UNIVERSIDAD DE CHILE

FACULTAD DE CIENCIAS FÍSICAS Y MATEMÁTICAS

DEPARTAMENTO DE INGENIERÍA ELÉCTRICA

A DISTRIBUTED PREDICTIVE SECONDARY CONTROL FOR VOLTAGE AND

FREQUENCY REGULATION, ECONOMIC DISPATCH AND IMBALANCE SHARING IN
ISOLATED MICROGRIDS

TESIS PARA OPTAR AL GRADO DE DOCTOR EN INGENIERÍA ELÉCTRICA

EN COTUTELA CON LA UNIVERSIDAD DE NOTTINGHAM

ALEX DARIO NAVAS FONSECA

PROFESORA GUÍA:
DORIS SÁEZ HUEICHAPAN

PROFESOR CO-GUÍA 1:
MARK SUMNER
PROFESOR CO-GUÍA 2:
CLAUDIO BURGOS MELLADO

MIEMBROS DE LA COMISIÓN:
DANIEL SBÁRBARO HOFER
ALAN WATSON
CÉSAR AZURDIA MEZA

SANTIAGO DE CHILE
2022
 RESUMEN DE LA TESIS PARA OPTAR
AL GRADO DE DOCTOR EN INGENIERÍA ELÉCTRICA
POR: ALEX DARIO NAVAS FONSECA
FECHA: 2022
PROF. GUÍA: DORIS SÁEZ HUEICHAPAN
PROF. CO-GUÍA 1: MARK SUMNER
PROF. CO-GUÍA 2: CLAUDIO BURGOS

CONTROL PREDICTIVO DISTRIBUIDO PARA EL NIVEL DE CONTROL SECUNDARIO

PARA LA REGULACIÓN DE TENSIÓN Y FRECUENCIA, DESPACHO ECONÓMICO Y
COMPARTICIÓN DE DESBALANCES EN MICRORREDES AISLADAS

Esta tesis se enfoca en estudiar la aplicación de control predictivo distribuido (DMPC) para el con-
trol secundario de microrredes (MGs) ac y microrredes híbridas ac/dc (H-MGs). Se proponen tres
estrategias de control que cumplen con la tarea principal del control secundario (restaurar frecuen-
cia y voltaje). Las estrategias propuestas incluyen objetivos complementarios en la formulación
según el tipo de MG estudiada. Además, estas estrategias pueden restaurar la frecuencia y el voltaje
a valores nominales o dentro de bandas de seguridad. La primera estrategia propuesta considera el
despacho óptimo de generadores distribuidos (DGs) en MGs ac balanceadas. La segunda estrategia
logra el despacho óptimo de ac DGs, dc DGs y gestiona la transferencia de potencia a través de
los interlinking converters (ILCs) basado en un criterio económico en H-MGs. La última estrategia
propuesta es capaz de gestionar la distribución de desbalances en MGs ac desbalanceadas. Las
principales características de las metodologías propuestas son el desarrollo de formulaciones nove-
dosas con funciones de costo multiobjetivo y modelos de predicción que representan las principales
dinámicas de los DGs y los ILCs (en el caso de las H-MGs). Finalmente, se validan los esquemas
DMPC propuestos experimentalmente, por hardware-in-the-loop y por simulación bajo escenarios
demandantes.

i
 RESUMEN DE LA TESIS PARA OPTAR
AL GRADO DE DOCTOR EN INGENIERÍA ELÉCTRICA
POR: ALEX DARIO NAVAS FONSECA
FECHA: 2022
PROF. GUÍA: DORIS SÁEZ HUEICHAPAN
PROF. CO-GUÍA 1: MARK SUMNER
PROF. CO-GUÍA 2: CLAUDIO BURGOS

A DISTRIBUTED PREDICTIVE SECONDARY CONTROL FOR VOLTAGE AND

FREQUENCY REGULATION, ECONOMIC DISPATCH AND IMBALANCE SHARING IN
ISOLATED MICROGRIDS

This thesis focuses on studying the application of distributed model predictive control (DMPC) for
the secondary control of ac microgrids (MGs) and hybrid ac/dc microgrids (H-MGs). Three control
strategies are proposed that fulfil the main task of the secondary control (restore frequency and volt-
age). The proposed strategies include complementary objectives in their formulation according to
the type of MG studied. Additionally, these strategies can restore frequency and voltage to nominal
values or within secure bands. The first proposed strategy considers the optimal dispatch of dis-
tributed generators (DGs) in balanced ac MGs. The second strategy achieves the optimal dispatch
of ac DGs, dc DGs and manages the power transference through interlinking converters (ILCs)
based on an economic criterion in H-MGs. The last proposed strategy is able to manage imbalance
sharing in unbalanced ac MGs. The main characteristics of the proposed methodologies are the
development of novel multi objective cost functions and prediction models that represent the main
dynamics of DGs and ILCs (in the case of H-MGs). Finally, extensive experimental, hardware-in-
the-loop and simulation studies validate the proposed DMPC schemes under demanding scenarios.
For the interested reader, an extended abstract is presented in Annexed A.

ii
 To the memory of my grandfathers Alonso and Humberto, you will always be in my heart.
To my wife, Verónica, the most important person in my life, thanks for your constant support and
encouragement in difficult times and for always being by my side.
To my parents, Marco and Mónica, for their unconditional love and support during my whole life.
To my brother Daniel and my sister Sheyla, I love you both with all my heart.

iii
Acknowledgements

I would like to express my sincere gratitude to my supervisor Prof. Doris Sáez for their uncondi-
tional encouragement and for always believe in me. This cotutelle program would not be possi-
ble without your support and motivation. My sincere gratitude to my co-supervisors Prof. Mark
Sumner and Prof. Claudio Burgos for their commitment, guidance and support during my stay in
Nottingham. My gratitude and thanks are also extended to my PhD committee members: Prof.
Alan Watson, Prof. Daniel Sbárbaro, Prof. Cesar Azurdia.

I wish to thank to all my friends from the University of Chile: Jacqueline, Diego O., Jorge
Pancho, Luis, Juan, Yeiner, Tomy, Daniel, Enrique, Felipe D. Matías, Erwin, Felipe H., Diego M.,
Manuel. Thank you for fruitful discussions and support during my studies.

I also thank to all my colleagues and friends from the Power Electronics, Machines and Control
group in Nottingham. I really enjoyed my stay in this beautiful city.

Last but not least, I would like to express my sincere gratitude to Secretaría de Educación
Superior, Ciencia, Tecnología e Innovación de Ecuador under Grant SENESCYT/ARSEQ-BEC-
0058482018. Agencia Nacional de Investigación y Desarrollo (ANID) in Chile under grant ANID
BECAS/DOCTORADO NACIONAL 2019-21190961. The support of the Instituto Sistemas Com-
plejos de Ingenieria (ISCI) ANID PIA/BASAL AFB180003 is also acknowledged. This work was
also supported by the Fondo Nacional de Desarrollo Científico y Tecnológico (FONDECYT) under
Grant 1170683 and by SERC-Chile ANID/FONDAP/15110019.

iv
Table of Content

List of Abbreviations xiii

List of Symbols xv

1 Introduction 1
1.1 Research motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 General objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.2 Specific objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Thesis structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Literature review 10
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Microgrids framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Control of ac MGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Primary control level in ac MGs . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Secondary control level in ac MGs . . . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Tertiary control level in ac MGs . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Control of dc MGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.1 Primary control level in dc MGs . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.2 Secondary control level in dc MGs . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Control of hybrid ac/dc microgrids . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.6 Distributed model predictive control for microgrids . . . . . . . . . . . . . . . . . 21
2.7 Distributed secondary control for microgrids in the literature . . . . . . . . . . . . 23
2.7.1 Distributed economic dispatch for ac microgrids . . . . . . . . . . . . . . 23
2.7.2 Distributed economic dispatch for dc microgrids . . . . . . . . . . . . . . 25

v
 2.7.3 Distributed economic dispatch for hybrid ac/dc microgrids . . . . . . . . . 26
2.8 Imbalance sharing in ac microgrids . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 The proposed DMPC scheme for frequency regulation and active power dispatch in
ac microgrids 31
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Centralised economic dispatch . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Proposed DMPC scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Dynamic models used for the design of the DMPC strategy . . . . . . . . . . . . . 35
3.4.1 Communication network model . . . . . . . . . . . . . . . . . . . . . . . 35
3.4.2 Dynamic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4.3 Discrete time models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5 Formulation of the distributed model predictive control . . . . . . . . . . . . . . . 38
3.5.1 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5.2 Predictive models and constraints . . . . . . . . . . . . . . . . . . . . . . 39
3.5.3 Formulation of the quadratic programming . . . . . . . . . . . . . . . . . 40
3.6 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.6.1 Experimental MG configuration . . . . . . . . . . . . . . . . . . . . . . . 42
3.6.2 Design parameters and test scenarios used to evaluate the DMPC . . . . . . 45
3.6.3 Scenario I (base case) - Load changes . . . . . . . . . . . . . . . . . . . . 46
3.6.4 Scenario II - Communication delay . . . . . . . . . . . . . . . . . . . . . 48
3.6.5 Scenario III - Communication link failure . . . . . . . . . . . . . . . . . . 49
3.6.6 Scenario IV - Plug-and-play . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4 The proposed DMPC scheme for frequency and voltage regulation within bands and
the economic dispatch of active and reactive power for hybrid ac/dc microgrids 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 The active power economic dispatch problem . . . . . . . . . . . . . . . . . . . . 54
4.3 The reactive power economic dispatch problem . . . . . . . . . . . . . . . . . . . 55
4.4 Communication structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.5 Proposed DMPC scheme for interlinking converters (ILCs) . . . . . . . . . . . . . 56
4.5.1 Dynamic models used for the design of the controller for interlinking con-
verters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.6 Formulation of the DMPC for ILCs . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.6.1 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.6.2 Prediction models and constraints . . . . . . . . . . . . . . . . . . . . . . 59

vi
 4.7 Proposed DMPC scheme for dc Generators . . . . . . . . . . . . . . . . . . . . . 61
4.7.1 Dynamic models used for the design of the controller for dc generators . . 62
4.8 Formulation of the DMPC for dc generators . . . . . . . . . . . . . . . . . . . . . 63
4.8.1 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.8.2 Predictive models and constraints . . . . . . . . . . . . . . . . . . . . . . 64
4.9 Proposed DMPC scheme for ac Generators . . . . . . . . . . . . . . . . . . . . . 65
4.9.1 Dynamic models used for the design of the controller for ac generators . . 67
4.10 Formulation of the DMPC for ac generators . . . . . . . . . . . . . . . . . . . . . 68
4.10.1 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.10.2 Predictive models and constraints . . . . . . . . . . . . . . . . . . . . . . 69
4.11 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.11.1 Design parameters and test scenarios used to evaluate the DMPC for H-MGs 74
4.11.2 Scenario I (base case) - Load changes . . . . . . . . . . . . . . . . . . . . 76
4.11.3 Scenario II - Combined communication link failures and Plug-and-Play . . 78
4.11.4 Scenario III - Comparison against a DAPI-based strategy without eco-
nomic dispatch for H-MGs . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.12 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5 The Proposed DMPC scheme for phase imbalance sharing and frequency and voltage
regulation within bands in ac microgrids 84
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2 Proposed DMPC scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3 Dynamic models used for the design of the DMPC strategy . . . . . . . . . . . . . 87
5.3.1 Droop control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.2 Phase angle model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.3.3 Power transfer models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.3.4 Phase voltage unbalance rate index . . . . . . . . . . . . . . . . . . . . . 89
5.4 Distributed MPC formulation for imbalance sharing . . . . . . . . . . . . . . . . . 89
5.4.1 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.4.2 Predictive models and constraints . . . . . . . . . . . . . . . . . . . . . . 91
5.5 Microgrid setup and simulation results . . . . . . . . . . . . . . . . . . . . . . . . 95
5.5.1 Scenario I (base case) - Unbalanced load changes . . . . . . . . . . . . . . 98
5.5.2 Scenario II - Communication delays . . . . . . . . . . . . . . . . . . . . . 103
5.5.3 Scenario III - Combined communication link failures and plug-and-Play . . 104
5.6 Hardware in the loop validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.7 Scalability and comparison with a DAPI-based controller . . . . . . . . . . . . . . 111
5.7.1 Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

vii
 5.7.2 Comparison with a distributed consensus-based controller for imbalance
sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

6 Conclusions and final remarks 119

6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2.1 Journal papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2.2 Conference papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

BIBLIOGRAPHY 123

Annexes 138

A Extended abstract 138

B Design of a reduced order nonlinear observer to estimate the voltage after a coupling
inductance 140

C Derivation of predictive linear models used as equality constraints in ac DGs 143

C.1 Continuous time model for equality constraints . . . . . . . . . . . . . . . . . . 143
C.2 Model discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
C.2.1 Droop equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
C.2.2 Phase angle equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
C.2.3 Power transfer equations . . . . . . . . . . . . . . . . . . . . . . . . . . 147
C.3 Prediction model for equality constraints . . . . . . . . . . . . . . . . . . . . . 149

D Derivation of predictive linear models used as inequality constraints in ac DGs 151

D.1 PVUR inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
D.2 Apparent power rating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

E Derivation of predictive linear models used as equality constraints in dc DGs 158

E.1 Continuous time model for equality constraints . . . . . . . . . . . . . . . . . . 158
E.1.1 Droop equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
E.1.2 Power transfer equation . . . . . . . . . . . . . . . . . . . . . . . . . . 159
E.2 Prediction model for equality constraints . . . . . . . . . . . . . . . . . . . . . 160

F Experimental microgrid setup 161

viii
List of Tables

2.1 Comparison between Centralised and Distributed Control . . . . . . . . . . . . . . 14

3.1 MG electrical parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.2 Power ratings and droop slopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 DG operating costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Controller parameters and weights . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.1 MG parameters and loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.2 DGs parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3 DMPC parameters and Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.1 MG parameters and loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5.2 Controller parameters and weights . . . . . . . . . . . . . . . . . . . . . . . . . . 97

ix
List of Figures

2.1 General topology for a hybrid ac/dc microgrid. a) ac microgrid. b) dc microgrid.

c) Interlinking converters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Control structures at the secondary control level. a) Centralised. b) Decentralised.
c) Distributed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Droop control for ac DGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Secondary control for ac DGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Droop control for dc DGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Secondary control for dc DGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.1 Distributed economic dispatch with frequency restoration for the i − th DG . . . . 34

3.2 Example of a MG with four DGs and its adjacency matrix . . . . . . . . . . . . . 36
3.3 a) MG Diagram, b) Experimental MG setup . . . . . . . . . . . . . . . . . . . . . 43
3.4 Load changes - base case: a) Active power contribution, b) Incremental cost con-
sensus, c) Frequency regulation, d) Total operation cost . . . . . . . . . . . . . . . 46
3.5 Load changes - base case: a) Reactive power contribution, b) Voltage regulation . . 47
3.6 Communication delays: a) Frequency regulation for τi j = 0.25s, b) Frequency reg-
ulation for τi j = 1s, c) Active power contribution for τi j = 0.25s, d) Active power
contribution for τi j = 1s, e) Incremental cost consensus for τi j = 0.25s, f) Incre-
mental cost consensus for τi j = 1s . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.7 Communication failure: a) Active power contribution, b) Incremental cost consen-
sus, c) Frequency regulation, d) Total operation cost . . . . . . . . . . . . . . . . . 50
3.8 Plug-and-Play test: a) Active power contribution, b) Incremental cost consensus, c)
Frequency regulation, d) Total Operation cost . . . . . . . . . . . . . . . . . . . . 51

4.1 Control diagram of DMPCi for ILCs. . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.2 Control diagram of DMPCi for dc DGs. . . . . . . . . . . . . . . . . . . . . . . . 62
4.3 Control diagram of DMPCi for ac DGs. . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Hybrid ac/dc MG topology for the validation of the DMPC scheme . . . . . . . . . 73

x
4.5 Load changes: a) Incremental cost consensus, b) Reactive marginal cost consensus,
c) Active power, d) Reactive power . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.6 Load changes: a) Active power through the ILCs, b) Frequency regulation, c) Aver-
age dc voltage regulation, d) Average ac voltage regulation. The dashed cyan lines
represent the predefined band limits for frequency and voltages. . . . . . . . . . . 77
4.7 Communication failure and plug-and-play test: a) Active power through the ILCs,
b) Frequency regulation, c) Average dc voltage regulation, d) Average ac voltage
regulation. The dashed cyan lines represent the predefined band limits for fre-
quency and voltages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.8 Communication failure and plug-and-play test: a) Incremental cost consensus, b)
Reactive marginal cost consensus, c) Active power, d) Reactive power . . . . . . . 79
4.9 Comparison between the proposed DMPC scheme and a DAPI-based method for
(τi j = 1s): a)-b) Active power for both methods, c)-d) Active power through the
ILCs for both methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.10 Comparison between the proposed DMPC scheme and a DAPI-based method for
(τi j = 1s): a) Total operating cost, b) Frequency regulation, c) Average dc voltage
regulation, d) Average ac voltage regulation. The dashed cyan lines represent the
predefined band limits for frequency and voltages. . . . . . . . . . . . . . . . . . . 82

5.1 General control diagram of DMPCi for imbalance sharing. . . . . . . . . . . . . . 85

5.2 Single-phase droop controller. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 Implemented MG simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.4 Base Case a) Normalised reactive power consensus - Phase a for load changes, b)
Normalised reactive power consensus - Phase b for load changes, c) Normalised
reactive power consensus - Phase c for load changes . . . . . . . . . . . . . . . . . 99
5.5 a) Three phase normalised active power consensus for load changes - Base Case,
b) Three-phase normalised reactive power consensus for load changes - Base Case 100
5.6 PVUR index of the voltage at the DGs output for load changes - Base Case. . . . . 101
5.7 a) Frequency regulation for load changes - Base Case, b) Average voltage regula-
tion for load changes - Base Case. The dashed cyan lines represent the predefined
band limits for both variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.8 Communication delay test: a) Normalised reactive power consensus - Phase a for
τi j = 0.25s, b) Normalised reactive power consensus - Phase a for τi j = 1s, c)
Three-phase normalised active power consensus for τi j = 0.25s, d) Three-phase
normalised active power consensus for τi j = 1s, e) Frequency regulation for τi j =
0.25s and τi j = 1s, f) Average voltage regulation for τi j = 0.25s and τi j = 1s. The
dashed cyan lines represent the predefined band limits for both latter variables. . . 103

xi
5.9 Communication failure and plug-and-play test. a) Frequency regulation, b) Average
voltage regulation. The dashed cyan lines represent the predefined band limits for
both variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.10 Communication failure and plug-and-play test. a) Normalised reactive power con-
sensus - Phase a, b) Normalised reactive power consensus - Phase b, c) Normalised
reactive power consensus - Phase c . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.11 OPAL-RT platform for HIL validation . . . . . . . . . . . . . . . . . . . . . . . . 107
5.12 HIL - Communication failure and plug-and-play test. a) Normalised reactive power
consensus - Phase a, b) Normalised reactive power consensus - Phase b, c) Nor-
malised reactive power consensus - Phase c . . . . . . . . . . . . . . . . . . . . . 108
5.13 HIL - Communication failure and plug-and-play test. a) Current and voltage at the
DG4 connection point, b) Current at the DG3 connection point, Yellow: Ia, Green:
Ib, Blue:Ic, Pink:Va - (20 A/Div, 100 V/Div ). . . . . . . . . . . . . . . . . . . . . 109
5.14 Implemented MG simulator for scalability and comparison scenarios . . . . . . . . 111
5.15 Scalability plug-and-play test. a) Normalised reactive power consensus - Phase a,
b) Normalised reactive power consensus - Phase b, c) Normalised reactive power
consensus - Phase c . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.16 Scalability plug-and-play test. a) Frequency regulation, b) Average voltage regula-
tion. The dashed cyan lines represent the predefined band limits for both variables. 114
5.17 Optimisation time for the scalability test . . . . . . . . . . . . . . . . . . . . . . . 115
5.18 Comparison between the proposed DMPC scheme and a DAPI-based method. a)-
b) Unbalanced Power for the two methods compared, c)-d) PVUR index of the
voltage at the DGs output for the two methods compared. . . . . . . . . . . . . . . 116
5.19 Comparison between the proposed DMPC scheme and a DAPI-based method. a)-
b) Normalised reactive power in the phase a for the two methods compared, c)-d)
Normalised reactive power in the phase b for the two methods compared. . . . . . 117
5.20 Comparison between the proposed DMPC scheme and a DAPI-based method for
τi j = 1s. a)-b) Normalised reactive power in the phase a for the two methods com-
pared, c)-d) Normalised reactive power in the phase b for the two methods compared.117

B.1 Electrical output circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

F.1 Topology of Triphase units PM15F120C and PM5F60R used to emulate the ac-
microgrid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

xii
List of Abbreviations

ADL Active damping loop.

APF Active power filter.
BESS Battery energy storage system.
CSI Current source inverter.
DAPI Distributed averaging proportional integral.
DER Distributed energy resources.
DG Distributed generator.
DMPC Distributed model predictive control.
ESS Energy storage systems.
EV Electric vehicle.
FCS-MPC Finite control set model predictive control.
FPGA Field-programmable gate array.
HIL Hardware in the loop.
H-MG Hybrid ac/dc Microgrid.
ICC incremental cost consensus
LC Inductive capacitive filter.
MAS Multi-agent system
MIMO Multiple input - multiple output.
MG Microgrid.
MPC Model predictive control.
PE Power electronics.
PI Proportional integral.
PLL Phase locked loop .

xiii
PR Proportional resonant.
PV Photovoltaic panels .
PVUR Phase voltage unbalance rate index.
PWM Pulse-width modulation.
QP Quadratic programming.
QSG Quadrature signal generator.
RH Rolling Horizon.
RT Real time.
RTT Real time target.
SISO Single input - single output.
SOC State of charge.
VSC Voltage source converter.
WT Wind turbine.

xiv
List of Symbols

x={a,b,c} Phases of the MG.

Qix Per-phase reactive power of DGi .
Qi Three-phase reactive power of DGi .
Vix Per-phase voltages at the output of DGi .
Vi Average voltage at the output of DGi .
Viabc DGi output voltage in the natural reference frame.
N Number of DGs.
Li Coupling inductor.
Bi Nominal admittance.
ωi Angular speed at the output of DGi .
θi Phase angle at the output of DGi .
V̂ixB Per-phase estimated voltages at the coupling point.
V̂iB Average estimated voltage at the coupling point.
ω̂iB Angular frequency at the coupling point.
θ̂iB Phase angle at the coupling point.
δ θi Phase angle deviation of DGi .
∆ω s,i Frequency control action variation.
∆V s,ix Per-phase voltage control action variations.
ωs,i Frequency control action.
Vs,ix Per-phase voltage control actions.
ω0 Nominal frequency.
Mpω,i Active droop slope.
Pi Three-phase active power contribution of DGi .
V0 Nominal voltage.
Mqv,i Reactive droop slope.

xv
A Adjacency matrix.
ai j Communication term between DGs.
τi j Communication delay.
τ̂i j Estimated communication delay.
Ny Prediction horizon.
Nu Control horizon.
λ Weighting parameter in the cost function.
Si max Maximum apparent power of DGi .
ωaux,i Auxiliary optimisation variable for frequency.
Vaux,i Auxiliary optimisation variable for voltage.
Tsec Predictive controller sample time.
ωi Local average frequency approximation.
ωmax Maximum limit for the local average frequency approximation.
ωmin Minimum limit for the local average frequency approximation.
Vi Local MG average voltage approximation.
Vmax Maximum limit for the local average voltage approximation.
Vmin Minimum limit for the Local average voltage approximation.
X p,i Predicted variables optimisation vector.
X∆,i Predicted control action sequence vector.

xvi
Chapter 1

Introduction

1.1 Research motivation

The way energy is generated around the world is going through significant changes to cope with
global warming and cut CO2 emissions. Governments and societies are setting stringent environ-
mental goals for this purpose. For instance, during the COP26, the objective of “NET ZERO" was
accorded [1]. Due to the advances in technology development and power electronics, currently,
it is cheaper to generate electricity from renewables than fossil fuels. The United Kingdom (UK)
is leading the energy transition from contaminating generation sources to more environmentally
friendly generation by integrating massively distributed generation mainly based on renewable re-
sources. Currently, renewables account for 43% of the UK’s domestic power generation [2]. More-
over, the UK has set the ambitious target to run 100% on renewables by 2035. Similarly, Chile is
the leading country in Latin America in the adoption of generation based on renewable resources.
Currently, 36.5 % of generation comes from renewables [3]. Moreover, during 2022, 86% of the
total installed and to be installed generation capacity correspond to solar or wind energy [4]. It is
worth noting that Chile has the best solar ration in the world due to its privileged location. In this
context, the study of microgrids (MGs) is essential, as MGs allow the integration and management
of traditional and renewable generation sources, such as wind and photovoltaic.

A MG is defined as “A group of interconnected loads and distributed energy resources (DER)

that acts as a single controllable entity with respect to the main grid" [5]. MGs can operate in grid-
connected mode and isolated mode. In grid-connected mode, frequency and voltage are fixed by
the main grid. In contrast, the isolated mode of operation poses a more complex scenario because
the number of available assets is limited, and the MG has to take care of everything by itself [6].
Nevertheless, isolated MGs are one of the most attractive options to bring electricity to communities
not connected to the national grid.

1
 In MGs, the generation resources are local, and transmission power losses are reduced [7]. MGs
mainly operate at the distribution level with low voltage levels, and can incorporate renewables
with energy storage systems (ESS) and ac and dc loads [7]. Smart loads, such as electric vehicles
(EV), can also be managed within a MG. To manage the generators and loads in MGs, efficient
and reliable control strategies need to be developed. In recent years, there has been an extensive
research effort to improve the management of MGs and enhance their capabilities [8,9]. Distributed
model predictive control (DMPC) has been proposed as a prominent solution for MGs management;
it is able to coordinate the MGs’ assets to achieve several global objectives simultaneously.

According to their specific electrical distribution structure, MGs can be classified as ac [10], dc
[11] and hybrid ac/dc MGs (H-MGs) [10]. Both ac and dc MGs can comprise renewable-based gen-
erators, such as wind turbines (WT) and photovoltaic panels (PV), as well as non-renewable-based
generators, such as diesel generators. Renewable-based generators tend to be connected through
power converters to the distribution network. On the other hand, H-MGs can reduce unnecessary
conversion stages and increase the power capacity and the reliability of the entire MG [12]. A H-
MG comprises an ac sub-MG and a dc sub-MG connected through interlinking converters (ILCs).
In principle, the MGs described can operate connected to the main grid or in an isolated mode of
operation [10].

To ensure the proliferation of MGs, there are still some pressing issues related to their control
and operation. Currently, there is an ongoing effort to ensure that MGs are secure, reliable, and
operate cost effectively.. Moreover, as MGs are low voltage networks, they inherently have unbal-
ance between their three phases [13]. These unbalances are the result of asymmetrical impedances
per phase and the constant turning on or off of single-phase loads. Unbalanced loads can cause a
reduction in the efficiency of grid assets and may affect the stability of the MG [14]. A detailed
discussion of the principles and state-of-the-art of MG’s control is presented in Chapter 2.

1.2 Problem statement

Microgrids are driving the integration of distributed generation (DG) units, and transforming the
traditional centralised power grid. MGs (including H-MGs) inherit the three-level hierarchical con-
trol structure of traditional power systems. The primary control level compensates for load changes
by quickly changing voltage and frequency deviations to maintain the MG’s stability (droop con-
trol), while the main task of the secondary level is to then slowly restore the aforementioned vari-
ables to their nominal values [15]. The tertiary level is typically in charge of the economic dispatch
of generation (usually based on cost and availability) and the coordination of neighbouring MGs
[15]. The economic dispatch of a MG consists of minimising the total generation cost of satisfying
the power demand. To achieve this goal, the output power of generators must be determined in

2
order to satisfy demand at the lowest cost while maintaining the generation equipment’s operating
constraints [16, 17]. However, latest research for dc MGs [18, 19], ac MGs [16, 17, 20] and H-MGs
[21, 22] has shown that, because of the vulnerability of isolated MGs to rapid fluctuations in gen-
eration and demand, the economic dispatch of DGs should be accomplished on a timescale that is
consistent with the secondary control level.

However, another pressing problem that affects the power quality of a MG is the presence of
unbalances, as MGs operate at low and medium voltage levels, i.e. distribution level [23, 24].
Unbalances can be compensated (mitigated) through additional hardware or shared among the DGs
that compose the MG. For the first option, active power filters (APFs) can be employed; however,
this option is effective only when the issue is concentrated at a specific node. The second approach
is a more viable option where the DGs’ controllers have the ability to share the imbalances by using
their available capacity.

As explained before, the compensation of imbalance can be achieved using APFs to compen-
sate for unbalanced currents or unbalanced voltages at specific points of the MG [25,26]. However,
APFs are not attractive in MGs since they constitute additional hardware and higher costs. A more
cost-effective solution is to embed imbalance compensation capabilities into the control schemes
of DG units that are already available in the MG [27–29]. Control schemes to improve the shar-
ing of unbalanced powers between the DGs of MGs are mainly based on droop control and use
virtual impedance loops. Virtual impedances are used to change the dynamic of the power con-
verter using loss-less software implemented impedances. This means that negative sequence (and
zero sequence for four-wire MGs) impedances are implemented to control the sharing of imbal-
ance between the DGs. The magnitude of these virtual impedances is controlled via decentralised
control schemes in [30–32], meaning that there is no coordination between the DG units (each DG
works autonomously based on variables measured locally). However, better performance could be
achieved via coordination between DG units (e.g. using centralised and distributed approaches).

In this sense, the magnitude of the negative sequence impedances is calculated in a coordinated
way by a secondary centralised controller in [33–36], while in [37–39] secondary distributed con-
trollers, based on consensus algorithms, are implemented. It is important to note that the sharing of
imbalance can only be achieved by increasing the voltage imbalance at the output of the DG unit.
Therefore, a control strategy should share unbalanced powers and, at the same time, regulate the
maximum unbalanced voltage at the DGs to fulfil the maximum values stated in the IEEE standard
1547-2018 [40].

Focusing on the secondary level, there are three types of control architectures: centralised, de-
centralised, and distributed [41]. Centralised controllers require communication between all the
DGs and ILCs (which is impractical for large systems), presenting a high computational burden

3
and a single-point-of-failure [10, 41]. On the other hand, decentralised controllers only handle
local information, and allow limited coordination between DGs and ILCs [10]. Conversely, dis-
tributed schemes present a more compelling solution, achieving global objectives via coordination
of DGs by considering only information from communicating neighbouring DGs [10, 41, 42], thus
facilitating a plug-and-play operation.

The vast majority of distributed control proposals at the secondary level are based on proportional-
integral (PI) controllers [16–19, 43]. Furthermore, most approaches in the literature assume fixed
operational set-points for voltage and frequency and do not take advantage of the flexibility as-
sociated with the secure operational bands defined in the IEEE standard 1547-2018 [40], which
suggests that DGs can operate normally as long as the frequency and the voltage are within 1% and
5% of their nominal values, respectively. Furthermore, in the case of H-MGs, existing approaches
are designed independently for either ac sub-MGs [16, 17, 44] or dc sub-MGs [18, 19], without
accounting for the specific aspects of H-MGs themselves. It is worth noting that all the aforemen-
tioned works assume that frequency and voltages must be set to fixed nominal values, giving up
flexibility in the microgrid control system. Also, being based on PI controllers, it is difficult to
achieve multiple objectives and cope with operation constraints [45, 46].

Due to the limitations of PI controllers at the secondary control level for MG and H-MG ap-
plications, DMPC has attracted the attention of the MG community. DMPC is based on a model
of a local system and the prediction of its behaviour over a prediction horizon. Each local con-
troller computes a control sequence based on its local measurements, and information received
from neighbouring controllers [47], reducing the computational burden. The information is up-
dated, and the process is repeated at each sample time (rolling horizon). The rolling horizon prop-
erty compensates for communication delays [20,46,48]. DMPC can model complex multi-variable
systems, control multiple objectives, and handle hard and soft constraints [45,46,48]. For the oper-
ation of MGs, DMPC can handle DGs and ILCs (for H-MGs), equipment power rating limits (hard
constraint), and can regulate variables such as frequency and voltages within secure bands (soft
constraints) instead of specific values, making the MG operation more flexible. For these reasons,
DMPC is one of the most prominent solutions for managing MGs.

Based on the motivation described above, this thesis proposes three control strategies at the
secondary level to address the main tasks of the secondary control level (restoring frequency and
voltage). These strategies are able to restore the frequency and voltage to nominal values, or within
secure bands that comply with IEEE standard 1547-2018 [40]. Moreover, the proposed strategies
include complementary objectives in their formulation depending on the type of MG studied. For
instance, the first proposed strategy considers the economic dispatch of DGs in balanced ac MGs.
The second strategy achieves the economic dispatch of ac DGs, dc DGs and manages the power
transference of ILCs based on an economic criterion in H-MGs. Finally, the last proposed strategy

4
is able to manage imbalance sharing in unbalanced ac MGs while the maximum unbalanced voltage
at the DGs output is regulated to fulfil the maximum values stated in the IEEE standard 1547-2018
[40].

Local prediction models based on droop control and power transference equations are devel-
oped for the proposed strategies. Moreover, multiobjective cost functions are formulated to address
global objectives via information sharing while important operational constraints, such as equip-
ment power rating, are met. As a result, all the proposed strategies, tested by simulation and exper-
iment, are able to operate in “plug-and-play" manner, are robust against communication issues and
present a low computational burden.

1.3 Hypotheses
The hypotheses that support this thesis are described as follows:

(i) Distributed model predictive control strategies can be designed and implemented at the sec-
ondary control level to tackle the main issues of MGs, which are the regulation of frequency
and voltage, economic dispatch, and phase imbalance sharing. These strategies can include a
detailed mathematical model of the dynamic of the DGs to solve the aforementioned issues.

(ii) It is possible to achieve economic dispatch, and restoration of voltage and frequency in ac
MGs at the secondary control level through distributed predictive controllers that share their
information to coordinate their control sequences. In this way, in addition to avoiding the
need to have a controller for each objective, the overall performance of the microgrid is
enhanced in terms of robustness and reliability.

(iii) It is possible to achieve economic dispatch in hybrid ac/dc MGs, along with the main ob-
jectives of the secondary control, which are frequency and ac voltage restoration on the ac
sub-MG and dc voltage restoration on the dc sub-MG. Additionally, is it possible to restore
these variables within secure bands that comply with the IEEE standard 1547-2018 [40], in-
stead of restoring them to nominal values. In this way, more flexibility is given to the the
hybrid ac/dc MG.

(iv) The interlinking converters in a hybrid ac/dc MG can communicate with ac DGs and dc
DGs to transfer power from the ac sub-MG to the dc sub-MG and vice-versa based on the
economic dispatch criterion. Moreover, power rating limits in DGs and ILCs can be included
in the predictive controller to avoid overloads.

(v) The inclusion of the dynamics of droop controllers, active power and reactive power transfer
models in a predictive controller will allow the optimisation of the power contribution of each

5
 DG to the microgrid. Furthermore, a distributed control structure based on measurements
and information sharing can avoid the complete modelling of a microgrid and face any load
variations within a microgrid’s physical capacity limits.

(vi) It is possible to share imbalances among DGs’ phases in unbalanced ac MGs, avoiding the
use of virtual impedance loops and without the need for adding extra power converters to the
MG through the use of a distributed predictive control scheme that also restores the frequency
and voltage. Moreover, it is possible to limit the voltage unbalance at the output of each DG
unit to comply with the IEEE standard 1547-2018 [40] within the same controller.

(vii) A set of distributed controllers can reduce the computational burden and provide a response
of equal quality to a centralised controller, and will allow plug-and-play operation (i.e. dis-
connect and reconnect generation units without the need for external intervention in the con-
trollers). They can also provide good performance when communication delays are present
in the system.

1.4 Objectives

1.4.1 General objective

This PhD research thesis aims to study the design, modelling, implementation and validation of
novel distributed predictive strategies for the secondary control level of MGs, which adds to the
frequency and voltage restoration, the objectives of economic dispatch and imbalance sharing.
Within the framework of this work, the following specific objectives are pursued.

1.4.2 Specific objectives

(i) To design, implement and validate experimentally a distributed predictive control strategy for
ac MGs to achieve economic dispatch of DGs and frequency restoration to nominal values.

(ii) To design, implement, and validate a distributed predictive control strategy for hybrid ac/dc
MGs to achieve economic dispatch of DGs and restore within bands the frequency and ac
voltage on the ac sub-MG and the dc voltage on the dc sub-MG.

(iii) To design, implement, and validate a distributed predictive control strategy for unbalanced
ac MGs to achieve imbalance sharing among DGs and restore within bands the frequency
and voltage while the unbalance at the DGs’ output is kept within the recommendation of the
IEEE standard 1547-2018 [40].

(iv) To validate the proposed DMPC strategies against the most demanding scenarios, which are

6
 communication delays, communication failures and disconnection/reconnection of DGs. The
validation is carried out via simulation, hardware-in-the-loop (HIL) and experimentally.

(v) To compare the performance of the proposed control strategies with the latest distributed
strategies reported in the literature.

1.5 Contributions
The work developed during this project has resulted in the publication of three journal papers sub-
mitted to top-tier indexed journals and two international conference papers. In addition, twelve
manuscripts have been published, with the candidate as a co-author. The details of these pub-
lications related to microgrids are listed in Chapter 6. The contributions of this thesis can be
summarised as follows:

For balanced ac MGs

(i) A novel DMPC scheme is proposed for active power economic dispatch and frequency
restoration where both objectives are achieved simultaneously.

(ii) The proposed controller neither requires the modelling of the entire MG nor the modelling of
the connected loads. Furthermore, this controller operates with the same usual measurements
used at the primary control level; thus, the number of physical measurements is reduced when
compared with previous predictive approaches [49].

(iii) The proposed DMPC scheme with the economic dispatch of DGs addresses communication
delays, loss of communications and plug-and-play scenarios without requiring changes in the
control structure, unlike centralised MPC schemes. This validation was carried out with the
experimental setup of the MGs laboratory at The University of Chile.

