, intermediate functional splits need to be further studied
to both keep the scalability benefits of C-RAN while resolving the bandwidth congestion
and allowing for effective CoMP coordination for cell-edge users.
To study the distribution of function processing between EC and CC, we model the
functional splitting of baseband processing chain for cells and users, as shown in Fig.
1.7. The baseband processing functions can be classified as cell processing (CP) and user
processing (UP) functions. The CP includes the functions within the physical layer that
are responsible for the signal processing associated with the cell. Few examples of CP
functions include serial to parallel encoding, FFT, cyclic prefix, and resource mapping.
Similarly, the UP includes a set of functions that are related to the physical layer and
some upper layer functions that are responsible for signal processing of each user in a cell.
Few examples of UP functions include antenna mapping and forward error correction.
According to Fig. 1.7, the functional split can either happen before Split 1 or after Split 7
or in between. When split happens at Split 1, then all the functions are centralized at the
CC resulting in C-RAN. When split happens after Split 7, all the functions are centralized
at EC resulting in DRAN. When Split happens in between, the functions above the split
Figure 1.7: Functional split model [5].
are placed at the CC and functions below the split are placed at the EC.
1.3.2.2 Network Architectures and Migration Costs

Despite the attractive advantages of C-RAN, its deployment brings considerable challenges
for the operators. The progress on C-RAN as well as its major challenges are discussed
in [64]. When all the network functionalities in the base band unit (BBU) are fully cen-
tralized, C-RAN may require approximately 160 Gb/s of throughput with 10µs − 250µs
of latency [65]. If lower physical layer (PHY) functions are kept co-located with the RUs,
these requirements become looser. For instance, the fronthaul transport link requirement
for lower level PHY functional, e.g., split 4 in Figure 1.7, is about 933 Mbps which is only
about 40% of a fully centralized scenario. Figure 1.7 shows that the baseband process-
ing functions contain a chain of functions that are categorized into cell-processing (CP)
and user-processing (UP) functions. CP functions are dedicated to signal processing and
physical layer functions, e.g., serial to parallel, common public radio interface (CPRI) en-
coding, and resource demapping. UP functions are dedicated to per user signal processing,
e.g., equalization and turbo coding. Using functional splitting in C-RAN, it is possible to
split this processing chain at certain points [4]. The requirements for the fronthaul link
between the cloud and radio unit depend on the functional split point. Maximum benefit in
terms of energy saving and virtualization is attained when all the functions are centralized
at the cloud which is equivalent to traditional C-RAN. However, this might not be possible
due to the limited fronthaul capacity, hence using different split points supported by the
cost-effective fronthaul is more preferable [66].
In Figure 1.7, the possible split points are depicted and in Table 1.1 the capacity re-
quirements in the fronthaul links are summarized [7, 66]. These requirements make the
passive optical networks the most suitable transport medium in the fronthaul [67]. There-
fore, from the design perspective new challenges should be considered. These challenges
include, (i) optimal placement of pools to centralize the processing, (ii) optimal design of
fronthaul to guarantee the requirements, and (iii) cost of the network.
Table 1.1: Fronthaul capacity requirement for different split points
Capacity requirements (Mbps)

Split point: 7 6 5 4 3 2 1
Downlink 151 151 173 933 1075 1966 2457.6
Uplink 48 49 452 903 922 1966 2457.6
To tackle such challenges, the authors in [68] proposed an energy efficient aggregation
network and a pool placement optimization problem, to minimize the total aggregation in-
frastructure power. The authors in [69] proposed an optimization problem to determine the
configuration of each DU pool, which minimizes the deployment cost. The fronthaul con-
straint, which must guarantee a certain latency, the infrastructure deployment, and the pool
placement, are among the main challenges of C-RAN. The authors in [70] presented an
overview of the fronthaul requirements, and proposed architectures and transmission tech-
nologies for next generation optical access networks. The authors in [71] addressed the
fronthaul constraint problem by proposing several solutions, such as signal compression
and quantization, coordinated signal processing, and radio resource allocation optimiza-
tion. The study in [72] investigated the impact of wired and wireless backhaul on the
delay performance of cellular networks. In their study, the infrastructure cost of backhaul
deployment was also considered.
In order to assess the monetary aspect of the network, a thorough cost model of the
network is required. The authors in [73] modeled the power consumption of backhaul
with different scenarios and investigated the impact of the backhaul network to TCO in
cellular networks. They demonstrated that if a power efficient backhaul solution such as
fiber optic architecture is adopted, the share of backhauling power consumption will be
small compared to the radio access part. The study in [74] attempted to model the total
cost of backhauling with two different technologies, i.e., microwave and fiber. Based on
this study, it was shown that fiber is the most promising backhaul technology with high
capacity and low delay.
To provide a techno economic framework of C-RAN architecture, all the network costs
must be considered. Therefore, the studies [32,75–79] take into account TCO as their eval-
uation metric. In [75], the authors presented a methodology to evaluate TCO for different
technologies. Their study mainly focused on migration cost of optical access technolo-
gies considering infrastructure and technology upgrade. The authors in [76] proposed a
TCO model that considers scalability, CapEx and OpEx, as well as the uncertainty in dif-
ferent stages such as demand and penetration index. They have shown that according to
resource designation, there is a great sensitivity in the maximum expected benefit. The
authors in [77] analyzed the TCO of network migration towards passive optical network,
concerning both infrastructure and technology upgrades. In their study, they considered
different migration starting times, customer penetration, node consolidation, and network
provider business roles in the fiber access networks. The study in [78] presented a compre-
hensive techno economic framework able to assess both the TCO and the business viability
of different deployments. The authors stated that TCO alone is not sufficient to understand
the profitability of different architectures and more parameters such as net present value
(NPV) and cash flow should be considered. The study in [79] modeled the costs of fiber
and microwave architectures and calculated the TCO for different geographical regions. In
their study, the optimal fronthaul design and functional splitting is missing.
The study in [32] formulated TCO minimization problem as a constraint programming
problem. In this study, the authors found the optimal functional split point for radio unit
and digital unit pair. However, the problem of finding optimal location for pools was not
resolved. The study in [80] provided a cost analysis of the deployment of the wireless
and optical x-haul segment for a fixed wireless access network. They considered RAN
functional splits with diverse requirements capacity and coverage requirements.
1.3.2.3 Content Placement
Caching is another approach that may not only relieve the fronthaul congestion but also
can reduce the content delivery latency. In caching, popular contents are cached into a
place closer to the users, e.g., in the edge cloud (EC), allowing user content demands to be
accommodated more easily and quickly. Since content access delay is an important factor
in caching problems, various algorithms and techniques have been proposed to incur lower
latencies. The challenges, paradigm, and potential solutions for caching are discussed
in [81]. In [82], cooperative hierarchical caching has been proposed to minimize the con-
tent access delay and boost the quality-of-experience (QoE) for end users. In [83], the
authors proposed caching algorithms to optimize the content caching locations and hence
reduce the delivery delay. In [84] the authors presented a caching structure and proposed
a cooperative multicast-aware caching strategy to reduce the average latency of delivering
content.
H-CRAN leverages the previous CC/EC structure with functional splitting in a three-
layer architecture to share the processing tasks between CC and EC. In [85], an end-to-
end delay model has been proposed for different functional split options. H-CRAN can
simultaneously employ caching and functional splitting to tackle the fronthaul bottleneck
problem. Although the existing content caching algorithms can reduce the service delay,
it is not easy to decide where to deploy the content caches since there is a trade-off in
balancing the centralized function processing and the distributed caching especially in H-
CRAN. It is worth noting that caching the content at the EC prevents us from functional
splitting since the content is already at EC and it is not meaningful to centralize processing
at CC. Due to this dependency, it is important to jointly decide whether to centralize or
distribute content caching together with network processing functions.
1.3.2.4 Network Slicing

The concept of network sharing was initially introduced by universal mobile telecommu-
nications systems (UMTS) in Rel.99. 3GPPP Working Group [86] where they identified
five scenarios in which different segments of the network could be shared, 1) multiple core
networks sharing a mutual RAN, 2) operators with different coverage area share their in-
dividual frequency licenses and RAN, 3) share coverage on specific regions, 4) sharing
spectrum, and 5) share core network. Network Slicing can be perceived as an extension of
network sharing. Network slicing aims to provide network resources to guarantee service
customization, isolation and multi-tenancy support on a common physical network infras-
tructure. A network slice is an isolated end-to-end network enabling virtualized and inde-
pendent logical networks, tailored to a specific requirement, on the same physical network
infrastructure. Therefore, the network architecture become capable of providing network
slicing [87].
The study in [87] provides a comprehensive survey on the slicing, its use cases, chal-
lenges, and benefits. The study in [88], investigates how network slicing can enable
multi-tenancy in 5G networks. It proposes a framework in which mobile virtual opera-
tor (MVNO), over-the-top (OTT), and vertical industries interact with the infrastructure
provider. The authors in [89], present a generic framework for network slicing by defining
three layers, namely, service layer, network function layer, infrastructure layer. Then a
separate layer, i.e., management and orchestration entity monitor and manages the afore-
mentioned layers. The white paper published by the Next Generation Mobile Networks
Alliance (NGMN) [90] introduced the three layers of network slicing as 1) service layer
that represents services for the end user; 2) network slice layer that provides the network
characteristics required by a service layer; and 3) resource layer that includes communi-
cation and computation resources, network architecture, and network functions. Mobile
operators orchestrate the network functions and associate the user services with network
resources to create an end-to-end slice according to the service-level agreement (SLA) re-
quirements. The network slice is subdivided into radio access network (RAN), transport
network, and core network, respectively.
One of the significant challenges for the realization of a network slice is to translate
the high-level description of the service to the concrete end-to-end slice in terms of in-
frastructure and network functions [91, 92]. There are numbers of studies that consider
network slicing. But their focus was mainly on slicing in RAN, routing, and in Core net-
work separately. However, network slicing incorporates the whole network from cloud to
RAN. Since different segments of the network are related to each other and have an im-
pact on each other it is of great importance to look at the end-to-end network when we
are performing network slicing. In the literature there are very few studies that consider
the problem of network slicing for RAN and core network jointly. In these studies, the
problem in RAN and core network is solved separately and the discussion on the impact
of fronthaul is missing. Moreover, the delay and power model of these studies are very
simplistic.
The studies in [91, 93–97] investigate the end-to-end network slicing. The authors
in [93] provide a comprehensive survey on end-to-end network slicing models. In [91],
1.4. RESEARCH QUESTIONS AND METHODOLOGIES 17
the authors study the challenges of mapping together the available resources in the various
layers of the network slicing architecture in order to compose an end-to-end slice. In [94],
the authors propose an end-to-end slicing approach in RAN, transport network, and core
network. However, the analysis of the impact of one on another segment is missing. In [95],
the authors propose an end-to-end network slicing framework for 5G resource allocation.
The proposed framework consists of RAN and core network. In RAN, they dynamically
allocate wireless resources to slices and base stations (BSs), and the aim is to maximize
the total rate in the system. The authors in [96] are dynamically allocating RAN resources,
i.e., bandwidth, and cloud resources, i.e, baseband processing resources, to each slice. The
objective of this framework is to minimize the violation of rates in each slice according to
their SLAs. The authors in [97] propose an algorithm to avoid over-provisioning resources
to the slices while still satisfying the latency constraints. They show that the bottleneck
for enhanced mobile broadband services is computing resources while the bottleneck for
URLLC services is the communication resources.
In the literature, despite the studies on the end-to-end resource allocation for each slice,
the end-to-end network slicing models are not yet well investigated. RAN slicing is one
of the key parts of network slicing. In RAN slicing, the communication resources, e.g.,
bandwidth, can be dynamically allocated to slices depending on their loads in the net-
work. In the literature, RAN slicing has been investigated from different perspectives. The
authors in [98], discuss how the RAN can be sliced to satisfy the heterogeneous service
requirements. In this study, different sharing concepts and RAN slicing challenges have
been described. In [99], the authors propose a network slicing-based architecture, focus-
ing particularly on the RAN, and elaborate how artificial intelligence (AI) can potentially
assist the network in terms of resource provisioning. The study in [100], formulates a dy-
namic optimization model considering power consumption and service quality. In their
model, they take into account the limited radio resources, power resources, and latency
constraints. The study in [101], formulates a framework for RAN slicing in which the
delay performance is improved by content caching schemes. The authors in [102] aim
to maximize bandwidth utilization while guaranteeing quality of service (QoS) of users
in their slice. The study in [103], proposes a two-timescale radio resource allocation for
network slicing where one timescale is for long-term dynamic characteristics of service
requests and the other one for adapting the allocated resources in short time periods. The
study in [104] proposes a two-timescale RAN slicing where one is for resource reservation
and the other one is for resource allocation within each slice in a C-RAN architecture.
1.4 Research Questions and Methodologies

1.4.1 Research Questions
Motivated by the conducted literature review, in this section we aim at specifying the re-
search gaps and highlight the open research questions in the area of green mobile network
management and network architecture design. The literature on green mobile networks
has been around for more than a decade, aiming at minimizing the energy consumption of
the different segments of the mobile networks including 2G/3G/4G/5G and beyond while
maintaining the users’ QoS. Most of the architecture designs and energy saving algorithms
in the literature are considering static traffic and hence neglecting the dynamicity of the
network. Solutions are missing that can enable practical implementation of energy sav-
ing features due to the unpredictable risk of putting network resources in sleep, which
can cause degradation of network performance. Today mobile networks are designed for
peak traffic, most of the time idle but not sleeping. Radio and computational resources
are under-utilized in conventional radio access network architectures where processing re-
sources are co-located with the radio sites. In this thesis, we will investigate the high level
research question (HRQ) of
"How to design green network architectures and dynamically allocate network resources
to minimize energy consumption considering the network state, traffic load, and demanded
QoS?" and describe steps and methodology to approach the solutions that can address the
high level research question and contribute to the answers to this question.
To address this question, we leverage machine learning based algorithms and real network
traffic obtained from a mobile network operator. We will explain our studied open re-
search questions, problem formulation, and our solution approach to fill the research gap
and respond to the need of improving the energy efficiency of the network by reducing
the energy consumption, improving the resource allocation schemes, and improving the
network architecture design.
In order to approach the high level question, we break it down into two sections. In the
first section, we first look into intent-based1 risk-aware energy saving methods to be used
in the network and we focus on the most energy consuming section of the network, i.e.,
base stations. One way of saving energy at the BSs is to put it into sleep when there is no
users attached to it. Selection of different sleep modes enables load adaptive sleep options
by deciding on when and how deep to sleep. Thus, we narrow down the research question
and put effort to answer the following detailed research questions:
• RQ1: How to dynamically put BSs into sleep modes when they are idle, i.e., when
and how deep to sleep, given the previous user arrivals and network load?
• RQ2: Depending on how deep BS sleeps, how can we design a network management
framework considering mobile operator’s intent, continuously analyze and monitor
the risk while using ML to make decisions?
The current network needs to be upgraded to meet the requirements of 5G and beyond
networks. As per literature, cloud-radio access network (C-RAN) based architectures are
promising architectures to support scalable network infrastructure growth to meet with the
need for higher data rates while maintaining the energy consumption as low as possible.
However, there are multiple network architecture options to realize the concept of C-RAN
depending on the traffic intensity and available optical Fiber infrastructure. Each network
architecture proposal has its own pros and cons. Therefore, there is a need for investigating
the proper choice of network architecture to upgrade to, in terms of migration costs and
energy consumption. Therefore, as the second part of this thesis, we study the network
1 Intent is defined as the goal and business objective of the mobile operator
1.4. RESEARCH QUESTIONS AND METHODOLOGIES 19
architectures from the end-to-end perspective and put effort on investigating different as-
pects of such architectures including the migration costs, energy consumption, delay, and
end-to-end resource allocation schemes. First, we answer the following questions:
• RQ3: What is the deployment cost of different C-RAN architectures? How sensitive
an architecture is to the price changes?
• RQ4: How artificial intelligence can help us to design a C-RAN fronthaul with min-
imum cost and energy consumption?
In order to address the fronthaul capacity limitations in C-RAN, hybrid C-RAN is

proposed by enabling network function splits and by adding another layer of processing
called edge cloud between the radio units and center cloud. However, hybrid C-RAN, as
one of the promising architectures capable of meeting the requirements of 5G and beyond
networks, requires tuning the end-to-end design parameters, such as functional split point
depending on the network status and demanded QoS by users. In order to satisfy the QoS
while minimizing the energy consumption, we consider a content placement problem in
hybrid C-RAN in which popular contents are stored at the edge cloud to reduce the delay
of providing the content to the users, while minimizing the energy consumption. This is a
challenging problem since placement of contents at the edge cloud disables centralization
of network functions and connected energy saving opportunity. In this context, we answer
the following research questions:
• RQ5: Is there a way to save energy while meeting with the delay requirements of
users by placing content to the edge cloud instead of centralized cloud?
• RQ6: How to allocate network resources in fronthaul, edge cloud, and centralized
cloud to minimize the energy consumption while meeting the users’ QoS?
5G and beyond networks are supporting and providing increasing number of services
with vast and heterogeneous QoS. Considering this trend, one architecture does not fit for
all service. Network slicing can help the operators to tailor their network for specific ser-
vices and dedicate network resources to the services so that all are served with satisfying
QoS. An end-to-end network slice spans across multiple network segments including ac-
cess network, core network and transport network. A network slice comprises dedicated
and/or shared resources, e.g., in terms of processing power, storage, and bandwidth and has
isolation from the other network slices. This advanced reservation of predefined resources
for each slice may result in under utilizing the network resources and possibly increase in
energy consumption. Hence it requires meticulous network management and resource allo-
cation schemes to allocate resources to each slice while maintaining the network’s energy
consumption as low as possible. In this thesis, we will investigate the following research
questions under the context of network slicing.
• RQ7: How to allocate computation and communication resources to each slice for
satisfying the QoS and minimizing energy consumption?
• RQ8: How machine learning capabilities can be used in slice admission for improv-
ing the performance of dynamic slicing in terms of energy saving?
• RQ9: How the designed algorithms can be trained over real network data to make
intelligent decisions about slices?
1.4.2 Research Methodologies

