Professional Documents
Culture Documents
Peer-to-Peer-Based
Grid Support Functions
for Smart Inverters
Distributed Voltage and Frequency Control
P2P
Hardware
Battery Energy
Storage System
P2P-Based Smart Inverter
Hamada Almasalma
August 2020
Peer-to-Peer-Based
Grid Support Functions
for Smart Inverters
Distributed Voltage and Frequency Control
Hamada ALMASALMA
August 2020
© 2020 KU Leuven – Faculty of Engineering Science
Uitgegeven in eigen beheer, Hamada Almasalma, Kasteelpark Arenberg 10, box 2445, B-3001 Leuven (Belgium)
Alle rechten voorbehouden. Niets uit deze uitgave mag worden vermenigvuldigd en/of openbaar gemaakt worden
door middel van druk, fotokopie, microfilm, elektronisch of op welke andere wijze ook zonder voorafgaande
schriftelijke toestemming van de uitgever.
All rights reserved. No part of the publication may be reproduced in any form by print, photoprint, microfilm,
electronic or any other means without written permission from the publisher.
“I cannot do all the good that the
world needs. But the world needs
all the good that I can do.”
Jana Stanfield
Preface
This work would have never succeeded without the dedicated support of several
people, whom I would like to thank personally.
First and foremost, I would like to sincerely thank my supervisor Prof. Geert
Deconinck for giving me the opportunity to carry out a doctoral research in
the field of peer-to-peer control and distributed optimization. During the past
four years and ten months, I had the valuable fortune of being able to enjoy his
guidance. He took the time to discuss and read, had great ideas, made numerous
useful comments, showed an open-mindedness and gave me the freedom and
confidence to delve into different areas that contributed to this PhD.
Secondly, I would like to express my gratitude to the Fonds voor Wetenschap-
pelijk Onderzoek (FWO), who believed in the successful outcome of my work
and decided back then to grant me a PhD scholarship. Thank you for the
support.
I would like to thank all the members of the examination committee, for sharing
their very useful thoughts and insights. Prof. Berlamont, thank you for chairing
the committee. Thank you Prof. Driesen and Prof. Suykens for being part
of my supervisory committee throughout the years. Thank you Prof. Caire
and Prof. Wu for being part of the jury and for the useful remarks that have
definitely made this thesis more complete.
I am also very grateful to the lab technicians Johan, Luc and Gert for assisting
me in the experimental works. I also would like to give special thanks to sweet
Electa staff: Katleen, Katja, Veerle and Veronica. They have eased my life by
fixing the administrative, financial and ICT problems.
Of course, I would like to thank all my colleagues and ex-colleagues at Electa
for making my PhD journey the most memorable one. I owe special thanks to
Jonas Engels who shared his experience in distributed and robust optimization
with me; all this work would not have carried out without his help. I am
i
ii PREFACE
also thankful for the good cooperation with Sander Claeys. Thanks for his
valuable contribution of the implementation and validation of the distributed
optimization-based voltage control algorithms. I also would like to thank Mudar
Abedrabbo for helping me with the implementation of the inverter control loops
and for his friendship.
I would like to thank all the members of the P2P-SmarTest project for their
collaboration, the constructive meetings and for sharing their insights and ideas.
I especially want to thank Ari Pouttu, Konstantin Mikhaylov and Jussi Haapola
(from Centre for Wireless Communications, University of Oulu, Finland) who
provided me with the device-to-device communication modules and the materials
of Section 4.5.
Finally, there are the people who were always there for me, who have supported
me before, during and will after. For them, I will finish in Arabic.
H@
ð QºË@ P AJ.« ùÖÞ AK. ÐY® K @ , éËAQË@
è YêË ùKAîE@ úΫ èQºð úÍAªK é<Ë@ YÔg YªK. ð
úÍA« É¿ úÍ AÓY¯ð ,QªË@ áÓ ÕÎªË @ Ik ú¯ AQ« áK YÊË@ áK QK QªË@ ø YË@ð úÍ@ QK Y® JË@
.
Èñ¢. AÒêË ZA«YË@ B@ ½ÊÓ @ C¯ , à B@ éJË@ IÊð AÒJ¯ é<Ë@ YªK. É®Ë@ AÒêË àA¿ð ,®Kð
áÓ úæJ¢« @ AÖÏ , àAJK úæk
AëYêk. ð AîD¯ð . ÐY® K @ AÒ» .ÉÒªË@ ákð QÒªË@
. . ð P úÍ@ QºËAK
. áªÓ ð YJ Qg úÍ I KA¿ ð éJJÒJË@ Aêm' A ð
úÍ ÐY¯ ð @ , A¯ðQªÓ úÍ øY @ ð @ , àñªË@ YK úÍ YÓ áÓ É¾Ë ð
ú¯ ÑîD @ ð @ , éjJ
.QK Y® JË@ ð QºË@
ËAg úæÓ éʯ ÉÒªË@ @ Yë PAm.' @
Hamada Almasalma,
August 2020.
Abstract
iii
iv ABSTRACT
v
vi BEKNOPTE SAMENVATTING
D2D Device-to-Device.
DC Direct Current.
DD Dual Decomposition.
DERs Distributed Energy Resources.
DisCOP Distributed Constraint Optimization.
DisCVC Distributed Coordinated Voltage Control.
DoD Depth of Discharge.
Dof Degrees of freedom.
DSO Distribution System Operator.
DVC Distributed Voltage Control.
vii
viii List of Abbreviations
KKT Karush-Kuhn–Tucker.
KPI Key Performance Indicator.
L Inductor.
LoRaWAN Long Range Wide Area Network.
LPWA Low Power Wide Area.
R Resistor.
R.Pi Raspberry Pi.
RTT Real Time Target Computer.
Notation
k · k1 First norm.
k · k2 Second norm.
Tr[·] Trace.
xi
xii NOMENCLATURE
General
v Voltage (V).
E Energy (kWh).
L Lagrangian Function.
q
γinv i
Reactive acceleration factor of inverter i.
(k)
e R-UP, bati Maximum energy capacity of the R-UP battery i (kWh).
(k)
e R-UP, bati Minimum energy capacity of the R-UP battery i (kWh).
(k)
βpi Coefficient of the reactive power policy i, related to the
control nodes’ active power consumption and PV power
(Chapter 6).
∆f (k)
R-UP
Minimum normalized frequency deviation.
(k)
∆P lx Maximum deviation from forecasted active power
consumption of household x (kW).
∆P (k)
pvi Maximum deviation from forecasted active power of PV
installation i (kW).
Counts
Sets
U Uncertainty set.
Y Observers set.
Contents
Abstract iii
Beknopte samenvatting v
List of Abbreviations ix
Nomenclature xi
Contents xix
1 Introduction 1
1.1 Context and Motivation . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives and Scope . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Outline and Contributions . . . . . . . . . . . . . . . . . . . . . 8
xix
xx CONTENTS
Bibliography 167
xxv
xxvi LIST OF FIGURES
6.1 The results of this figure are obtained based on the implemen-
tation of the policy-based DisCVC proposed in Chapter 5, in a
simulation over 104 scenarios. The scenarios are presented in
Subsection 6.4.2. The DisCVC is used to regulate voltage profiles
of the 62-node network presented in Subsection 6.4.1. This figure
presents the charging power of BESS number 31. . . . . . . . . 119
6.2 Energy content of the 13.5 kWh battery over 730 frequency
scenarios. Battery parameters: maximum power= 7 √kW,
initial charge= 6.5 kWh, charge/discharge efficiency= 0.9.
The energy content E is calculated based on: E (k+1) =
(k) + (k) +
E + ∆t ηi pFCR − ηdis
(k) ch 1
−pFCR , where ηich and ηidis are,
i
respectively, the charge and discharge efficiencies , ∆t is the time
+
step duration and [·] ≡ max(·, 0). . . . . . . . . . . . . . . . . . 121
6.3 General overview of the proposed control framework. . . . . . . 124
6.4 (a) Unregulated voltage profile of a prosumer in a network with
high PV penetration, (b) Expected upward and downward reserve
capacity profile. In this example, the inverter reactive power and
the VC battery power are used to solve the voltage rise problem,
therefore, the upward reserve capacity is zero in Area 2. . . . . 126
6.5 Normalized positive and negative frequency deviation ∆f of
730 days of historical frequency data (2017-2018). The dotted
lines show the maximum and minimum normalized frequency
deviations. ∆f is calculated based on (6.1b). The frequency data
are from the CE synchronous region. . . . . . . . . . . . . . . . 129
6.6 Illustration of energy and power capacity allocation of battery i. 136
6.7 Schematic diagram of the network used in this case study. . . . 149
6.8 Aggregated power capacities of the downward-FCR, upward-FCR,
voltage rise, energy management policies of R-DN and R-UP
batteries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.9 Power capacity reserved for charging the VC battery 31. . . . . 152
LIST OF FIGURES xxix
6.10 (a) Voltage profiles of node 31 with no control, (b) Voltage profiles
of node 31 with voltage and frequency control. The dotted lines
show the maximum and minimum voltage limits . . . . . . . . 153
6.11 (a) Apparent power magnitude of inverter 31, (b) Energy content
of battery 31. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
xxxi
Chapter 1
Introduction
Over the last decade, the worldwide deployment of photovoltaic (PV) systems
has been increasing exponentially due to the cost reductions in renewables,
advances in digital technologies, the supporting governments’ policies and
the increased customer awareness. According to the IEA’s1 2019’s Trends in
Photovoltaic Applications report [1], at the end of 2018 the globally cumulative
PV capacity reached over half a TW. It is also predicted that the global PV
capacity will continue to grow by 9% every year in the next 50 years [2],
according to the 2019 study released by the International Renewable Energy
Agency (IRENA). This rapid growth of PV penetration in power system helps
the nation to fulfill their electrical energy needs without increasing fossil fuel
consumption, which largely contributes to reducing greenhouse gas emissions
and mitigating climate change effects.
Although PV systems are a key technology for implementing the shift to a
decarbonized energy supply, the current distribution networks were not designed
to incorporate this type of generation. Consequently, grid operators are facing
several problems in controlling distribution networks. One of the most common
problems is voltage deviation from acceptable range defined by the current
standards (e.g. EN 50160 [3]). For example, high PV generation during low
demand might reverse power flows in the grid and create voltage rise issues,
while sudden drop in PV generation during peak demand period might create
voltage drop issues. Furthermore, intermittent and unpredictable nature of PV
1 International Energy Agency
1
2 INTRODUCTION
BESSs [37], the authors are interested in combining distributed voltage control
with frequency control.
This thesis aims to tackle the aforementioned challenges with respect to over-
voltage and under-voltage problems of distribution networks with high PV
penetration. Owning to the drawbacks of centralized and decentralized voltage
control systems, the thesis focuses on distributed voltage control systems. The
main objective of this thesis is twofold. The first is to develop P2P-based GVSFs
for smart PV inverters and PV-battery inverters2 . The aim is to enable smart
inverters to communicate with each other in a P2P fashion and regulate grid
voltages cooperatively; without relying on a central decision-making controller.
The second is to develop frequency support functions (FSFs) for PV-battery
inverters, and combine them with the GVSFs. The aim is to enable PV-battery
systems to provide simultaneously frequency containment reserve service (to
their synchronous area) and DVC service (to their distribution network). There
are five underlying objectives to reach the main objective, as discussed in the
following paragraphs.
The first goal of the thesis is to help the readers understand the linkage of the
thesis work to the big picture of P2P-based concept in distribution networks.
To this end, this thesis will first look at different architectures of distribution
networks control. It will then present a general P2P-based control architecture.
The research questions addressed in this objective are:
outputs of both PV panels and batteries; charging batteries with either PV panels or the grid.
6 INTRODUCTION
The third goal of the thesis is to illustrate the practical implementation of DVC
systems. To this end, this thesis will develop a laboratory-based microgrid
equipped with P2P-based PV inverters. The research question addressed in
this objective is:
The fourth goal of the thesis is to reduce the complexity and communication
burden of DVC systems. To this end, this thesis will develop a DVC system
enabled by linear policies. The idea is to let smart inverters solve, each time step
in real-time, a set of linear equations rather than solving a complex optimization
problem. The research question addressed in this objective is:
P2P-SmarTest Project
This research fits within the wider scope of the EU funded Peer-to-Peer Smart
Energy Distribution Networks (P2P-SmarTest) project3 , in which the author
actively contributed. The overall objective of this project was to employ P2P
approaches to ensure the integration of demand side flexibility and the optimum
operation of distributed energy resources, while maintaining power balance
and the quality and security of the supply. The focus was to develop and
experimentally validate intelligent P2P-based control, communication and
trading algorithms. The contributions of the KU Leuven, and specifically
the ELECTA research group, in this project were:
Main Contributions
Outline
Chapter 2
Towards a new control paradigm for distribution networks
Chapter 3 Chapter 5
Real-time
Optimization-based Offline robust Policy-based
distributed
distributed voltage optimization distributed voltage
optimization
control control
Development of a new
methodology to simplify the
Chapter 6
Chapter 4 implementation of distributed
Simultaneous provision
Laboratory voltage control systems
of voltage and
implementation frequency control
Chapter 7
Conclusions and future work
The contribution of this chapter is four-fold. First, the chapter modifies the state-
of-the-art DD algorithm to implement a fully distributed DD method. It then
applies the fully distributed DD method to design a DVC system. Second, the
chapter speeds up the convergence of the DVC system significantly by applying
the Jacobi-Proximal ADMM method (JP-ADMM). Third, three strategies are
proposed in this chapter to improve the performance of the proposed algorithms:
feedback strategy; anti-windup strategy; and active/standby modes. Fourth,
the chapter integrates a push-sum gossip protocol with the DD-based and JP-
10 INTRODUCTION
This chapter proposes a DVC system enabled by robust linear policies. The
proposed voltage control consists of an offline coordinator and real-time DVC.
In the offline coordinator, the voltage control problem is formulated as a
chance-constrained stochastic optimization problem, where the control actions
are expressed as linear function of the uncertain households’ active power
consumption and PV power. In real-time, the inverters communicate with each
other and apply the robust policies to regulate voltage profiles within limits.
Simulations over 36500 scenarios are used to demonstrate the robustness of the
proposed voltage control system. The work of this chapter has been published
in IEEE ACCESS, vol. 8, pp. 124939-124948, 2020 [42].
This chapter concludes the work and presents recommendations for future
research.
This research was funded by the P2P-SmarTest project from its beginning in
November 2015 until the end of 2016. After that, the research was awarded with a
PhD fellowship by the Research Foundation – Flanders (Fonds Wetenschappelijk
Onderzoek -Vlaanderen, FWO). The work of Chapter 2 and the development of
the DD-based DVC algorithm in Chapter 3 were funded by the P2P-SmarTest
project. The development of the JP-ADMM-based DVC algorithm in Chapter
3 and the work of Chapters 5 and 6 were funded by FWO. The work of Chapter
4 was funded by both the P2P-SmarTest project and FWO.
Chapter 2
Peer-to-Peer Control of
Distribution Networks
2.1 Introduction
13
14 PEER-TO-PEER CONTROL OF DISTRIBUTION NETWORKS
installed DERs [44]. This active monitoring and control of the distribution grid
is commonly referred to as an essential part of the smart grid, which is regarded
to be key in the future integration of electricity consumers, generators and those
that do both (prosumers).
Current distribution networks are not designed to accommodate a large amount
of DERs. A large penetration of DERs may create problems to maintain the
quality of supply to all customers connected to the distribution network. Besides,
the intermittent nature of DERs can create issues with the second-by-second
balance of demand and supply. However, by coordinating the DERs, these
issues could be resolved without the need for additional investments in grid
infrastructure [45].
In literature [46, 47, 48], the idea of a microgrid is an often mentioned alternative
to controlling the whole distribution grid with a large amount of DERs. The
main idea is that, when there are many DERs in a wide network, it can be very
complex and difficult to control. Thus, a potential way to manage this complexity
is by breaking down the entire grid into smaller microgrids, containing only
a limited amount of DERs. This chapter elaborates this idea and proposes
an operational control paradigm for the future distribution grid, based on the
concept of microgrids.
When considering such a microgrid and the coordination of multiple microgrids,
different control methods can be found in literature [49, 50, 51]. This chapter
proposes a taxonomy of these control methods, from fully centralized to
completely decentralized. A fully distributed control method, the P2P control
architecture, is further elaborated in this chapter, as it is a promising way for
future control of the distribution grid. Since the DERs are typically highly
distributed, operated by many different owners and with different objectives, it
is desirable that the microgrid control system operates in a highly distributed
way as well. Besides, a robust control system is needed that does not depend
on a single point of failure, as most of the more centralized control methods do.
The rest of the chapter is organized as follows: Section 2.2 introduces the most
prominent issues with regard to the integration of DERs in the electrical grid,
together with an elaboration of the microgrid concept that would be able to
overcome these issues. Section 2.3 proposes a categorization for the different
architectures of microgrid control. Section 2.4 then proposes a new control
paradigm for the distribution grid, based on the microgrid concept and the P2P
architecture. Finally, the chapter is concluded in Section 2.5.
ISSUES WITH THE INTEGRATION OF DERS IN THE ELECTRICAL GRID 15
This section summarizes the most prominent issues with regard to the integration
of DERs in the current electrical grid. Both voltage issues and frequency or
stability issues are discussed. Other type of issues, such as harmonics, security
issues and power fluctuations are not discussed here. The microgrid concept is
presented as a possible solution to overcome these issues.
the control actions for all units [68]. The advantage of a centralized control
system is that the central system receives all necessary data of the microgrid,
and based on all available information the multi-objective controller can achieve
globally optimal performance. As there is only one controller, this results in a
high controllability of the system. However, this high performance comes at
a cost. First of all, the computational burden is heavy, as the optimization is
computed based on a large amount of information. Moreover, a centralized
controller is a single point of failure and redundancy of the central controller is
expensive. The loss of communication with the central controller may cause
a shutdown of the overall system. Besides, as all system states and boundary
conditions have to be known at the central point, this requires a high quality
of communication from all DERs to the central point of control. There is also
the concern that the owners of the different DERs are not willing to hand over
control of their resources to a third party. Finally, central systems are usually
regarded as not being very scalable and system maintenance requires complete
shutdown. To overcome these issues, more distributed control architectures are
developed, as described in the subsections below.
Centralized controller
Optimization
Prosumer Prosumer
Prosumer Prosumer
Prosumer
Prosumer Prosumer
Centralized controller
: Optimization
Aggregator Aggregator
Prosumer Prosumer
Prosumer Prosumer
Prosumer
Prosumer Prosumer
by letting each agent responsible for computing the Lagrangian multipliers associated with
the local PCC voltage.
ARCHITECTURES OF MICROGRID CONTROL 21
Coordinator
: Optimization
Prosumer Prosumer
Prosumer Prosumer
Prosumer Prosumer
: Optimization
Prosumer
Prosumer
Prosumer
Prosumer
Prosumer
Finally, there also exist control architectures without any form of communication.
This chapter classifies them as local control architectures. In this case, optimal
operation of the microgrid is rather difficult, as it is impossible to know the
complete state of the grid and all operational boundary conditions of the DERs.