(iv) The DMPC includes as equality constraints the droop, the active power transfer, and the phase
angle models to predict the behaviour of each DG. Additionally, terminal values and inequal-
ity constraints contribute to bound the feasible solution space. This reduces the optimisation
time and enables the control strategy to be implemented in real-time controllers.

For hybrid ac/dc MGs

(i) A cooperative DMPC scheme that considers the interaction among ILCs, ac DGs and dc DGs
in isolated H-MGs is proposed. This DMPC scheme controls the H-MG as a single entity
instead of three separated systems, giving redundancy to communications and improving the

7
 controller’s dynamic response.

(ii) This work considers a DMPC strategy to control variables to specific values and within oper-
ation bands for H-MGs. For this purpose, novel cost functions are proposed to control ILCs,
ac DGs and dc DGs. The proposed DMPC for secondary control can achieve accurate con-
sensus objectives, i.e., economic dispatch of active power and reactive power. While through
the use of soft constraints, the following variables are considered as flexible: the frequency
and average voltage of the ac sub-MG and the average voltage of the dc sub-MG. They
are regulated within predefined bands that comply with the IEEE standard 1547-2018 [40].
Furthermore, by including maximum power rating constraints in the DMPC formulations,
physical saturation (overloading) of ac DGs, dc DGs, and ILCs is prevented.

(iii) The proposed DMPC considers the existence of multiple ILCs. Where the cost function of
the ILCs achieves simultaneously the cost-effective operation of the H-MG, averts overload-
ing ILCs and avoids circulating currents. Moreover, to predict the behaviour of ILCs and
DGs, the dynamic models that rule each of them are included as equality constraints in their
respective optimisation problems.

(iv) Extensive simulation studies validate that the proposed DMPC has good performance under
load change, plug-and-play of DGs and ILCs, and communication delays scenarios. Also,
this work shows that implementing the economic dispatch at the secondary level reduces the
operation costs of a H-MG considerably.

For unbalanced ac MGs

(i) This work proposes a DMPC scheme for unbalanced MGs. The proposal improves the shar-
ing of imbalances among DG phases in ac MGs, avoiding the use of virtual impedance loops
(as this methodology has many drawbacks [23]) and without the need for adding additional
power converters to the MG.

(ii) With the DMPC mathematical model, which includes soft constraints, the proposed imbal-
ance sharing control scheme can achieve accurate control of some variables in the MG,
whereas other variables are controlled within more relaxed predefined bands. Specifically,
the frequency and average voltage are regulated within predefined bands. This produces a
more flexible control system than those reported in [30, 33–39, 50–53], which look for an
accurate sharing of all the variables of the MG.

(iii) The proposed DMPC approach can improve the sharing of both three-phase active and reac-
tive power, and single-phase reactive power between the DG units. This can not be included
in methods based on virtual impedance loops, as will be shown in Section 5.7.2.

8
 (iv) The proposed control scheme achieves the sharing of imbalance, reducing the single-phase
voltage deviations at the output of each DG unit, when compared with methods based on the
virtual impedance loop. The proposed approach has better performance in the presence of
time delays in the communication network and complies with the plug-and-play capability.
The good performance of the proposed controller was validated via HIL and simulation tests.

1.6 Thesis structure

The remainder of this thesis is divided in the following chapters:

Chapter 2 presents a background of MGs’ control and an extensive review of secondary control
techniques for MGs and H-MGs reported in the literature. Particular attention is given to consensus-
based distributed schemes and distributed predictive schemes for economic dispatch of DGs and
imbalance sharing.

Chapter 3 proposes a predictive secondary control strategy for ac MGs that simultaneously
achieves economic dispatch and restores the MG frequency to its nominal value. The mathematical
formulation of this proposal is detailed. Furthermore, the experimental validation of this technique
is addressed.

Chapter 4 proposes a predictive secondary control strategy for hybrid ac/dc MGs that simul-
taneously achieves economic dispatch of active and reactive power and restores frequency and ac
voltage on the ac sub-MG and dc voltage on the dc sub-MG within secure bands. The mathematical
formulations for ac DGs, dc DGs and ILCs of this proposal are detailed. Furthermore, simulation
validation of this technique is addressed as well as a performance comparison against other reported
technique in the literature.

Chapter 5 proposes a predictive secondary control strategy for unbalanced ac MGs that si-
multaneously achieves single-phase imbalance sharing and restores frequency and voltage within
secure bands. The mathematical formulation of this proposal is detailed. Furthermore, simula-
tion and hardware-in-the-loop validation of this technique are addressed as well as a performance
comparison against other reported technique in the literature.

Chapter 6 presents the conclusions, suggests possible future research lines on the topics ad-
dressed in this work, and presents a summary of the works published during the PhD studies.
Specifically, the manuscripts derived from the work presented in this thesis and MG-related papers
where the candidate has contributed are presented.

9
Chapter 2

Literature review

2.1 Introduction
This chapter presents the state-of-the-art related to control strategies proposed for microgrids (MGs).
Particular attention is given to consensus-based distributed strategies and distributed model predic-
tive (DMPC) schemes. Also, proposals for economic dispatch and imbalance sharing are discussed
extensively.

This chapter is organised as follows: Section 2.2 provides the background of the hierarchi-
cal control structure used in MGs. Section 2.3 provides a detailed explanation of the controllers
involved in the hierarchical control structure for ac MGs. Similarly, Section 2.4 presents the con-
trollers for the hierarchical control structure for dc MGs. The control of hybrid ac/dc MGs is
presented in Section 2.5.

Then the latest reported schemes at the secondary control level for the aforementioned types of
MG are discussed in Section 2.7. Special emphasis is given in this section to approaches that solve
economic dispatch. Section 2.8 presents the main approaches to solve imbalance sharing in MGs.
Then, the related solutions based on DMPC for the secondary level are discussed in Section 2.6.
Finally, Section 2.9 summarises the state of the art and main benefits of the proposed strategy.

2.2 Microgrids framework

Microgrids (MGs) are an essential feature of future power systems, as they enable the full inte-
gration of distributed energy resources (DERs). These DERs are mostly based on renewable-based
generation sources and aim to gradually leave behind the generation based on fossil fuels and decar-
bonise the planet. Harvesting electricity from renewables is now a viable option as their generation

10
prices decrease day by day. Moreover, MGs can become the key enablers to achieve goal number 7
of the United Nations, which states: ensure access to affordable, reliable, sustainable and modern
energy for all. As mentioned before, a MG acts as a single controllable entity from the grid point
of view and has its own generation and storage resources with clearly defined electrical boundaries
[5]. MGs aim to achieve adequate active and reactive power sharing among distributed generation
(DG) units, while frequency and voltage levels must be kept within safe levels [54]. Therefore,
MGs represent an attractive solution in many applications, such as terrestrial, naval or aerospace
electrical grids, due to their controllability, capability to include DGs and flexibility [55].

MGs can operate in both grid-connected or island modes. In the grid-connected mode, the
MG can trade energy with the main grid based on the existing power deficit or surplus. Islanded
MGs can operate independently, but they can reconnect to the main grid when necessary. On the
other hand, MGs that do not have a connection with the grid because of technical, geographical or
economic constraints are usually called isolated MGs [6]. Isolated MGs must be controlled with
extreme care because the number of DGs available to tackle voltage and overloading problems is
limited [6]. Isolated MGs must autonomously regulate voltage amplitude and frequency through
the MG’s control system. For most DERs, power electronics (PE) interfaces are needed to connect
the DERs to a MG. The sinusoidal rectifier is the most common configuration used [15]. These
DERs must be controlled in a coordinated manner with care, as they do not have inertia, which can
influence the stability of a MG.

According to their specific electrical distribution structure, MGs can be classified as ac MGs
[10], dc MGs [11] and Hybrid ac/dc MGs (H-MGs) [10]. Isolated ac MGs operate on a grid form-
ing/supporting scheme where at least one DG works as a voltage source converter (VSC), regulating
the MG voltage magnitude and frequency [56]. A general topology of an ac MG is presented in
Fig. 2.1a. On the other hand, dc MGs neither require frequency regulation nor reactive power con-
trol, thus reducing the system’s operational complexity. Additionally, dc DGs do not need to be
synchronised to the utility grid [10]. A general topology of a dc MG is presented in Fig. 2.1b. Both
ac and dc MGs can comprise renewable-based generators, such as wind turbines (WT) and pho-
tovoltaic panels (PV), as well as nonrenewable-based generators, such as diesel generators. These
generators can be connected directly to the distribution network (depending on their nature) or
through PE interfaces.

Conversely, H-MGs merge the benefits of both kinds of MGs, reducing unnecessary conversion
stages and increasing the power capacity and the reliability of the entire H-MG [12]. A H-MG
comprises an ac sub-MG (Fig. 2.1a) and a dc sub-MG (Fig. 2.1b) connected through interlinking
converters (Fig. 2.1c). The ILCs need to be controlled adequately to guarantee power flow between
sub-MGs. Moreover, the infrastructure of existing ac MGs can be re-utilised for the deployment of
H-MGs.

11
Figure 2.1: General topology for a hybrid ac/dc microgrid. a) ac microgrid. b) dc microgrid. c)
Interlinking converters.

12
 Typically, the control tasks of an isolated MG are split into three control levels, where each
control level operates at a different time-scale [15]. The first level, which is the fastest of all (mil-
liseconds), maintains the stability of the MG and ensures correct power sharing [54]. This control
level is comprised of inner current controllers, outer voltage controllers, and droop control loops;
the latter allows variations of active/reactive power to be reflected as variations of frequency/volt-
age, as in large power systems [15]. Droop control produces deviations in frequency and voltage
amplitude. At a slower time scale (seconds), the setpoint values of frequency and voltage can be
restored; the controllers used to manage this are known as the secondary control level [10]. The
long-term tasks of a MG (seconds to minutes), such as economic dispatch of DGs and coordination
of MGs with the main grid, are performed by the tertiary control [57].

It is worth noting that most of the control solutions proposed in MGs have been derived from
large-scale electric power systems. However, latest research for dc MGs [18, 19], ac MGs [16, 17,
20] and H-MGs [21, 22] has demonstrated that due to the susceptibility of isolated MGs to fast
changes in generation and demand, the economic dispatch should be performed in a time-scale
consistent with that of the secondary control level for it to be adequate. This is because MGs
possess low inertia due to the use of renewables and power electronics interfaces [58,59]. For these
reasons, the secondary control level will be studied in this thesis.

There are typically three types of control architecture proposed for secondary control; these
are centralised, decentralised and distributed [10], as shown in Fig. 2.2. Centralised control (see
Fig. 2.2a) can give a global solution, but it has a common point of failure in the communication
network and presents a high computational burden. Decentralised controllers do not need a com-
munication channel because each local controller takes actions based on their own measurements
(see Fig. 2.2b), but an optimal solution is difficult to achieve [10]. Finally, distributed controllers
can achieve global objectives through information sharing (see Fig. 2.2c). They are robust against
communication failures and allow plug-and-play connection for the various distributed resources
[60]. A detailed comparison between centralised and distributed control is presented in Table 2.1.

13
Figure 2.2: Control structures at the secondary control level. a) Centralised. b) Decentralised. c)
Distributed.

Table 2.1: Comparison between Centralised and Distributed Control

Features Centralised control Distributed control

Application Small microgrids. Large microgrids.

Microgrid topology is required. Microgrid topology is not required.

High computational cost. Low computational cost.

Operation
Scalability requires changes in control. Easy to scale.
[60, 61]
Decisions based on Decisions based on local measurements

communications with all DGs. and the communication with neighbours.

Single point of failure. Plug-and-play capability.

Reliability is degraded. Reliability is maintained.

Powerful hardware Embedded controllers.

Design and implementation
Complex algorithms. Easier algorithms.
[10]
Information from all units is required. Handle local and shared information.

Low-bandwidth communication and High-bandwidth communication

low communication complexity. and high communication complexity.

14
2.3 Control of ac MGs

2.3.1 Primary control level in ac MGs

The primary control level is in charge of voltage stabilisation and power sharing by fixing the
current and voltage at the DGs’ output. Usually, it is composed of an inner current loop (faster), a
voltage outer loop (slower) and droop controllers. A bandwidth separation of at least ten times is
needed between these controllers for their operation [62]. The current and voltage control loops can
be proportional integral (PI) controllers or proportional resonant (PR) controllers. When using PI
controllers, the variables are converted into the dq synchronous reference frame [57]. PR controllers
can be implemented directly in the abc natural reference frame or the αβ stationary reference
frame. Among the advantages of PR controllers over PI controllers, these controllers allow the
implementation of positive and negative sequence loops at the same time as well as the independent
control of phases [26].

The droop controllers dictate the main dynamic of this control level. The droop controllers
emulate the behaviour of classical synchronous machines (mechanical/electrical). An LCL filter
can be placed at the DGs’ output, where the second coupling inductance can be set to ensure a
predominantly inductive impedance [63, 64], as will be shown later in Section 3.3, and classical
droop controllers can be used at the primary control level. To achieve active power sharing, DG
units modify the frequency of the microgrid (see Fig. 2.3a), whereas to reach proper reactive power
sharing, DG units modify their voltage amplitudes (see Fig. 2.3b). As ESS are out of the scope
of this work, Fig. 2.3a considers unidirectional power. Although communication-based controllers
are proposed for this control level, these are not widely used because they present higher imple-
mentation costs and complexity [15, 65]. Moreover, they may complicate plug-and-play behaviour
[15, 65].

The equations that represent the droop controllers for frequency (ωi ) - active power (Pi ) and
voltage (Vi ) - reactive power (Qi ) of DGi are given in (2.1).

ωi (t) = ω0 + M pω,i Pi (t)

(2.1)
Vi (t) = V0 + Mqv,i Qi (t)

where ω0 and V0 are the setpoint frequency and voltage, respectively. M pω,i and Mqv,i are the
droop slopes of DGi . These parameters can be designed as shown in (2.2).

ω0 − ωmin Vmax −Vmin

M pω,i = − , Mqv,i = − (2.2)
Pmax Qmax − Qmin

15
 Figure 2.3: Droop control for ac DGs

where ωmin , Vmin and Vmax are the frequency and voltage amplitude limits, and Pmax , Qmin and
Qmax are the active and reactive power limits of DGi .

The main advantage of the droop control method is that communication is not required; instead,
the frequency and the voltage magnitude measurement indicates the overall deficit of active and
reactive power in the MG, respectively [66]. Distributed resources are automatically plug-and-play;
thus, no coordination is necessary, and the units can automatically adjust their set points to meet
the overall needs of the MG. Conversely, the main disadvantage of the droop method is that, due to
its operating principle, deviations in frequency and voltage are created; thus, additional controllers
with a time-scale separation are needed so that they do not compromise the droop behaviour.

Note that if frequency and voltage are not restored, the variability of loads and generation based
on renewable resources may cause considerable frequency excursions and, in the worst case, may
lead a microgrid to instability. This is why great emphasis is given to studying techniques that
guarantee voltage and frequency restoration in a microgrid, as described in the following section.

2.3.2 Secondary control level in ac MGs

The secondary control level has a lower bandwidth than the droop controller. Therefore, the com-
munication network bandwidth is lower. The main tasks of the secondary control level are to re-
store the frequency and the voltage amplitude. The secondary controller shifts vertically the droop
curves, as shown in Fig. 2.4. As explained before, there are three possible control structures at
this level which are centralised, decentralised and distributed controllers. Since the MG frequency
is a global variable, the frequency controller can measure the frequency at any node. However,

16
since the voltage is not a global variable, it is necessary to define the voltage used to implement the
secondary controller.

Figure 2.4: Secondary control for ac DGs

Furthermore, in a MG, the existence of line impedance creates voltage drops. Hence, regulating
all nodal voltages at the nominal value V0 does not allow appropriate power-sharing among DGs
[67]. For this reason, we aim for an average bus voltage regulation across the MG. In this sense,
the proposed control scheme achieves the control of reactive power while regulating the average
bus voltage across the MG [8, 49].

2.3.3 Tertiary control level in ac MGs

This control level is generally called Energy Management System (EMS) and operates on a large
time scale. It usually achieves the economic dispatch of the MG and its coordination with the main
grid when the MG is operating in grid-connected mode. Centralised and distributed controllers are
mainly used at this control level [65, 68]. They typically require the solving an offline optimisation
problem, although recently, MPC methods have been proposed at this level [69, 70]. Nevertheless,
these optimisation problems use external prediction models, which are subject to uncertainty in
both the power load and the generated power. Prediction models based on neural networks and
fuzzy systems are the most common models used [71].

In an effort to achieve real-time operation of the EMS, [69, 70] propose a variable resolution
(sample time) EMS, where the first prediction minutes are predicted with high accuracy (5 min),
and the resolution decreases as the forecasted horizon is extended (60 min) to include an entire
day of operation in the optimisation. These formulations consider optimal power flow and unit

17
commitment separately, and for proper operation, these EMS need the mathematical formulation
of the whole microgrid. Using the same approach described previously, [72] includes demand
side management (DSM), reduces CO2 emissions [73] and controls phase imbalance in an isolated
microgrid [74]. The three previously mentioned works reported better performance in terms of an
improved solution, less energy curtailed, reduced peak demand and improved load factor, but at the
cost of higher computational time.

The performance of control strategies based on MPC at the tertiary level has a strong dependency
on the accuracy of the prediction models, which in turn depends on the sampling time and aggrega-
tion level [75]. Contrary to bulk power systems, MGs can present generation fluctuations and rapid
load changes. This could cause the predictions at this control level to deviate from the MG’s real
operating condition. For this reason latest research for dc MGs [18, 19], ac MGs [16, 17, 20] and
H-MGs [21, 22] has demonstrated that the economic dispatch should be performed in a time-scale
consistent with that of the secondary control level. A review of the state-of-the-art for the control
of MGs is presented in Section 2.7.

2.4 Control of dc MGs

The control structure of dc MGs is similar to that described for ac MGs previously. The primary
control level is based on a decentralised droop control, while the secondary control level restores the
voltage deviation caused by the primary control. The tertiary control level exchanges energy with
the main grid when a connection is possible. In the next section, only the primary and secondary
control are discussed, as the tertiary control is similar to the one described for ac MGs.

2.4.1 Primary control level in dc MGs

The primary control level sets the current and voltage at the dc DGs’ output and shares power
among the dc DGs. These DGs are usually interfaced through dc-to-dc converters [76]. It is
composed of an inner current loop (faster), voltage outer loop (slower) and droop controllers; the
current and voltage control loops are usually proportional integral (PI) controllers. A bandwidth
separation of at least ten times is needed between these controllers for their operation [62].

As mentioned before, the main dynamic of this control level is dictated by the droop control. For
dc DGs, current-voltage and power-voltage droop curves have been proposed [10]. The power P-
voltage (V ) droop allows the direct control of the power supplied by DGs and eases the management
of dc DGs in hybrid ac/dc MGs.

In this case, to achieve active power sharing, DG units modify the output voltage (see Fig. 2.5).
As ESS are out of the scope of this work, Fig. 2.5 considers unidirectional power.

18
 Figure 2.5: Droop control for dc DGs

The equation that represents the droop controller Vi − Pi of DGi is given in (2.3).

Vi (t) = V0 + M pv,i Pi (t) (2.3)

where V0 is the setpoint voltage, and M pv,i is the droop slope of DGi . The slope can be designed
as shown in (2.4).

V0 −Vmin
M pv,i = − (2.4)
Pmax

where Vmin is the minimum voltage limit, and Pmax is the active power limit of DGi .

2.4.2 Secondary control level in dc MGs

The secondary control has a lower bandwidth than the droop controller. The main task of the
secondary control level is to restore voltage magnitude [10]. The secondary controller shifts the
droop curves vertically, as shown in Fig. 2.6. As explained before, there are three possible control
structures at this level: centralised, decentralised and distributed controllers (see Fig. 2.2) [8, 10].

Since the voltage is not a global variable, it is necessary to define the voltage used to implement
the secondary control. In a MG, the existence of line resistance creates voltage drops. Hence,
regulating all nodal voltages at the nominal value V0 does not enable appropriate power-sharing
among the dc DGs. For this reason, we aim at an average bus voltage regulation across the dc MG.

19
In this sense, the proposed control scheme achieves the sharing of active power while regulating
the average bus voltage across the MG [8, 49]. A review of the state-of-the-art of MG’s control is
presented in Section 2.7.

Figure 2.6: Secondary control for dc DGs

2.5 Control of hybrid ac/dc microgrids

As described in Section 2.2, a hybrid ac/dc microgrid (H-MG) is composed of an ac sub-MG and
a dc sub-MG connected through bidirectional interlinking converters (ILCs), as shown in Fig. 2.1.
The controllers of the sub-MGs were explained in Section 2.3 and Section 2.4. Therefore, only the
control of the ILC is explained in the following.

ILCs control the power transference in a H-MG. There are two main approaches for controlling
the power transferred through an ILC. The first one is based on the difference of the normalised
deviations of the primary variables with respect to minimum and maximum allowed values, i.e.,
frequency of the ac sub-MG and voltage of the dc sub-MG [77, 78], which is fed to a PI controller.
However, this approach losses accuracy when a secondary control regulates these variables.

The second approach includes the power management task at the secondary control level [21,
22, 38, 43, 79], where diverse objectives can be defined, such as power sharing [38, 43] and eco-
nomic dispatch [21, 22]. Most of the existing approaches consider that the H-MG is composed of
three independent systems [21, 22, 38, 77–79] (ac sub-MG, dc sub-MG, and ILCs) and neglect the
dynamic interactions between them, which can degrade the dynamic response of the controllers.
Furthermore, these formulations do not consider limits for the operation of the DGs and the ILCs.
It is worth noting that no distributed model predictive control approaches for economic dispatch of

20
H-MGs have been reported in the literature as yet. A review of the state-of-the-art of MG’s control
is presented in Section 2.7.

2.6 Distributed model predictive control for microgrids

Proportional integral (PI) controllers tend to be used at the secondary level [67]. However, thanks to
advances in communication speeds and improved hardware capabilities of controllers, model pre-
dictive control (MPC) is being incorporated at the secondary control level, improving the overall
performance of microgrids as these controllers are multi-objective and robust against communica-
tion delays [49, 80]. MPC is becoming one of the most successful advanced control techniques
implemented in industry [81, 82] due to its ability to handle complex systems with input and state
constraints. MPC is emerging as a useful control strategy in the microgrids community [49, 68, 83]
because it is a multi-variable constrained control scheme which obtains a real-time solution at each
sample instant (rolling horizon). MPC consist of an objective function and a model of the process
being controlled. The control includes physical and dynamic constraints [84] which, together with
the objective function, are used to perform an online optimisation of the overall system. Among
the advantages of MPC, the following are highlighted: modelling of complex systems, optimal so-
lution, handling of communication delays, and better control of transients [49, 83, 85]. It is worth
noting that MPC needs a good model of the system for the controller to work properly.

Reported applications of MPC to the MG’s secondary control can be divided into centralised
and distributed MPC algorithms. Centralised secondary MPC controllers consider the MG as a
single system and perform an online optimisation in order to guarantee a good performance, to-
gether with voltage, frequency and power regulation [68, 84]. However, a large communication
network and high computational capability are necessary to compute the control actions. On the
other hand, Distributed MPC (DMPC) solves local optimisation problems with information shared
through a communication network, reducing the computational burden and traffic over the commu-
nication network. DMPC directly controls individual loads or DGs connected to the MG, requiring
only partial information of the overall MG status. In this way, the computational burden is dis-
tributed across the MG, and large communication networks are avoided [49, 83, 85]. As DMPC is
a distributed strategy, only the complexity of the communication network is increased.

MPC can compensate for communication delays and uncertainties due to its principle of opera-
tion, i.e., the rolling horizon mechanism [46, 86]. At each sample time, the controller receives new
measurements and information from neighbouring units, then new control action sequences are
computed, but only the first control actions are applied to the system. Then the whole procedure
is repeated, which means that a feedback mechanism is incorporated by using new measurements
and neighbouring information to update the optimisation problem for the next sample time.

21
 It has been reported in [48, 49, 86, 87], that conventional controllers at the secondary level, such
as proportional-integral-based (PI-based) ones, may not be robust enough to guarantee good and
stable operation in the presence of uncertain and relatively large communication delays. Indeed,
traditional PI-based techniques are more sensitive to delays. In such controllers, the closed-loop
performance is sacrificed to endure the effects of delays [88].

The larger the delay, the more the performance is sacrificed, giving a slow dynamic response.
In these controllers, if the bandwidth is not reduced, a poorer behaviour will result, i.e., significant
overshoot and oscillations. Therefore, classical PI-based techniques which rely on simple gain
changes are often not appropriate for systems with uncertain and potentially large delays. This
phenomenon will be corroborated in this thesis through comparison tests.

Previous studies have demonstrated via small-signal modelling [87], simulation work [89] and
experimentally [20, 49] that MPC and DMPC can handle a wider range of delays. MPC reduces
the oscillations in the system, while other control strategies can have a predominant oscillating
behaviour. These studies concluded that MPC is more robust in terms of the maximum unknown
delay allowed. Therefore, in the presence of communication delays, MPC is able to compute an
optimal control sequence that minimises the cost function to obtain a smooth dynamic response in
the MG.

Recently, distributed control schemes at the secondary control level based on DMPC have been
reported for MGs [49, 80, 90, 91]. The authors of [80] present a feedback linearization DMPC
for frequency and voltage restoration, considering the voltage and current at the LC filter output
as state variables. The authors of [90] present a DMPC controller for voltage restoration, while
frequency is restored using a variation of the distributed averaging proportional-integral (DAPI)
controller with a finite-time observer. The authors of [91] present a DMPC controller based on
a consensus version of the alternating direction method of the multipliers algorithm to regulate
frequency in a networked MG system by manipulating the voltages of voltage-sensitive loads. The
authors of [49] propose and validate experimentally a DMPC for frequency and voltage restoration
using droop models and power transfer models; also active and reactive proportional power sharing
is considered based on the concepts of [67]. The power sharing is based on consensus over the
active and reactive power contributions from each generation unit in the MG using an adjacency
matrix. Using external measures this controller avoids the necessity to model the MG topology.
Only [92] proposes a DMPC for economic dispatch and frequency restoration via simulation at
the secondary level. The work of [92] includes both operation and maintenance costs within its
formulation. However, it does not use consensus for the economic dispatch, and it assumes an ideal
communication network. Nevertheless, none of the DMPC methods proposed for the secondary
level take into account the overall economic performance of the MG via a consensus strategy.

22
 All of the previously reported works proposing secondary MPC or DMPC controllers for MGs
have been developed considering balanced MGs [49, 68, 83–85]. However, when looking at the
secondary control level, the low voltage ac MGs used for the distribution of electrical energy are
inherently (phase) unbalanced systems since they usually have to feed unbalanced loads, leading
to significant challenges for the secure and reliable operation of the MG. Adding this to the usual
aims of voltage and frequency regulation and the improvement in the sharing of both active and
reactive power introduces additional control challenges to the MG secondary control [33, 93].

2.7 Distributed secondary control for microgrids in the litera-

ture
In distributed control, the consensus and cooperative control algorithms have attracted the attention
of the MG community [60]. The multi-agent system (MAS) cooperative control method, in par-
ticular, aims to achieve system objectives cooperatively by mimicking the behaviour of biological
phenomena [94]. Controllers based on MAS at the secondary level have been reported. For exam-
ple, the authors of [67] proposed the distributed averaging proportional integral (DAPI) controller,
which uses PI controllers for frequency and voltage restoration with an adjacency matrix that rep-
resents the communication network of the MG. The advantage of this controller is its distributed
structure; however, active power is shared proportionally among DG units.

The authors of [95] present a finite-time control approach to restore voltage and a DAPI con-
troller to restore frequency; the former controller guarantees a fixed time convergence for voltage
restoration and decouples the secondary controllers. A variation of the DAPI controller is presented
in [96] where the integrative part of the control actions are shared among the DGs. This controller
improves the transient of the proportional active power sharing among the DG units. However, as
these controllers are based on PI controllers, they neither include dynamic models of the DGs nor
include their physical power limits.

2.7.1 Distributed economic dispatch for ac microgrids

There is a clear tendency to integrate the economic dispatch of DGs at the secondary level and to
implement distributed controllers in isolated MGs [58, 67, 97]. In a distributed fashion [16, 58, 98]
demonstrated that the economic dispatch can be integrated into the same time-scale of frequency
and voltage restoration. The literature distinguishes between two main approaches to achieve the
distributed economic dispatch in MGs. The first one employs the distributed gradient method [99],
which directly calculates a global incremental cost through a consensus algorithm, whereas the
second one uses the incremental cost consensus (ICC) concept in which the incremental cost is
estimated [98, 100]. The first option requires several communication iterations per sample time to

23
find a solution which is hard to achieve in experimental microgrids. On the other hand, the ICC
presents a better approach because communication among DGs is required once per sample time.
In the following the related approaches based on the ICC are discussed.

The authors of [16] reformulate the optimisation problem to achieve the Karush-Kuhn-Tucker
(KKT) conditions of linear optimal power flow. Then, a control action from a PI controller is added
to a droop controller to achieve the economic dispatch of DGs. The authors of [98] achieved active
power and reactive power dispatch, including voltage and frequency regulation in a decomposition
of the optimisation problem. Nevertheless, these publications only provide simulation results. The
authors of [101] proposed a MAS economic resource management in an isolated microgrid. The
control strategy is tested under time-varying load conditions. Conversely, in [102] a two-stage
control strategy is proposed to achieve an economical operation and restore frequency. In contrast,
in [103], a grid-connected distributed resource management is presented. Two control levels are
needed to achieve adequate economic dispatch. As the microgrid considered is connected to the
distribution network, the regulation of frequency and voltage is not considered in the formulation.
Nevertheless, these strategies do not solve an actual optimisation problem and most of them do not
consider operating limits in their formulation.

The authors of [104, 105] propose a distributed economic dispatch that includes power balance
and ramp constraint restrictions using a distributed primal-dual consensus algorithm; the advantage
of the previous technique is its speed of convergence. Time-varying communication delays are
considered in the distributed optimisation through the decomposition of [106], but in this work,
renewable resources are considered as constants. In [100], a distributed finite-time economic dis-
patch is proposed; this technique guarantees a fixed convergence time, but its formulation does
not consider the regulation of voltage and frequency. Finally, [107] depicts a multi-agent system
(MAS) frequency regulation at a minimum cost controller and a threshold-based demand response
controller; this technique disconnects electrical load in emergency cases.

By contrast, [108] includes the minimisation of environmental objectives within its formulation,
and it is tested in the presence of communication delays and noisy communication channels. In
[109], BESS modelling is included in the formulation of distributed economic dispatch to provide
the service of arbitrage and spinning reserve. Although the controllers proposed present different
benefits, none of them is tested experimentally. As the previously described distributed controllers
only use various single-input single-output (SISO) PI controllers and rely on a fixed control law,
they do not guarantee an optimal solution, as they do not consider real-time changes in the operation
of the MG [45]. Furthermore, most studies do not solve an optimisation problem and do no not
consider physical constraints in their formulation. In addition, they do not provide compensating
mechanisms to address communication delays in the communication network.

24
Economic dispatch of reactive power

Usually, only the economic dispatch of active power is considered. However, neglecting the related
reactive power costs might result in increased operating costs and deviations from the optimal
solutions for dispatch. For instance, [110] co-optimised the dispatch of active and reactive power
in distribution networks. The results show that a reduction of more than 12 % of energy losses
could be obtained if reactive power is appropriately managed. The reactive power cost function, in
a deregulated electricity market, can be represented by the weighted coefficients of the cost function
of active power [111]. The weights are obtained through the power factor from the relation of the
triangle of powers.

Recent works [111–114] have included the dispatch of reactive power along with active power
in their formulation. The works of [111,112] consider a centralised controller at the tertiary control
level to address both active and reactive economic dispatch. The results show that considering the
costs associated with reactive power reduces the total operation cost. The work of [114] presents
a distributed approach at the tertiary level; the controller only requires local measurements and
information exchange with its neighbouring buses. This paper verifies that a distributed approach
can reduce the computational burden and avoid communication losses while the solution’s quality
is preserved compared to a centralised technique. A distributed PI approach based on the same
concept of the previous methods was proposed in [113] for isolated ac MGs. However, none of
these controllers considers equipment power limits, and only results for load changes are provided.
Moreover, these controllers do not consider the existence of hybrid ac/dc MGs.

2.7.2 Distributed economic dispatch for dc microgrids

The concept of incremental cost consensus (ICC) discussed for ac MGs has been extended for
dc MGs [18, 19, 99, 115, 116]. In this type of MG, the secondary controller restores the (aver-
age) voltage while ensuring economic dispatch of DGs by modifying the droop voltage control.
In [99] economic dispatch of generation and local voltage regulation is achieved. The authors of
[18] achieve economic dispatch of generation and average voltage regulation; however, separate
communication networks are used to communicate the power and voltage references. In a similar
fashion, [19] extends the ICC concept to multiple dc MGs. In [115] economic dispatch is per-
formed without considering voltage restoration. Nevertheless, none of these controllers considers
equipment power rating limits in their formulations. Only [116] considers power limits by using
saturators.

The authors of [117] use droop control to estimate the MG load and generation from uncontrol-
lable DGs. However, several parameters need to be adequately adjusted for this strategy to work.
The effects of constant communication delays in distributed schemes have also been analysed using

25
simulation work in [118]. Next, the works reported for H-MGs are presented.

2.7.3 Distributed economic dispatch for hybrid ac/dc microgrids

The economic dispatch of H-MGs has been usually solved through a centralised optimisation prob-
lem. These strategies can incorporate uncertainties in market prices [119] or uncertainties in gen-
eration and consumption [120]. The authors in [121, 122] include the power limits of the units in
the centralised optimisation problem. Nevertheless, these approaches are subject to a single point
of failure and do not provide reliability under communication failures.

Conversely, the authors of [123] proposed a two-level distributed controller for the economic
dispatch of H-MGs that depends on the estimation of the MG’s load. The incremental cost is in-
cluded in the droop controllers of the primary level while the secondary level restores the deviations
caused by the droop controllers. However, when these variables are restored the loading conditions
of ac and dc sub-MGs are hard to estimate. A relative loading index (RLI) is used to extract the
hidden loading status of each sub-MG and generate the power reference for the ILC. Nevertheless,
this method is highly dependent on the quality of the estimation of the RLI. The works of [117,124]
also include the economic dispatch within the droop controllers design but also present erroneous
behaviour when a secondary controller is included.

Better performance can be achieved by integrating the economic dispatch in distributed ap-
proaches at the secondary level [21, 22]. For instance, [21] proposes a unified controller for eco-
nomic dispatch and frequency and voltage restoration. First, ICC within sub-MGs is achieved
through information sharing between the DGs; the ILC then equalises the incremental costs of the
sub-MGs. However, the ILC’s communication scheme, stability, and parameter design are not pro-
vided. In [22], a similar approach is presented, where an event-based control scheme is utilised
to design the consensus protocol. This protocol reduces the communication burden of the MG.
However, these approaches consider that the H-MG is composed of three independent systems(ac
sub-MG, dc sub-MG, and ILCs) and neglect the dynamic interactions between them, which can
degrade the controllers’ dynamic response.

Most approaches in the literature assume fixed operational set-points for voltage and frequency
[16–19, 43], and do not take advantage of the flexibility associated with the secure operational
bands defined in the IEEE standard 1547-2018 [40], which suggests that DGs can operate normally
as long as the frequency and the voltage are within 1% and 5% of their nominal values, respectively.
It is worth noting that all the works mentioned above assume that frequency and voltages must be
set to fixed setpoint values, giving up flexibility in the microgrid control system. Additionally,
existing approaches are designed independently for either ac sub-MGs [16, 17, 44] or dc sub-MGs
[18,19], without accounting for the particular features of H-MGs. Also, as they are strategies based

26
on PI controllers, it is difficult to achieve multiple parallel objectives and cope with operational
constraints [45, 46].

To the best of the authors’ knowledge, the problem of combined economic dispatch of active
and reactive power as well as frequency and voltage restoration within secure bands in isolated
H-MGs using DMPC has not been thoroughly studied in the existing literature. Furthermore, in
most studies, equipment power constraints are not taken into account: these constraints can pre-
vent equipment damage. For these reasons, this thesis adopts a distributed cooperative scheme
for controllers based on MPC. Each controller seeks to dispatch its controllable unit based on its
generation cost and to restore the frequency and voltage amplitude deviations on the ac sub-MG
and the voltage magnitude on the dc sub-MG. Furthermore, as the proposed controller includes a
detailed model of the dynamics of the primary control level (controlled system), it is able to deal
with variable and uncertain delays [48, 86]. Another characteristic of the proposed DMPC is that
the variation of the control actions is penalised in the cost function. This penalisation prevents
abrupt changes in the MG operating point, and a smooth dynamic response is obtained.

2.8 Imbalance sharing in ac microgrids

As MGs operate at the distribution level, their power quality is affected by the presence of unbal-
ance between the phases [23, 24]. Unbalances directly affect the MGs’ operation if they are not
dealt with properly, as they can lead to inefficient operation and trigger protection systems. Un-
balances are mainly caused by the presence of single-phase loads. Moreover, the integration of
renewable energy resources is putting more stress on MGs, as these can be connected across all
three phases or only on one single phase [24].

The secondary control schemes reported in the literature for unbalanced MG operation usu-
ally include three features (according to their control objectives): (i) Compensation of imbalance
at particular points of the MG, (ii) Improving the sharing of unbalanced powers between the DG
units of the MG, and (iii) Simultaneous compensation of imbalance and improvement of unbal-
anced power-sharing. These aims can be addressed using decentralised, centralised and distributed
approaches.

The compensation of imbalance can be achieved using active power filters (APFs) to compensate
unbalanced currents or unbalanced voltages at specific points of the MG [25, 26]. However, APFs
are not attractive in MGs since they constitute additional hardware and higher costs. Another more
cost-effective solution is to embed imbalance compensation capabilities into the control schemes
of DG units that are already available in the MG [27–29]. For instance, in [27] a master/slave
based approach is proposed. A supervisory controller calculates the compensation effort (in terms
of current) for each slave converter to compensate for imbalances at sensitive load buses.