This section will present a brief explanation of the research approaches followed in this
dissertation in order to answer the above mentioned research questions. In order to conduct
our research, we mathematically model the studied system and tune the parameters and
validate it using the measurement data obtained from the real network.
The obtained real network data was from a multiple base stations in the live network
of a Swedish operator. The original data is not clean and could not be used in its original
form. We clean the data and process it to prepare it for our studies. Due to ethical consid-
erations and to ensure trust, accountability, mutual respect, and fairness, we re not able to
show the obtained data in our research results. Therefore, we normalize the data and re-
move the sensitive information in the data set, such as locations of BSs, users information.
Moreover, we keep the data in a safe local storage device with a limited access. Then, we
formulate the optimization problems that can address our research problems. In order to
solve the problems, we leverage tools from optimization theory and artificial intelligence
and propose algorithms to solve the problems. We evaluate the the proposed algorithms by
doing simulations. We compare our solutions with multiple baselines such as optimal so-
lutions, an existing solution in the literature, lower/upper bounds, and plain system without
any energy conservation schemes.
In particular, in the first part of our research studies, we focus on the minimizing the
BS energy consumption. In this context, we mathematically model power consumption of
the operational and idle mode of a BS. Then, in order to address the research questions, we
define the problem of minimizing the power consumption of the BS while minimizing the
users’ QoS degradation. We propose a AI-based algorithm to solve the problem. We eval-
uate the proposed algorithm by simulation in python and compare our results with lower
bound, non-causal optimal solution and one existing solution in the literature. In the sec-
ond part, we focus on the network architectures where we need to mathematically model
the network in terms of delay and energy performance. With the help of optimization the-
ory, we formulate the problems of energy saving, fronthaul design, and network resource
allocation. The formulated problems are hard to solve and requires approximations and
algorithm design. Therefore, we benefit from integer linear programming algorithms and
AI/ML tools to solve the formulated problems. Then, we investigate these models and find
solutions to address the research questions. In these studies, we solve the problems using
both optimization theory and AI based techniques. In particular, we used reinforcement
learning and deep learning to solve the complex problems. Simulations are done in Python
and Matlab environment and for the system level simulations we use Monte carlo simu-
lations. We compare the results with (non-scalable) optimal solutions and system without
energy saving features.
1.5. THESIS CONTRIBUTION 21
Figure 1.8: The break down of thesis topics
1.5 Thesis Contribution

To put our studies into the context of current research, and based on the raised research
questions an overview of the thesis contribution is listed as follows:
• Energy Saving at the Base Stations

In mobile networks, base stations (BSs) have the largest share in energy consump-
tion. To reduce BS energy consumption, BS components with similar (de)activation
times can be grouped and put into sleep during their times of inactivity. The deeper
and the more energy saving a sleep mode (SM) is, the longer (de)activation time it
takes to wake up, which incurs a proportional service interruption. Therefore, it is
challenging to timely decide on the best SM, bearing in mind the daily traffic fluc-
tuation and imposed service level constraints on delay/dropping. In this section, we
leverage reinforcement learning techniques, i.e., SARSA and DQN, and propose so-
lutions to decide which SM to choose given time and BS load. We use real mobile
traffic obtained from a BS in Stockholm to evaluate the performance of the proposed
algorithm.
– Paper 1: M. Masoudi, et al, "Green Mobile Networks for 5G and Beyond,"

IEEE Access, 2019.
– Paper 2: M. Masoudi, et al, "Reinforcement Learning for Traffic-Adaptive
Sleep Mode Management in 5G Networks," IEEE PIMRC, 2020.
– Paper 3: M. Masoudi, et al, "Digital Twin Assisted Risk-Aware Sleep Mode
Management Using Deep Q-Networks," under revision in IEEE Transactions
on Vehicular Technology.
– Patent 1: M. Masoudi, et al, "NETWORK NODE AND METHOD PER-

FORMED THEREIN," Accepted.
• Cost-Effective Migration towards C-RAN with Optimal Fronthaul Design

Radio Access Networks (RANs) are facing a rapid increase in traffic demand, due
to the increasing number of devices connected to the network, and to the higher
quality of service requested by the users. Network operators must be able to provide
the required capacity to fulfill such demand. C-RAN has been recently proposed
as a promising architecture to increase network capacity, while improving energy
efficiency, and providing scalability. The main idea of C-RAN is to centralize the
digital units in a shared location responsible for coordinated signal processing and
management, which is called DU pool, while keeping the radio units at the cell
sites. Such centralization allows to share maintenance costs and energy consumption
among several digital units. However, C-RAN requires an extensive modification to
the current infrastructure, which results in a significant deployment cost. Prior to
this study, the formulation of a Total Costs of Ownership (TCO) model to identify
the migration cost of C-RAN as well as the formulation of a DU pool placement
optimization problem to minimize the TCO has been done in [33]. In this section,
to support increasing demand on capacity, the impact of upgrading C-RAN on the
TCO and energy consumption is investigated.
– Paper 4: M. Masoudi, S.S. Lisi, and C. Cavdar, "Cost-effective migration to-

wards virtualized C-RAN with scalable fronthaul design," IEEE Systems jour-
nal 2020.
• Content Placement and Functional Split Optimization in Virtual C-RAN:

In cellular networks, caching popular contents in the edge can reduce the backhaul
cost, access latency, and energy consumption. Since x-haul capacity becomes a
bottleneck, especially in ultra dense networks of 5G, we can equip the edge clouds
with caches to reduce the usage of fronthaul links. On the other hand, due to the
capacity limitation of the fronthaul, the functional split technique has been recently
proposed to split the baseband processing chain at multiple points. Therefore, the
cloud and the processing chain is divided in two hierarchical layers: central cloud
and edge cloud. In this study, we investigate the joint impact of content placement
and functional split on the energy consumption of network. In particular, we aim
at minimizing the total energy consumption of network, by jointly optimizing the
functional split point and content placement, while satisfying users QoS. The users’
QoS is defined as a minimum delay which users can tolerate until receiving their
requested files.
– Paper 5: A. Siram, M. Masoudi, A. Alabbasi, and C. Cavdar, "Joint optimiza-

tion of cache placement and functional split in H-CRAN architecture," IEEE
PIMRC 2019.
1.6. THESIS ORGANIZATION 23
• Joint Communication and Computation Resource Allocation for End-to-End

Network Slicing in C-RAN:
The network slicing concept will allow the deployment of virtual network slices on
top of the same physical infrastructure. In this study, we will propose a model to
estimate the impact of slice deployment and operation on the energy consumption
and delay. This study will be facilitated by the use of collected network data and ar-
tificial intelligence techniques. In addition, ML based algorithm will be employed to
develop online energy-efficient optimization algorithms to allocate communication
and computation resources which are usually too complex to be tackled by current
optimization methods.
– Paper 6: M. Masoudi, et al, "Energy-Optimal End-to-End Network Slicing in

Cloud-Based Architecture," Submitted to IEEE Open Journal of the Commu-
nications Society.
1.6 Thesis Organization

The reminder of this thesis is organized into two chapters as follows:
• Chapter 2 investigates the problem of reducing BSs’ energy consumption and man-
aging the linked risk of using AI algorithms in the BSs while maintaining the users’
QoS.
• Chapter 3 investigates three aspects of the network management and architecture
design, i.e., migration costs to new architectures, energy saving and improving delay
performance by content placement, and end-to-end network slicing.
• Chapter 4 concludes the research studies in this thesis and shed lights on the limita-
tions and future extensions of the conducted research studies.
Chapter 2
AI Assisted Green Mobile Networks
Mobile network operators anticipate a surge in their network energy consumption due to
exponentially increasing network traffic and introduction of new services [4]. Therefore,
in order to make their network sustainable, energy efficiency becomes a key issue in their
future deployments. Studies show that base stations (BSs) consume a major share of the
overall network energy consumption [51]. Thus, reducing BSs’ energy consumption can
boost the energy efficiency of the network. In particular, base station sleeping is one of
the promising ways of reducing energy consumption in the mobile networks, however, it
may lead to incurring longer service delay to the users. Due to variation of the traffic
profile during a day, network can be low loaded at some times. During the day, there are
multitudes but short durations in which no user is connected to the BS while BS is active
and consumes energy. In such idle durations, BS can put some of its components into sleep
and hence save energy. In this chapter, we examine the potentials of energy saving as well
as the challenges of activating sleep modes in 5G networks.
2.1 BS Sleep Mode Management

In this study, we consider a non-stand alone deployment in which there are two BSs,
namely capacity BS (CapBS) and coverage BS (CovBS). The CovBS can be an LTE BS
and provides ubiquitous coverage and cell specific signaling. The CapBS is responsible
for providing high data transmission rate, on-demand services, and user based signaling.
According to [10], the base station (BS) hardware components can be grouped accord-
ing to their activation/deactivation delays into 4 groups that can be shut down together to
save power whenever they are not needed, forming 4 different levels of SMs.
1. SM 1: This SM is the fastest, and the most shallow, whose duration is comparable
to the symbol time T s. Whenever a BS does not have any traffic over the entire band
of sub-carriers during symbol time per antenna, the PA can be switched off to save
power. This mode is available and known in current technologies as micro-scale
discontinuous transmission (µDTX).
25
26 CHAPTER 2. AI ASSISTED GREEN MOBILE NETWORKS
Table 2.1: Power and minimum durations from [9, 10]
Mode Active
SM1 SM2 SM3
100% 0%
Pow. (W) 702.6 114.5 76.5 8.6 6.0
Dur. (sec) T s: fast modes (FMs) 1 msec 10 msec
2. SM 2: A slightly deeper sleep level can be reached by switching off one more set
of components with slower (de)activation transition time than that of SM 1, if this
longer delay can be afforded by the system. In this mode, the transition is on the
transmission time interval (TTI) scale, i.e., 1 msec (a duration of a sub-frame consti-
tuting 2 resource blocks (RBs) or 14 resource elements (REs) assuming frequency
division duplexing (FDD) Frame Structure 1 [105])
3. SM 3: This mode has a deeper sleeping level but with a minimum duration of a 10
msec frame (10 sub-frames).
4. SM 4: This is the deepest SM with a minimum duration of 100 frames (1 sec).
In this study, we consider SM1-SM3. Table 2.1 presents the power consumption and min-
imum duration of the each SMs for macro-base stations with max power 49 dBm over 3
sectors with a bandwidth of 20 MHz [10].
2.2 Power Model

Base station is composed of multiple components such as power amplifiers (PA), analog
front-end (FE), digital baseband (BB), digital control (DC), power supply (PS). Therefore,
the BS’s power consumption can be defined as the summation of its component’s power
consumption.
Pc = PPA + PFE + PBB + PDC + PPS . (2.1)
The FE includes amplifiers and filters (analog baseband and RF), up/down-conversion mix-
ers, frequency synthesizer and digital to analog/analog to digital converter (DAC/ADC).
BB includes baseband filtering, up/down-sampling, FFT/IFFT, digital pre-distortion, digi-
tal compensation of system non-idealities, modulation and demodulation, channel encod-
ing and decoding, channel estimation, synchronization and processing for Multiple Input
Multiple Output and equalization. DC contains platform control processor, backbone se-
rial link interface and MAC and network layer processors. PS includes AC/DC converter,
DC/DC converters and active cooling.
The adopted BS power consumption model in this thesis follows the one implemented
by imec in [9] and reported in [10, 63]. The BS can operate in three modes, namely Fast
mode (FM), SM2, and SM3. FM mode merges three cases: 1) active with 0 to 100% load,
2) idle with 0% load, and 3) SM1. In this mode, BS makes decision and hops instanta-
neously into any operational mode depending on the arriving traffic without any need for
optimization. The remaining two sleep modes incur a transition delay from the time of
2.3. TRAFFIC MODEL 27
decision until completion of the action which causes additional queuing delay of traffic
arriving in the transition period.

Pidle + Pt
 If BS is in active mode
Pc = (2.2)

PSMi If BS is in SMi, i ∈ {1, 2, 3}

The BS consumes transient power for switching, between operational modes. Denote the
T
total power consumption of switching in duration of T by Psw , Power consumption per
switching by PSW , and the number of switching in duration of T by Fm , then we have
T
Psw = Fm Psw (2.3)
2.3 Traffic Model

For traffic model, we use a data set from a Swedish mobile network operator. The data
set is an anonymized averaged data rate for both uplink and downlink, collected by the
operator from multiple BSs in Stockholm for two months in 2018, with one sample taken
every 5 minutes. This coarse-grained data set guarantees the credibility of our daily traffic
pattern analysis and modeling. This data set can provide us with long term dependencies of
network traffic, however, this information is too coarse to provide information about very
short time instances comparable to sleep mode duration which is in order of mili seconds.
On the other hand, it is not possible to obtain data in such fine-grained granularity due to the
limitations in the deployed equipment at BSs measuring the KPIs. Therefore, we should
adopt an interrupted Poisson process (IPP) model to generate traffic data with short term
dependencies while maintaining the daily pattern and long term dependencies of traffic
data. The details of data generation scheme is provided in Paper E.
2.3.1 Dataset Description

Cellular traffic varies hourly depending on the daily activity of users in the area. Under-
standing the traffic patterns of cellular networks is extremely valuable for better network
resource management. In this study, we use a data set from a Swedish mobile network
operator. The data set is an anonymized averaged data rate for both uplink and downlink,
collected by the operator from multiple BSs in Stockholm for two months in 2018, with
one sample taken every 5 minutes. This coarse-grained data set guarantees the credibility
of our daily traffic pattern analysis and modeling. This data set can provide us with long
term dependencies of network traffic, however, this information is too coarse to provide
information about very short time instances comparable to sleep mode duration which is
in order of milliseconds. On the other hand, it is not possible to obtain data in such fine-
grained granularity due to the limitations in the deployed equipment at BSs measuring the
KPIs. Therefore, we should adopt a model to generate traffic data with short term depen-
dencies while maintaining the daily pattern and long term dependencies of traffic data.
2.3.2 Distribution of Traffic

The data collected by the operator needs to be preprocessed due to the existence of the
incomplete information in the provided data rates at some time instances. Moreover, the
information on the average provided rate cannot be directly used in our study, since we need
the demand per user. The preprocessing includes the following steps. First, we eliminate
the empty and incomplete information from the data set. Then, we focus on the aggregated
downlink traffic of one of such BSs in one location in Stockholm, only considering the
FDD downlink band for simplicity, and without loss of generality. Very similar studies can
be performed to simultaneously take into account all different BS bands, considering both
uplink and downlink. This coarse grained original data set is further used as follows to
obtain the required information to generate short term traffic data.
For data preparation, we first collate the average data rates of same time index over all
days into one set. Denote the set of data at time index i by Oi :
Oi = {o1 , o2 , . . . , oκ }, for i ∈ T (2.4)
where κ is 60 in our study since we have data for 60 days and T denotes the set of the time
index of data during the day. The cardinality of set T is 24 ∗ 60/ts where ts is sampling
time which is 5 minutes in our study. Then, for each set, i.e., each 5 minutes, we calculate
the mean, i.e., E(Oi ), and variance, i.e., V ar(Oi ), of the original data. Finally, based on
derived statistics, we can generate demanded data rate for each time step and each user,
using the steps explained in the next subsection.
It is worth mentioning that this coarse grained time granularity data provides no infor-
mation about the time of arrivals or the time duration between two arrivals which is a key
parameter to determine the idle time of the BS. Therefore, an accurate and yet tractable
model is required to generate the random arrivals of the users.
2.3.3 Bursty Behavior of Traffic Data

In order to model the user arrivals, various distributions such as the Poisson distribution
[106], log normal distribution [60], hyper exponential distribution [59], and interrupted
Poisson process (IPP) [107] have been used in the literature. Among these models, IPP
and hyper exponential distribution can model the bursty behavior of incoming traffic while
they are mathematically tractable [107]. It can be shown that the IPP is stochastically
equivalent to a hyper exponential H2 renewal process [108].
In IPP, as illustrated in Fig. 2.1, arrivals have two states, i.e., ON and OFF states. In
ON state, arrivals follow the Poisson model with parameter λ and in OFF state there is no
arrival. Assume that the transition rate from ON to OFF (OFF to ON) is denoted by ζ (τ ),
the steady state probabilities are derived as,
τ ζ
PON = , POFF = (2.5)
τ +ζ τ +ζ
2.4. RISK AWARE SLEEP MODE MANAGEMENT 29
τ
On Off
ζ
Figure 2.1: Data traffic model as interrupted Poisson process [6].
2.3.4 Data Generation

In previous two subsections, we derived the statistics of the number of arrivals and the data
sets. In this section we explain how to map the parameters of the IPP model so that it
matches with the statistics of the original data set. Denote the data rate request per arrival
by ψj and the aggregated data rate request per given time T by Ψ.
U
X
Ψ = ψj (2.6)
j=1
E(Ψ) = E(U )E(ψ) (2.7)
V ar(Ψ) = E(U )V ar(ψ) + V ar(U )E(ψ)2 . (2.8)
In Equation (2.6), ψ-s are independent random variables and for a sufficiently large value
of U , e.g., more than 30, Ψ is normally distributed, regardless of distribution of ψ-s, with
mean and variance of E(Ψ) and V ar(Ψ), respectively.
In order to generate the user arrivals and their data rates with the original data set
statistics, the model parameters, i.e., τ, ζ, λ, E(ψ), and V ar(ψ), should be set in a way
that Equations (2.7) and (2.8) hold.
2.4 Risk Aware Sleep Mode Management

In this section, we propose a framework for risk-aware sleep mode management. AI or ma-
chine learning algorithms are prone to anomalous and unknown data and their performance
can be inadequate or sub-optimal. A mechanism is needed to define the risk, monitor the
performance of AI with respect to the input traffic to find out whether re-training is needed.
As shown in Fig.2.2, we use a DT that gets the network data, communicates and mimics
the performance of the real system including the AI module, and assesses the risk. Based
on this assessment and the risk in the real network, we decide whether to use the AI, re-
train the network, or temporarily deactivate the AI module. In the following we explain
this approach in detail. First, we introduce the concept of risk associated to the BS sleeping
algorithms. We explain the DT model for BS ASM management to estimate the risk. Then,
we combine the intelligent sleep mode management and DT and explain the framework for
risk-aware BS sleep mode management as is depicted in Fig.2.3.
Real Use AI
System
Policy Risk
Network Estimated Assessment
Data Parameters Retrain
risk
DT Deactivate
Figure 2.2: Digital twin assisted decision making [6].
2.4.1 Risk Definition