Nevertheless, the absence of any communication in this architecture will ensure
that the grid is still controlled when all communication channels fail. As primary
frequency control should be able to operate even when communication fails,
this is often implemented as a local control architecture [85]. Additionally, local
PROPOSED P2P-BASED CONTROL PARADIGM 23
High voltage
Load
G Distributed
generator
P2P communication
G Microgrid
G
G G
G
Low voltage
G
G G
G
G G
G
agents obtain new information about the new state of the grid through the
P2P communication, and decide on a new action based on this new state. The
successive grid states should converge to a desired steady state of the grid.
Appropriate control algorithms are needed to achieve this.
Epidemic Algorithms
Epidemic or gossiping algorithms are used for scalable and efficient data
dissemination in distributed P2P networks, without a central controller [89].
They mimic the spread of a contagious disease. Each agent in the distributed
system sends new information it has received to other agents rather than to
a server or cluster of servers in charge of forwarding it. In turn, each of these
agents forwards the information to other selected agents, and so on. Gossiping is
often used as a synonym for epidemic algorithms [90]. Gossiping protocols have
been used in many publications to enable fully distributed control systems. In
PROPOSED P2P-BASED CONTROL PARADIGM 25
The proposed P2P-based architecture can be used to enable P2P energy trading.
P2P energy trading concept provides local and regional energy producers with
options to trade energy fairly within the neighbourhood, within the community
and within the vicinity of the distribution system. This will fundamentally
change the current peer-to-grid paradigm where any surplus of local production
can only be sold to transmission grids, and transform consumers position from
energy/price takers to energy/price makers.
P2P energy trading has been discussed widely in literature. A comprehensive
review of the state-of-the-art in research on P2P energy trading techniques is
presented in [93]. A four-layer system architecture of P2P energy trading is
proposed in [94] to identify and categorize the key elements and technologies
involved in P2P energy trading. In [95], a motivational psychology framework is
introduced for P2P energy trading with an objective to improve the participation
of prosumers. The study proposed in [96] investigates the optimal sizing problem
of battery energy storage systems in a P2P energy sharing network considering
different ownership structures. An optimization model is proposed in [97] to
maximize the economic benefits for rooftop PV-battery system in a P2P energy
trading environment. In [98], a methodology is proposed based on sensitivity
analysis to assess the impact of P2P transactions on the grid and to guarantee
an exchange of energy that does not violate grid constraints.
2.5 Conclusion
This chapter discusses the major issues with regard to the integration of DERs
in the electrical grid and the concept of microgrids as a possible solution
to overcome the complexity of controlling distribution networks. As there
are different methods for organizing the control of these microgrids found in
literature, this chapter presents a classification of these methods, from highly
centralized to fully decentralized architectures. The drawbacks of the centralized
control and the advantages of the distributed control architectures have been
discussed as a motivation to propose a new control paradigm for the distribution
grid, P2P-based control paradigm. In this paradigm, the distribution grid is
decomposed into coupled microgrids. P2P-based control is deployed in each
microgrid to coordinate the local DERs in a fully distributed manner, and
enable the microgrids to coordinate themselves, through coupling agents, in a
distributed manner as well.
Chapter 3
Real-Time
Optimization-Based
Distributed Voltage Control
27
28 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
3.1 Introduction
within a reasonable time, making voltage profiles comply with the European
standard EN 50160 [3].
Distributed optimization-based voltage control algorithms appearing in literature
are mostly based on the DD method and the Alternating Direction Method of
Multipliers (ADMM) [80]. DD and ADMM decompose a coupled optimization
problem into sub-problems, suitable for distributed control. DD and ADMM
apply the theory of Lagrangian multipliers and duality. ADMM also applies
penalty function methods. DD and ADMM have been used in many publications
to decompose a centralized optimization-based voltage control system into sub-
systems that interact with each other in a distributed manner [27, 28, 107, 108].
Most of these algorithms require a large number of iterations to converge.
Hence, these proposed algorithms fail in regulating voltage profiles within a
reasonable time. To tackle this issue, our work develops a novel fast distributed
optimization-based voltage control algorithm. The algorithm uses a change in
reactive power and active power curtailment of some participating PV inverters
installed in the grid to regulate voltage profiles within allowed limits.
The work of this chapter is four-fold. First, we modify the state-of-the-art
DD algorithm to implement a fully distributed DD method. We then apply
the fully distributed DD method to design a fully distributed voltage control
system. Second, we speed up the convergence of the distributed voltage control
system significantly by applying the JP-ADMM method. Third, three strategies
are proposed in this chapter to improve the performance of our algorithms:
feedback strategy; anti-windup strategy; and active/standby modes. Fourth, we
integrate a push-sum gossip protocol with the DD-based and JP-ADMM-based
voltage control systems, to enable P2P data interchange between the inverters.
A real low voltage network is used to study the convergence speed of the DD-
based, Jacobi ADMM-based, and JP-ADMM-based distributed voltage control
algorithms, to see if they are fast enough to solve a voltage problem in real-
time. Based on the convergence speed study, we propose the use of JP-ADMM
(with the push-sum gossip protocol) for distributing a centralized optimization-
based voltage control algorithm. As shown in Figure 3.1, the gossip-based
JP-ADMM (G-JP-ADMM) algorithm is hosted by the grid voltage support
function (GVSF), and integrated with inner control loops of the PV inverter.
In G-JP-ADMM, the GVSFs are updated in parallel; proximal penalization
functions are used to preserve the convergence; and acceleration factors are used
to speed up the convergence. Two factors for accelerating the convergence of
the JP-ADMM are proposed in this chapter: one related to the reactive power
compensation; and one related to the active power curtailment. To the best of
our knowledge, this work is the first work proposing the use of G-JP-ADMM
for designing a P2P-based GVSF.
30 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
Smart inverter
P2P-based
communication
Voltage support
function Scope of the
G-JP-ADMM chapter
DC AC
Control Control
DC AC PCC
PV Filter
Bus DC
Grid
Figure 3.1: General block diagram of smart PV inverter with P2P-based grid
voltage support function (PCC: point of common coupling).
2 2
(k) (k)
X
minimize cp,invi ∆pinvi + cq,invi ∆qinvi (3.1a)
(k) (k)
∆pinv ,∆qinv ,∀i∈I i∈I
i i
Subject to: ∀i ∈ I
(k)
v ≤ vi ≤v (3.1b)
(k) (k) (k)
f vi , ∆pinvd , ∆qinvd = 0, ∀d ∈ I (3.1c)
(k) (k)
(−cri ) (ppv
i ) ≤ ∆pinvi ≤ 0 (3.1d)
(k) (k) (k)
− ∆q invi ≤ ∆qinvi ≤ ∆q invi (3.1e)
r 2
(k)
2 (k)
∆q invi = (sinvi ) − (ppv
i ) (3.1f)
where I is the set of inverters participating in the voltage control. i and d are used
to indicate inverters within the set I. The objective function (3.1a) minimizes
the total cost of all changes in inverter active power pinvi and inverter reactive
power qinvi needed to maintain voltages within the maximum voltage limit v
and the minimum voltage limit v. The total cost is the sum of the quadratic
2
(k)
cost functions of the individual inverters: cp,invi ∆pinvi represents the cost
(k)
function of a change in active power of inverter i with an amount ∆pinvi , while
2
(k)
cq,invi ∆qinvi represents the cost function of a change in reactive power of
(k) (k) (k)
inverter i with an amount ∆qinvi . ∆pinvi and ∆qinvi have to be within specified
limits, defined by the constraints (3.1d)-(3.1f), to respect the maximum apparent
power magnitude sinvi of inverter i, and the maximum allowed curtailment
cri . ppv
i is the active power generation of the PV module connected to inverter
(k)
i. vi in (3.1b)-(3.1c) is the expected magnitude of the PCC voltage after
(k) (k)
applying ∆pinvi and ∆qinvi . The constraint (3.1c) represents the non-linear
(k)
relation between vi and the active/reactive power changes of the inverters
participating in voltage control. The non-linear relation can be obtained from
the AC power flow equations [109].
32 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
cp,invi and cq,invi are constant factors used to prioritize the use of reactive power,
while active power curtailment is performed only as a last resort. In this chapter,
cp,invi is set to be greater than cq,invi in order to use most reactive power before
starting to use active power curtailment [110].
A centralized controller solves the voltage control problem (3.1a)-(3.1f) in three
steps (see Figure 3.2). First, at time step k, all input variables are gathered;
these include PCC voltages and PV active power generation measured by the
GVSFs. Secondly, the central optimizer solves the optimization problem (3.1a)-
(3.1f) to optimally coordinate the inverters by minimizing the total change
in reactive power compensation and active power curtailment. Finally, the
controller sends the new set-points to the GVSFs. The major drawback of this
approach is that a failure in the central coordinator completely disables the
voltage control. Therefore, a methodology is proposed in this chapter to convert
the centralized optimization-based voltage control problem into a distributed
one.
2-way communication
1 2 i
pv (k)
sinv1, ( p1 ) , v1(k) pinv(k) , qinv(k) pinv(k) , qinv(k) pinv(k) , qinv(k)
1 1 2 2 i i
pv (k)
sinv2, ( p2 ) , v2 (k)
(k)
pv
sinvi , ( pi ) , vi(k)
It is worth to point out that network losses and voltage imbalances are not
considered in the optimization problem. Two constraints can be added to
the optimization problem to limit network losses incurred by reactive power
provision, and minimize voltage imbalances. Minimizing network losses and
voltage imbalances is out of scope of this chapter.
BRANCH FLOW MODEL 33
Since the DD method and ADMM are iterative methods, the time step k of the
distributed optimization-based voltage control system is considered as a control
iteration number.
The use of the AC power flow model in (3.1c) results in a non-convex optimization
problem that is difficult to handle in distributed optimization. To avoid the non-
convexity, many works use linearised power flow equations [111, 112]. Voltage
sensitivities can be used to formulate a linear relation between controlled voltages
(k) (k) (k)
vi , and control variables ∆pinvd and ∆qinvd (∀d ∈ I). In Figure 3.3, a simple
distribution system is shown. The system consists of an inverter d connected to
a bus i via a power line having a resistance Rdi and a reactance Xdi . Based on
the distributed power flow equations of Baran and Wu [113], it can be shown
(k)
that the voltage magnitude vi can be described by the following equation.
(k)
vi = (3.2a)
v
u
u 2 + X 2 ) (p(k) )2 + (q (k) )2
(Rdi
u (k) di inv d inv d
t(v )2 − 2 Rdi p(k) + Xdi q (k) +
d invd invd (k) 2
(vd )
(k−1) (k)
Here, pinvd and pinvd denote active powers of inverter d at the control iterations
(k−1) (k)
k − 1 and k, respectively. qinvd and qinvd denote reactive powers of inverter d
at the control iterations k − 1 and k, respectively.
(k) (k)
For a small change ∆pinvd and ∆qinvd , a first order approximation can be used
to simplify (3.2a) as follows:
vd(k) vi (k)
Rdi Xdi
AC
DC
d pinv(k) +
(k)
j qinv i
d d
in active and reactive powers) into a given network, the voltage sensitivities are
then determined by measuring the change in voltage.
32 31 30 26 28 29 0 300m
Household with PV installation
25
F E
27
13 15 17 19 21 23 33 34 36 38 40 41 43
G
D
10/0.4 kV
250 kVA 14 16 18 20 22 24 35 37 39 42 44
A B
47 49 51 53 55 57 59 61
3 5 7 9 11
1 H
C 45
46 48 50 52 54 56 58 60 62
2 4 6 8 10 12
Figure 3.4: Schematic diagram of the simulated network. All lengths are drawn
to scale.
project [116].
36 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
The main goal of this chapter is to design a voltage control system that does not
rely on a centralized controller. To do so, the centralized optimization problem
(3.1a)-(3.1f) has to be decomposed into sub-optimization problems that can be
solved locally by the GVSFs.
The objective function (3.1a) is basically a sum of separate cost functions, one
for each participating inverter. The constraints (3.1d)-(3.1f) are local, meaning
(k) (k)
that they only influence the local decision variables ∆pinvi and ∆qinvi , and
therefore these constraints can be distributed easily, one for each inverter. On
the other hand, the constraints (3.1b) and (3.1c) are complicating (coupling)
constraints and cannot be distributed, as the voltage magnitude of each PCC
DUAL-DECOMPOSITION-BASED DISTRIBUTED VOLTAGE CONTROL 37
Subject to: ∀i ∈ I
(k) (k)
λinvi ≥ 0 , λinvi ≥ 0 (3.6b)
(k)
(k) (k) (k)
λinvi vi − v + λinvi −vi + v (3.7)
(k)
where vi is given by (3.5). λinvi is the Lagrangian multiplier associated with
the complicating constraint vi ≤ v, whereas λinvi is the Lagrangian multiplier
associated with the complicating constraint −vi ≤ −v. Because of the Karush-
Kuhn–Tucker conditions (KKT), the Lagrangian multipliers cannot be smaller
than zero [122]. One can notice that the optimization problem (3.6a)-(3.6c) is
penalized when the complicating constraints are violated, and rewarded when
these constraints are satisfied strictly.
Keeping the Lagrangian multipliers fixed makes the Lagrangian function L(k)
a sum of separate objective functions, one for each participating inverter.
38 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
Below we prove that the solution of (3.9) equals the solution of the original
optimization problem (3.1a)-(3.1f).
Proof : We define the feasible region Ψ by the complicating constraints set
(k) (k) (k) (k)
{vi ≤ v, −vi ≤ −v, ∀i ∈ I}. If ∆pinvi and ∆qinvi (∀i ∈ I) are within the
DUAL-DECOMPOSITION-BASED DISTRIBUTED VOLTAGE CONTROL 39
feasible region Ψ, then all the complicating constraints are non-positive. In this
case, the Lagrangian function L(k) can be written as:
2 2
(k) (k)
X
L(k) = cp,invi ∆pinvi + cq,invi ∆qinvi +
i∈I
(k) (k)
λinvi × negative value + λinvi × negative value (3.11)
(k) (k)
If ∆pinvi and ∆qinvi (∀i ∈ I) are not within the feasible region Ψ, then all the
complicating constraints are positive. In this case, the Lagrangian function L(k)
can be written as:
2 2
(k) (k)
X
L(k) = cp,invi ∆pinvi + cq,invi ∆qinvi +
i∈I
(k) (k)
λinvi × positive value + λinvi × positive value (3.12)
Based on the above cases, maximum of the Lagrangian function can be computed
as follows:
(
(k) (k)
f (k) ∆pinvi , ∆qinvi , ∀i ∈ I ∈ Ψ,
maximize L = (k)
(k) (k) (3.13a)
(k)
λinv ≥0 ∞ /Ψ
∆pinvi , ∆qinvi , ∀i ∈ I ∈
i
(k)
λinv ≥0
i
∀i∈I
2 2
(k) (k)
X
f (k) = cp,invi ∆pinvi + cq,invi ∆qinvi (3.13b)
i∈I
(k) (k)
Subject to: ∆pinvi , ∆qinvi , ∀i ∈ I ∈ Ψ (3.14)
(3.14) can be rewritten as follows:
minimize maximize L(k) = minimize f (k)
(k) (k)
∆pinv ∈Υ λinv ≥0
i i
(k) (k)
∆qinv ∈Υ λinv ≥0
i i
∀i∈I ∀i∈I
40 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
Subject to:
(k) (k)
∆pinvi , ∆qinvi , ∀i ∈ I ∈ Ψ
(k) (k)
∆pinvi , ∆qinvi , ∀i ∈ I ∈ Υ (3.15)
One can notice that (3.15) is the same as the original problem (3.1a)-(3.1f).
Hence, the optimization problem (3.9) gives the same solution as the original
optimization problem (3.1a)-(3.1f).
The optimization problem (3.9) is a decomposable problem and gives the same
solution as the original optimization problem (3.1a)-(3.1f). However, minimizing
maximize L(k) is not an easy task, as maximize L(k) is ill-conditioned 2 . As
(k) (k) (k) (k)
λinv ,λinv ≥0 λinv ,λinv ≥0
i i i i
∀i∈I ∀i∈I
shown in (3.13a), if the decision variables are slightly outside the feasible region,
then this would lead to maximize L(k) → ∞.
(k) (k)
λinv ,λinv ≥0
i i
∀i∈I
The solution of maximize minimize L(k) gives a lower bound on the problem
(k) (k)
λinv ≥0 ∆pinv ∈Υ
i i
(k) (k)
λinv ≥0 ∆qinv ∈Υ
i i
∀i∈I ∀i∈I
(3.9). The above relation is called the minimax theorem. Below we prove that
relation (3.16) holds.
Proof : For fixed Lagrangian multipliers, one can always claim that the value
(k) (k)
of the Lagrangian function for any given ∆pinvi and ∆qinvi (∀i ∈ I) ∈ Υ is
2 A mathematical problem is ill-conditioned if a small change in input leads to a large change
in the output. This can lead to computational problems. For example, if an optimization
problem is ill-conditioned, the solution exists, but it is very difficult to find [123].
DUAL-DECOMPOSITION-BASED DISTRIBUTED VOLTAGE CONTROL 41
(k) (k)
greater than or equal to its minimum over ∆pinvi and ∆qinvi (∀i ∈ I) ∈ Υ:
(k) (k)
∀∆pinvi , ∆qinvi (∀i ∈ I) ∈ Υ L(k) (∆p, ∆q, λ) ≥ minimize L(k) (∆p, ∆q, λ)
(k)
∆pinv ∈Υ
i
(k)
∆qinv ∈Υ
i
∀i∈I
| {z }
g (k) (λ)
(3.17)
minimize L(k) (∆p, ∆q, λ) can be defined as the function g (k) (λ), that depends
(k)
∆pinv ∈Υ
i
(k)
∆qinv ∈Υ
i
∀i∈I
(k) (k) (k) (k)
only on λinvi , λinvi (∀i ∈ I). Now we fix ∆pinvi and ∆qinvi (∀i ∈ I), to
(k) (k)
maximize both sides of (3.17) over λinvi ≥ 0, λinvi ≥ 0 (∀i ∈ I):
maximize L(k) (∆p, ∆q, λ) can be defined as the function Ĺ(k) (∆p, ∆q), that
(k)
λinv ≥0
i
(k)
λinv ≥0
i
∀i∈I
(k) (k)
depends only on ∆pinvi , ∆qinvi (∀i ∈ I). g (k) (λ∗ ) in (3.18) is a constant value.
(k) (k)
As Ĺ(k) (∆p, ∆q) ≥ g (k) (λ∗ ) for all possible values of ∆pinvi , ∆qinvi (∀i ∈ I)
(k) (k)
∈ Υ, its minimum over all possible values of ∆pinvi and ∆qinvi (∀i ∈ I) ∈ Υ is
greater than g (k) (λ∗ ):
The gap between the dual problem (3.21) and the primal problem (3.1a)-(3.1f)
(or the equivalent relaxed problem (3.9)) is called the duality gap [126]. If the
equality
holds, then we say strong duality holds. Strong duality holds if and only if the
duality gap is equal to zero [127]. This is as opposed to weak duality, which
holds if the duality gap is not zero:
The basic idea of the DD method is to decompose the dual problem into
sub-problems, which are then coordinated by a high level master problem, by
means of the Lagrangian multipliers. This approach results in a distributed
control system that has a degree of centralization. The local controllers (GVSFs)
perform an optimization by themselves. However, a central agent still exists,
inherently resulting in a single point of failure.