27
 Control schemes to improve the sharing of unbalanced powers between the DGs of MGs are
mainly based on droop control and use virtual impedance loops. This means that negative sequence
impedances are implemented to control the sharing of imbalance between the DGs. The magnitude
of these virtual impedances is controlled via decentralised control schemes in [30–32], meaning
that there is no coordination between the DG units (each DG works autonomously based on vari-
ables measured locally). However, better performance could be achieved via coordination between
DG units (centralised and distributed approaches). In this sense, the magnitude of the negative
sequence impedances are calculated in a coordinated way by a secondary centralised controller in
[33–36], while in [37–39] secondary distributed controllers, based on consensus algorithms, are
implemented.

It is worth noting that these papers (describing centralised and distributed systems) quantify
the DGs’ imbalance by defining three-phase unbalanced powers (calculated based on three-phase
power theories [38, 39]), aiming to improve the sharing of these powers. However, as shown in
[53] and [23], when unbalanced ac MGs are considered, the improvement in three-phase power
sharing does not ensure that the single-phase powers are appropriately shared. In this scenario,
overloading may occur in some of the DG phases, causing inappropriate behaviour in the DG and
load shedding, which could affect the overall security and reliability of the MG.

Another approach to improve the sharing of imbalance is the addition of power converters to
the three-phase MG to directly manage the DG phase power balance and prevent overloading of
single phases [50–52]. For instance, in [52], a multi-objective formulation is proposed to achieve
per-phase imbalance sharing in a three-phase MG. An additional power converter is placed at the
point of common coupling for managing single-phase power. However, the addition of extra power
converters into the MG increases the cost of this approach making the proposals reported in [50–52]
not cost-effective.

When looking for MPC-based methods to manage imbalance, only decentralised finite control
set model predictive control (FCS-MPC) has been reported at the MG’s primary control level [125–
127]. The authors of [125] propose a decentralised FCS-MPC, where the imbalance is managed
by an external loop that shares the negative sequence reactive power. The works of [126, 127]
present FCS-MPC methods to improve imbalance sharing. FCS-MPC methods may produce a
variable switching frequency because FCS-MPC does not use a modulator. They demand a high
computational burden, as they operate at the primary control level; this is because the evaluation of
the cost function is usually realised for all the switching states of the power converter. Furthermore,
these are usually decentralised methods that compute local solutions; thus, an optimal solution is
hard to obtain. To the authors’ best knowledge, no DMPC strategy for imbalance sharing at the
secondary control level has been proposed.

28
 To summarise, imbalance sharing methods based on virtual impedance loops [30, 33–39] do not
ensure a proper sharing of single-phase powers. This issue is discussed in more depth in Section
5.7.2. The solution proposed in [50–52] of adding power converters to the MG to achieve single-
phase power management is not cost-effective. Finally, FCS-MPC-based methods used for the
primary level for imbalance sharing require extensive processing capabilities, and since they are
based on a decentralised approach, do not provide a cooperative solution.

To avoid the drawbacks of these approaches, we propose a novel secondary DMPC control
scheme for improving the sharing of imbalances. This proposal avoids the use of virtual impedance
loops; moreover, it does not require additional converters in the MG. In this case, contrary to the
reported MPC-based methods for imbalance sharing, the proposed control algorithm achieves a
global solution via consensus objectives which do not require an extensive computational burden
(thanks to the distributed approach). The imbalance sharing is achieved by controlling the single-
phase reactive powers of the DGs.

In particular, the proposed strategy uses a modified single-phase Q − V droop scheme, where
one additional secondary control action is introduced per phase. With the multi objective DMPC,
it is possible to formulate a predictive controller that considers the dynamic behaviour of the MG
main variables. This technique uses a reduced number of control actions to achieve all the control
objectives (see Section 5.2). Furthermore, the proposed secondary control scheme can regulate
simultaneously the imbalance sharing and power quality of each DG. This has not previously been
explored extensively: the works published in [33, 39] are the only ones reported so far.

These simultaneous objectives are of paramount importance since, as discussed in [33, 34],
imbalance sharing methods increase the voltage imbalance at the output of the DGs. Therefore,
these imbalances should be regulated to avoid power quality issues as defined by IEEE standard
1547-2018 [40]. A well-used index to determine the power quality of a MG is the phase voltage
unbalance rate index (PVUR), which is defined Section 5.3.4. Thus, a control technique must
improve imbalance sharing without exceeding the maximum PVUR value recommended in the
IEEE standard 1547-2018 for PVUR [40].

2.9 Discussion
This section presented the hierarchical control structures of ac MGs, dc and H-MGs. The cor-
responding controllers at each control level and their functionalities were explained. The main
advantages of distributed control against centralised and decentralised control were also discussed.
A review of the current state-of-the-art in distributed control approaches applied to isolated ac, dc
and hybrid ac/dc MGs was presented. Special emphasis was given to consensus-based approaches
and distributed model predictive control approaches.

29
 The latest consensus-based strategies for economic dispatch were presented. Their operation
principles, main characteristics and limitations were detailed. Moreover, imbalance sharing meth-
ods reported so far were also discussed and analysed in detail.

The review carried out shows that despite the advances in control strategies, the economic dis-
patch of H-MGs and the sharing of imbalances have not been studied thoroughly. Moreover, there
are still research gaps that DMPC-based strategies can solve to improve the entire system’s reliabil-
ity and operation. On the other hand, DMPC presents several advantages against classical PI-based
approaches. This is because DMPC models the main dynamics of the controlled system and allows
the inclusion of hard and soft constraints, allowing to represent physical limitations and achieve a
more flexible regulation of certain variables.

The distributed structure of the predictive controllers facilitate the plug-and-play operation while
better performance against communication delays and communication failures is secured. Another
advantage of the distributed structure of the proposed strategies is their easy scalability. In addi-
tion, the proposed strategies are easy to deploy in current operating systems as they use the usual
measurements of the primary control level and do not need the addition of extra power convert-
ers. Instead, all the desired functionalities are incorporated directly into the formulations of the
predictive controllers. The following chapters present the proposed controllers.

30
Chapter 3

The proposed DMPC scheme for frequency

regulation and active power dispatch in ac
microgrids

3.1 Introduction

In this Chapter, the proposed distributed model predictive control (DMPC) strategy for economic
dispatch and frequency restoration for isolated ac microgrids (MGs) is detailed. In the DMPC
scheme of this chapter, each DG can economically dispatch active power by employing an incre-
mental cost consensus (ICC) approach and can also restore the MG’s frequency to its nominal value
whilst satisfying physical constraints.

The main challenge for this Chapter is to develop the local models and cost function required to
tackle simultaneously the aforementioned objectives. A local prediction model of the DGs primary
control level is included in each DMPC to predict its future behaviour. These are in the form of
droop control and active power transfer models. In addition, the DMPC strategy has to be able
to recognise and update its calculations when the MG is subject to external phenomena including,
communication delays, communication failures and the disconnection/reconnection of DGs.

This chapter is organised as follows: Section 3.2 presents the derivation of the incremental cost
from the usual centralised economic dispatch. Then, its extension for the distributed scheme is de-
signed. Section 3.3 introduces the proposed strategy. The dynamic prediction models are presented
in Section 3.4. The DMPC formulation is then described in Section 3.5. The experimental setup
and validation tests are explained in detail in Section 3.6. Finally, Section 3.7 summarises the main
benefits of the proposed strategy.

31
3.2 Centralised economic dispatch
Consider a three-phase balanced ac MG with a set of N DGs, where N = {1, ..., N}. The traditional
centralised approach to economically dispatching active power in a MG can be expressed as the
optimisation problem in (3.1).

N
minimise
P
∑Ci(Pi) (3.1a)
i=1

N
subject to ∑Pi=PD (3.1b)
i=1

Economic dispatch establishes the lowest cost dispatch of controllable DG units, whilst ensuring
that the total load is met. It is composed of a cost function that minimises a quadratic cost (3.1a)
subject to the power balance constraint (3.1b). N is the set of DGs in the MG, P = {Pi : i ∈ N}, Pi is
the active power contribution of DG i, PD is the total MG load, and Ci (Pi ) is a convex cost function
described in (3.4a). The traditional centralised approach to economic dispatch (3.1) requires a
unique central controller, and its failure could compromise the economic dispatch of the entire
MG . It presents a high computational burden [8, 10]. Therefore, a distributed control is a better
alternative for higher reliability and security for the operation of the MG [128].

Assuming strong duality holds e.g. Slater’s constraint qualification condition holds, the problem
may be expressed through its Lagrange dual [129]. The Lagrangian function of the economic
dispatch problem is:

L (Pi , η) = ∑Ni=1 Ci (Pi ) + η (PD − ∑Ni=1 Pi ) (3.2)

where the Lagrange multiplier η is associated with the power balance constraint. The Karush–Kuhn–Tucker
(KKT) stationary condition for the problem is defined in (3.3a) [16]. From (3.3a) it is possible to
establish that at the optimal point, the incremental cost (IC) function is defined by (3.3b).

∂L
=∇Ci(Pi )−η=0 i∈N (3.3a)
∂ Pi

η=∇Ci(Pi ) i ∈ N (3.3b)

The generation cost function for the i − th DG is stated in (3.4a) [59], where ai , bi and ci are

32
constant cost parameters defined in Section 3.6.2, and Pi is its active power contribution. Therefore,
by using (3.4a) in (3.3b), the IC for the i − th DG is given by (3.4b).

Ci(Pi )=ai Pi2+bi Pi+ci (3.4a)

ηi(Pi )=2ai Pi+bi (3.4b)

The economic dispatch problem redistributes the power contribution of all DGs to reach the
same optimal dispatch point η (3.3b) (KKT stationary condition), where η corresponds to the
(unique) dual variable associated with the demand-supply balance equation of the MG’s economic
dispatch problem (3.1b). Therefore, a distributed cooperative predictive scheme can be designed
that ensures the condition of ηi = η j = η in steady state, where η j is the IC of neighbouring DGs.
Note that this distributed scheme intrinsically meets the demand-supply balance.

Based on the IC, a new distributed predictive cooperative control strategy is designed, aiming
to provide an economic dispatch of active power and at the same time regulating the frequency.
This DMPC control strategy is based on the work presented in [49]. It is worth noting that in [49]
economic dispatch is not considered. Also, in this first proposal the DMPC formulation for the
coupling between voltage regulation and reactive power sharing is not considered. However, the
DAPI controller reported in [67] is used to solve for voltage restoration and reactive power sharing.

3.3 Proposed DMPC scheme

The following explanations and mathematical analysis are made for the i −th DG, as the analysis is
analogous for the rest of the DGs. The proposed controller neither depends on the MG’s electrical
topology nor on adjacent physical measurements i.e. only the typical measurements at the output
filter and voltage observers are required. Therefore, the number of buses and distribution lines are
irrelevant. At the DG’s output, an LCL filter is used, where the second inductance (Li ) is designed
to ensure a predominantly inductive impedance [63, 64], as shown in the electrical configuration of
Fig. 3.1. This allows the decoupling of active and reactive power and the implementation of the
classic droop controllers in the power converters [8].

To predict the power contribution of the i − th DG, the phase angle deviation (δ θi ) between the
local unit and the MG is determined [130]. For this, the voltage measurement (Vi ) at the output of
the LC filter is used. Then using a phase locked loop (PLL), its frequency (ωi ) and phase angle (θi )
are estimated. In addition, the voltage (V̂iB ) at the connection bar node (after the coupling induc-

33
Figure 3.1: Distributed economic dispatch with frequency restoration for the i − th DG

34
tances Li ) is estimated by virtual meters based on non-linear reduced-order observers to reduce the
hardware, while the performance of the controller is maintained. Then, the MG’s frequency (ω̂iB )
and phase angle (θ̂iB ) are estimated using a PLL, as shown at the bottom of Fig. 3.1.

The voltage observer is based on [131], and its main advantage is its linear dynamic for the
estimation error. Therefore, the observer gains can be tuned via pole placement, improving the
transient performance of the observer and the convergence rate of the estimation error. It also has a
low computational burden. The basis of the observer is detailed as follows. The observer works in
B = V sin(θ ) and V̂ B = V cos(θ ),
the α − β framework. The estimated states are defined by V̂α,i m i β ,i m i
where Vm depends on the abc − αβ transformation used. The observer’s state space formulation is
obtained from Kirchhoff’s voltage law, whereas its inputs are the measured values of Vi and ii (at
the LCL filter), both in the α − β framework. A detailed explanation of the development of the
voltage observer is presented in the Annexed B.

The controller scheme for the i − th DG is depicted in Fig. 3.1. It is configured as a voltage
source converter (VSC) with its respective LCL output-filter. Two control layers are highlighted.
The primary control level (cyan box in Fig. 3.1) is made up of ω − P and V − Q droop controllers,
outer voltage (slower) and inner current (faster) cascaded Proportional-Resonant (PR) controllers.
The voltage observer also operates at this control level.

At the secondary level (orange box in Fig. 3.1), the predictive controller for economic dis-
patch and frequency restoration is presented. This controller receives as inputs the local mea-
surements/estimates (Pi (k), Vi (k), ωi (k), θi (k), V̂iB (k), ω̂iB (k), θ̂iB (k)) of the i − th DG unit and the
results of the optimisation problems of communicated neighbouring units X p,i j . The controller has
two outputs, which are the frequency control action (vector ∆ωs,i ) and the results of the local op-
timisation problem X p,i (vector of predicted values), both defined in Section 3.5.3. Whereas the
former passes through a discrete integrator to ensure zero error in steady-state, the latter is sent via
the communication network. Note that both objectives of this proposal (frequency restoration and
economic dispatch) are achieved with the same control action ∆ωs,i . Next, the dynamic models that
rule the behaviour of the DMPC scheme are presented.

3.4 Dynamic models used for the design of the DMPC strategy

3.4.1 Communication network model

As distributed control schemes require information exchange, a full-duplex communication net-
work is considered. This network allows the consensus objectives to be achieved through coopera-
tion between the MG DGs [132]. To accomplish a consensus objective, all distributed controllers
must converge to the same steady-state value, also known as a consensus value. This communi-

35
cation structure considers both latency and connectivity phenomena. Latency represents the time
interval (τi j ) for a data packet to be transmitted from source to destination, whereas the connectivity
is represented by the N × N adjacency matrix A. The terms ai j (3.5) of a non-negative A represent
information flow among DGs at time instant k.




1 Data from DG j arrives at DGi at k

ai j (k) = 0 Data from DG j does not arrive at DGi at k (3.5)



0 j=i

where k = nTsec , n ∈ Z+ , and Tsec is the sample time of the controller.

As the communication is bidirectional, the associated graph is undirected. Thus, τi j = τ ji and

ai j =a ji [42, 132, 133]. A fully meshed communication network can be used as primary setup; how-
ever, this topology can change as long as at least one communication path exists between all DGs
in the MG (i.e., there is a spanning tree) [42, 132, 133]. An example of the adjacency matrix of a
MG with four DGs is depicted in Fig. 3.2. This matrix encodes the communication topology of the
MG.

Figure 3.2: Example of a MG with four DGs and its adjacency matrix

The adjacency matrix A is initialised at the beginning of the simulation based on communication
topology at that sample time. Then, this matrix is updated (by verifying the communication links)

36
at each sample time based on the information received on each DG from its directly communicating
neighbouring DG units. Note that each DG only knows (interacts with) its direct communication
links and not the entire network. Furthermore, an asynchronous communication protocol is used;
thus, no global clock is necessary to ensure that the sharing of information is globally synchronised
[134, 135].

3.4.2 Dynamic models

To model the dynamics of the MG, the frequency at each node, along with the active power flow,
are considered. Since these variables are coupled, in this work they are modelled using the droop
(3.6), phase angle difference (3.7) and power transfer (3.8) equations . In particular, the i − th DG
unit is modelled as the node at the output of its LCL filter, as shown at the bottom of Fig. 3.1. The
i − th DG is connected to the MG through the second inductance (Li ) of the LCL filter.

The droop model for frequency ωi (t) - active power Pi (t) for the local i − th DG is presented in
(3.6).

ωi (t) = ω0 + Mpω,i Pi (t) + ωs,i (t) (3.6)

where ω0 is the nominal frequency, Mpω,i is the droop slope that defines the linear relation
between the frequency and the active power Pi (t), and ωs,i (t) is the secondary control action (after
the discrete-time integrator, see Fig. 3.1) for the i-th unit. This relationship allows the DGs to
interpret active power changes in the MG by producing a frequency deviation; furthermore, through
this model, the primary and secondary control levels interact.

To determine the power contribution from the i − th DG to the MG, the phase angle difference
is determined. The phase angle difference (δ θi ) between the local DG unit and the MG through a
coupling inductance (Li ) of the LCL output filter is defined in (3.7) [130].

Z t
B
ωi (τ) − ω̂iB (τ) dτ

δ θi (t) = θi (t) − θ̂i (t) = (3.7)
0

where θi , ωi are the phase angle and frequency before the coupling inductance (Li ), and θ̂iB and ω̂iB
are the phase angle and frequency after the coupling inductance (Li ), respectively.

The power transfer equation (3.8) is included to govern the active power contribution from each
DG unit to the MG.

Pi (t) = BiVi (t)V̂iB (t) sin (δ θi (t)) (3.8)

37
 where Vi (t) and V̂iB (t) are the voltages before and after the coupling inductance Li , and Bi =
1/(Li · ω0 ) [130]. Note that this model only requires local measurements and the estimation of the
voltage (V̂iB ) after the coupling inductance to predict the active power contribution. Therefore, the
use of an admittance matrix is avoided with this formulation, ensuring the plug-and-play capability
of the DG units.

3.4.3 Discrete time models

Prior to defining the prediction models, a discretisation of the models in (3.6), (3.7) and (3.8) is
needed. These models are discretised using the forward Euler method, where k = nTsec , n ∈ Z+ ,
and Tsec is the sample time of the controller. As integrators are placed at the output of the predictive
controllers to ensure zero error in steady-state (See Fig. 3.1) [45], the incremental operator (3.9)
is applied to (3.6); thus, the optimisation problem is expressed as a function of the control action
variation (∆ωs,i ).

∆ f (k) = [ f (k) − f (k − 1)] (3.9)

Additionally, a Taylor expansion is applied around the measured/estimated point {ωi (k), ω̂iB (k),
Vi (k), V̂iB (k), δ θi (k), Pi (k)} to linearise the power transfer model (3.8). The resulting linear discrete
models are shown in (3.10). A detailed explanation of the discretisation and linearisation processes
is given in Annexed C. Note that having linear models ensures that the computational burden will
remain within the recommendations for the secondary control level.

ωi (k+1) = ωi (k)+Mpω,i[Pi (k+1)−Pi (k)]+∆ωs,i (k) (3.10a)

δ θi (k+1) = δ θi (k)+Tsec ωi (k+1)−ω̂iB (k)

 
(3.10b)

Pi (k+1) = Pi (k)+[δ θi (k+1)−δ θi (k)]BiVi (k)V̂iB (k)cos(δ θi (k)) (3.10c)

The models described in this section are included in the in DMPC strategy in the following
section.

3.5 Formulation of the distributed model predictive control

Unlike [43, 49, 67], where the aim was that all generation units contribute to the power-sharing of
active power proportionally to their maximum power rating, in this work, a more cost-effective
formulation is proposed that meets the economic dispatch, where the most economical generation

38
units are the ones that contribute most to the power-sharing as long as power limits are not exceeded.
Additionally, this proposed controller restores the MG frequency to its nominal value.

DMPC uses the discrete-time model of the system, presented in (3.10), to predict the behaviour
of the DGs over a prediction horizon (Ny ), and a sequence of control actions (Nu ) is calculated
through a numerical optimisation problem that minimises a cost function (see (3.11) in Section
3.5.1). The predicted sequences for the variables and control actions are contained in the vector Xi
(defined in Section 3.5.3), which is the solution to the optimisation problem. Only the first control
action is applied to the system, and the optimal control problem is repeated at each sample time
with updated measures[45]. A challenge in the implementation of MPC strategies at the secondary
control level is the definition of an optimisation problem with a low computational burden which
can be solved in a short sample period [46, 49]. The optimisation problem and how it is solved are
detailed in the next section.

3.5.1 Cost function

The multiobjective cost function is stated in (3.11) and is composed of three weighted terms, where
each term seeks a specific objective.

Ny Nu
Ji (k)=∑λ1i (ω i (k+m)−ω0 )2+∑λ2i (∆ωs,i (k+m−1))2
m=1 m=1
N Ny
(3.11)

2
+ ∑ ∑λ3iai j (k) ηi(k+m)−η j (k+m−τ̂i j )
j=1, j̸=i m=1

The first term represents the restoration of the average frequency (ω i ) to its nominal value,
which is calculated only with the information communicated from other DGs. The second term
penalises the control action sequence required to carry out (at the same) time both the regulation
and consensus objectives; note also, the overshoot and settling time are adjusted with this term.
The third term achieves the economic dispatch through a cooperative consensus over the predicted
ICs between the local DG and the neighbouring (communicating) DGs. Therefore, the condition
ηi = η j = η in steady state is enforced within the cost function. The terms λ1i , λ2i and λ3i are the
tuning parameters explained in Section 3.6, and τ̂i j is the estimation of the communication delay,
which is assumed to be one sample period at the secondary level.

3.5.2 Predictive models and constraints

The set of equations of (3.10) is generalised in (3.12) to predict the response of the i − th DG at
k + m steps ahead, where m ∈ Z+ . Note that although the coefficients produced in the linearisation
are updated each sample time, they are constant during the optimisation.

39
 ωi (k+m)=ωi (k+m−1)+Mpω,i[Pi (k+m)−Pi (k+m−1)]+∆ωs,i (k+m−1) (3.12a)

δ θi (k+m)=δ θi (k+m−1)+Tsec ωi (k+m)−ω̂iB (k)

 
(3.12b)

Pi (k+m)=Pi (k)+[δ θi (k+m)−δ θi (k)]BiVi (k)V̂iB (k)cos(δ θi (k)) (3.12c)

In addition to the previous models, a set of operational constraints are included in the DMPC
formulation. These are equality constraints to ensure appropriate performance of the controller and
inequality constraints to guarantee the solution is within the equipment power ratings of each DG.

Equation (3.13a) represents a local average frequency approximation (ω i ). This average is

calculated with only the information communicated from the other DGs (ω j ), which is determined
by the adjacency term ai j . Also, the term τ̂i j is the estimate of the time-delay. These two terms
provide robustness against communication failures and latency effects, respectively. The terminal
constraint (3.13b) is included to guarantee convergence of the distributed predictive scheme to the
tracked value (nominal frequency) at the end of the prediction horizon Ny [47].

N
ωi (k+m)+∑ai j (k)ω j (k+m−τ̂i j )
j=1
ω i (k+m)= N
(3.13a)
1+∑ai j (k)
j=1

ω i (k+Ny )=ω0 (3.13b)

The generalisation of the incremental cost (IC) model for the prediction horizon is expressed in
(3.14). This term is used in the cost function to achieve the economic dispatch of DGs, as will be
shown in (3.11).

ηi (k + m) = 2ai Pi (k + m) + bi (3.14)

Finally, the active power contribution of the i −th DG is limited to its physical maximum power
rating through the inequality constraint (3.15); hence, the space solution is bound.

Pi,min (k) ≤ Pi (k + m) ≤ Pi,max (k) (3.15)

3.5.3 Formulation of the quadratic programming

The proposed DMPC comprises a quadratic cost function, linear equality constraints and linear
inequality constraints; thus, it is convex and can be synthesised in a canonical quadratic program-

40
ming (QP) formulation. Due to the convexity of the proposed controller, the global minimum can
be achieved. The cost function (3.11) with its respective equality and inequality constraints (3.12)-
(3.15) are included in a QP formulation (3.16) via the matrices/vectors Hi , Ai , Bi , Aeq,i , Beq,i , Fi .

1
minimise Ji (k) := XTi Hi Xi + FiT Xi
Xi 2
subject to Ai Xi ≤ Bi (3.16)

Aeq,i Xi = Beq,i

The output of the QP problem is (3.17), where the set of predicted variables is contained in X p,i
(3.18) and the optimal control sequence is given by X∆,i (3.19). Note that thanks to the distributed
structure of the controller, the number of predicted variables is fixed. Therefore, the computational
burden does not increase when new DGs are introduced into the MG. This is of high importance
for the scalability of control techniques at the secondary control level.

Xi ={X p,i , X∆,i } (3.17)

X p,i ={ω i (k + m), ωi (k + m), δ θi (k + m),

N
(3.18)
y
Pi (k + m), ηi (k + m)}m=1

X∆,i = {∆ωs,i (k + m − 1)}Nu

m=1 (3.19)

Following the principle of MPC, only the first predicted control action of X∆,i is applied in
the frequency-active power droop controller, i.e. ∆ωs,i (k). The optimisation is then repeated each
sample time with updated measurements/estimations. Moreover, the predictions (X p,i ) are shared
to achieve the cooperative objectives.

To compute the QP problem (3.16), the QPKWIK algorithm is used, which is a stable variation
of the classic active-set method [136]. This solver is able to generate C++ code to run on the
experimental setup. The methodology to solve the DMPC scheme is described in Algorithm 1,
which details all the necessary steps to achieve the cooperative objectives.

41
Algorithm 1 DMPC solution for DGi
Inputs: Measurements and estimations:{ωi (k), ω̂iB (k),Vi (k), V̂iB (k),
δ θi (k), Pi (k)}
Received information:Xi j , ∀ j = {1, ..., p}
Outputs: Xi , ∆ωs,i (k)
Initialisation :
1: Compute matrix/vector coefficients of Hi , Ai , Bi , Aeq,i , Beq,i , Fi
2: for every k do
3: Compute adjacency terms ai j according to the received information.
4: According to the received information, compute the sum of frequency and incremental cost from (3.13a) and (3.11).
5: Update matrices/vector Hi , Ai , Bi , Aeq,i , Beq,i , Fi from (3.16) according to the results of step 4 and the measurements/estimations
{ωi (k), ω̂iB (k),Vi (k), V̂iB (k), δ θi (k), Pi (k)}.
6: Solve QP problem using QPKWIK algorithm.
7: if Xi is feasible and t < k + Tsec then
8: Extract ∆ωs,i (k) from Xi .
9: else
10: ∆ωs,i (k) = 0.
11: end if
12: Update controller outputs and send Xi to neighbour DGs if it is feasible
13: end for

3.6 Experimental results

3.6.1 Experimental MG configuration

The performance of the proposed DMPC control strategy was assessed in a case study using the
experimental MG configuration illustrated in Fig. 3.3a. This experimental MG was built in the
MGs Control Lab at the University of Chile1 . The DGs are emulated by real-time controlled power
units from the Triphase® company, part of National Instruments. The MG comprises three ac DG
units, which are emulated by the PM15F120 (DG1 and DG2 ) and PM5F60 (DG3 ) Triphase®, as
shown in Fig. 3.3b. Each DG unit is controlled by a real-time target (RTT) computer, in which
the DMPC control algorithm is uploaded. A detailed explanation of the experimental setup is
provided in Annexed F. Table 3.1 presents the MG electrical parameters. Furthermore, Table 3.2
and Table 3.3 present the DG parameters, including their operating costs which were obtained from
[16]. In Table 3.3, ai , bi and ci are the constant generating cost parameters used in the quadratic
cost function (3.4b). These parameters represent the generating cost of the DGs and their operating
point efficiency.

1 https://www.die.cl/sitio/home/investigacion/laboratorios/laboratorio-de-control-de-micro-redes/

42
 a)
Text

DG1

Direct
Communication
Optical
Communication

DG2

DG3

b)

PM15F120:
DG1 DG2
RTT1
Z1

RTT2
Z2+Z3

RTT3

PM5F60:
DG3
RTT: Real-time target

Figure 3.3: a) MG Diagram, b) Experimental MG setup

43
 Table 3.1: MG electrical parameters
Parameter Description Value

Tprim [s] Primary level sample period 1/16E3

Z1 [Ω] Load 1 35 (0.64 KW)

Z2 [Ω] Load 2 15 (1.5 KW)

Z3 [Ω] Load 3 22 (1.02 KW)

Li [mH] Coupling inductance 2.5

Li j [mH] Distribution line inductance 2.5

ω0 [rad/s] Nominal frequency 100π

V0 [V] Nominal voltage (peak) 150

ωc [rad/s] Droop controller cutoff frequency 2π

Table 3.2: Power ratings and droop slopes

Parameter Description DG1 -DG3

Pmax [Kw] Maximum power rating 2.1

rad
M pω sW P − ω droop coefficients -2.38E-4
V
Mqv VAR Q −V droop coefficients -4.8E-3

Table 3.3: DG operating costs

Parameter DG1 DG2 DG3

a[$/kW 2 ] 0.264 0.4 0.5

b[$/kW ] 0.067 0.1 0.125

c[$] 0 0 0

44
3.6.2 Design parameters and test scenarios used to evaluate the DMPC

Table 3.4 presents the DMPC design parameters and weighting factors. All the design parameters
were selected aiming to reduce the overall computational burden. This is because the computational
burden increases significatively with the sample time, and the prediction and control horizons [47].
While the sample time was selected considering the open loop rise time (Tr = 0.7s) of the MG’s
frequency (operating with only the primary control enabled) as Tsec = 0.7/14 = 0.05s [137], the
prediction and control horizons were selected as 5 samples because with these values, the traf-
fic over the communication network is reduced. The weighting factors were tuned following the
guidelines in [88], i.e., looking for a trade-off between the control objectives and, if needed giving
more importance of one objective compared to the rest of the objectives. The minimum estimated
delay corresponds to one sample period at the secondary level sample time.

Table 3.4: Controller parameters and weights

Parameter Description Value

Tsec [s] Controller sample time 0.05

τ̂i j [s] Estimated communication delay 0.05

Ny Prediction horizon 5

Nu Control horizon 5
s 2
λ1i [( rad ) ] Average frequency error 18.5E2
s 2
λ2i [( rad ) ] Frequency control action 4.7E4

λ3i [( W1 )2 ] Active power dispatch 1.01E-4

The controller was tested under four scenarios using the experimental MG, shown in Fig. 3.3.
The first scenario presents the DMPC’s performance when the MG experiences load changes. The
second scenario shows the behaviour of the MG when there are latency effects over the commu-
nication network. The third scenario shows the effects of a failure in the communication network.
Finally, the last scenario validates the Plug-and-Play capability, where DG2 is disconnected and
reconnected from/to the MG. These four test scenarios were selected because these are the most
common phenomena that a controller at the secondary level confronts [49, 134]. A distributed
controller has to perform well against communication issues, such as communication delays and
failures. In addition, the possibility of disconnecting and reconnecting DGs to the MG without
changes in the programming of the controllers is essential. Moreover, as voltage restoration and
reactive power sharing are not the focus of this first proposal, their results are presented just for the
base case.

45
3.6.3 Scenario I (base case) - Load changes
This scenario presents the performance of the proposed controller in the MG when there are several
load impacts. During the whole test, the adjacency matrix is not changed and is represented by
(3.20).

   
a11 a12 a13 0 1 1
A(k) =  a21 a22 a23  =  1 0 1  (3.20)
   

a31 a32 a33 1 1 0

The test starts with the primary control enabled (internal loops and droop control) and two loads
connected (Z1 and Z2 ) at different nodes to observe that without the DMPC controller, the DG units
share active power equally (see Fig. 3.4a before 10s) and the frequency deviates from its nominal
value (see Fig. 3.4c before 10s) with an operating cost of 0.8293 $/h (American dollars/hour),
shown in Fig. 3.4d (before 10 s).

DG1 DG2 DG3 DG1 DG2 DG3

F1 F2 F3 Cost

Figure 3.4: Load changes - base case: a) Active power contribution, b) Incremental cost consensus,
c) Frequency regulation, d) Total operation cost

At t = 10s, the secondary controller is enabled, so the active power is redistributed, as observed
in Fig. 3.4a, and the frequency is restored to its nominal value, as observed in Fig. 3.4c. Once
the controller is enabled the active power is redispatched according to the DG’s operating costs in
Table 3.3 via the consensus over the incremental cost (η), as shown in Fig. 3.4b. The operating
cost is reduced from 0.8293 $/h to 0.7696 $/h. This cost reduction may seem small in monetary
terms due to the small size of the MG; however, it could be significant in percentage terms in a
larger MG. Indeed the reduction cost is 7.2%. As DG1 is the least expensive unit, it takes the

46
majority of the load followed by DG2 , which has an intermediate cost. Furthermore, DG3 takes
the lowest load as it is the most expensive unit to operate. At 30 seconds, the MG is subjected to
its total load (i.e. Z3 is connected). Finally, the loads Z3 and Z2 are disconnected at t = 50 and
t = 70, respectively. Fig. 3.4a, Fig. 3.4b and Fig. 3.4c show that during all the load perturbations
the controller provides a smooth response without large overshoots and with a settling time below
3 seconds for both objectives. Moreover, the proposed DMPC reduces the operating cost (as shown
in Fig. 3.4d) during the entire test.

In addition, it is shown that the proposed DMPC controller does not affect the performance of
the DAPI controller for normalised reactive power sharing and voltage restoration, as shown in
Fig. 3.5a and Fig. 3.5b, respectively. It is worth noting that there is better regulation of voltage than
reactive power sharing. This is because these two objectives are opposed in the DAPI controller
[67], and more weight was given to regulation of voltage.

DG1 DG2 DG3 V1 V2 V3

Figure 3.5: Load changes - base case: a) Reactive power contribution, b) Voltage regulation

47
3.6.4 Scenario II - Communication delay
This test presents the performance of the proposed controller when there is a constant delay (τi j )
in all links of the communication network, whilst the estimated delay (τ̂i j ) is kept constant at one
sample, as shown in (3.13a) and (3.11). For each test, the same load perturbations of scenario I
are applied. Two cases were considered: a) small time-delay (τi j =0.25s) and b) large time-delay
(τi j = 1s). Note that the worst-case scenario represents a delay of 20 samples, which is four times
the prediction horizon Ny .

The results of the controller performance are presented in Fig. 3.6. The frequency restoration is
the most affected variable when communication delays are present. Fig. 3.6a and Fig. 3.6b show the
frequency of each DG for the cases of small time-delay (τi j =0.25s), and large time-delay (τi j =1s),
respectively. It is observed that the larger the time-delay the larger the overshoot and settling time.
However, these two parameters are still small, even for the worse case-scenario, i.e. for (τi j =1s) the
overshoot is negligible (less than 0.3%) and the settling time is below ten seconds.

F1 F2 F3 F1 F2 F3

DG1 DG2 DG3 DG1 DG2 DG3

DG1 DG2 DG3 DG1 DG2 DG3

Figure 3.6: Communication delays: a) Frequency regulation for τi j = 0.25s, b) Frequency regu-
lation for τi j = 1s, c) Active power contribution for τi j = 0.25s, d) Active power contribution for
τi j = 1s, e) Incremental cost consensus for τi j = 0.25s, f) Incremental cost consensus for τi j = 1s

The active power dispatch for the same two time-delays is presented in Fig. 3.6c Fig. 3.6d, while

48
the IC consensus for both time-delays is shown Fig. 3.6e Fig. 3.6f. For both variables, it is observed
that the settling time is practically unaffected; even in the worst case (τi j =1s), the settling time is
lower than five seconds. The overshoot slightly increases, as the time-delay increases. This is seen
most when the controller is activated; nevertheless, it is still negligible. From these results, it is
possible to establish that the DMPC is robust against communication delays over and above the
prediction horizon. This is because the DMPC uses the rolling horizon property, which determines
the appropriate control sequence even with past information from neighbouring DGs [46].

3.6.5 Scenario III - Communication link failure

To analyse the performance of the controller against communication link failures, the following
test was carried out. The test begins with two loads connected at different nodes (Z1 and Z2 ). At
t = 10s, the controller is enabled. A communication failure is forced at t = 30s between DG1 and
DG2 , so the adjacency matrix is modified as shown in (3.21), and the control algorithm identifies
automatically the failure by calculating (3.13a) and (3.11) only with the information received. Z3
is connected at t = 50s and disconnected at t = 70s. Finally, the communication link is restored at
t = 90s.

   
a11 a12 a13 0 0 1
A(k) =  a21 a22 a23  =  0 0 1  (3.21)
   

a31 a32 a33 1 1 0

The results are shown in Fig. 3.7. It is observed that the controller performance is not impaired,
and the control objectives are achieved. Therefore, the DMPC strategy is robust against commu-
nication failures. Nevertheless, the transient response is different, specifically, the settling time is
increased to nearly ten seconds. This is because the communication matrix (A) is not complete
(3.21), and the control objectives are directly related to known information from the neighbour-
ing DGs. A detailed explanation of the effects of the communication network on the controller
performance is presented in Section 5.5.3 and Section 5.7.1.

49
 DG1 DG2 DG3 DG1 DG2 DG3

F1 F2 F3 Cost

Figure 3.7: Communication failure: a) Active power contribution, b) Incremental cost consensus,
c) Frequency regulation, d) Total operation cost

3.6.6 Scenario IV - Plug-and-play

This test presents the controller’s response when an unscheduled failure occurs in a specific DG.
The test starts with two loads connected at t = 0s (Z1 and Z2 ) and the controller is enabled and at
t = 10s, where the adjacency matrix is represented by (3.20). At t = 30s DG2 is taken out of service,
i.e. DG2 is disconnected from both the electrical system and the communication network. Thus,
the adjacency matrix is modified as shown in Fig. 3.8c at t = 30s. The MG continues operating
with DG1 and DG3 connected. Next at t = 50s, the total load is connected. At t = 70s, after a
synchronisation routine, DG2 is reconnected to the MG. Finally, Z3 is disconnected at t = 90s.

Note that although DG2 is disconnected from the MG, it is not turned off. Only its secondary
control is disabled, but its primary control continues operating. When DG2 is disconnected or
reconnected the adjacency matrix is updated, and the remaining controllers identify this failure by
calculating (3.13a) and (3.11) only with the information received. Therefore, the remaining DG
units optimise the consensus terms with only the operating units.

