Delft University of Technology

Fast calculation of electrical transients in power systems after a change of topology

Thomas, Romain

DOI
10.4233/uuid:40d94dfa-75f8-4b72-b313-3e72c782b9f9
Publication date
2017
Document Version
Final published version
Citation (APA)
Thomas, R. (2017). Fast calculation of electrical transients in power systems after a change of topology.
https://doi.org/10.4233/uuid:40d94dfa-75f8-4b72-b313-3e72c782b9f9

Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.

Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.

This work is downloaded from Delft University of Technology.

For technical reasons the number of authors shown on this cover page is limited to a maximum of 10.
FAST CALCULATION OF ELECTRICAL TRANSIENTS IN
POWER SYSTEMS AFTER A CHANGE OF TOPOLOGY
FAST CALCULATION OF ELECTRICAL TRANSIENTS IN
POWER SYSTEMS AFTER A CHANGE OF TOPOLOGY

Proefschrift

ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus prof. ir. K.C.A.M. Luyben,
voorzitter van het College voor Promoties,
in het openbaar te verdedigen op donderdag 30 november 2017 om 10:00 uur

door

Romain T HOMAS

Diplôme d’Ingenieur en Génie Electrique et Automatique

Institut National Polytechnique de Toulouse
Ecole Nationale Supérieur d’Electrotechnique,
d’Electronique, d’Hydraulique et des Télécomucations,
Toulouse, France,
geboren te Quimper, Frankrijk
Dit proefschrift is goedgekeurd door de
Promotor: Prof. ir. L. van der Sluis
Promotor: Prof. dr. ir. C. Vuik
Copromotor: Dr. D.J.P. Lahaye

Samenstelling promotiecommissie:
Rector Magnificus, voorzitter
Prof. ir. L. van der Sluis, Technische Universiteit Delft
Prof. dr. ir. C. Vuik, Technische Universiteit Delft
Dr. D.J.P. Lahaye, Technische Universiteit Delft

Onafhankelijke leden:
Prof. dr. S. Schöps, Technische Universität Darmstadt
Prof. dr. ing. A.J.M. Pemen, Technische Universiteit Eindhoven
Prof. dr. ir. C.W. Oosterlee, CWI & Technische Universiteit Delft
Prof. dr. ir. H.X. Lin, Technische Universiteit Delft
Prof. dr. P. Palensky, Technische Universiteit Delft, reservelid

Keywords: Power system, Elctrical transient, Modeling nethods, Ordinary differ-

ential equations, Integration methods, Runge-Kutta methods, linear
solvers
Printed by: Ridderprint, the Netherlands
Front & Back: Tout au bout du fil, Anne et Florence Ramos & A venir, Florence Ramos

Copyright © 2017 by R. Thomas

ISBN 978-94-6299-791-2
An electronic version of this dissertation is available at
http://repository.tudelft.nl/.
Without mathematics, there’s nothing you can do.
Everything around you is mathematics.
Everything around you is numbers.
Shakuntala Devi
 C ONTENTS

Samenvatting xiii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Research objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Objectives for the modeling approach of a power system . . . . . . 7
1.3.2 Objectives for the numerical approach. . . . . . . . . . . . . . . . 7
1.4 Research approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Outline of the thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Electrical background 11
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Maxwell equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Main lumped elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Kirchhoff’s laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Kirchhoff’s current law . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Kirchhoff’s voltage law . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Norton’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Graph theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Modeling methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.7.1 Nodal analysis method. . . . . . . . . . . . . . . . . . . . . . . . 17
2.7.2 Modified nodal analysis method . . . . . . . . . . . . . . . . . . . 22
2.7.3 Cut-set method . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.8 Power system elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.8.1 Switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.2 Generator model . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.3 Load model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.8.4 PI-section model. . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.8.5 Double PI-section model . . . . . . . . . . . . . . . . . . . . . . 29
2.8.6 Arc models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Mathematical background 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Ordinary differential equations and differential algebraic equations . . . . 33
3.3 Numerical integration methods . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.1 Euler methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

vii
viii C ONTENTS

3.3.2 Trapezoidal rule . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.3.3 Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.1 Linear solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.2 Newton-Raphson method . . . . . . . . . . . . . . . . . . . . . . 46
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4 Switching action network calculation 51
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Block modeling method methodology . . . . . . . . . . . . . . . . . . . 53
4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . 54
4.2.3 Block model representation and parameters . . . . . . . . . . . . . 55
4.2.4 First example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.3 Connecting block model matrices . . . . . . . . . . . . . . . . . . . . . 58
4.3.1 First step: Full admittance matrix . . . . . . . . . . . . . . . . . . 60
4.3.2 Second step: Reduction of the full admittance matrix . . . . . . . . 61
4.3.3 Third step: Updating the connecting block matrices . . . . . . . . . 62
4.4 Example of an analytical solution of a switch block . . . . . . . . . . . . . 62
4.4.1 Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.4.2 Block-modeling solution . . . . . . . . . . . . . . . . . . . . . . . 63
4.5 Time-dependent and non-linear electrical networks . . . . . . . . . . . . 65
4.5.1 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . 65
4.5.2 Introduction to some time-dependent and non-linear electrical net-
works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5 Numerical analysis of the block modeling method 71
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2 Linear power systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.1 Description of test cases . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.2 Results and discussions . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Non-linear power systems . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3.1 Test cases presentation. . . . . . . . . . . . . . . . . . . . . . . . 75
5.3.2 Presentation of the results and discussion . . . . . . . . . . . . . . 76
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6 Inaccurate solver for the simulation of power systems 81
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Definition of problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.3 Approximate linear solver definition . . . . . . . . . . . . . . . . . . . . 82
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
C ONTENTS ix

7 Approximate solver for the simulations of power system 87

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.2 Test case presentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.3 Results and discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.3.1 Fixed time-step . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.3.2 Adaptive time-stepping strategy . . . . . . . . . . . . . . . . . . . 97
7.4 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8 Conclusion 103
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
8.2 Answer to the research question . . . . . . . . . . . . . . . . . . . . . . 103
8.3 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
8.4 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Acknowledgements 107
Curriculum Vitæ 109
List of Publications 111
 S UMMARY

A P ower system is composed of various components such as generators, transform-

ers, transmission lines, switching devices and loads. They have their mathematical
model and graphical representation. Sometimes, a power system’s change of topology
occurs due to events like short circuits, lightning striking a transformer, or a reconfigu-
ration of the transmission system.
In this thesis, a new way of simulating large scale power systems is presented from
the modeling point of view. In the literature, a lot of modeling methods and mathe-
matical tools are available to tackle this subject. However, this thesis mainly focuses on
the time domain simulation of large scale power systems - and in particular, transients
which appear after a change of topology.
A change of topology in electrical networks impact time domain simulations on two
levels. The first impact is that it is necessary to update or re-compute the set of equa-
tions. The computation time of this action on the topology can be significant - espe-
cially for large scale power systems. The second impact of this change of topology is the
transient that will occur. Usually, this change will impose to numerically compute fast
oscillations in currents and voltages until they reach a new steady state.
The novelty of this thesis is that it tackles the effect of switching action in terms of
the modeling approach and the mathematical tools that can be used during the time
integration method. For these reasons, the thesis focuses on two issues. The first is about
the modeling approach of electrical networks, when switching devices are included. The
second aspect examines the numerical solver for large power systems. The goal of this
thesis is to compute the overall time domain simulation of a large scale power system
faster.
The modeling approaches and the mathematical tools, which are needed to under-
stand the contributions of the thesis, are described in Chapters 2 and 3, respectively.
Chapter 2 demonstrates the basics of the electrical engineering theorems and modeling
approaches of electrical networks. In Chapter 3, mathematical definitions and tools are
defined. This literature review helps circumvent the problem of numerical simulation of
electrical powers when a switching action occurs.
In Chapter 4, a new approach for modeling power system is shown, called the block
modeling method. The methodology is based on the fact that mathematical descriptions
of every power system components are known. The connection between them is made
by a combination of switching devices.
Two physical properties of power systems need to be taken into account for this ap-
proach. The first property is that power system components have strayed capacitors
between a phase and the ground. They then help to connect the power system com-
ponents mathematically. However, Kirchoff’s voltage law makes the parallel connection
of capacitors a computational problem. For this reason, the switching devices are non-
ideal and they are represented by a resistance of small value when they are in a closed