In this subsection, we modify the state-of-the-art DD method by letting the
GVSFs compute the Lagrangian multipliers locally, to coordinate themselves
in a P2P manner. We apply the push-sum gossip protocol to enable the P2P
coordination [129].
State-of-the-Art Dual Decomposition:
One can show that the Lagrangian function L(k) can be decomposed into nc local
Lagrangian functions, where nc is the number of smart inverters participating
(k−1) (k−1)
in voltage control. For λinvi , λinvi (∀i ∈ I) computed at the control iteration
k − 1, and (vimeas )(k−1) (∀i ∈ I) measured at the control iteration k − 1, L(k)
can be decomposed as follow:
(k)
X
L(k) = Linvi (3.24a)
i∈I
2 2
(k) (k) (k)
Linvi = cp,invi ∆pinvi + cq,invi ∆qinvi +
(k−1)
(k−1) (k) (k)
X p q
λinvd − λinvd vd,i ∆pinvi + vd,i ∆qinvi +
d∈I
(k−1)
(k−1)
λinvi − λinvi 2(vimeas )(k−1) − v − v (3.24b)
p q
Here, vd,i and vd,i represent the influence of active power and reactive power
(respectively) of inverter i on the PCC voltage vd . As shown in Figure 3.5, the
DD method seperates the dual problem (3.21) into two levels. At the lower
44 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
∀i ∈ I :
(k)
minimize Linvi (3.25a)
(k) (k)
∆pinv ,∆qinv
i i
Subject to:
(k−1) (k)
(−cri ) (ppv
i ) ≤ ∆pinvi ≤ 0 (3.25b)
(k) (k) (k)
− ∆q invi ≤ ∆qinvi ≤ ∆q invi (3.25c)
r 2
(k)
2 (k−1)
∆q invi = (sinvi ) − (ppv
i ) (3.25d)
(k−1)
where (ppv
i ) is the active power generation (of the PV installation i)
measured at the control iteration k − 1.
Master problem
(k-1) (k-1)
λinv 1
λinv n
c
qinv(k) qinv(k)
1 nc
Sub-problem Sub-problem
1 nc
At the higher level, the central agent solves the master problem:
The gradient of g (k) (λ) consists of 2nc components, since g (k) (λ) has 2nc
Lagrangian multipliers. It can be shown that the gradient of g (k) (λ) is:
(k)
∇λinv g (k) (λ) = vi − v , (∀i ∈ I)
i
(k) (k)
X p q
= (vimeas )(k−1) + vi,d ∆pinvd + vi,d ∆qinvd − v (3.27a)
d∈I
(k)
∇λinv g (k) (λ) = −vi + v , (∀i ∈ I)
i
(k) (k)
X p q
= −(vimeas )(k−1) − vi,d ∆pinvd + vi,d ∆qinvd + v
d∈I
(3.27b)
When we use the dual ascent to solve the DD-based voltage control problem,
we get the following simple algorithm:
repeat
Local problem: solve the sub-problems in parallel:
(k) (k) (k)
Find ∆pinv1 and ∆qinv1 that minimize Linv1 .
..
.
46 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
Here, α is the step size of the dual ascent that may vary throughout the
k
iterations. The max operator in the master problem is used to guarantee that the
Lagrangian multipliers are greater than or equal to zero (KKT conditions). With
suitable choice of αk and certain assumptions, the DD method is guaranteed to
converge to an optimal solution [131]. However, as will be demonstrated later
in this chapter, the convergence of the DD method often tends to be slow in
practice.
One can notice that the update of the Lagrangian multipliers, at the control
iteration k, depends on the local PCC voltages (vimeas )(k−1) (∀i ∈ I), measured
at the control iteration k − 1. Meaning that the GVSFs should apply the
control variables at each control iteration, and not at the end of the convergence.
Applying the control variables and updating voltages measurement at each
control iteration is called the iterative feedback strategy (IFS).
There are two approaches regarding applying the control variables. First,
applying the control variables only after the control algorithm converges. Second,
applying the control variables at each iteration. The latter is called IFS (see
Figure 3.6). IFS has been used in [132], and is applied in this chapter to
improve the performance of the distributed voltage control algorithm. In IFS, at
each control iteration k, the GVSF i dynamically adjusts the inverter outputs,
measures the PCC voltage (vimeas )(k) . IFS helps in increasing the voltage
quality. Without IFS, the voltage during the convergence would stay beyond
the accepted limits, without getting closer to these limits until the end of
convergence. Moreover, IFS helps in correcting the error of linearizing the
branch flow model (3.5). Additionally, if the PV generation and/or consumption
changes during the computation, the GVSFs would implement outdated control
variables.
DUAL-DECOMPOSITION-BASED DISTRIBUTED VOLTAGE CONTROL 47
pinv(k) , qinv(k)
Sub-problems i i Master problem
i=1,2, ,nc
Dual decomposition
(k-1) (k-1)
λinvi
, λinv i
i=1,2, ,nc
Feedback
(k-1)
(vi meas ) Grid
i=1,2, ,nc
∀i ∈ I
(k)
(k−1)
λinvi = max 0, λinvi + αk (vimeas )(k) − v (3.28a)
48 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
(k) (k−1)
λinvi = max 0, λinvi + αk −(vimeas )(k) + v (3.28b)
Here, (vimeas )(k) is the magnitude of the PCC voltage i measured in the middle
of the control iteration k.
(k-1) (k-1)
λinvi
, λinv i
i=1,2, ,nc
Feedback
(k)
Grid (vi meas )
i=1,2, ,nc
Figure 3.7: Illustration of the DD-based voltage control with IFS applied in the
middle of the control iteration.
From Figure 3.7, one can notice that there is no feedback from the grid to the
(k−1) (k−1)
sub-problems. This is because the part (λinvi − λinvi )(2(vimeas )(k−1) − v − v)
(k) (k)
in (3.24b) is constant, and can be removed when we minimize Linvi over ∆pinvi
(k)
and ∆qinvi .
(k ) (kG ) (k) (k ) (k )
Si G consists of four vectors: λinvi , λinvi , µinvGi and σ invGi . Each vector has nc
entries, each entry belongs to one of the GVSFs. As will be discussed later, the
(k )
vector µinvGi is used to detect the Lagrangian windup. In step 3 of the algorithm,
(k )
the vector σ invGi is used to make sure that the other GVSFs have filled their
matrices before the GVSF stops the gossiping iterations of the control iteration
(k )
k. In step 7, the GVSF detects a complete filling of its own Si G by observing
(kG )
the entries of the vector Wi .
(k ) (k )
Each entry of Si G represents a sum and Wi G is the weight of this sum
(entries of the same row have the same weight). After stopping the gossiping
50 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
iterations, the correct values of the Lagrangian multipliers and the power limit
indicators can be obtained by dividing each entry of the first three vectors by its
(k ) (k )
own weight, as shown in step 12. The matrix Si G and the vector Wi G are
always transmitted in pairs, so if a packet gets lost, all the other pairs continue
to have the correct values.
The proposed control method continuously repeats two steps. First, the GVSFs
optimize their control decisions in parallel; exceeding the voltage limit has a
fixed price, set by the Lagrangian multipliers. Secondly, the prices are updated;
if one of the controlled voltages exceeds the voltage limit, then the corresponding
price will be increased. This makes the over-voltage (or under-voltage) more
expensive, and in the next iteration the GVSFs should decide to use more
compensation to reduce the over-voltage. However, the optimization might
be infeasible; there might not be enough compensation available to resolve
the over-voltage issues. In such a case, the price will keep on increasing until
the inverters saturate. Past this point, increasing the price further has no use
and will slow down the response when the problem becomes feasible again.
This effect is similar to ’windup’ in classical control theory. The update of the
Lagrangian multipliers are disabled when all inverters reach their operational
limits, which is referred to hereafter as the anti-windup strategy.
Active and standby modes are proposed in this chapter to decide when to
activate/deactivate the GVSFs. There is no need to operate the GVSFs
continuously. When the controlled voltages are within the accepted limits,
then there is no need to activate the GVSFs, whereas all the GVSFs should be
activated when one (or more) of the controlled voltages is beyond the accepted
limits. After activating the GVSFs, the Lagrangian multipliers can be used to
know when to go to standby mode. The GVSFs can be deactivated when the
Lagrangian multipliers of all the GVSFs return back to zero. Zero values of
the Lagrangian multipliers indicate that the complicating constraints (3.1b) are
satisfied strictly. This means that voltages are below the maximum limit, and
above the minimum limit.
DUAL-DECOMPOSITION-BASED DISTRIBUTED VOLTAGE CONTROL 51
In this subsection, we combine all the ideas and equations, presented in the
previous subsections, in one algorithm. Algorithm 2 is proposed in this chapter
to implement a distributed voltage control system provided by GVSFs of smart
PV inverters.
The algorithm includes different strategies that improve the performance of
the fully distributed DD-based voltage control system: active and standby
modes; iterative feedback; and anti-windup. The push-sum gossip Algorithm 1
is applied in Algorithm 2 to enable P2P data interchange between the GVSFs.
(k)
In step 14, the GVSF sets µinvi = 1 when the inverter reaches its power limits.
(k)
In case of voltage rise, µinvi is set to 1 when maximum reactive power absorption
and maximum active power curtailment are used. In case of voltage drop, on
(k)
the other hand, µinvi is set to 1 when maximum reactive power injection is used.
In step 16, when all the inverters reach their power limits, each GVSF freezes
its Lagrangian multipliers to avoid windup.
Numerical tests on the simulated network of Figure 3.4, over 104 scenarios of
daily residential demand and PV profiles (5 s resolution), are performed to
study convergence speed of the DD-based distributed voltage control system
(Algorithm 2). The daily residential demand and PV profiles are generated
for a weekday in July, to incorporate the effect of high PV generation on grid
voltages. The time step of each control iteration k is 5 s, and the time step of
each gossiping iteration is 100 ms. The fastest convergence of Algorithm 2 is
achieved when the step size α is set to a constant value of 0.3. Algorithm 2
diverges when α is higher than 0.3.
Figure 3.8 shows the average convergence speed of the DD-based distributed
voltage control system with number of agents3 varying from 10 to 60. The
average convergence speed is calculated by taking the mean of convergence
speeds of different time steps for the 104 scenarios. One can notice that the
average convergence speed varies from 8.4 minutes to 65 minutes4 . When 60
agents participate in the voltage control, Algorithm 2 needs more than one hour
3 Distributed voltage control system with 10 agents, for example, means that only the PV
inverters of the last 10 nodes of Figure 3.4 (nodes 53-62) participate in the voltage control.
4 One minute represents 12 iterations (60 s/5 s (5 s is the time step of one control iteration))
52 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
(k)
n o
(k)
13: or ∆qinvi = ∆q invi , then
(k)
14: µinvi ← 1 (power limit indicator)
15: end if
(k)
16: if all the entries of the vector µinvi are 1, then
(k) (k−1) (k) (k−1)
17: λinvi ← λinvi , λinvi ← λinvi (anti-windup)
18: else
(k) (k)
19: Update λinvi and λinvi based on (3.28a) and (3.28b)
20: end if
(k) (k) (k)
21: Execute Algorithm 1 to update the vectors λinvi , λinvi and µinvi
(k) (k)
22: if all the entries of the vectors λinvi and λinvi are zero, then
23: Go to step 2 (standby mode)
24: else
25: k ← k + 1, repeat steps 8-25
26: end if
27: else
28: Stay in standby mode
29: end if
less than 10 minutes). It does not make sense to spend one hour, or even 30
minutes, to solve a single voltage problem in real-time. Therefore, we conclude
from this study that the DD method can be used to implement a distributed
voltage control system with few agents (e.g. 10 agents or less). We also see
from this study that the DD method has more potential in offline distributed
optimization (e.g. offline coordination of the agents) than online distributed
optimization.
The classic ADMM algorithm [133, 134] is a two-block system, it cannot be used
to decompose a centralized optimization problem into sub-problems that can be
solved in parallel. The 2-block ADMM can be extended to: 1) multi-block Gauss-
Seidel ADMM (GS-ADMM); and 2) multi-block Jacobi ADMM (J-ADMM).
Different forms of ADMM are reviewed in [135]. GS-ADMM is not amenable
for parallelization as the blocks are updated one after another (sequentially).
Additionally, the convergence is not guaranteed [136]. J-ADMM updates all
the blocks in parallel. The convergence of J-ADMM can be guaranteed under
certain assumptions [137], but it has a slow rate of convergence. Our simulation
results show that the convergence of the J-ADMM algorithm is even slower than
the convergence of the DD algorithm. In [138], a mathematical method called
JP-ADMM is developed. The work of [138] shows that the 2-block ADMM
can be extended to parallel multi-block ADMM and preserve a convergence
(at a rate of o(1/k)) by adding proximal terms to the augmented Lagrangian
function. The use of proximal terms to preserve the convergence of the parallel
multi-block ADMM has been also discussed in [139].
In this section, the Jacobi-Proximal ADMM (JP-ADMM) method is applied to
relax the optimization problem (3.1a)-(3.1f), and decompose it into sub-problems
that can converge quickly. These sub-problems are then solved separately by
the GVSFs in parallel, imposing only the local constraints. The motivation
behind using the ADMM is driven by its superior convergence, and its ability
to solve voltage problems within a reasonable time.
JP-ADMM applies an iterative optimization, in a Jacobi-fashion, of the Proximal
Augmented Lagrangian Function (PALF), followed by a steep ascent update
of the Lagrangian multipliers. In this chapter, JP-ADMM is integrated with
the push-sum gossip protocol to enable the GVSFs to communicate with each
other, and exchange their control variables and Lagrangian multipliers in a P2P
fashion. With P2P communication, the GVSFs coordinate themselves and make
the correct control decision in every particular situation to maintain voltages
within the required limits.
The PALF of the objective function (3.1a) and the complicating constraints
(3.1b)-(3.1c) can be formulated as:
(k) 2
ρinvi ρinvi 2
(k)
+ max 0, U invi + max 0, U invi
2 2
!
τinvi (k) (k−1) 2 τinvi (k) (k−1) 2
+ ∆pinvi − ∆pcinvi + ∆qinvi − ∆qinv
c
2 2 i
(3.31)
(k) 2
Here, ρinvi > 0 is the augmented penalty factor. max 0, U invi and
2
(k)
max 0, U invi are one-sided quadratic penalty functions used to penalize
(k)
the objective function only when vi is higher than v or less than v.
(k) (k−1) 2
τinvi > 0 is the proximal penalization factor. ∆pinvi − ∆pcinvi and
2
(k) (k−1)
∆qinvi − ∆qinv are the proximal terms. These terms penalize the
c
i
(k) (k)
deviation of the control variables at the k-th iteration (∆pinvi and ∆qinvi ) from
the control decisions (∆pcinvi )(k−1) and (∆qinv
c
i
)(k−1) that have been calculated
at the previous iteration (superscript c: constant). The proximal terms are used
to preserve the convergence of the extended parallel multi-block JP-ADMM. To
derive the sub-problems, let’s first define the following:
(k)
X p q
Cd,i = vd,j (∆pcinvj )(k−1) + vd,j (∆qinv
c
j
)(k−1) (3.33)
j∈I
j6=i
(k)
(k−1)
(k) (k−1) (k)
X
LPALF
invi = finvi + λinvd − λinvd Yd,i
d∈I
ρinvd 2
(k) (k)
+ max 0, (vdmeas )(k−1) + Yd,i + Cd,i − v
2
!
ρinvd 2
(k) (k)
+ max 0, −(vdmeas )(k−1) − Yd,i − Cd,i + v
2
nc
τinvi > ρinvi − 1 (3.35)
2 − γinvi
The acceleration factor γinvi is used to have a steep ascent update of the
Lagrangian multipliers, which helps in increasing the speed of convergence. In
ADMM-BASED DISTRIBUTED VOLTAGE CONTROL 57
end for
this work, two acceleration factors are proposed: 1) the reactive acceleration
q p
factor γinv i
; and 2) the curtailment acceleration factor γinv i
. The acceleration
q
factor γinvi is used to accelerate the convergence of the controlled voltages to
p
v (or v) when reactive power compensation is used, whereas γinv i
is used
to accelerate the convergence when active power curtailment is used. As
active power curtailment is penalized more than reactive power compensation,
convergence of the controlled voltages to the accepted limits using curtailment
is slower than the convergence when using reactive power compensation. Hence,
p q
γinv i
should be greater than γinv i
.
(k) (k)
26: if all the entries of the vectors λinvi and λinvi are zero, then
27: Go to step 2 (standby mode)
28: else
29: k ← k + 1, repeat steps 8-29
30: end if
31: else
32: Stay in standby mode
33: end if
p
In step 13, the switching to the curtailment acceleration factor γinv i
occurs
when most of the available reactive power capacity is used (90%) and the active
power curtailment has begun to be used (10%). In step 18, the GVSF sets
(k)
µinvi = 1 when the inverter reaches its power limits. In case of voltage rise,
(k)
µinvi is set to 1 when maximum reactive power absorption and maximum active
(k)
power curtailment are used. In case of voltage drop, on the other hand, µinvi
is set to 1 when maximum reactive power injection is used. In step 20, when
all the inverters reach their power limits, each GVSF freezes its Lagrangian
multipliers to avoid windup.
The maximum and minimum voltage limits are set to 207 V and 253 V,
respectively. The sensitivity of the controlled voltages to the change in the
active/reactive power is calculated based on the constant linear voltage model
presented in Section 3.3.
To prioritize the use of reactive power, while active power curtailment is
performed only as a last resort, the reactive power penalization factor cq,invi is
set to 1 and the curtailment penalization factor cp,invi is set to 200 (the tuning
is based on trial and error).
q p
To tune γinv i
and γinv i
, we did two experiments. The first experiment was
q p
to tune γinvi , whereas the second experiment was to tune γinv i
. In the first
experiment, we disabled the active power curtailment, to regulate voltage profiles
q
using only reactive power. γinv i
was initialized to 1. After that, we increased
q
γinvi gradually till undesirable oscillation was noticed in voltage profiles. In the
60 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
The low load and high PV generation result in reversed power flows. The
reversed power flows cause a voltage rise beyond the EN 50160 limits. Figure
3.10 shows voltage profiles of the 62 households (phase to neutral voltages).
One can see that most of voltage profiles exceed the voltage limit v = 253 V.
To solve the voltage rise problem, the P2P-based GVSFs of the 20 inverters
ANALYSING THE PERFORMANCE OF THE JP-ADMM-BASED DISTRIBUTED VOLTAGE CONTROL
ALGORITHM 61
Generation
Consumption
are activated. The regulated voltage profiles are shown in Figure 3.11. The
existing violations of the over-voltage limit are mitigated. The voltage peaks,
resulting from steep change in PV generation, fall beyond the limits, because
the inverters need some time to react to the change (convergence time).
activating the GVSFs, which means that the voltage quality of the voltage
profiles has been improved significantly.