Fig. 3.8a presents the active power contribution. It is observed that when DG2 is disconnected
although both DG units increase their contributions, DG1 takes the majority of the load; however,
its maximum power rating is not exceeded. Fig. 3.8b shows that the IC consensus is achieved
during all the disturbances. However, when DG2 is disconnected the MG operating cost increases
(see Fig. 3.8b at t = 30s). This is because DG3 , which is the most expensive, increases it power
contribution. Similarly, Fig. 3.8b presents the frequency restoration. It is observed that the operat-
ing DGs restore the frequency adequately during the whole test without overshoots or long settling

50
 DG1 DG2 DG3 DG1 DG2 DG3

F1 F2 F3 Cost

Figure 3.8: Plug-and-Play test: a) Active power contribution, b) Incremental cost consensus, c)
Frequency regulation, d) Total Operation cost

times. Therefore, the proposed DMPC scheme has a Plug-and-Play capability and always respects
the physical power rating of the DGs.

3.7 Discussion
This Chapter presented a novel distributed predictive control strategy to cope with economic dis-
patch and frequency regulation for isolated ac microgrids. The proposed controller is able to main-
tain frequency regulation and economic dispatch simultaneously, respecting the maximum power
limits of each DG unit. The proposed strategy achieves the economic dispatch of DGs in a dis-
tributed fashion through the incremental cost consensus concept. Moreover, the proposed strategy
considers only the communication links between neighbouring communicated DGs.

The dynamic performance of the controller is evaluated experimentally and discussed under
four test scenarios. The controller effectively tackles the disconnection and reconnection of DGs
and communication link issues, i.e., communication delays and communication failures. It is veri-
fied that the rolling horizon property of the DMPC scheme compensates for large communication
delays. Moreover, the distributed structure of the controller provides robustness against communi-
cation link failures.

This proposed strategy and its experimental validation were presented in the journal paper:

A. Navas-Fonseca, J. S. Gomez, J. Llanos, E. Rute, D. Saez, and M. Sumner, “Distributed

51
Predictive Control Strategy for Frequency Restoration of Microgrids Considering Optimal Dis-
patch," in IEEE Transactions on Smart Grid, vol. 12, no. 4, pp. 2748-2759, July 2021, doi:
10.1109/TSG.2021.3053092, [Q1-IF 8.960 - published ].

The following chapter formulates the application of the DMPC strategy for the operation of
hybrid ac/dc microgrids and the inclusion of voltage and reactive power control in the proposed
predictive control strategy for ac DGs. For the operation of hybrid ac/dc MGs (H-MGs), predictive
models and cost functions for the control of dc DGs and interlinking converters (ILCs) need to be
developed. Additionally, new models and additional objectives are included in the formulation for
ac DGs to control reactive power and voltage restoration. A detailed explanation of the DMPC
scheme for the operation of H-MGs is provided in the following chapter.

52
Chapter 4

The proposed DMPC scheme for frequency

and voltage regulation within bands and the
economic dispatch of active and reactive
power for hybrid ac/dc microgrids

4.1 Introduction

In this Chapter, the distributed model predictive control (DMPC) strategy is proposed for the eco-
nomic dispatch of active and reactive power and frequency and voltage regulation within bands for
isolated hybrid ac/dc microgrids (H-MGs). In this DMPC scheme, all the DGs of the H-MG can
achieve active power economic dispatch while the DGs of the ac sub-MG can also provide reactive
power economic dispatch. In contrast to the previous approach, the frequency and ac voltage on
the ac sub-MG and the dc voltage on the dc sub-MG are regulated within secure bands through the
use of soft constraints. Additionally, equipment operating constraints are not exceeded.

The main challenge for this chapter is to design the local dynamic models and cost functions of
ac generators, dc generators and interlinking converters (ILCs) to tackle these objectives simulta-
neously. A local prediction model is included in each DMPC to predict its future behaviour, i.e.,
droop control, and active and reactive power transfer models. Moreover, the DMPC strategy has to
be able to recognise and update its calculations when the MG is subject to external phenomena such
as, communication delays, communication failures and the disconnection/reconnection of DGs and
ILCs.

This chapter is organised as follows: Section 4.2 presents the derivation of the incremental

53
cost from the usual centralised economic dispatch for H-MGs. In a similar fashion, Section 4.3
shows the derivation of the reactive marginal cost (RMC). Section 4.4 describes the communication
structure of the H-MG. Section 4.5 introduces the proposed strategy for ILCs. Then, the DMPC
formulation for ILCs is described in Section 4.6. The proposed strategy for dc DGs is presented in
Section 4.7, and its DMPC formulation is detailed in Section 4.8. The proposed strategy for ac DGs
is presented in Section 4.9, and its DMPC formulation is described in Section 4.10. The simulation
setup and validation tests are explained in detail in Section 4.11. Finally, Section 4.12 summarises
the main benefits of the proposed strategy.

4.2 The active power economic dispatch problem

The usual centralised approach to economically dispatching active power in a H-MG can be ex-
pressed as the optimisation problem in (4.1) [21]. The objective function minimises a quadratic
cost function subject to the power balance constraint with P = {Pi : i ∈ Nac ∪ Ndc }, where Nac
and Ndc are the sets of ac and dc DGs, respectively. Pi is the active power contribution of DG i.
The aggregated active power load on the ac and dc sub-MGs are PDac and PDdc , respectively. The
quadratic cost function for DGi is expressed in (4.2), where ai , bi and ci are the cost coefficients of
DGi , defined in Section 4.11.

minimise ∑ Ci (Pi ) (4.1a)

P i∈Nac ∪Ndc

subject to ∑ Pi=PDac+PDdc (4.1b)

i∈Nac ∪Ndc

Ci (Pi ) = ai Pi2 + bi Pi + ci (4.2)

However, as demonstrated in [17–19, 44], this optimisation problem when operated in a cen-
tralised framework is susceptible to a single-point-of-failure and has a high computational burden
if it is solved using a centralised approach. For these reasons, a distributed controller is a more
reliable and secure solution. The optimisation problem of (4.1) can be expressed in a distributed
way through the incremental cost criterion [16, 17, 20]. Following the same procedure explained
previously for ac MGs in Section 3.2, the Lagrangian function of (4.1) can be expressed as follows.

54
 L (Pi , η) = ∑ Ci (Pi )
i∈Nac ∪Ndc
! (4.3)
ac dc
+ η PD + PD − ∑ Pi
i∈Nac ∪Ndc

∂ L (Pi , η) ∂ Ci (Pi )
=0 ⇐⇒ η= , i ∈ Nac ∪ Ndc (4.4)
∂ Pi ∂ Pi

From the stationary condition (4.4), the incremental cost (IC) or Lagrange operator η can be
obtained. To accomplish the economic dispatch of active power in the H-MG, all DGs must achieve
the same IC, i.e., η = ηi = η j where i, j ∈ Nac ∪ Ndc .

4.3 The reactive power economic dispatch problem

As the total generation costs of ac DGs are associated with both the active and reactive power
supplied, it is necessary to co-optimise the production of reactive power in the ac sub-MG [138].
For this purpose, similar to the IC principle described in Section 4.2, the reactive marginal cost (Ψi )
in (4.5) is proposed to minimise the production of reactive power in a distributed fashion [111,112].

Ψi = 2a′i Si,res Qi + b′i i ∈ Nac (4.5)

where Qi is the reactive power of ac DG i and Si,res is the residual apparent power capacity
of the i − th ac DG as a function of its rated capacity, i.e., Si,res = (Si,max − Si )/Si,max . The cost
coefficients a′i = ai sin2 (φ ) and b′i = bi sin(φ ) depend on the active power cost parameters ( defined
in Section 4.11) and the power factor angle φ . To accomplish the economic dispatch of reactive
power in the ac sub-MG, all ac DGs must achieve the same reactive marginal cost (RMC), i.e.,
Ψ = Ψi = Ψ j where i, j ∈ Nac . Based on the principles of incremental cost (IC) and reactive
marginal cost (RMC), a novel cooperative DMPC strategy is proposed. This strategy minimises
the operational costs of both active and reactive power while the frequency and voltage variables
on the H-MG are regulated within predefined bands. The development of the distributed predictive
strategies for ILCs, dc DGs and ac DGs is explained in detail in the following sections.

4.4 Communication structure

Although ac DGs, dc DGs and ILCs have different operating principles, they can interact with each
other [43]. In this way, information can be shared globally to achieve cooperative control objectives
via a fully-meshed communication network. Consider a H-MG composed of a set of N nodes (DGs

55
or ILCs), where N = Nac ∪ Ndc ∪ NILC . The subsets of ac DGs, dc DGs, and ILCs are represented for
Nac = {1, ..., Nac }, Ndc = {1, ..., Ndc }, and NILC = {1, ..., NILC }, respectively. Each node (DG or ILC)
includes a model of the full-duplex communication network. This model considers both latency
and connectivity issues. Latency (represented in sampling periods) is characterised as the end to
end communication delay (τi j ), i.e., total time for a data packet to be transmitted from source to
destination. Connectivity represents the information flow among nodes at time instant k and is
stated by the N × N non-negative adjacency matrix A (defined in Section 4.11). The entries ai j of
the adjacency matrix A are 1 if there is communication between node j and node i at time instant k
or 0 otherwise, where k = nTsec , n ∈ Z+ , and Tsec is the DMPC sample time.

As a full-duplex communication network is used, the associated communication graph is undi-

rected. Thus, τi j = τ ji and ai j =a ji [42]. We consider that the undirected graph in this work is
connected, which implies that there must be at least one communication path between any two
nodes (i.e., there is a spanning tree). Therefore, the H-MG’s topology can vary as long as there is
at least one communication path between all its nodes.

4.5 Proposed DMPC scheme for interlinking converters (ILCs)

As stated in the introduction section, ILCs transfer active power bidirectionally between the ac and
dc sub-MGs. In this way, to achieve the economic dispatch of active power in a H-MG, the ILC
should equalise the incremental costs (ICs) of both sub-MGs, i.e., the condition ηiac = η dc j =η
must hold, where i ∈ Nac ∧ j ∈ Ndc and η is the optimal IC value. The control diagram of the ILCh
with h ∈ NILC is shown in Fig. 4.1.

Two control levels are distinguished. As a back-to-back configuration is used for each ILC,
the primary level comprises a current controller on both ac and dc sides. The proposed DMPC is
presented at the secondary level (orange box in Fig. 4.1). The DMPC receives as inputs the local
active power measurement (PhILC ) and the active power predictions of communicated ac DGs, dc
DGs, and ILCs. The controller has two outputs, which are the active power variation (vector ∆PhILC )
and the results of the local optimisation problem XILC
p,h (vector of predicted values), both defined in
Section 4.6.1. The former passes through a discrete integrator to ensure zero error in steady-state,
Whereas the latter is sent via the communication network. Furthermore, the ILCs can be connected
to any electrical node in the H-MG as they share information with the DGs of both sub-MGs.

56
Figure 4.1: Control diagram of DMPCi for ILCs.

57
4.5.1 Dynamic models used for the design of the controller for interlinking
converters
ILC ) to ILC from the
Considering that the losses in ILCh are negligible, the power contribution (Ph,i h
ac ILC
i − th ac DG (Pi ) can be computed as (4.6a). Similarly, the power contribution (Ph, j ) to ILCh
from the j − th dc DG (Pjdc ) can be calculated as (4.6b).

Piac (k) = Piac (k − 1) + Ph,i

ILC
(k − 1) ∀ i ∈ Nac (4.6a)

Pjdc (k) = Pjdc (k − 1) − Ph,ILC

j (k − 1) ∀ j ∈ Ndc (4.6b)

These models are the basis for determining the power transference for each ILC that comprises
the H-MG. Note that (4.6a) considers that the ac sub-MG is receiving active power form the dc
sub-MG through the ILC, whilst (4.6b) considers that the dc sub-MG is providing power to the ac
sub-MG through the ILC.

4.6 Formulation of the DMPC for ILCs

The following explanations and mathematical analysis is performed for the h − th ILC, as the
analysis is analogous for the rest of the ILCs. The optimisation problem with its cost function and
constraints are described in Section 4.6.1 and Section 4.6.2, respectively. The predicted variables
and control actions sequence of this optimisation problem are contained in the vector XILC
h (defined
in Section 4.6.2), which is the solution to the optimisation problem (4.7). Only the first control
action is applied to the system, and the optimal control problem is repeated at each sample time
with updated measures[45]. The optimisation problem and how it is solved is detailed in the next
section.

4.6.1 Cost function

The cost function of the DMPC scheme for ILCh is composed of three terms and is described in
(4.7) where Ny and Nu are the prediction and the control horizons, respectively.

58
 Nu
JhILC (k) = ∑ λ1h (∆PhILC (k + m − 1))2
k=1
Ny  2
ac dc
+∑ ∑ a (k)a
∑ 2h hi h j
λ (k) η i (k + m − τ̂hi ) − η j (k + m − τ̂hj ) (4.7)
i∈Nacj∈Ndc k=1
Ny  2
PhILC (k + m) PILC (k + m − τ̂hl )
+∑ ∑ λ3hahl (k) ILC − l
Ph max PlILC
max
l∈NILC k=1

The first term penalises the variation of the control action sequence to minimise the control
effort and improve the controller’s transient behaviour. The second term equalises the ICs of the
sub-MGs by transferring power from the the most economical side to the most expensive side.
When all DGs achieve a consensus on the IC, the economic dispatch equilibrium is reached. This
objective is strengthened in each DG controller, as will be explained in Section 4.7 and Section
4.9. The third term ensures that when there are multiple ILCs, the power transferred per ILC is
proportional to its maximum power rating with l ∈ NILC , thus, avoiding overloading the ILCs. The
terms λ1h to λ3h are positive tuning parameters explained in Section 4.11. Note that the consensus
objectives (economic dispatch of DGs and power sharing of ILCs) are updated only with the pre-
dicted information from neighbouring agents (connected), represented by the terms ahi (k), ah j (k)
and ahl (k), and the estimated delays τ̂hi , τ̂h j and τ̂hl .

4.6.2 Prediction models and constraints

The set of prediction models included in the DMPC scheme of ILCh are described as follows.
The previous discrete time models of (4.6) are generalised for k + m steps ahead in (4.8a) and
(4.8b), where m ∈ Z+ . Additionally, the incremental operator (3.9) is applied to derive the power
ILC ).
contributions as a function of their variation (∆Ph,i

Piac (k + m) =2Piac (k + m − 1) − Piac (k + m − 2)

(4.8a)
ILC
+ ∆Ph,i (k + m − 1) ∀ i ∈ Nac

Pjdc (k + m) =2Pjdc (k + m − 1) − Pjdc (k + m − 2)

(4.8b)
− ∆Ph,ILC
j (k + m − 1) ∀ j ∈ Ndc

To obtain the active power reference variation (∆PhILC ) for ILCh , the power contributions from
each communicated DG are added in (4.9a). Where the terms aih (k) and a jh (k) represent the
communication from the ac DGi and from the dc DG j to ILCh , respectively. A discrete time
integrator is used to obtain the power reference for ILCh , ensuring zero error in steady-state, as
shown in Fig. 4.1. Note that PhILC > 0 indicates that power is flowing from the dc sub-MG to the

59
ac sub-MG. Conversely, PhILC < 0 indicates that power is flowing from the ac sub-MG to the dc
sub-MG.

∆ PhILC (k + m − 1) = ∑ ILC
aih (k)∆Ph,i (k + m − 1) =
i∈Nac
(4.9a)
∑ a jh (k)∆Ph,ILC
j (k + m − 1)
j∈Ndc

The incremental cost (IC) models for ac DGs and dc DGs are presented in (4.10a) and (4.10b),
respectively.

ηiac (k + m) = 2ai Piac (k + m) + bi ∀ i ∈ Nac (4.10a)

η dc dc
j (k + m) = 2a j Pj (k + m) + b j ∀ j ∈ Ndc (4.10b)

Finally, the ILC’s maximum active power rating model is included in (4.11) to guarantee that
the power reference required is within the equipment limits of ILCh .

ILC
Ph,min ≤ PhILC (k + m − 1) ≤ Ph,max
ILC
(4.11)

The proposed DMPC comprises a quadratic cost function (4.7), linear equality constraints and
linear inequality constraints (4.8)-(4.11); thus, it is convex and can be synthesised in a canonical
quadratic programming (QP) formulation. Moreover, the methodology to solve the DMPC scheme
is similar to the one described in Algorithm 1. The optimisation vector of the QP problem, XILC
h in
(4.12), comprises the predicted variables XILC ILC
p,h and the control decisions X∆,h presented in (4.13)

and (4.14), respectively.

XILC
h ={XILC ILC
p,h , X∆,h }
(4.12)

dc N
XILC ac ILC
p,h ={Pi (k + m), Pj (k + m), Ph (k + m), ηiac (k + m), η dc y
j (k + m)}k=1 (4.13)

Nu
XILC ILC
∆,h = {∆Ph (k + m − 1)}k=1 (4.14)

The predicted variables are sent to the communication network to achieve the consensus objec-
tives and the first control decision, ∆PhILC (k), is applied to ILCh , after passing through an integrator

60
(see Fig. 4.1). At each sample time the optimisation problem is computed with updated measures
( forming a rolling horizon) [47]. The following section explains the development and formulation
of the DMPC strategy for dc DGs, which interacts with the DMPCs of the ILCs described in this
section and the DMPCs of the ac DGs described in Section 4.9.

4.7 Proposed DMPC scheme for dc Generators

The control scheme for the dc DGi is depicted in Fig. 4.2, where the primary control (lower half)
is based on P − V droop control [18, 43] and the proposed DMPC scheme for secondary level
is presented in the orange box. The local measurements/estimates (Pidc (k), Vidc (k), V̂idc,B (k)) and
the results of the optimisation problems from connected neighbouring units are the inputs of the
DMPC. The controller has two outputs, which are the voltage control action (vector ∆Vs,i dc ) and the

results of the local optimisation problem Xdc

p,i (vector of predicted values), both defined in Section
4.8.1. The former passes through a discrete integrator to ensure zero error in steady-state, Whereas
the latter is sent via the communication network.

61
 Figure 4.2: Control diagram of DMPCi for dc DGs.

4.7.1 Dynamic models used for the design of the controller for dc generators
To model the dynamics of the dc DGs, the voltage at each node, along with the active power flow,
are considered. Since these variables are coupled, in this work, they are modelled using the droop
(4.15) and active power transfer (4.16) equations.

The droop model for P −V for the local i − th dc DG is given in (4.15).

Vidc (t) = V0dc + Mpv,i Pidc (t) +Vs,idc (t) (4.15)

where Vidc is the output voltage of dc DGi , Pidc is the active power transferred to the MG, com-
puted by (4.16), V0dc is the nominal voltage of the dc sub-MG, M pv,i is the droop slope, and Vs,i
dc is

the secondary control action.

Equation (4.16) determines the power contribution of DGi to the MG. This equation is reformu-
lated to avoid dependence on the output current (Iidc ) of DGi .

62
 Pidc (t) = Vidc (t)Iidc (t) = GiVidc (t)(Vidc (t) − V̂idc,B (t)) (4.16)

Note that by using (4.16) a complete electrical model of the MG is not needed. Moreover, (4.16)
only uses local measurements and the voltage estimation V̂idc,B after the coupling resistor Ri (with
Gi = 1/Ri ).

4.8 Formulation of the DMPC for dc generators

The following explanations and mathematical analysis are performed for the i − th dc DG, as the
analysis is analogous for the rest of dc DGs. The optimisation problem with its cost function and
constraints are described in Section 4.8.1 and Section 4.8.2, respectively. The predicted variables
and control actions sequence are contained in the vector Xdci (defined in Section 4.8.2), which is the
solution to the optimisation problem (4.17). Only the first control action is applied to the system
dc (k)), and the optimal control problem is repeated at each sample time with updated measures
(∆Vs,i
[45]. The optimisation problem and how it is solved is detailed in the next section.

4.8.1 Cost function

The multiobjective cost function is stated in (4.17) and is composed of four quadratic weighted
terms, where each term seeks a specific objective.

Ny  2
dc dc dc
Ji (k) = ∑ ∑ λ1i a i j (k) η i (k + m) − η j (k + m − τ̂ij )
j∈Ndc m=1
Ny  2
+ ∑ ∑ λ2iai j (k) ηidc (k + m) − η ac
j (k + m − τ̂i j ) (4.17)
j∈Nac m=1
Ny h i Nu
dc
+ ∑ λ3i (Vaux,i (k + m))2 + ∑ λ4i (∆Vs,idc (k + m − 1))2
m=1 k=1

The first term achieves the consensus over the ICs within the dc DGs, while the second term
performs the consensus for the ICs of the ac DGs. The latter objective only works when the ILCs
are enabled; the controller verifies the ILCs status (1:ON, 0:OFF) at each sample time. The third
term achieves the regulation of the average dc voltage within a band by penalising the auxiliary
dc . This term temporally relaxes the average dc voltage regulation constraint (4.21b),
variable Vaux,i
allowing the average dc voltage to take transient values outside its predefined band when the MG is
disturbed. The last term penalises the control effort to achieve all the previous objectives with good
transient behaviour. The terms λ1i to λ4i are positive tuning parameters explained in Section 4.11.
Note that the cooperative objectives (IC consensus with dc DGs and with ac DGs) are updated only

63
with the predicted information of connected neighbouring DGs, represented by the terms ai j (k),
and the estimated delays τ̂i j .

4.8.2 Predictive models and constraints

The following equations present the prediction models included as equality and inequality con-
straints in the DMPC formulation for dc DGs. Models (4.18) and (4.19) are the discretised versions
of (4.15) and (4.16) via the forward Euler method, where in the former, the incremental operator
(3.9) was applied and in the latter a Taylor expansion around the measured/estimated point {Vidc (k),
V̂idc,B (k), Pidc (k)} is used. The derivation of this predictive model is detailed in Annexed E.

Vidc (k + m) =Mpv,i [Pidc (k + m) − Pidc (k + m − 1)]

(4.18)
+Vidc (k + m − 1) + ∆Vs,idc (k + m − 1)

Pidc (k + m) =[Vidc (k + m) −Vidc (k)]Gi [2Vidc (k) − V̂idc,B (k)]

(4.19)
+ Pidc (k)

Model (4.20a) represents the IC of the i −th dc DG, while a local average dc voltage approxima-
dc dc
tion (V i ) is computed in (4.20b). Note that V i also depends on the communication terms ai j (k)
and the estimated delay (τ̂i j ).

ηidc (k+m)=2ai Pidc(k+m)+bi (4.20a)

Vidc (k+m)+ ∑ ai j (k)V jdc (k+m−τ̂i j )

dc j∈Ndc
V i (k+m)= (4.20b)
1+ ∑ ai j (k)
j∈Ndc

The active power contribution is limited within the DG’s power rating in (4.21a). Finally, the
soft constraint (4.21b) works in conjunction with the cost function (see third term of (4.17)) to
regulate the average dc voltage in a predefined band within the recommendation of IEEE standard
1547-2018 [40] and avoid unfeasible solutions [47].

dc
Pi,min≤Pidc (k+m)≤Pi,max
dc
(4.21a)

dc dc dc dc
V min≤V i (k+m)+Vaux,i (k+m)≤V max (4.21b)

64
 dc is an auxiliary variable that acts as a slack variable.
Vaux,i

The proposed DMPC is synthesised in a QP formulation. The methodology to solve the DMPC
scheme is similar to the one described in Algorithm 1. The optimisation vector of the QP problem,
Xdc
i in (4.22), comprises the predicted variables Xdc dc
p,i and the control decisions X∆,i presented in

(4.23) and (4.24), respectively.

Xdc dc dc
i ={X p,i , X∆,i }
(4.22)

dc N
Xdc dc dc dc y
p,i ={V i (k + m),Vi (k + m), Pi (k + m), ηi (k + m)}k=1
(4.23)

Nu
Xdc dc
∆,i = {∆Vs,i (k + m − 1)}k=1 (4.24)

The predicted variables are sent to the communication network and the first control decision,
dc
∆Vs,i (k), is applied to the i − th dc DG, after passing through an integrator (see Fig. 4.2). At

each sample time the optimisation problem is computed with updated measures (forming a rolling
horizon) [47]. The following section explains the development and formulation of the DMPC
strategy for ac DGs, which interacts with the DMPCs for dc DGs described in this section and the
DMPCs of ILCs described Section 4.5.

4.9 Proposed DMPC scheme for ac Generators

Unlike the DMPC strategy presented in Chapter 3, which considers active power economic dispatch
and frequency regulation, the following DMPC scheme for ac DGs also considers the economic
dispatch of reactive power and the regulation of ac voltage. Moreover, this strategy regulates both
frequency and voltage within secure bands following the recommendations of the IEEE standard
1547-2018 [40], instead of restoring these variables to nominal values. The following explanations
and mathematical analysis are performed for the i − th ac DG, as the analysis is analogous for the
rest of ac DGs.

The control scheme for the ac DGi is depicted in Fig. 4.3, where the primary controller is
based on frequency-active power (ω − P) and voltage-reactive power (V − Q) droop controllers,
as well as outer voltage (slower) and inner current (faster) cascaded Proportional-Resonant (PR)
controllers. At the DG’s output, an LCL filter is placed, where the second inductance (Li ) is
set to ensure an impedance predominantly inductive [63, 64]. The proposed DMPC for the sec-

65
Figure 4.3: Control diagram of DMPCi for ac DGs.

66
ondary level is presented in the orange box. Its inputs are the local measurements/estimates
(Pi (k), Viac (k), ωi (k), θi (k), V̂iac,B (k), ω̂iB (k), θ̂iB (k)) and the results of the optimisation problems
of connected neighbouring units.

To model the behaviour of ac DGs, the voltage and frequency at each node, along with the active
and reactive power flows, are considered. Since these variables are coupled, in this thesis they are
modelled using the droop, power transfer and phase angle equations. In particular, the i − th ac
DG unit is modelled as the node at the output of its LCL filter, as shown at the bottom of Fig. 4.3,
through its voltage Viac , its angular speed ωi , and its phase angle θi . The i − th unit is connected to
the rest of the MG through inductor Li .

Unlike previous approaches [49] where external sensors are used, a sensorless scheme is em-
ployed to estimate the voltage after the coupling inductor Li (at the connection bus); thus, only
the usual voltage and current measurements at the LCL filter are needed. The estimated voltages
V̂iac,B are computed using a reduced-order state observer based on [139] and explained in detail
in Annexed B. Finally, the angular frequency and phase angle at the coupling point ω̂iB and θ̂iB ,
respectively, are estimated using phase-locked loops (PLLs).

4.9.1 Dynamic models used for the design of the controller for ac generators
The dynamic models for the angular speed (3.6), phase angle difference (3.7) and active power
transference (3.8) introduced before in Chapter 3 are also used in this controller. Thus, in this
section, only the new dynamic models included in the DMPC formulation are presented.

The V − Q droop model, which allows that a variation in the reactive power (Qi ) be reflected as
a voltage variation (Vi ) is presented in (4.25).

Viac (t) = V0ac + Mqv,i Qi (t) +Vs,iac (t) (4.25)

where Viac is the output voltage of ac DGi , Qi is the active power transferred to the MG, com-
puted by (4.26), V0ac is the nominal voltage of the ac sub-MG, Mqv,i is the droop slope, and Vs,i
ac is

the secondary control action.

The reactive power contribution of ac DGi is computed in (4.26).

h i
Qi (t) = Bi (Viac (t))2 −Viac (t) V̂iac,B (t) cos (δ θi (t)) (4.26)

where, Bi =1/(ω0 Li ), is the nominal admittance of the coupling inductor Li .

67
4.10 Formulation of the DMPC for ac generators
In a similar fashion to the models presented in Section 3.4.2, the models of Section 4.9.1 are
discretised using the forward Euler method. Furthermore, the incremental operator (3.9) is applied
ac ) and
in (4.25) to express the optimisation problem as a function of the control action variation (∆Vs,i
a Taylor expansion is applied in (4.26) to linearise the reactive power transfer model. The procedure
to obtain the all the prediction models is detailed in Annexed C. The optimisation problem with its
cost function and constraints are described in Section 4.10.1 and Section 4.10.2, respectively. The
predicted variables and control actions sequence are contained in the vector Xac i (defined in Section
4.10.2), which is the solution to the optimisation problem (4.27). Only the first control action is
applied to the system, and the optimal control problem is repeated at each sample time with updated
measures [45]. The optimisation problem and how it is solved are detailed in the next section.

4.10.1 Cost function

The multiobjective cost function is stated in (4.27) and is composed of seven quadratic weighted
terms, where each term seeks a specific objective.

Ny

2
Jiac (k) = ∑ ∑ λ1iai j (k) ηiac (k + m) − η ac
j (k + m − τ̂i j )
j∈Nac m=1
Ny  2
ac dc
+ ∑ ∑ λ2iai j (k) ηi (k + m) − η j (k + m − τ̂i j )
j∈Ndc m=1
Ny

2
+ ∑ ∑ λ3iai j (k) Ψi (k + m) − Ψ j (k + m − τ̂i j ) (4.27)
j∈Nac m=1
Ny

+ ∑ λ4i (ωaux,i (k + m))2 + λ5i (Vaux,i

ac
(k + m))2
 
m=1
Nu 
+ ∑ λ6i (∆Vs,i (k + m − 1))2 + λ7i (∆ωs,i (k + m − 1))2

m=1

The first term achieves the consensus over the ICs within the ac DGs, while the second term
performs the consensus for the ICs of the dc DGs. The latter objective works only when the ILCs
are enabled; the controller verifies the ILCs status (1:ON, 0:OFF) at each sample time. The third
term performs the consensus over the RMC, hence guaranteeing the economic dispatch of reactive
power. The fourth and fifth terms regulate the frequency and the average ac voltage within bands
ac , respectively. These terms temporally relax the
by penalising the auxiliary variables ωaux,i Vaux,i
frequency constraint (4.38) and the average ac voltage constraint (4.39), allowing these variables
to take values outside their predefined bands for a short period of time.

68
 The sixth term penalises any variations of the voltage control action, and the seventh term pe-
nalises any variation of the frequency control action. The controller achieves all the previous ob-
jectives with good transitory behaviour with only two control actions (∆Vs,iac and ∆ω ). The terms
s,i
λ1i to λ7i are positive tuning parameters explained in Section 4.11. Note that the cooperative ob-
jectives (IC consensus with ac DGs and with dc DGs and RMC consensus) are updated only with
the predicted information of connected neighbouring DGs, represented by the terms ai j (k), and the
estimated delays τ̂i j .

4.10.2 Predictive models and constraints

The linear discrete time prediction models included in the controller to rule its behaviour are de-
tailed as follows. Model (4.28) represents the incremental cost (IC) prediction model for the i − th
ac DG while (4.29) is its reactive marginal cost (RMC) prediction model; both models were derived
in Section 4.2 and Section 4.3, respectively.

ηiac (k + m) = 2ai Piac (k + m) + bi (4.28)

Ψi (k + m) = 2a′i (k) Si,res (k) Qi (k + m) + b′i (k) (4.29)

Model (4.30) represents the ω − P droop control, and model (4.31) is the V − Q droop control.

ωi (k + m) =ωi (k + m − 1) + M pω,i [Piac (k + m) − Piac (k + m − 1)]

(4.30)
+ ∆ωs,i (k + m − 1)

Viac (k + m) =Viac (k + m − 1) + Mqv,i [Qi (k + m) − Qi (k + m − 1)]

(4.31)
+ ∆Vs,iac (k + m − 1)

Model (4.32) calculates the phase angle deviation (δ θi ) due to the inductance Li of the LCL
output filter (see at the bottom of Fig. 4.3). This model is used to predict the active and reactive
power contributions of ac DGi to the MG, where ω̂iB (k) is the frequency estimated after Li .

δ θi (k + m) = δ θi (k + m − 1) + Tsec ωi (k + m) − ω̂iB (k)

 
(4.32)

The active power contribution from ac DGi to the MG is computed by the linearised model
(4.33).

69
 Piac (k + m) = Piac (k) + [Viac (k + m) −Viac (k)] BiV̂iac,B (k) sin (δ θi (k))
(4.33)
+ [δ θi (k + m) − δ θi (k)] BiViac (k) V̂iac,B (k) cos (δ θi (k))

where Bi = 1/(Li · ω0 ), Viac (k) and V̂iac,B (k) are the voltage measurements and estimations
before and after the inductance Li , Piac (k) is the active power measurement, and δ θi (k) is the phase
angle deviation measurement. It is worth noting that, unlike (3.12c), a linear term is added to the
active power prediction model. This is because the formulation considers the DG’s voltage control
(see Annexed C).

Similarly, the reactive power contribution from ac DGi to the MG is determined by the linearised
model (4.34), where Qi (k) is the reactive power measurement.

Qi (k + m) = Qi (k) + [δ θi (k + m) − δ θi (k)] BiViac (k) V̂iac,B (k) sin (δ θi (k))

h i (4.34)
+ [Viac (k + m) −Viac (k)] Bi 2Viac (k) − V̂iac,B (k) cos (δ θi (k))

The active and reactive power contributions are bounded within the DG’s power rating (Si,max )
through the linearised triangular constraint in (4.35). The procedure to obtain the power rating
constraint is detailed in Annexed D.

|Piac (k)| + |Qi (k)| + sign (Piac (k)) [Piac (k + m) − Piac (k)]
(4.35)
+ sign (Qi (k)) [Qi (k + m) − Qi (k)] ≤ Si,max

ac
Local approximations of the average frequency (ω i ) and the average ac voltage (V i ) are com-
puted by models (4.36) and (4.37), respectively. Note that both average approximations are updated
only with the predicted information from connected neighbouring ac DGs, represented by the terms
ai j (k), and the estimated delays τ̂i j .

ωi (k + m) + ∑ ai j (k)ω j (k + m − τ̂i j )
j∈Nac
ω i (k + m) = (4.36)
1 + ∑ ai j (k)
j∈Nac

Viac (k + m) + ∑ ai j (k)V jac (k + m − τ̂i j )

ac j∈Nac
V i (k + m) = (4.37)
1 + ∑ ai j (k)
j∈Nac

Finally, soft constraints (4.38) and (4.39) work together with the cost function (see fourth and
fifth terms of (4.27)) to maintain both average frequency and average ac voltage within predefined

70
bands according to the recommendation of IEEE standard 1547-2018 [40] and avoid unfeasible
ac .
solutions [47] by using the auxiliary variables ωaux,i , Vaux,i

ω min ≤ ω i (k + m) + ωaux,i (k + m) ≤ ω max (4.38)

ac ac ac ac
V min ≤ V i (k + m) +Vaux,i (k + m) ≤ V max (4.39)

The proposed DMPC is synthesised in a QP formulation, similar to (3.16). Moreover, the

methodology to solve the DMPC scheme is similar to the one described in Algorithm 1. The
optimisation vector of the QP problem, Xac i in (4.40), comprises the predicted variables Xac p,i and

the control decisions Xac∆,i presented in (4.41) and (4.42), respectively. The former is sent to the

communication network and the first control decisions of the latter, ∆Vs,iac (k), and ∆ωs,i (k), are ap-
plied to ac DGi , after passing through discrete integrators (see Fig. 4.3). At each sample time the
optimisation problem is computed with updated measures (rolling horizon) [47].

Xac ac ac
i ={X p,i , X∆,i }
(4.40)

Xac ac ac
p,i = [ηi (k + m), ωi (k + m), δ θi (k + m),Vi (k + m),
ac
ωaux,i (k + m),Vaux,i (k + m), Piac (k + m), Ψi (k + m), (4.41)
ac Ny
Qi (k + m), ω i (k + m),V i (k + m)]m=1

Nu
Xac ac
∆,i = [∆Vs,i (k + m − 1), ∆ωs,i (k + m − 1)]m=1 (4.42)

In summary, all the proposed DMPCs in this chapter have quadratic cost functions with linear
constraints; thus, their respective minimums can be reached [45, 47]. Note that in models (4.19),
(4.29) and (4.32) to (4.35) all the variables at time instant k are local measurements and estima-
tions produced in the discretisation and linearisation of the continuous time models; they are not
predicted variables. Hence, all the previous models are linear. The distributed structure inherently
addresses communication issues and plug-and-play scenarios. Moreover, each DG or ILC performs
its own optimisation locally; thus, the computational burden is reduced and does not increase when
more DGs or ILCs are added to the H-MG. The controllers identify when the ILCs are enabled
before performing the economic dispatch in the entire H-MG. In contrast, when there are no ILCs
available, the sub-MGs work independently, i.e. the economic dispatch is performed in each sub-
MG, and the voltage (and frequency for the ac sub-MG) is regulated within appropriate bands.

71
4.11 Simulation results
To evaluate the performance of the proposed DMPC scheme, the H-MG in Fig. 4.4 was simu-
lated using the electrical parameters described in Table 4.1. The PLECS blockset® was used to
build the MG electrical model, and the Matlab/Simulink® environment was used to implement the
controllers.

The system comprises an ac sub-MG with 5 ac DGs, a dc sub-MG with 5 dc DGs and two
interlinking converters (ILCs) that connect both sub-MGs. The ac distribution lines are inductive-
resistive (RL), while the dc distribution lines are resistive (R). There are four RL loads on the ac
sub-MG and four R loads on the dc sub-MG. The ILCs’ maximum power ratings are given in
Table 4.1 while the cost and operating parameters of the ac DGs and dc DGs are given in Table 4.2.
In Table 4.2, ai , bi and ci are the constant generating cost parameters used in the quadratic cost
function (4.2). These parameters represent the generating cost of the DGs and their operating point
efficiency. Note that in Fig. 4.4 the dashed lines represent the communication network, and its
associated adjacency matrix A is also shown. The subsections of the matrix A in cyan represent the
communication links within the ac sub-MG or within the dc sub-MG. In contrast, the subsections
in red represent the communication links between sub-MGs, and the subsections in green are the
communication links to the ILCs from both sub-MGs. From left to right (and top to bottom), the
elements of A belong to ac DGs, dc DGs and ILCs, respectively.

The MG electrical model was built in PLECS blockset®, and the primary and secondary con-
trollers were implemented in Matlab/Simulink® environment. The primary level consists of the
droop, inner voltage and current controllers. The inner voltage and current controllers were imple-
mented as self-tuning proportional-resonant controllers for ac variables and proportional-integral
controllers for dc variables. The following simplifications were considered in the simulator since
their high-bandwidth dynamics are much faster than that of the studied controllers, and they are not
relevant on the time-scale of the proposed controllers:

• The DGs are simulated as regulable voltage sources.