xi
xii S UMMARY

position.
Finally, these properties makes the computation of the system of equations of the
electrical network possible. The voltage at the terminal of each power system compo-
nent is known,as is the topology of switching devices. Then, the computation of the
current through all switching devices can be effectuated. Those currents are then used
to calculate the derivative of the voltage across all strayed capacitors.
This chapter offers an introduction to non-linear elements in power systems. Non-
linear elements considered in this thesis are a change in function in time of a value of a
lumped element, such as an inductor whose inductance varies over time. This evolution
in time of a lumped element value does not introduce new differential variables. More-
over, the substitution of switching devices by arc models is shown too. This change adds
new non-linear differential equations. This section of the chapter subsequently focuses
on the computation of the Jacobian matrix.
Chapter 5 describes and discusses the utilization of the block modeling approach
on several linear and non-linear electrical networks. In the first instance, linear electri-
cal systems of different sizes, including several switching actions, are studied and com-
pared with an electrical power system software used for fast transient simulations. In the
second instance, non-linear electrical networks, where several arc models replace some
switching devices, are investigated. In particular, the effect on the overall computation
time with the way of calculating the Jacobian matrix is also investigated.
Usually, power systems are stiff due to the difference between the small and large
time constants due to the topology and the value of the different lumped elements.
Then, the implicit integration method is recommended. This fact then implies the solv-
ing of systems of equations where the computation time increases factorially with their
sizes. Chapter 6 presents the numerical study of the utilization of inexact solvers, instead
of direct or iterative solvers for the simulation of large stiff power systems.
Chapter 7 presents the different electrical networks of various sizes and stiffnesses.
These test cases allow the study of the effect of the sizes, stiffness, and time-stepping
strategy when inexact solvers are employed. This study focuses on the computation of
the first transient of the power system when no energies are involved, to the next steady
state. For this reason, there is no switching action included in the studies. However, the
block modeling method is applied to compute the system of differential equations.
Finally, Chapter 8 offers the conclusions and recommendations for future work.
 S AMENVATTING

E LEKTRICITEITSNETTEN bestaan uit verschillende componenten, zoals generatoren,

transformatoren, transmissielijnen, schakelapparaten en belastingen. Ze hebben
elk hun wiskundige model en grafische weergave. Soms kan de topologie van een elek-
triciteitsnet veranderen door gebeurtenissen zoals kortsluiting, blikseminslag in een
transformator of een her-configuratie van het transmissie en distributie systeem.
In dit proefschrift wordt een nieuwe manier gepresenteerd van het simuleren van
grootschalige elektriciteitsnetten vanuit een modelleringsbenadering. In de literatuur
zijn er veel modellen en wiskundige instrumenten beschikbaar om dit onderwerp aan
te pakken. Dit proefschrift richt zich echter vooral op de simulatie van tijdsdomeinen
van grootschalige elektriciteitsnetten - en in het bijzonder transiënten die verschijnen
na een verandering van topologie.
Een verandering van topologie in elektrische netwerken beïnvloedt tijdsdomeinsim-
ulaties op twee niveaus. De eerste impact is dat het nodig is om de set vergelijkingen bij
te werken of opnieuw te berekenen. De berekeningstijd van deze actie op de topologie
kan significant zijn - vooral voor grootschalige elektriciteitsnetten. De tweede impact
van deze verandering van topologie is de transient die zal optreden. Gewoonlijk zal deze
wijziging het noodzakelijk maken om snelle oscillaties in stromen en spanningen nu-
meriek te berekenen tot ze een nieuwe stabiele toestand bereiken.
Het vernieuwende van dit proefschrift is dat de thesis het effect van wisselwerking
aanpakt tussen de modelleringsbenadering en de wiskundige hulpmiddelen die tijdens
de tijdintegratiemethode kunnen worden gebruikt. Daarom richt het proefschrift zich
op twee problemen. De eerste gaat over de modelleringsbenadering van elektrische
netwerken wanneer schakelaars worden meegenomen. Het tweede aspect onderzoekt
de numerieke oplosser voor grote netten. Het doel van dit proefschrift is om de algehele
tijddomein simulatie van een grootschalige elektriciteitsnetten sneller te berekenen.
De modelbenaderingen en de wiskundige hulpmiddelen die nodig zijn om de bijdra-
gen van het proefschrift te begrijpen, worden beschreven in respectievelijk hoofdstuk 2
en 3. Hoofdstuk 2 toont de basis van de elektrotechnische theorieën en modelleringsbe-
naderingen van elektrische netwerken. In hoofdstuk 3 worden wiskundige begrippen en
gereedschappen gedefinieerd.
In hoofdstuk 4 wordt een nieuwe benadering voor het modelleren van het elek-
triciteitsnetten getoond, genaamd de blokmodelleringsmethode. De methodologie is
gebaseerd op het feit dat wiskundige beschrijvingen van alle componenten van het elek-
triciteitsnet bekend zijn. De verbinding tussen de blokken wordt gemaakt door een com-
binatie van schakelaars.
Voor deze aanpak moet rekening worden gehouden met twee fysieke eigenschappen
van het elektriciteitsnet. De eerste eigenschap is dat componenten van het elektriciteit-
snet altijd een zekere capaciteit hebben tussen een fase en de grond. Zij helpen dan
mathematisch de componenten te verbinden. De spanningswet van Kirchhoff verbiedt

xiii
xiv S AMENVATTING

parallelle aansluiting van condensatoren. Om deze reden worden de schakelaars niet

als ideaal gezien en worden ze gerepresenteerd door een weerstand van kleine waarde
wanneer ze in een gesloten positie staan.
Tenslotte maken deze eigenschappen de berekening van het stelsel van vergeli-
jkingen van het elektrische netwerk mogelijk. De spanning aan de klemmen van elke
netwerk component is bekend, evenals de topologie van schakelaars. Vervolgens kan
de berekening van de stroom door alle schakelapparaten worden bewerkstelligd. Deze
stromen worden dan gebruikt om de afgeleide van de spanning te berekenen over alle
capaciteiten.
Dit hoofdstuk geeft voorts een inleiding tot niet-lineaire elementen in elektriciteit-
snetten. Niet-lineaire elementen die in dit proefschrift worden beschouwd zijn een
verandering van waarde in de tijd van een netwerk element, zoals een spoel waarvan
de inductiviteit in de loop van de tijd varieert. Verder wordt ook de vervanging van
schakelaars door boogmodellenbesproken. Deze voegen nieuwe niet-lineaire differ-
entiaalvergelijkingen toe. Dit gedeelte van het hoofdstuk richt zich vervolgens op de
berekening van de Jacobiaanse matrix.
Hoofdstuk 5 beschrijft het gebruik van de blokmodelleringsbenadering op verschil-
lende lineaire en niet-lineaire elektrische netwerken. Op de eerste plaats worden lineaire
elektrische netwerken van verschillende grootte en diverse schakelacties bestudeerd en
vergeleken bestaande software die wordt gebruikt voor snelle transiënte simulaties. Op
de tweede plaats worden niet-lineaire elektrische netwerken onderzocht, waar verschil-
lende boogmodellen de schakelaars vervangen. In het bijzonder wordt ook de invloed op
de totale berekeningstijd van de berekeningsmethode voor de Jacobiaanse matrix onder-
zocht.
Meestal zijn elektriciteitsnetten wiskundig beschouwd stijf vanwege het verschil
tussen de kleine en grote tijdsconstanten die bepaald worden door het netwerk en de
waarde van de verschillende Netwerkelementen. Dan wordt de impliciete integratie
methode aanbevolen. Dit feit betekent dat het oplossen van systemen van vergelijkingen
waar de berekeningstijd een factor met hun grootten toeneemt. Hoofdstuk 6 presenteert
de numerieke studie van het gebruik van niet exacte oplossers, in plaats van directe of
iteratieve oplossers voor de simulatie van grote stijve elektriciteitsnetwerken.
Hoofdstuk 7 presenteert de verschillende elektrische netwerken van verschillende
groote en stijfheid. Deze testcases maken het mogelijk om het effect op de grootte sti-
jfheid en tijdstapstrategie te bestuderen wanneer inexacte oplossers worden gebruikt.
Ten slotte geeft hoofdstuk 8 de conclusies en aanbevelingen voor toekomstig onder-
zoek op het terrein van de berekening van snelle verschijnselen in hoogspanningsnetten
 1
I NTRODUCTION