Z tend =24:00
Ei = max (vimeas − v , 0) dt (3.38)
tstart =00:00
without control
with control
Figure 3.12: Voltage quality metric E of the 62 households with and without
voltage control.
Figure 3.13a shows the Lagrangian multipliers λinv62 5 of smart inverter No. 62.
The GVSF of inverter 62 is in standby mode between 00:00 and 08:40; when
voltages are within the limits. The GVSF starts to regulate voltages at 08:40
(active mode), when voltages start to violate the maximum limit. The GVSF
returns back to standby mode and the Lagrangian multiplier returns back to
5λ is zero for the entire day, since there are no voltage drop problems.
inv62
ANALYSING THE PERFORMANCE OF THE JP-ADMM-BASED DISTRIBUTED VOLTAGE CONTROL
ALGORITHM 63
zero at 16:35, when voltages return back to normal values due to solar irradiance
reduction. Figure 3.13b shows the control variables of smart inverter No. 62.
The average reactive power inverter 62 absorbs during the day is around -4
kvar, whereas the average active power the inverter curtails is around -0.5 kW.
∆pinv62 and ∆qinv62 return to zero when the Lagrangian multipliers return back
to zero.
pinv [kW]
62
qinv [kvar]
62
(a) (b)
Figure 3.13: (a) Lagrangian multiplier λinv62 of smart inverter No. 62, (b)
Active power curtailment and reactive power compensation of smart inverter
No. 62.
Figure 3.14 shows a comparison of the convergence speed of the proposed JP-
ADMM algorithm (Algorithm 4), the J-ADMM algorithm, and the DD algorithm
(Algorithm 2). The J-ADMM algorithm can be implemented as presented in
Algorithm 4, but without including the proximal terms and the acceleration
factors [138]. As discussed earlier, the best results for the JP-ADMM algorithm
(in terms of convergence speed) occur when γ q = 500, γ p = 104 , and ρ = 0.001
(for all the GVSFs). For the case study of this subsection, the best results for
the J-ADMM algorithm occur when ρ is set to 0.1 for all the GVSFs. The
J-ADMM algorithm does not converge when ρ is higher than 0.1. The fastest
convergence of the DD algorithm is achieved when the step size α is set to 0.3
for all the GVSFs. The DD algorithm diverges when α is higher than 0.3.
The three algorithms are implemented at each GVSF of the 20 inverters, and
are executed at 14:00, the time when voltage profiles have the highest voltage
rise. Figure 3.14 presents the convergence of the PCC voltage of smart inverter
64 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
JP-ADMM, q =500, p
=10 4 , =0.001
J-ADMM, =0.1
DD, =0.3
No. 62 (worst voltage). We can see that the JP-ADMM algorithm is clearly
the fastest one among the compared algorithms. The JP-ADMM algorithm
returns the PCC voltage back to the limit in less than 10 iterations, whereas the
J-ADMM algorithm needs around 500 iterations, and the DD algorithm needs
arround 370 iterations to bring the PCC voltage back to the limit. From this
study, we can also conclude that the J-ADMM algorithm is even slower than
the DD algorithm. It is worth to mention again that the JP-ADMM enjoys a
fast preserved convergence thanks to the proposed acceleration factors and the
applied proximal penalization functions.
The linear voltage model presented in Section 3.3 is used to predict voltages of
(k) (k)
the control iteration k before applying the control decisions ∆pinvi and ∆qinvi
(∀i ∈ I). The optimization of the control decisions depends highly on how
accurate the prediction is. The prediction made by equation (3.5) is quite
accurate. Figure 3.15 shows a comparison between the predicted voltage and
the actual one of smart inverter No. 62. The algorithm is executed for two hours,
from 14:00 till 16:00 (period of worst voltage problem). The predicted voltage
is calculated by equation (3.5), whereas the measured voltage is calculated by
exact power flow equations (using MATPOWER). One can notice the high
accuracy of the linear voltage model. The maximum error is around 0.3%.
The high accuracy of the prediction is due to two reasons; first, the algorithm of
the proposed method is an iterative one. The system takes small steps towards
the final solution. At each iteration, the change in reactive power and active
ANALYSING THE PERFORMANCE OF THE JP-ADMM-BASED DISTRIBUTED VOLTAGE CONTROL
ALGORITHM 65
Predicted voltage
Measured voltage
Figure 3.15: Predicted voltage vs. actual voltage of smart inverter No. 62.
power curtailment is small. For a small change, the first order approximation of
equation (3.2a) is quite accurate, and second, thanks to the feedback strategy,
the system corrects the error in its decision at each control iteration. This means
that the error does not accumulate. Figure 3.16 presents a comparison between
the predicted voltage and actual voltage for two cases, with and without the
feedback strategy. With the feedback strategy the error is around 0.3%, whereas
without the feedback strategy the error is around 1.2%. One can notice that
without the feedback strategy the algorithm brings the actual voltage to a value
lower than the maximum limit. This means that the voltage is over-regulated,
and the amount of the absorbed reactive power and curtailed active power is
more than needed.
(a) (b)
Figure 3.16: (a) Voltage regulation with the feedback strategy, (b) Voltage
regulation without the feedback strategy.
(a) (b)
Figure 3.17: (a) Voltage profiles of the 124-bus network without voltage control,
(b) Voltage profiles of the 124-bus network with voltage control.
(62-bus network) and voltage convergence of smart inverter No. 124 (124-bus
network)6 , at 14:00. In case of 20 agents (62-bus network), the algorithm takes
8 iterations to bring the voltage back to the limit, whereas the algorithm takes
15 iterations to bring the voltage back to the limit in case of 50 agents (124-bus
network). Hence, increasing the number of agents does not lead to exponential
increase in the number of iterations. This is because in G-JP-ADMM, each
agent treats the decision variables of other agents as constants.
62-bus network
124-bus network
Figure 3.20 shows the number of GVSFs each GVSF needs to communicate
with, to know the control decisions and the Lagrangian multipliers of all the
GVSFs. In case of 20 smart inverters participating in voltage control, each
GVSF needs to communicate with 8 to 13 GVSFs. In case of 50 smart inverters
participating in voltage control, each GVSF needs to communicate with 10 to
14 GVSFs. The number (of GVSFs to be communicated with) is not fixed since
each agent contacts random agent at each gossiping iteration. The results of
6 Smart inverter No. 124 is at the end of the extended feeder DH.
68 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
JP-ADMM
J-ADMM
DD
Figure 3.20 demonstrate the fact that each GVSF needs to communicate with
some GVSFs (not all) to know the status of all other GVSFs. The results also
demonstrate the scalability of the proposed push-sum gossip protocol. The
gossip-based protocol enjoys exponential rapid spread of information. Increasing
the number of smart inverters does not significantly increase the communication
burden.
20 agents
50 agents
Discussion
The time needed to finish one control iteration of JP-ADMM, J-ADMM and
DD is around 5 seconds7 . Based on the results presented in the 124-bus case
study, the proposed JP-ADMM needs around 1.25 minutes (5s × 15 iterations)
to bring the voltage of bus No. 62 to the accepted limit, whereas J-ADMM
needs around 86.6 minutes (5s × 1040 iterations), and DD needs around 59.4
minutes (5s × 713 iterations). We conclude from this analysis that the proposed
JP-ADMM algorithm succeeds in regulating the voltage profiles to comply with
the European standard EN 50160, whereas DD and J-ADMM fail to comply
with the standard EN50160, since they fail to maintain all 10 min mean values
of voltages within the range [207 V, 253 V].
The problem treated in this chapter is a convex problem. For non-convex
problems, we would like to refer the readers to the coupled local minimizers
(CLM) method [140, 141]. CLM is intended for fast global optimization
of non-convex problems and speeds up simple gradient descent schemes in
convex problems. In CLM, a cooperative search mechanism is set up using a
population of local optimizers, which are coupled during the search process by
synchronization constraints.
3.8 Conclusion
This chapter presents two novel distributed voltage control methods for PV
inverters to expand their features with added grid voltage support functions.
In the first method, the state-of-the-art dual decomposition (DD) method is
modified to a fully distributed dual decomposition, by decomposing its master
problem into local sub-problems. The fully distributed dual decomposition is
then used to convert a centralized optimization-based voltage control system
into a distributed one. Simulation results show that the DD-based distributed
voltage control system suffers from slow convergence.
In the second method, we succeeded in speeding up the convergence of the
distributed voltage control system, by applying the Jacobi-Proximal Alternating
Direction Method of Multipliers (JP-ADMM). The JP-ADMM enjoys a fast
preserved convergence thanks to the proposed acceleration factors and the
applied proximal penalization functions. For a distributed voltage control
7 Based on the performance of the workstation, and assuming the following: a)1 second as
the time needed to measure the PV generation and the voltage, b) 50 ms as the response
time of the inverter inner control loop, c) 100 ms communication latency
70 REAL-TIME OPTIMIZATION-BASED DISTRIBUTED VOLTAGE CONTROL
Laboratory Implementation of
Peer-to-Peer Voltage Control
System
4.1 Introduction
71
72 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
The architecture of the P2P voltage control testbed is depicted in Figure 4.1.
The testbed consists of four different layers which interact with each other: (1)
microgrid layer; (2) control layer; (3) communication layer; and (4) monitoring
1 The testbed was developed in the P2P-SmarTest project. The communication modules of
the testbed have been built and supplied by the Centre for Wireless Communications (CWC,
University of Oulu).
TESTBED ARCHITECTURE 73
Monitoring agent
Data Acquisition
Monitoring
(11)
VM R. Pi VM R. Pi VM R. Pi
(9) (10) (9) (10) (9) (10)
P2P communication
D2D D2D
Actuator agent
(7) (7)
Actuator agent
(5)
Voltage measurement
Voltage measurement
Voltage measurement
(5)
Control
GVSF GVSF
(6) (6)
L R L R
Microgrid
(3) (3)
DC power DC power
(2) supply
(1) (2) supply
(1)
Inverter Inverter
The overall schematic of the testbed2 is depicted in Figure 4.2. The microgrid,
shown in Figure 4.3, is connected to the main grid through 400 V (line-to-line
voltage (L-L)), 64 A busbar.
(7)
D2D
Actuator agent
(7) (7) (7)
D2D D2D D2D
(5)
(8) (8) (8)
Observer agent Observer agent Observer agent
VM VM VM RTT (12)
1
R. Pi 2 R. Pi 3 R. Pi
(10) (10) (10)
(9) (9) (9)
L1 R1 L2 R2 (1)
Inverter 2
Node 0
Node 1
Node 2
(3) (3)
(7)
D2D
Grid (400 V L-L) DC Power
From other (1) Supply
agents Router
(2)
R. Pi Inverter 1
RTT
DC Power
(11) Supply
(2)
Figure 4.2: Schematic of the P2P voltage control testbed (VM: voltmeter, D2D:
device-to-device communication module, R: resistor, L: inductor, RTT: real
time target computer, R.Pi: raspberry pi computer, L-L: line-to-line, labels
(1)-(12) indicate the different parts of the testbed, labels (1)–(11) same as in
Figure 4.1).
The P2P voltage control algorithm regulates voltages within allowed limits
based on an optimization problem. The algorithm uses a minimum change
2 Short video on the testbed: https://www.youtube.com/watch?v=CgZn-A5Ejww
76 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
(12) (2)
Grid
(7)
(1) (10)
(5)
(3)
(9)
(2)
(9)
the active power. Therefore, the actuator agent can take two actions: reducing
(k)
the active power (by an amount ∆pinvi ) and injecting or absorbing reactive
(k)
power (by an amount ∆qinvi ).
(k)
Each actuator agent i solves the following optimization problem to find ∆pinvi
(k)
and ∆qinvi , at each time step k:
2 2
(k) (k)
argmin cp,invi ∆pinvi + cq,invi ∆qinvi + (4.1a)
(k) (k)
∆pinv ,∆qinv
i i
X (k−1)
p (k) q (k)
λy − λ(k−1)
y vy,i ∆pinvi + vy,i ∆qinvi
y∈Y
Subject to:
(k) (k)
(−cri ) (ppv
i ) ≤ ∆pinvi ≤ 0 (4.1b)
(k) (k) (k)
− ∆q invi ≤ ∆qinvi ≤ ∆q invi (4.1c)
r 2
(k)
2 (k) (k)
∆q invi = (sinvi ) − (ppv
i ) + ∆pinvi (4.1d)
Here, i ∈ I is the number (index) of the actuator agent (I is the set of actuators
(control nodes) participating in voltage control). y ∈ Y is the number (index)
of the observer agent (Y is the set of observers participating in voltage control).
2
(k)
cp,invi ∆pinvi represents the quadratic cost of a change in active power
2
(k) (k)
of inverter i with an amount ∆pinvi at time step k, while cq,invi ∆qinvi
represents the quadratic cost of a change in reactive power of inverter i with an
(k)
amount ∆qinvi at time step k. cp,invi and cq,invi are constant factors used to
(k) (k)
penalize the control variables ∆pinvi and ∆qinvi . These factors define priorities
for the control actions. It is supposed that reactive power control of the inverter
is cheaper than cutting its active power. Therefore, cp,invi should be greater
than cq,invi in a sense that gives priority of the control action to the reactive
power. When the reactive power of the inverter is not sufficient, active power
curtailment of the inverter will be used to regulate system voltages. In our
control system, we set cp,invi = 200 and cq,invi = 1, ∀i ∈ I. Active power
curtailment can be penalized more to minimize its use, but having higher cp,invi
would decrease the speed of convergence when the curtailment is used to return
voltages back to limits.
78 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
p q
vy,i and vy,i are the sensitivity of the voltage at bus y to the change in the
active power and reactive power of inverter i, respectively. As discussed in
Chapter 3, these voltage sensitivities can be calculated based on the impedance
and topology of the microgrid. cri is the curtailment factor. In this chapter,
cri is set to 30%, ∀i ∈ I. In reality, cri can be set based on how much the
(k)
prosumer would like to curtail the active power. (ppv i ) represents the active
power generation of the emulated PV module connected to inverter i, at time
step k. sinvi is the rated apparent power of inverter i.
(k−1)
λy and λ(k−1)
y are the control signals of violating the maximum and minimum
(respectively) allowed voltages at bus y. They are calculated at the previous time
step k − 1 and considered in the optimization of time step k. Mathematically
speaking, they represent the Lagrangian multipliers. Each observer measures
the voltage at its bus and updates these control signals based on the following
equations:
(k)
(k−1)
(k)
λy = max 0, λy + α vymeas −v
(4.2)
meas (k)
λy = max 0, λy
(k) (k−1)
− α vy −v
(k)
where λy and λ(k)
y are the updated control signals calculated at time step k,
(k)
and considered in the optimization of time step k +1. vymeas is the measured
(k) (k)
voltage at bus y after applying the decisions ∆pinvi and ∆qinvi . v and v are
the maximum and minimum allowed voltages, respectively. We set v = 1.1 p.u.
(per unit) and v = 0.9 p.u., according to the European standard EN 50160 [3].
The parameter α is the step size of the dual decomposition method. Because of
the Karush-Kuhn–Tucker conditions (KKT), the Lagrangian multipliers cannot
be smaller than zero. This explains the use of maximum operator in (4.2).
The control algorithm goes through the following steps:
1. Each observer agent measures the voltage. If the voltage exceeds the upper
voltage limit v, it will increase λy . If the voltage is lower than the upper
limit, it will decrease λy , at most until it reaches zero. A similar procedure
applies to λy . The parameter α determines how large the updates to the
control signals will be.
2. Actuator agents receive updates of λy and λy periodically. They will
adjust their compensation to take new values of the control signals into
account.
3. The voltage changes due to the actions of actuator agents. Observer
agents update again their λy and λy , and the whole process repeats.
CONTROL OF THE RAPID PROTOTYPING INVERTER WITH GRID VOLTAGE SUPPORT FUNCTION
79
From this explanation, it is clear that this process is based on feedback. As long
as the voltage problem persists, observer agents will increase control signals to
get more compensation from actuator agents. The effect of α is similar to a
gain in control theory. The trade-off in its selection is similar: a low value can
lead to slow convergence, while a too large value can lead to instability.
LCL Filter
Voltage sensors
Bypass resistors
DC Voltage sensor DC Bus
Ethernet with an on-board PC-based Real Time Target (RTT), which controls
the Triphase power electronics as shown in Figure 4.5. Python has been used to
code a software that manages the interface with D2D communication modules,
fetches the Lagrangian multipliers from D2D modules, stores the PV profiles,
and solves the quadratic optimization problem (4.1a)-(4.1d). The software also
manages the interface with MATLAB.
MATLAB exposes an interface to the Python software, which allows the Python
software to directly execute scripts in MATLAB. The Python software uses a
(k) (k)
MATLAB script to push updates on the PV profiles, and ∆pinvi and ∆qinvi
setpoints to Simulink. MATLAB also manages the interface with the RTT to
control the switches and fan of the PM15FM30C.
The P2P voltage control algorithm represents a high level control system to
coordinate inverters in a distributed way. For the inverter to be able to follow
the regulation of the P2P control algorithm, an internal control system has
to be implemented and integrated with the GVSF. We have implemented a
state-of-the-art current control loop, Phase-Locked-Loop (PLL), and Kalman
filter to drive the inverter. The inner control system of the inverter is depicted
in Figure 4.6.
The inverter uses current-mode control to control the active and reactive power.
The line current is tightly regulated by the current control loop, through the
inverter AC-side terminal voltage. Then, the active and reactive power are
controlled by the phase angle and the amplitude of the inverter current with
respect to a rotating frame that is synchronized with the point of common
CONTROL OF THE RAPID PROTOTYPING INVERTER WITH GRID VOLTAGE SUPPORT FUNCTION
81
(7)
(12)
(1)
Figure 4.5: Actuator agent setup (labels same as in Figures 4.1 and 4.2).
coupling (PCC) voltage using PLL. A Kalman filter is placed in front of the PLL
in order to ensure that the PLL input at all times matches an ideal sinusoidal
waveform as closely as possible, even when the voltage is highly distorted by
the presence of harmonics. This ensures fast and low distortion operation of
the PLL. Kalman filter is used in this work because it efficiently deals with the
uncertainty of tuning its parameters.
The reference setpoints of active and reactive power are calculated based on
the PV profiles and the decisions of the GVSF as shown in Figure 4.6; then
the reference setpoints of active and reactive power are converted into d-q
(direct-quadrature) reference setpoints of the three phase current, and these
d-q setpoints are used by the current control loop as a reference to control the
d-component and q-component of the three phase current in order to follow the
reference setpoints of active and reactive power.
The details of the design and tuning of Kalman filter are presented in [148].
Chapter 8 in [149] presents details of the design and tuning of proportional-
integral (PI) controllers used by the current control loop and PLL. The d-q
transformation (of the current, voltage and power) can be also found in the
same chapter.