In this study, risk refers to a situation in which SM management algorithm takes sleep
decisions resulting in delays in connection setup time of future user arrivals. The more
users experiencing delay, the higher value of risk should be associated to the situation.
Therefore, if the risk could be measured in advance, BS can avoid delaying large number
of users. For this purpose, we utilize the DT concept as a virtual representation of the BS
sleeping process.
In Fig.2.3, we illustrate the proposed a DT assisted risk aware SM management frame-
work. In this scheme, the physical network provides the required information, e.g., du-
ration of sleeping, ON/OFF state duration, sleeping time, and arrival rate. The observed
information is fed to the virtual network to update the model parameters and predict the
RDM in the network using Equation (2.11). Hence, the DT can predict the risk utilizing
the Markov model and its updated parameters. The Markov model together with the up-
date parameter step construct the virtual network. The output of virtual network is used to
predict the risk. This risk is compared with the threshold set by the operator and the actual
risk (which will be available in future). If the predicted risk is higher than the threshold,
the SMs should be deactivated. If the predicted risk is higher than the actual one, it is
possible that the algorithm needs to be retrained. Otherwise, if RDM is below the pre-
defined threshold, BS can activate the SMs and benefit from BS sleep mode management
algorithms. This algorithm is designed and explained in Paper 3. Using the risk monitor-
ing procedure in Fig.2.3, the operators can make sure that their sleep mode management
algorithm does not take sleep decisions when there are user arrivals or is not missing the
opportunity of energy saving when it is safe to save energy.
2.4.2 Risk Management

In this section, we propose a framework for risk-aware sleep mode management. First, we
introduce the concept of risk associated to the BS sleeping algorithms. We quantify the
risk and formulate it based on the proposed hidden Markov model for BS ASMs. Then,
Deactivate
SMs
Activate SMs
if not activated
else Threshold < RDMdt
RDMa
Physical Calculate current Check for re-
SMA RDM
Network RDM training data
RDMdt RDMa > RDMdt
Traffic
Data
Threshold
Retrain
Parameter Markov Predict
Operator
update Model RDM
Virtual Network
Digital Twin
Figure 2.3: Digital twin based risk aware sleep mode management [6].
we propose a solution for BS ASM management to estimate the risk. Finally, we propose a
framework for risk-aware BS sleep mode management to combine intelligent sleep mode
management and hidden Markov model based risk estimation.
In this thesis, risk refers to a situation in which SM management algorithm takes sleep
decisions resulting in delays in connection setup time of future user arrivals. The more
users experiencing delay, the higher value of risk should be associated to the situation.
Therefore, if the risk could be measured in advance, BS can avoid delaying large number of
users. For this purpose, we utilize the the hidden Markov model, as a virtual representation
of a process, in our case BS sleeping. The model is continuously updated from real-time
data, and uses machine learning and reasoning to help decision-making. Therefore, it
helps us understand the present and predict the future performance of the BS sleeping
performance metric, e.g., sleeping duration, probability of delaying users, and the risk for
each state of the network. In the following we explain the Markov model for BS sleep
mode management algorithm.
2.4.3 Digital Twin Model
In this study, as is depicted in Fig.2.3, the digital twin has three main parts, 1) parameter up-
date, 2) network model, and 3) prediction. The former two constructs the virtual network,
representative of the physical network. The model is continuously updated from real-time
data, and uses machine learning and reasoning to help decision-making. The network is
modeled as hidden Markov process and in the prediction phase we use the updated virtual
network to estimate and predict the future performance metric of the BS sleeping, e.g.,
sleeping duration, probability of delaying users, and the risk for each state of the network.
In the following we explain the Markov model for BS sleep mode management algorithm.
λ λ λ λ λ
A1,1 S3,1 S1,1 A1,1 A2,1 ... AM,1
µ µ µ µ µ
ζ τ ζ τ a1,3 ζ τ ζ τ ζ τ ζ τ
µ ...
A1,2 S3,2 S1,2 A1,2 A2,2 AM,2
a3,1
µ µ µ µ
a2,3 a1,2
a3,2 a2,1
A1,2 S2,2
µ
τ ζ τ ζ
λ
A1,1 S2,1
µ
Figure 2.4: Markov model for advanced sleep mode management [6].
2.4.4 Virtual Network Model: A Hidden Markov Process

In the DT, we require a realistic analytical model that can estimate the actual behavior of
ML based BS sleep mode management as a virtual network. In particular, we model the
system as a hidden Markov model1 depicted in Fig. 2.4 where the states, input traffic, and
model parameters are obtained by the interaction with the physical network environment.
The environment is defined as the load in the system and level of sleeping. In the following,
we present the hidden Markov model, required information, performance metrics, and a
framework for risk monitoring in the system.
The Markov model should encompass all operating states of the BS to be able to predict
the performance of the BS sleep modes. The states contain information on 1) sleep mode
level that the BS is in, 2) number of users in active mode, and 3) whether there are arrivals
or not (ON state or OFF state as depicted in Fig. 2.1). We model the state transition
diagram of the BS as a Markov model illustrated in Fig. 2.4. For instance, state S(i, j),
i ∈ {1, 2, 3} and j ∈ {1, 2}, denotes the state that the BS is in SMi and the traffic arrivals
are in the state j where j = 1 means ON state and j = 2 means OFF state. If the BS
is in state Am,j , m ∈ {1, 2, . . . , M } and j ∈ {1, 2}, then it is active and m number of
users are in the BS. We assume that maximum M number of users can be served and the
BS can change its state according to the state transition diagram. When all users in the BS
are served, BS goes to one of the SMi-s and stays there until a new user arrives. Before
arrivals of the users, the state changes from S(i, 2), OFF state to S(i, 1), ON state. When
a new user arrives, BS goes from state S(i, 1) to state A(1, 1) switching from SMi to FM;
however, it must wait for the wake-up time of SMi, the duration to activate the sleeping
components. When the BS is in any of the sleep modes there is a transition possibility from
1 An IPP model plotted in Fig. 1 is a Poisson-emission hidden Markov model with two hidden states [107]
in which one of the states has zero emission rate. Therefore, the state diagram depicted in Fig. 2.4 is a hidden
Markov model.
any of the sleep mode state, i.e., S(i, j) i ∈ {1, 2, 3} to the other SMs if j =2, there is no
traffic arrival.
2.4.5 Performance Metrics

The proposed Markov model can enable us to derive the steady state probabilities and
later the calculation of the risk value. In this section, according to Fig. 2.4, given the
model parameters, and the power model, we derive the sleeping probability, and we define
a new parameter quantifying the risk of taking wrong action. Let’s define the following
probabilities:
• νi,j be the probability that BS is in state S(i, j).
• um,j be the probability that BS is in state A(m, j).
where S(i, j) is the state in which BS is in sleep mode i and the arrival state is j and
A(m, j) is the state in which BS is active and there are m number of users being served at
the BS and the arrival state is j, and j = 1 means ON state and j = 2 means OFF state.
2.4.5.1 Probability of Sleeping

The probability that the BS is in sleep mode is summation of the probabilities that BS is in
SMi which is given by,
2 X
X 3
Vs = νi,j (2.9)
i=1 j=1
In [6], we have written the balance equations for Markov model in Fig.2.4. The balance
equations are used to calculate the steady state probabilities, i.e., νi,j and um,j .
2.4.5.2 Number of Switching

Each time BS switches from one sleeping state to another, it adds extra energy consumption
or cost to the BS. Therefore, it is crucial to keep track of the number of switchings which
is denoted by Fm and is defined as the number of times that BS switches between SMs or
activate/deactivates its components. Each time we pass from active to sleep mode in Fig.
2.4, i.e., go to SMi, later BS has to take a go to FM action. The number of switchings has
two terms. The first term counts the number of switching to SMs and wake-ups. When BS
is sleeping and arrivals are in the OFF state, it can switch between SMs. The second term
counts the number of switchings between SMs.
2.4.5.3 Risk of Decision Making

When BS is in SM and new users arrive, the new coming users must wait until the BS
wakes up and goes to the active mode again. Denoted by Us , the average number of users
in the BS when BS is any of SMs is calculated as,
Us = λν1,1 + λν2,1 + λν3,1 (2.10)
where λ is user arrival rate and νi,1 , i ∈ {1, 2, 3} are the probabilities of being in SMi
while arrivals are in ON state. From the operators’ perspective, the smallest possible value
for Us is more preferable since minimum number of users experience delay until they are
being served. Although this metric can reflect the favor of the operators, it is incapable of
measuring the performance of the BS sleeping algorithms. For instance, when the network
is busy, i.e. λ is high, the probability of sleeping will be low. However, one inappropriate
sleeping decision may result in a large number of users experiencing delay. On the other
hand, in off peak hours, when λ is low and Vs is high, we may have very few users experi-
encing delay. Therefore, the parameter Us cannot solely reflect the higher risk of incurring
high delay at peak hours. To tackle this issue, a novel metric called risk of decision making
(RDM) is introduced to measure the risk when we take a non-optimal action by taking into
account the Us and the probability of sleeping (in both ON and OFF state). Risk of wrong
decision is formulated as,
RDM = E(Number of users per time unit|BS be in SMs)

Us λ(ν1,1 + ν2,1 + ν3,1 )
= = . (2.11)
Vs Vs
When the network is in peak hours, Us might be high and Vs is low, hence there is not
much room to save energy. Since the network is crowded, any wrong sleeping decision
may result in a large number of users experiencing delay, and hence RDM is higher as
indicated in Equation (2.11). In the off peak hours, Us might be low and Vs is high. Since
less number of users are in the network, RDM is lower which is indicated in Equation
(2.11). RDM value can also be used to determine the right time/moments to deactivate the
BS sleeping algorithm, in order to avoid delaying large number of users. There are two
scenarios in which RDM will be high.
High RDM in Busy Hours: At peak hours, when the network is crowded, it is not
beneficial to put BS into sleep modes due to the risk of incurring delay. Therefore, the
BS sleeping algorithm can be disabled temporarily until the sleeping becomes beneficial
again. It is true that the learning algorithm may learn these moments and avoids sleeping
in such events, but it is possible that the algorithm choose the wrong action due to many
reasons such as abnormal behavior of traffic, lack of proper training, and change of traffic
pattern.
High RDM due to Abnormal Behavior of Traffic: When the behavior of users or the
traffic pattern changes, e.g., arrival rate changes, and the algorithm is not previously trained
for it, the sleeping decision may yield to unnecessary wake-ups or inappropriate sleepings.
It takes some time for the BS to realize the abnormal behavior of the input traffic which
results in unnecessary energy consumption and/or unacceptable incurred delay. When any
non-optimal action is taken, it takes some time for the algorithm to converge and update its
policy accordingly. Using RDM, we can prevent incurring such delays by comparing the
actual RDM in the network with a reference or expected value of risk. The reference value
2.5. BS SLEEP MODE MANAGEMENT PROBLEM 35
can be dynamically set by the operator depending on the hours of the day, and assured
QoS, etc. Using the same data sets as the one used for training the algorithms, the RDM
can be pre-calculated as a reference value. Then, the actual RDM in the network can be
always compared with this reference value. If the experienced value of RDM is above
the expected one, it means the current traffic pattern is different than the expected traffic.
Therefore, the BS should disable the sleeping features, i.e., SMs, until either the algorithm
is trained over new data set or the traffic behavior becomes normal again.
Interaction with DT and Physical Network:
After defining the DT, underlying Markov model, and the performance metrics, we now
explain how the parameters of the hidden Markov model within the DT can be obtained
and updated. The parameters of hidden Markov model is function of user arrival patterns,
rates, ON state duration, and OFF state duration. Therefore, these parameters can change
during the day. In order to obtain the parameters of the model, i.e., λ, τ , ζ, from the
training data set, we can use a well-known backward-forward algorithm, i.e., Baum-Welch
algorithm [109]. The Baum-Welch algorithm finds the maximum likelihood estimate of
the parameters of a hidden Markov model given a set of observed input sequences. This
algorithm computes the statistics of the input traffic sequence O = {o1 , o2 , . . . , oT }, and
then updates the maximum-likelihood estimate of the model parameters, i.e., ON/OFF
state duration and arrival rate. The procedure of Baum-Welch algorithm to extract the IPP
parameters is defined in [107].
2.5 BS Sleep Mode Management Problem

2.5.1 Action Set and Environment
The state of the system at time index i, is denoted as si . This state is composed of current
traffic and encoded information about previous traffic which is encoded and fed back by a
neural network module, explained in Paper E. The current traffic is the number of requested
PRBs between late time users are served until and time index i. Each action has a transition
period, and the system is said to be in the last state until the new state is in action. The
action set denoted by A is defined as A = {F M, SM 2, SM 3}. Here, F M and SM k,
with k ∈ {2, 3}, denote the go-to-fast-mode and go-to-SMk mode , respectively.
The system environment information at time index i is fully captured via the following
5-tuple, ei = (pb, lb, pc, lc, ltot) where pb and pc are the basic and capacity BS power
consumption, lb and lc represent the basic and capacity BS accommodated load in RBs,
while ltot is the total traffic load that needs to be served at the ith time index, respectively.
That is ltot = lb + lc + di where di represents the number of delayed RBs at the ith time
index, which is the difference between the arriving traffic and total served traffic. We also
define the total power consumption at the ith time index as pt = pb + pc.
2.5.2 Reward Function Design

When the CapBS is operating in fast mode (FM), no serving delay takes place. However,
the system may miss the opportunity of saving more energy if deeper SM is possible. We
define a normalized energy saving metric as

+
(pSM 1 − pi )
r̃i,p = , (2.12)
pSM 1 − pSM 3
where (x)+ is a operator that takes the value of x if it is positive and is zero otherwise,
pi , i ∈ A, is the power consumption of SMs and rp ∈ [0, 1].
When the system receives a request while it is in one of the deep SMs, i.e., di 6= 0, or
in the case when di = lc = 0, the normalized delaying penalty can be calculated as
di
ri,d = − , (2.13)
lmaxc
which takes 0 when di = 0, and 1 when di = lmaxc, respectively. When the system is in
FM with lc 6= 0, although no power saving is attained, there should exist a reward for the
avoided delay, i.e., (di = 0)&(lc 6= 0), the reward is
lc
rd = . (2.14)
lmaxc
The incurred reward function at the end of the ith time index due to the action taken at the
end of time index i − 1 can now be plausibly defined as
ri = (1 − α)rp + αrd, (2.15)
where α is a weight parameter to prioritize power saving or serving delay. When an action
is taken, the system will be frozen for the minimum duration of that action. Therefore, the
incurred reward is calculated at the end of such duration.
2.6 Selected Results

An event-based numerical simulation is implemented in Python, running the proposed
adaptive algorithm and its baseline counterparts. We compare the performance of the SMA
algorithm with optimal BS sleeping (OBS) as an upper bound on energy saving gain and
fixed sleep mode where only SM 1 can be used. In these baselines all users are served in-
stantaneously with no incurred delay. We also compare the results with SARSA algorithm
explained in [63]. In OBS, the BS is fully aware of all future incoming users and their
activity times. Scanning all inactivity periods, BS can fill the inactive time with the proper
SM. We follow the procedure explained in [63].
The provided data sets needs to be processed to be used in our simulation tools. In Fig.
2.5a, we depict the normalized average load for each hour of a day for original data set and
the generated load. The comparison shows that the generated data preserves the average
daily behavior of the original data. In Fig. 2.5b we show the arrivals in time scale of 5
seconds.
The parameter α prioritizes the power saving or incurred delay in the reward function.
In Fig. 2.6 we present the performance of the proposed algorithm, i.e., SMA, with regards
2.6. SELECTED RESULTS 37
1.2
Generated load 1
Normalized averaged traffic load
1 Load from data set
Normalized user request

0.8
0.8
0.6
0.6
0.4
0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 0 1 2 3 4 5
Hour of day Time in Second
(a) Average original and generated load. (b) User arrivals in 5 seconds.
Figure 2.5: Network load generation as input to the network for τ = 0.1 and ζ = 0.5 [6].
to the α. The higher α is, the less weight is assigned to energy saving. We compare the
performance of SMA with OBS algorithm defined in [63], as an upper bound on energy
saving gain. OBS is always optimal and non-causal due to the knowledge of future infor-
mation. For α = 0 the total reward includes only energy saving reward. SMA can achieve
optimal energy saving but at cost of incurring delay to the users. When latency is more pri-
oritized, the BS becomes more conservative to choose deeper SMs and prefers shallower
SMs. Therefore, less energy is saved in favor of less incurred delay. SMA outperforms
fixed SM where only SM1 is in use, in which BS cannot benefit from deeper SMs and
miss the energy saving opportunities. Moreover, SMA performs better than SARSA algo-
rithm because SMA can find better long-short term dependencies in traffic data and hence
leverage this information to find a better energy saving policy.
The Fig. 2.8 illustrates the absolute energy consumption of the SMA compared to the
three cases: 1) when no sleeping is used, 2) when only SM1 is used, and 3) the optimal
energy saving approach. During the rush hours the SMA mostly chose to go to the shal-
lowest SM hence the performance is very close to the only SM1. Moreover, during these
hours there is not room to save energy and hence it is meaningful to be more conservative
about delaying users and let the BS be either awake or benefit from the SM1.
Fig. 2.7 shows the normalized energy saving performance of the SMA and OBS over
24 hours. The values are normalized with regards to the maximum energy saving of OBS
algorithm. Hours between 2:00-6:00 AM are the best hours of a day with the highest
energy saving potentials. The least opportunity for energy saving happens in the evening
at about 18:00. It is worth mentioning the data set under consideration is for an area which
is dominated by the industrial buildings. The traffic pattern and hence the energy saving
patterns depend on the type of area.
In Fig. 2.9, we plot the performance of risk aware BS sleeping algorithm. Using the
procedure in Fig. 2.3, BS calculates the risk value and when the risk is high it temporarily
1
0.9
Normalized Energy Saving
0.8
0.7 SMA
Sarsa
0.6 OBS
0.5 Only SM1
0.4
0.3
0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Alpha
Figure 2.6: Energy saving vs delay trade-off [6].
1
OBS
SMA
0.8
Normalized Energy Saving
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24
Hours
Figure 2.7: Normalized daily profile of energy saving [6].
deactivates SMs. When SMs are deactivated, BS calculates the potential risks and when
the risk is small again for a while, i.e., for duration of T , it activates the SMs again.
2.7. SUMMARY 39
600
No SM
550 Only SM1
Optimal
500
SMA
450
Energy Consumption (K joule)
400
350
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour
Figure 2.8: Daily profile of energy consumption.
7
RDM Threshold
BS RDM Value
6
Risk of Decision Making (RDM)
SMA Status
5
SMA Status
3
Active
1
De-active
0
10 20 30 40 50 60 70 80 90 100 110
Time (s)
Figure 2.9: Performance of risk management algorithm [6].
2.7 Summary
In this study, we propose a framework to generate the user arrivals based on the real data
set. We develop a deep learning based algorithm to decide at what time which level of
sleep mode should be chosen. The objective of the proposed algorithm is to maximize
the reward function, i.e., linear combination of normalized energy saving and delay re-
ward. Simulation results show that considerable energy saving can be achieved using the
proposed algorithm. The results show that for busy hours of a day or when in the reward
function delay is more prioritized, the algorithm tends to choose more shallow SMs, i.e.,
SM1 which reduces the energy saving. On the other hand, for low traffic hours or when
energy saving is more prioritized, deeper SMs, i.e., SM3 is chosen more frequently, result-
ing in more energy saving and more incurred delay to the users. Therefore, comparing to
the optimal BS sleeping, at low traffic hours, the proposed solution achieves near optimal
result, while for busy hours, the proposed solution achieves less, yet considerable, energy
saving. Comparing to the tabular methods, e.g., SARSA, the proposed solution has a better
energy saving performance for a similar reward function (similar priority for energy sav-
ing and incurred delay). This means that compared to the tabular method, the proposed
solution learns better the dynamicity of the network, e.g., here user arrivals.
Chapter 3
AI Assisted Network Architecture