• The modulation techniques (PWM, SVM) of the converters are not considered.

• The switching of the switching devices is not considered.

72
Figure 4.4: Hybrid ac/dc MG topology for the validation of the DMPC scheme

Table 4.1: MG parameters and loads

Parameter Description Value

Tprim Primary level sample time [s] 1/(16 · 103 )

Li i − th ac DG coupling inductance [mH] 2.5

Zi j ac sub-MG distribution lines [Ω] 0.7 + j2.5 · 2π

Ri i − th dc DG coupling resistor [Ω] 0.67

Ri j dc sub-MG distribution lines [Ω] 0.78

Z1 ; Z2 ac sub-MG loads [kVA] 5.7+j2.3; 1.5+j0.2

Z3 ; Z4 ac sub-MG loads [kVA] 2+j0.5; 2.3+j0.3

R1 ; R2 ; R3 ; R4 dc sub-MG loads [kW ] 3.36; 1.32; 1.13; 1.11

ILC ; PILC
P1,max 2,max ILCs power rating 5.0; 3.0

ω0 ac sub-MG nominal frequency 2π · 50

V0ac ac sub-MG nominal voltage 220

V0dc dc sub-MG nominal voltage 400

73
 Table 4.2: DGs parameters
dc DGs parameters

Parameter DGdc dc dc dc dc
1 DG2 DG3 DG4 DG5

a [$/kW h2 ] 0.35 0.37 0.46 0.51 0.52

b [$/kW h] 1.8 2.9 2.0 2.6 1.6

c [$] 0 0 0 0 0

Power capacity (Pi,max ) [kW] 2.5

V
P −V droop coefficient (Mpv ) [ W ] −1.2 · 10−2

ac DGs parameters

Parameter DGac ac ac ac ac
1 DG2 DG3 DG4 DG5

a [$/kW h2 ] 0.39 0.44 0.49 0.55 0.66

b [$/kW h] 2.9 2.3 2.4 1.5 1.9

c [$] 0 0 0 0 0

Power capacity (Si,max ) [kVA] 3

P − ω droop coefficient (Mpω ) [ rad

sW ] −3.33 · 10−4
V
Q −V droop coefficient (Mqv ) [ VAR ] −1.5 · 10−2

4.11.1 Design parameters and test scenarios used to evaluate the DMPC for
H-MGs

The DMPC design parameters and weighting factors are presented in Table 4.3. All the design
parameters were selected to reduce the computational burden. The sample time was selected con-
sidering the open loop rise time (Tr = 0.7s) of the ac sub-MG’s frequency (operating with only the
primary control enabled) as Tsec = 0.7/14 = 0.05s [137]. The weighting factors were tuned follow-
ing the guidelines in [88], i.e., looking for a trade-off between the control objectives and, if needed
giving more importance to one objective over the rest of the objectives. The average ac voltage and
dc voltage are limited to a band of ±5 [V] of their nominal values V0ac and V0dc , respectively, while
the frequency is limited to a band of ±π [rad/s] of the nominal value ω0 . Although the limits are
fixed for all the test scenarios, these can be modified as long as they comply with IEEE standard
1547-2018 [40].

74
 Table 4.3: DMPC parameters and Weights
Parameter Description Value

Tsec [s] Controller sample time 0.05

τ̂i j [s] Estimated communication delay 0.05

Ny Prediction horizon 10

Nu Control horizon 10

DMPC for ILC weights

λ1h [(kW h/$)2 ] Active power dispatch for DGs 4 · 100

λ2h [-] Active power consensus for ILCs 4 · 104

λ3h [(1/kW )2 ] ILC control action (∆PhILC ) 5 · 103

DMPC for dc DGs parameters and weights

[Vmax ;V min] Average voltage predefined band [395,405]

λ1i [(kW h/$)2 ] Active power dispatch with dc DGs 3.5 · 10−4

λ2i [(kW h/$)2 ] Active power dispatch with ac DGs 3.5 · 10−4

λ3i [(1/V )2 ] Average voltage regulation 7 · 103

λ4i [( V1 )2 ] Voltage control action variation 16 · 102

DMPC for ac DGs parameters and weights

[Vmax ;V min] Average voltage predefined band [215,225]

[ωmax , ωmin ] [rad/s] Frequency predefined band [101π, 99π]

λ1i [(kW h/$)2 ] Active power dispatch with ac DGs 2.1 · 10−3

λ2i [(kW h/$)2 ] Active power dispatch with dc DGs 2.1 · 10−3

λ3i [(kVAR/$)2 ] Reactive power dispatch 4.2 · 101

s 2
λ4i [( rad ) ] Average frequency regulation 3.8 · 106

λ5i [(1/V )2 ] Average voltage regulation 5.0 · 105

s 2
λ6i [( rad ) ] Frequency control action variation 4.5 · 106

λ7i [( V1 )2 ] Voltage control action variation 1.9 · 104

The controller was tested under three scenarios using the H-MG simulator, shown in Fig. 5.3.
The first scenario presents the DMPC’s performance when the H-MG experiences ac and dc load

75
changes on both sub-MGs. The second scenario shows the behaviour of the MG when two issues
are present simultaneously. These issues are communication path failures and disconnection and
connection of ac DGs, dc DGs and ILCs. The last scenario presents a comparison between the
proposed DMPC and a reported technique based on the distributed averaging proportional-integral
(DAPI) controller. For this scenario, both controllers are tested when the entire communication
network has communication delays. These three test scenarios were selected because these are the
most common phenomena that a controller at the secondary level confronts [49,134]. A distributed
controller must perform well against communication issues, such as communication delays and
failures. In addition, the possibility of disconnecting and reconnecting DGs and ILCs to the H-MG
without changes in the programming of the controllers is essential.

4.11.2 Scenario I (base case) - Load changes

This test evaluates the performance of the proposed DMPC scheme when the MG is subject to ac
and dc load steps. For the whole test the communication network does not vary, i.e., the adjacency
matrix A is kept constant and is described in Fig. 4.4. The test starts with both ac DGs and dc DGs
enabled and working only with the primary control while the ILC is disabled. Hence, the ac sub-
MG and the dc sub-MG are working separately, and its dynamic is fixed by the droop controller.
Additionally Z1 , Z2 , R1 , and R2 are connected at their respective nodes (see Fig. 4.4). Note that
as the DMPC controllers are disabled, there is neither consensus in the incremental cost (IC) nor
consensus in the reactive marginal cost (RMC) (see Fig. 4.5.a and Fig. 4.5.b before t = 10s); there-
fore, active and reactive power are not economically dispatched (see Fig. 4.5.c and Fig. 4.5.d before
t = 10s). Furthermore, the frequency and voltages of the H-MG are outside of the established bands
(see Fig. 4.6.b, Fig. 4.6.c and Fig. 4.6.d before t = 10s).

At t = 10s, the predictive controllers for the sub-MGs are enabled, but the ILCs are still disabled,
so there is no power transference between the sub-MGs yet. Therefore, the DMPC controllers
optimise the DGs’ performance locally, i.e., the second terms (objectives) in the cost functions of
(4.17) and (4.27) are disabled. It is observed that on the ac sub-MG, consensus on both the IC and
RMC are achieved (see Fig. 4.5.a and Fig. 4.5.b at t = 10s). The consensus values for IC and RMC
are achieved as all the ac DGs achieve the same value in steady-state. Thus, both active and reactive
power are redispatched economically, considering generating costs (see Fig. 4.5.c and Fig. 4.5.d at
t = 10s). Moreover, both frequency and average ac voltage are regulated within the established
bands (see Fig. 4.6.b and Fig. 4.6.d at t = 10s). Similarly, on the dc sub-MG, the consensus on the
IC is achieved, hence active power is redispatched based on the DGs’ generation costs, as shown
in Fig. 4.5.a and Fig. 4.5.c at t = 10s. Additionally, the average dc voltage is regulated within its
established band (see Fig. 4.6.c at t = 10s).

The ILCs and their DMPC strategies are enabled at t = 30s, but initially only the IC consensus

76
 DGac
1
DGac
2
DGac
3
DGac
4
DGac
5
DGdc
1
DGdc
2
DGdc
3
DGdc
4
DGdc
5

Figure 4.5: Load changes: a) Incremental cost consensus, b) Reactive marginal cost consensus, c)
Active power, d) Reactive power

F average

P ILC
1
P ILC
2

Vdc
average
Vac
average

Figure 4.6: Load changes: a) Active power through the ILCs, b) Frequency regulation, c) Average
dc voltage regulation, d) Average ac voltage regulation. The dashed cyan lines represent the prede-
fined band limits for frequency and voltages.

objective (second term) in (4.7) is enable. Thus, the ILCs equalise the ICs on both sub MGs to
achieve global economic dispatch (see Fig. 4.5.a at t = 30s), but they do not transfer active power
proportionally to their power rating (see Fig. 4.6.a at t = 30s), which could cause overloading in
the ILCs. Fig. 4.5.c shows that the DGs on the dc sub-MG increase their power contribution to
transfer power to the ac sub-MG because they have lower generation costs (see Table 4.2) while

77
the ac DGs decrease theirs, as they are more expensive. At t = 50s, the power consensus on the
ILCs is enabled (third term) in (4.7). Therefore, both ILCs transfer power proportionally to their
power rating, as shown in Fig. 4.6.a at t = 50s.

Then, at t = 70s the total ac load (Z3 and Z4 ) is connected. As the dc DGs are cheaper, they
increase their power contribution more than the ac DGs (see Fig. 4.5.c at t = 70s) to transfer more
through the ILCs (see Fig. 4.6.a at t = 70s) and achieve the consensus value (see Fig. 4.5.a at t =
70s). This takes the frequency and voltages outside of their bands for a short time; however, these
variables are regulated immediately by the DMPC controllers, as shown in Fig. 4.6.b, Fig. 4.6.c,
and Fig. 4.6.d at t = 70s. Finally, at t = 90s, R3 and R4 are connected; hence, the H-MG is subject
to its total load. Due to the dc sub-MG increasing their load consumption and the dc DGs being
cheaper, this new load is mostly supplied by its local DGs. Moreover, as two dc DGs almost reach
their power rating limit, the ac DGs also increase their power contribution to supply the remaining
load. Therefore, the power transferred through the ILCs is reduced.

It is observed that the economic dispatch of active and reactive power is fulfilled under this de-
manding operating condition. In addition, even though the disturbances are sudden, the frequency
and voltages are regulated within their bands at all times. The results show that the proposed DMPC
scheme achieves all the objectives without large overshoots and with settling times below 8 seconds
for all the scenarios tested, and all the equipment constraints have been respected

4.11.3 Scenario II - Combined communication link failures and Plug-and-

Play
This test evaluates the performance of the DMPC scheme when two of the most common failures
occur, i.e., communication link failures, and disconnection and reconnection of DGs and ILCs.
The test starts with the predictive controllers of the ac DGs, dc DGs and ILCs enabled. Loads Z1 ,
Z2 and R1 to R4 are connected at their respective nodes (see Fig. 4.4). A communication failure
is forced at t = 10s between the communication channels of DGac dc ac dc
1 to DG1 , DG3 to DG3 and
DGac dc
5 to DG5 (see red dashed lines in Fig. 4.4). Note that when these communication links fail,
the H-MG is coordinated through the communication links to the ILCs (see green dashed lines in
Fig. 4.4), and the DGs only communicate with the neighbouring DGs within their corresponding
sub-MGs. Furthermore, the adjacency matrix is modified when the communication links fail.

The control algorithms of dc DGs and ac DGs automatically identify the failure and change
the consensus calculations. Specifically, the second terms in (4.17) for dc DGs and (4.27) for ac
DGs, respectively, are not considered in the controller. The predictive controllers do not suffer any
noticeable deterioration in performance, as shown in Fig. 4.7 and Fig. 4.8 at t = 10s. Note that the
communication links are not reestablished for the remain of the test. Then at t = 20s, t = 40s, and

78
 F average

P ILC
1
P ILC
2

Vdc
average
F average

Figure 4.7: Communication failure and plug-and-play test: a) Active power through the ILCs, b)
Frequency regulation, c) Average dc voltage regulation, d) Average ac voltage regulation. The
dashed cyan lines represent the predefined band limits for frequency and voltages.

DGac DGac DGac DGac DGac DGdc DGdc DGdc DGdc DGdc
1 2 3 4 5 1 2 3 4 5

Figure 4.8: Communication failure and plug-and-play test: a) Incremental cost consensus, b) Re-
active marginal cost consensus, c) Active power, d) Reactive power

79
t = 60s three unscheduled failures occur in the H-MG, i.e. ILC2 , DGac dc
1 , and DG1 are disconnected,
respectively, from both the electrical system and the communication network. The MG continues
operating with the remaining connected DGs and ILC until t = 80s, where DGdc 1 is reconnected
ac
to the MG. In a similar way, DG1 , after a synchronisation routine, and ILC2 are reconnected at
t = 100s and t = 120s, respectively.

The results in Fig. 4.7 and Fig. 4.8 show that the controller performance is not affected by both
phenomena, and the remaining operating DGs and ILCs achieve the consensus objectives, even
when both failures are present at the same time. Nevertheless, the transient response is different;
both overshoot and settling time increase slightly. This is because the adjacency matrix is not
complete, and the consensus objectives depend on the known information of the neighbouring DGs
and ILCs. Note that when DGdc ac
1 and DG1 are reconnected (see Fig. 4.8 at t = 80s and t = 100s),
these DG units achieve the consensus objectives in a longer time. Next the performance of the
proposed strategy against a DAPI-based strategy under the presence of communication delays is
evaluated.

4.11.4 Scenario III - Comparison against a DAPI-based strategy without

economic dispatch for H-MGs
This section presents a comparison study between the proposed DMPC strategy and the reported
technique in [43]. The DAPI-based method in [43] shares active and reactive power proportionally
to the DGs’ power rating without considering the DGs generation costs and restores frequency and
voltage to nominal values. The same simulator and adjacency matrix (A) presented in Fig. 4.4 is
used for this comparison study. The selected scenario is communication delays, one of the most
common phenomena in distributed controllers. A communication delay of one second is applied
in the entire communication network (τi j = 1s) for the whole test. For this test, both the proposed
DMPC strategy and the DAPI strategy [43] start with all their functionalities enabled and all the
loads connected except R3 and R4 . Fig. 4.9 presents on the left-hand side the results of active
power contribution and active power transference through the ILCs for [43], while the results of
these variables for the proposed DMPC are on the right-hand side. Conversely, Fig. 4.10 presents
in each graph a performance comparison of both control strategies.

At t = 10s, the H-MG is subject to its total load, i.e., R3 and R4 are connected. It is observed
that the DAPI controller shares power proportionally while the proposed DMPC dispatches the DGs
considering their operation costs (see Fig. 4.9.a and Fig. 4.9.b at t = 10s). The proposed DMPC
transfers more active power through the ILCs than the DAPI controller (see Fig. 4.9.c and Fig. 4.9.d
at t = 10s); this is because the DMPC achieves the economical dispatch point by dispatching more
power from the dc DGs, as these are cheaper (see Table 4.2). Furthermore, when the H-MG has
its total load connected, the operation cost is reduced by up to 4.55%, i.e., from 62.25 $/h to

80
59.42 $/h, as shown in Fig. 4.10.a. Regarding frequency and voltage, the DAPI controller restores
these variables to nominal values while the proposed DMPC provides more flexibility to the H-MG
and regulates them only when they are outside their established bands, as shown in Fig. 4.10.b,
Fig. 4.10.c and Fig. 4.10.d. It is crucial to note that the DAPI controller is highly affected by the
large communication delay (τi j = 1s); this controller presents larger overshoots and long settling
times compared to the DMPC in all their controlled variables. This poor performance is because
the DAPI controller does not provide a delay compensation mechanism. On the other hand, the
proposed DMPC controller has a delay estimation and uses a rolling horizon, which corrects the
control action sequences and compensates for the effects of delays [65].
DGac DGac DGac DGac DGac DGdc DGdc DGdc DGdc DGdc
1 2 3 4 5 1 2 3 4 5

P ILC P ILC P ILC

1
P ILC
2
1 2

Figure 4.9: Comparison between the proposed DMPC scheme and a DAPI-based method for (τi j =
1s): a)-b) Active power for both methods, c)-d) Active power through the ILCs for both methods.

Finally, at t = 30s and at t = 50s, Z2 and Z4 , and Z3 are disconnected, respectively. The results
during these events are consistent with the previously described performance. In summary, the
communication delay affects the behaviour of [43] significantly by increasing its overshoots and
settling time, while the proposed DMPC is slightly affected with negligible overshoots. Moreover,
the proposed DMPC has a lower operating cost during the entire test. This is because the proposed
DMPC uses the DGs of the H-MG cost-effectively. The cost reduction may seem small in monetary
terms due to the small size of the H-MG, but it could be significant in percentage terms in a larger H-
MG. Furthermore, the DMPC scheme can simultaneously regulate variables to specific values and
within bands, providing more flexibility to the H-MG while equipment constraints are satisfied.
Additionally, the DMPC can tackle more control objectives with fewer control actions, while in
the DAPI method, a new controller needs to be designed (added) for each control objective to be
addressed.

81
 Proposed DMPC Method reported in [3]

Cost DMPC Cost DAPI F DMPC F DAPI

average average

dc-DMPC
Vaverage
dc-DAPI
Vaverage Vac-DMPC
average
Vac-DAPI
average

Figure 4.10: Comparison between the proposed DMPC scheme and a DAPI-based method for
(τi j = 1s): a) Total operating cost, b) Frequency regulation, c) Average dc voltage regulation, d)
Average ac voltage regulation. The dashed cyan lines represent the predefined band limits for
frequency and voltages.

4.12 Discussion
This Chapter presented a novel DMPC strategy for isolated H-MGs to tackle simultaneously the
economic dispatch of both active power and reactive power and the regulation within bands of
frequency and voltage (fulfilling IEEE standard 1547-2018 [40]). Specifically, the frequency and
the average ac voltage on the ac sub-MG and the average dc voltage on the dc sub-MG are regu-
lated within bands. The DMPC scheme considers the H-MG as a single entity by modelling the
behaviour and interaction of ac DGs, dc DGs and ILCs. Furthermore, the dynamic performance
of the DMPC scheme is evaluated and discussed for load impacts and communication issues, such
as communication delays and communication link failures. The distributed structure of the control
scheme allows the disconnection and reconnection DGs and ILCs. In all the tests, the DMPC ful-
fils its objectives without exceeding the maximum power rating limits, preventing any overloading
of DGs and ILCs. In summary, the simulation results demonstrated the main advantages of the
proposed strategy, which are:

• Active and reactive economic dispatch is achieved among ac and dc DGs when a load is
connected to either side of the H-MG and when a DG or ILC is disconnected and reconnected
to the H-MG.

• The frequency and ac voltage in the ac sub-MG and the dc voltage on the dc sub-MG are
restored within secure bands in steady-state under all the analysed cases.

82
 • The ILCs transfer active power based on the DGs’ generation costs, thus reducing the total
operation cost of the H-MG.

• The DMPC scheme operates appropriately with a reduced communication network, where
DGs and ILCs are communicated only with neighbouring agents. Moreover, global objec-
tives are achieved through information sharing.

• DGs and ILCs can easily be connected to or disconnected from the H-MG, and the control
structure, demonstrating the plug-and-play capability of the proposed strategy.

• The superior behaviour of the DMPC scheme against usual DAPI-based controllers is demon-
strated under communication issues.

The development of these strategies and their validation were presented in the following papers:

A. Navas-Fonseca, C. Burgos-Mellado, J. S. Gómez, E. Espina, J. Llanos, D. Sáez, M. Sum-

ner, and D. E. Olivares, “Distributed predictive secondary control with soft constraints for optimal
dispatch in hybrid AC/DC microgrids," in IEEE Transactions on Smart Grid, 2022 , [Q1-IF 8.960
- under review ].

A. Navas-Fonseca, C. Burgos-Mellado, J. Gomez, J. Llanos, E. Espina, D Saez, M. Sumner,

"Distributed Predictive Control using Frequency and Voltage Soft Constraints in AC Microgrids
including Economic Dispatch of Generation " IECON 2021 – 47th Annual Conference of the IEEE
Industrial Electronics Society, 2021, pp. 1-7,
doi: 10.1109/IECON48115.2021.9589500.

A. Navas-Fonseca, C. Burgos, E. Espina, E. Rute, D. Saez, M. Sumner, "Distributed Predictive

Secondary Control for Voltage Restoration and Economic Dispatch of Generation for DC Micro-
grids," 2021 IEEE Fourth International Conference on DC Microgrids (ICDCM), 2021, pp. 1-6,
doi: 10.1109/ICDCM50975.2021.9504612.

This chapter concludes the application of DMPC for the economic dispatch of DGs in H-MG.
The next chapter presents the application of DMPC to solve the other pressing problem considered
in this thesis: the presence of unbalance between the phases in MGs. To this end, new predic-
tion models and a novel cost function are proposed. The following chapter provides a detailed
explanation of the proposed control strategy’s development and validation.

83
Chapter 5

The Proposed DMPC scheme for phase

imbalance sharing and frequency and
voltage regulation within bands in ac
microgrids

5.1 Introduction

In this Chapter, the distributed model predictive control (DMPC) strategy for imbalance sharing
and the regulation of frequency and voltage within bands for isolated ac microgrids (MGs) is pro-
posed. In the DMPC scheme of this chapter, each DG will achieve per-phase reactive power shar-
ing, three-phase active power sharing and will also restore the MG’s frequency and voltage within
recommended operating bands that comply with IEEE standard 1547-2018 [40] while operating
equipment constraints are satisfied.

The main challenge for the techniques proposed in this Chapter is to design the local models
and a cost function to tackle simultaneously the control objectives proposed. A local prediction
model of the DG’s primary control level is included in each DMPC to predict its future behaviour,
i.e., the active power - frequency droop control and the active power transfer models as well as
per-phase reactive power - voltage droop control and per-phase reactive power transfer models.
Moreover, the DMPC strategy has to be able to recognise and update its calculations when the
MG is subject to external phenomena such as communication delays, communication failures and
disconnection/reconnection of DGs.

This chapter is organised as follows: Section 5.2 introduces the proposed strategy for imbalance

84
sharing. The dynamic prediction models are presented in Section 5.3. Then, the DMPC formulation
is described in Section 5.4. The MG simulator and validation tests are explained in detail in Section
5.5. Additionally, HIL validation is presented in Section 5.6, and a comparison analysis is detailed
in Section 5.7. Finally, Section 5.8 summarises the main benefits of the proposed strategy.

5.2 Proposed DMPC scheme

In this section, a three-phase unbalanced ac MG composed of N DGs is considered. Within the
general structure of a MG, a local model for each DG is considered. Fig. 5.1 shows the control of
the i − th DG with i={1, · · · , N}. The following explanations and mathematical analysis are made
for the i − th DG, as the analysis is analogous for the rest of the DGs.

Figure 5.1: General control diagram of DMPCi for imbalance sharing.

In this work, each DG is connected to an LCL filter, where the second inductance (Li ) is set to
ensure an impedance which is predominantly inductive [63, 64], as shown in the electrical config-
uration of Fig. 5.1. The primary control of the i−th DG is shown at the bottom of Fig. 5.1. The

85
primary control is achieved by means of a droop equation for the active and reactive power of each
unit. The P − ω droop equation takes into account the three-phase active power of each DG to
calculate the angular speed ωi . The droop equation for the reactive power is defined for each phase
x with x={a, b, c} as shown in Fig. 5.1.

This novel approach allows the introduction of small imbalances in the output voltage (Vx ) of the
DG units with x={a, b, c}, enabling unbalanced power sharing. Indeed, as stated in the introduction
section, the sharing of imbalance can only be achieved by increasing the voltage imbalance at the
DG units’ output. This issue is taken into account for the proposed method, which enables the
sharing of unbalanced powers and at the same time regulates the maximum unbalanced voltage
at the DGs to enable the maximum values stated in IEEE standard 1547-2018 [40]. In this work,
the Phase Voltage Unbalance Rate index (PVUR) is used to quantify the level of imbalance in the
output voltage of the DG units (see Section 5.3.4). To create the single-phase Q − V droops, a
Quadrature Signal Generator (QSG) is applied in the voltage and current measurements of each
phase, as shown in Fig. 5.2.

Figure 5.2: Single-phase droop controller.

The QSG creates a 90-degree shifted signal when applied to sinusoidal signals, thus, creating
the virtual β components from the voltage and the current measurements, while only the α compo-
nents are actually measured. These QSGs are implemented using all-pass filters, and they present
good performance when operating around the MG nominal frequency [23, 53]. The output of the
reactive power droop control is the DG output voltage in the natural reference frame Viabc . The
DG voltage and current are then regulated through outer and inner cascaded Proportional-Resonant

86
(PR) controllers (see Fig. 5.1).

At the top of Fig. 5.1 the DMPC controller of DGi is presented. This controller receives as inputs
the local estimates and measurements ({ωi (k), ω̂iB (k),Vi (k), Vix (k), V̂iB (k), V̂ixB (k), δ θi (k), Pi (k), Qix (k)}
with x={a, b, c}) from the primary control level and state variable predictions from neighbour
units, connected via a communication network. The DMPC outputs are the frequency variation,
the per-phase voltage variations (vectors ∆ω s,i and ∆V s,ix with x={a, b, c}), and the predictions of
the local optimisation problem X p,i , defined in Section 5.4. While the control actions are processed
by discrete-time integrators to ensure zero error in steady-state, the predictions are sent directly
through the communication network. It is worth noting that the same communication principle
described previously in Section 3.4.1 is used in this control strategy.

5.3 Dynamic models used for the design of the DMPC strategy
To model the dynamics of the MG, the voltage and frequency at each node, along with the active
and reactive power flows, are considered. Since these variables are coupled, in this thesis they are
modelled using the droop, power transfer and phase angle equations. In particular, the i − th DG
unit is modelled as the node at the output of its LCL filter, as shown at the bottom of Fig. 5.1,
through the per-phase voltages Vix with x={a, b, c}, its angular speed ωi , and its phase angle θi .
The i − th unit is connected to the rest of the MG through inductor Li . Unlike previous approaches
[49] where external sensors are used, a sensorless scheme is employed to estimate the per-phase
unbalanced voltage after the coupling inductor (at the connection bus); thus, only the usual voltage
and current measurements at the LCL filter are needed. The per-phase estimated voltages V̂ixB with
x={a, b, c} are computed using a reduced-order state observer based on [139] and explained detail
in Annexed B. Also, the average voltage magnitude of the phases V̂iB is calculated. Finally, the
angular frequency and phase angle at the coupling point ω̂iB and θ̂iB , respectively, are estimated
using PLLs.

5.3.1 Droop control

To share the active and reactive power between the units belonging to the MG, the angular speed
and output voltage of each DG are computed using droop control. The droop models are included in
the DMPC scheme because droop controllers rule the MG’s behaviour, and through these models,
the secondary control interacts with the primary control. For the i − th unit, the active power
is regulated through its instantaneous angular frequency ωi (t) through the ω − P droop shown
previously in (3.6).

It should be pointed out that a single-phase P-ω droop scheme is not considered in this work
because this approach may produce differences in each phase frequency, especially during the

87
transients, which could be a drawback if the MG is feeding three-phase loads, such as motors [53].

The reactive power is shared evenly per phase. Consequently, it is necessary to regulate the
reactive power independently for each output phase. In this case, the droop control law for the
phase x={a, b, c} is shown in (5.1).

Vix (t) = V0 + Mqv,i Qix (t) +Vs,ix (t) (5.1)

The reactive power is regulated by controlling the magnitude of the output voltages Vix with
x={a, b, c}. In addition, the nominal voltage V0 is the same for all output phases, and the droop
slope for the reactive power Qix (t) is Mqv,i . Finally, the outputs of the secondary control are Vs,ix (t)
for phases x={a, b, c} of the i − th unit (after the discrete-time integrators, see Fig. 5.1).

The average voltage magnitude of the phases Vi (t) of the i −th unit is also included, and is given
by:

1
Vi (t) = Vix (t) x={a, b, c} (5.2)
3∑x

5.3.2 Phase angle model

To estimate the active and reactive power transferred from the i − th unit to the rest of the MG,
its deviation angle δ θi (t) is employed. The phase angle δ θi (t) is defined as the angular difference
between the output of the LC filter of the i − th unit and the node after the coupling inductor Li .
This model was presented in (3.7)

5.3.3 Power transfer models

To estimate the active and reactive power contribution from the i − th unit to the MG, the power
flow through its coupling inductor (Li ) is considered. By using this approach it is possible to avoid
the use of the admittance matrix of the whole MG, reducing modelling complexity and facilitating
the plug-and-play operation. The active power from DGi to the rest of the MG was presented in
(3.8). Conversely, the reactive power needs to be calculated for each output phase x={a, b, c} to
adequately share reactive power in each phase:

Qix (t) = Bi Vix2 (t) −Vix (t) V̂ixB (t) cos (δ θi (t))
 
(5.3)

where, Bi =1/(ω0 Li ), is the nominal admittance of the coupling inductor Li .

88
 The total reactive power Qi (t) of the i − th unit is also included as the sum of the single-phase
reactive powers previously defined. This model is used to limit the power contribution of DGi
within the DMPC formulation.

Qi (t) = ∑ Qix (t) x={a, b, c} (5.4)

x

5.3.4 Phase voltage unbalance rate index

As discussed in the introduction section, imbalance sharing methods should be designed to achieve
simultaneously the imbalance sharing and the regulation of the voltage at the output of the DG
units. Currently, only a few works have explored this topic [33, 39]. In this sense, the proposed
predictive controller addresses this issue by using the phase voltage unbalance rate index (PVUR)
(5.5) to quantify the unbalance at the i − th DG’s output [39].

1
PVURi (t) = max {|Vix (t) | −Vi (t)} (5.5)
Vi (t) x={a,b,c}

This index is regulated in each DG unit to meet IEEE standard 1547-2018 [40], where a max-
imum voltage imbalance of 5% is allowed. In (5.5), |Vix (t) | and Vi (t) are the per-phases voltage
magnitudes and the average voltage magnitude among the phases at t, defined in (5.1) and (5.2),
respectively.

5.4 Distributed MPC formulation for imbalance sharing

The main objective of the proposed DMPC is to share imbalances through a single-phase approach
between the DGs that comprise the MG whilst the unbalanced voltage at the converters’ output is
regulated below the maximum allowed PVUR to meet the IEEE power quality standard [40]. A
consensus in the three-phase active power is also sought in this strategy. Furthermore, instead of
restoring the average frequency and the average voltage to their nominal values, a more flexible
objective is proposed. These variables are kept within predefined bands and only restored when
they are outside their band, and this is achieved through the use of soft constraints. Note that by
regulating the MG average voltage, it is possible to have good reactive power-sharing, as the voltage
nodes can have different voltage levels, which are close to the MG average voltage.

5.4.1 Cost function

The cost function of the proposed DMPC comprises ten weighted terms and is presented in (5.6).

89
 3Qia (k + m) 3Q ja (k + m − τ̂i j ) 2
N Ny  
Ji (k) = ∑ ∑ λ1i ai j (k) −
j=1, j̸=i m=1 |Si max | |S j max |
3Qib (k + m) 3Q jb (k + m − τ̂i j ) 2
N Ny  
+ ∑ ∑ λ2i ai j (k) −
j=1, j̸=i m=1 |Si max | |S j max |
3Qic (k + m) 3Q jc (k + m − τ̂i j ) 2
N Ny  
+ ∑ ∑ λ3i ai j (k) −
j=1, j̸=i m=1 |Si max | |S j max |
Pi (k + m) Pj (k + m − τ̂i j ) 2
N Ny  
+ ∑ ∑ λ4i ai j (k) − (5.6)
j=1, j̸=i m=1 |Si max | |S j max |
Ny

+ ∑ λ5i (ωaux,i (k + m))2 + λ6i (Vaux,i (k + m))2

 
m=1
Nu 
+ ∑ λ7i (∆Vs,ia (k + m − 1))2 + λ7i (∆Vs,ib (k + m − 1))2

m=1
Nu
+ ∑ λ7i (∆Vs,ic (k + m − 1))2 + λ8i (∆ωs,i (k + m − 1))2
 
m=1

The first to the third terms achieve a normalised consensus in the reactive power contribution in
phases a, b and c, respectively, which are used to produce the sharing of imbalance. The fourth term
seeks consensus for the the three-phase normalised active power. The fifth and sixth terms penalise
the auxiliary variables ωaux,i , Vaux,i . These variables act as slack variables to keep both average
voltage and average frequency within predefined bands while an unfeasible solution is avoided.
Terms fifth and sixth work in conjunction with the soft constraints (5.15a) and (5.15b), respectively.
The seventh to ninth terms penalise any variations of the voltage control actions in each phase, and
the tenth term penalises any variation of the frequency control action. The objectives of sharing
reactive power per phase and restoring the average voltage are achieved with the control actions of
the terms seven to nine, while the objectives of sharing active power and restoring frequency are
achieved by the control action of tenth term. These control actions are applied in their respective
droop controllers (see Fig. 5.1). The terms λ1i to λ7i are positive tuning parameters explained in
Section 5.5.

By penalising the control action variations, the control effort is minimised, and the transient
performance of the controller is improved. Note that all the consensus objectives consider the
communication terms ai j (k) and the estimated time delay τ̂i j , which is defined as one sample period
at the secondary level. As the consensus objectives in each DG are optimised considering the
predictions from communicated neighbouring DGs, the regulation is global for the entire MG.

90
5.4.2 Predictive models and constraints
The optimisation problem incorporates the dynamic models presented in Section 5.3, which are
included as equality constraints based on (3.6)-(3.8), (5.1)-(5.4), and inequality constraints based
on (5.5). All these models are discretised and generalised for the prediction horizon (Ny ) and a se-
quence of control actions for the control horizon (Nu ) is calculated through a numerical optimisation
problem. The discretisation is carried out through the forward Euler method, where k = nTsec , n ∈
Z+ , and Tsec is the sample time of the controller. Then, they are generalised for k + m steps ahead,
where m ∈ Z+ . Furthermore, the incremental operator (3.9) is applied in models (3.6) and (5.1) to
express the optimisation problem as a function of the variation of the frequency control action and
the per-phase voltage control actions ∆ωs,i and ∆Vs,ix with x={a, b, c}, respectively. The non-linear
power transfer models (3.8), (5.3), and the PVUR (5.5) model are linearised via a Taylor expan-
sion around the measured/estimated point {ωi (k), ω̂iB (k),Vi (k),Vix (k),V̂iB (k), V̂ixB (k), δ θi (k), Pi (k),
Qix (k)} with x={a, b, c} before their discretisation.

The predictive models of the ω − P (3.6) and Qix −Vix (5.1) droop controllers, are presented in
(5.7a) and (5.7b), respectively.

ωi (k+m)=ωi (k+m−1)+Mpω,i[Pi (k+m)−Pi (k+m−1)]+∆ωs,i (k+m−1) (5.7a)

Vix (k+m)=Vix (k+m−1)+Mqv,i[Qix (k+m)−Qix (k+m−1)]+∆Vs,ix (k+m−1) (5.7b)

The predictive model of the average voltage, described previously in (5.2), is presented in (5.8).

1
Vi (k + m) = Vix (k + m) x={a, b, c} (5.8)
3∑x

The predictive model of the phase angle difference, described previously in (3.7), is presented
in (5.9).

δ θi (k + m) = δ θi (k + m − 1) + Tsec ωi (k + m) − ω̂iB (k)

 
(5.9)

The predictive models of the three-phase active power, described previously in (3.8), and the
per-phase reactive power, described previously in (5.3), are presented in (5.10a) and (5.10b), re-
spectively.

Pi (k+m)=Pi(k)+[Vi(k+m)−Vi(k)]BiV̂iB(k)sin(δ θi(k))
(5.10a)
+[δ θi(k+m)−δ θi(k)]BiVi(k)V̂iB(k)cos(δ θi(k))

91
 Qix(k+m)=Qix(k)+[Vix(k+m)−Vix(k)]Bi 2Vix(k)−V̂ixB(k)cos(δ θi(k))
 
(5.10b)
+[δ θi(k+m)−δ θi(k)]BiVix(k)V̂ixB(k)sin(δ θi(k))

The predictive model of the three-phase reactive power, described previously in (5.4), is pre-
sented in model (5.11).

Qi (k + m) = ∑ Qix (k + m) x={a, b, c} (5.11)

x

As the max operator for (5.5) cannot be included directly in the DMPC formulation, all the pos-
sible cases of the operator are included through a set of linear inequality constraints (see equations
(5.12a), (5.12b), (5.12c)); thus, the PVUR at the DG’s output is regulated in the solution to comply
with standard IEEE 1547-2018 [40]. Note that although the coefficients produced in the lineari-
sation are updated each sample time, they are constant during the optimisation and not computed
within the controller.

Kiaa (k)[Via (k+m)−Via (k)]+Kiab (k)[Vib (k+m)−Vib (k)]

(5.12a)
+Kiac (k)[Vic (k+m)−Vic (k)]+Fia (k)≤PVUR∗ (k)

Kiba (k)[Via (k+m)−Via (k)]+Kibb (k)[Vib (k+m)−Vib (k)]

(5.12b)
+Kibc (k)[Vic (k+m)−Vic (k)]+Fib (k)≤PVUR∗ (k)

Kica (k)[Via (k+m)−Via (k)]+Kicb (k)[Vib (k+m)−Vib (k)]

(5.12c)
+Kicc (k)[Vic (k+m)−Vic (k)]+Fic (k)≤PVUR∗ (k)

The coefficients of (5.12a) at time k produced in the linearisation process are expressed in (5.13).
The coefficients of (5.12b) and (5.12c) have a similar structure to the coefficients of (5.12a), and
are detailed in Annexed D.