It is worth mentioning that we have not used a voltage control loop, because
the voltage of the DC bus is fixed by the DC source. For the inverter to be able
to inject power from the DC side to the AC side, the DC bus has to be charged
82 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
LCL Filter
Ab
Vdc
Ac
PWM
10 Vabc
PCC
Iabc Fundamental+Harmonics
Vabc
abc
wt w
Kalman Filter
dq
Fundamental
Vabc
PWM gating pulses wt abc
dq
ma , mb , mc Id Iq Vd Vq
Vq
VDC Current control loop wt PLL
PI controller PI controller
f
Idref Iqref
Saturation block
P ref Idref
Q ref Iqref
P ref Q ref
Inner control system
PV profile + (4)
+
(k) (k)
p inv q inv
Label(7) Actuator agent
GVSF (5)
Lagrangian multipliers (6)
Figure 4.6: Inner control system of the inverter integrated with the GVSF
(labels same as in Figures 4.1 and 4.2).
4.5.1 Background
Konstantin Mikhaylov and Jussi Haapola (Centre for Wireless Communications, University of
Oulu) have provided the materials of this section [150, 151].
84 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
Stack Stack
on on
bottom top
Mini USB
connector (gives Reset Switch
power and acts
as data
interface
Fully assembled devices
during testing
USB-UART
USB UART Main UART Radio
submodule
cable board submodule
[FTDI VCP]
module can be configured to periodically report the complete table (i.e., the
data from all other agents) to its agent.
The developed firmware implements a multi-stage error detection and correction
system. In case of non-critical errors (e.g., wrong format of UART commands
from the physically connected agent), the module recovers automatically. In
case of severe mistakes (detected by the software or if the software hangs), the
module reset procedure is initiated. After reset, the most recent state of the
module is recovered.
Initializations
Initialization Initialization
Critical Critical
errors errors
Critical errors
Read Commands
Data
Reboot
Synchronized Protocol
The radio thread handles control over the radio transceiver and implements
a synchronized radio protocol. The synchronized protocol is a simple slotted
protocol, where each of the D2D modules is assigned a periodic time slot for
transmission of its data and receiving the transmissions from the other modules
in their respective slots, as shown in Figure 4.9.
The parameters of the protocol, namely the number of slots (M) in the
superframe and the duration of each slot (T-slot) are hardcoded in the firmware
and cannot be changed without reprogramming the module. Each module
DEVICE-TO-DEVICE COMMUNICATION 87
uses for its transmission the slot with the number equaling to its programmed
identifier (i.e., a module with ID 1 will send in slot 1, etc.). Empirically it was
found out that the need of using low-speed UART interface between the main
module and the radio transceiver chipset and the slow operation of the chipset
itself introduces substantial overheads (e.g., packet transmission, switching
between transmit and receive, etc.). Due to this reason, the duration of one slot
cannot be set below 150 ms.
When enabled the first time, the D2D module based on this protocol first
scans the radio channel for several superframe periods. If it does not find
any transmissions and it has data to send, it will start the transmission right
away. If during scanning a module finds some transmissions ongoing, it will
use this transmission as a reference for defining its designated slot. After each
superframe, a module adjusts its synchronization. As a reference point for
adjusting the synchronization, each module uses the timestamp of the packet
with minimum identifier not exceeding the identifier of the module. If such a
reference is not available, no compensation is applied. As a practical example,
module 1 transmitting in slot 1 never adjusts its synchronization. If modules 2,
3 and 4 hear transmission of module 1, they will adjust their synchronization
based on it. If module 5 does not hear module 1 but hears modules 2 and 3, it
will adjust its synchronization based on the transmission of module 2.
T-superframe
T-slot T-slot
slot 1 slot 2 slot 3 slot M slot 1
Time
Module 1
TX RX TX
Time
Module 2
RX TX RX
Time
To test the performance of the P2P distributed voltage control system, one
needs to create a voltage rise (or drop) problem and solve it in a P2P fashion. To
create a voltage rise problem in a laboratory-based microgrid, a high-power
injection from the inverters back to the grid can be used. Alternatively, the
impedance of the feeder depicted in Figure 4.2 can be oversized to create such
a problem with low-power injection. In the following experiments, R1 and R2
(of the microgrid) are set to 8 Ω, L1 and L2 (of the microgrid) are set to 5 mH.
Figure 4.10 shows the generation profile applied at both inverters. The active
power generation starts at zero, and increases to a maximum of 1200 W. At the
higher generation, the voltage is expected to rise above the maximum voltage
limit. To comply with the European standard EN 50160, voltage limits v and v
are enforced to be ±10% of the nominal phase voltage.
Two experiments are carried out to compare voltage profiles with and without
voltage control. The comparison helps in quantifying the performance of the
P2P voltage control.
Ac�ve power [W]
Time [s]
node 0
node 1
node 2
Maximum voltage
Voltage [V]
Time [s]
Figure 4.11: Voltage profiles without voltage control for the 3 nodes of the
microgrid (the generation profile causes over-voltages up to 1.145 p.u.)
The inverters apply the same generation profile, but now the agents execute the
distributed voltage control algorithm. This leads to the voltage profile shown
by Figure 4.12. When an increase in generation causes an over-voltage issue,
the agents bring voltages back to the defined limits (±10%) within 3 min.
node 0
node 1
node 2
Maximum voltage
Voltage [V]
Time [s]
Figure 4.12: Voltage profiles with voltage control for the 3 nodes.
The actions of the observer and actuator agents are reflected in Figure 4.13.
The evolution of the control signals over time are presented in Figure 4.13c,
and 4.13d. The control signals for under-voltages (λ) are zero, because no
under-voltages beyond the limits occur during this experiment. The control
signals for over-voltages (λ) however, increase sharply after an increase in the
voltages above the maximum voltage limit. One can notice that the control
signals λ return back to zero when the voltages return back to normal values
90 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
inverter 1 inverter 1
inverter 2
[Var]
inverter 2
Δpinv[W]
Δq inv
node 0 node 0
node 1 node 1
node 2 node 2
λ λ
Figure 4.13: (a) Active power curtailment of inverter 1 and inverter 2; (b)
Reactive power compensation of inverter 1 and inverter 2; (c) Control signals
for over-voltages; (d) Control signals for under-voltages.
When the measured voltage vymeas is less than v, λy starts to decrease till it
reaches zero. One can explain this based on (4.2). The Lagrangian multipliers
drop back to zero because the underlying profile of the inverters change. The
active power injection drops, and the voltage drops with it. The Lagrangian
multipliers adapt to the new situation. When both λy and λy are zero at each
observer, the optimization problem (4.1a)-(4.1d) can be written as:
2 2
(k) (k)
argmin cp,invi ∆pinvi + cq,invi ∆qinvi + zero
(k) (k)
∆pinv ,∆qinv
i i (4.3)
(k)
One can notice that the solution of the above optimization problem is: ∆pinvi = 0
(k)
and ∆qinvi = 0. Hence, a stop mechanism can be designed to stop the solver of
the optimization problem whenever the Lagrangian multipliers are zero at each
observer. This should decrease the computational burden of the algorithm.
There are three key performance indicators (KPIs) considered in this work:
convergence time; voltage quality; and communication delays.
The first KPI, convergence time, is a measure for how long it takes the algorithm
to solve the voltage problem. Voltage quality reflects how well the control
algorithm can mitigate the voltage rise (or drop) problems. Finally, the
communication delays depend on the communication infrastructure. Below
follows an explanation of how each of these KPIs is quantified in practice.
Convergence Time
The voltage control algorithm is online and adjusts itself continuously. When
a change in the generation profile occurs, there are two possibilities: either
there is a voltage problem or not. If there is no voltage problem, the control
algorithm stays idle. However, if there is a voltage problem, then the agents
start to undertake action. The observer agents change the control signals until
the voltage problems are resolved. If they succeed, then the control signals
converge to a stable value, and the voltages converge to a value within the
limits. In this chapter, we define the convergence time as the time it takes
from a moment when the voltage exceeds the limits until the moment when the
voltage is restored within the limits. As demonstrated in Figure 4.12, it takes
the algorithm around 3 min to regulate the voltages within the defined limits,
which is an acceptable time for voltage problems.
It is worth mentioning that the intervention time of the interface protection relay
of Triphase inverter is much less than the convergence time. The intervention
time of the interface protection relay of Triphase inverter is less than 1 ms.
The Triphase inverter is configured to trip at 280 V. This means that there
is 27 V as a voltage margin, since the algorithm starts regulating the voltage
when the PCC voltage is higher than 253 V. Hence, the inverter in our setup is
able to correct the voltages before reaching 280 V. If an inverter trips at 253 V
(maximum voltage defined by the standard EN 50160), then v of the proposed
algorithm should be set to a value lower than 253 V (i.e., 240 V), in a way to
92 LABORATORY IMPLEMENTATION OF PEER-TO-PEER VOLTAGE CONTROL SYSTEM
make sure that the convergence time is sufficient to correct the PCC voltages
before reaching the maximum voltage at which the inverter trips.
Voltage Quality
XZ tend
E= max vymeas − v , 0 + max v − vymeas , 0 dt (4.4)
y∈Y tstart
Maximum E
voltage
Minimum
Vymeas
voltage
t
Figure 4.14: Voltage quality metric: sum of the surfaces above and below the
voltage limits.
Table 4.1 shows a comparison between the regulated and the unregulated voltage
profiles based on the voltage quality metric. E is the sum of the metrics E0 , E1
and E2 of the nodes 0, 1, and 2, respectively. The P2P voltage control reduced
the metric E from 58.724 to 2.633. E of the regulated voltage profiles is slightly
higher than zero, because it takes the algorithm some time until it has resolved
the voltage issues.
Communication Delays
For the observer agents, the delay is defined as the time between consecutive
updates of their control signals, which they broadcast periodically to the
actuator agents. For the actuator agents, the delay is defined as the time
between consecutive updates of the setpoints which are sent to the Triphase
power hardware. Figure 4.15 shows the delays between the iterations of the
control algorithm, for each agent individually.
RESULTS OF THE EXPERIMENT 93
Table 4.1: Voltage quality metrics of the regulated and unregulated voltage
profiles
Ey
vymeas E
Node 0 Node 1 Node 2
Figure 4.15: Communication delays: The observer agents update the control
signals every 1.5 s, with very little deviation. The actuator agents issue their
updates more slowly, with a significant difference between both actuator agents.
4.7 Conclusion
The testbed presented in this chapter provides realistic and pragmatic solution
for evaluating P2P smart grid applications. The testbed is used to evaluate
the performance of the P2P voltage control system experimentally. It can also
be used to implement and evaluate other P2P systems. For example, a dual-
decomposition-based P2P trading algorithm can replace the dual-decomposition-
based voltage control algorithm, to study the performance of a P2P trading
system.
Chapter 5
Robust Policy-Based
Distributed Voltage Control
Provided by PV-Battery
Inverters
95
96 ROBUST POLICY-BASED DISTRIBUTED VOLTAGE CONTROL PROVIDED BY PV-BATTERY
INVERTERS
5.1 Introduction
Related Work
but extends the voltage control policies to be used for enabling real-time
distributed voltage control. To the best of our knowledge, this work is the
first work proposing the use of linear policies for enabling real-time distributed
coordinated voltage control.
The remainder of the chapter is organised as follows: Section 5.2 presents a
general overview of the proposed voltage control system. The optimization
problem treated in this chapter is formulated in Section 5.3 based on a robust
optimization technique and linear models of control policies, voltages, inverters
and batteries. Section 5.4 demonstrates the robustness of the proposed voltage
control system and its ability to solve voltage problems. Finally, the chapter is
concluded in Section 5.5.
Control policies
Day-ahead
Real time control
p p
PV Grid voltage invi i
ppv support function q q
i invi i
Measurement
point Measurement
p point
p Inverter control li
bat i q Control
li
Smart PV-Battery Inverter Node
Battery
Household Load
Control policies
Control policies
PV-Battery
PV-Battery System
System
In the policy-based DisCVC, the linear control policies are deployed at each
individual PV-battery inverter. Once the uncertainty is revealed, the inverters
communicate with each other and apply the push-sum gossip protocol to compute
the sum of the PV power generation and households’ active power consumption
of their nodes. Each inverter then applies its control policy to compute the
reactive power and battery power needed to maintain voltages within accepted
limits.
This Section presents the optimization problem the offline coordinator solves
on a day-ahead basis.
100 ROBUST POLICY-BASED DISTRIBUTED VOLTAGE CONTROL PROVIDED BY PV-BATTERY
INVERTERS
The objective of the offline coordinator is to compute reactive power and battery
power control policies that minimize the expected cost of operating the voltage
control system while keeping voltage profiles within accepted limits; inverters
within their capacity constraints; and batteries within their energy and power
constraints. The expected cost of operating the voltage control system is the
sum of two costs: expected cost of the reactive power compensation; and
batteries degradation cost. As we want to make sure that the risk of breaching
the constraints (of voltages, inverters and batteries) is as small as possible,
the probability of violating these constraints is set on the order of 10−4 . This
results in the following stochastic optimization problem with chance constraints
(probabilistic constraints):
" #
X 2 2
min E c2qi ∆t2 kQinvi k2 + c2pi ∆t2 kPbati k2 (5.1a)
i∈I
V2 ∈ Rnt ×nnodes , ∀i ∈ I :
Pr v 2 6 V2 6 v 2 > 1 − (5.1b)
Pr pbat 6 Pbati 6 pbati > 1 − (5.1c)
i
Here, E[·] denotes the expected value operator. nt is the number of time steps
with duration ∆t (K is the time steps set). nnodes is the number of network’s
nodes. We consider two types of nodes in this chapter: control nodes where
the PV-battery inverters and loads are connected; and passive nodes with only
loads. Control nodes belong to the set I, passive nodes belong to the set J and
network’s nodes belong to N .
h iT h iT
(1) (n ) (1) (n )
Qinvi = qinvi , · · ·, qinvti , Pbati = pbati , · · ·, pbatt i denote reactive power
(k)
vector of inverter i and battery power vector of battery i, respectively. qinvi
(k)
and pbati are respectively the reactive power value (of inverter i) and active
power value (of battery i) at time step k ∈ K. The objective function (5.1a)
DESIGN OF THE ROBUST OFFLINE COORDINATOR 101
minimizes the expected total quadratic cost of reactive power injection and
battery degradation cost. cqi and cpi in (5.1a) are respectively the reactive
power compensation cost (cent/kvarh) and battery degradation cost per kWh
(cent/kWh). As shown in subsection 5.3.7, the expected value of the objective
function is approximated using a second-order moment matrix. The constraints
(5.1b)-(5.1e) are the chance constraints and is the probability of violating
these constraints.
The constraint (5.1b) is used to maintain nodal voltage magnitudes within
the maximum voltage limit v and the minimum voltage limit v. V2 =
T
(v ) , · · ·, (v(nht ) )2 denotes vectori of squared nodal voltage magnitudes,
(1) 2
(k) (k) (k) (k)
where (v )2 = (v1 )2 , · · ·, (vnnodes )2 . vx is the voltage magnitude of node
x ∈ N at time step k ∈ K. Computation of voltages is presented in subection
5.3.4.
The constraints (5.1c), (5.1d) are used to keep batteries within their energy
h iT
(1) (n )
and power limits. Ebati = Ebati , · · ·, Ebatti is the energy content vector of
(k)
battery i, where Ebati is the energy content of battery i at time step k ∈ K.
ebati and ebati are respectively the maximum and minimum energy capacity of
battery i. pbati and pbat are the maximum and minimum power of battery i,
i
respectively. The computation of the battery energy content, at different time
steps, is discussed in subsection 5.3.6.
The constraint (5.1e) is used to keep inverters within their capacity limits.
h iT
(1) (n )
Sinvi = Sinvi , · · ·, Sinvti is the apparent power vector of inverter i, where
(k)
Sinvi is the apparent power magnitude of inverter i at time step k ∈ K. sinvi and
sinvi are respectively the maximum and minimum apparent power magnitude
of inverter i. The inverter capacity model is presented in subsection 5.3.5.
The models, approximations, and reformulation needed to solve the optimization
problem (5.1a)-(5.1e) are presented in the following subsections. Trading off
modelling accuracy against computational tractability, we resort to linearized
models. Robust optimization is then used to construct a tractable optimization
problem.
(k)
pvi = ppvi + ξpvi ∆ppvi , ∀i ∈ I , ∀k ∈ K
p̃(k) (5.3)
(k) (k)
(k) (k)
q̃lx = tan (acos (PF)) p̃lx , ∀x ∈ N , ∀k ∈ K (5.4)
The uncertainty definition (5.2)-(5.4) allows the forecast error to vary between
their minimum and maximum values, leading to conservative solutions.
Therefore, budget uncertainty is applied in this chapter to control the degree
of conservatism. As we would like to formulate the constraints (5.1b)-(5.1e)
as linear constraints, a polyhedral uncertainty set Ξ with budget Γ is used to
model the uncertainty [169].
n o
U = P̃l,pv ∈ Rnt (nnodes +nc ) |P̃l,pv = Pl,pv + I∆P ξ l,pv (5.5)
Where ξ l,pv ∈ Ξ , ∀k ∈ K :
h iT h iT
(1) (nt ) (1) (nt )
P̃l,pv = P̃l,pv , · · · , P̃l,pv , Pl,pv = Pl,pv , · · ·, Pl,pv ,
h i
(k) (k) (k)
P̃l,pv = p̃l1 , · · · , p̃ln , p̃(k) (k)
pv1 , · · ·, p̃pvn ,
nodes c
DESIGN OF THE ROBUST OFFLINE COORDINATOR 103
h i
(k) (k) (k)
Pl,pv = pl1 , · · ·, pln , p(k) (k)
pv1 , · · ·, ppvn ,
nodes c
h iT
(1) (nt )
I∆P = diag [∆Pl,pv ] = diag ∆Pl,pv , · · · , ∆Pl,pv ,
h i
(k) (k) (k)
∆Pl,pv = ∆pl1 , · · · , ∆pln , ∆p(k) (k)
pv1 , · · · , ∆ppvn ,
nodes c
n o
Ξ = ξ l,pv ∈ Rnt (nnodes +nc ) |kξ l,pv k1 6 Γ ,
h iT
(1) (nt )
ξ l,pv = ξ l,pv , · · · , ξ l,pv ,
h i
(k) (k) (k)
ξ l,pv = ξl1 , · · · , ξln (k)
, ξpv 1
(k)
, · · · , ξpv n
nodes c
Here, nc is the number of control nodes. The operator diag [∆Pl,pv ] creates
a square diagonal matrix with the elements of the maximum deviation vector
∆Pl,pv on the main diagonal. The vector P̃l,pv includes the uncertain households’
active power consumption and PV power generation for nt time steps. The
(k)
elements of P̃l,pv are defined in (5.2), (5.3). The households’ active power
consumption and PV power generation forecast data for nt time steps are
contained in the vector Pl,pv .
The convention in this chapter is that the power is positive when it is consumed
(k)
and negative when it is injected. In our simulations, the elements of P̃l,pv and
(k) (k)
Pl,pv are positive. Hence, a minus sign will be introduced in front of p̃pvi and
(k)
ppvi .