Design and Management
Trends like softwarization and centralization are being widely attracted by both the re-
search community and standardisation bodies. In particular, centralization in the radio
access network (RAN) has been discussed in the context of Cloud RAN (C-RAN), where
the base band processing unit (BBU), also known as digital unit (DU), is decoupled from
the radio unit (RU) and pooled in a central cloud. In this chapter, we investigate the C-
RAN based network architectures that can enable 5G to provide multitudes of services to
the users. In this chapter, we first provide the end-to-end delay and power consumption
model of the network. Then, we focus on two aspects of the network. First, we investigate
the cost of migrating to the new architectures. In the next step, with the aim at minimizing
the power consumption of the network where we investigate two methods of improving
the provided QoS to the users. In the first method, we try to reduce the delay of delivering
contents to the users, and in the second method we tailor the end-to-end network resources
for each services with heterogeneous QoS requirements. In the following, we start with
the explanation of the investigated architecture and its delay and power consumption.
3.1 Cloud RAN Network Architecture Power and Delay Models

Various advantages are enabled through centralization of RAN functions including en-
hanced cooperative solutions, improved load balancing and RAN sharing. On the other
hand, it can raise critical challenges such as migration cost to the new architectures, de-
livering low latency services, and requiring high bandwidth in the transport layer [110].
To address these issues, different network architectures and techniques has been proposed.
For instance, to reduce the bandwidth requirements in the fronthaul some lower layer net-
work functions may remain close to the RU, reducing the level of centralization which
means that the processing functions should be split [110] and be distributed between the
two layers of processing. Therefore, a new layer of processing is required in the network,
responsible for processing the partial network functions. This layer is connected to the cen-
41
CHAPTER 3. AI ASSISTED NETWORK ARCHITECTURE DESIGN AND
42 MANAGEMENT
ter cloud via high speed low latency transport link. Therefore, the transport link is a vital
components of the network architecture. We call the transport link as X-haul which can be
Fronthaul or Midhaul. Moreover, This link is using CPRI or eCPRI to transport data to the
cloud and can be implemented with different technologies such as passive optical network
(PON), Ethernet network, or time sensitive network. Each of these technologies have their
own pros and cons which make them a better solution in a specific situation.
3.1.1 Conventional C-RAN Architecture

In conventional C-RAN architecture, as depicted in Fig. 3.1, the digital units (DUs) are
residing in a central cloud (CC) and are connected to the RUs via fronthaul links. Fronthaul
is a transport network that connects the RUs to the central cloud. The central cloud is
responsible for the base band processing. The main issue with this architecture is that
the fronthaul links may not be enough to transport the I/Q signals from RUs to the DUs.
Therefore, we need to lessen the load from the fronthaul. This can be done by adding
another processing layer, called edge cloud, between the RUs and central clouds.
Figure 3.1: Basic C-RAN architecture.
3.1.2 Hybrid C-RAN

In Fig. 3.3, we illustrate a hybrid architecture with a central cloud and multiple edge
clouds, where DUs are deployed at both central cloud and edge cloud (EC). Thus, the base-
band processing can be flexibly distributed over central clouds and edge clouds. Moreover,
both CC and EC can be equipped with memory which holds popular files for later caching.
H-CRAN is a three-layer architecture, which consists of cell layer, edge cloud, and central
cloud. Cell layer consists of cells that are being densified, each serving several UEs.
The edge cloud can be simply a DU co-located with a RU. The DU is responsible for
a part of base band processing while the rest of processing is done at the CC. The link
between EC and RU, the fronthaul link, can be coaxial cable, fiber, or Ethernet link. The
link between EC and CC, the midhaul link, can be fiber links using different technologies
3.1. CLOUD RAN NETWORK ARCHITECTURE POWER AND DELAY MODELS43
such as time-wavelength division multiplexing passive optical network (TWDM-PON) or

Ethernet base networks. This architecture is depicted in Fig. 3.2.
Common protocol radio interface (CPRI) connects EC and CC and delivers traffic via
a very high-speed connection links known as midhaul/fronthaul. On one hand, these links
have limited capacity, on the other hand, in 5G networks, the required midhaul/fronthaul
capacity is getting higher. Therefore, the links between the RUs and DUs will see an in-
crease in traffic. To ensure that the demands are not reaching the limits of the bandwidth
when 5G is implemented, evolved CPRI (eCPRI) is introduced. The eCPRI specification
supports 5G and enables increased efficiency in order to meet the needs foreseen for 5G
mobile networks. In contrast to CPRI, the eCPRI specification supports more flexibility
in the positioning of the functional split inside the PHY of the BSs. It provides flexible
radio data transmission through a packet-based midhaul/fronthaul network, for example
IP or Ethernet. Packet-based Ethernet midhaul/fronthaul can decrease the cost of network
because it enables traffic aggregation and switching which is cheaper than, e.g., deploying
wavelength-division multiplexing (WDM)-based networks. Packet-based transport archi-
tecture is of high interest for industry [111].
The eCPRI protocol can balance important components of latency, throughput, and
reliability requirements for advanced 5G applications. Besides, eCPRI is an open interface,
allowing carriers to work together in a more complimentary way, which can bring about
better connected and faster networks across the world [112]. There are multiple advantages
of eCPRI, compared to CPRI, summarized as follows:
• Ten-fold reduction of required bandwidth
• Required bandwidth can scale according to the to the user plane traffic
• Ethernet can carry eCPRI traffic and other traffic simultaneously, in the same switched
network
• A single Ethernet network can simultaneously carry eCPRI traffic from several sys-
tem vendors.
Figure 3.2: Two layer CRAN architecture.

44 MANAGEMENT
In H-CRAN, the EC can co-locate with RUs. Another alternative is to connect multiple
RUs to an EC. In this architecture, a group of cells/RUs are connected to a EC as an
aggregation point. The fronthaul between a cell and an EC can be wireless links, e.g.
mmWave links. The ECs are connected to CC via midhaul using TWDM-PON technology.
Edge cloud layer and central cloud layer are deployed with DUs, which are containers for
virtualized functions. EC is usually less energy efficient than CC, because the number of
DUs at the CC is larger than that in EC. This architecture is illustrated in Fig. 3.3.
Figure 3.3: Multi layer C-RAN architecture.
3.1.3 Network Delay Model

The total incurred delay for a user is composed of transmission delay in midhaul/fronthaul,
transmission delay in RAN, processing delay at CC and radio site, propagation delay in
midhaul/fronthaul and is modeled for user u that belongs to RU r as:
tot M,tx R,tx pr
Du,r =Du,r + Du,r + Du,r + Dp
N
X sw
+ (Dkq + Df + Dksw ) (3.1)

k=1
M,tx R,tx
where Du,r is the transmission delay in midhaul/fronthaul, Du,r is transmission delay
pr p
in RAN, Du,r is processing delay, D is propagation delay in midhaul/fronthaul which is
dependent on the distance and is a fixed value, Nsw is the number of switches between
CC and radio site, Dkq is the queuing delay which is function of load on switch k, Df is
fabric delay which is constant and is defined as delay caused by Ethernet switch circuits
with typical value of 5 − 10 µs [113], and Dksw is switching delay due to load from user u
belongs to on switch k.
Here, we focus on eCPRI protocol and Ethernet based architecture. More information
other architecture can be found in [5,7]. In eCPRI protocol, based on spliting point a group
of packets are transmitted at each burst period. To calculate transmission delay for each
user we need to calculate how many periods we need to transfer the requested content of
the user as follows:
Table 3.1: List of parameters and their definitions.
Parameter Definition
U, R Sets of users, and RUs, respectively
II
Rr D The midhaul data rate for RU r
tot
Du,r Total delay of user u in Ur
M,tx
Du,r Transmission delay of user u in Ur in midhaul/fronthaul
R,tx
Du,r Transmission delay of user u in Ur in RAN
pr
Du,r Total processing delay of user u in Ur
sw
Dr,k Switching delay in RU r on switch k
dpr
x,u,r Processing delay of user u in Ur at x ∈ {CC, BS}
x
Cu,r Total required Gops for user u in Ur at x ∈ {CC, BS}
pr
Cx,r Total allocated GOPS of RU r at x ∈ {CC, BS}
ECC , EM F , ErBS Energy consumption of the CC, midhaul/fronthaul, and BS/RU r
CC
Pcool Cooling power consumption at the CC
CC
PDU,min Idle power consumption of a DU at the CC
CC
Pproc Processing power consumption at the CC
sw
Pr,k Power consumption of switch k in RU r
BS
Pstatic Static power consumption of a BS
PrDU,min Minimum processing power consumption of a DU in BS r
PrDU,max Maximum processing power consumption of a DU in BS r
Prtx Downlink transmission power for a single user in Ur
pr
νr = CCC,r Allocated processing resources in RU r
& '
IID
M,tx
Lu,r Ru,r Lpacket
Du,r = sc n
, (3.2)
Lpacket δsc Nu,r mod RL
sc
where Lu,r is the file size of user u connected to RU r, Lpacket is the packet size, δsc Nu,r nmod
is the respective downlink data rate with nmod being the modulation index, e.g., nmod = 4
for 16-QAM and δsc is the subcarrier spacing (in Hz), which depends on the 5G numerol-
IID
ogy given in Table 3.2. Ru,r is the split point data rate, and RL is the line rate which is
dependent on interface medium, e.g., fiber links.
R,tx
The transmission delay in RAN, Du,r , is calculated as
R,tx slot
Du,r = Nu,r Tslot (3.3)
slot
where Tslot is 5G time slot duration given in Table 3.2 and Nu,r is the required number
46 MANAGEMENT
Table 3.2: 5G frame structure [11].
Subcarrier Spacing Frame Subframe Slot Symbol

1 71
15 ×2n KHz 10 ms 1 ms 2n ms 2n µs
of time slots for user u that is served by RU r, and is calculated as

slot Lu,r
Nu,r = sc × n
(3.4)
Nu,r sym × nmod
where nsym is number of symbols per time slot.

The total processing delay is modeled as:
pr
Du,r = dpr pr
CC,u,r + dBS,u,r
CC BS
Cu,r Cu,r
= pr + pr (3.5)
CCC,r /|Ur | CBS,r /|Ur |
where dprx,u,r is the processing delay for slice s, RU r, user u and x = CC means at the CC
x
and x = BS means at the BS. Cu,r is defined as the total required Giga operations (Gops)
pr
for user u, slice s, and RU r at x = {CC, BS}. Cx,r is the total allocated processing
resources, i.e., Giga operations per second (GOPS), to RU r at location x ∈ {CC, BS}.
We assume the processing resources are equally shared among the users in RU r. The
x
total available processing resources is limited at the CC and the BS. The amount of Cu,r is
determined according to the split point and it is given as
X
x
Cu,r = PFu,r,i , x ∈ {CC, BS} (3.6)
i∈Zx
where Zx includes the processing functions at x ∈ {CC, BS}. Utilizing the information
about the required amount of processing resources for each communication function, PFi
is calculated as follows:
Y xact si,x
PFi = Ci,ref , (3.7)
xref
x∈X
where xact and xref are the system input parameters/resources under actual scenario and
the reference scenario, respectively, and X is the set of all possible tuning parameters. The
exponent si,x highlights the impact of changing the input parameter on the required Gops
for the communication function sub-component.
The information to calculate the required Gops is summarized in [8, 114]. All the
necessary digital sub-components that contribute to the overall processing delay and their
corresponding Gops are summarized in [114].
The switching delay of switch k for users is formulated as,
Lk
Dksw = , k = 1, . . . , Nsw (3.8)
Rsw
where Ls,k is the load on switch k and Rsw is the switch rate.
3.1.4 Network Power Model

Total energy consumption of network is composed of energy consumption at the CC, mid-
haul/fronthaul, and at the BSs. Total power consumption of the network can be expressed
as,
X
Ptotal = PCC + PM F + PrBS (3.9)
r∈R
where PCC , PM F , and PrBS , are the power consumption of the CC, midhaul/fronthaul,
and each RU r, respectively.
3.1.4.1 Central Cloud Power Model

Power consumption at the CC can be calculated as
1 CC CC

PCC = EDU + Pcool (3.10)
ζCC
CC
where ζCC is power efficiency at CC. Pcool is the power consumption the cooling system
CC
at the CC. PDU is the power consumption of the DUs and is given as follows:
CC CC CC CC
PDU = NDU PDU,min + Pproc (3.11)
CC
where PDU,min is the fixed power consumption of a single DU when there is no processing
CC
and Pproc is the processing power consumption of the CC.
3.1.4.2 Midhaul/Fronthaul Power Model

The total power consumption of the midhaul/fronthaul transport network is given as
N
X sw
PM F = Pksw (3.12)
k=1
where Pksw is the switch power consumption corresponding to switch k. It is composed of

load-dependent and load-independent power consumption and is calculated as,
sw
Pstatic
Pksw = + ρsw Lk (3.13)
|S|
sw
where Pstatic is the load-independent power consumption, |S| is number of slices, and ρSw
is a load dependent power consumption of the switch which is multiplied by total amount
of service’s data to be processed by the switch. Lk is total volume of data per switch k.
48 MANAGEMENT
3.1.4.3 Edge Cloud Power Model
Energy consumption of edge cloud is very similar to the central cloud. However, one may
assume that less powerful devices may be deployed at the edge clouds compared to the
central cloud. Power consumption at the EC can be formulated as
1 EC EC

PEC = EDU + Pcool (3.14)
ζEC
CE
where ζEC is power efficiency at EC. Pcool is the power consumption the cooling system
EC
at the EC. PDU is the power consumption of the DUs and is calculated similar to Eq. 3.11.
3.1.4.4 Base Station Power Model

Energy consumption of the BS includes the energy consumption of RU r and its corre-
sponding DU which is calculated as,
1
ErBS = BS
+ PrDU

Pstatic
ζBS
1 X tx
+ Pu,r (3.15)
ζBS
u∈Ur
with
PrDU = PrDU,min + PrDU,max − PrDU,min × ηrBS

(3.16)
BS
where ζBS is the radio site efficiency, Pstatic is a load-independent power consumption
DU
including circuit power and cooling. Pr is the power consumption of the DU at BS r
and ηrBS shows what portion of total DU’s processing capacity at BS r. Pu,r
tx
is the power
R,tx
transmitted to user u. Du,r is the RAN transmission time from (3.3).
3.2 Migration Costs towards C-RAN

In this study, the migration cost of C-RAN from today’s D-RAN is evaluated. A DU pool
placement optimization problem is formulated to minimize the migration cost, which is
modeled by calculating the TCO, defined as the sum of build-out costs, i.e. CapEx, and
operation and maintenance costs plus electric bills, i.e. OpEx. The TCO of C-RAN is
compared over time to that of traditional D-RAN, which is assumed to only include OpEx,
as it is already deployed. This analysis is intended to understand the migration cost of
C-RAN, and the amount of time needed to compensate for such cost, under different cost
assumptions. In order to meet the increasing capacity demand, possible future upgrades of
C-RAN are investigated. An optimization framework is proposed, to minimize the TCO of
upgrading the network. This analysis is intended to understand and compare the upgrade
capabilities of C-RAN and D-RAN, in terms of TCO. However the solution to the ILP
problem is optimal, it is not scalable. Hence, we develop a scalable heuristic algorithm
3.2. MIGRATION COSTS TOWARDS C-RAN 49
based on genetic algorithm (GA) to solve the problem. Although the GA does not guar-
antee the optimality, it is scalable. We investigate the optimality gap of this algorithm as
well. Finally, the impact of the proposed frameworks on the performance of JT as one of
the CoMP techniques is investigated. The main contributions of this study are as follows:
• The formulation of a TCO model to identify the migration cost of C-RAN.
• The formulation of a DU pool placement optimization problem and a network up-

grade optimization problem to minimize the TCO.
• Developing the GA-based scalable heuristic algorithm to solve the ILP problem in
[33]. Investigation of optimality gap between scalable solution and optimal solution.
• The evaluation of joint transmission performance in C-RAN under the proposed

framework.
3.2.1 System Model

C-RAN fronthaul is assumed to employ a TWDM PON technology. In TWDM PON both
wavelength and time division are performed, to exploit the bandwidth efficiency of WDM
with the cost efficiency of TDM. Two aggregation levels are assumed in the fronthaul.
The first aggregation level is composed of passive splitters, while the second aggregation
level is composed of AWG filters. The radio units access the first aggregation level in
time division, and the second aggregation level in wavelength division. The two main
components of the fronthaul architecture are the ONU, located at the cell site, and the OLT,
located at the pool site. These components perform electrical to optical signal conversion
and vice versa. The OLT considered in this architecture provides single line rates of 10
Gbps, and it is composed of a wavelength division multiplexer and a LC for each digital
unit. CPRI standard sets a constraint on the round trip time between radio units and digital
units, in order to achieve seamless baseband signal transmission. This constraint translates
into a maximum ONU-OLT distance between 20 and 40 km [115]. AWG and splitters
splitting ratios are denoted as N and M , respectively. The splitting ratio of an AWG is
defined as the number of wavelengths supported by a single fiber in TWDM PON. While
the splitting ratio of a splitter is defined as the number of radio units supported by a single
wavelength. According to CPRI requirements, for a system where radio units are provided
with LTE coverage of 3 sectors with 20 MHz bandwidth and 2x2 MIMO configuration,
the minimum line rate must be set to 2.4576 Gbps [116]. Thus, a maximum number of
four CPRI flows can be transmitted in a 10 Gbps link, which leads to Mmax = 4. It is
assumed that a fiber can achieve a maximum rate of 80 Gbps, hence the maximum number
of wavelengths per fiber in the TWDM PON is Nmax = 8. The DU pool is considered as a
shared location where the digital units are stacked on each other. The digital units coincide
with the processing units residing at the cell site in a traditional distributed topology.
50 MANAGEMENT
3.2.2 Pool Placement Problem
In this section, we investigate the DU pool placement optimization problem. Given a set of
radio units, the goal is to find the optimal number and location of pools, digital units, and
splitters, such that all the radio units are connected to a pool and the TCO is minimized.
For simplicity, the AWG are assumed to be located in the same locations as the pools,
hence they are not included in the optimization problem. The problem is modeled as an
ILP problem, and it is formulated as follows:
Given: set of radio units and digital units; candidate locations of pools and splitters; capac-
ity of pools, digital units, AWG and splitters; maximum distance allowed between pools,
splitters and radio units; cost and energy consumption values.
Find: optimal locations of pools, digital units, and splitters; connections between pools,
splitters, and radio units.
Constraints: limitation related to the Common Public Radio Interface (CPRI)
Objective: minimize the TCO. To solve this problem optimally, we use ILP which gives
optimal solution but is not scalable, hence, we design a scalable heuristic algorithm at cost
of sacrificing optimality,
3.2.3 Solution Methodology