1.1. M OTIVATION
O N September 4, 1882, in the South of Manhattan, New York City, United States of
America, the inhabitants and workers illuminated their households and offices via
electrical lamps for the first time [1, 2]. This huge project was led by T. Edison. At the
beginning of the project, a study had shown two main advantages of using electricity
for lighting. The first one was economical because an electrical lamp costed lesser than
candles or gas to illuminate a room. The second advantage was the brightness - electrical
lights had a higher luminosity than flames, and they significantly increased the comfort
for the eyes.
The power plant to produce the electrical power imagined by T. Edison was initially
composed of a single generator. Then later, with the expansion of the electrical network
and more interconnected households and buildings, five new generators were added.
On the inauguration day, T. Edison described the lifetime of the generators as: "They will
go on forever unless stopped by an earthquake" [3]. At that time, the scientific community
believed that electricity might impact the creation of earthquakes [4].
Since that day, power systems permit the production, transportation and supply of
electrical energy [5, 6]. For example, in T. Edison’s power system design, the genera-
tors provided the electrical power. Then, this energy was transported via an electrical
network to the different light bulbs, and finally, these lamps transformed this energy
into light. This design is still applied in today’s power systems, and are composed of
power plants that produce electrical power. This electrical power is distributed to the
consumers via an interconnection of overhead transmission lines and underground ca-
bles.
In order to produce electrical power, power plants need a primary source of energy -
such as a dam (hydro power plants), gas, oil and coal (thermal power plants), or uranium
(nuclear power plants). They respectively contributed 48.8%, 12.8%, and 28.8% of the
European Union’s production in 2013, which was 3.10 million Gigawatt-hours (GWh).
Finally, the remaining 9.6% electric power produced was due to wind, solar, and geother-
mal primary energies (renewable energy)[7].

1
 2 1. I NTRODUCTION

The energy is transported and distributed to the consumers via overhead transmis-
1 sion lines, underground cables, and transformers [8]. The distance between power
plants and consumers, and the power to be consumed, determines the design of the
transmission and distribution system. The transmission system corresponds to the net-
work between the power plants and the electrical sub-stations near the cities. Then,
the distribution system corresponds to the network between these sub-stations and the
households. The difference between the transmission and distribution system is the
level of voltage needed to reduce the losses during the long distance transportation of
energy. Transformers are used to change the voltage level between both systems.
There are around 552 million consumers in Europe. In 2013, the total consumption
of electric power in the European Union was 2.77 million GWh [9]. The European Union
defined three groups of users: industrial organizations, which consumed 1 million GWh,
the transportation systems (e.g. trains), which consumed 0.66 million GWh, and third,
private households, administrations, and commerce - which used 1.22 million GWh that
year.
In general, a power system can be divided into two separate entities: the electric util-
ity, which provides the electrical power and usually owns a distribution network, and the
transmission system operator, which manages the transmission system. It coordinates
the production and the demand for electric energy with the different electric utilities
and other transmission system operators in real-time, and also actively participates in
the electrical market. Finally, a power system operator conducts a multitude of projects
to ensure the transportation of energy for the future.
In order to play its role, the transmission system operator needs a lot of informa-
tion. First, it requires a complete and accurate model of its electrical network. Then, it
needs to collect data to predict the consumption and the production of energy along the
power system. With all this information, it uses a supervisory control and data acquisi-
tion which monitors and manages the whole power system in real-time [10]. Besides, it
utilizes this information for its research for its future transmission system. Finally, the
transmission system operator works on different time scales, as shown in Figure 1.1.

Planning

Maintenance

Real-time
control
Slow
transients
Fast
transients
µs

m
s

y
ur

ar
th
m

da
ho

ye
on
m

Time

Figure 1.1: The power system events that occur, and their time frame

According to Figure 1.1, various actions that could occur on the electrical network
have different time scales (from microseconds to several years). The definitions of these
acts are:

• Planning: Consists of designing future electrical networks. Usually, the design is

1.1. M OTIVATION 3

defined in collaboration with the electric utilities and their view on the future con-
sumption of electrical energy; 1
• Maintenance: Permits the change of components that malfunction, or have
reached the end of their life, for a better sustainability of the power system;

• Real-time control: Allows the control and operation of the network, from the data
along the network, in order to reduce losses and increase the efficiency of the sys-
tem;

• Slow transients: For example, after a short-circuit, the recovery of the optimal fre-
quency and voltages can be slow according to the power system’s topology;

• Fast transients: For instance, just after clearing a short-circuit via a circuit breaker,
a very rapid increase of voltage can occur between the extremities of the circuit
breaker.

The study of fast transients is necessary for the sustainability of the power system, and
it is effectuated during the planning stage. For example, the design of circuit-breakers,
transmission lines, and cables depends on the instantaneous maximal currents and volt-
ages that they can support [11]. Then, the evolution in time of currents and voltages can
be obtained by a time domain simulation for a better sustainability of the electrical net-
work. For these reasons, this thesis focuses on those phenomena.
For the most accurate time domain simulation, the mathematical model of the power
system should be as precise as possible. In order to obtain the mathematical expression,
Maxwell’s equations can be applied [12]. However, this method requires a comprehen-
sive understanding of the physics involved, and a significant computational power. For
this reason, the circuit theory has been developed for de-complexing Maxwell’s equa-
tions and predicting the electrical behavior of an electrical circuit [13].
When the circuit theory is applied to a power system, the simplest model consists of
using lumped elements (e.g. resistances, inductances, capacitances, and sources). Addi-
tionally, every lumped element has its own graphical representation and mathematical
description [14]. Their electrical behavior depends on a relation between the current
through them and the voltage across them [15]. Besides, this relationship can be linear
or non-linear. For example, the utilization of an arc model to represent the interruption
of current in a circuit breaker introduces non-linear equations [16].
The circuit theory consists of a multitude of laws and theorems [17–20]. The most
useful rules are Kirchhoff’s laws [17]. Kirchhoff’s first law concerns the current interac-
tions of different, lumped elements that are connected. Kirchhoff’s second law repre-
sents the voltage interaction between lumped elements. Then, when they are applied,
the network equations are found. The Norton-Thevenin theorem [18, 19] or Millman
theorem [20] can be used to either simplify the graph or find the voltage at a single point.
From these laws and theories, several methods are available to model any power sys-
tems based on circuit theory. Traditional modeling approaches were developed 50 years
ago for the time domain simulation. These methods are the nodal analysis [21], the mod-
ified nodal analysis [22], and the cut set methods [23]. A brief description of these meth-
ods is given below:
 4 1. I NTRODUCTION

• After applying the nodal analysis method, the network equations are found and
1 they can be solved easily. This method is based on Kirchhoff’s current law. More-
over, this approach imposes a fixed time-step. Finally, simulations of large scale
power systems use this algorithm;

• The modified nodal analysis method uses Kirchhoff’s current and voltage laws to
obtain a set of differential algebraic equations. Then, this method allows the split-
ting of the time interval with various time-steps. However, this approach is re-
stricted in terms of an algorithm to obtain the time domain solution due to its
mathematical formulation;

• A set of ordinary differential equations is obtained by the use of the cut-set

method. Unlike the modified nodal analysis, it allows the use of any algorithm
in order to obtain the time domain solution. However, with this approach, the
methodology to get the mathematical representation of a power system grows
faster than its size .