 
3Via (k)−[Via (k)+Vib (k)+Vic (k)]
Fia (k)= (5.13a)
[Via (k)+Vib (k)+Vic (k)]

 
[3][Vib (k)+Vic (k)]
Kiaa (k)= (5.13b)
[Via (k)+Vib (k)+Vic (k)]2

92
  
[−3Via (k)]
Kiab (k)= (5.13c)
[Via (k)+Vib (k)+Vic (k)]2

 
[−3Via (k)]
Kiac (k)= (5.13d)
[Via (k)+Vib (k)+Vic (k)]2

The following operational constraints are also included in the optimisation. The models (5.14a)
and (5.14b) compute local approximations of average frequency and MG average voltage based
on the predictions of frequency and voltage at the connection bus of DGi and the predictions of
frequency and voltage at their respective connection buses of the communicating neighbouring
DGs, respectively. These models consider the communication terms ai j (k) and the estimated time
delay τ̂i j , i.e. only the information received through the communication network is used.

N
ωi (k+m)+∑ai j (k)ω j (k+m−τ̂i j )
j=1
ω i (k+m)= N
(5.14a)
1+∑ai j (k)
j=1

N
Vi (k+m)+∑ai j (k)V j (k+m−τ̂i j )
j=1
V i (k+m)= N
(5.14b)
1+∑ai j (k)
j=1

The soft constraints (5.15a) and (5.15b) work in conjunction with the previous defined models
and the auxiliary variables ωaux,i , Vaux,i of the objective function (5.6) to achieve the objective to keep
both average voltage and average frequency within predefined bands while an unfeasible solution
is avoided.

ω min ≤ ω i (k + m) + ωaux,i (k + m) ≤ ω max (5.15a)

V min ≤ V i (k + m) +Vaux,i (k + m) ≤ V max (5.15b)

As these auxiliary variables are penalised in the cost function, see (5.6), they will temporally
relax the average frequency and average voltage inequality constraints, allowing these variables to
take values outside their predefined band for a short time. The optimisation problem is relaxed
by applying these constraints, and a feasible solution is guaranteed [47], as long as the demanded

93
power is within the physical capacity of the MG.

The maximum apparent power capacity of DGi is also included through the triangular linearised
constraint (5.16) to limit the solution within the physical capacity of DGi (see Annexed D for the
detailed procedure to obtain this model).

|Pi (k)| + |Qi (k)| + sign (Pi (k)) [Pi (k + m) − Pi (k)]

(5.16)
+ sign (Qi (k)) [Qi (k + m) − Qi (k)] ≤ Smax

The proposed DMPC controller has a quadratic cost function, linear equality constraints and lin-
ear inequality constraints. Therefore, the optimisation problem is convex and can be synthesised in
a canonical quadratic programming (QP) formulation. The optimisation vector of the QP problem,
Xi in (5.17), comprises the predicted variables X p,i and the control decisions X∆,i .

Xi ={X p,i , X∆,i } (5.17)

The predicted variables are presented in (5.18) with x={a, b, c}. These prediction variables are
shared through the communication network to achieve the consensus objectives.

X p,i = [ωi (k + m), δ θi (k + m),Vi (k + m),Vix (k + m),

ωaux,i (k + m),Vaux,i (k + m), Pi (k + m), (5.18)
N
y
Qi (k + m), Qix (k + m), ω̄i (k + m), V̄i (k + m)]m=1

The predicted control decisions are presented in (5.19) with x={a, b, c}. Then, only the first
control decisions of ∆Vs,ix for phases x={a, b, c} and ∆ωs,i pass through integrators and are applied
to the system (see Fig. 5.1), and the optimisation problem is repeated at each sample time with
updated measures (rolling horizon) [45].

X∆,i = [∆Vs,ix (k + m − 1), ∆ωs,i (k + m − 1)]Nu

m=1 (5.19)

The QP problem is solved using the QPKWIK Matlab built-in algorithm, which is a stable
variation of the classic active-set method [136]. Moreover, the methodology to solve the DMPC
scheme is similar to the one described in Algorithm 1. This algorithm details all the necessary steps
to obtain a cooperative solution among the DGs that form the MG. Given that the cost function
represented in (5.6) is convex and QP is used to solve the optimisation problem, the controller will
find the optimum of the objective function at each sample time [47].

94
5.5 Microgrid setup and simulation results
The MG simulator implemented to test the performance of the DMPC scheme is shown in Fig. 5.3,
and Table 5.1 presents its electrical parameters.

DG1 DG3

DG2 DG4
Adjacency matrix

Communication
link

Figure 5.3: Implemented MG simulator

It comprises four power electronic DG units with different power ratings, different coupling
inductances and transmission lines with different impedances. The DGs have the following ap-
parent power capacities: S1max = 12.5[KVA] for converter 1, S2max = 0.9S1max for converter 2,
S3max = 0.8S1max for converter 3, and S4max = 0.7S1max for converter 4. Based on this config-
uration, the following droop coefficients are used for implementing the P − ω and Q − V droop
controllers and the predictive controllers: Mpω,1 = 1.6 · 10−4 rad/sW and Mqv,1 = 3.2 · 10−3V /VAR,
Mpω,2 = 0.9Mpω,1 and Mqv,2 = 0.9Mqv,1 , Mpω,3 = 0.8Mpω,1 and Mqv,3 = 0.8Mqv,1 , Mpω,4 = 0.7Mpω,1 and
Mqv,4 = 0.7Mqv,1 . The coupling inductances are L1 = 2.5[mH], L2 = 1.1L1 , L3 = 1.2L1 , L4 = 1.3L1
for DG1 to DG4 , respectively. Transmission lines are L34 = 2.5[mH], L12 = 1.1L34 , L24 = 1.2L34 .

The MG electrical model is built with the PLECS blockset®, whereas the primary and secondary
controllers are implemented in Matlab/Simulink® environment. Each DG unit possesses at the
primary level a three-phase active power - frequency droop controller, per phase reactive power -
voltage droop controllers, and self-tuning voltage and current PR controllers in the abc reference
frame [26] (see the bottom of Fig. 5.1). The following simplifications were considered in the

95
 Table 5.1: MG parameters and loads
Description Parameter Value

Primary level sample period Tprim [s] 1/(16 · 103 )

Nominal frequency ω0 [rad/s] 2π · 50

Nominal voltage (peak) V0 [V] 220

Droop controller cutoff frequency ωc [rad/s] 10π

Positive sequence current [A] 69.23

Unbalanced load 1 (Z1)
Negative sequence current [A] 23.76

Positive sequence current [A] 27.81

Unbalanced load 2 (Z2)
Negative sequence current [A] 1.58

Positive sequence current [A] 26.8

Unbalanced load 3 (Z3)
Negative sequence current [A] 1.05

simulator since their high-bandwidth dynamics are much faster than that of the studied controllers,
and they are not relevant on the time-scale of the proposed controllers:

• The DGs are simulated as regulable voltage sources.

• The modulation techniques (PWM, SVM) of the converters are not considered.

• The switching of the switching devices is not considered.

Table 5.2 presents the DMPC design parameters and the weighting factors. The parameters
were chosen aiming to reduce the computational effort. This is because the computational burden
is directly affected by the sample time, and prediction and control horizons [47]. The sample
time was selected considering the frequency and active power open loop rise time (Tr = 0.7s) as
Tsec = 0.7/14 = 0.05s [137]. The prediction and control horizons were selected as 10 samples
because with these values the controller always finds a solution within the sample time, and the
traffic over the communication network is reduced. The weighting factors were tuned following
the guidelines in [88], i.e., looking for a trade-off between the control objectives and, if needed
giving more importance to one objective over the rest of the objectives. The frequency and average
voltage are limited to a band of 1% and 5% with respect to their nominal values (ω0 and V0 ),
respectively, as recommended in [40]. The PVUR limit (PVUR∗ in equations (5.12a), (5.12b),
(5.12c)) was selected as 4% to meet the converter’s output voltage quality standard (below 5%)
[40]. These limits are fixed for all the test-scenarios; however, these can be modified, as long as

96
 Table 5.2: Controller parameters and weights
Description Parameter Value

Controller sample time Tsec [s] 0.05

Estimated communication delay τ̂i j [s] 0.05

Prediction horizon Ny 10

Control horizon Nu 10

Average voltage predefined band [Vmax ,Vmin ] [V ] [231,209]

Frequency predefined band [ωmax , ωmin ] [rad/s] [101π, 99π]

Maximum PVUR limit PVUR∗ [%] 4

VA 2
Phase a - Reactive power consensus λ1i [( VAR ) ] 9.5
VA 2
Phase b - Reactive power consensus λ2i [( VAR ) ] 8.0
VA 2
Phase c - Reactive power consensus λ3i [( VAR ) ] 1.7

Active power consensus λ4i [( VA 2

W ) ] 2.1
s 2
Average frequency regulation λ5i [( rad ) ] 3.8 · 102

Average voltage regulation λ6i [( V1 )2 ] 5.0

Per phase voltage control actions λ7i [( V1 )2 ] 3.0

s 2
Frequency control action λ8i [( rad ) ] 5.0 · 101

97
they are within the recommendations of the IEEE standard.

Three case-scenarios test the performance of the DMPC scheme. The first scenario consists
of connecting unbalanced loads at different nodes. The second scenario tests the controller when
there are both a short and a large constant time-delay over the communication network. Finally,
the last scenario tests the controller when two failures occur in conjunction, i.e. communication
link failures and a DG is disconnected/reconnected from/to the MG. The communication network
for the first two tests is represented by the adjacency matrix A(k) (see Fig. 5.3), which remains
constant for the whole test. Only in the last scenario is the adjacency matrix changed according to
the events of the test. These three test scenarios were selected because these are the most common
phenomena that a controller at the secondary level confronts [49, 134]. A distributed controller
must perform well against communication issues, such as communication delays and failures. In
addition, the possibility of disconnecting and reconnecting DGs to the MG without changes in the
programming of the controllers is essential.

5.5.1 Scenario I (base case) - Unbalanced load changes

This test verifies the performance of the DMPC on the MG when there are several unbalanced load
impacts at different nodes. The MG starts with Z1 connected and the primary control enabled, i.e.
droop controllers, and PR controllers enabled. Note that without the DMPC both per phase nor-
malised reactive power and three-phase normalised active power are shared unevenly (see Fig. 5.4
and Fig. 5.5 before 10 s). This is because the DGs have different power ratings and different droop
slopes. At t = 10s, the predictive controllers are enabled, so the power consensus objectives are
achieved in less than 7 seconds.

As discussed in the introduction section, to share the imbalance, it is necessary to induce small
imbalances in the output voltage of the DGs. Therefore, the maximum unbalanced voltage allowed
in the MG must be regulated to avoid power quality issues. This regulation is achieved by the
inequality constraints (5.12a), (5.12b), (5.12c), which limit the maximum allowed PVUR in the
voltage at DG’s output. In this test, the maximum allowed PVUR in each DG is set to 4%. Fig. 5.6
shows that the closest DG to the load impact reaches the PVUR limit; however, it is never surpassed.
At t = 30s, both Z2 and Z3 are connected; thus, the MG is subject to its total load.

This event takes the average frequency outside its band, and the DMPC restores this variable
inside the band immediately, as shown in Fig. 5.7a at t = 30s. This approach makes flexible the
behaviour of the frequency and average voltage by restoring them only when it is strictly necessary,
instead of restoring these variables to their nominal values at each sample time, as reported in
previous approaches. Finally, at t = 50s, Z3 is disconnected. During all the load perturbations the
controller presents a smooth response and all the objectives are achieved without large overshoots

98
 DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

Figure 5.4: Base Case a) Normalised reactive power consensus - Phase a for load changes, b)
Normalised reactive power consensus - Phase b for load changes, c) Normalised reactive power
consensus - Phase c for load changes

99
 DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

Figure 5.5: a) Three phase normalised active power consensus for load changes - Base Case, b)
Three-phase normalised reactive power consensus for load changes - Base Case

100
 DG1 DG2 DG3 DG4

Figure 5.6: PVUR index of the voltage at the DGs output for load changes - Base Case.

and with settling times below 7 seconds. Furthermore, none of the constraints are violated.

Note that regardless of the approach used (regulate to fixed values or regulate to a band), tem-
poral violations will always occur. These are due to external physical events, such as connec-
tion/disconnection of loads or connection/disconnection of generation units, and are not related to
the control system. In this sense, our proposed DMPC control system ensures quick recovery from
those temporary violations following the guidelines established by the IEEE 1547-2018 standard
[40].

101
 F Average

VAverage

Figure 5.7: a) Frequency regulation for load changes - Base Case, b) Average voltage regulation
for load changes - Base Case. The dashed cyan lines represent the predefined band limits for both
variables.

102
5.5.2 Scenario II - Communication delays
This scenario verifies the performance of the controllers when there is a constant delay (τi j ) over the
entire communication network, whilst the estimated delay (τ̂i j ) is kept constant at one sample. Two
cases are considered: a) small time-delay (τi j =0.25s) and b) large time-delay (τi j = 1s). Note that
the worst-case scenario represents a 20-sample delay, which is two times larger than the prediction
horizon (Ny ). A delay of one second is considered to be a large delay (see [16, 17]); such a delay
may be due to weather conditions or line of sight requirements in rural/remote areas [134]. For this
test, the same load changes considered in scenario I are applied.

DG1 DG2 DG3 DG4 DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

Figure 5.8: Communication delay test: a) Normalised reactive power consensus - Phase a for τi j =
0.25s, b) Normalised reactive power consensus - Phase a for τi j = 1s, c) Three-phase normalised
active power consensus for τi j = 0.25s, d) Three-phase normalised active power consensus for
τi j = 1s, e) Frequency regulation for τi j = 0.25s and τi j = 1s, f) Average voltage regulation for
τi j = 0.25s and τi j = 1s. The dashed cyan lines represent the predefined band limits for both latter
variables.

Fig. 5.8 shows the test results. Fig. 5.8a shows the reactive power in phase a for a delay of 0.25s,
while Fig. 5.8b shows the same information for a delay of 1s. The rest of the phases are omitted
as they present the same behaviour. Fig. 5.8c and Fig. 5.8d present the results for the active power
sharing for a delay of 0.25s and 1s, respectively. Fig. 5.8e and Fig. 5.8f present the results for
the average frequency regulation and average voltage regulation for both delays. The results show

103
that the DMPC is robust against communication delays, and the delay affects the overshoot and the
settling time of the consensus variables: the single phase reactive power is slightly affected when
two unbalanced loads are connected simultaneously at different points in the MG at t = 30s, and
the active power consensus is the variable most affected by delays.

The more the delay the larger the overshoot and the settling time. It is observed that although
the frequency and average voltage are taken outside the defined limits, the proposed controller is
capable of regulating these variables within their predefined bands. This is due to the inequality
constraints (5.15a) and (5.15b) that allow temporary violations. Therefore, the DMPC presents a
good performance against communication delays even when the delay is two times the prediction/-
control horizons. This is because the rolling horizon scheme compensates the delay in the shared
information by correcting the control action sequences [46].

5.5.3 Scenario III - Combined communication link failures and plug-and-

Play

This test presents the behaviour of the predictive controllers when two of the most demanding
scenarios are present in the MG at the same time. The test starts with the controllers enabled and
Z1 connected. A communication failure is forced at t = 10s between the communication channels
of DG1-DG3 and DG2-DG4, so the MG continues operating with four communication channels,
and the MG adjacency matrix is modified (see A(k) in Fig. 5.9a at t = 10s).

The control algorithm automatically identifies the failure and and changes the consensus (5.6)
and average (5.14a), (5.14b) calculations . At t = 30s an unscheduled failure occurs, i.e. DG4 is
disconnected from both the electrical system and the communication network. The MG continues
operating with the remaining connected DGs until t = 50s, when Z2 is connected. At t = 70s, after a
synchronisation routine, DG4 is reconnected to the MG. Note that when DG4 is disconnected there
is only one communication path among the remaining DGs, which is the worst communication
scenario for distributed controllers [132] (see A(k) in Fig. 5.9a at t = 30s).

Fig. 5.9 and Fig. 5.10 show the test results. When the communication failure occurs (t = 10s),
the predictive controllers do not suffer noticeable deterioration; only the frequency and voltage
present a slight variation, as shown in Fig. 5.9. Furthermore, the controller performance is not
affected by both phenomena, and the remaining operating DGs achieve the consensus objectives,
even when both failures are present at the same time. Nevertheless, the transient response is differ-
ent, both overshoot and settling time are increased slightly. This is because the adjacency matrix is
not complete, and the consensus objectives depend on the known information of the neighbouring
DGs.

104
 F Average

VAverage

Figure 5.9: Communication failure and plug-and-play test. a) Frequency regulation, b) Average
voltage regulation. The dashed cyan lines represent the predefined band limits for both variables.

105
 DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

Figure 5.10: Communication failure and plug-and-play test. a) Normalised reactive power consen-
sus - Phase a, b) Normalised reactive power consensus - Phase b, c) Normalised reactive power
consensus - Phase c

106
 Note that when DG4 is reconnected (see Fig. 5.10 at t = 70s), this DG unit achieves the consen-
sus objectives in a higher time. The time response of the proposed DMPC scheme depends on the
density of the adjacency matrix A. This implies that the dynamic response is slow when A is sparse
(most of the elements of A are zero, meaning few communication channels), whereas the dynamic
response is fast when the density of A is high (most of the elements of A are one, meaning more
communication channels). Furthermore, it has been reported that the convergence time is not re-
lated directly to the number of DGs that form the MG, and the convergence time will not be affected
as long as the new DGs are properly communicated [42]. Indeed, if more well-communicated DGs
are added, the converge speed will be improved [42].

5.6 Hardware in the loop validation

The proposed DMPC scheme has been implemented and validated via hardware in the loop (HIL)
to illustrate the physical simulation’s fidelity. For this purpose, the real-time (RT) platform OP4510
OPAL-RT power grid digital simulator was used. This widely used FPGA-based platform allows
HIL validation, as it assigns an independent processor’s core for control tasks and another core for
system simulation tasks. The OPAL-RT platform is shown in Fig. 5.11. The MG simulator with
four DGs of Fig. 5.3 was implemented in the OPAL platform. The DGs implemented in the OPAL
include the DMPC controllers along with the primary and droop controllers (see Section 5.2 for a
detailed explanation).

HIL platform

Scope

Figure 5.11: OPAL-RT platform for HIL validation

To corroborate the results obtained in the simulation section, the test presented in Section Sec-
tion 5.5.3 was selected for the HIL validation because this is the most demanding test, as it com-
bines the failure of two communication links and the disconnection and reconnection of a DG (to
see the events that occur in this test see Section Section 5.5.3).

107
 DG DG DG DG
1 2 3 4

DG1 DG2 DG3 DG4

DG1 DG2 DG3 DG4

Figure 5.12: HIL - Communication failure and plug-and-play test. a) Normalised reactive power
consensus - Phase a, b) Normalised reactive power consensus - Phase b, c) Normalised reactive
power consensus - Phase c

108
Figure 5.13: HIL - Communication failure and plug-and-play test. a) Current and voltage at the
DG4 connection point, b) Current at the DG3 connection point, Yellow: Ia, Green: Ib, Blue:Ic,
Pink:Va - (20 A/Div, 100 V/Div ).

109
 The HIL results are consistent with the simulations. For example, Fig. 5.12 shows the single-
phase reactive powers in the OPAL platform. When comparing this figure with the one obtained by
simulations (Fig. 5.10), it can be seen that they are similar.

Furthermore, this section presents waveforms that cannot be obtained from simulations. For
instance, Fig. 5.13 shows the single-phase currents in the natural reference frame at the connection
points of DG4 and DG3 (see Fig. 5.3) when DG4 is reconnected at around t = 70s. Fig. 5.13a
shows the single-phase currents of DG4 during the reconnection of DG4, together with the voltage
on one phase. When the DMPC on DG4 is enabled, a good dynamic response without overshoots
is appreciated. On the other hand, Fig. 5.13b shows the current injected by DG3 during the recon-
nection of DG4. Note that in this image, after reconnecting DG4, the current of DG3 decreases
since DG4 takes part of the MG load.

110
5.7 Scalability and comparison with a DAPI-based controller
In this section, the scalability of the proposed DMPC and a comparison study with the reported
technique in [39] (based on the widely used virtual impedance method) are provided. To test both
scenarios a MG simulator with eight DGs was implemented (see Fig. 5.14), as four new generators
were added, i.e. DG5 , DG6 , DG7 and DG8 .

DG1 DG5

DG2 DG6

DG3 DG7

DG4 DG8

Communication
link

Figure 5.14: Implemented MG simulator for scalability and comparison scenarios

These generators have the same droop slopes, power capacities, and coupling inductors of DG1 ,
DG2 , DG3 and DG4 , respectively. Furthermore, the same controller and weighting parameters pre-

111
sented in Table 5.2 are configured in the new generators, and a reduced number of communication
links is used for the information sharing (see the blue dashed lines in Fig. 5.14). In the same fash-
ion, the transmission lines were duplicated (see Section 5.5). Note that loads Z5 and Z6 were added
with the same values of Z1 and Z2 , respectively (see Table 5.1).

5.7.1 Scalability

Scalability is crucial for a distributed control strategy, and its dynamic behaviour is directly affected
by the communication topology of its DGs (agents) and not necessarily related to the number of
DGs [42]. Moreover, its behaviour can be analysed through the eigenvalues of the Laplacian matrix
L, which is defined as L = D − A. Where A is the adjacency matrix (defined in Section 3.4.1), and D
is a diagonal matrix formed by the sum of the elements in each row of the adjacency matrix A, i.e.,
D = diag∑Nj=1 ai j . The Laplacian matrix L is symmetric for undirected graphs, and its eigenvalues
are nonnegative real [42]. In particular, the control strategy convergence speed depends on the
Laplacian second eigenvalue, which is known as the Fiedler Eigenvalue [42].

The following test evaluates the proposal’s scalability. The test starts with two loads connected
and five DGs operating with their DMPCs enabled. Then, DG8 , DG7 and DG6 are connected (after
a synchronisation routine) to the microgrid at around t = 10s, t = 40s and t = 70s, respectively.
The remaining two loads are added at t = 100s. In order for the control algorithm to work an
initial configuration is needed when a new DG is introduced for the first time to the MG, as is the
case of the majority of distributed consensus techniques. All operating DGs need to know their
neighbouring power capacities (S jmax ), and the number of DGs that form the MG (N).

This does not compromise the Plug-and-Play capability of our proposal, as DGs can be discon-
nected or reconnected at any time. Then, the DMPC algorithm automatically updates the consensus
terms on (5.6) and (5.14) in all operating DGs to solve the optimisation problem. Note that thanks
to the distributed structure of the controller, the number of predicted variables is fixed. Therefore,
the computational burden does not increase when new DGs are introduced into the MG. This is of
high importance for the scalability of control techniques at the secondary control level.

The performance of the controller is depicted in Fig. 5.15 and Fig. 5.16, where it is observed
that due to the reduced density of the adjacency matrix A, the new DGs achieve the consensus
objectives in a higher settling time (around 5 seconds) with a smooth transitory response, and the
average voltage and frequency are kept within the predefined operating band in all the events in this
test.

The optimisation time for the DMPC scheme is presented in Fig. 5.17. It is observed that all
predictive controllers find a solution at around 0.01 seconds, which is well below the sample time

112
 DG1 DG2 DG3 DG4 DG5 DG6 DG7 DG8

Figure 5.15: Scalability plug-and-play test. a) Normalised reactive power consensus - Phase a, b)
Normalised reactive power consensus - Phase b, c) Normalised reactive power consensus - Phase c

113
 F Average

VAverage

Figure 5.16: Scalability plug-and-play test. a) Frequency regulation, b) Average voltage regulation.
The dashed cyan lines represent the predefined band limits for both variables.

114
(0.05 seconds). It should be noted that due to the distributed structure of the predictive scheme,
the number of optimisation variables is fixed (see (5.18) and (5.19)). Fig. 5.17 shows that the time
required to obtain a solution does not increase when DG8 , DG7 , and DG6 are connected to the MG
at t = 10s, t = 50s and t = 70s, respectively. These tests were performed on a 9th generation Intel
Core i7 3.6GHz computer with 32GB of RAM.

DG1 DG2 DG3 DG4 DG5 DG6 DG7 DG8

Figure 5.17: Optimisation time for the scalability test

5.7.2 Comparison with a distributed consensus-based controller for imbal-

ance sharing
A comparison between our proposal and the work of [39] is presented. This comparison is suitable
because both control techniques include consensus objectives to improve imbalance sharing in a
distributed fashion, regulate the PVUR of the DGs’ output voltage and use an adjacency matrix
to represent their communication topology. The work of [39] is based on the concept of virtual
impedance and, the imbalance sharing is achieved via a consensus in the three-phase unbalanced
power defined by the Conservative Power Theory (CPT) [26]. In the following tests, both control
strategies have a PVUR limit of 4%.

The behaviour of the unbalanced power sharing and PVUR at the DGs’ output voltage for [39]
in the presence of load changes are presented on the left side of Fig. 5.18.

It is observed that although this technique improves the sharing of unbalanced power, a high
PVUR at the DGs output voltage is present, which is outside of the desired limit. This is because the
DG units that are far from where the loads are connected, do not increase their PVUR. Furthermore,
the single-phase reactive powers’ behaviour for [39] is presented on the left side of Fig. 5.19. Phase
c is omitted as it presents the same behaviour. These results demonstrate that methods based on
virtual impedance and defined in the sequence components domain, where the consensus is defined

115
only considering magnitudes and not sequence phase angles, do not guarantee good sharing in
the phases (i.e., phase a to phase c). On the other hand, the proposed strategy’s performance
is presented on the right side of Fig. 5.18 and Fig. 5.19. It is observed that this technique has
better performance for unbalanced power sharing with a reduced PVUR that is always below the
established limit (4%). Furthermore, the reactive power in the phases is shared properly.
DG1 DG2 DG3 DG4 DG5 DG6 DG7 DG8

Figure 5.18: Comparison between the proposed DMPC scheme and a DAPI-based method. a)-b)
Unbalanced Power for the two methods compared, c)-d) PVUR index of the voltage at the DGs
output for the two methods compared.

Another advantage of the proposed technique is its resilience under communication delays. This
is verified by applying a constant delay of one second (τi j = 1s) on the entire communication
network and testing the performance of both strategies. The results for [39] are presented on the
left side of Fig. 5.20, whilst the results of our proposal are depicted on the right side. Phase c is
omitted as it presents the same behaviour. It is observed that [39] is highly affected under large
delays by presenting oscillations in its behaviour when the MG load condition changes. Whereas,
the proposed DMPC is slightly affected in the transitory response; nevertheless, the consensus
objectives are achieved regardless of the delay. This is because the rolling horizon property and
the delay estimation of the DMPC scheme correct the control actions sequence [46], while the
consensus technique of [39] does not posses a delay compensation property.

116
 DG1 DG2 DG3 DG4 DG5 DG6 DG7 DG8

Figure 5.19: Comparison between the proposed DMPC scheme and a DAPI-based method. a)-b)
Normalised reactive power in the phase a for the two methods compared, c)-d) Normalised reactive
power in the phase b for the two methods compared.

DG1 DG2 DG3 DG4 DG5 DG6 DG7 DG8

Figure 5.20: Comparison between the proposed DMPC scheme and a DAPI-based method for
τi j = 1s. a)-b) Normalised reactive power in the phase a for the two methods compared, c)-d)
Normalised reactive power in the phase b for the two methods compared.

117
5.8 Discussion
This Chapter presented a novel distributed predictive control strategy to cope with per-phase power
imbalance sharing and active power sharing in ac isolated MGs. The proposed DMPC scheme
is able to achieve all the consensus control objectives simultaneously, while the imposed physical
constraints are respected. The dynamic performance of the controller was evaluated and discussed
under three of the most demanding test scenarios. Simulation and HIL results verify the good
performance of the rolling horizon scheme for communication network issues, PVUR limit restric-
tions and the disconnection and reconnection of DGs. Furthermore, the proposal’s scalability and a
comparison study with the usual consensus technique based on the virtual impedance method were
evaluated. In summary, the results demonstrated the main advantages of the proposed strategy,
which are:

• Per-phase reactive power is shared proportionally among DGs when a load is connected to
the MG and when a DG is disconnected from and reconnected to the H-MG.

• The frequency and voltage are restored within secure bands in steady-state under all the
analysed cases.

• The DMPC scheme operates appropriately with a reduced communication network, where
DGs are communicated only with neighbouring DGs. Moreover, global objectives are achieved
through information sharing.

• DGs can easily be connected to or disconnected from the MG, and the control structure,
demonstrating the plug-and-play capability of the proposed strategy.

• The superior behaviour of the DMPC scheme against usual DAPI-based controllers is demon-
strated under communication issues.

This strategy was validated through hardware-in-the-loop (HIL) and simulation tests in the jour-
nal paper:

A. Navas-Fonseca, C. Burgos, J. S. Gómez, F. Donoso, L. Tarisciotti, D. Sáez, R. Cárdenas D.,

and M. Sumner, “Distributed Predictive Control for Imbalance Sharing in ac Microgrids,” in IEEE
Transactions on Smart Grid, vol. 13, no. 1, pp. 20-37, Jan. 2022, doi:
10.1109/TSG.2021.3108677, [Q1-IF 8.960 - published ].

118
Chapter 6

Conclusions and final remarks

Microgrids are positioning themselves as the key enablers for integrating distributed generation and
energy storage systems. Nevertheless, to harness their benefits, new and more sophisticated control
systems must be developed. Distributed model predictive control (DMPC) is one of the leading
approaches to managing microgrids (MGs). In this context, the main objective of this PhD thesis
was to propose distributed predictive cooperative strategies for secondary control in MGs. These
strategies combined the benefits of distributed control and model predictive control, giving a better
overall performance of the MG.

Based on the results, it can be stated that DMPC provides the flexibility to achieve the main task
of the secondary control level (restoring frequency and voltage) along with complementary objec-
tives such as minimisation of operational cost, or imbalance sharing, within the same predictive
controller formulation. The proposed DMPC schemes include detailed models of the dynamics of
the generators (and interlinking converters) including equality constraints and physical limits as in-
equality constraints. In addition, the rolling horizon property of the proposed DMPCs compensates
for communication delays. These characteristics allow the proposed DMPC schemes to perform
better than traditional controllers. The main findings of this PhD thesis are highlighted as:

For balanced ac MGs

• It is demonstrated that DMPC can minimise the MG’s operating cost and restore the fre-
quency to its nominal value over a similar time scale. The response of the controller is
enhanced by including power rating limits, local dynamic models of the DGs, and a model
of the communication network model. In addition, the controller does not need knowledge
of the MG topology.

• Extensive experimental tests validate the performance of the proposed control scheme against

119
 sudden load changes, communication delays, communication failures, as well as disconnec-
tion and reconnection of DGs.

For hybrid ac/dc MGs

• It was shown that the frequency, ac voltage and dc voltage on hybrid ac/dc MGs can be re-
stored to within secure bands. This strategy gives more flexibility to the operation of MGs
where frequency and voltages are restored only when they are outside predefined bands in-
stead of being restored to specific nominal values at each sample time, as is the case for the
vast majority of approaches proposed in the literature.

• The proposed control strategy regulates the frequency of all ac DGs (a global variable). By
contrast, since the voltage is not a global variable, the proposed control scheme regulates the
average ac voltage and dc voltage, allowing different voltages in the voltage buses of the MG
and employing accurate reactive power control.

• Since the proposed strategy is implemented in each DG or ILC, the computational burden
required for the predictive controllers is reduced and does not increase when more DGs or
ILCs are added to the MG. This is because the number of optimisation variables is fixed.
In addition, consensus objectives are achieved through information sharing and coordination
between DGs and ILCs. The controllers operate with the usual measurements at the primary
level, and no additional measurements are required.

• Extensive simulation studies verify the performance of the proposed control scheme against
sudden ac and dc load changes, communication delays, communication failures, as well as
disconnection and reconnection of ac DGs, dc DGs and ILCs. A detailed comparison study
was carried out in terms of performance under communication delays.

For unbalanced ac MGs

• Phase imbalance sharing between DGs is controlled through the control of single-phase re-
active power. The proposed distributed model predictive control scheme does not require
virtual impedance loops or the inclusion of additional power converters for managing single-
phase reactive power between distributed generators. In fact, with the proposed technique,
the sharing of imbalance is performed directly in terms of single-phase reactive power and
without adding extra power converters into the microgrid.

• As part of sharing imbalance through the single-phase reactive power, the DMPC bounds the
unbalanced voltage at the DGs’ output, fulfilling the recommendation of the IEEE standard
1547-2018 [40]. Moreover, power rating limits are also considered in the formulation.

120
 • Extensive hardware-in-the-loop and simulation studies verify the performance of the pro-
posed control schemes against sudden load changes, communication delays, communication
failures, as well as disconnection and reconnection of DGs. Tests to show he controller’s
scalability and a detailed comparative performance study under communication delays were
also carried out.

In summary, it is verified that currently distributed predictive control is one of the most promis-
ing strategies for MG management, as they allow the achievement of global objectives through
information sharing. Moreover, they allow the pursuit of more flexible objectives using soft con-
straints. They also allow modelling and the use of limits in states and inputs within the formulation.
All the proposed distributed predictive strategies are easy to scale, and they do not increase the com-
putational burden when more agents are added; this is because the number of predicted variables is
fixed.

6.1 Future work

A few areas that could form the basis of future research are detailed in the following.

• A direct future path could be the design of a controller that combines economic dispatch and
imbalance sharing in hybrid ac/dc MGs. Also, energy storage systems and specific renewable
energy services should be considered in the formulation.

• The proposed control strategy could be extended to consider congestion management of the
distribution line currents. The losses on the ILCs and the losses on the distribution lines could
also be considered in the formulation.

• A current hot research topic is the study of the effects of cyber-attacks on distributed con-
trollers. A common cyber-attack is known as the false data injection attack (FDIA). FDIAs
can appear in sensors, communication links, and actuators. Therefore, a direct extension of
the proposed DMPC schemes would be to incorporate new detection methods and counteract
mechanisms to tackle cyber-attacks.

• Different tuning methodologies for the weighting parameters could be explored. For instance,
heuristic algorithms, such as particle swarm optimisation (PSO) or genetic algorithms (GA),
could be used.

• An open research topic is the study of theoretical stability on distributed model predictive
control. Concepts like terminal costs and terminal invariant sets could be incorporated into
the formulation.

121
6.2 Publications

6.2.1 Journal papers

[1] A. Navas-Fonseca, J. S. Gomez, J. Llanos, E. Rute, D. Saez, and M. Sumner, “Distributed

Predictive Control Strategy for Frequency Restoration of Microgrids Considering Optimal
Dispatch," in IEEE Transactions on Smart Grid, vol. 12, no. 4, pp. 2748-2759, July 2021,
doi: 10.1109/TSG.2021.3053092 (Q1, impact factor: 8.960).

[2] A. Navas-Fonseca, C. Burgos-Mellado, J. S. Gómez, F. Donoso, L. Tarisciotti, D. Sáez, R.

Cárdenas D., and M. Sumner, “Distributed Predictive Control for Imbalance Sharing in AC
Microgrids,” in IEEE Transactions on Smart Grid, vol. 13, no. 1, pp. 20-37, Jan. 2022, doi:
10.1109/TSG.2021.3108677 (Q1, impact factor: 8.960).

[3] A. Navas-Fonseca, C. Burgos-Mellado, J. S. Gómez, E. Espina, J. Llanos, D. Sáez, M.

Sumner, and D. E. Olivares, “Distributed predictive secondary control with soft constraints
for optimal dispatch in hybrid AC/DC microgrids," in IEEE Transactions on Smart Grid,
2022 (under review, Q1, impact factor: 8.960).

[4] E. Rute, A. Navas-Fonseca, J. S. Gómez, E. Espina, C. Burgos-Mellado, D. Sáez, M. Sumner

and Diego Muñoz-Carpintero, “Distributed Predictive Control for Secondary Level in Hybrid
ac/dc Microgrids,” in Journal of Emerging and Selected Topics in Power Electronics,
doi: 10.1109/JESTPE.2022.3157979 (Q1, impact factor: 4.472).

[5] M. Martinez-Gomez, A. Navas-Fonseca, M. E. Orchard, S. Bozhko, C. Burgos-Mellado and

R. Cardenas, “Multi-Objective Finite-Time Control for the Interlinking Converter on Hybrid
AC/DC Microgrids," in IEEE Access, vol. 9, pp. 116183-116193, 2021, doi:
10.1109/ACCESS.2021.3105649 (Q1, impact factor: 3.367).

6.2.2 Conference papers

[1] A. Navas-Fonseca, C. Burgos, E. Espina, E. Rute, D. Saez, and M. Sumner, “Distributed Pre-
dictive Secondary Control for Voltage Restoration and Economic Dispatch of Generation for
DC Microgrids," 2021 IEEE Fourth International Conference on DC Microgrids (ICDCM),
2021, pp. 1-6, doi: 10.1109/ICDCM50975.2021.9504612.

[2] A. Navas-Fonseca, C. Burgos-Mellado, J. Gomez, J. Llanos, E. Espina, D Saez, and M.

Sumner, “Distributed Predictive Control using Frequency and Voltage Soft Constraints in
AC Microgrids including Economic Dispatch of Generation,” 2021 47th Annual Conference
of the IEEE Industrial Electronics Society (IECON), Toronto, Canada, 2021.

122
[3] E. Espina, A. Navas-Fonseca, J. S. Gómez, R. Cárdenas D, “Experimental Performance
Evaluation of a Distributed Secondary Control Strategy for Hybrid ac/dc-Microgrids in the
event of Communication Loss/Delay”, 2021 IEEE, 23rd European Conference on Power
Electronics and Applications.

[4] R. Bustos, L. Marin, A. Navas-Fonseca, D. Saez, “Demand Side Management based on

Fuzzy Prediction Intervals for Microgrids," 2022 IEEE World Congress on Computational
Intelligence (WCCI), [accepted for publication].

[5] M. Martinez-Gomez, R. Cardenas D., A. Navas-Fonseca, and E. Rute, “A Multi-Objective

Distributed Finite-Time Optimal Dispatch of Hybrid Microgrids," IECON 2020 The 46th
Annual Conference of the IEEE Industrial Electronics Society, 2020, pp. 3755-3760, doi:
10.1109/IECON-43393.2020.9254375.

[6] E. Espina, C. Burgos-Mellado, J. S. Gómez, J. Llanos, E. Rute, A. Navas-Fonseca, M.