(k)
X (k)
qinvi = L(k)
qi p̃li − p̃(k)
pvi (5.6)
i∈I
(k)
Where Lqi is the coefficient of the reactive power policy of inverter i at time
step k, contained in the diagonal matrix Lqi ∈ Rnt ×nt . The vector P̃∆ includes
the total summation of households’ active power consumption and PV power
(k)
generation of all control nodes at each time step. With Lpi as the coefficient
of battery power policy (of battery i at time step k) contained in the diagonal
matrix Lpi ∈ Rnt ×nt , the battery power policy can be defined as:
In this chapter, the linear branch flow model shown in (5.9) is applied. The
model is based on the DistFlow method developed in [173]. According to [174],
the linearized branch flow model tends to introduce a small relative error of
1-5% when used for calculating power flows of real distribution networks.
(k) 2
(v ) = RP(k) + XQ(k) + vo2 1nnodes , ∀k ∈ K (5.9)
node, then their active and reactive power can be defined as:
(k) (k) (k)
pi = p̃li − p̃(k)
pvi + pbati
∀i ∈ I (5.10)
(k) (k) (k)
qi = q̃li + qinvi
(k) (k)
pj = p̃lj
∀j ∈ J (5.11)
(k) (k)
qj = q̃lj
The magnitude of the inverter’s apparent power at time step k can be given by
the following equation:
q
(k) (k) (k)
Sinvi = (pinvi )2 + (qinvi )2 , ∀i ∈ I , ∀k ∈ K (5.12)
Here, ηich and ηidis are respectively the charge and discharge efficiency of battery
+
i. The operator [·] ≡ max(·, 0) introduces integer variables to the constraint
(5.1d), which results in a mixed integer chance constraint. As mixed integer
chance constraints lead to a high computational complexity, a heuristic approach
is proposed in [176] to get rid of integer variables of a battery model providing
frequency containment reserve. In this chapter, we adapt the heuristic approach
to be used for a battery model providing voltage control:
!
+ +
(k+1) (k) (k) (k)
Ebati = Ebati + ∆t pbati (k) − − pbati (k) , (5.16)
η η
(k+1) (k) (k)
Ebati = Ebati + ∆t pbati ,
η (k)
(k) (k)
(
(k)
ch (k)
ηi pbati if pli 6 ppvi
pbati = (k) (k)
η (k)
1 (k)
p
η dis bati
if pli > ppvi
i
The if condition in (5.16) is based on the fact that the controller tends to
discharge the battery when the load is higher than the PV generation (to
mitigate voltage drop) and charge the battery when the PV generation is higher
than the load (to mitigate voltage rise). Our simulation results show that this
heuristic approach does not lead toviolation of the constraint
(5.1d).
In fact,
(k) (k) (k)
(5.16) is the same as (5.15) if sign p̃li − p̃pvi = −sign pbati . To include
the initial energy content Ebat
0
i
, we can reformulate (5.16) as:
k−1
X
(k) (t)
Ebati = Ebat
0
i
+ ∆t pbati , k = 2, 3, · · · , nt (5.17)
η (t)
t=1
(1)
Ebati = Ebat
0
i
Where creplacementi is the battery replacement cost, Ncycle life is the battery cycle
life and DoD is the depth-of-discharge. The denominator of (5.18) gives the
charging and discharging kWh the battery can tolerate over its serviceable life.
Based on the defined linear control policies, the objective function (5.1a) can
be reformulated as:
" #
X 2 2
E c2qi ∆t2 kQinvi k2 + c2pi ∆t2 kPbati k2 = (5.19)
i∈I
X h 2
i h 2
i
c2qi ∆t2 E kLqi P̃∆ k2 + c2pi ∆t2 E kLpi P̃∆ k2
i∈I
h 2
i
E kLpi P̃∆ k2 = E Tr LTpi Lpi P̃∆ P̃∆T = Tr LTpi Lpi M
i∈I
Here, the matrix P̃∆scen ∈ Rnscen ×nt contains nscen scenarios of the random vector
P̃∆ . P̃∆ ,mean ∈ R1×nt is the sample mean of P̃∆scen , and cov P̃∆scen ∈ Rnt ×nt is
Based on the previous linear models, one can show that the chance constraints
(5.1b)-(5.1e) can be written as:
By correct design of Ξ, the solution of (5.23) can ensure that the probability in
(5.22) is larger than or equal to 1 − . Unfortunately, the probability guarantee
is only applicable to independent variables with zero mean, which is not the
case when considering ξ l,pv . However, one can obtain independent variables
with zero mean by applying a whitening transformation [179]:
ξ indp = l,pvmean = W ξ l,pv
zero mean
(5.24)
l,pv W ξ l,pv − ξ
b = APl,pv + BI ξ
A ,
∆P l,pvmean
n o
Ξ = ξ indp
l,pv ∈ R nt (nnodes +nc )
|kξ indp
k
l,pv 1 6 Γ
Using standard duality techniques, it can be shown that the constraint (5.25)
with a polyhedral uncertainty set can be reformulated as linear constraints
CASE STUDY 109
(5.26b)-(5.26d). We ask the readers to refer to property 4.3 in [169] for details
about robust counterpart of linear constraints with polyhedral uncertainty set.
For more robust approaches on BESSs and DERs, we ask the readers to refer to
[180]. Below is the complete optimization problem the robust offline coordinator
solves on a day-ahead basis:
X
min c2qi ∆t2 Tr LTqi Lqi M + c2pi ∆t2 Tr LTpi Lpi M (5.26a)
i∈I
aϕ + Γ λϕ 6 cϕ (5.26b)
λϕ > I ∆P W −1 µϕ (5.26c)
− µϕ 6 bTϕ 6 µϕ (5.26d)
Group G3
16 14 15
13
Control node
i V VI
Passive node
VII
IV
7 8 9 10 11 12 17 18 19 20 21 22
Grid Group G2 Group G4
II
I
III VIII
1 2 3 4 5 6 23 24 25 26 27 28 29 30 31
Group G1 Group G5
Figure 5.2: Schematic diagram of the network used in this case study.
VII belong to group G4 and between IV and VIII belong to group G5. The
parameters of each group are listed in Table 5.1.
To comply with the European standard EN 50160 on power quality [3], the
voltage limits v and v are enforced to be ±10% of the nominal phase voltage
230 V, resulting in v= 253 V and v = 207 V. The probability of violating
voltages, inverters and batteries constraints is set to 10−4 . The number of
linear constraints ns used to approximate the conic constraint (5.13) is set to
10 (relative error = 0.0123).
G1 G2 G3 G4 G5
sinv (kVA) 5 8 10 12 15
Ebat
0
(kWh) 2.4 3.6 4.1 4.9 6.75
DoD (%) 80 80 80 80 80
the reader to refer to [179] for details about generating scenarios of random
dependent variables using Cholesky decomposition.
365 daily PV profiles (with one-minute resolution) are generated for each control
node, based on the model presented in [184]. The generated PV profiles are
considered as the forecasted PV profiles. The maximum PV forecast error is set
to 40 %. 100 PV scenarios for each control node for each day of the simulation
are generated based on eq. (5.3). The random dependent variables in (5.3)
are generated based on the covariance matrix of the normalized forecasted PV
profiles and the Cholesky decomposition.
112 ROBUST POLICY-BASED DISTRIBUTED VOLTAGE CONTROL PROVIDED BY PV-BATTERY
INVERTERS
Forecasted
data
Push-sum MATPOWER
gossip
Summations of
households load
and PV power Voltage
profiles
Active and reactive
Control policies power of voltage control
Figure 5.3: General overview of the MATLAB simulation used in the case study.
one-minute time step is considered for each day. One control policy is considered
for each 15 minutes simulation in MATPOWER.
First, the voltage profiles are computed for the 36500 scenarios without applying
the linear control policies. Figure 5.4 shows voltage profiles of control node 31
(phase to neutral voltage). One can see that most of voltage profiles of different
scenarios exceed the maximum voltage limit. When the linear control policies
are applied, it is noticed that the existing violations of the over-voltage limit
are eliminated for the 36500 scenarios, which demonstrates the robustness of
the proposed policy-based DisCVC. Figure 5.5 shows voltage profiles of control
node 31 (at the end of the feeder) after applying the policies.
Figure 5.6a and Figure 5.6b show, respectively, the apparent power magnitude
and reactive power of inverter 31 for the 36500 scenarios. Figure 5.6c and
Figure 5.6d show the energy content and charge/discharge power of battery 31
for all the scenarios. One can see that limits of the battery and inverter are
respected. As the cost of reactive power compensation cq is cheaper than the
cost of battery degradation cp , we notice that the reactive power of inverter 31
used in regulating voltage profiles is higher than the charge/discharge power of
battery 31.
(a) (b)
(c) (d)
Figure 5.6: (a) Apparent power magnitude of inverter 31, (b) Reactive power
of inverter 31, (c) Energy content of battery 31, (d) Charge/discharge power of
battery 31.
CONCLUSION 115
5.5 Conclusion
Simultaneous Provision of
Voltage and Frequency
Control by PV-Battery
Systems
117
118 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
as a case study. Simulations over 104 scenarios are used to demonstrate the
robustness of the proposed control methodology. The contents of this chapter
are based upon the journal paper:
6.1 Introduction
6.1.1 Motivation
Smart inverters with advanced grid support mechanisms and data communica-
tion capabilities are being deployed to tackle the negative impact of increased
PV penetration on grid voltage [188]. These inverters have the ability to absorb
(or inject) reactive power and curtail PV power to maintain grid voltage within
accepted limits. Behind-the-meter battery energy storage systems (BESSs)
can also play an important role in regulating voltage profiles of distribution
networks with high PV penetration [31]. BESSs can be paired with smart PV
inverters to control their energy output in response to violation of grid voltage
limits. Smart inverters, for example, can charge BESSs during PV peak period
to mitigate voltage rise.
To enable PV-battery inverters to actively participate in voltage regulation,
various voltage control strategies of different complexity and data transfer needs
have been proposed in literature. As has been discussed in the previous chapters,
one of the effective control strategies that enables smart PV-battery inverters
to regulate grid voltage is distributed coordinated voltage control (DisCVC)
[40]. In DisCVC, smart inverters communicate with each other in a peer-to-peer
(P2P) fashion to solve a voltage control problem in a distributed way without
relying on a central decision-making controller.
As shown in Figure 6.1, when BESSs are used to support PV inverters in
regulating voltage profiles, they are inactive or partially used most of the time.
Their unused capacity could be used to provide other services, allowing them
to generate additional revenues. Since frequency control has been identified
as one of the highest value services for BESSs [189], we are interested in
combining voltage control with frequency containment reserve (FCR) (also
referred to as primary frequency control). Moreover, it can be a necessity for
PV-battery systems that provide frequency control to provide voltage control
as well; providing frequency control services with several BESSs connected to
INTRODUCTION 119
Figure 6.1: The results of this figure are obtained based on the implementation
of the policy-based DisCVC proposed in Chapter 5, in a simulation over 104
scenarios. The scenarios are presented in Subsection 6.4.2. The DisCVC is used
to regulate voltage profiles of the 62-node network presented in Subsection 6.4.1.
This figure presents the charging power of BESS number 31.
p(k)
FCR
= r∆f (k) , (6.1a)
(k)
−fnom
f
if ∆fdb <| f (k) − fnom |< ∆f ,
∆f
1 if f (k) − fnom ≥ ∆f ,
∆f (k) = (6.1b)
−1
if f (k) − fnom ≤ −∆f ,
0 otherwise.
Here, k is the time step of the frequency control, ∆f is the frequency deviation
at which maximum FCR capacity r needs to be activated, and ∆fdb is the
frequency deadband in which no FCR reaction is required. In the Continental
Europe (CE) synchronous area, ∆f = 200 mHz and ∆fdb = 10 mHz.
Non-delivery or non-availability of a committed FCR reserve capacity will result
in penalties charged by the TSO. Providers can lose all revenues made from
FCR in case of not meeting the FCR requirements, or can be excluded from
further participation in the market in case delivery is consistently not available
or insufficient.
As BESSs have proven great potential in providing FCR [37], many TSOs
opened their networks for BESSs and released new FCR regulations for assets
with limited energy reservoirs [192]. BESSs are expected to entirely take over
the role of FCR providers from thermal power plants. Nevertheless, designing
a control system that enables BESSs to provide FCR is not straightforward.
Due to the limited capacity of batteries, ensuring the FCR reserve capacity r is
available during the contracted period is not possible without a management
strategy that keeps energy content of BESSs within limits. This is because the
frequency signal has a non-zero energy content over short time periods, and
efficiency losses of batteries decrease their energy content when being charged
or discharged. Figure 6.2 demonstrates the fact that a BESS cannot be used to
provide FCR without an energy management strategy. In Figure 6.2, a 13.5
kWh residential battery is used to provide FCR based on (6.1a). Two years of
frequency data (2017-2018) with 1 minute resolution, from the CE synchronous
region [193], are used to study the energy content of the 13.5 kWh battery
while providing FCR. One can clearly notice the violation of the maximum and
minimum energy content limits. Hence, an energy management strategy is a
must.
1 The convention in this chapter is that the power is positive when it is consumed (e.g.,
charging power of the BESS) and negative when it is injected (e.g., discharging power of the
BESS).
INTRODUCTION 121
Figure 6.2: Energy content of the 13.5 kWh battery over 730 frequency
scenarios. Battery parameters: maximum
√ power= 7 kW, initial charge= 6.5
kWh, charge/discharge efficiency=
0.9. The energy content E is calculated
(k) + (k) +
based on: E (k+1)
= E + ∆t ηi pFCR − ηdis −pFCR
(k) ch 1
, where ηich and
i
ηidis are, respectively, the charge and discharge efficiencies , ∆t is the time step
+
duration and [·] ≡ max(·, 0).
The German TSOs have defined control strategies [194], also referred to as
degrees of freedom (Dof), which can be utilized to keep energy content of
batteries within operational range during FCR provision. These strategies
presume that BESSs can deviate slightly from the required FCR Power. These
strategies are: 1) overdelivery: providers are allowed to provide FCR power that
is anywhere between 100% and 120% of the instantaneous FCR requirement,
2) deadband utilization: usually FCR assets are not obliged to activate their
reserve capacity within the range ±10 mHz around the nominal frequency, a
battery can use this
deadband for charging and discharging by using a power
within the range 0, p(k) in case of charging, or within the range , 0 in
(k)
FCR
−p FCR
case of discharging. A comparison of these control strategies is made in [194],
in which authors conclude that overdelivery and deadband utilization are not
sufficient to regulate the energy contents of BESSs within limits while providing
FCR. Providers, who wish to use these strategies, should combine them with an
additional management strategy, or use a single battery management strategy
that can regulate energy content of batteries within limits while providing FCR.
BESSs create value for prosumers and grid operators, but leave significant
untapped value on the table. Currently, most BESSs are deployed for one of
122 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
VC battery p
VC, bat i
Household load
R-UP battery p
R-UP, bati
p
R-DN, bati
R-DN battery
Control policies,
Control policies, reserve capacity
reserve capacity Control node
Control node
local information yields to a poor performance [26], each voltage control policy
is designed as a linear function of the local active powers of the household’s
load and PV installation, and active powers of the households’ loads and PV
installations connected to the nodes participating in voltage control (control
nodes). Each GVSF needs to know the sum of the households’ active power
consumption and PV power generation of the nodes participating in voltage
control, to be able to compute its own control actions. To compute the sum
of the households’ active power consumption and PV power generation in a
fast and efficient way, a P2P-based push-sum gossip protocol is applied in this
chapter [90]. The push-sum gossip protocol enables the PV-battery inverters to
perform a distributed summation with a moderate communication overhead.
The FSF computes the upward-FCR power of the R-UP battery, and the
downward-FCR power of the R-DN battery. The FSF hosts two linear
management policies: an R-UP battery management policy that maintains
energy content of the R-UP battery within limits while providing upward
frequency regulation; and an R-DN battery management policy that maintains
CONTRIBUTION AND GENERAL OVERVIEW OF THE PROPOSED CONTROL FRAMEWORK 125
energy content of the R-DN battery within limits while providing downward
frequency regulation.
deviation, whereas in the downward-FCR, batteries are charged in response to the positive
frequency deviation.
3 It is worth to point out that the priority of the control system is to solve voltage rise
(a)
(b)
Figure 6.4: (a) Unregulated voltage profile of a prosumer in a network with high
PV penetration, (b) Expected upward and downward reserve capacity profile.
In this example, the inverter reactive power and the VC battery power are used
to solve the voltage rise problem, therefore, the upward reserve capacity is zero
in Area 2.
battery4 cannot be charged and discharged at the same time. Since there is no
voltage rise problem in Area 1 and Area 3, the R-UP battery can be discharged
in these areas to provide upward-FCR. The voltage in Area 1 is closer to the
maximum limit than the voltage in Area 3, therefore, the upward-FCR capacity
in Area 1 is expected to be less than the upward-FCR capacity in Area 3.
As there is no voltage drop problem in the example of Figure 6.4, the R-DN
battery can be charged in the three areas to provide downward-FCR. The
downward-FCR capacity depends on the amount of reactive power capacity
4 In this chapter, if the word “battery” is not preceded by the three abbreviations: VC;
R-UP; and R-DN, then this word means the real battery and not one of the three virtual
batteries.
PROBLEM FORMULATION 127
that can support the R-DN battery to absorb more FCR power without causing
the network violating voltage operation limits. In Area 2 of Figure 6.4a, the
downward-FCR capacity is expected to be at minimum, since most of the
inverter capacity is expected to be occupied by the PV power and the reactive
power of the voltage control system. The downward-FCR capacity in Area 3 is
expected to be less than the downward-FCR capacity in Area 1. This is because
the voltage in Area 3 is closer to the minimum voltage limit than the voltage in
Area 1.
For sake of simplicity, the power capacity reserved for the energy management
policy of the example of Figure 6.4b is set to a constant value of 1 kW. In the
proposed optimization problem, the power capacity of the energy management
policy is considered as a variable capacity to compute the minimum sufficient
power capacity. It is clear from Figure 6.4b that the energy management policy
will have to be designed carefully, as more power capacity for the management
policy will mean that less FCR capacity can be sold to the TSO, while on the
other hand the power capacity for the management policy should be sufficient
to ensure the energy content of the battery remains within limits.
The sources of uncertainty in the control problem of the offline controller include
households’ active and reactive power consumption, PV power generation, and
(k)
the grid frequency. The uncertain household’s active power consumption p̃lx ,
(k)
and the uncertain PV power generation p̃pvi can be bounded as:
(k) (k)
0 ≤ p̃lx ≤ plx
∀x ∈ N , ∀k ∈ K (6.2)
(k) (k) (k)
plx = plx + ∆plx
0 ≤ p̃(k) (k)
pvi ≤ ppvi
∀i ∈ I, ∀k ∈ K (6.3)
pvi = ppvi + ∆ppvi
p(k) (k) (k)
(k) (k)
Where plx and ppvi are, respectively, the maximum uncertain household’s
active power consumption, and the maximum uncertain PV power generation.