In this study, we model the pool placement problem as an ILP problem. This problem can
be solved optimally; however, the solution is not scalable. Therefore, for larger problem
sizes, the problem cannot be solved in a polynomial time and hence we need a scalable
solution. On the other hand, scalable solutions usually cannot necessarily guarantee the
optimality of the solution. Hence, there is a trade-off between scalability of the solution
and optimality of the result. In this study, together with optimal solution, in which we
use IBM ILOG CPLEX solver, we propose an scalable solution using Genetic algorithm
(GA). For GA, we need to design a chromosome. Each chromosome corresponds to a
solution to the problem at each run. Better chromosomes have higher chance to survive and
maintain themselves more often in the reproduction phases, namely mutation and crossover
which are two main operators in GA. In crossover phase, two chromosomes, known as
parents, are picked randomly. Then, we generate two offsprings by combining the parents
chromosomes. In mutation phase, we randomly pick a chromosome and change one or
multiple points of the chromosome randomly. Then, we sort all the generated as well
as existing solutions based on the problem objective. Here, we explain the chromosome
design and the GA operators and GA steps.
For the network design, we encode the individuals as depicted in Figure 3.4. In the
individual encoder, rows represent pools’ locations and columns represent the radio units.
The intersection of rows and columns (pools and RUs) consists of a vector showing the
DU activation. If any of the DUs are active, then the corresponding pool location becomes
active and the active DU is responsible for serving the RU. Additionally, knowing the num-
ber of active RUs in each pool gives the number of active digital units, since each digital
unit has limited capacity of serving RUs. Furthermore, the required number of splitters
to implement such a network can be obtained by figuring out the number of active DUs,
3.2. MIGRATION COSTS TOWARDS C-RAN 51
(a) The chromosome encoder.
(b) The mutation and crossover phase example
Figure 3.4: An encoder example for 3 RUs, 3 pools, and 5 DUs per pool. RU 1 and 2 are
connected to DU 1 and 2 in Pool 1 respectively, pool 2 serves RU 3 and Pool 3 is inactive.
RUs, and pools. It is worth mentioning that some constraints such as distance constraint
are considered when the individuals are selected, therefore, the calculated number of active
splitters are based on feasible solutions and hence we do not consider all of them in the
individual formation phase.
3.2.4 Selected Results

We solve the DU pool optimization problem using IBM ILOG CPLEX Optimization Stu-
dio and genetic algorithm. The former can solve the ILP problem optimally but is not
scalable while the latter is scalable and can be used for larger network size. We com-
pare the resulting TCO of C-RAN scenarios with D-RAN, to assess the migration cost of
C-RAN. The assumptions for the simulation of this study can be found in [5].
To show the impact of functional splitting, TCO for C-RAN with functional splitting
is plotted in Fig. 3.6. Since the RU sides have processing units, the network faces with
higher costs incurred by extra power consumption and benefits from less required capacity
and equipment in the architecture.
52 MANAGEMENT
106
2
Pool build-out
Equipment
Fiber
1.5 OpEx
0.5
0
AN FS AN FS AN AN
-R th -R th -R -R
ldC , Wi ldC , Wi eD lD
fie AN fie AN v ca
n R n R wa pti
ow C- ee C- cro O
Br ie ld Gr ie ld Mi
nf nf
B row G ree
Figure 3.5: TCO of Different Network Architectures [5].
10 6
4
C-RAN, Alpha=0.25
C-RAN, Alpha=1
3.5 Optical D-RAN
Microwave D-RAN
C-RAN, Alpha=0.25, With FS
C-RAN, Alpha=1, With FS
3
2.5
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18 20
Years
Figure 3.6: TCO of C-RAN and D-RAN over the years, with TCO minimization, for
greenfield pool deployment, with and without FS [5].
Fig. 3.6 shows the impact of TCO of functional splitting and availability of fibers in-
3.3. CONTENT PLACEMENT AND FUNCTIONAL SPLIT OPTIMIZATION 53
frastructures to the operators. When more fiber is available, the cross over points (years
that the cost of D-RAN is more than that of C-RAN) is lower meaning that in a short
term the costs of C-RAN will be lower than cost of D-RAN for the operators. Moreover,
the TCO of C-RAN with functional splitting is lower that conventional C-RAN, because
C-RAN with functional splitting has much lower OpEx while the CapEx are similar.
Fig. 3.5 depicts the breakdown of TCO for green/brown field C-RAN with/without
functional splitting and optical/microwave D-RAN. The initial investment of greenfield
fiber deployment is much higher than that of brownfield deployment but it gives us more
flexibility to design the optimal fronthaul with lower OpEx. With enabling functional
splitting, TCO is decreased since the OpEx and fiber leasing cost is reduced. Compared to
D-RAN, both greenfield and brownfield fiber deployment require more initial investment.
However, this cost is paid to decrease the OpEx. The reduction in OpEx means reduction in
annual cost and hence after a certain crossover time (in case of brownfield fiber deployment
about 4 years), C-RAN becomes an economically viable solution.
3.2.5 Conclusion
In this part, we investigate the deployment cost of different C-RAN architectures. We
propose an ILP optimization problem to minimize the TCO of C-RAN network and an
AI-based algorithm to solve the problem. We show that the fiber-rich operators should
migrate to conventional C-RAN while for fiber-short operators migrating to the C-RAN
with functional splitting is a more reasonable choice. We also show that C-RAN with
functional splitting had lower TCO and crossover time compared to C-RAN. Furthermore,
in terms of upgradability, C-RAN was more cost efficient compared to D-RAN.
3.3 Content placement and Functional Split Optimization

In cellular networks, caching popular contents in the edge can reduce the backhaul cost,
access latency and energy consumption. Since fronthaul capacity becomes a bottleneck,
especially in ultra dense networks of 5G, we can equip the edge clouds with caches to
reduce the usage of fronthaul links [117]. The transportation link between edge cloud
and central cloud is referred to as "midhaul". The motivation of dual-site processing is
two-fold: a) by partial baseband processing at edge cloud, bandwidth requirement can be
significantly relaxed for midhaul; b) by equipping RSs with general-purpose processors,
traffic load can be terminated at edge cloud to relieve traffic load in core networks [118].
In this study, we investigate the joint impact of content placement and functional split
on the energy consumption of network. In particular, we aim at minimizing the total energy
consumption of network, by jointly optimizing the functional split point and cache place-
ment, while satisfying users QoS. The users’ QoS is defined as a minimum delay which
users can tolerate until receiving their requested files. The main contributions of this study
is as follows:
• The formulation of a experienced delay of delivering requested contents to users and

energy consumption of network considering both QoS of users and functional split
in the network.
54 MANAGEMENT
• Investigating the impact of content placement and function split on the energy con-
sumption of network.
3.3.1 System Model

According to Fig. 3.3, we consider a hybrid architecture with a central cloud and multiple
edge clouds. We assume that the CC and EC are equipped with memory units and DUs so
that the popular files can be stored and network can enable virtualized functional process-
ing. In this setup, computational resources can be shared by connected RUs. The flexibility
of the EC/CC functional processing also depends on the placement of the requested files.
Hence, sharing infrastructure equipment based on the optimal content placement creates
a trade-off between the consumed power, midhaul bandwidth and experienced delay. Our
quest hence becomes (1) whether to fully perform our baseband processing at the CC, the
EC, or select an intermediate split point and (2) whether to place the content at the edge or
central cloud given a user request with a delay threshold. In this setup, each user request
a content and network must deliver the content within a given threshold. The content can
be retrieved from EC, which is less energy efficient but has a better delay performance,
or from CC, which is more energy efficient but has worse delay performance. Therefore,
together with functional split we need to find the best location of contents to improve the
delay performance. This issue is formulated in the following section.
3.3.2 Problem Formulation

In this problem, we aim at minimizing the power consumption of the network by jointly
optimizing the content placement and functional split. To consider QoS of users, we guar-
antee the total delay of a user should be less than the delay threshold. Also, the functional
split can occur only once either at cell processing or user processing. If all functional pro-
cessing are centralized at CC then corresponding file cannot be placed at EC. If the file is
optimally placed at EC, then function processing must be placed at EC. The problem is
modeled as an ILP problem, and is formulated as follows:
Given: Number of: cells, edge clouds (ECs), users. Delay threshold, maximum capacity
of: cache at each EC, cache at central cloud, DU processing, and fronthaul links.
Constraints: Service oriented delay, x-haul limited capacity
Find: Functional split of a user, active DUs to host user/cell processing at edge/central
cloud, optimal capacity of cache at edge cloud, best place to cache contents either at ECs
or CCs.
Objective: Power consumption of the network.

For this study, we invesit we assume maximum number of 95 users are connected to the
RUs of the corresponding cells. All the simulation parameters are presented in paper D.
FSCP is compared with 2 reference cases where files are cached at EC (1) and CC (2).
In Fig. 3.7 we investigate the total power consumption of the network. When all
contents are at the EC, it is not possible to centralize functions and hence more energy is
3.3. CONTENT PLACEMENT AND FUNCTIONAL SPLIT OPTIMIZATION 55
Figure 3.7: Power consumption of the network [7].
consumed. Moreover, the capacity of the edge cloud is limited, therefore, after a point, no
more users can be served and a saturation point will be observed in the power consumption.
When the delay requirement is less stringent, we have more room for centralizing functions
and hence more energy is saved.
In Fig. 3.8, we have depicted the average delay experienced by the users. Fetching the
content from the CC experiences more delay than other cases due to the higher distance
between the CC and the users. FSCP-80-20 has up to 20% lower delay compared to FSCP
due to the content sharing. Comparing this trend with that of Fig. 3.7, one can see the trade
off between the experienced delay and the the total power consumption.
3.3.4 Conclusion
In this subsection, we show that the proposed content placement algorithm can reduce the
content delivery delay to the users. On the other hand, we can save energy in the network by
turning off the extra unused equipment such as digital units, and line cards. The simulation
results show that the popular contents for delay stringent services can be cached at the EC,
leaving the rest in the center cloud for better energy and QoS performance.
56 MANAGEMENT
Figure 3.8: Average delay of the users vs percentage of active users [7].
3.4 Energy Efficient End-to-End Network Slicing
5G and beyond networks will support a wide range of services with diverse and heteroge-
neous service requirements. The service demands are evolving while the requirements are
very diverse and heterogeneous. Therefore, one architecture does not fit for all applications
and services. The current network architecture utilizes a relatively monolithic network and
transport framework to accommodate these emerging services. It is anticipated that the
current architecture is not flexible and scalable enough to efficiently support a wide range
of services each with their own specific set of quality-of-service requirements [119]. Net-
work slicing is a promising technology for realizing this vision. With network slicing, the
network is partitioned into multiple dedicated virtual networks tailored and customized for
specific services. Network slicing makes it possible to create fit-for-purpose virtual net-
works with varying degrees of freedom for each service [120]. It enables service-oriented
resource allocation by tailoring the network resources, e.g., bandwidth and processing, to
specific services [121]. Hence, each service can have its own customized network. In net-
work slicing, resources are reserved in advance for each service. Therefore, the resource
utilization may not be energy efficient. In fact, by network slicing the energy consumption
of the network will increase and this is the cost network should pay to support wide and
heterogeneous services.
In the literature of network slicing, the main focus is to provide and guarantee the
required QoS to the users. However, the energy consumption of such approach is not
3.4. ENERGY EFFICIENT END-TO-END NETWORK SLICING 57
well-investigated. On the other hand, most of the studies are focusing on RAN slicing
where only RAN resources are dedicated and reserved for a slice. In terms of energy
consumption, one may increase the bandwidth in RAN to reduce the energy consumption
in the transmission phase. However, this does not necessarily lead to reducing the total
energy consumption of the network. Because more bandwidth requires higher data rate
and data transport in mid/fronthaul and more processing at the cloud side. Therefore, the
energy consumption performance should be evaluated from end-to end perspective. In
the literature of network slicing, the problem of end-to-end network modeling is not well
investigated. In this study, we propose end-to-end delay and energy model for the end-to-
end network slicing in Ethernet-based C-RAN architecture and propose an optimization
problem to minimize the networks’ energy consumption by allocating the orthogonal re-
sources to each slice. We guarantee the slice QoS is satisfied by imposing constraints to
the problem. The main contributions of this study are summarized as follows:
• We highlight the underlying trade-offs between communication and computation

resources allocations and their mutual impacts on each other.
• We formulate an optimization problem to minimize the energy consumption of the

network by jointly allocation of communication and computation resources to each
slice. We take into account the delay requirement of each slice, limited capacity of
Front/mid haul, and limited processing and bandwidth resources in the cloud and
RAN.
• Through simulations, we highlight the impact of 5G numerology on the required

communication and computation resources for each slice. Furthermore, we investi-
gate the interplay between communication and computation resources.
3.4.1 Network Architecture for Slicing

In C-RAN-based architectures, the RUs are distributed in a network while the DUs are
separated from them and pooled in a CC. We assume a network with maximum number of
CC,max CC
NDU DUs, which NDU of them are active, are pooled at the CC. These parameters
and the remaining ones are summarized in Table 3.1. Each BS is equipped with a RU and
single DU for low-PHY processing. BSs are connected with DUs in the CC via an evolved
common protocol radio interface (eCPRI). In our study, we assume a fiber-based Ethernet
interface that supports four lanes of 25 Gbps transport data rate [122]. The connection from
each RU to its co-located single DU is called “fronthaul link” whereas the “midhaul link”
is the connection from the BSs at the radio site to the CC.
The eCPRI standardization [122] supports specific functional splits. In this study, we
consider downlink operation and assume the functional split IID is implemented in the
network. This split point requires less capacity in the midhaul and allows us to benefit
from PHY layer techniques such as coordinated multipoint (CoMP) [123]. The data rate
for the functional split IID envisioned in eCPRI standard is function of assigned number of
sub carriers, the symbol time, number of antennas, and number of quantization bits. The
utilized midhaul interface must be chosen to support such a bit rate.
58 MANAGEMENT
3.4.1.1 Slice Demand
In this study, we define the slice as a network with dedicated computation and communi-
cation resources which use the transport network to carry the corresponding traffic. Each
user in a cell is interested in a particular service. Each service belongs to a slice depend-
ing on its delay requirement. The network slice is configured to satisfy a predefined delay
threshold with dedicated end-to-end resources to meet the respective requirements. We
assume that the cardinality of the set of the users is determined by the predicted maximum
number of users that will be active in the network simultaneously. We take into account
four types of slices namely, critical machine-type communications (MTC) (which includes
delay stringent services), eMBB delay sensitive, eMBB delay tolerant, and massive MTC.
Critical MTC services have the most strict delay requirements with typically small file
sizes. The delay requirement for these types of services are below 20 ms. eMBB delay
sensitive services may tolerate higher delay but still below 100 ms. eMBB delay tolerant
services are less sensitive to large delays but still need to receive their data with an ap-
propriate latency. Depending on the application, delay between 100 ms and 500 ms can be
acceptable for these services. For massive MTC services, the number of connected devices
can be much higher than other services but the delay requirements can be very strict as in
factory automation use case. Other types of massive MTC services are tolerant to larger
delays as in smart home or smart parking use cases. We assume that for each slice a certain
amount of data is to be delivered to the corresponding users within a delay threshold. This
delay threshold is representative for the slice’s and users’ QoS.
3.4.2 Slice Model

Each slice is composed of processing resources at edge and center cloud, switches, fron-
thaul links and equipment, and bandwidth in RAN as communication resources. We ded-
icate a part of network as a slice for each service. Depending on the class and priority of
services, they might use isolated or shared slices. In isolated slices, we dedicate the nec-
essary resources to the service and these resources are not shared with any other services.
The shared slices are used by multiple services and resources are shared between them.
Within each slice, we assume that there is a resource allocation module that allocates the
resources to the users.
Given the traffic for each service, we calculate the required processing per cell per
service to calculate the processing delay. Then, given the total tolerable latency and latency
in each segment of the network, e.g., delay in fronthaul and processing delay, we derive
the maximum tolerable delay in RAN in which we assign the radio resources to satisfy
this delay requirement of the services. The total delay budget to serve a user with specific
service is given based on the type of service. Among all terms of delay, which has been
described earlier, in this study we focus on two delay terms in RAN and processing. The
more resources is assigned in each part the less delay is incurred for that specific service
and leaving more delay budget in the other section. It is worth mentioning that allocating
more resources to each slice will limit the available resources for other slices and imposes
extra energy consumption to the network. In this study, we assume fixed the split point, i.e.,
Figure 3.9: Simple overview of slice manager
Figure 3.10: Computation and communication resources assigned to each slice
ID defined in eCPRI standardization [122], for each slices. This point of split mitigates
the fronthaul capacity requirements while allowing to benefit from PHY layer techniques
such as joint transmission/reception coordinated multipoint (CoMP) [66].
60 MANAGEMENT
Slice Energy Model: Each slice is composed of set of RUs, switches, assigned sub-
carriers, and DUs. The energy consumption per slice is given by,
CC CC CC

Pcool + NDU PDU,min T
Esslice =
ζCC |S|
CC X X N sw
Pproc X
+ dpr
CC,u,s,r + sw
Ps,k T
ζCC
r∈R u∈Us,r k=1
BS
1 X Pstatic T DU
+ + Ps,r T
ζBS |S|
r∈R
1 X X sc tx R,tx
+ Nu,s,r Ps,r Du,s,r , (3.17)
ζBS
r∈R u∈Us,r
where the static power is distributed equally among the slices. It is worth mentioning that
different sharing of constant power consumption is possible among each slice or service
categories. Equal sharing assigns equal share of power to each slice, however, it intro-
duces a non-affordable cost burden to the slice with small loads which might be essential,
such as voice, or yet newly introduced ones. Another approach is proportional sharing in
which the constant power consumption is shared proportional to the requested data rate
of each slice/service. It requires quantifying the amount of traffic produced by each ser-
vice category and then sharing the power consumption between them. In this case, we are
penalizing slices or services with high loads which are major driving forces for network
operation and revenue. Another approach is to use Shapely value in which it yields to
smaller share for large service categories than proportional sharing and a smaller share for
small service categories than equal sharing [4].
3.4.3 Minimization of Network Energy Consumption