Finally after a change of topology, these modeling methods require the updation of the
whole system of equations.
For example, a change of topology can happen when a switching action occurs [24].
Then, the system of equations needs to be updated. Also, this change of topology may
create very fast oscillations in voltages and currents across the power system [11]. These
oscillations can have a frequency of up to 10MHz, and decrease over time to reach a new
steady state. Consequently, a switching action requires a lot of computational resources
- first, to update the system of equations, and second, to compute the time signals of the
high-frequency oscillations.
Now, the power system equations are known and the time domain simulation is re-
quired. For the modified nodal analysis method and the cut set method, two methods
are available to compute the time evolution of the current and voltage across the electri-
cal network. The first method is the analytical solution; however, this method only works
on linear power systems. The second method is the numerical approach, and it can be
utilized for any power system.
The analytical solution of a set of linear differential equations can be calculated by
computing the time constant or by using the Laplace transform and inverse Laplace
transform [25–27]. These two methods are computationally expensive for large sets of
linear equations. As a consequence, they are not recommended for the study of large
power systems.
Numerical integration methods - such as the Euler methods, the Trapezoidal rule,
or the Runge-Kutta methods - are applied to obtain the time domain solution [28–30].
These methods are usually less time consuming than analytical approaches, especially
for large power systems. However, the solution is an approximation of the analytical
solution and is based on the Taylor expansion [31]. Finally, this estimation depends on
the properties of the numerical integration method applied.
The time domain solution of the numerical integration method depends mainly on:

• The time-step used:

1.1. M OTIVATION 5

The choice of the time-step is important for the numerical solution. On the one
hand, it influences the accuracy of the simulation, and on the other , it is the dom- 1
inant factor for the computation time, since the computation time of an overall
simulation depends on the number of steps. Then, when the time-step is small,
the computation time is larger than when a bigger time-step is applied;

• The type of the numerical integration method:

There are two types of methods - explicit and implicit. Explicit methods have a
quick process of calculation, while implicit methods require more computational
power per time-step. In the case of a stiff system of equations, an implicit method
- in general - allows a larger time-step and can be more accurate than an explicit
method;

• The order of the numerical integration method:

It corresponds to the higher derivative of the Taylor expansion. Then, the order
also plays a role in the error at each time-step, which is proportional to the time-
step power of the order of the method.

From the characteristics of a power system, the choice of a numerical integration

method can be made. First, it should not have a high order, because it requires a large
computation time and not a small order due to the error made at each time-step - there-
fore, second-, third-, or fourth-order methods are often chosen. Second, power systems
are stiff. This results in high-frequency oscillations created after a switching action, and
consequently, implicit methods are recommended. Third, some numerical integration
methods allow adaptive time-stepping algorithms[29]. They change the time-step ac-
cording to the change in stiffness of the variables to integrate.
Implicit methods require more computation time per time-step than explicit meth-
ods. This increase in complexity is due to the linear systems of equations to solve in each
time-step[30]. In the literature, two types of solvers are used: linear solvers for linear sys-
tems of equations, and non-linear solvers for non-linear systems of equations [32, 33].
Unless they require a large computational power, they always converge to the steady-
state for any time-step applied.
The category of the linear solver is composed of two subcategories - direct or iterative
methods [34]. The direct methods are derived from the Gaussian elimination, such as
LU or Cholesky factorizations. Small systems of equations usually require such solvers.
Otherwise, iterative solvers are applied, like GMRES [35]- which is a Krylov subspace
method [32].
Non-linear solvers are mainly based on the Newton-Raphson iteration process[33].
They permit a convergence to the solution after several iterations when, during the pro-
cess, a linear system of equations has to be solved. Therefore, linear solvers are used.
Also, some features have to be taken into account to speed up the computation. The
first feature is to write the system of equations in a matrix form. This form contains many
zeros; then, the sparse notation can be employed [36]. Folowing this, only non-zero
elements and their indexes are stored. Finally, the numerical arithmetic is faster. The
second useful feature is to use a re-ordering method such as the Approximate Minimum
 6 1. I NTRODUCTION

Degree (AMD) method, for example[37]. It permits a reduction in the number of non-
1 zero elements during the factorization. As a consequence, it decreases the computation
time needed to solve the system of equations.
Presently, most new studies are looking into hardware or co-simulation [38, 39] to
overcome the computational challenge. Present-day hardware studies delve into the
use of graphical processor units (GPUs). Co-simulation studies instead consider using a
multitude of computers to perform a single simulation.
The utilization of GPU typically gives a relative speed-up for large power systems
under consideration, because arithmetic operations such as matrix-vector operations
can be easily parallelized[38]. However, there usually are one or several linear systems
of equations to solve at each time-step. Furthermore, it is used for very large power
systems.
The co-simulation of power systems could also give impressive results by using a
multitude of computers[39]. However, the work of matching results at the interface be-
tween computers can lead to a numerical instability and inaccuracy of the time domain
solution.
The computation time of large power systems is a challenge. The computation time
is related to their size. Besides, their size increases every day. A lot of studies are per-
formed on this topic because faster simulations imply that more simulations can be
made [40–42]. Then, new developments can be achieved quickly.
The computational challenge is not posed only to use of numerical integration
method and the solver. It is a combination of the previous facts as well as the method
to compute the set of equations. Today, with the increased of power electronics due to
the utilization of renewable energy - and therefore - switching actions, new modeling
methods which tackle these problems have to be developed.

1.2. P ROBLEM DEFINITION

As explained in the previous section, modeling methods rely on algorithms developed
50 years ago. After a switching action, it becomes computationally expensive due to the
change of topology. As a consequence, the first research question of this thesis to be
answered is:
Is it possible to design a new methodology to model a power system which allows
the use of adaptive time-stepping strategy with a relatively small computation time
for re-computing or updating the network equations in the case of a switching event
within an ordinary differential equation formulation?
The second research question is related to the association of the linear solver and
the numerical integration method. Usually, during the integration process, solving the
system of equations takes the most computation time. As a consequence, the time do-
main solution computation depends on the size, stiffness, tolerances, and topology of
the power system under consideration. Then, the second research question of this the-
sis to be answered is:
How strong is the interaction between the time loop of a numerical, semi-explicit,
diagonally implicit integration method of an ordinary differential equation and the
solver for obtaining the time domain solution of a power system?
1.3. R ESEARCH OBJECTIVES 7

1.3. R ESEARCH OBJECTIVES 1

The research objectives divide the previous research questions into two categories - the
modeling approach of a power system, and the numerical integration aspect of the time
domain solution. Finally, the following sections describe the two research objectives.

1.3.1. O BJECTIVES FOR THE MODELING APPROACH OF A POWER SYSTEM

• To review the state-of-the-art modeling methods for power system simulations,
with a focus on traditional approaches;

• To develop a new approach to view a power system, especially in terms of ordinary

differential equations when a switching action occurs;

• To be able to replace switching devices by non-linear elements such as arc models,

for which the Jacobian matrix is easy to assemble.

1.3.2. O BJECTIVES FOR THE NUMERICAL APPROACH

• To review the state-of-the-art numerical tools, such as integration methods and
linear solvers, which could be used for the simulation of power systems;

• To analyze the use of inaccurate linear solvers for the implicit numerical integra-
tion of ordinary differential equations for power system simulations.