(k)
plx is defined in (6.2) as the sum of the forecasted household’s active power
(k) (k)
consumption plx ∈ R≥0 and the maximum deviation ∆plx from the forecast of
(k)
the household’s active power consumption. In (6.3), ppvi is defined as the sum
(k)
of the forecasted PV power generation ppvi ∈ R≥0 and the maximum deviation
(k) (k) (k)
∆ppvi from the PV forecast. ∆plx and ∆ppvi can be approximated based on
historical data of forecast error.
(k)
The uncertain household’s reactive power consumption q̃lx can be defined as
(k)
function of p̃lx and a power factor (PF):
(k) (k)
q̃lx = tan (acos (PF)) p̃lx , ∀x ∈ N , ∀k ∈ K (6.4)
(k)
A normalized positive frequency deviation ∆f
f
R-DN (related to the downward-
(k)
FCR), and normalized negative frequency deviation ∆f f
R-UP (related to the
upward-FCR) can be bounded as:
(k) (k)
0 ≤ ∆f
f
R-DN ≤ ∆f R-DN , ∀k ∈ K (6.5a)
(k)
∆f (k)
R-UP
≤ ∆f
f
R-UP ≤ 0, ∀k ∈ K (6.5b)
(k)
Here, ∆f R-DN and ∆f (k)
R-UP
are the maximum normalized frequency deviation
and minimum normalized frequency deviation at time step k, respectively. In
PROBLEM FORMULATION 129
(k)
this chapter, ∆f R-DN and ∆f (k)
R-UP
are set based on historical frequency data.
(k)
One can set ∆f R-DN to 1, and ∆f (k)
R-UP
to −1 ∀k ∈ K, which is correct. However,
from our experience, this will lead to conservative solutions, e.g., low contracted
reserve capacity. As shown in Figure 6.5, the historical maximum normalized
frequency deviation can be much less than 1, and the historical minimum
normalized frequency deviation can be much higher than −1.
Figure 6.5: Normalized positive and negative frequency deviation ∆f of 730 days
of historical frequency data (2017-2018). The dotted lines show the maximum
and minimum normalized frequency deviations. ∆f is calculated based on
(6.1b). The frequency data are from the CE synchronous region.
(κ)
∆f (k) = (6.6)
(k) (κ)
(f ) −fnom
if ∆fdb <| (f (k) )(κ) − fnom |< ∆f ,
∆f
1 if (f (k) )(κ) − fnom ≥ ∆f ,
−1 if (f (k) )(κ) − fnom ≤ −∆f ,
0 otherwise.
κ = 1, · · · , nf ,
!
(k)
(1) + (nf ) +
∆f R-DN = max ∆f (k)
,··· , ∆f (k)
,
130 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
!
(1) − (nf ) −
∆f (k)
R-UP
= min ∆f (k)
,··· , ∆f (k)
(κ)
Here, ∆f (k) is the normalized frequency deviation number κ related to
the frequency measurement number κ (f (k) )(κ) at time step k. The operator
+ −
[·] ≡ max(·, 0), and [·] ≡ min(·, 0).
To get immunity against service unavailability, and violating operational
constraints, we look for solutions which are feasible for any realization of the
uncertain data in a predefined uncertainty set U. In this chapter, we consider a
polyhedral description of the uncertainty set of the form:
U = {ζ ≥ 0 : Hζ ≤ h} , (6.7)
H ∈ R2nζ ×nζ and h ∈ R2nζ can be defined to represent the constraints (6.2)-
(6.5b), where nζ = nt (2 + nc + nnodes ) is the number of uncertain variables for nt
times steps. The vector ζ ∈ Rnζ includes the uncertain households’ active power
consumption, PV power generation, normalized positive frequency deviations,
and normalized negative frequency deviations for nt time steps:
h (1) (nt ) (1) (nt )
ζ = P̃ l , · · · , P̃ l , P̃ pv , · · · , P̃ pv , (6.8a)
(1) (n ) (1) (n ) iT
∆f f t f t
R-DN , · · · , ∆f R-DN , ∆f R-UP , · · · , ∆f R-UP
f f
(k)
h i
(k) (k)
P̃ l = p̃l1 , · · · , p̃ln , ∀k ∈ K (6.8b)
nodes
(k)
h i
P̃ pv = p̃(k)
pv1 , · · · , p̃(k)
pvn , ∀k ∈ K (6.8c)
c
set. The goal of robust optimization is to find solutions that are immune to
uncertainty. In static robust optimization, decisions must be made before the
realization of the uncertain data. Ben-Tal et al. extended the robust optimization
framework to dynamic settings by proposing the adjustable robust optimization
(ARO) technique [170], in which some decision variables are allowed to be
computed after the realization of the uncertain data, thus leading to better
objective value and less conservative solutions. This performance is achieved
at higher computational burden, as ARO is usually NP-hard to solve. Ben-Tal
et al. proposed the method of affinely adjustable robust optimization (AARO)
to construct a tractable optimization problem. In AARO, decision variables
are restricted to be affine functions of the uncertain data [202]. Accordingly,
control policies proposed in this chapter are designed as linear functions of the
uncertain data.
As has been discussed in subsection 6.2.1, we consider four linear control policies
for each PV-battery system: reactive power policy that gives reactive power
setpoints; active power policy that gives charge/discharge power setpoints of
the VC battery; R-UP battery management policy that gives charge power
setpoints needed to maintain the energy content of the R-UP battery within
limits while providing upward-FCR; and R-DN battery management policy that
gives discharge power setpoints needed to maintain the energy content of the
R-DN battery within limits while providing downward-FCR.
To regulate voltage profiles, each PV-battery system adjusts its reactive power
and VC battery power linearly as function of control nodes’ active power
consumption and PV power generation, and normalized frequency deviations.
Given the possibility that the frequency deviation can be either positive or
negative for each time step, we propose two types of reactive power policy:
reactive power policy that takes into account the effect of responding to positive
frequency deviations; and reactive power policy that takes into account the
effect of responding to negative frequency deviations:
∀i ∈ I, ∀k ∈ K :
(k)
X (k)
(k) (k)
q̃R-DN, invi = βp(k)
i pvi + βR-DNi ∆f R-DN
p̃li − p̃(k) f (6.9a)
i∈I
(k)
X (k)
(k) (k)
q̃R-UP, invi = βp(k)
i pvi + βR-UPi ∆f R-UP
p̃li − p̃(k) f (6.9b)
i∈I
132 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
(k) (k)
where q̃R-DN, invi and q̃R-UP, invi are the uncertain reactive powers of inverter
i (at time step k) that take into account the effect of downward-FCR and
(k)
upward-FCR, respectively, on mitigating voltage problem. βpi is the coefficient
of the reactive power policy (of inverter i at time step k) related to the active
powers of households’ loads and PV installations connected to the control nodes.
(k) (k)
βR-DNi and βR-UPi are, respectively, the coefficients of the reactive power policy
related to the positive and negative normalized frequency deviations at time
step k. The two types of the reactive power policy can be defined in a vectorized
form as:
∀i ∈ I :
h iT
(1) (nt )
Q̃R-DN, invi = q̃R-DN, invi , · · · , q̃R-DN, invi
T
(1) (n )
ξ R-DN = ∆f
f , · · · , f t
∆f ,
R-DN R-DN
T
(1) (nt )
ξ R-UP = ∆f R-UP , · · · , ∆f R-UP
f f
(k)
where β pi , β R-DNi and β R-UPi are diagonal matrices ∈ Rnt ×nt with the βpi ,
(k) (k)
βR-DNi and βR-UPi coefficients (∀k ∈ K) contained in their diagonal elements,
respectively. When the frequency deviation is positive, the real-time controller
uses the reactive power policy (6.9a), whereas when the frequency deviation is
negative, the real-time controller uses the reactive power policy (6.9b). The two
policies are equal when the frequency deviation is zero, or within the deadband
The active power policy that is responsible for charging and discharging VC
batteries can be defined as:
h iT
(1) (nt )
P̃ VC, bati = p̃VC, bati , · · · , p̃VC, bati = Z pi ξ ∆p , ∀i ∈ I (6.11)
PROBLEM FORMULATION 133
(k)
where p̃VC, bati is the uncertain active power of the VC battery i at time step k.
(k)
Z pi is a diagonal matrix ∈ Rnt ×nt , with zpi (∀k ∈ K) the coefficients of the
active power policy contained in its diagonal elements.
For each time step k, the coefficients of the reactive and active power policies
are calculated considering the following cases:
2. In the case the offline controller predicts a voltage drop problem: the
same logic as above is applied.
3. In the case the offline controller predicts no voltage problem: the upward-
FCR and downward-FCR capacities are maximized considering constraints
related to batteries, inverters and voltages. The offline controller sets
(k) (k) (k)
βpi and zpi to zero. βR-UPi is calculated to avoid any voltage rise
(k)
problem that may arise from discharging R-UP batteries, whereas βR-DNi
is calculated to avoid any voltage drop problem that may arise from
charging R-DN batteries.
k−1 (k)
(k)
X
(κ) (κ)
p̃m-DN, bati = mR-DNi ∆f
f
R-DN , ∀i ∈ I, ∀k ∈ K
κ=1
(k)
h iT
(1) (nt )
P̃ m-DN, bati = p̃m-DN, bati , · · · , p̃m-DN, bati
(k)
(κ)
with mR-DNi , κ = 1, · · · , k − 1 (∀k ∈ K), the coefficients of the R-
DN battery management policy i, contained in the lower triangular matrix
M R-DNi ∈ Rnt ×nt with zeros on the diagonal.
(k)
(κ)
With mR-UPi , κ = 1, · · · , k − 1 (∀k ∈ K), as the coefficients of the R-
UP battery management policy i, contained in the lower triangular matrix
M R-UPi ∈ Rnt ×nt with zeros on the diagonal, the power of the R-UP battery
(k)
management policy p̃m-UP, bati (∀k ∈ K) can be defined as:
(k)
h iT
(1) (nt )
P̃ m-UP, bati = p̃m-UP, bati , · · · , p̃m-UP, bati
As shown in Figure 6.6, the offline controller allocates fractions of the energy
and power capacity of each BESS to the voltage and frequency control services.
For each time step k, the energy content of each battery is divided into three
parts that provide upward-FCR, downward-FCR, and voltage control. The
constraints (6.14a)-(6.14h) allocate the maximum, minimum, and initial energy
capacity for each part:
∀k ∈ K, ∀i ∈ I
Stochastic constraints:
(k) (k) (k)
eVC, bati 6 ẼVC, bati 6 eVC, bati (6.14a)
pbat e bat i
i
(k)
r DN
Charge power
(k) (k)
i Downward-FCR Downward-FCR e R-DN, bat i e R-DN, bat i
(k)
pm-UP, bat
i R-UP management policy R-DN management policy
p (k)
VC, bat i Voltage rise
Voltage rise
(k) (k) (k)
pVC, bat
0 e VC, bat i e VC, bat
i Voltage drop i
Voltage drop
(k)
Discharge power
r (k)
UP Upward-FCR Upward-FCR (k) (k)
i e R-UP, bat i e R-UP, bat i
p bat i e bat i
Power capacity allocation Energy capacity allocation
Other constraints:
(k) (k) (k)
e VC, bati ≥ 0, e R-UP, bati ≥ 0, e R-DN, bati ≥ 0 (6.14d)
(k) (k)
UP battery i, and R-DN battery i at time step k. e VC, bati , e R-UP, bati and
(k)
e R-DN, bati denote minimum energy capacities of the VC battery i, R-UP
battery i, and R-DN battery i at time step k. Ebat 0
i
, E 0VC, bati , E 0R-UP, bati
and E R-DN, bati denote the initial energy contents of the battery i, VC battery
0
∀k ∈ K, ∀i ∈ I
Stochastic constraints:
(k) (k)
p(k)
VC, bat
6 p̃VC, bati 6 pVC, bati (6.15a)
i
(k) (k)
0 6 p̃m-UP, bati 6 pm-UP, bati (6.15b)
(k)
p(k)
m-DN, bat
6 p̃m-DN, bati 6 0 (6.15c)
i
Other constraints:
(k) (k)
rUPi ≥ 0, rDNi ≥ 0 (6.15d)
(k)
p(k)
VC, bat
− rUPi + p(k)
m-DN, bat
≥ pbat (6.15e)
i i i
FCR. In (6.15a), the power of the VC battery varies between negative power
(k) (k)
capacity pVC, bati (discharge power), and positive power capacity pVC, bati
(charge power). The VC battery can be charged to reduce the reverse power
flow from PV installations during PV peak period, which helps mitigate voltage
rise. During load peak period, the VC battery can be discharged to feed a local
load, which helps mitigate voltage drop.
We consider a simple discrete battery model for the three virtual batteries:
1 h (k)
h i+ i+
(k+1) (k) (k)
ẼVC, bati = ẼVC, bati + ∆t ηi p̃VC, bati − dis −p̃VC, bati
ch
(6.16a)
ηi
1 (k)
(k+1) (k) ch (k)
ẼR-DN, bati = ẼR-DN, bati + ∆t ηi p̃R-DNi + dis p̃m-DN, bati ,
ηi
1 (k)
(k+1) (k) (k)
ẼR-UP, bati = ẼR-UP, bati + ∆t ηich p̃m-UP, bati + dis p̃R-UPi ,
ηi
Here, ηich and ηidis are, respectively, the charge and discharge efficiencies of
(k) (k)
battery i. p̃R-DNi and p̃R-UPi are, respectively, the uncertain downward-FCR
+
power and upward-FCR power at time step k. In (6.16a), the operator [·] ≡
max(·, 0) introduces integer variables to the stochastic constraint (6.14a), which
results in a mixed integer stochastic constraint. As mixed integer stochastic
constraints lead to a high computational complexity, a heuristic approach is
proposed in [176] to get rid of integer variables of a battery model providing
FCR. In this chapter, we adapt the heuristic approach to be used for a battery
model providing voltage control:
+ +
(k+1) (k) (k) (k)
ẼVC, bati = ẼVC, bati + ∆t p̃VC, bati − − p̃VC, bati ,
η (k) η (k)
(k+1) (k) (k)
ẼVC, bati = ẼVC, bati + ∆t p̃VC, bati , (6.17)
η (k)
PROBLEM FORMULATION 139
The if condition in (6.17) is based on the fact that the controller tends to
discharge the battery when the load is higher than the PV generation (to
mitigate voltage drop) and charge the battery when the PV generation is higher
than the load (to mitigate voltage rise). Our simulation results show that
this heuristic approach does not lead to violation of the constraint
(6.14a).
In
(k) (k) (k)
fact, (6.17) is the same as (6.16a) if sign pli ,meas − ppvi ,meas = −sign pbati ,
(k) (k)
where pli ,meas and ppvi ,meas are, respectively, the real-time measured active
power consumption and PV power generation at time step k.
Degradation Cost
Here, cpi is the battery degradation cost per kWh (cent/kWh), creplacementi
is the battery replacement cost, Ncycle life is the battery cycle life, and DoD
is the depth-of-discharge. The denominator of (6.18) gives the charging and
discharging kWh the real battery can tolerate over its serviceable life.
The active and reactive power passing through inverter i is limited by its
apparent power capacity:
q
(k) (k)
(p̃R-DN, invi )2 + (q̃R-DN, invi )2 6 sinvi , ∀i ∈ I, ∀k ∈ K,
(k) (k) (k)
p̃R-DN, invi = −p̃(k)
pvi + p̃R-DNi + p̃VC, bati +
(k) (k)
p̃m-DN, bati + p̃m-UP, bati (6.19a)
q
(k) (k)
(p̃R-UP, invi )2 + (q̃R-UP, invi )2 6 sinvi , ∀i ∈ I, ∀k ∈ K,
140 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
(k) (k) (k)
p̃R-UP, invi = −p̃(k)
pvi + p̃R-UPi + p̃VC, bati +
(k) (k)
p̃m-DN, bati + p̃m-UP, bati (6.19b)
(k) (k)
− sinvi 6 cos(ϑφ)p̃R-UP, invi + sin(ϑφ)q̃R-UP, invi 6 sinvi ,
π
φ= , ϑ = 1, · · · , nS (6.20)
nS
The reactive power cost cqi (cent/kvarh), considered in this chapter, corresponds
to the approximative compensation cost of the additional inverter losses due to
the reactive power utilization and is provided by [175].
In this chapter, a linear branch flow model is applied. The model is based on
the DistFlow method developed in [173]. According to [174], the linearized
branch flow model tends to introduce a small relative error of 1-5% when used
PROBLEM FORMULATION 141
∀k ∈ K :
(k) (k)
(ṽ (k) )2R-DN = RP̃I + X Q̃I + vo2 1nnodes (6.21a)
(k)
h (k) (k)
i (k) h (k) (k)
i
P̃I = P̃ R-DN, node ; P̃ passive , Q̃I = Q̃R-DN, node ; Q̃passive
(k) (k)
(ṽ (k) )2R-UP = RP̃II + X Q̃II + vo2 1nnodes (6.21b)
(k)
h (k) (k)
i (k) h (k) (k)
i
P̃II = P̃ R-UP, node ; P̃ passive , Q̃II = Q̃R-UP, node ; Q̃passive
(k) (k)
p̃passivej = p̃lj
∀j ∈ J (6.23)
(k) (k)
q̃passivej = q̃lj
As has been discussed in subsection 6.3.2, in the case VC batteries are predicted
to be charged at time step k (to help in solving voltage rise problems), the offline
(k) (k)
controller sets rUPi and βR-UPi to zero. In the case VC batteries are expected
to be discharged at time step k (to help in solving voltage drop problems), the
(k) (k)
offline controller sets rDNi and βR-DNi to zero. The following constraints are
PROBLEM FORMULATION 143
(k) (k)
then rUPi = 0 and βR-UPi = 0
(k)
If pVC, bati = 0 and p(k)
VC, bat
< 0 , (6.25b)
i
(k) (k)
then rDNi = 0 and βR-DNi = 0
The solver used in the case study (YALMIP) transforms the above logic
constraints into mixed integer linear constraints using Big-M strategy [203].
2
k M R-DNi + IRDNi ξ R-DN k +
| {z }2
5
#
2 X
k M R-UPi + IRUPi ξ R-UP k (RUPi + RDNi ) (6.26)
−
| {z }2 i∈I
6 | {z }
7
Here, E[·] denotes the expected value operator, and k · k2 is the second norm.