The aim of this section is to propose a method to efficiently assign resources with minimum
energy consumption. To minimize the energy, the duration of request must be taken into
account. Duration of request is defined as the duration which dedicated network resources
are assigned to a specific service. In our study, we assume that this duration is estimated in
the demand prediction stage. Since the service requirements are vastly heterogeneous, the
assigned resources within a demand duration or slice window should enough to support the
required QoS for each service.
In this study and under the C-RAN architecture, we formulate a radio and network
resource allocation problem containing two sub-problems. The first problem is centralized
joint communication and computation resource allocation for each slice in the network.
Depending on the number of users per slice, number of slices, and their QoS requirement,
the slice manager decides on the required bandwidth and processing resources and reserves
them for each slice. The second problem is how to dynamically allocate the reserved
resources to the users in each slice.
Figure 3.11: Network architecture and slice manager [8].
The first problem problem is joint computation and communication resource allocation
for network slicing. The slice manager solves the optimization problem (which is explained
in Paper 6). At each time instance, we will solve the following problem and create services
slices.
Given: Number of users at each time with their demanded QoS, requested services,
architecture and its parameters such as line rates, processing capacities, bandwidth, switch
rate, etc.
Objective: Minimize the energy consumption of the end to end network
Constraints: Experienced delay of each user, used bandwidth, used processing capac-
ity, used rate in the fronthaul
Optimization variables: bandwidth for each service in each cell, processing resources
for each service in the center cloud.
Once the problem is solved, we can calculate the number of subcarriers per allocated
bandwidth. We assume the Gops at CC is a linear function of the bandwidth, we can
express the required processing as a function of the bandwidth. We express the number of
active DUs as a function of the processing capacity of a single DU at CC and the required
processing resources.
3.4.3.1 Intra-slice resource allocation problem

Once the network slicing problem is solved, the reserved bandwidth and the allocated pro-
cessing is given for each slice and RU. However, there might be several and different users
at each time instant in the network. Therefore, the radio resources within the slice should
be allocated to the users, so that their QoS is satisfied [15]. Given the reserved communica-
tion and computation resources for each slice, an adaptive radio resource allocation scheme
perform the resource allocation to the real-time user arrival and departures [104]. The re-
sponsibility of the radio resource allocation scheme is to obtain the information on the
reserved RAN resources to the slice and allocate them to the slice users to keep the RAN
energy consumption as low as possible while maintaining the QoS of the users [124]. If
62 MANAGEMENT
No
Yes
Slice Requirements Reconfigure Slice? QoS Monitoring
Slice Resource Allocation
Slice Demand Prediction Radio Resources

Resource Allocation per Slice
Processing Resources
Slice Manager
Figure 3.12: Slicing life cycle management with QoS guarantee [8].
satisfying the users’ QoS is not feasible, then slice reconfiguration may be required which
is explained in the following subsection.
3.4.3.2 Slicing life cycle management
When resources are reserved for each slice, the slice manager needs to know whether the
provided resources are enough for satisfying users’ requirements. Therefore, the network
should continuously monitor the QoS to make sure that the users’ QoS are satisfied. In Fig.
3.12, we show how the slice manager interacts with the network. The slice manager gets
the information of all slice requirements and their corresponding load prediction and ac-
cordingly reserves the communication and computation resources for each slice and casts
this information to each RU. At each RU and for each slice, the assigned resources are
allocated to the users. The network examines the experienced users’ QoS. When network
status changes, e.g., we have more users than expected, the provided resources in the slice
may not be enough to meet the QoS requirements and the slice resources should be recon-
figured. If BS cannot serve the users with their QoS, it submits the slice reconfiguration
request to the slice manager and raise the flag if slice reconfiguration is required. The slice
manager reconfigures the allocated resources per slice based on the new predictions and
update the RUs with the new assigned resources so that they can serve users with their
required QoS within each slice.
On the other hand, if the resources are not utilized for the slice window, again RU
can submit the slice reconfiguration request to release the assigned resources in favor of
providing more resources to other services, improve the resource utilization, and decrease
the energy wasted for keeping the network resources idle. With this approach, we can
dynamically manage the life cycle of each slice.

In this section, we provide the selected results to evaluate the performance of the proposed
end-to-end slicing method, i.e., slice-based resource allocation (SBRA), and compare it
with the baselines. In this study, we assume that the expected number of users in each slice
is provided to the slice manager in advance. Therefore, the respective optimization problem
is solved to reserve resources for the future. We consider a network with |R|= 4 number of
RUs and |S|= 4 slices namely, critical MTC, eMBB delay sensitive, eMBB delay tolerant,
and massive MTC. The maximum tolerable delays for these slices are 10, 50, 100, 150 ms,
respectively. The users’ file sizes for each slice are 0.1, 2, 2, 0.1 MB, respectively. We
assume that the 5G NR numerology index is 0, otherwise stated. We assume that each BS
is equipped with a single RU and DU, and the slice duration (slice window) is T = 1 s.
We assume that the processing limit of DU of the BS is the same as that of the DUs at the
CC. Moreover, the processing resources of the BSs in (3.5) are allocated proportional to
the load of each slice.
We compare the proposed solution with two cases: RAN slicing and user-based re-
source allocation (UBRA). With RAN slicing we only allocate the bandwidth to each slice
and the computation resources are assumed to be fixed and given. We assume that the
allocated computation resources are enough to serve the peak load and this amount is fixed
for all loads. Comparing with this baseline, we can highlight the impact of considering
computation resource allocation. For UBRA, we create an end-to-end slice per user and
we allocate processing resources in CC and bandwidth in RAN to each user. Comparing
the SBRA with UBRA, we can demonstrate how the proposed SBRA can decrease the
complexity of the network slicing and how close the solution is to optimal solution.
In Fig. 3.13, we illustrate the hourly energy consumption of the network for two cases
of SBRA and RAN slicing. To generate this figure, we use the real network data from
BSs in rural area of Uppsala county. We extracted the number of users belonging to the
eMBB delay sensitive and eMBB delay tolerant services in each hour of a weekday as is
depicted in the right hand side axis of Fig. 3.13. Since the information on the massive
MTC and the critical MTC services is not available in the data set, we set the number of
critical MTC users to 1 and the number of connected massive MTC devices to 100. As
can be seen from this figure, SBRA performs better in off-peak hours of the day because
it will utilize less processing resources and hence can turn off the unnecessary equipment.
However, in RAN slicing the processing resources can be only in idle mode, consuming
energy. On the other hand, in the peak hours of the day, the network energy consumption
difference is smaller because most of resources are in use in both methods. However, still
SBRA slightly outperforms RAN slicing due to end-to-end resource allocation.
Fig. 3.14a compares the allocated processing resources per user per slice for SBRA.
According to this figure, the SBRA will give us an estimate of the average required re-
sources for the next slice window. This figure illustrates that critical MTC slice which has
the tightest delay constraints requires more processing resources per user than the other
slices. Massive MTC users requires the least amount of processing resources due to their
relaxed delay constraints. Fig. 3.14b compares the total allocated processing resources per
each slice. The total traffic demand in each slice, depicted in Fig. 3.14c, and their delay
64 MANAGEMENT
20
SBRA
520 18
RAN Slicing
eMBB Delay Sensitive
eMBB Delay Tolerant 16
500
Energy consumption (J)
14
Number of users
12
480
10
8
460
6
440 4
420 0
2 4 6 8 10 12 14 16 18 20 22 24
Hour
Figure 3.13: Hourly profile of expected energy consumption and number of users for
eMBB delay sensitive and eMBB delay tolerant slices [8].
requirements determine the total required processing at the CC. For instance, in this study
since the number of massive MTC devices is higher than other slices, in total it requires
more processing resources. Please note that each massive MTC device requires much less
processing resources as is depicted in Fig. 3.14a. On the other hand critical MTC de-
vices, despite their lower load than other slices, they require higher processing resources
than eMBB slices. Because critical MTC slice has the most stringent delay requirement
and hence higher processing resources should be assigned to it, so that the QoS is satis-
fied. It is worth mentioning that the file size of each user may differ in UBRA, and thus
the required processing resources may differ from SBRA. Therefore, compared to SBRA,
less processing resources may be assigned to the users in UBRA for one slice and more
processing resources be assigned to other slices.
For UBRA the joint communication and computation resource allocation problem is
more complex in comparison to the proposed end-to-end slicing SBRA. The reason is that
the number of optimization variables as well as the number of constraints for users’ QoS
linearly scales with number of users while for SBRA the number of constraints is in the
order of the number of slices. Therefore, it takes more time to solve the problem of UBRA
as is depicted in Fig. 3.15.
3.4.5 Conclusion
In this study, we investigated the problem of joint communication and computation re-
source allocation to create an end-to-end network slice in a C-RAN based architecture. We
1%
Critical MTC
Massive MTC
Delay Sensitive 8%
Delay Tolerant 8%
25%
32%
23% 22% 37%
65%
37%
4% 38%
(a) Allocated processing re- (b) Allocated processing re- (c) Proportion of traffic de-
sources per user. sources per slice. mands per slice.
Figure 3.14: Allocated processing resources to each slice and user [8].
35
SBRA
UBRA
30
25
Time(Seconds)
20
15
10
0
1 2 3 4 5 6 7 8 9
Number of users
Figure 3.15: Time complexity performance comparison of SBRA and UBRA [8].
formulated an optimization problem to allocate communication and computation resources

to each slice. The objective of the problem is to minimize the energy consumption of the
network while satisfying the delay requirement of each slice. We took into account the
limited available bandwidth in the RAN and the limited amount of processing resources at
the CC. The results show that end-to-end network slicing is more energy efficient than only
RAN slicing due to joint consideration of the communication and computation resources.
With the proposed approach significant energy and bandwidth saving can be achieved com-
pared to the RAN slicing in low load scenarios. Moreover, we showed that despite the less
traffic corresponding to the delay stringent services, they occupy considerable share of the
66 MANAGEMENT
processing resources in the center cloud. Finally, We showed that user based slicing re-
quired more communication and computation resources and has higher time complexity
compared to the proposed slice based resource allocation.
3.5 Summary
In this section, we investigate three different aspect of the mobile network architecture
design. We propose an end to end delay and power (energy model) for the C-RAN based
architecture. Then we evaluate the migration cost from D-RAN to C-RAN. We show that it
is more cost-efficient for the fiber-rich network operator to migrate to C-RAN architecture.
The reason is that they can use their existing fiber infrastructure as the front/mid haul.
In this case, the cross-over time is lower than that of fiber-short operators. Fiber-short
operators, should benefit from functional splitting so that they require less capacity in the
front/mid haul. After examining the network migration cost, we investigate two different
methods of improving the the users’ QoS. In the first method, we improve the the delay
performance of delivering a file to the users by caching them at the edge cloud, closer to
the user. For this study, we minimize the power consumption of the network, by finding
the best location of each file, and the best functional splitting point for the users while
satisfying their required delay performance. In the last study, the aim is to serve multiple
services with heterogeneous required QoS. For this purpose, we create multiple service-
based customized network that are made on top of same network infrastructure. Each
customized network is know as slice and the method is called network slicing. In this
method we assign end to end resources to each slice so that they can serve their users with
the requires QoS. The result showed that end-to-end network slicing can save more energy
compared to the only RAN slicing.
Chapter 4
Conclusions and Contributions
4.1 Conclusions
Mobile networks are a major energy consumer accounting for 1 − 2% of the global en-
ergy consumption. As we move ahead into 5G and beyond, the energy consumption will
be increasing considering new demands on service providers to increase network capac-
ity, extend geographical coverage, and deploy advanced technology use cases. In order
to assure the sustainability of the mobile networks, it is of great importance to improve
the network energy efficiency. In this thesis, we investigated the potential techniques to
improve the 5G and beyond network’s energy efficiency. We divide the contribution of the
thesis into two parts, i.e., 1) AI assisted green mobile network where we reduce the energy
consumption of the BSs in the network, and 2) AI assisted network architecture design and
management where we investigate the end-to-end network design and resource allocation
to improve the energy efficiency of the networks.
In the first part of thesis reducing the BSs energy consumption is under the focus. BSs
are one of the most energy consuming components of the network. Hence, we can reduce
the energy consumption by sleeping the BS when it is idle. Then, the question of when
and how deep a BS should sleep so that the energy consumption reduces without adverse
impact on the users’ QoS. To tackle these issues, we defined the RQ1-2 as:
• RQ1: How to dynamically put BSs into sleep modes when they are idle, i.e., when
and how deep to sleep, given the previous user arrivals and network load?
• RQ2: Depending on how deep BS sleeps, how can we design a network manage-
ment framework continuously analyze and monitor the risk while using ML to make
decisions?
To handle the issues in RQ1-2, we have devised a novel framework to define the risk
linked to the intention of mobile operator and translated this into network KPIs. We created
a novel network management framework with a digital twin continuously analyzing the
probability of risk and orchestrating the usage of ML. We propose an ML-based technique
67
68 CHAPTER 4. CONCLUSIONS AND CONTRIBUTIONS
to save energy at the BS, utilizing advanced sleep modes, while maintaining the users’ QoS.
We have combined analytical model based approaches with ML. Model based approaches
are used for risk analyses together with the operator traffic data to create a digital twin
which can mimic the behavior of a real BS with advanced sleep modes. Digital twin is
used for assessing the risk, continuously monitoring the performance by controlling the
usage of ML to make sleep mode decisions. Our study shows that there is a potential
to reduce base stations’ energy consumption by more than 50% when network operates at
10% average load. At around 20−30% average load, BSs can enter in sleep for a 70−80%
of time, leading to a gain in energy consumption of more than 20%. We show that using the
proposed method with achieving near optimal results can be achieved in low traffic hours
and sub-optimal energy saving in busy hours.
Today large telco players achieved a consensus on an open RAN architecture based
on hybrid C-RAN studied in this thesis. We studied and modeled end-to-end energy con-
sumption and delay of a cloud native network architecture based on virtualized cloud RAN
forming foundations of open RAN. Moreover, both network architecture design and the
utilizing network energy saving features, playing key roles in improving the energy effi-
ciency of the network meanwhile supporting multitudes of heterogeneous services.
Current network architectures cannot meet the requirements of the 5G and beyond
systems. Therefore, there is a need to devise and design new energy efficient architectures
and migrate to it. However, migration to new architectures is costly and imposes new
constraints and challenges to be tackled. To address these issues, we define the following
research questions:
• RQ3: What is the deployment cost of different C-RAN architectures? How sensitive
an architecture is to the price changes?
• RQ4: How artificial intelligence can help us to design a C-RAN fronthaul with min-
imum cost and energy consumption?
In order to answer the RQ3-4, we model the migration cost, in terms of both OPEX and
CAPEX, from a conventional distributed RAN architecture to a C-RAN architecture with
economic viability analyses of a virtualized cloud native architecture considering the future
traffic forecast. It is known that C-RAN is energy efficient compared to D-RAN due to the
cooling efficiency and processing multiplexing gains in the centralized cloud however it is
not clear under what conditions it becomes also cost efficient considering the infrastructure
cost of fronthaul and fiber links. We solved the optimal fronthaul design problem in terms
of costs and energy consumption in C-RAN networks. For this study, we formulate the
problem as the ILP and propose an AI algorithm for the optimal fronthaul design. We
show that the fiber-rich operators should migrate to conventional C-RAN while for fiber-
short operators migrating to the C-RAN with functional splitting is a more reasonable
choice.
• RQ5: Is there a way to still save energy while meeting with the delay requirements
of users by placing content to the edge cloud instead of centralized cloud?
4.2. DISCUSSIONS AND FUTURE WORK 69
• RQ6: How to allocate the network resources in fronthaul, edge cloud, and central-
ized cloud to minimize the energy consumption while meeting the users’ QoS?
To address the RQ5-6, in a multi-layer hybrid C-RAN architecture, we solved the prob-
lem of joint content placement and functional split optimization by minimizing network
energy consumption while meeting with the QoS requirements of mobile users. The chal-
lenge here is the trade-off between energy consumption and delay optimization. Energy
efficiency is achieved when all network functions are centralized. To optimize the delay,
content needs to be placed at the edge-cloud requiring network function processing at the
edge. In this study, it is demonstrated that utilizing the content caching at edge cloud to-
gether with functional splitting is beneficial in terms of content access delay but at cost of
power consumption. Joint optimization of function splitting and content cashing allows us
to find a compromise in between delay and power consumption.
• RQ7: How to allocate computation and communication resources to each slice for
satisfying the QoS and minimizing energy consumption?
• RQ8: How to design a network management framework for slice admission to con-
tinuously maintain the performance of dynamic slicing in terms of energy saving?
Finally, to address the RQ7-8, we investigate the problem of end-to-end network slicing
by joint allocation of communication and computation resources in C-RAN networks. We
investigate the tradeoff between communication and computation resource. In this study,
end-to-end network slicing is optimized for the first time considering the energy consump-
tion in addition to the QoS guaranteeing. Most network slicing studies consider only radio
access network resources and the goal of slicing is to guarantee the performance for spe-
cific services which in turn increases the energy consumption of the network. Intuitively
energy consumption goes down if more bandwidth resources is allocated to users when
RAN segment of the network is considered. Thanks to our end-to-end energy consumption
model it is demonstrated that this is not always true since increasing bandwidth alloca-
tion increases processing energy consumption in the cloud and fronthaul segment of the
network.
To sum up, in the first part of this thesis a new obstacle of practical usage and risk
of AI for managing mobile networks is treated under the context of advanced sleep mode
management. A new risk-aware network management framework combining model based
approaches with ML is proposed considering the intent of the mobile operator as an input.
The concept of digital twin is formalized under this framework. In the second part of
the thesis, cloud native network architectures are studied based on C-RAN by developing
end-to-end cost, energy consumption and delay models and jointly optimizing performance
together with energy consumption founding the basis for the future open RAN architecture.
4.2 Discussions and Future Work