1.4. R ESEARCH APPROACH

To satisfy the research problems and objectives, the following research approach has
been used. In the beginning, a literature study on the modeling approaches and nu-
merical analysis methods is performed. This survey reveals the actual advantages and
drawbacks of the various modeling methods and mathematical tools being used.
Then from the literature review, a new approach to model power systems is devel-
oped. This process benefits from the different methods - without their inconveniences.
Then, the implementation of non-linear elements, such as arc models, is considered for
this method. Finally, for testing purposes, various power systems are studied.
As explained previously, a large number of numerical integration methods are avail-
able. As a consequence, the relation between solvers and numerical integration method
is investigated. Then, this relationship provides the possibility of using special linear
solver, instead of regular solvers, for the study of large linear power systems without
compromising the accuracy of the time domain solution.

1.5. O UTLINE OF THE THESIS

The thesis chapters are defined as follows:

• Chapter 2 - This chapter contains the literature review of the state-of-the-art, from
the circuit theory to the principal modeling method (the nodal analysis, the mod-
ified nodal analysis, the cut-set method). Moreover, it also contains some power
system components.
 8 R EFERENCES

• Chapter 3 - This chapter presents the different notions of the mathematical anal-
1 ysis of this thesis. A brief introduction to the ordinary differential equations is
given. Afterwards, basic integration methods are illustrated. Finally, introductions
to solve systems of equations are shown.

• Chapter 4 - A new approach to model power systems is presented in this chapter,

which is later refered to as the block modeling method. First, a detailed demon-
stration is given, and then, a simple example is studied. Finally, the utilization of
arc models is described for the analytical Jacobian matrix computation.

• Chapter 5 - Various different power systems resulting from the block modeling
method are studied. Both linear and non-linear power systems are discuted.

• Chapter 6 - The use of inaccurate linear solvers during the numerical integration of
power systems is investigated in this chapter. Then, advantages and consequences
are drawn.

• Chapter 7- Here, several power systems of different sizes and stiffness are inves-
tigated when approximate solvers are used during the integration process. The
study employs a fixed and adaptive time-step.

• Chapter 8 - The main conclusions, contributions, and recommendations for fur-

ther research are offered in this chapter.

R EFERENCES
[1] Edison’s light turned on, downtown building supplied from the sation in Pearl street,
The Sun , New York edition (Sep 5, 1882).

[2] The Edison electric light at last, The Brooklyn Daily Eagle (Sep 5, 1882).

[3] Miscellaneous city news; Edison’s electric light. the times building illuminated by
electricity. The New York Times (Sep 5, 1882).

[4] Earthquakes, The New York Times (Oct 2, 1882).

[5] P. Schavemaker and L. van der Sluis, Electrical Power System Essentials (John Wiley
and Sons, Chichester, 2008).

[6] C. C. Antonio Gomez-Exposito, Antonio J . Conejo, Electric Energy Systems, Analysis

and Operation (CRC Press city, Boca Raton, 2008).

[7] Eurostat Statistics Explained, Electricity production, consumption and market

overview, (2016).

[8] S. A. Nasar and F. Trutt, Electric Power Systems (CRC Press, Boca Raton, 1998).

[9] Eurostat, Electricity consumption by industry, transport activities and house-

holds/services (gwh), (2009-2014).
R EFERENCES 9

[10] B. Theraja and A. Theraja, Text Book of Electrical Technology: Volume 3: Transmis-
sion, Distribution and Utilization (S. Chand & Company Ltd, New Delhi, 2013). 1
[11] L. van der Sluis, Transients in Power Systems (John Wiley & Sons, Chichester, 2001).

[12] J. C. Maxwell, A dynamical theory of the electromagnetic field, Philosophical Trans-

actions of the Royal Society of London 155, 459 (1865).

[13] R. DeCarlo and P.-M. Lin, Linear Circuit Analysis, 2nd ed. (OXFORD UNIVERSITY
PRESS, New York, 2001).

[14] L. Chua, C. Deseor, and E. Kuh, Linear and nonlinear circuits (McGraw-Hill Book
Company, New York, 1987).

[15] B. Theraja and A. Theraja, Text Book of Electrical Technology: Volume 1: Basic Elec-
trical Engineering in S.I. Units (S. Chand & Company Ltd, New Delhi, 2004).

[16] M. Kapetanovic, High voltage circuit breakers (Falculty of Electrotechnical Engi-

neering, Sarajevo, Sarajevo, 2011).

[17] G. Kirchhoff, Ueber den durchgang eines elektrischen stromes durch eine ebene, ins-
besondere durch eine kreisförmige, Annalen der Physik 140, 497 (1845).

[18] E.L.Norton, Design of finite networks for uniform frequency characteristic.Technical

Report TM26-0-1860, Tech. Rep. (Bell Laboratories, 1926).

[19] L. Thévenin, Sur un nouveau théoreme d’électricité dynamique [on a new theorem of
dynamic electricity], CR des Séances de l’Académie des Sciences 97, 159 (1883).

[20] J. Millman, A useful network theorem, proceeding of the IRE 28, 413 (1940).

[21] H. W. Dommel, Digital computer solution of electromagnetic transients in single-

and multiphase networks, Power Apparatus and Systems, IEEE Transactions on
PAS-88, 388 (1969).

[22] C. W. Ho, A. E. Ruehli, and P. A. Brennan, The modified nodal approach to network
analysis, Circuits and Systems, IEEE Transactions on 22, 504 (1975).

[23] E. Kuh and R. Rohrer, The state-variable approach to network analysis, Proceedings
of the IEEE 53, 672 (1965).

[24] R. Smeets, L. Van der Sluis, M. Kapetanovic, D. F. Peelo, and A. Janssen, Switching in
electrical transmission and distribution systems (John Wiley & Sons city, Chichester,
2014).

[25] J. Vlach and K. Singhal, Computer methods for circuit analysis and design (Van Nos-
trand Reinhold, New York, 1983).

[26] P. Dyke, Convolution and the solution of ordinary differential equations, in An Intro-
duction to Laplace Transforms and Fourier Series (Springer city, 2014) pp. 39–82.

[27] J. Stewart, Calculus: early transcendentals (Cengage Learning, Belmont, 2015).

 10 R EFERENCES

[28] U. M. Ascher and L. R. Petzold, Computer Methods for Ordinary Differential Equa-
1 tions and Differential-Algebraic Equations, 1st ed. (Society for Industrial and Ap-
plied Mathematics, Philadelphia, 1998).

[29] W. Hundsdorfer and J. Verwer, Numerical Solution of Time-Dependent Advection-

Diffusion-Reaction Equations (Springer, New York, 2003).

[30] K. Brenan, S. Campbell, and L. Petzold, Numerical Solution of Initial-Value Prob-

lems in Differential-Algebraic Equation (North-holland, Amsterdam, 1989).

[31] P. D. Lax and M. S. Terrell, Calculus with applications (Springer, New York, 2014).

[32] Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd ed. (Society for Industrial
and Applied Mathematics, Philadelphia, 2003).

[33] J. M. Ortega and W. C. Rheinboldt, Iterative solution of nonlinear equations in sev-

eral variables, Vol. 30 (Society for Industrial and Applied Mathematics, New York,
1970).

[34] W. Hackbusch, Iterative solution of large sparse systems of equations, Vol. 95

(Springer Science & Business Media, New York, 2012).

[35] Y. Saad and M. H. Schultz, GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems, Society for Industrial and Applied Mathemat-
ics: Journal on Scientific and Statistical Computing 7, 856 (1986).

[36] Å. Björck, Numerical methods in matrix computations (Springer, Cham, 2015).

[37] P. R. Amestoy, T. A. Davis, and I. S. Duff, An approximate minimum degree ordering

algorithm, SIAM Journal on Matrix Analysis and Applications 17, 886 (1996).

[38] Z. Zhou and V. Dinavahi, Parallel massive-thread electromagnetic transient simula-

tion on gpu, IEEE Transactions on Power Delivery 29, 1045 (2014).