Parts 1, 2 and 3 of the objective function represent sum of quadratic uncertain
reactive powers of inverter i for nt time steps. Uncertain reactive powers of
parts 1, 2 and 3 are responsible for solving voltage problems expected to be
144 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
caused by the reverse power flow from PV units, charging R-DN batteries, and
discharging R-UP batteries, respectively. Parts 4, 5 and 6 of the objective
function represent sum of quadratic uncertain active powers of battery i for
nt time steps. Uncertain active powers of part 4 are the uncertain charging
and discharging powers of the VC battery i. Uncertain active powers of part 5
represent the uncertain charging powers of the downward-FCR and uncertain
discharging powers of the R-DN management policy. Uncertain active powers
of part 6 represent uncertain discharging powers of the upward-FCR and
uncertain charging powers of the R-UP management policy. IRDNi in part
5 denotes a diagonal matrix ∈ Rnt ×nt with the vector of downward reserve
h iT
(1) (n )
capacities RDNi = rDNi , · · · , rDNt i contained in its diagonal elements. IRUPi
in part 6 denotes a diagonal matrix ∈ Rnt ×nt with the vector of upward reserve
h iT
(1) (n )
capacities RUPi = rUPi , · · · , rUPti contained in its diagonal elements. Part
7 is included in the objective function to maximize the upward and downward
reserve capacities.
It can be shown that the objective function (6.26) can be written as:
h i X
2
min E kzζk2 − (RUPi + RDNi ) (6.27)
i∈I
where z ∈ R6nc nt ×nζ can be defined to represent the first six parts of the
objective function. A closed-form of the expected value in (6.27) is not readily
available. A second-order moment matrix can be used h to
i approximate the
expected value, as discussed in [178]. If M = E ζζ T
is a second-order
moment matrix of the random vector ζ, then:
h i h i
2
E kzζk2 = E Tr zT zζζ T = Tr zT zM (6.28)
Here, the matrix ζ scen ∈ Rnscen ×nζ contains nscen scenarios of the random vector
ζ. µ ∈ R1×nζ is the sample mean of ζ scen , and cov (ζ scen ) is the covariance
matrix of ζ scen .
PROBLEM FORMULATION 145
Based on the previous linear models, one can show that the stochastic constraints
(6.14a)-(6.14c), (6.15a)-(6.15c), (6.20), (6.24a), and (6.24b) can be written as:
Aζ ≤ B, (6.30)
maxAζ 6 B (6.31)
ζ∈U
Using standard duality techniques, it can be shown that the constraint (6.31)
with a polyhedral uncertainty set U can be reformulated as the linear constraints
(6.32b)-(6.32d). We ask the readers to refer to lemma 2 in [204] for details
about robust counterpart of linear constraints with polyhedral uncertainty set.
Below is the complete optimization problem the offline controller solves on a
day-ahead basis:
X
min Tr zT zM − (RUPi + RDNi ) (6.32a)
i∈I
subject to:
Robust constraints :
πh ≤ B (6.32b)
πH ≥ A (6.32c)
π≥0 (6.32d)
The elements of the matrix π ∈ R(2nc nt (6+2ns )+4nnodes nt )×2nζ are auxiliary
variables.
146 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
subject to:
subject to:
Θ (ζ) ≥ 0 (6.34b)
where Áζ = Aζ − B. The elements of the vector Θ (ζ) are auxiliary variables,
and λ (ζ) is a vector of dual decisions. The dual feasible set of the problem
(6.34a)-(6.34d) can be restricted to combine only linear dual decisions. To this
end, we require the dual decisions to be representable as λ (ζ) = Λζ. Based on
this, the objective function (6.34a) can be reformulated as:
X
min max Tr zT zM − (RUPi + RDNi ) +
Λ≥0
i∈I
h i
E ζ T ΛT Áζ + Θ (ζ) (6.35a)
X
= min max Tr zT zM − (RUPi + RDNi ) +
Λ≥0
i∈I
h i
Tr ΛT E Áζ + Θ (ζ) ζ T (6.35b)
The maximization in (6.35b) can be carried out explicitly. This yields to the
following approximate problem.
X
min Tr zT zM − (RUPi + RDNi ) (6.36a)
i∈I
subject to:
h i
E Áζ + Θ (ζ) ζ T = 0 (6.36b)
Θ (ζ) ≥ 0 (6.36c)
subject to:
Á + θ = 0 (6.37b)
H − hoT1 M θ T ≤ 0 (6.37c)
148 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
Group G3
16 14 15
13
Control node
i V VI
Passive node
VII
IV
7 8 9 10 11 12 17 18 19 20 21 22
Grid Group G2 Group G4
II
I
III VIII
1 2 3 4 5 6 23 24 25 26 27 28 29 30 31
Group G1 Group G5
Figure 6.7: Schematic diagram of the network used in this case study.
G1 G2 G3 G4 G5
sinv (kVA) 5 8 10 12 15
Ebat
0
(kWh) 2.4 3.6 4.1 4.9 6.75
DoD (%) 80 80 80 80 80
Figure 6.8 shows aggregated power capacities of the 31 BESSs reserved for
the downward-FCR, upward-FCR, voltage rise, energy management policies
of R-DN and R-UP batteries. The upward reserve capacity is defined as a
positive value in (6.15d); it is presented as a negative power in Figure 6.8 to
compare it to the maximum discharging power. One can notice that the upward
reserve capacity is zero from 10:00 till 14:30, this is because VC batteries are
expected to be charged over this period. The highlighted positive power area
between 10:00 and 14:30 shows the aggregated power capacity reserved for
solving voltage rise problems. Figure 6.9, for example, shows the power capacity
reserved for charging the VC battery 31 over the period 10:00-14:30. The
aggregated downward reserve capacity is at minimum at 10:45, since most of the
inverters capacity is occupied by the PV power, VC batteries power, and reactive
power. The highlighted negative power area in Figure 6.8 shows the aggregated
power capacity reserved for the management policy of R-DN batteries. The
highlighted positive area between 00:00 and 10:00 and between 14:30 and 24:00
shows the aggregated power capacity reserved for the management policy of
R-UP batteries. There is no power capacity reserved for the management policy
of R-UP batteries between 10:00 and 14:30. Reactive power is expected to solve
the minor voltage drop problem shown in Figure 6.10a, therefore, there is no
power capacity reserved for solving voltage drop problems.
Voltage profiles are computed for the 104 scenarios without activating the
voltage and frequency control system. Figure 6.10a shows voltage profiles of
control node 31 (phase to neutral voltage). One can see that most of voltage
profiles of different scenarios exceed the maximum voltage limit. When linear
voltage control policies are applied, with frequency control enabled, it is noticed
that the existing violations of the over-voltage limit are eliminated for the 104
scenarios, which demonstrates the robustness of the proposed voltage control
policies. Figure 6.10b shows voltage profiles of control node 31 (at the end of the
feeder) after applying voltage control policies, with frequency control enabled.
Figure 6.11a and Figure 6.11b show, respectively, the apparent power magnitude
of inverter 31 and energy content of battery 31 for all the scenarios. One can
see that the capacity limits of the inverter and battery are respected for all
the scenarios. The R-UP and R-DN energy management policies succeed in
maintaining the energy content of battery 31 within limits while providing
downward-FCR and upward-FCR, which demonstrates the robustness of the
proposed energy management policies.
152 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
Figure 6.9: Power capacity reserved for charging the VC battery 31.
(a) (b)
Figure 6.10: (a) Voltage profiles of node 31 with no control, (b) Voltage profiles
of node 31 with voltage and frequency control. The dotted lines show the
maximum and minimum voltage limits
(a) (b)
Figure 6.11: (a) Apparent power magnitude of inverter 31, (b) Energy content
of battery 31.
bounds to the upper bounds. Results show that the optimality gap of the
proposed robust control methodology varies between 8% and 13% for the 104
scenarios, which is an acceptable optimality gap. We would like to remind the
reader that obtaining an optimal solution is intractable. Hence, we trade-off
optimality for computational tractability.
154 SIMULTANEOUS PROVISION OF VOLTAGE AND FREQUENCY CONTROL BY PV-BATTERY
SYSTEMS
6.5 Conclusion
155
156 GENERAL CONCLUSIONS AND FUTURE WORK
This section summarizes the answers the thesis provides to the research questions
posed in Section 1.2.
ANSWERS TO THE RESEARCH QUESTIONS 157
Chapter 2
Chapter 3
In Subsection 3.5.11, the thesis studies convergence speed of the DD-based DVC
system. The thesis performs 104 numerical tests on the simulated network of
Figure 3.4. For each test, the number of smart inverters participating in voltage
control varies from 10 to 60. Results show that the average convergence speed
of the different experiments varies linearly from 8.4 minutes (10-agent case)
158 GENERAL CONCLUSIONS AND FUTURE WORK
One of the main findings in the thesis is that the direct extension of the sequential
2-block ADMM to parallel multi-block ADMM (the so-called Jacobi ADMM)
without additional modifications or assumptions slows down the convergence
significantly. As demonstrated in Subsection 3.7.2, Jacobi ADMM is even slower
than the DD algorithm.
Chapter 4
Chapter 5
Chapter 6
Inverter Manufacturers
Today’s smart inverter are equipped with communication-less local grid support
functions or centralized grid support functions or both. As has been discussed
in this thesis, local grid support functions suffer from degraded performance and
centralized grid support functions have poor reliability and scalability. Hence,
it is recommended for inverter manufacturers to place emphasis on the design,
development and experimental validation of P2P-based grid support functions.
Standardization Bodies
To enable DERs, electric vehicles, and smart loads to interact with each other
and control distribution networks in a P2P fashion, appropriate regulatory
frameworks must be developed first. Hence, it is recommended for the
standardization bodies to develop technical specifications and amend grid codes
for the provision of P2P-based control and management services. Roles and
responsibilities of the DSOs and prosumers in the P2P-based paradigm should
be clearly defined.
FUTURE WORK 163
Voltage Unbalance
P2P-Based Platform
This thesis focuses on developing grid support functions for PV and PV-battery
systems. An interesting future research would be to investigate how the proposed
grid support functions can be extended to build a scalable P2P-based platform
(Figure 7.1) that integrates renewable energy sources, energy storage systems,
smart controllable loads, electric vehicles, on-load tap changer, etc. Enabling
different types of DERs and network assets to communicate with each other in
a P2P fashion would create a very powerful control system.
Grid
On-load
tap-changer
Inverter Controlling
GSF active power
Solar panel
P2P
Inverter
Controlling smart
loads Controlling active GSF
GSF Battery
and reactive power
Inverter
167
168 BIBLIOGRAPHY
[29] J. Li, Z. Xu, J. Zhao, and C. Zhang, “Distributed online voltage control
in active distribution networks considering PV curtailment,” IEEE
Transactions on Industrial Informatics, vol. 15, no. 10, pp. 5519–5530,
2019.
[30] I. Ranaweera, O.-M. Midtgård, and M. Korpås, “Distributed control
scheme for residential battery energy storage units coupled with PV
systems,” Renewable Energy, vol. 113, pp. 1099–1110, 2017.
[109] D. Das, D. Kothari, and A. Kalam, “Simple and efficient method for load
flow solution of radial distribution networks,” International Journal of
Electrical Power & Energy Systems, vol. 17, no. 5, pp. 335–346, 1995.
[110] F. Olivier, P. Aristidou, D. Ernst, and T. Van Cutsem, “Active
management of low-voltage networks for mitigating overvoltages due
to photovoltaic units,” IEEE Transactions on Smart Grid, vol. 7, no. 2,
pp. 926–936, 2016.
[111] S. Weckx, R. D’Hulst, and J. Driesen, “Voltage sensitivity analysis of a
laboratory distribution grid with incomplete data,” IEEE Transactions
on Smart Grid, vol. 6, no. 3, pp. 1271–1280, 2015.
[137] B. He, L. Hou, and X. Yuan, “On full jacobian decomposition of the
augmented Lagrangian method for separable convex programming,” SIAM
Journal on Optimization, vol. 25, no. 4, pp. 2274–2312, 2015.
[138] W. Deng, M.-J. Lai, Z. Peng, and W. Yin, “Parallel multi-block ADMM
with o(1/k) convergence,” Journal of Scientific Computing, vol. 71, no. 2,
pp. 712–736, 2017.
[139] B. He, H.-K. Xu, and X. Yuan, “On the proximal Jacobian decomposition
of ALM for multiple-block separable convex minimization problems and
its relationship to ADMM,” Journal of Scientific Computing, vol. 66,
no. 3, pp. 1204–1217, 2016.
[140] J. A. Suykens, J. Vandewalle, and B. De Moor, “Intelligence and
cooperative search by coupled local minimizers,” International Journal of
Bifurcation and Chaos, vol. 11, no. 08, pp. 2133–2144, 2001.
[141] J. Suykens and J. Vandewalle, “Coupled local minimizers: alternative
formulations and extensions,” in Proceedings of the 2002 International
Joint Conference on Neural Networks, vol. 3, pp. 2039–2043, IEEE, 2002.
[142] I. Kouveliotis-Lysikatos, D. Koukoula, I. Vlachos, A. Dimeas,
N. Hatziargyriou, and S. Makrynikas, “Decentralised distribution system
operation techniques: results from the Meltemi community smart grids
pilot site,” CIRED-Open Access Proceedings Journal, vol. 2017, no. 1,
pp. 1673–1677, 2017.
[143] S. Iacovella, P. Vingerhoets, G. Deconinck, N. Honeth, and L. Nordstrom,
“Multi-agent platform for grid and communication impact analysis of
rapidly deployed demand response algorithms,” in Energy Conference
(ENERGYCON), 2016 IEEE International, pp. 1–6, IEEE, 2016.
[144] B. Dupont, P. Vingerhoets, P. Tant, K. Vanthournout, W. Cardinaels,
T. De Rybel, E. Peeters, and R. Belmans, “LINEAR breakthrough project:
Large-scale implementation of smart grid technologies in distribution
grids,” in 3rd IEEE PES Innovative Smart Grid Technologies Europe
(ISGT Europe), pp. 1–8, IEEE, 2012.
[145] K. Vanthournout, K. De Brabandere, E. Haesen, J. Van den Keybus,
G. Deconinck, and R. Belmans, “Agora: Distributed tertiary control
of distributed resources,” in Proceedings of the 15th Power Systems
Computation Conference, pp. 1–7, 2005.
[146] K. J. Macken, K. Vanthournout, J. Van den Keybus, G. Deconinck, and
R. J. Belmans, “Distributed control of renewable generation units with
integrated active filter,” IEEE Transactions on Power Electronics, vol. 19,
no. 5, pp. 1353–1360, 2004.
BIBLIOGRAPHY 181
[157] J. Markkula and J. Haapola, “Ad hoc LTE method for resilient smart
grid communications,” Wireless Personal Communications, vol. 98, no. 4,
pp. 3355–3375, 2018.
[158] U. Raza, P. Kulkarni, and M. Sooriyabandara, “Low power wide area
networks: An overview,” IEEE Communications Surveys & Tutorials,
vol. 19, no. 2, pp. 855–873, 2017.
[159] K. Mikhaylov and J. Petäjäjärvi, “Design and implementation of the
plug&play enabled flexible modular wireless sensor and actuator network
platform,” Asian Journal of Control, vol. 19, no. 4, pp. 1392–1412, 2017.
[160] IEA, “Renewables 2019: Market analysis and forecast from 2019 to 2024,
IEA, Paris.” https://www.iea.org/reports/renewables-2019. Accessed:
24-12-2019.
[161] IEA (2020), “Tracking energy integration: Energy storage, IEA, Paris.”
https://www.iea.org/reports/tracking-energy-integration-2020/energy-
storage. Accessed: 28-07-2020.
[162] H. Almasalma, J. Engels, and G. Deconinck, “Dual-decomposition-based
peer-to-peer voltage control for distribution networks,” CIRED - Open
Access Proceedings Journal, vol. 2017, no. 1, pp. 1718–1721, 2017.
[163] K. S. Ayyagari, N. Gatsis, and A. F. Taha, “Chance constrained
optimization of distributed energy resources via affine policies,” in
2017 IEEE Global Conference on Signal and Information Processing
(GlobalSIP), pp. 1050–1054, IEEE, 2017.
[164] R. A. Jabr, “Robust volt/var control with photovoltaics,” IEEE
Transactions on Power Systems, vol. 34, no. 3, pp. 2401–2408, 2019.
[165] R. A. Jabr, “Linear decision rules for control of reactive power by
distributed photovoltaic generators,” IEEE Transactions on Power
Systems, vol. 33, no. 2, pp. 2165–2174, 2017.
[166] W. Lin and E. Bitar, “Decentralized stochastic control of distributed
energy resources,” IEEE Transactions on Power Systems, vol. 33, no. 1,
pp. 888–900, 2017.
[167] E. E. Elattar, J. Goulermas, and Q. H. Wu, “Electric load forecasting
based on locally weighted support vector regression,” IEEE Transactions
on Systems, Man, and Cybernetics, Part C (Applications and Reviews),
vol. 40, no. 4, pp. 438–447, 2010.
[168] N. Zhang, C. Kang, E. Du, and Y. Wang, Analytics and Optimization for
Renewable Energy Integration, ch. 5, pp. 96–98. CRC Press, 2019.
BIBLIOGRAPHY 183
[191] T. Kern, S. Roon, “New auction design for frequency containment reserve
since 1 July 2019.” Available at: https://www.ffegmbh.de/en/areas-
of-expertise/scientific-analysis-of-system-and-energy-markets/106-
BIBLIOGRAPHY 185
electricity-market/892-new-auction-design-for-frequency-containment-
reserve-since-1-july-2019-first-interim-results. Accessed: 10 March
2020.
[199] Y. Shi, B. Xu, D. Wang, and B. Zhang, “Using battery storage for peak
shaving and frequency regulation: Joint optimization for superlinear gains,”
IEEE Transactions on Power Systems, vol. 33, no. 3, pp. 2882–2894, 2017.
[200] O. Mégel, J. L. Mathieu, and G. Andersson, “Stochastic dual dynamic
programming to schedule energy storage units providing multiple services,”
in 2015 IEEE Eindhoven PowerTech, pp. 1–6, IEEE, 2015.
[201] J. Chen, R. Zhu, M. Liu, G. De Carne, M. Liserre, F. Milano, and
T. O’Donnell, “Smart transformer for the provision of coordinated
186 BIBLIOGRAPHY
Hamada Almasalma
Born on 27th December, 1987 in Beit Jala (Bethlehem Governorate), Palestine.
187
List of Publications
Journal publications
• H. Almasalma, and G. Deconinck, “Simultaneous Provision of Voltage
and Frequency Control by PV-Battery Systems,” IEEE ACCESS, vol. 8,
2020 (Early Access).
• H. Almasalma, and G. Deconinck, “Robust Policy-Based Distributed
Voltage Control Provided by PV-Battery Inverters,” IEEE ACCESS, vol.
8, pp. 124939-124948, 2020.
• H. Almasalma, S. Claeys, and G. Deconinck, “Peer-to-peer-based
integrated grid voltage support function for smart photovoltaic inverters,”
Applied Energy, vol. 239, pp. 1037–1048, 2019.
International Conferences
• J. Haapola, S. Ali, C. Kalalas, J. Markkula, N. Rajatheva, A. Pouttu, J.
M. M. Rapún, I. Lalaguna, F. V. Gallego, J. A. Zarate, G. Deconinck,
H. Almasalma, J. Wu, C. Zhang, E. P. Muñoz, and F. D. Gallego.
Peer-to-peer energy trading and grid control communications solutions.
IEEE 87th Vehicular Technology Conference (VTC Spring), Porto, 2018.
189
190 LIST OF PUBLICATIONS