In the journey to beyond 5G networks, energy efficiency will be one of the key pillars
of network design and management. On the other hand, in recent years it is possible to
70 CHAPTER 4. CONCLUSIONS AND CONTRIBUTIONS
obtain mobile networks data. Utilizing the available data can significantly help the mobile
operators and vendors, to design, optimize, and maintain their network. Moreover, thanks
to the advancement of the field of artificial intelligence and processing technologies which
can handle processing extensive tasks, it is now possible to benefit from the AI techniques
to assist the network management and improve the performance metrics. In this thesis, we
cover many aspects of improving the network energy efficiency in the mobile networks.
However, there are nevertheless some limitations that this thesis and can be improved and
extended.
Intelligent Energy Saving at Base Stations
Various energy saving features are already implemented in the base stations. However,
due to the trade off between provided QoS and energy saving, the mobile operators are
very conservative about activating the energy saving feature. In the studies in [6, 49, 63],
we take into account the activating the base sleep modes to save energy while minimizing
the adverse impact on the users’ QoS. However, in [6, 63] single BS is considered and the
problem of multiple BSs with distributed learnign is still open. Moreover in these studies,
the impact of interference is neglected. In fact, with interference from the neighboring
cells the required time to serve users will increase, hence less time will be left for BS to
sleep. The impact of interference on the decision can be investigated as one extension of
the works. In [49], we adapt the BS’s configuration to increase the potential sleeping time
of BS considering the neighboring interference, however, the decisions on the sleep modes
are based on average values and can be adapted dynamically to enhance the energy sav-
ing performance. In all above studies, the underlying assumption is that the algorithms are
trained over a generic BS. However, depending on the hour of the day, and the geographical
location of the BSs, the solution of the proposed algorithms may be different. Hence, it is
important to incorporate the spatio temporal data into the analysis and provide the solution
depending on the time and position of the BS. Since BSs are serving users who are using
different services, and services have heterogeneous QoS demands, some delay stringent
services cannot tolerate the imposed delay caused by BS sleeping. Hence, the heteroge-
neous service requirements as well as allocating resource blocks in the time-frequency grid
have impact on the decisions and can be considered as a research direction to improve the
performance of our studies.
Risk Aware Network Management
It is significantly crucial to make sure that the algorithms that are implemented in the
network do not cause any unexpected performance degradation to the network. This issue
becomes even more critical when AI based algorithms are utilized because there is always
a risk associated to the AI algorithms. In this thesis, we addressed the importance of risk
and performance degradation analysis in the network in [6, 8] where we investigated this
issue. However, this can be further analyzed for the AI based algorithms to be utilized in
the network. For instance, in the content placement problem [7], the content delivery time
to the user is a key performance metric, and the impact of any AI algorithms on this metric
should be quantified and analyzed. Similar analysis is required for any AI assisted energy
saving algorithms, to avoid the risk of performance degradation in the network. Therefore,
one research direction is to extend the analysis of energy saving methods by adding another
control layer to manage and monitor the risk.
4.2. DISCUSSIONS AND FUTURE WORK 71
Dynamic Network Management

Network traffic varies during the day and also the pattern can vary depending on the
day of the week. If we design the network according to the peak traffic, we can guarantee
that there will be enough resources during the day so that all traffic can be served with their
QoS (expect for unexpected events or anomaly behavior of the traffic). However, with this
static approach network resources will be underutilized. Moreover, this approach is not
energy efficient since many network components can be put to sleep mode, consuming
much less energy than the idle mode. This motivates the researchers to design dynamic
algorithms so that they can adapt their resources according to the network status. In par-
ticular, in [7, 8] we proposed two static method to save energy in the network. In both
studies, network operator’s data can be used to show the energy saving opportunities and
their impact on the QoS during the day for each location type. In particular, in [7], the file
and service popularity can be considered to adapt the caching policy and improve the con-
tent delivery time. In [8], network data can be categorized based on the popular services
per time and per location so that the operator be able to assign resources in advance for
that specific geographical location. Since the end to end slicing includes RAN resources,
one can benefit from machine learning approach to train the global network by the use
of the local information at each node for instance using federated learning or distributed
learning methods [125]. These approaches can be very useful because network data is not
concentrated in a single network node and is distributed over multiple places and network
layers.
Bibliography
[1] “Mobile data traffic outlook,” https://www.ericsson.com/en/mobility-report/

dataforecasts/mobile-traffic-forecast, accessed: 2020.
[2] “5G technology: Mastering the magic triangle,” https://www.reply.com/en/

industries/telco-and-media/5g-mastering-the-magic-triangle, accessed: 2020.
[3] A. Osseiran, F. Boccardi, V. Braun, K. Kusume, P. Marsch, M. Maternia, O. Que-

seth, M. Schellmann, H. Schotten, H. Taoka et al., “Scenarios for 5G mobile and
wireless communications: the vision of the metis project,” IEEE communications
magazine, vol. 52, no. 5, pp. 26–35, 2014.
[4] M. Masoudi et al., “Green mobile networks for 5G and beyond,” IEEE Access,
vol. 7, pp. 107 270–107 299, 2019.
[5] M. Masoudi, S. S. Lisi, and C. Cavdar, “Cost-effective migration toward virtualized

C-RAN with scalable fronthaul design,” IEEE Systems Journal, vol. 14, no. 4, pp.
5100–5110, 2020.
[6] M. Masoudi, E. Soroush, J. Zander, and C. Cavdar, “Digital twin assisted risk-aware
sleep mode management using deep Q-networks,” submitted to IEEE Transactions
on Vehicular Technology, 2022.
[7] A. Sriram, M. Masoudi, A. Alabbasi, and C. Cavdar, “Joint functional splitting

and content placement for green hybrid CRAN,” in 2019 IEEE 30th Annual In-
ternational Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC). IEEE, 2019, pp. 1–7.
[8] M. Masoudi, D. Özlem Tuğfe, J. Zander, and C. Cavdar, “Energy-optimal end-to-

end network slicing in cloud-based architecture,” submitted IEEE Open Journal of
Communications Society, 2022.
[9] imec, GreenTouch Project, “Power model for wireless base stations.” [Online].
Available: https://www.imec-int.com/en/powermodel
[10] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof power model
for cellular base stations,” in Proc. IEEE VTC-Spring’15, May 2015, pp. 1–7.
73
74 BIBLIOGRAPHY
[11] A. A. Zaidi, R. Baldemair, H. Tullberg, H. Bjorkegren, L. Sundstrom, J. Medbo,

C. Kilinc, and I. Da Silva, “Waveform and numerology to support 5G services and
requirements,” IEEE Communications Magazine, vol. 54, no. 11, pp. 90–98, 2016.
[12] M. Masoudi, “Energy and delay-aware communication and computation in wireless

networks,” 2020.
[13] M. Masoudi, H. Zaefarani, A. Mohammadi, and C. Cavdar, “Energy efficient re-

source allocation in two-tier OFDMA networks with QoS guarantees,” Wireless
Networks, pp. 1–15, 2017.
[14] M. Masoudi, B. Khamidehi, and C. Cavdar, “Green cloud computing for multi
cell networks,” in Wireless Communications and Networking Conference (WCNC).
IEEE, 2017.
[15] M. Masoudi, H. Zaefarani, A. Mohammadi, and C. Cavdar, “Energy efficient re-

source allocation in two-tier OFDMA networks with QoS guarantees,” Wireless
Networks, vol. 24, no. 5, pp. 1841–1855, 2018.
[16] M. Masoudi and C. Cavdar, “Device vs edge computing for mobile services: Delay-
aware decision making to minimize power consumption,” IEEE Transactions on
Mobile Computing, 2020.
[17] M. Masoudi, A. Azari, and C. Cavdar, “Low power wide area IoT networks: Re-
liability analysis in coexisting scenarios,” IEEE Wireless Communications Letters,
2021.
[18] S. K. Gowtam Peesapati, M. Olsson, M. Masoudi, S. Andersson, and C. Cavdar,

“An analytical energy performance evaluation methodology for 5G base stations,”
in 2021 17th International Conference on Wireless and Mobile Computing, Net-
working and Communications (WiMob)(WiMob 2021), 2021.
[19] Ö. T. Demir, M. Masoudi, E. Björnson, and C. Cavdar, “Cell-Free massive MIMO

in virtualized CRAN: how to minimize the total network power?” in 2022 IEEE In-
ternational Conference on Communications (ICC), Seoul, South Korea, May 2022.
[20] A. Azari and M. Masoudi, “Interference management for coexisting internet of

things networks over unlicensed spectrum,” Ad Hoc Networks, p. 102539, 2021.
[21] M. Masoudi, A. Azari, E. A. Yavuz, and C. Cavdar, “Grant-free radio access IoT
networks: Scalability analysis in coexistence scenarios,” in 2018 IEEE International
Conference on Communications (ICC). IEEE, 2018, pp. 1–7.
[22] A. Azari, M. Masoudi, and C. Cavdar, “Optimized resource provisioning

and operation control for low-power wide-area IoT networks,” arXiv preprint
arXiv:1804.09464, 2018.
BIBLIOGRAPHY 75
[23] A. Azari, M. Masoudi, C. Stefanovic, and C. Cavdar, “IoT networks with grant-
free access: Reliability and energy efficiency analysis in coexistence deployments,”
submitted to IEEE IoT Journal, 2022.
[24] N. Zabetian, A. Mohammadi, and M. Masoudi, “Energy-efficient power allocation

for device-to-device communications underlaid cellular networks using stochastic
geometry,” Transactions on Emerging Telecommunications Technologies, vol. 30,
no. 12, p. e3768, 2019.
[25] A. Kuhestani, A. Mohammadi, and M. Masoudi, “Joint optimal power allocation

and relay selection to establish secure transmission in uplink transmission of un-
trusted relays network,” IET Networks, vol. 5, no. 2, pp. 30–36, 2016.
[26] M. Ganjalizadeh et al., “Translating cyber-physical control application requirements

to network level parameters,” in IEEE International Symposium on Personal, Indoor
and Mobile Radio Communications, 2020.
[27] M. Ganjalizadeh et al., “Impact of correlated failures in 5G dual connectivity archi-

tectures for urllc applications,” in 2019 IEEE Globecom Workshops, 2019.
[28] M. Tavana, A. Rahmati, V. Shah-Mansouri, and B. Maham, “Cooperative sensing

with joint energy and correlation detection in cognitive radio networks,” IEEE Com-
munications Letters, vol. 21, no. 1, pp. 132–135, 2016.
[29] G. Vallero, D. Renga, M. Meo, and M. A. Marsan, “RAN energy efficiency and fail-
ure rate through ANN traffic predictions processing,” Computer Communications,
vol. 183, pp. 51–63, 2022.
[30] M. Tavana, V. Shah-Mansouri, and V. W. Wong, “Congestion control for bursty

M2M traffic in LTE networks,” in 2015 IEEE International Conference on Commu-
nications (ICC). IEEE, 2015, pp. 5815–5820.
[31] “How to optimize total cost of ownership (TCO) to acceler-

ate 5G network evolution ,” www.samsung.com/global/business/
networks/insights/blog/how-to-optimize-total-cost-of-ownership-tco-to\
-accelerate-5g-network-evolution-part-one/, accessed: 2021-09-10.
[32] X. Wang, L. Wang, S. E. Elayoubi, A. Conte, B. Mukherjee, and C. Cavdar, “Cen-

tralize or distribute? a techno-economic study to design a low-cost cloud radio ac-
cess network,” in 2017 IEEE International Conference on Communications (ICC).
IEEE, 2017.
[33] S. S. Lisi, A. Alabbasi, M. Tornatore, and C. Cavdar, “Cost-effective migration

towards C-RAN with optimal fronthaul design,” in 2017 IEEE International Con-
ference on Communications (ICC), May 2017, pp. 1–7.
76 BIBLIOGRAPHY
[34] A. A. W. Ahmed, J. Markendahl, and C. Cavdar, “Interplay between cost, capacity

and power consumption in heterogeneous mobile networks,” in 2014 21st Interna-
tional Conference on Telecommunications (ICT). IEEE, 2014, pp. 98–102.
[35] A. A. W. Ahmed, K. Chatzimichail, J. Markendahl, and C. Cavdar, “Techno-

economics of green mobile networks considering backhauling,” in European Wire-
less 2014; 20th European Wireless Conference. VDE, 2014, pp. 1–6.
[36] E. Dinc, M. Vondra, S. Hofmann, D. Schupke, M. Prytz, S. Bovelli, M. Frodigh,

J. Zander, and C. Cavdar, “In-flight broadband connectivity: Architectures and busi-
ness models for high capacity air-to-ground communications,” IEEE Communica-
tions Magazine, vol. 55, no. 9, pp. 142–149, 2017.
[37] E. Dinc, M. Vondra, and C. Cavdar, “Total cost of ownership optimization for direct
air-to-ground communication networks,” IEEE Transactions on Vehicular Technol-
ogy, vol. 70, no. 10, pp. 10 157–10 172, 2021.
[38] “Intelligent and agile: capex planning in the 5G era ,” https://assets.ey.com/content/

dam/ey-sites/ey-com/en_gl/topics/tmt/ey-intelligent-and-agile-capex-planning.
pdf, accessed: 2021-09-10.
[39] ETSI, “Environmental engineering (EE); methodology for environmental life cy-
cle assessment (LCA) of information and communication technology (ICT) goods,
networks and services,” ETSI ES 203 199 V1.3.1, 2015.
[40] I. 14040, “Environmental management – life cycle assessment – principles and

framework,” 2006.
[41] E.-E. Commission et al., “International reference life cycle data system (ILCD)
handbook general guide for life cycle assessment detailed guidance,” Institute for
Environment and sustainability, 2010.
[42] “More connectivity and less energy consumption,” https://www.ericsson.com/

en/blog/2018/4/new-research-more-connectivity-and-less-energy-consumption,
accessed: 2018.
[43] “How to estimate carbon emissions in mobile networks: a

streamlined approach,” https://www.ericsson.com/en/blog/2021/5/
how-to-estimate-carbon-emissions-from-mobile-networks, accessed: 2020.
[44] R. Yanggratoke, F. Wuhib, and R. Stadler, “Gossip-based resource allocation for

green computing in large clouds,” in 2011 7th International Conference on Network
and Service Management. IEEE, 2011, pp. 1–9.
[45] C. Cavdar, “Energy-efficient connection provisioning in WDM optical networks,”

in Optical Fiber Communication Conference. Optical Society of America, 2011,
p. OWI3.
BIBLIOGRAPHY 77
[46] F. Yaghoubi, M. Mahloo, L. Wosinska, P. Monti, F. S. Farias, J. C. Costa, and

J. Chen, “Techno-economic and business feasibility analysis of 5G transport net-
works,” Optical and Wireless Convergence for 5G Networks, pp. 273–295, 2019.
[47] A. Conte, A. Feki, L. Chiaraviglio, D. Ciullo, M. Meo, and M. A. Marsan, “Cell

wilting and blossoming for energy efficiency,” IEEE Wireless Communications,
vol. 18, no. 5, pp. 50–57, 2011.
[48] M. A. Marsan, L. Chiaraviglio, D. Ciullo, and M. Meo, “Optimal energy savings in

cellular access networks,” in 2009 IEEE International Conference on Communica-
tions Workshops. IEEE, 2009, pp. 1–5.
[49] S. K. Gowtam Peesapati, M. Olsson, M. Masoudi, S. Andersson, and C. Cavdar, “Q-

learning based radio resource adaptation for improved energy performance of 5G
base stations,” in 2021 IEEE 32nd Annual International Symposium on Personal,
Indoor and Mobile Radio Communications (PIMRC), 2021.
[50] F. Yaghoubi, M. Furdek, A. Rostami, P. Öhlén, and L. Wosinska, “Design and relia-
bility performance of wireless backhaul networks under weather-induced correlated
failures,” IEEE Transactions on Reliability, 2021.
[51] ETSI, “Energy consumption and CO2 footprint of wireless networks,” Report
RRS05-024, 2011.
[52] G. Vallero, M. Deruyck, M. Meo, and W. Joseph, “Base station switching and edge
caching optimisation in high energy-efficiency wireless access network,” Computer
Networks, vol. 192, p. 108100, 2021.
[53] G. Andersson, A. Västberg, A. Devlic, and C. Cavdar, “Energy efficient heteroge-

neous network deployment with cell DTX,” in Proc. IEEE ICC’16. IEEE, 2016,
pp. 1–6.
[54] P. Piunti, C. Cavdar, S. Morosi, K. E. Teka, E. Del Re, and J. Zander, “Energy
efficient adaptive cellular network configuration with QoS guarantee,” in Proc. IEEE
ICC’15. IEEE, 2015, pp. 68–73.
[55] S.-E. Elayoubi, L. Saker, and T. Chahed, “Optimal control for base station sleep
mode in energy efficient radio access networks,” in Proc. IEEE INFOCOM’11.
IEEE, 2011, pp. 106–110.
[56] G. K. Tran, H. Shimodaira, R. E. Rezagah, K. Sakaguchi, and K. Araki, “Practical