[39] B. Zhang, W. Deng, L. Ruan, T. Wang, J. Quan, Q. Cao, Y. Teng, W. Wang, Y. Yuan, and
L. Li, Circuit cosimulation of 500-kv transformers in ac/dc hybrid power grid, IEEE
Transactions on Applied Superconductivity 26, 1 (2016).

[40] F. Wang and M. Yang, Fast electromagnetic transient simulation for over-voltages of
transmission line by high order radau method and v-transformation, IET Genera-
tion, Transmission Distribution 10, 3639 (2016).

[41] S. Chiniforoosh, H. Atighechi, and J. Jatskevich, A generalized methodology for dy-

namic average modeling of high-pulse-count rectifiers in transient simulation pro-
grams, IEEE Transactions on Energy Conversion 31, 228 (2016).

[42] A. A. van der Meer, M. Gibescu, M. A. M. M. van der Meijden, W. L. Kling, and
J. A. Ferreira, Advanced hybrid transient stability and emt simulation for vsc-hvdc
systems, IEEE Transactions on Power Delivery 30, 1057 (2015).
 2
E LECTRICAL BACKGROUND

2.1. I NTRODUCTION
E LECTRICAL engineering is a complex and broad subject. This field discusses topics
ranging from the interaction between magnetic and electromagnetic fields to the
control of power systems. It also incorporates the understanding of electronic compo-
nents and computer sciences. As specified in the previous chapter, this thesis focuses on
the time domain solution of power systems. Thus, only the basics of electrical engineer-
ing are necessary.
To simulate a power system, its equations need to be computed. As a consequence,
the fundamental idea of modeling power systems is to apply the circuit theory that sim-
plifies the Maxwell equations [1, 2]. Next according to the circuit theory, several ap-
proaches to model power systems are available in the literature. These approaches are
the nodal analysis method [3], the modified nodal analysis method [4] and, the cut-set
method [5]. The studies of large power systems usually use the nodal analysis method,
with the two other modeling methods being used less commonly for such computation.
Other methods, such as co-simulation methods, are available [6]. However, they are not
studied in this thesis.
The methodology applied to simplify the Maxwell equations is shown in Section 2.2.
Then, a brief description of the various lumped elements is given in Section 2.3. Next, in
Section 2.4, Kirchhoff’s laws are demonstrated under some assumptions. Then, Norton’s
theorem and an introduction to the graph theory applied to circuit theory are treated in
Section 2.5 and Section 2.6. Finally, the various approaches to model power systems are
described in Section 2.7, along with some power system elements in Section 2.8. Sec-
tion 2.9 provides information about the drawbacks, and the advantages of the modeling
approaches.

2.2. M AXWELL EQUATIONS

The goal of the Maxwell equations, which were developed by J.C. Maxwell, is to predict
the electrical behavior of voltages and currents of an electrical circuit [1]. The displace-

11
 12 2. E LECTRICAL BACKGROUND

ment of electrical charges creates voltages and currents, and by consequence, electrical
and magnetic fields. This interaction between them is quantified by the Maxwell equa-
tions, as shown in Table 2.1.

Integral equations Differential equations

2 Gauss’ law
Ò
∂ΩÒDd S =
Ð
ρdV 5·D = ρ
Gauss’ law for magnetism ∂Ω B d S = 0 5·B = 0
Maxwell-Faraday equation l = − ddt Σ B d S
∂Σ E dÎ 5 × E = − ∂B
H Î
∂t
d
Ampere’s circuital law ∂Σ H d l = Σ J d S + d t Σ Dd S 5 × H = J + ∂D
H Î
∂t

Table 2.1: Maxwell equations

Equations of Table 2.1 are related to each other by additional mathematical formula-
tions about the constitution of the different media, as summarized in Table 2.2.

Dielectric medium Magnetic medium conductor

D = ²E B = µH E = ρe J
J = σE

Table 2.2: Media’s constitutive equations

Then, Figure 2.1 illustrates the schematic representation of the Maxwell equations
from Tables 2.1 and 2.2.

5× µ
∂t D + J H B
Ampere

∂t ρe ∂t

² −5×
D E ∂t B
Faraday

Figure 2.1: Representation of the Maxwell’s equations

The wavelength (λ) of a periodical electrical wave (periodic signal) is defined such as
[7]:

v
λ= , (2.1)
f
where v is the speed of light (v ≈ 300.000.000m/s) and f the frequency of the signal [Hz].
The wavelength sheds light on the possible utilization of lumped elements (e.g. re-
sistances) to model the electrical circuit instead of the Maxwell equations (Table 2.1). In
order to use lumped elements, the size of the network studied strictly needs to be smaller
than λ. Consequently, only the left side and right side of Figure 2.1 are taken into account
with lumped elements, without interaction between them.
2.3. M AIN LUMPED ELEMENTS 13

2.3. M AIN LUMPED ELEMENTS

Lumped elements used in this thesis have two terminals. The current entering through
one terminal is the same as the current leaving the other terminal [2]. Their physical
dimensions are small compared to the wavelength of the frequencies that excite them.
A mathematical relation exists between the voltage at their terminals and the current
through the lumped element. These relations can be extrapolated from the Maxwell
2
equations (Table 2.1) and the medium constitution (Table 2.2). Each type of lumped el-
ement has its own graphical representation. Then, when several lumped elements con-
nected constitute an electrical diagram, the interconnection point of their terminals is
called a node. Table 2.3 shows the most common lumped elements.

Lumped Time domain

Symbol
element equations
R
i R (t )
Resistance v R (t ) = Ri R (t )
v R (t )
L
i L (t )
Inductance v L (t ) = L dditL
v L (t )
C
i C (t )
d vC
Capacitance i C (t ) = C dt

vC (t )
e(t )
Voltage source i e(t )

j (t )
Current source j (t )

Table 2.3: Lumped elements characteristics

Finally, another important lumped element is the ground (Figure 2.2). It symbolizes
the 0V reference. An electrical network can have several ground symbols. Moreover, they
are all connected together.

Figure 2.2: Ground symbol

2.4. K IRCHHOFF ’ S LAWS

The Kirchhoff current and voltage laws were formulated by G. Kirchhoff in 1845 [8]. They
govern the interaction between voltages and currents of an electrical circuit. To satisfy
 14 2. E LECTRICAL BACKGROUND

the Maxwell equations (Table 2.1 ), the wavelength of current and voltage need to be
much bigger than the size of the different lumped elements. As a consequence, the elec-
trical charges are transported almost instantaneously through the various components,
according to these laws.

2 2.4.1. K IRCHHOFF ’ S CURRENT LAW

Kirchhoff’s current law states that the algebraic sum of the currents entering a node is
equal to the algebraic sum of the currents leaving the node, as shown in Eq. 2.2.
X X
ie = il (2.2)
cur r ent s ent er i ng cur r ent s l eavi ng

This law forbids the series connection of current sources. For example, when two or
more current sources, which have different values of current injection, are in series, they
do not satisfy Kirchhoff’s current law because the current sum at each node is not zero.

2.4.2. K IRCHHOFF ’ S VOLTAGE LAW

Kirchhoff’s voltage law states that the algebraic sum of the voltage drop of all lumped
elements of a closed loop is equal to 0V, as shown in Eq. 2.3.
X
uj = 0 (2.3)
number o f l umped el ement s o f a cl ose l oop

This law disallows the parallel connection of two voltage sources. For example, when
two or more voltage sources, which impose several potentials, are in parallel, they do not
satisfy Kirchhoff’s voltqge law because the sum of the voltage of a loop is not zero.