Martínez, R. Cárdenas, and D. Sáez, “Experimental Hybrid AC/DC-Microgrid Prototype for
Laboratory Research," 2020 22nd European Conference on Power Electronics and Applica-
tions (EPE’20 ECCE Europe), 2020, pp. 1-9, doi: 10.23919/EPE20ECCE-Europe43536.2020.9215751.

[7] T. Roje, A. Navas-Fonseca, M. Urrutia, P. Mendoza-Araya and G. Jiménez-Estévez, “Ad-

vanced lead-acid battery models for the state-of-charge estimation in an isolated microgrid,"
2019 IEEE CHILEAN Conference on Electrical, Electronics Engineering, Information and
Communication Technologies (CHILECON), 2019, pp. 1-6,
doi: 10.1109/CHILECON47746.2019.8987985.

123
BIBLIOGRAPHY

[1] “What were the outcomes of COP26?” [Online]. Available: https://commonslibrary.

parliament.uk/what-were-the-outcomes-of-cop26/

[2] “UK targets power from 100% renewable sources by

2035 | S&P Global Commodity Insights.” [Online]. Avail-
able: https://www.spglobal.com/commodityinsights/en/market-insights/latest-news/
energy-transition/100421-uk-targets-power-from-100-renewable-sources-by-2035

[3] “2022 comienza en Chile con la fotovoltaica generando el 20% de

la electricidad - ACERA - AG.” [Online]. Available: https://acera.cl/
2022-comienza-en-chile-con-la-fotovoltaica-generando-el-20-de-\la-electricidad/

[4] “86% de los MW que entrarán a Chile este año

son renovables.” [Online]. Available: https://www.revistaei.cl/2022/
05/05/un-86-de-los-mw-que-entraran-a-operar-en-chile-este-ano-provienen-de\
-energia-solar-o-eolica/

[5] Department of Energy office of Electricity Delivery and Energy Reliability, “Summary re-
port : 2012 DOE microgrid workshop,” U.S. Dep. Energy, Tech. Rep., Jul. 2012.

[6] D. E. Olivares, A. Mehrizi-Sani, A. H. Etemadi, C. A. Cañizares, R. Iravani, M. Kazerani,

A. H. Hajimiragha, O. Gomis-Bellmunt, M. Saeedifard, R. Palma-Behnke, G. A. Jiménez-
Estévez, and N. D. Hatziargyriou, “Trends in microgrid control,” IEEE Trans. Smart Grid,
vol. 5, no. 4, pp. 1905–1919, Jul. 2014.

[7] R. Lasseter, “Microgrids,” in 2002 IEEE Power Engineering Society Winter Meeting. Con-
ference Proceedings (Cat. No.02CH37309), vol. 1, 2002, pp. 305–308 vol.1.

[8] E. Espina, J. Llanos, C. Burgos-Mellado, R. Cárdenas-Dobson, M. Martínez-Gómez, and

D. Sáez, “Distributed control strategies for microgrids: An overview,” IEEE Access, vol. 8,
pp. 193 412–193 448, 2020.

124
 [9] K. Ahmed, M. Seyedmahmoudian, S. Mekhilef, N. M. Mubarak, and A. Stojcevski, “A re-
view on primary and secondary controls of inverter-interfaced microgrid,” Journal of Mod-
ern Power Systems and Clean Energy, vol. 9, no. 5, pp. 969–985, Sep. 2021.

[10] S. K. Sahoo, A. K. Sinha, and N. K. Kishore, “Control Techniques in AC, DC, and Hybrid
AC–DC Microgrid: A Review,” IEEE J. Emerg. Sel. Top. Power Electron., vol. 6, no. 2, pp.
738–759, Jun. 2018.

[11] T. Dragicevic, X. Lu, J. C. Vásquez, and J. M. Guerrero, “DC microgrids - Part I: A review of
control strategies and stabilization techniques,” IEEE Trans. Power Electron., vol. 31, no. 7,
pp. 4876–4891, Sep. 2016.

[12] F. Nejabatkhah and Y. W. Li, “Overview of Power Management Strategies of Hybrid AC/DC
Microgrid,” IEEE Trans. Power Electron., vol. 30, no. 12, pp. 7072–7089, Dec. 2015.

[13] C. D. Burgos Mellado, “Control strategies for improving power quality and PLL stability
evaluation in microgrids (PhD thesis),” Ph.D. dissertation, University of Nottingham, 2019.

[14] E. Nasr-Azadani, C. A. Canizares, D. E. Olivares, and K. Bhattacharya, “Stability analysis

of unbalanced distribution systems with synchronous machine and DFIG based distributed
generators,” IEEE Transactions on Smart Grid, vol. 5, no. 5, pp. 2326–2338, Sep. 2014.

[15] J. M. Guerrero, J. C. Vásquez, J. Matas, L. G. de Vicuña, and M. Castilla, “Hierarchical con-

trol of droop-controlled ac and dc microgrids–a general approach toward standardization,”
IEEE Trans. Ind. Electron., vol. 58, no. 1, pp. 158–172, Aug. 2011.

[16] J. Llanos, D. E. Olivares, J. W. Simpson-Porco, M. Kazerani, and D. Saez, “A novel dis-

tributed control strategy for optimal dispatch of isolated microgrids considering congestion,”
IEEE Trans. Smart Grid, vol. 10, no. 6, pp. 6595–6606, Nov. 2019.

[17] G. Chen and Z. Guo, “Distributed secondary and optimal active power sharing control for
islanded microgrids with communication delays,” IEEE Trans. Smart Grid, vol. 10, no. 2,
pp. 2002–2014, Mar. 2017.

[18] M. Zaery, P. Wang, W. Wang, and D. Xu, “Distributed Global Economical Load Sharing for
a Cluster of DC Microgrids,” IEEE Trans. Power Syst., vol. 35, no. 5, pp. 3410–3420, Sep.
2020.

[19] R. Babazadeh-Dizaji and M. Hamzeh, “Distributed Hierarchical Control for Optimal Power
Dispatch in Multiple DC Microgrids,” IEEE Syst. J., vol. 14, no. 1, pp. 1015–1023, Mar.
2020.

125
[20] A. Navas-Fonseca, J. S. Gomez, J. Llanos, E. Rute, D. Saez, and M. Sumner, “Distributed
Predictive Control Strategy for Frequency Restoration of Microgrids Considering Optimal
Dispatch,” IEEE Trans. Smart Grid, vol. 12, no. 4, pp. 2748–2759, Jul. 2021.

[21] W. Feng, J. Yang, Z. Liu, H. Wang, M. Su, and X. Zhang, “A unified distributed control
scheme on cost optimization for hybrid ac/dc microgrid,” in 2018 IEEE 4th South. Power
Electron. Conf. (SPEC), Dec. 2018, pp. 1–6.

[22] Z. Li, Z. Cheng, J. Si, and S. Li, “Distributed event-triggered hierarchical control to improve
economic operation of hybrid ac/dc microgrids,” IEEE Trans. Power Syst, pp. 1–1, 2021.

[23] C. Burgos-Mellado, J. Llanos, E. Espina, D. Saez, R. Cárdenas, M. Sumner, and A. Wat-

son, “Single-phase consensus-based control for regulating voltage and sharing unbalanced
currents in 3-wire isolated ac microgrids,” IEEE Access, vol. 8, pp. 164 882–164 898, Sep.
2020.

[24] A. S. Vijay, S. Doolla, and M. C. Chandorkar, “Unbalance mitigation strategies in micro-

grids,” IET Power Electron., vol. 13, no. 9, pp. 1687–1710, Jul. 2020.

[25] V. Gali, N. Gupta, and R. Gupta, “Mitigation of power quality problems using shunt ac-
tive power filters: A comprehensive review,” in 2017 12th IEEE Conference on Industrial
Electronics and Applications (ICIEA). IEEE, Jun. 2017, pp. 1100–1105.

[26] C. Burgos-Mellado, C. Hernández-Carimán, R. Cárdenas, D. Sáez, M. Sumner,

A. Costabeber, and H. K. Morales Paredes, “Experimental evaluation of a cpt-based four-leg
active power compensator for distributed generation,” IEEE Trans. Emerg. Sel. Topics Power
Electron., vol. 5, no. 2, pp. 747–759, Nov. 2017.

[27] A. Mortezaei, M. G. Simões, M. Savaghebi, J. M. Guerrero, and A. Al-Durra, “Cooperative

control of multi-master–slave islanded microgrid with power quality enhancement based on
conservative power theory,” IEEE transactions on smart grid, vol. 9, no. 4, pp. 2964–2975,
Jul. 2016.

[28] H. M. Munir, R. Ghannam, H. Li, T. Younas, N. A. Golilarz, M. Hassan, and A. Siddique,

“Control of distributed generators and direct harmonic voltage controlled active power fil-
ters for accurate current sharing and power quality improvement in islanded microgrids,”
Inventions, vol. 4, no. 2, p. 27, May 2019.

[29] D. I. Brandao, T. Caldognetto, F. P. Marafão, M. G. Simões, J. A. Pomilio, and P. Tenti,

“Centralized control of distributed single-phase inverters arbitrarily connected to three-phase
four-wire microgrids,” IEEE Transactions on Smart Grid, vol. 8, no. 1, pp. 437–446, Jan.

126
 2016.

[30] X. Zhou, F. Tang, P. C. Loh, X. Jin, and W. Cao, “Four-leg converters with improved com-
mon current sharing and selective voltage-quality enhancement for islanded microgrids,”
IEEE Trans. Power Deliv., vol. 31, no. 2, pp. 522–531, Jun. 2016.

[31] B. Liu, Z. Liu, J. Liu, R. An, H. Zheng, and Y. Shi, “An adaptive virtual impedance control
scheme based on small-ac-signal injection for unbalanced and harmonic power sharing in
islanded microgrids,” IEEE Transactions on Power Electronics, vol. 34, no. 12, pp. 12 333–
12 355, Dec. 2019.

[32] Y. Karimi, H. Oraee, and J. M. Guerrero, “Decentralized method for load sharing and power
management in a hybrid single/three-phase-islanded microgrid consisting of hybrid source
pv/battery units,” IEEE Transactions on Power Electronics, vol. 32, no. 8, pp. 6135–6144,
Aug. 2016.

[33] C. Burgos-Mellado, R. Cardenas, D. Saez, A. Costabeber, and M. Sumner, “A control al-

gorithm based on the conservative power theory for cooperative sharing of imbalances in
four-wire systems,” IEEE Trans. Power Electron., vol. 34, no. 6, pp. 5325–5339, Sep. 2019.

[34] L. Meng, F. Tang, M. Savaghebi, J. C. Vásquez, and J. M. Guerrero, “Tertiary control of

voltage unbalance compensation for optimal power quality in islanded microgrids,” IEEE
Trans. on Energy Conversion, vol. 29, no. 4, pp. 802–815, Jul. 2014.

[35] M. Savaghebi, A. Jalilian, J. C. Vásquez, and J. M. Guerrero, “Secondary control for voltage
quality enhancement in microgrids,” IEEE Trans. Smart Grid, vol. 3, no. 4, pp. 1893–1902,
Jul. 2012.

[36] D. I. Brandao, L. S. Araujo, A. M. S. Alonso, G. L. dos Reis, E. V. Liberado, and F. P.

Marafão, “Coordinated control of distributed three-and single-phase inverters connected to
three-phase three-wire microgrids,” IEEE Trans. Emerg. Sel. Topics Power Electron., pp.
1–1, 2019.

[37] L. Meng, X. Zhao, F. Tang, M. Savaghebi, T. Dragicevic, J. C. Vásquez, and J. M. Guerrero,

“Distributed voltage unbalance compensation in islanded microgrids by using a dynamic
consensus algorithm,” IEEE Trans. Power Electron., vol. 31, no. 1, pp. 827–838, Jan. 2016.

[38] J. Zhou, S. Kim, H. Zhang, Q. Sun, and R. Han, “Consensus-based distributed control for
accurate reactive, harmonic, and imbalance power sharing in microgrids,” IEEE Trans. Smart
Grid, vol. 9, no. 4, pp. 2453–2467, Jul. 2018.

127
[39] C. Burgos-Mellado, J. J. Llanos, R. Cardenas, D. Saez, D. E. Olivares, M. Sumner, and
A. Costabeber, “Distributed control strategy based on a consensus algorithm and on the
conservative power theory for imbalance and harmonic sharing in 4-wire microgrids,” IEEE
Trans. Smart Grid, vol. 11, no. 2, pp. 1604–1619, Mar. 2020.

[40] IEEE Standard Association, IEEE Std. 1547-2018. Standard for interconnection and inter-
operability of distributed energy resources with associated electric power systems interfaces.
IEEE, 2018.

[41] Y. Khayat, Q. Shafiee, R. Heydari, M. Naderi, T. Dragičević, J. W. Simpson-Porco, F. Dör-

fler, M. Fathi, F. Blaabjerg, J. M. Guerrero, and H. Bevrani, “On the secondary control
architectures of ac microgrids: An overview,” IEEE Trans. Power Electron., vol. 35, no. 6,
pp. 6482–6500, Jun. 2020.

[42] F. L. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative Control of

Multi-Agent Systems, ser. Communications and Control Engineering. London: Springer
London, 2014. [Online]. Available: http://link.springer.com/10.1007/978-1-4471-5574-4

[43] E. Espina, R. Cardenas-Dobson, J. W. Simpson-Porco, D. Saez, and M. Kazerani, “A

Consensus-Based Secondary Control Strategy for Hybrid AC/DC Microgrids with Exper-
imental Validation,” IEEE Trans. Power Electron., vol. 36, no. 5, pp. 5971–5984, May 2021.

[44] Q. Shafiee, V. Nasirian, J. C. Vasquez, J. M. Guerrero, and A. Davoudi, “A multi-functional

fully distributed control framework for ac microgrids,” IEEE Trans. Smart Grid, vol. 9, no. 4,
pp. 3247–3258, Jul. 2018.

[45] C. Bordons, F. Garcia-Torres, and M. A. Ridao, Model predictive control of microgrids,

1st ed., ser. Advances in Industrial Control. Springer International Publishing, 2020.

[46] J. Hu, Y. Shan, J. M. Guerrero, A. Ioinovici, K. W. Chan, and J. Rodriguez, “Model predic-
tive control of microgrids – an overview,” Renew. Sust. Energ. Rev., vol. 136, p. 110422, Feb.
2021.

[47] E. F. Camacho and C. Bordons, “Constrained model predictive control,” in Model Predictive
Control, 2nd ed., ser. Advanced Textbooks in Control and Signal Processing. London:
Springer London, 2007.

[48] O. Babayomi, Z. Zhang, T. Dragicevic, R. Heydari, Y. Li, C. Garcia, J. Rodriguez, and

R. Kennel, “Advances and opportunities in the model predictive control of microgrids: Part
II–Secondary and tertiary layers,” Int. J. Electr. Power Energy Syst., vol. 134, p. 107339, Jan.
2022.

128
[49] J. S. Gomez, D. Saez, J. W. Simpson-Porco, and R. Cardenas, “Distributed predictive control
for frequency and voltage regulation in microgrids,” IEEE Trans. Smart Grid, vol. 11, no. 2,
pp. 1319–1329, Mar. 2020.

[50] Q. Sun, J. Zhou, J. M. Guerrero, and H. Zhang, “Hybrid three-phase/single-phase microgrid

architecture with power management capabilities,” IEEE Trans. Power Electron., vol. 30,
no. 10, pp. 5964–5977, Oct. 2015.

[51] S. A. Raza and J. Jiang, “Intra-and inter-phase power management and control of a res-
idential microgrid at the distribution level,” IEEE Trans. Smart Grid, vol. 10, no. 6, pp.
6839–6848, Nov. 2019.

[52] D. I. Brandao, W. M. Ferreira, A. M. S. Alonso, E. Tedeschi, and F. P. Marafão, “Optimal

multiobjective control of low-voltage ac microgrids: Power flow regulation and compensa-
tion of reactive power and unbalance,” IEEE Trans. Smart Grid, vol. 11, no. 2, pp. 1239–
1252, Mar. 2020.

[53] E. Espina, R. Cárdenas-Dobson, M. B. Espinoza, C. Burgos-Mellado, and D. Saez, “Cooper-

ative regulation of imbalances in three-phase four-wire microgrids using single-phase droop
control and secondary control algorithms,” IEEE Trans. Power Electron., vol. 35, no. 2, pp.
1978–1992, Feb. 2020.

[54] J. Vásquez, J. Guerrero, M. Savaghebi, E. G. Carrasco, and R. Teodorescu, “Modeling, anal-

ysis, and design of stationary reference frame droop controlled parallel three-phase voltage
source inverters,” IEEE Trans. Ind. Electron., vol. 60, pp. 1271–1280, Apr. 2013.

[55] S. Parhizi, H. Lotfi, A. Khodaei, and S. Bahramirad, “State of the art in research on micro-
grids: A review,” IEEE Access, vol. 3, pp. 890–925, Jun. 2015.

[56] G. S. Rawat and Sathans, “Survey on DC microgrid architecture, power quality issues and
control strategies,” in 2018 2nd Int. Conf. Inven. Syst. Control, no. Icisc. IEEE, Jan. 2018,
pp. 500–505.

[57] A. Bidram and A. Davoudi, “Hierarchical structure of microgrids control system,” IEEE
Trans. Smart Grid, vol. 3, no. 4, pp. 1963–1976, Dec. 2012.

[58] F. Dörfler, J. W. Simpson-Porco, and F. Bullo, “Breaking the hierarchy: Distributed control
and economic optimality in microgrids,” IEEE Trans. Control Netw. Syst., vol. 3, no. 3, pp.
241–253, Sep. 2016.

[59] Z. Li, Z. Cheng, J. Liang, J. Si, L. Dong, and S. Li, “Distributed event-triggered secondary

129
 control for economic dispatch and frequency restoration control of droop-controlled AC mi-
crogrids,” IEEE Trans. Sustain. Energy, vol. 11, no. 3, pp. 1938–1950, Jul. 2020.

[60] X. Feng, A. Shekhar, F. Yang, R. E. Hebner, and P. Bauer, “Comparison of hierarchical con-
trol and distributed control for microgrid,” Electr. Power Components Syst., vol. 45, no. 10,
pp. 1043–1056, Jul. 2017.

[61] T. Dragičević, X. Lu, J. C. Vasquez, and J. M. Guerrero, “Dc microgrids—part i: A review

of control strategies and stabilization techniques,” IEEE Transactions on Power Electronics,
vol. 31, no. 7, pp. 4876–4891, Jul 2016.

[62] Q. Shafiee, T. Dragičević, J. C. Vasquez, and J. M. Guerrero, “Hierarchical control for mul-
tiple dc-microgrids clusters,” IEEE Transactions on Energy Conversion, vol. 29, no. 4, pp.
922–933, Dec 2014.

[63] M. Ben Said-Romdhane, M. W. Naouar, I. S. Belkhodja, and E. Monmasson, “An Improved

LCL Filter Design in Order to Ensure Stability without Damping and Despite Large Grid
Impedance Variations,” Energies 2017, Vol. 10, Page 336, vol. 10, no. 3, p. 336, Mar. 2017.

[64] Y.-J. Kim and H. Kim, “Optimal design of LCL filter in grid-connected inverters,” IET Power
Electron., vol. 12, no. 7, pp. 1774–1782, Jun. 2019.

[65] J. Hu, J. M. Guerrero, and S. Islam, Model predictive control for microgrids from power elec-
tronic converters to energy management, ser. IET energy engineering series 199. London:
The Institution of Engineering and Technology, 2021.

[66] A. Mehrizi-Sani and R. Iravani, “Secondary control for microgrids using potential functions
: Modeling issues,” in Conference on Power Systems, 2009.

[67] J. W. Simpson-Porco, Q. Shafiee, F. Dorfler, J. C. Vasquez, J. M. Guerrero, and F. Bullo,

“Secondary frequency and voltage control of islanded microgrids via distributed averaging,”
IEEE Trans. Ind. Electron., vol. 62, no. 11, pp. 7025–7038, Nov. 2015.

[68] A. Parisio, C. Wiezorek, T. Kyntäjä, J. Elo, K. Strunz, and K. H. Johansson, “Cooperative

MPC-based energy management for networked microgrids,” IEEE Trans. Smart Grid, vol. 8,
no. 6, pp. 3066–3074, Aug. 2017.

[69] D. E. Olivares, C. A. Cañizares, and M. Kazerani, “A centralized energy management system

for isolated microgrids,” IEEE Transactions on Smart Grid, vol. 5, no. 4, pp. 1864–1875,
July 2014.

[70] D. E. Olivares, C. A. Canizares, and M. Kazerani, “A centralized optimal energy manage-

130
 ment system for microgrids,” in 2011 IEEE Power and Energy Society General Meeting,
July 2011, pp. 1–6.

[71] O. Cartagena, S. Parra, D. Muñoz-Carpintero, L. G. Marín, and D. Sáez, “Review on fuzzy

and neural prediction interval modelling for nonlinear dynamical systems,” IEEE Access,
vol. 9, pp. 23 357–23 384, 2021.

[72] B. V. Solanki, A. Raghurajan, K. Bhattacharya, and C. A. Cañizares, “Including smart loads

for optimal demand response in integrated energy management systems for isolated micro-
grids,” IEEE Transactions on Smart Grid, vol. 8, no. 4, pp. 1739–1748, Jul. 2017.

[73] B. V. Solanki, K. Bhattacharya, and C. A. Cañizares, “A sustainable energy management

system for isolated microgrids,” IEEE Transactions on Sustainable Energy, vol. 8, no. 4, pp.
1507–1517, Oct 2017.

[74] B. V. Solanki, C. A. Cañizares, and K. Bhattacharya, “Practical energy management systems

for isolated microgrids,” IEEE Transactions on Smart Grid, pp. 1–1, Sep. 2018.

[75] I. A. Sajjad, G. Chicco, and R. Napoli, “Effect of aggregation level and sampling time
on load variation profile — a statistical analysis,” in MELECON 2014 - 2014 17th IEEE
Mediterranean Electrotechnical Conference, Apr 2014, pp. 208–212.

[76] M. Shahbazi and A. Khorsandi, “Power Electronic Converters in Microgrid Applications,”

Microgrid: Advanced Control Methods and Renewable Energy System Integration, pp. 281–
309, jan 2017.

[77] P. C. Loh, D. Li, Y. K. Chai, and F. Blaabjerg, “Autonomous operation of hybrid microgrid
with ac and dc subgrids,” IEEE Trans. Power Electron., vol. 28, no. 5, pp. 2214–2223, May
2013.

[78] H. J. Yoo, T. T. Nguyen, and H. M. Kim, “Consensus-based distributed coordination control

of hybrid AC/DC microgrids,” IEEE Trans. Sustain. Energy, vol. 11, no. 2, pp. 629–639,
Apr. 2020.

[79] J.-W. Chang, G.-S. Lee, S.-I. Moon, and P.-I. Hwang, “A novel distributed control method
for interlinking converters in an islanded hybrid ac/dc microgrid,” IEEE Trans. Smart Grid,
vol. 12, no. 5, pp. 3765–3779, Sep. 2021.

[80] Z. Guo, S. Li, and Y. Zheng, “Feedback linearization based distributed model predictive
control for secondary control of islanded microgrid,” Asian J. Control, vol. 22, no. 1, pp.
460–473, Jan. 2020.

131
[81] G. A. Papafotiou, G. D. Demetriades, and V. G. Agelidis, “Technology readiness assessment
of model predictive control in medium-and high-voltage power electronics,” IEEE Trans.
Ind. Electron., vol. 63, no. 9, pp. 5807–5815, Jan. 2016.

[82] A. Andersson and T. Thiringer, “Assessment of an improved finite control set model predic-
tive current controller for automotive propulsion applications,” IEEE Trans. Ind. Electron.,
vol. 67, no. 1, pp. 91–100, Feb. 2019.

[83] F. Mehmood, B. Khan, S. M. Ali, and J. A. Rossiter, “Distributed model predictive based
secondary control for economic production and frequency regulation of MG,” IET Control
Theory Appl., vol. 13, no. 17, pp. 2948–2958, Nov. 2019.

[84] L. Liang, Y. Hou, and D. J. Hill, “Design guidelines for MPC-based frequency regulation
for islanded microgrids with storage, voltage, and ramping constraints,” IET Renew. Power
Gener., vol. 11, no. 8, pp. 1200–1210, Jul. 2017.

[85] Y. Du, J. Wu, S. Li, C. Long, and S. Onori, “Coordinated energy dispatch of autonomous
microgrids with distributed MPC optimization,” IEEE Trans. Ind. Informat., vol. 15, no. 9,
pp. 5289–5298, Feb. 2019.

[86] J. Rossiter, A First Course in Predictive Control. CRC Press, 2018. [Online]. Available:
https://books.google.cl/books?id=q30StAEACAAJ

[87] C. Ahumada, R. Cárdenas, D. Sáez, and J. M. Guerrero, “Secondary Control Strategies

for Frequency Restoration in Islanded Microgrids With Consideration of Communication
Delays,” IEEE Transactions on Smart Grid, vol. 7, no. 3, pp. 1430–1441, May 2016.

[88] J. Rossiter, Model-Based Predictive Control: A Practical Approach (Control Series), 1st ed.
CRC Press, 2017.

[89] R. Holiš and V. Bobál, “Model predictive control of time-delay systems with measurable
disturbance compensation,” in 2015 20th International Conference on Process Control (PC),
2015, pp. 209–214.

[90] G. Lou, W. Gu, Y. Xu, M. Cheng, and W. Liu, “Distributed mpc-based secondary voltage
control scheme for autonomous droop-controlled microgrids,” IEEE Trans. Sustain. Energy,
vol. 8, no. 2, pp. 792–804, Apr. 2017.

[91] K. Liu, T. Liu, Z. Tang, and D. J. Hill, “Distributed MPC-based frequency control in net-
worked microgrids with voltage constraints,” IEEE Trans. Smart Grid, vol. 10, no. 6, pp.
6343–6354, Nov. 2019.

132
 [92] Z. Guo, H. Jiang, Y. Zheng, and S. Li, “Distributed model predictive control for efficient
operation of islanded microgrid,” in 2017 Chinese Automation Congress (CAC). Institute
of Electrical and Electronics Engineers Inc., Oct. 2017, pp. 6253–6258.

[93] R. Reginatto and R. A. Ramos, “On electrical power evaluation in dq coordinates under
sinusoidal unbalanced conditions,” IET Gener. Transm. Distrib., vol. 8, no. 5, pp. 976–982,
May 2014.

[94] S. V. Viscido, J. K. Parrish, and D. Grünbaum, “The effect of population size and number
of influential neighbors on the emergent properties of fish schools,” Ecol. Modell., vol. 183,
no. 2, pp. 347 – 363, Apr. 2005.

[95] F. Guo, C. Wen, J. Mao, and Y. Song, “Distributed secondary voltage and frequency restora-
tion control of droop-controlled inverter-based microgrids,” IEEE Trans. Ind. Electron.,
vol. 62, no. 7, pp. 4355–4364, Jul. 2015.

[96] B. Wei, Y. Gui, S. Trujillo, J. M. Guerrero, and J. C. Vásquez, “Distributed average integral
secondary control for modular ups systems-based microgrids,” IEEE Trans. Power Electron.,
vol. 34, no. 7, pp. 6922–6936, Jul. 2019.

[97] Y. Han, K. Zhang, H. Li, E. A. A. Coelho, and J. M. Guerrero, “Mas-based distributed

coordinated control and optimization in microgrid and microgrid clusters: A comprehensive
overview,” IEEE Trans. Power Electron., vol. 33, no. 8, pp. 6488–6508, Aug. 2018.

[98] G. Chen and Z. Guo, “Distributed secondary and optimal active power sharing control for
islanded microgrids with communication delays,” IEEE Trans. Smart Grid, vol. 10, no. 2,
pp. 2002–2014, Mar. 2019.

[99] C. Li, J. C. Vasquez, and J. M. Guerrero, “Convergence analysis of distributed control for
operation cost minimization of droop controlled DC microgrid based on multiagent,” in Con-
ference Proceedings - IEEE Applied Power Electronics Conference and Exposition - APEC,
vol. 2016-May. Institute of Electrical and Electronics Engineers Inc., may 2016, pp. 3459–
3464.

[100] G. Chen and Z. Zhao, “Delay Effects on Consensus-Based Distributed Economic Dispatch
Algorithm in Microgrid,” IEEE Transactions on Power Systems, vol. 33, no. 1, pp. 602–612,
may 2017.

[101] T. Zhao and Z. Ding, “Distributed agent consensus-based optimal resource management for
microgrids,” IEEE Transactions on Sustainable Energy, vol. 9, no. 1, pp. 443–452, Jan 2018.

133
[102] Z. Wang, W. Wu, and B. Zhang, “A fully distributed power dispatch method for fast fre-
quency recovery and minimal generation cost in autonomous microgrids,” IEEE Transac-
tions on Smart Grid, vol. 7, no. 1, pp. 19–31, Jan 2016.

[103] Y. Xu and Z. Li, “Distributed optimal resource management based on the consensus al-
gorithm in a microgrid,” IEEE Transactions on Industrial Electronics, vol. 62, no. 4, pp.
2584–2592, Apr 2015.

[104] X. He, J. Yu, T. Huang, and C. Li, “Distributed power management for dynamic economic
dispatch in the multimicrogrids environment,” IEEE Transactions on Control Systems Tech-
nology, vol. 27, no. 4, pp. 1651–1658, Jul 2019.

[105] C. Li, X. Yu, T. Huang, and X. He, “Distributed optimal consensus over resource allocation
network and its application to dynamical economic dispatch,” IEEE Transactions on Neural
Networks and Learning Systems, vol. 29, no. 6, pp. 2407–2418, Jun 2018.

[106] B. Huang, L. Liu, Y. Li, and H. Zhang, “Distributed optimal energy management for mi-
crogrids in the presence of time-varying communication delays,” IEEE Access, vol. 7, pp.
83 702–83 712, Jun. 2019.

[107] P. Lü, J. Zhao, J. Yao, and S. Yang, “A decentralized approach for frequency control and eco-
nomic dispatch in smart grids,” IEEE Journal on Emerging and Selected Topics in Circuits
and Systems, vol. 7, no. 3, pp. 447–458, Sep. 2017.

[108] M. Hamdi, M. Chaoui, L. Idoumghar, and A. Kachouri, “Coordinated consensus for smart
grid economic environmental power dispatch with dynamic communication network,” IET
Generation, Transmission Distribution, vol. 12, no. 11, pp. 2603–2613, Jun 2018.

[109] H. Xing, Z. Lin, M. Fu, and B. F. Hobbs, “Distributed algorithm for dynamic economic
power dispatch with energy storage in smart grids,” IET Control Theory Applications,
vol. 11, no. 11, pp. 1813–1821, Jul 2017.

[110] A. Gabash and P. Li, “Active-reactive optimal power flow in distribution networks with em-
bedded generation and battery storage,” IEEE Transactions on Power Systems, vol. 27, no. 4,
pp. 2026–2035, 2012.

[111] Y. Zhao, M. Irving, and Y. Song, “A cost allocation and pricing method for reactive power
service in the new deregulated electricity market environment,” in 2005 IEEE/PES Trans-
mission Distribution Conference Exposition: Asia and Pacific, 2005, pp. 1–6.

[112] S. M. Sadek, W. A. Omran, M. A. M. Hassan, and H. E. A. Talaat, “Data driven stochas-

134
 tic energy management for isolated microgrids based on generative adversarial networks
considering reactive power capabilities of distributed energy resources and reactive power
costs,” IEEE Access, vol. 9, pp. 5397–5411, 2021.

[113] M. M. Gomez, C. Burgos Mellado, and R. C. Dobson, “Distributed control for a cost-based
droop-free microgrid,” in 2020 IEEE 21st Workshop on Control and Modeling for Power
Electronics (COMPEL), 2020, pp. 1–7.

[114] I. Khan, Z. Li, Y. Xu, and W. Gu, “Distributed control algorithm for optimal reactive power
control in power grids,” International Journal of Electrical Power & Energy Systems, vol. 83,
pp. 505–513, 2016.

[115] V. Nasirian, F. L. Lewis, and A. Davoudi, “Distributed optimal dispatch for DC distribution
networks,” in 2015 IEEE 1st Int. Conf. Direct Curr. Microgrids, ICDCM 2015. Institute of
Electrical and Electronics Engineers Inc., Jul. 2015, pp. 97–101.

[116] S. Moayedi and A. Davoudi, “Unifying Distributed Dynamic Optimization and Control of
Islanded DC Microgrids,” in IEEE Trans. Power Electron., vol. 32, no. 3. Institute of
Electrical and Electronics Engineers Inc., Mar. 2017, pp. 2329–2346.

[117] J. Hu, J. Duan, H. Ma, and M.-Y. Chow, “Distributed adaptive droop control for optimal
power dispatch in dc microgrid,” IEEE Transactions on Industrial Electronics, vol. 65, no. 1,
pp. 778–789, 2018.

[118] H. Han, H. Wang, Y. Sun, J. Yang, and Z. Liu, “Distributed control scheme on cost optimisa-
tion under communication delays for DC microgrids,” IET Gener. Transm. Distrib., vol. 11,
no. 17, pp. 4193–4201, Nov 2017.

[119] A. Hussain, V.-H. Bui, and H.-M. Kim, “Robust optimal operation of ac/dc hybrid micro-
grids under market price uncertainties,” IEEE Access, vol. 6, pp. 2654–2667, 2018.

[120] A. Maulik and D. Das, “Optimal power dispatch considering load and renewable generation
uncertainties in an AC–DC hybrid microgrid,” IET Gener. Transm. Distrib., vol. 13, no. 7,
pp. 1164–1176, apr 2019.

[121] P. Buduma, M. K. Das, S. Mishra, and G. Panda, “Robust power management and control
for hybrid ac-dc microgrid,” in 2020 3rd International Conference on Energy, Power and
Environment: Towards Clean Energy Technologies, 2021, pp. 1–6.

[122] K. Zhang, M. Su, Y. Sun, P. Wu, Z. Luo, and H. Han, “A novel distributed control for hybrid
ac/dc microgrid with consideration of power limit,” in 2021 IEEE 12th Energy Conversion

135
 Congress Exposition - Asia (ECCE-Asia), 2021, pp. 1856–1859.

[123] P. Lin, C. Jin, J. Xiao, X. Li, D. Shi, Y. Tang, and P. Wang, “A distributed control architec-
ture for global system economic operation in autonomous hybrid ac/dc microgrids,” IEEE
Transactions on Smart Grid, vol. 10, no. 3, pp. 2603–2617, 2019.

[124] P. Yang, M. Yu, Q. Wu, P. Wang, Y. Xia, and W. Wei, “Decentralized economic operation
control for hybrid ac/dc microgrid,” IEEE Transactions on Sustainable Energy, vol. 11, no. 3,
pp. 1898–1910, 2020.

[125] Z. Zhao, J. Zhang, B. Yan, R. Cheng, C. S. Lai, L. Huang, Q. Guan, and L. L. Lai, “Decen-
tralized Finite Control Set Model Predictive Control Strategy of Microgrids for Unbalanced
and Harmonic Power Management,” IEEE Access, vol. 8, pp. 202 298–202 311, Oct. 2020.

[126] J. Liu, Y. Miura, and T. Ise, “Cost-Function-Based Microgrid Decentralized Control of Un-
balance and Harmonics for Simultaneous Bus Voltage Compensation and Current Sharing,”
IEEE Trans. Power Electron., vol. 34, no. 8, pp. 7397–7410, Aug. 2019.

[127] S. R. Mohapatra and V. Agarwal, “Model Predictive Control for Flexible Reduction of Ac-
tive Power Oscillation in Grid-Tied Multilevel Inverters under Unbalanced and Distorted
Microgrid Conditions,” IEEE Trans. Ind. Appl., vol. 56, no. 2, pp. 1107–1115, Mar. 2020.

[128] M. Yazdanian and A. Mehrizi-Sani, “Distributed control techniques in microgrids,” IEEE

Trans. Smart Grid, vol. 5, no. 6, pp. 2901–2909, Nov. 2014.

[129] S. Boyd and L. Vandenberghe, Convex optimization. USA: Cambridge University Press,
2004.

[130] P. Kundur, Power system stability and control. McGraw-Hill, 1994, vol. 7.

[131] A. E. Leon and J. A. Solsona, “Design of reduced-order nonlinear observers for energy
conversion applications,” IET Control Theory Appl., vol. 4, no. 5, pp. 724–734, May 2010.

[132] F. Bullo, Lectures on network systems, 1st ed. Kindle Direct Publishing, 2020,
with contributions by J. Cortes, F. Dorfler, and S. Martinez. [Online]. Available:
http://motion.me.ucsb.edu/book-lns

[133] W. Ren, R. W. Beard, and E. M. Atkins, “Information consensus in multivehicle cooperative

control,” IEEE Control Syst., vol. 27, no. 2, pp. 71–82, Jul. 2007.

[134] I. Serban, S. Cespedes, C. Marinescu, C. A. Azurdia-Meza, J. S. Gomez, and D. Saez Hue-

ichapan, “Communication requirements in microgrids: A practical survey,” IEEE Access,

136
 vol. 8, pp. 47 694–47 712, Mar. 2020.

[135] Z. Wang, S. Mei, F. Liu, P. Yi, and M. Cao, “Asynchronous Distributed Power Control of
Multi-Microgrid Systems Based on the Operator Splitting Approach,” arXiv, Oct 2018.

[136] C. Schmid and L. T. Biegler, “Quadratic programming methods for reduced hessian sqp,”
Comput. Chem. Eng., vol. 18, no. 9, pp. 817–832, Sep. 1994.

[137] K. J. Åström and B. Wittenmark, Adaptive Control. Addison-Wesley, 1989.

[138] Y. jun Zhang and Z. Ren, “Optimal reactive power dispatch considering costs of adjusting
the control devices,” IEEE Trans. Power Syst, vol. 20, no. 3, pp. 1349–1356, Aug. 2005.

[139] J. Solsona and A. Leon, “Design of reduced-order nonlinear observers for energy conversion
applications,” IET Control Theory Appl., vol. 4, no. 5, pp. 724–734, Jul. 2010.