evaluation of on-demand small cell ON/OFF based on traffic model for 5G cellular
networks,” in Proc. IEEE WCNC’16. IEEE, 2016.
[57] M. Meo, D. Renga, and Z. Umar, “Advanced sleep modes to comply with delay con-
straints in energy efficient 5G networks,” in 2021 IEEE 93rd Vehicular Technology
Conference (VTC2021-Spring). IEEE, 2021, pp. 1–7.
78 BIBLIOGRAPHY
[58] F. E. Salem, A. Gati, Z. Altman, and T. Chahed, “Advanced sleep modes and their
impact on flow-level performance of 5G networks,” in Proc. IEEE VTC-Fall’17,
Sep. 2017.
[59] F. E. Salem, Z. Altman, A. Gati, T. Chahed, and E. Altman, “Reinforcement learning
approach for advanced sleep modes management in 5G networks,” in Proc. IEEE
VTC-Fall’18, Chicago, USA, Aug. 2017.
[60] F. E. Salem, T. Chahed, Z. Altman, and A. Gati, “Traffic-aware advanced sleep
modes management in 5G networks,” in Proc. IEEE WCNC’19, Marrakech, Mo-
rocco, Apr. 2019.
[61] A. El-Amine, M. Iturralde, H. A. H. Hassan, and L. Nuaymi, “A distributed Q-
Learning approach for adaptive sleep modes in 5G networks,” in Proc. IEEE
PIMRC’19. IEEE, 2019, pp. 1–6.
[62] H. Pervaiz, O. Onireti, A. Mohamed, M. A. Imran, R. Tafazolli, and Q. Ni, “Energy-
efficient and load-proportional eNodeB for 5G user-centric networks: A multilevel
sleep strategy mechanism,” IEEE Vehicular Technology Magazine, vol. 13, no. 4,
pp. 51–59, 2018.
[63] M. Masoudi, M. G. Khafagy, E. Soroush, D. Giacomelli, S. Morosi, and C. Cav-
dar, “Reinforcement learning for Traffic-Adaptive sleep mode management in 5G
networks,” in Proc. IEEE PIMRC’20, London, United Kingdom, Aug. 2020.
[64] I. Chih-Lin, J. Huang, R. Duan, C. Cui, J. X. Jiang, and L. Li, “Recent progress
on C-RAN centralization and cloudification,” IEEE Access, vol. 2, pp. 1030–1039,
2014.
[65] D. Pliatsios, P. Sarigiannidis, S. Goudos, and G. K. Karagiannidis, “Realizing
5G vision through cloud RAN: technologies, challenges, and trends,” EURASIP
Journal on Wireless Communications and Networking, vol. 2018, no. 1, p. 136,
May 2018. [Online]. Available: https://doi.org/10.1186/s13638-018-1142-1
[66] “Functional splits and use cases for small cell virtualization,” Small Cell Forum
release, Tech. Rep., 2016.
[67] S. Matoussi, I. Fajjari, S. Costanzo, N. Aitsaadi, and R. Langar, “A user centric
virtual network function orchestration for agile 5G Cloud-RAN,” in 2018 IEEE In-
ternational Conference on Communications (ICC). IEEE, 2018.
[68] N. Carapellese, M. Tornatore, and A. Pattavina, “Energy-efficient baseband unit
placement in a fixed/mobile converged WDM aggregation network,” IEEE J. Select.
Areas Commun., vol. 32, no. 8, pp. 1542–1551, Aug. 2014.
[69] A. Asensio, P. Saengudomlert, M. Ruiz, and L. Velasco, “Study of the centralization
level of optical network-supported cloud RAN,” in 2016 International Conference
on Optical Network Design and Modeling (ONDM), May 2016.
BIBLIOGRAPHY 79
[70] K. Tanaka and A. Agata, “Next-generation optical access networks for C-RAN,” in
Optical Fiber Communication Conference, Mar. 2015.
[71] M. Peng, C. Wang, V. Lau, and H. V. Poor, “Fronthaul-constrained cloud radio
access networks: insights and challenges,” IEEE Wireless Commun., vol. 22, no. 2,
pp. 152–160, Apr. 2015.
[72] D. C. Chen, T. Q. Quek, and M. Kountouris, “Backhauling in heterogeneous cellu-
lar networks: Modeling and tradeoffs,” IEEE Transactions on Wireless Communi-
cations, vol. 14, no. 6, pp. 3194–3206, 2015.
[73] A. A. Widaa Ahmed, K. Chatzimichail, J. Markendahl, and C. Cavdar, “Techno-
economics of green mobile networks considering backhauling,” in European Wire-
less 2014; 20th European Wireless Conference, May 2014.
[74] M. Mahloo, P. Monti, J. Chen, and L. Wosinska, “Cost modeling of backhaul for
mobile networks,” in Communications Workshops (ICC), 2014 IEEE International
Conference on. IEEE, 2014, pp. 397–402.
[75] C. M. Machuca, M. Kind, K. Wang, K. Casier, M. Mahloo, and J. Chen, “Methodol-
ogy for a cost evaluation of migration toward NGOA networks,” IEEE/OSA Journal
of Optical Communications and Networking, vol. 5, no. 12, pp. 1456–1466, 2013.
[76] A. Peralta, E. Inga, and R. Hincapié, “Optimal scalability of FiWi networks based
on multistage stochastic programming and policies,” Journal of Optical Communi-
cations and Networking, vol. 9, no. 12, pp. 1172–1183, 2017.
[77] K. Wang, C. Masmachuca, L. Wosinska, P. Urban, A. Gavler, K. Brunnstrom, and
J. Chen, “Techno-economic analysis of active optical network migration toward
next-generation optical access,” IEEE/OSA Journal of Optical Communications and
Networking, vol. 9, no. 4, pp. 327–341, 2017.
[78] F. Yaghoubi, M. Mahloo, L. Wosinska, P. Monti, F. de Souza Farias, J. C. W. A.
Costa, and J. Chen, “A techno-economic framework for 5G transport networks,”
IEEE wireless communications, vol. 25, no. 5, pp. 56–63, 2018.
[79] H. Frank, R. Tessinari, Y. Zhang, Z. Gao, C. Meixner, S. Yan, and D. Simeonidou,
“Resource analysis and cost modeling for end-to-end 5G mobile networks,” in Op-
tical Network Design and Modelling (ONDM), 10 2019.
[80] C. Ranaweera, P. Monti, B. Skubic, E. Wong, M. Furdek, L. Wosinska, C. M.
Machuca, A. Nirmalathas, and C. Lim, “Optical transport network design for 5G
fixed wireless access,” Journal of Lightwave Technology, vol. 37, no. 16, pp. 3893–
3901, Aug 2019.
[81] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborative mobile edge
computing in 5G networks: New paradigms, scenarios, and challenges,” IEEE Com-
munications Magazine, vol. 55, no. 4, pp. 54–61, 2017.
80 BIBLIOGRAPHY
[82] T. X. Tran and D. Pompili, “Octopus: A cooperative hierarchical caching strategy

for cloud radio access networks,” in 2016 IEEE 13th International Conference on
Mobile Ad Hoc and Sensor Systems (MASS). IEEE, 2016, pp. 154–162.
[83] J. Kwak, Y. Kim, L. B. Le, and S. Chong, “Hybrid content caching in 5G wireless
networks: Cloud versus edge caching,” IEEE Transactions on Wireless Communi-
cations, vol. 17, no. 5, pp. 3030–3045, 2018.
[84] X. Huang, Z. Zhao, and H. Zhang, “Latency analysis of cooperative caching with
multicast for 5G wireless networks,” in 2016 IEEE/ACM 9th International Confer-
ence on Utility and Cloud Computing (UCC), 2016.
[85] A. Alabbasi and C. Cavdar, “Delay-aware green hybrid CRAN,” in Modeling and
Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), 2017. IEEE,
2017, pp. 1–7.
[86] 3GPP, “Study on Radio Access Network(RAN)Sharing enhancements,” 3rd

Generation Partnership Project (3GPP), Technical Specification (TS) 22.852, 09
2014, version 13.1.0. [Online]. Available: https://portal.3gpp.org/desktopmodules/
Specifications/SpecificationDetails.aspx?specificationId=668
[87] I. Afolabi, T. Taleb, K. Samdanis, A. Ksentini, and H. Flinck, “Network slicing and
softwarization: A survey on principles, enabling technologies, and solutions,” IEEE
Communications Surveys & Tutorials, vol. 20, no. 3, pp. 2429–2453, 2018.
[88] K. Samdanis, X. Costa-Perez, and V. Sciancalepore, “From network sharing to

multi-tenancy: The 5G network slice broker,” IEEE Communications Magazine,
vol. 54, no. 7, pp. 32–39, 2016.
[89] J. Ordonez-Lucena, P. Ameigeiras, D. Lopez, J. J. Ramos-Munoz, J. Lorca, and

J. Folgueira, “Network slicing for 5G with SDN/NFV: Concepts, architectures, and
challenges,” IEEE Communications Magazine, vol. 55, no. 5, pp. 80–87, 2017.
[90] N. Alliance, “Description of network slicing concept,” NGMN 5G P, vol. 1, no. 1,

2016.
[91] X. Foukas, G. Patounas, A. Elmokashfi, and M. K. Marina, “Network slicing in

5G: Survey and challenges,” IEEE Communications Magazine, vol. 55, no. 5, pp.
94–100, 2017.
[92] R. Pasquini and R. Stadler, “Learning end-to-end application QoS from openflow
switch statistics,” in 2017 IEEE Conference on Network Softwarization (NetSoft).
IEEE, 2017, pp. 1–9.
[93] M. Chahbar, G. Diaz, A. Dandoush, C. Cérin, and K. Ghoumid, “A comprehensive

survey on the E2E 5G network slicing model,” IEEE Transactions on Network and
Service Management, vol. 18, no. 1, pp. 49–62, 2020.
BIBLIOGRAPHY 81
[94] X. Li, R. Ni, J. Chen, Y. Lyu, Z. Rong, and R. Du, “End-to-end network slicing in
radio access network, transport network and core network domains,” IEEE Access,
vol. 8, pp. 29 525–29 537, 2020.
[95] T. Li, X. Zhu, and X. Liu, “An end-to-end network slicing algorithm based on deep
Q-learning for 5G network,” IEEE Access, vol. 8, pp. 122 229–122 240, 2020.
[96] H. Chergui and C. Verikoukis, “Offline SLA-constrained deep learning for 5G net-
works reliable and dynamic end-to-end slicing,” IEEE Journal on Selected Areas in
Communications, vol. 38, no. 2, pp. 350–360, 2020.
[97] H.-T. Chien, Y.-D. Lin, C.-L. Lai, and C.-T. Wang, “End-to-end slicing with opti-
mized communication and computing resource allocation in multi-tenant 5G sys-
tems,” IEEE Transactions on Vehicular Technology, vol. 69, no. 2, pp. 2079–2091,
2019.
[98] S. E. Elayoubi, S. B. Jemaa, Z. Altman, and A. Galindo-Serrano, “5G RAN slicing
for verticals: Enablers and challenges,” IEEE Communications Magazine, vol. 57,
no. 1, pp. 28–34, 2019.
[99] X. Shen, J. Gao, W. Wu, K. Lyu, M. Li, W. Zhuang, X. Li, and J. Rao, “AI-assisted
network-slicing based next-generation wireless networks,” IEEE Open Journal of
Vehicular Technology, vol. 1, pp. 45–66, 2020.
[100] L. Feng, Y. Zi, W. Li, F. Zhou, P. Yu, and M. Kadoch, “Dynamic resource allocation
with RAN slicing and scheduling for uRLLC and eMBB hybrid services,” IEEE
Access, vol. 8, pp. 34 538–34 551, 2020.
[101] H. Xiang, S. Yan, and M. Peng, “A realization of fog-RAN slicing via deep re-
inforcement learning,” IEEE Transactions on Wireless Communications, vol. 19,
no. 4, pp. 2515–2527, 2020.
[102] Y. Sun, S. Qin, G. Feng, L. Zhang, and M. Imran, “Service provisioning framework
for RAN slicing: user admissibility, slice association and bandwidth allocation,”
IEEE Transactions on Mobile Computing, 2020.
[103] Y. Cui, X. Huang, P. He, D. Wu, and R. Wang, “A two-timescale resource alloca-
tion scheme in vehicular network slicing,” in 2021 IEEE 93rd Vehicular Technology
Conference (VTC2021-Spring). IEEE, 2021, pp. 1–5.
[104] H. Zhang and V. W. Wong, “A two-timescale approach for network slicing in QoS,”
IEEE Transactions on Vehicular Technology, vol. 69, no. 6, pp. 6656–6669, 2020.
[105] C. Cox, An Introduction to LTE: LTE, LTE-Advanced, SAE, VoLTE and 4G Mobile
Communications, 2nd ed. Wiley Publishing, 2014.
[106] X. Guo, Z. Niu, S. Zhou, and P. Kumar, “Delay-constrained energy-optimal base sta-
tion sleeping control,” IEEE Journal on Selected Areas in Communications, vol. 34,
no. 5, pp. 1073–1085, 2016.
82 BIBLIOGRAPHY
[107] J. Liu, B. Krishnamachari, S. Zhou, and Z. Niu, “DeepNap: Data-driven base station
sleeping operations through deep reinforcement learning,” IEEE Internet of Things
Journal, vol. 5, no. 6, pp. 4273–4282, 2018.
[108] J. E. Beyer and B. F. Nielsen, “Predator foraging in patchy environments: the inter-
rupted poisson process (IPP) model unit,” DANA-CHARLOTTENLUND-, vol. 11,
1996.
[109] A. McCallum, “Hidden markov models baum welch algorithm,” Introduction to

Natural Language Processing CS585; University of Massachusetts Amherst: Mas-
sachusetts, USA, 2004.
[110] G. Mountaser, M. L. Rosas, T. Mahmoodi, and M. Dohler, “On the feasibility of

MAC and PHY split in cloud RAN,” in 2017 IEEE Wireless Communications and
Networking Conference (WCNC), March 2017, pp. 1–6.
[111] L. M. Larsen, A. Checko, and H. L. Christiansen, “A survey of the functional splits

proposed for 5G mobile crosshaul networks,” IEEE Communications Surveys & Tu-
torials, vol. 21, 2018.
[112] “CPRI vs eCPRI: What Are Their Differences and Meanings to 5G?” https:
//community.fs.com/blog/cpri-vs-ecpri-differences-and-meanings-to-5g.html, ac-
cessed: 2020-04-04.
[113] H. Lehpamer, Introduction to power utility communications. Artech House, 2016.
[114] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof power model
for cellular base stations,” in Vehicular Technology Conference (VTC Spring), 2015
IEEE 81st. IEEE, 2015, pp. 1–7.
[115] China Mobile Research Institute, “C-RAN the road towards green RAN,” White
Paper, October 2011.
[116] Ericsson AB, Huawei Technologies Co. Ltd, NEC Corporation, Alcatel Lucent, and
Nokia Networks, “Common public radio interface (CPRI); interface specification
v7.0,” Oct. 2015.
[117] D. Liu and C. Yang, “Energy efficiency of downlink networks with caching at base
stations,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 4, pp.
907–922, April 2016.
[118] X. Wang, A. Alabbasi, and C. Cavdar, “Interplay of energy and bandwidth consump-
tion in CRAN with optimal function split,” in 2017 IEEE International Conference
on Communications (ICC), May 2017, pp. 1–6.
[119] N. Alliance, “NGMN 5G P1 requirements & architecture work stream end-to-end

architecture description of network slicing concept,” 2016.
BIBLIOGRAPHY 83
[120] B. Henrik, L. Jan, C. Angelo, and a. Thomas, “Applied network slicing scenarios in
5G,” Ericsson Technology Review, 2021.
[121] R. Su, D. Zhang, R. Venkatesan, Z. Gong, C. Li, F. Ding, F. Jiang, and Z. Zhu,
“Resource allocation for network slicing in 5G telecommunication networks: A
survey of principles and models,” IEEE Network, vol. 33, no. 6, pp. 172–179, 2019.
[122] “Common public radio interface: eCPRI interface specification,” Standard, 2019.
[Online]. Available: http://www.cpri.info/downloads/eCPRI_v_2.0_2019_05_10c.
pdf
[123] S. Abdelwahab, B. Hamdaoui, M. Guizani, and T. Znati, “Network function virtu-
alization in 5G,” IEEE Communications Magazine, vol. 54, no. 4, pp. 84–91, 2016.
[124] M. Yan, G. Feng, J. Zhou, Y. Sun, and Y.-C. Liang, “Intelligent resource scheduling
for 5G radio access network slicing,” IEEE Transactions on Vehicular Technology,
vol. 68, no. 8, pp. 7691–7703, 2019.
[125] A. Alabbasi, M. Ganjalizadeh, K. Vandikas, and M. Petrova, “On cascaded federated
learning for multi-tier predictive models,” in 2021 IEEE International Conference
on Communications Workshops (ICC Workshops), 2021, pp. 1–7.

Data Driven AI Assisted Green Network Design and Management FULLTEXT01

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Data Driven AI Assisted Green Network Design and Management FULLTEXT01

Uploaded by

Copyright:

Available Formats

kth royal institute

Doctoral Thesis in Information and Communication Technology

Data Driven AI Assisted Green

Stockholm, Sweden 2022

Doctoral Thesis in Information and Communication Technology

Printed by: Universitetsservice US-AB, Sweden 2022

The energy consumption of mobile networks is increasing due to an increase in

Energiförbrukningen i mobilnäten ökar ständigt i takt med ökade trafikvolymer

List of Acronyms xiii

2 AI Assisted Green Mobile Networks 25

2.4 Risk Aware Sleep Mode Management . . . . . . . . . . . . . . . . . . . 29

3 AI Assisted Network Architecture Design and Management 41

4 Conclusions and Contributions 67

1.1 Ericsson traffic forecast by 2026 [1]. . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Data traffic model as interrupted Poisson process [6]. . . . . . . . . . . . . . 29

3.1 Basic C-RAN architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.11 Network architecture and slice manager [8]. . . . . . . . . . . . . . . . . . . 61

1 List of frequently used acronyms and definitions . . . . . . . . . . . . . . . . xiv

1.1 Fronthaul capacity requirement for different split points . . . . . . . . . . . 14

2.1 Power and minimum durations from [9, 10] . . . . . . . . . . . . . . . . . . 26

3.1 List of parameters and their definitions. . . . . . . . . . . . . . . . . . . . . 45

Table 1: List of frequently used acronyms and definitions

1.1 Background and Motivation

Figure 1.1: Ericsson traffic forecast by 2026 [1].

1.2 Challenges of Mobile Networks

1.2.1 Surge in Traffic Data and Subscribers

Figure 1.2: Ericsson mobile user subscriptions forecast by 2026 [1].

1.2.2 High Energy Consumption

Figure 1.3: Three main 5G services [2].

1.2.3 High Total Cost of Ownership

1.2.4 Smart Network Management

1.2.5 Carbon Footprint and Life Cycle Assessment

Figure 1.4: Performance requirements comparison between 4G and 5G [3].

The cradle-to-grave approach to embodied energy consumption of the complete pro-

1.2.5.2 Carbon Footprints

1.2.5.3 Thesis Contribution to LCA and Carbon Foot Print

1.3 Literature Study

1.3.1 Green Networks

1.3.2 Network Architecture Design

Figure 1.6: Evolution of RAN (DRAN/C-RAN/H-CRAN) [4].

of capacity limitation in the fronthaul considering the delay constraints.

1.3.2.1 Functional Split

Figure 1.7: Functional split model [5].

1.3.2.2 Network Architectures and Migration Costs

Capacity requirements (Mbps)

1.3.2.3 Content Placement

1.3.2.4 Network Slicing

1.4 Research Questions and Methodologies

In order to address the fronthaul capacity limitations in C-RAN, hybrid C-RAN is

1.4.2 Research Methodologies

Figure 1.8: The break down of thesis topics

1.5 Thesis Contribution

• Energy Saving at the Base Stations

– Paper 1: M. Masoudi, et al, "Green Mobile Networks for 5G and Beyond,"

– Patent 1: M. Masoudi, et al, "NETWORK NODE AND METHOD PER-

• Cost-Effective Migration towards C-RAN with Optimal Fronthaul Design

– Paper 4: M. Masoudi, S.S. Lisi, and C. Cavdar, "Cost-effective migration to-

• Content Placement and Functional Split Optimization in Virtual C-RAN:

– Paper 5: A. Siram, M. Masoudi, A. Alabbasi, and C. Cavdar, "Joint optimiza-

• Joint Communication and Computation Resource Allocation for End-to-End

– Paper 6: M. Masoudi, et al, "Energy-Optimal End-to-End Network Slicing in

1.6 Thesis Organization

AI Assisted Green Mobile Networks

2.1 BS Sleep Mode Management