2.5. N ORTON ’ S THEOREM

E.L. Norton formulated this theorem in 1926 [9]. It transforms an arbitrary source, com-
posed of lumped elements, into a non-ideal current source. For example, the Norton
theorem is applied to Figure 2.3a, which represents a two-port terminal, non-ideal volt-
age source.

e T h(t )
ZT h

(a) Non-ideal voltgae source (before applying the Norton’s theorem)

j N (t ) ZN

(b) Non-ideal current source (After applying the Norton’s theorem)

Figure 2.3: Application of the Norton’s Transformation

2.6. G RAPH THEORY 15

e
Then after the transformation, Figure 2.3b is found where j N (t ) = ZT h(t ) and Z N =
Th
ZT h . Moreover, the dual of this theorem is called the Thevenin theorem [10].

2.6. G RAPH THEORY

Graph theory is a useful tool for computations with the Kirchhoff laws [11, 12]. First, an 2
electrical circuit is composed of nodes and lumped elements. Second, a graph consists
of nodes (n) and branches (b). In fact, an electrical node corresponds to a graph node,
and a lumped element to a branch. In a graph, the branches are oriented, and their
orientations show an arbitrary current flow for the computation of the Kirchhoff laws.
Z
Then, the incidence matrix (KC L ∈ b×n−1 ) can be obtained by applying Kirchhoff’s
current law at each node of a graph, except to the ground node, usually denoted node 0.
Thus, Kirchhoff’s current law can be written as:

KC L i = 0, (2.4)

where the vector i = [i 1 i 2 ... i b ]T corresponds to the current through each branch and

 +1 if branch i j is connected and oriented to the node k



KC L k j = −1 if branch i j is connected and not oriented to the node k .

0 if the branch i j is not connected to the node k


In addition, a graph can be denoted, such as G(V, E ), where the vertices V repre-
sent the node, and the edges E represent the node connection of lumped elements. On
G(V, E ), a spanning tree T is constructed. This tree includes, by definition, all the ele-
ments of the graph. The set of edges of G(V, E ) not belonging to T defines the co-tree
denoted by L.
For example, the following electrical network is studied for the notions of edges, ver-
tices, tree, and co-tree.

R4

R1 R2

E (t ) R3 R5

Figure 2.4: Electrical network example

The electrical network, shown in Figure 2.4, is composed of six lumped elements and
five nodes. The graph representation of the previous electrical network is presented in
Figure 2.5.
 16 2. E LECTRICAL BACKGROUND

i 2 (R 1 ) i 3 (R 2 )
1 2 3
i 3 (R 4 )
i 1 (E (t ))
i 4 (R 3 )

2 0 i 6 (R 5 )
4
Figure 2.5: Graph of the Figure 2.4

From Figure 2.5, the mathematical formulation of V and E are:

V = {0, 1, 2, 3, 4} , (2.5)
E = {(0, 1) , (1, 2) , (2, 3) , (2, 4) , (3, 0) , (4, 0)} . (2.6)

In addition, another vector N ame E can be defined by naming the edges of the graph
from the lumped representation, such as:

N ame E = {E (t ), R 1 , R 2 , R 3 , R 4 , R 5 } . (2.7)

In addition, the current flows from the first node to the second node of each edge. From
this notation, the incidence matrix of Figure 2.5 can be computed, and is:

1 −1 0 0 0 0
 
 0 1 −1 0 −1 0 
KC L = 
 0 0
 (2.8)
1 −1 0 0 
0 0 0 0 1 −1
Now, the spanning tree is obtained by splitting all the loops of the graph. Then, the
co-tree is obtained by the removing eges of the graph for obtaining the spanning tree.
The spanning tree and co-tree of the graph, shown in Figure 2.6, is given in the following
figure.
Using the same methodology as the graph, the spanning tree T (VT , E T ) of Figure 2.6a
can be defined by:

VT = {0, 1, 2, 3, 4} , 0 (2.9)
E T = {(0, 1) , (1, 2) , (2, 3) , (4, 0)} , (2.10)
N ame E T = {E (t ), R 1 , R 2 , R 5 } . (2.11)

Then, the co-tree definition is:

VL = {0, 1, 2, 3, 4} , (2.12)
E L = {(2, 4) , (3, 0)} , (2.13)
N ame E L = {R 3 , R 4 } . (2.14)

Finally, a graph can be defined as:

V = VT ∪ VL , (2.15)
E = ET ∪ EL , (2.16)
N ame E = N ame E T ∪ N ame E L . (2.17)
2.7. M ODELING METHODS 17

i 2 (R 1 ) i 3 (R 2 )
1 2 3
i 1 (E (t ))

0 i 6 (R 5 )
4 2
(a) Spanning tree of the graph of the Figure 2.5

1 2 3
i 5 (R 4 )
i 4 (R 3 )

0 4
(b) Co-tree of the graph of the Figure 2.5

Figure 2.6: Tree and co-tree of the graph of the Figure 2.5

2.7. M ODELING METHODS

This section offers a brief introduction to the three most important modeling methods:

• The nodal analysis method;

• The modified nodal analysis method;

• Cut-set method.

2.7.1. N ODAL ANALYSIS METHOD

The nodal analysis method was created by H.W. Dommel, who started to work on it at
the Munich Institute of Technology (Germany) in the 1960s [3]. He continued his work at
BPA (Bonneville Power Administration, USA) with Scott-Meyer, his collaborator in those
days on the Electromagnetic Transients Program (EMTP). They published the source
code in the public domain - which explains the popularity of this method.

M ATHEMATICAL EXPRESSION
A power system composed of n nodes and b branches (lumped elements) is under con-
sideration. Then, the mathematical expression at each time-step of the time domain
simulation of the nodal analysis method is:

Y u = i, (2.18)

R
where the matrix Y ∈ n−1×n−1 is the admittance matrix, i ∈ n−1 represents the exci- R
tation of the power system under consideration at each node, and u ∈ n−1 represents R
the voltage between the nodes and the ground (from u 10 to u (n−1)0 where u 10 = u 1 − u 0 ).
The vector u is called the nodal voltage vector. Moreover, the admittance matrix needs
to be updated when the power system’s topology changes.
 18 2. E LECTRICAL BACKGROUND

According to Eq. 2.18, the admittance matrix represents the topology of the power
system under consideration. The vector j describes the current to inject at each node.
For this reason, all voltage sources have to be converted into non-ideal current sources
by means of Norton’s theorem (Section 2.5). As a consequence, the vector u needs to be
computed by solving the linear system of equations (Section 3.4.1). Two examples will
2 demonstrate the simplicity and drawbacks of the nodal analysis method. The first exam-
ple is an arbitrary sample electrical diagram composed only of a current source, a voltage
source, and three resistors. The second example shows the mathematical expression of
the nodal analysis method applied to an RLC circuit.

Example 1
The first example consists of the following electrical circuit [13].

e(t )

R2
1 2
R1
j (t ) R3

Figure 2.7: Electrical diagram 1

In Figure 2.7, a voltage source e(t ) is used. As explained previously in Section 2.5,
the non-ideal voltage source (e(t ) and R 2 in series) needs be replaced by its non-ideal
current source equivalent ( e(t )
R 2 and R 2 in parallel) according to Norton’s theorem. Then,
the transformed electrical circuit is:
2.7. M ODELING METHODS 19

1
R 2 e(t )

2
R2
1 2
R1
j (t ) R3

Figure 2.8: Norton’s theorem transformation of Figure 2.7

The sample electrical diagram of Figure 2.8 is only composed of resistances and cur-
rent sources. Thus, Kirchhoff’s current law is applied at each node. For node number
one, Kirchhoff’s current law reads as:

1 1 1
− j (t ) + (u 10 − u 20 ) + (u 10 − u 20 ) + e(t ) = 0. (2.19)
R1 R2 R2
A similar equation for node number two gives:

1 1 1
u 20 + (u 20 − u 10 ) +G 2 (u 20 − u 10 ) − e(t ) = 0. (2.20)
R3 R1 R2
Equations 2.19 and 2.20 can be combined in the matrix form of the Eq. 2.18, such as:

1 1
− R11 − R12 j (t ) − R12 e(t )
" #· " #
R1 + R2 u 10
¸
= (2.21)
− R11 − R12 1 1
R1 + R2 + R3
1
u 20 1
R 2 e(t )

Example 2

Now, a sample RLC network is studied (Figure 2.9) [14].