[140] “Triphase.” [Online]. Available: https://triphase.com/products/

[141] E. Espina, C. Burgos-Mellado, J. S. Gomez, J. Llanos, E. Rute, A. Navas F., M. Martínez-

Gómez, R. Cárdenas, and D. Sácz, “Experimental hybrid ac/dc-microgrid prototype for labo-
ratory research,” in 2020 22nd European Conference on Power Electronics and Applications
(EPE’20 ECCE Europe), 2020, pp. 1–9.

137
Annexes

Annexed A

Extended abstract

Microgrids (MGs) are the cornerstone for a new model of electrical generation and distribution
based on renewable resources. However, managing the operation of a MG is a challenging and
complex task due to the characteristics of the various types of renewable sources and interactions
between different types of generating equipment. In this context, some of the most pressing prob-
lems in MGs are associated with guaranteeing that they operate in a cost-effective way and that
they correctly manage the quality of the supply. Therefore, new and more reliable control strate-
gies need to be developed for the management of microgrids. Distributed model predictive control
(DMPC) is positioned as one of the best solutions for MGs as it can model complex systems and
address multiple objectives simultaneously.

Traditionally MGs have been controlled via a three-level hierarchical structure, where each level
operates at a different time scale. The primary control level is the fastest and aims to maintain the
stability of the MG and ensures correct power sharing. The secondary control level restores the
variables modified by the primary control level. The tertiary control level is the slowest and aims
for economic dispatch ( i.e. aiming for the lowest monetary cost of generated energy) of the MG
and correct coordination with the main grid. However, isolated MGs are prone to fast changes in
generation and demand whilst having a slow time response at the tertiary control level. The latest
research suggests that this control should be performed on a time-scale comparable to that used at
the secondary control level. In addition, as the power references sent by the tertiary control level
tend to be updated with a slower sample time, the power limits of distributed generators (DGs) used
in the system can be exceeded.

This thesis, therefore, focuses on the application of DMPC schemes for the secondary control
level for ac MGs and hybrid ac/dc microgrids (H-MGs - composed of an ac sub-MG and a dc sub-

138
MG connected through interlinking converters (ILCs)). The main characteristics of the proposed
methodologies are the development of novel multi objective cost functions and prediction models
that correctly represent the main dynamics of the DGs and the ILCs (in the case of H-MGs) in the
formulation. Three control strategies are proposed that fulfil the main task of the secondary control
level (i.e. restoring frequency and voltage). These strategies are able to restore the frequency and
voltage to nominal values or within secure bands. The first proposed strategy considers the eco-
nomic dispatch of DGs in a balanced ac MG. The second strategy achieves the economic dispatch
of ac DGs, dc DGs and manages the power transference of ILCs based on an economic criterion in
H-MGs. The third strategy manages the sharing of phase imbalance in an unbalanced ac MG. The
proposed strategies all include important operating constraints, e.g., power limits due to convertor
ratings.

Extensive experimental, hardware-in-the-loop and simulation studies are used to validate the
proposed DMPC schemes for the most common operating scenarios in MGs, namely, load changes,
requirement for plug-and-play DGs and ILCs (in H-MGs), and communication link failures and
communication delays. Finally, the controllers’ scalability has been investigated, and comparison
studies have also been performed to highlight the advantages of the proposed schemes over other
reported distributed schemes.

139
Annexes

Annexed B

Design of a reduced order nonlinear

observer to estimate the voltage after a
coupling inductance

In this appendix, we provide an explanation of the application of the reduced-order non-linear

observer proposed in [139] to estimate the voltage (V̂iB ) after the coupling inductance (Li ). For
a complete description of the demonstration of the observer, the reader is encouraged to read the
aforementioned work.

Consider a class of nonlinear system given by

ẋ = F (xa , u) x + g (xa , u) (B.1)

where x ∈ Rn×1 is the state vector and u ∈ Rm×1 is the input vector, with F ∈ Rn×n and g ∈ Rn×1 .
The state vector can be partionated as x = [xa xb ]T , where xa ∈ Rna ×1 contains measurable variables
and xb ∈ Rnb ×1 contains non-measurable variables. The representation of (B.1) can be rewritten as
follows.

" # " #" # " #

ẋa N (xa , u) M (xa , u) xa ga (xa , u)
= + (B.2)
ẋb R (xa , u) S (xa , u) xb gb (xa , u)

The previously described non-linear observer of reduced-order is used to estimate the voltage
(V̂iB )
after the coupling inductance (Li ), as shown at the bottom of Fig. B.1. The observer works in

140
 B = V sin(θ ) and V̂ B = V cos(θ ), where V depends on the
the α − β framework. Considering V̂α,i m i β ,i m i m
abc − αβ transformation used. Equation (B.3), which represents the estimated states, is obtained
deriving both expressions. Where ωo is the nominal frequency.

To the
Microgrid

DGi

Figure B.1: Electrical output circuit

V̇ˆα,i
B = V ω cos(θ ) = ω V̂ B
m o i o β ,i
(B.3)
V̇ˆβB,i = −Vm ωo sin(θi ) = −ωoV̂α,i
B

Equation (B.4), which represents the measured estates, is obtained applying the Kirchhoff’s
voltage law to the circuit of Fig. B.1. Where Ri represents the cable resistance, and is assumed as
Ri = 0.01Ω.

Li i̇α,i = −Riα,i +Vα,i − V̂ α, iB

(B.4)
Li i̇β ,i = −Riβ ,i +Vβ ,i − V̂ β , iB

The measured states (xa ) and estimated states (xb ) are presented in Equation (B.5). These are
obtained expressing (B.3) and (B.4) in the required form of the observer (B.2). Where its inputs
are the measured values of the output voltage Vi and the output current ii (both at the output of the
LCL filter, before Li , and in the α − β framework).

" # " #" # " #" # "V #

i̇α,i − RLii 0 iα,i − L1i 0 V̂α,iB α,i
Li
= + + Vβ ,i
i̇β ,i 0 − RLii iβ ,i 0 − L1i V̂βB,i L
| {z } | {z } | {z } | {z } | {z } | {zi }
ẋa" # " N # " xa# " M #" x#b " #ga (B.5)
ˆ B
V̇α,i 0 0 iα,i 0 ωo V̂α,iB
0
ˆ B
= + B
+
V̇β ,i 0 0 iβ ,i −ωo 0 V̂β ,i 0
| {z } | {z } | {z } | {z } | {z } |{z}
ẋb R xa S xb gb

The structure of the observer is presented in (B.6).

141
 ξ˙ =Ar (ξ + Gw) + Br
(B.6)
x̂b = ξ + Gw

Where w is a transformation that depends on the measured variables to obtain a linear dynamic
of the error.
" # " ωo i #
ω1 −Li (iα,i + gvβ ,i )
w = T(xa ) = = ω i (B.7)
ω2 −Li (iβ ,i − ogvα,i )

The estimation error dynamic Ar (B.8) is obtained through pole placement so that the observer
is able to follow the phase of the estimated voltages, and it is faster than the secondary controller.
Finally, the gains gv were placed at -31500, and Br is represented in (B.9).

" #
−gv 0
Ar = −G = (B.8)
0 −gv

∂T
Br = Rxa + gb − G (Nxa + ga ) (B.9)
∂ xa

142
Annexes

Annexed C

Derivation of predictive linear models used

as equality constraints in ac DGs

C.1 Continuous time model for equality constraints

The set of equations (3.6)-(3.8) and (5.1)-(5.4) is rewritten as (C.1). As it was mentioned, (C.1)
characterises frequency and voltage droop controllers, phase angle deviation and, the active/reactive
power transferred from the i-th DG to the microgrid.

ωi (t)=ω0+Mpω,i Pi (t)+ωs,i (t) (C.1a)

Z t
B
ωi (τ)−ω̂iB (τ) dτ

δ θi (t) = θi (t)−θ̂i (t) = (C.1b)
0

Pi (t) = BiVi (t)V̂iB (t)sin(δ θi (t)) (C.1c)

1
Vi(t)= (Via(t)+Vib(t)+Vic(t)) (C.1d)
3

Via (t)=V0+Mqv,i Qia (t)+Vs,ia (t) (C.1e)

143
 Vib (t)=V0+Mqv,i Qib (t)+Vs,ib (t) (C.1f)

Vic (t)=V0+Mqv,i Qic (t)+Vs,ic (t) (C.1g)

Qi(t)=Qia(t)+Qib(t)+Qic(t) (C.1h)

h i
Qia(t)=Bi Via(t)2−Via(t)V̂iaB(t)cos(δ θi(t)) (C.1i)

h i
Qib(t)=Bi Vib(t)2−Vib(t)V̂ibB(t)cos(δ θi(t)) (C.1j)

h i
Qic(t)=Bi Vic(t)2−Vic(t)V̂icB(t)cos(δ θi(t)) (C.1k)

Proof. Balanced Case

Assumption 1 A balanced microgrid satisfies:

Via (t) = Vib (t) = Vic (t) = Vix (t) ; Vs,ia (t) = Vs,ib (t) = Vs,ic (t) = Vs,ix (t) ;
V̂iaB (t) = V̂ibB (t) = V̂icB (t) = V̂ixB (t) (C.2)

Therefore:
from (C.1h) to (C.1k)
h i
Qi (t) =Bi Via (t)2 −Via (t) V̂iaB (t) cos (δ θi (t)) +
h i
Bi Vib (t)2 −Vib (t) V̂ibB (t) cos (δ θi (t)) +
h i
2 B
Bi Vic (t) −Vic (t) V̂ic (t) cos (δ θi (t)) (C.3)

h i
Qi (t) = 3Bi Vix (t)2 −Vix (t) V̂ixB (t) cos (δ θi (t)) (C.4)

144
from (C.1d) to (C.1g)

1
Vi (t) = [V0 + Mqv,i Qia (t) +Vs,ia (t) +V0 + Mqv,i Qib (t) +Vs,ib (t) +V0 ] +
3
1
[V0 + Mqv,i Qic (t) +Vs,ic (t)] (C.5)
3

1
Vi (t) = [3V0 + 3Mqv,i Qix (t) + 3Vs,ix (t)] = Vix (t) (C.6)
3

C.2 Model discretisation

This section details the discretisation of equations (C.1). We use the forward Euler method defined
by (C.7). In this case, it is considered that k = nTsec , n ∈ Z+ , and Tsec is the sample time used at
secondary control level.

d f (t)
Tsec = [ f (k + 1) − f (k)] = ∆ f (k + 1) (C.7)
dt
t=k

The discretisation process of each equation (model) is detailed as follows:

C.2.1 Droop equations

The linearisation process for the frequency droop equation (C.1a) is detailed as follows. The same
procedure is applied to voltage droop equations (C.1e), (C.1f) and (C.1g)

From (C.1a), it is possible to rewrite ωs,i (t) in function of ∆ωs,i (t).

ωi (t) = ω0 + Mpω,i Pi (t) + ωs,i (t) (C.8)

Z
1
ωi (t) = ω0 + Mpω,i Pi (t) + ∆ ωs,i (t)dt
Tsec
Deriving both sides and applying forward Euler method:

dωi (t) dPi (t) 1

= Mpω,i + ∆ω (t)
dt dt Tsec s,i

dωi (t) dPi (t) 1

= Mpω,i + ∆ω (k)
dt dt Tsec s,i
t=k t=k

ωi (k + 1) − ωi (k) = Mpω,i [Pi (k + 1) − Pi (k)] + ∆ωs,i (k)

ωi (k + 1) = ωi (k) + Mpω,i [Pi (k + 1) − Pi (k)] + ∆ωs,i (k) (C.9)

145
 □

C.2.2 Phase angle equation

The linearisation process for the phase angle deviation model is shown from (B.10) to (B.12). Two
procedures that reach the same result are detailed.
Z t
B
ωi (τ) − ω̂iB (τ) dτ

δ θi (t) = θi (t) − θ̂i (t) = (C.10)
0

Procedure 1 Z t
ωi (τ) − ω̂iB (τ) dτ

δ θi (t) =
0

Deriving both sides and applying forward Euler method

dδ θi (t) d
hZ t   i
B
= ωi (τ) − ω̂i (τ) dτ
dt dt 0
t=k t=k

Z k+1  Z k
B
ωi (τ) − ω̂iB (τ) dτ
 
δ θi (k + 1) − δ θi (k) = ωi (τ) − ω̂i (τ) dτ −
0 0
Z k+1 
ωi (τ) − ω̂iB (τ) dτ

δ θi (k + 1) − δ θi (k) =
k

δ θi (k + 1) − δ θi (k) = Tsec [ωi (k) − ω̂iB (k)]

δ θi (k + 1) = δ θi (k) + Tsec [ωi (k) − ω̂iB (k)] (C.11)

Procedure 2

Considering Z t
θi (t) = ωi (τ), dτ
0

Deriving both sides and applying forward Euler method

dθi (t) d
hZ t i
= ωi (τ), dτ
dt dt 0
t=k t=k

Z k+1 Z k
θi (k + 1) − θi (k) = ωi (τ) dτ − ωi (τ) dτ
0 0
Z k+1
θi (k + 1) − θi (k) = ωi (τ)dτ
k

θi (k + 1) − θi (k) = Tsec ωi (k)

146
Then, from (A.1b)
δ θi (k + 1) = θi (k + 1) − θ̂iB (k + 1)

δ θi (k + 1) = [θi (k) + Tsec ωi (k)] − [θ̂iB (k) + Tsec ω̂iB (k)]

Re-ordering the terms

δ θi (k + 1) = [θi (k) − θ̂iB (k)] + [Tsec ωi (k) − Tsec ω̂iB (k)]

δ θi (k + 1) = δ θi (k) + Tsec [ωi (k) − ω̂iB (k)] (C.12)

C.2.3 Power transfer equations

Due to power transfer equations (B.1c) and (B.1i)-(B.1k) are non-linear, these are linearised via a
Taylor expansion around the measured/estimated point p(k) = {ωi (k), ω̂iB (k),Vi (k),Vix (k), V̂iB (k)
, V̂ixB (k), δ θi (k), Pi (k), Qi (k)} with x={a, b, c}. Then, the forward Euler discretisation is applied to
the linearised equations. The same procedure is applied in equations (B.1c) and (B.1i)-(B.1k), but
only the the procedure for (B.1c) is shown in the following.

Pi (t) = BiVi (t)V̂iB (t) sin (δ θi (t)) (C.13)

Linearising

∂ Pi (t) ∂ Pi (t)
Pi (t) =Pi (k) + [Vi (t) −Vi (k)] + [V̂ B (t) − V̂iB (k)]+ (C.14)
∂Vi
p(k)
∂ V̂iB p(k) i

∂ Pi (t)
[δ θi (t) − δ θi (k)]
∂δθ
p(k)

Pi (t) = Pi (k) + KV [Vi (t) −Vi (k)] + KV̂ B [V̂iB (t) − V̂iB (k)] + Kδ θ [δ θi (t) − δ θi (k)]
i

where
KV = BiV̂iB (k) sin(δ θi (k))

KV̂ B = BiVi (k) sin(δ θi (k))

i

Kδ θ = BiVi (k)V̂iB (k) cos(δ θi (k))

Deriving both sides and evaluating at t = k

Pi (t) Pi (k) d d d
= + KV [Vi (t) −Vi (k)] + KV̂ B [V̂iB (t) − V̂iB (k)] + Kδ θ [δ θi (t) − δ θi (k)]
dt dt dt i dt dt

147
 dPi (t) dVi (t) dV̂iB (t) dδ θi (t)
= KV + KV̂ B + Kδ θ
dt dt i dt dt
t=k t=k t=k t=k

Pi (k + 1) − Pi (k) =KV [Vi (k + 1) −Vi (k)] + KV̂ B [V̂iB (k + 1) − V̂iB (k)]+ (C.15)
i

Kδ θ [δ θi (k + 1) − δ θi (k)]

Assuming V̂iB is constant

V̂iB (k + 1) − V̂iB (k) = 0

Pi (k + 1) = Pi (k) + KV [Vi (k + 1) −Vi (k)] + Kδ θ [δ θi (k + 1) − δ θi (k)]

Then

Pi (k + 1) =Pi (k) + [BiV̂iB (k) sin(δ θi (k))][Vi (k + 1) −Vi (k)]+

[BiVi (k)V̂iB (k) cos(δ θi (k))][δ θi (k + 1) − δ θi (k)] (C.16)

Therefore, the linear-discrete time model used to state the predictive model is summarised in
(B.17).

ωi (k+1)=ωi (k)+Mpω,i[Pi (k+1)−Pi (k)]+∆ωs,i (k) (C.17a)

δ θi (k+1)=δ θi (k)+Tsec ωi (k+1)−ω̂iB (k)

 
(C.17b)

Pi (k+1)=Pi (k)+[Vi(k+1)−Vi(k)]BiV̂iB(k)sin(δ θi(k))+

[δ θi (k+1)−δ θi (k)]BiVi (k)V̂iB (k)cos(δ θi (k)) (C.17c)

1
Vi(k+1)= (Via(k+1)+Vib(k+1)+Vic(k+1)) (C.17d)
3

Via (k+1)=Via (k)+Mqv,i[Qia (k+1)−Qia (k)]+∆Vs,ia (k) (C.17e)

Vib (k+1)=Vib (k)+Mqv,i[Qib (k+1)−Qib (k)]+∆Vs,ib (k) (C.17f)

Vic (k+1)=Vic (k)+Mqv,i[Qic (k+1)−Qic (k)]+∆Vs,ic (k) (C.17g)

148
 Qi(k+1)=Qia(k+1)+Qib(k+1)+Qic(k+1) (C.17h)

Qia(k+1)=Qia(k)+[Via(k+1)−Via(k)]Bi 2Via(k)−V̂iaB(k)cos(δ θi(k)) +

 

[δ θi(k+1)−δ θi(k)]BiVia(k)V̂iaB(k)sin(δ θi(k)) (C.17i)

Qib(k+1)=Qib(k)+[Vib(k+1)−Vib(k)]Bi 2Vib(k)−V̂ibB(k)cos(δ θi(k)) +

 

[δ θi(k+1)−δ θi(k)]BiVib(k)V̂ibB(k)sin(δ θi(k)) (C.17j)

Qic(k+1)=Qic(k)+[Vic(k+1)−Vic(k)]Bi 2Vic(k)−V̂icB(k)cos(δ θi(k)) +

 

[δ θi(k+1)−δ θi(k)]BiVic(k)V̂icB(k)sin(δ θi(k)) (C.17k)

C.3 Prediction model for equality constraints

From (C.17) it is possible to state the predictive model (C.18) to be used as a set of equality con-
straints in the optimisation problem. Note that the measured/estimated coefficients at t = k are
considered constant along the prediction horizon.

ωi (k+m)=ωi (k+m−1)+Mpω,i[Pi (k+m)−Pi (k+m−1)]+∆ωs,i (k+m−1) (C.18a)

δ θi (k+m)=δ θi (k+m−1)+Tsec ωi (k+m)−ω̂iB (k)

 
(C.18b)

Pi (k+m)=Pi (k)+[Vi(k+m)−Vi(k)]BiV̂iB(k)sin(δ θi(k))+

[δ θi (k+m)−δ θi (k)]BiVi (k)V̂iB (k)cos(δ θi (k)) (C.18c)

1
Vi(k+m)= (Via(k+m)+Vib(k+m)+Vic(k+m)) (C.18d)
3

Via (k+m)=Via (k+m−1)+Mqv,i[Qia (k+m)−Qia (k+m−1)]+∆Vs,ia (k+m−1) (C.18e)

Vib (k+m)=Vib (k+m−1)+Mqv,i[Qib (k+m)−Qib (k+m−1)]+∆Vs,ib (k+m−1) (C.18f)

149
Vic (k+m)=Vic (k+m−1)+Mqv,i[Qic (k+m)−Qic (k+m−1)]+∆Vs,ic (k+m−1) (C.18g)

Qi(k+m)=Qia(k+m)+Qib(k+m)+Qic(k+m) (C.18h)

Qia(k+m)=Qia(k)+[Via(k+m)−Via(k)]Bi 2Via(k)−V̂iaB(k)cos(δ θi(k)) +

 

[δ θi(k+m)−δ θi(k)]BiVia(k)V̂iaB(k)sin(δ θi(k)) (C.18i)

Qib(k+m)=Qib(k)+[Vib(k+m)−Vib(k)]Bi 2Vib(k)−V̂ibB(k)cos(δ θi(k)) +

 

[δ θi(k+m)−δ θi(k)]BiVib(k)V̂ibB(k)sin(δ θi(k)) (C.18j)

Qic(k+m)=Qic(k)+[Vic(k+m)−Vic(k)]Bi 2Vic(k)−V̂icB(k)cos(δ θi(k)) +

 

[δ θi(k+m)−δ θi(k)]BiVic(k)V̂icB(k)sin(δ θi(k)) (C.18k)

150
Annexes

Annexed D

Derivation of predictive linear models used

as inequality constraints in ac DGs

Inequality constraints are stated to bound the feasible solution space, considering operational
requirements. This section presents the procedure to derive inequalities (5.12) and (5.16).

D.1 PVUR inequality

The phase voltage unbalance rate index (PVUR) is defined as (D.1) to quantify the voltage un-
balance at the i−th DG’s output, where Vi (t) is defined as the average voltage magnitude among
phases at t. It is required to preserve PVURi (t) below to its maximum value PVUR∗ . Thus, a linear
approximation is required to include this model into the optimisation problem.

max {[|Via (t)| −Vi (t)], [|Vib (t)| −Vi (t)], [|Vic (t)| −Vi (t)]}
PVURi (t) = (D.1)
Vi (t)

PVURi (t) ≤ PVUR∗ (D.2)

Considering that Vix (t) ≥ 0, from (D.1) and (D.2) a set of three inequalities (one per phase) can
be stated:

Vix (t) −Vi (t)

Fix (Via (t),Vib (t),Vic (t)) = ≤ PVUR∗ ∀x = a, b, c (D.3)
Vi (t)

151
Therefore, the linear approximation is defined as (D.4):

Fix (Via (t),Vib (t),Vic (t)) ≈Fix (Via (k),Vib (k),Vic (k))+
∂ Fix (Via (t),Vib (t),Vic (t))
 
[Via (t) −Via (k)] +
∂Via (t) t=k
∂ Fix (Via (t),Vib (t),Vic (t))
 
[Vib (t) −Vib (k)] +
∂Vib (t) t=k
∂ Fix (Via (t),Vib (t),Vic (t))
 
[Vic (t) −Vic (k)] (D.4)
∂Vic (t) t=k

Solving for x = a, and considering Vi (t) = Via (t)+Vib3(t)+Vic (t) and ∂Vi (t)
∂Via (t) = ∂Vi (t)
∂Vib (t) = ∂Vi (t)
∂Vic (t) = 1
3

∂ Fia (Via (t),Vib (t),Vic (t))

   
∂ Via (t) −Vi (t)
=
∂Via (t) t=k ∂Via (t) Vi (t) t=k
 
∂Vi (t) ∂Vi (t)
[1 − ∂V (t) ][Vi (t)] − [Via (t) −Vi (t)] ∂V
ia ia (t) 
= 2
Vi (t)
t=k
 
∂Vi (t)
[Vi (t)] − [Via (t)] ∂V (t)
ia
= 
Vi2 (t)
t=k
" V (t)+V (t)+V (t) #
ia ib ic 1
[ 3 ] − [V ia (t)] 3
= Via (t)+Vib (t)+Vic (t) 2
[ 3 ] t=k
 
[3][Via (t) +Vib (t) +Vic (t)] − [3Via (t)]
=
[Via (t) +Vib (t) +Vic (t)]2 t=k

∂ Fia (Via (t),Vib (t),Vic (t))

   
[3][Vib (k) +Vic (k)]
= (D.5)
∂Via (t) t=k [Via (k) +Vib (k) +Vic (k)]2

152
 ∂ Fix (Via (t),Vib (t),Vic (t))
   
∂ Via (t) −Vi (t)
=
∂Vib (t) t=k ∂Vib (t) Vi (t) t=k
 
∂Vi (t) ∂Vi (t)
[− ∂V (t) ][Vi (t)] − [Via (t) −Vi (t)][ ∂V ]
ib ib (t) 
= 2
Vi (t)
t=k
" #
1
−[Via (t)][ 3 ]
=
Vi2 (t)
t=k
 
[−3Via (k)]
= (D.6)
[Via (k) +Vib (k) +Vic (k)]2

∂ Fix (Via (t),Vib (t),Vic (t))

   
∂ Via (t) −Vi (t)
=
∂Vic (t) t=k ∂Vic (t) Vi (t) t=k
 
∂Vi (t) ∂Vi (t)
[− ∂V (t) ][Vi (t)] − [Via (t) −Vi (t)][ ∂V ]
ic ic (t) 
=
Vi2 (t)
t=k
" #
1
−[Via (t)][ 3 ]
=
Vi2 (t)
t=k
 
[−3Via (k)]
= (D.7)
[Via (k) +Vib (k) +Vic (k)]2

Replacing in (D.4)

 
3Via (k) − [Via (k) +Vib (k) +Vic (k)]
Fia (Via (t),Vib (t),Vic (t)) ≈ +
[Via (k) +Vib (k) +Vic (k)]
 
[3][Vib (k) +Vic (k)]
[Via (t) −Via (k)] +
[Via (k) +Vib (k) +Vic (k)]2
 
[−3Via (k)]
[Vib (t) −Vib (k)] +
[Via (k) +Vib (k) +Vic (k)]2
 
[−3Via (k)]
[Vic (t) −Vic (k)]
[Via (k) +Vib (k) +Vic (k)]2

Regarding inequations (D.3) for x = a, (D.8) has the linear representation defined by (D.9).

Via (t) −Vi (t)

≤ PVUR∗ (D.8)
Vi (t)

153
 Fia (k) + Kiaa (k) [Via (t) −Via (k)] + Kiab (k) [Vib (t) −Vib (k)] +
Kiac (k) [Vic (t) −Vic (k)] ≤ PVUR∗ (D.9)

where the coefficients produced in the linearisation are presented in (D.10).

 
3Via (k) − [Via (k) +Vib (k) +Vic (k)]
Fia (k) =
[Via (k) +Vib (k) +Vic (k)]
 
[3][Vib (k) +Vic (k)]
Kiaa (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[−3Via (k)]
Kiab (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[−3Via (k)]
Kiac (k) = (D.10)
[Via (k) +Vib (k) +Vic (k)]2

Similarly, solving for x = b and x = c:

Vib (t) −Vi (t)

≤ PVUR∗ (D.11)
Vi (t)

Vic (t) −Vi (t)

≤ PVUR∗ (D.12)
Vi (t)

Then

Fib (k) + Kiba (k) [Via (t) −Via (k)] + Kibb (k) [Vib (t) −Vib (k)] +
Kibc (k) [Vic (t) −Vic (k)] ≤ PVUR∗ (D.13)

Fic (k) + Kica (k) [Via (t) −Via (k)] + Kicb (k) [Vib (t) −Vib (k)] +
Kicc (k) [Vic (t) −Vic (k)] ≤ PVUR∗ (D.14)

Where the coefficients produced in the linearisation for phase b and phase c are presented in

154
(D.15) and (D.16), respectively.
 
3Vib (k) − [Via (k) +Vib (k) +Vic (k)]
Fib (k) =
[Via (k) +Vib (k) +Vic (k)]
 
[−3Vib (k)]
Kiba (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[3][Via (k) +Vic (k)]
Kibb (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[−3Vib (k)]
Kibc (k) = (D.15)
[Via (k) +Vib (k) +Vic (k)]2

 
3Vic (k) − [Via (k) +Vib (k) +Vic (k)]
Fic (k) =
[Via (k) +Vib (k) +Vic (k)]
 
[−3Vic (k)]
Kica (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[−3Vic (k)]
Kicb (k) =
[V (k) +Vib (k) +Vic (k)]2
 ia 
[3][Via (k) +Vib (k)]
Kicc (k) =
[Via (k) +Vib (k) +Vic (k)]2
(D.16)

Therefore, the prediction model for PVUR inequality constraints is defined by (D.17), which is
the same set of inequalities stated by (5.12). Note that the measured/estimated coefficients at t = k
are considered constant along the prediction horizon.

Fia (k)+Kiaa (k) [Via (k + m) −Via (k)] + Kiab (k) [Vib (k + m) −Vib (k)] +
Kiac (k) [Vic (k + m) −Vic (k)] ≤ PVUR∗ (D.17a)

Fib (k)+Kiba (k) [Via (k + m) −Via (k)] + Kibb (k) [Vib (k + m) −Vib (k)] +
Kibc (k) [Vic (k + m) −Vic (k)] ≤ PVUR∗ (D.17b)

Fic (k)+Kica (k) [Via (k + m) −Via (k)] + Kicb (k) [Vib (k + m) −Vib (k)] +
Kicc (k) [Vic (k + m) −Vic (k)] ≤ PVUR∗ (D.17c)

155
D.2 Apparent power rating
Apparent power is used as a constraint to ensure that each DG operates within its physical limits,
according to the inequality (D.18). As there is not a linear relationship among apparent, active and
reactive powers, it is required to linearise and discretise (D.18).

|Si (t)| = (Pi (t)2 + Qi (t)2 )1/2 < Smax (D.18)

Linearising |Si (t)| around the measured/estimated point

p(k) = {ωi (k), ω̂iB (k),Vi (k),Vix (k), V̂iB (k), V̂ixB (k), δ θi (k), Pi (k), Qi (k)}
with x={a, b, c} :

∂ |Si (t)| ∂ |S (t)|

|Si (t)| ≈ (Pi (k)2 + Qi (k)2 )1/2 + (Pi (t) − Pi (k)) + i (Qi (t) − Qi (k))
∂ Pi ∂ Qi
1
≈ (Pi (k)2 + Qi (k)2 )1/2 + (Pi (k)2 + Qi (k)2 )−1/2 (2Pi (k))(Pi (t) − Pi (k))
2
1
+ (Pi (k)2 + Qi (k)2 )−1/2 (2Qi (k))(Qi (t) − Qi (k)) (D.19)
2

Note that (D.19) is not feasible when Pi (k) = Qi (k) = 0; therefore, applying the triangular inequal-
ity, the polytopic inner approximation (D.20) is achieved. This approximation is feasible for any
operation point.

(Pi (t)2 + Qi (t)2 )1/2 < |Pi (t)| + |Qi (t)| = f (Pi , Qi ) (D.20)

Note that if the right side of inequality (D.20) is less than Smax , the left side will also be. There-
fore, linearising:

∂ f (Pi , Qi )
|Pi (t)| + |Qi (t)| ≈|Pi (k)| + |Qi (k)| + (Pi (t) − Pi (k))+
∂ Pi
∂ f (Pi , Qi )
(Qi (t) − Qi (k))
∂ Qi
≈|Pi (k)| + |Qi (k)| + sign(Pi (k))(Pi (t) − Pi (k))+
sign(Qi (k))(Qi (t) − Qi (k)) (D.21)

In this case the discretisation can be directly derived. Then, extending along the prediction hori-
zon (D.22) is obtained, which is the inequality (4.35) and (5.16). Note that the measured/estimated
coefficients at t = k are considered constant along the prediction horizon.

156
|Pi (k)| + |Qi (k)| + sign(Pi (k))[Pi (k + m) − Pi (k)]
+sign(Qi (k))[Qi (k + m) − Qi (k)] < Smax (D.22)

157
Annexes

Annexed E

Derivation of predictive linear models used

as equality constraints in dc DGs

E.1 Continuous time model for equality constraints

Equations (4.15) and (4.16) are rewritten as (E.1). As it was mentioned, (E.1) characterises the
voltage droop controller and the active power transferred from the i − th dc DG to the microgrid.

Vidc (t)=V0dc+Mpv,i Pidc (t)+Vs,idc (t) (E.1a)

 
Pidc (t) = GiVidc (t) Vidc (t)−V̂idc,B (t) (E.1b)

E.1.1 Droop equation

The linearisation process for the droop equation (E.1a) is described as follows.

From (E.1a), it is possible to rewrite Vs,idc (t) in function of ∆Vs,idc (t).

Vidc (t) = V0dc + Mpv,i Pidc (t) +Vs,idc (t) (E.2)

Z
dc dc dc 1
Vi (t) = V0 + Mpv,i Pi (t) + V dc (t)dt
∆ s,i
Tsec

158
Deriving both sides and applying the forward Euler method:

dVidc (t) dPidc (t) 1 dc

= Mpv,i + ∆V (t)
dt dt Tsec s,i

dVidc (t) dPidc (t) 1 dc

= Mpv,i + ∆V (k)
dt dt Tsec s,i
t=k t=k
h i
Vidc (k + 1) −Vidc (k) = Mpv,i Pidc (k + 1) − Pidc (k) + ∆Vs,idc (k)
h i
Vi (k + 1) = Vi (k) + Mpv,i Pi (k + 1) − Pi (k) + ∆Vs,idc (k)
dc dc dc dc
(E.3)

E.1.2 Power transfer equation

Due to power transfer, equation (D.1b), is non-linear, this is linearised via a Taylor expansion
around the measured/estimated point p(k) = {Vidc (k), V̂idc,B (k), Pidc (k)}. Then, the forward Euler
discretisation is applied to the linearised equation.

 
dc dc dcdc dc,B dc
Pi (t) = Vi (t)Ii (t) = GiVi (t) Vi (t) − V̂i (t) (E.4)

Linearising

∂ Pidc (t) ∂ Pidc (t)

Pidc (t) =Pidc (k) + [V dc
(t) −V dc
(k)] + [V̂idc,B (t) − V̂idc,B (k)]
∂Vidc p(k) i i
∂ V̂idc,B p(k)

Pidc (t) = Pidc (k) + KV [Vidc (t) −Vidc (k)] + KV̂ dc,B [V̂idc,B (t) − V̂idc,B (k)]
i

where h i
dc dc,B
KV = Gi 2Vi (k) − V̂i (k)

KV̂ dc,B = −GiVidc (k)

i

Deriving both sides and evaluating at t = k

Pidc (t) Pidc (k) d d

= + KV [Vidc (t) −Vidc (k)] + KV̂ dc,B [V̂idc,B (t) − V̂idc,B (k)]
dt dt dt i dt

dPidc (t) dVidc (t) dV̂idc,B (t)

= KV + KV̂ dc,B
dt dt i dt
t=k t=k t=k

159
 Pidc (k + 1) − Pidc (k) =KV [Vidc (k + 1) −Vidc (k)] + KV̂ dc,B [V̂idc,B (k + 1) − V̂idc,B (k)]
i

Assuming V̂idc,B is constant

V̂idc,B (k + 1) − V̂idc,B (k) = 0

Pidc (k + 1) = Pidc (k) + KV [Vidc (k + 1) −Vidc (k)]

Then

Pidc (k + 1) =Pidc (k) + Gi [2Vidc (k) − V̂idc,B (k)][Vidc (k + 1) −Vidc (k)] (E.5)

E.2 Prediction model for equality constraints

From (E.3) and (E.5), it is possible to state the predictive models (E.6). These models are used
as a set of equality constraints in the optimisation problem. Note that the measured/estimated
coefficients at t = k are considered constant along the prediction horizon.

h i
Vidc (k+m)=Vidc (k)+Mpv,i Pidc (k+m)−Pidc (k) +∆Vs,idc (k+m−1) (E.6a)

Pidc (k+m)=Pidc (k)+Gi [2Vidc (k)−V̂idc,B (k)][Vidc (k+m)−Vidc (k)] (E.6b)

160
Annexes

Annexed F

Experimental microgrid setup

This appendix presents the main characteristics of the experimental setup implemented to vali-
date the proposed DMPC of Chapter 3. This experimental setup was built at The Microgrid Control
Lab (Laboratorio de Control de Microrredes) of the University of Chile1 . The used equipment cor-
responds to power electronics industrial modules assembled and commercialised by the Triphase®
company, currently part of National Instruments.

The Triphase Distributed Power System (DPS) creates a scalable, adaptable, and open platform
for power system application rapid prototyping and power-hardware-in-the-loop testing. Ac mi-
crogrids, dc microgrids, motor drives, battery testing and storage, and many more uses are among
the target applications. The DPS product line also offers signal processors and software libraries
for real-time control, along with power amplifiers, power converters, and power measurements. In
addition, triphase solutions may be customised to the user’s requirements since they are open and
fully (re)programmable [140]. The Triphase units and the experimental microgrid are described as
follows.

The experimental setup of Fig. 3.3 comprises the PM15F120 and PM5F60 Triphase® power
modules. The internal components of the power modules PM15F120 and PM5F60 are presented
in Fig. F.1.a and Fig. F.1.b, respectively. As seen in Fig. F.1.a, the PM15F120 is composed of
4 VSCs (VSC1, VSC2, VSC3, VSC4). This unit allows emulating two independent three-phase
ac-DGs using VSC3 and VSC4. These ac-DGs are DG1 and DG2 in the experimental setup of
Section 3.6.1. In contrast, the PM5F60 is composed of 2 inverters (VSC1, VSC2) as shown in
Fig. F.1.b. This unit allows emulating one three-phase ac-DG using VSC2. Therefore, it is possible
to emulate a three-phase three-wire ac-microgrid composed of three ac-DGs. These power units

1 https://www.die.cl/sitio/home/investigacion/laboratorios/laboratorio-de-control-de-micro-redes/

161
have a real-time target (RTT) based on Unix as a control unit.

Figure F.1: Topology of Triphase units PM15F120C and PM5F60R used to emulate the ac-
microgrid

A conventional wired TCP/IP network is used to connect the engineering station and the RTT,
being possible to manage the RTTs from the same engineering station via Matlab/Simulink® in-
terface. In contrast, the microgrid setup possesses a dedicated communication network for infor-
mation sharing among DGs through an optical ring. Considering that this is a fast channel, several
communication phenomena can be emulated, and their impact on the microgrid performance can
be evaluated.

Regarding the topology of the MG presented in Fig. 3.3, the impedance lines are emulated
through inductors; thus, the resistive component of the lines is small and corresponds to the induc-
tance’s resistance. Moreover, the loads connected to the MG are resistive load banks. For more
information about all the equipment available at The Microgrid Control Lab of the University of
Chile, the interested reader is encouraged to review the conference paper [141], where a detailed
explanation is provided.

162