R L

E (t ) C

Figure 2.9: Sample RLC network

 20 2. E LECTRICAL BACKGROUND

In fact, the electrical diagram of Figure 2.9 gives a problem because inductances
and capacitances impose differential equations (Table 2.3), and so, the time integration
method needs to be used. Then, the Trapezoidal rule has been chosen. This integration
method is used to convert differential equations into algebraic equations (Section 3.3.2).
Therefore, inductances and capacitances can be represented by a control current source
2 with a parallel resistance, as shown in Figure 2.10, according to Eq. 3.19 to 3.22.

L
C
i C (t − ∆t ) i L (t − ∆t )
i C (t ) i L (t )

∆t 2L
2C ∆t

(a) (b)

Figure 2.10: Capacitance (a) and inductance (b) transformation

According to Figure 2.10, inductance and capacitance models are related to the time-
step. Furthermore, it is necessary to constantly compute the current through all induc-
tances and capacitances (called historical current) to initialize their current sources for
the next time-step [15].
Now, the first step is to replace the voltage source, the inductance, and the capaci-
tance by their equivalent models, as shown in Figure 2.11.

i L (t − ∆t ) i C (t − ∆t )
1 2

E (t )
R (t ) R
RL RC

Figure 2.11: Sample RLC network after transformation

Then, the following set of equations is found after applying the nodal analysis
method to Figure 2.11.

1 E (t )
+ ∆t − ∆t u 10 − i L (t − ∆t )
· ¸· ¸ · ¸
R 2L 2L = R (2.22)
− ∆t
2L
∆t
2L + 2C
∆t
u 20 −i c (t − ∆t ) + i L (t − ∆t )
2.7. M ODELING METHODS 21

G RAPH THEORY APPLIED TO THE NODAL ANALYSIS METHOD

Graph theory is a useful tool to obtain the mathematical expression of the nodal analysis
method. As shown in Section 2.6, it is possible to write Kirchhoff’s current law in a matrix
form. However, a branch represents the parallel associations of a current source and
resistance. According to Figure 2.11, Figure 2.12 shows its graph only where Norton’s
theorem has been employed. 2
i2
1 2
i1
i3

0
Figure 2.12: Graph of the Figure 2.11

In Figure 2.12, branch i 1 represents the current source and its parallel resistance.
Then, branch i 2 describes the current source of the inductance and its parallel resis-
tance. Finally, branch i 3 represents the current source of the capacitance and its parallel
resistance. Thus, Kirchhoff’s current law is applied to Figure 2.12, and it is defined as:

i1
 
−1 1 0
· ¸
 i 2  = 0, (2.23)
0 −1 1
i3
where, the matrix KC L ∈ Rn×b is:
−1 1 0
· ¸
KC L = . (2.24)
0 −1 1
Moreover, to obtain Eq. 2.18, it is necessary to compute matrix G and vector i sour ce
such as:

KC L GKCTL u = −KC L i sour ce (2.25)

where

G j if branch i j represents a resistance

½
Gj j = ,
0 else
and

j k if the branch i k represents a current source

½
i sour ce k = .
0 else
Matrix G and vector i sour ce corresponding to Figure 2.11 are:
¡£ 1 ∆t 2C ¤¢
• G = diag R 2L ∆t ;
E (t )
¤T
• i sour ce = .
£
R i L (t − ∆t ) i c (t − ∆t )
From this information, Eq. 2.22 is found.
 22 2. E LECTRICAL BACKGROUND

2.7.2. M ODIFIED NODAL ANALYSIS METHOD

The modified nodal analysis method was developed by C.W. Ho, A.E. Ruehli, and P.A.
Brennan [4]. This approach is mainly applied to simulate transient stabilities of power
systems or power electronics, and it is based on the Kirchhoff’s laws. The mathemati-
2 cal expression of this method can be obtained by using graph theory as well. Unlike the
nodal analysis method, this formulation allows the use of a time-stepping strategy with
respect to the nodal analysis. Consequently, computation time may decrease. The mod-
ified nodal analysis method can have at least two forms, index 1 and 2 (Section 3.2). To
simulate the time domain solution for both indexes, an implicit numerical integration
method can be used (Section 3.2).

M ATHEMATICAL EXPRESSION
A power system composed of n nodes, b branches, n e voltage sources, n L inductances,
and nC capacitances is under consideration. Thus, the mathematical expression of the
modified nodal analysis method is:

d x(t )
M + Ax(t ) + g (t ) = 0, (2.26)
dt

where matrix M ∈ R(n−1+n +n )×(n−1+n +n ) is the mass matrix.

L e L e Matrices M and A ∈
Rn−1+n +n
(n−1+n L +n e )×(n−1+n L +n e )
are related to the topology of the power system. Vector x ∈
R L e contains algebraic and differential variables. Also, vector g (t ) ∈ R
n−1+n L +n e

represents the excitation of the power system. Moreover after a change of topology, it is
necessary to update matrix A.
According to Figure 2.9, five equations can be found. The first three equations are
related to Kirchhoff’s current law (Eq. 2.27 to 2.29). Then, the fourth equation is related
to Kirchhoff’s voltage law (Eq. 2.30). The last equation imposes the potential across
u 10 due to the voltage source because it is located between node 1 and the ground (Eq.
2.31)[14].

u 10 − u 20
0 =i E (t ) − (2.27)
R
u 10 − u 20
0= − iL (2.28)
R
d u 30
0 =i L − i C = i L −C (2.29)
dt
d iL
0 =E (t ) − (u 10 − u 20 ) − v L − vC = u 20 − L − u 30 (2.30)
dt
0 =E (t ) − u 10 (2.31)

Thus, it is possible to write the previous equations into a matrix form, such as:
2.7. M ODELING METHODS 23

  1 1

0 0 0 0 0 u˙10 −R 0 0 1 u 10
  
R
1
 0 0 0 0 0  u˙20  
  R
 − R1 0 −1 0 
 u 20 
u˙30
  
0 0 −C 0 0 + 0 0 0 1 0 u 30
   
2

i˙L
    
0 0 0 −L 0   0 1 −1 0 0 iL
   

0 0 0 0 0 i˙E −1 0 0 0 0 iE
0
 
 0 
 
 0
+ =0 (2.32)

 0 
E (t )

G RAPH THEORY APPLIED TO THE MODIFIED NODAL ANALYSIS METHOD

As for the nodal analysis methods, a matrix form can be obtained from the graph theory
[16]. Firstly, matrices L l i st and C l i st , which are necessary for the computation of the
modified nodal analysis set of equations from the graph theory, are developed - such
as matrix G of the nodal analysis. Then, vector e sour ce is computed according to vector
i sour ce definition of the nodal analysis too. Finally, the modified nodal analysis set of
equations can be divided into three subsets of equations. The first subset is related to
Kirchhoff’s current law. The second subset is linked to Kirchhoff’s voltage law, and the
last subset is tied to the voltage source node relation. Then, Figure 2.13 shows the graph
used for the modified nodal analysis of the electrical diagram of Figure 2.9.

i2 i3
1 2 3
i1
i4

0
Figure 2.13: Graph of the Figure 2.9

The first subset of equations can be written as:

KC L C l i st KCTL u̇ + KC L GKCTL u + (KC L L l i st )i L + (KC L E l i st )i e + KC L i sour ce (t ) = 0, (2.33)

R
where vector u ∈ n−1 is composed of voltages u 10 to u (n−1)0 , vector i L ∈ n−1 repre- R
sents the current through the different inductances, and vector i e ∈ n−1 represents the R
current through the different voltage sources. The second subset of equations is:

− L l i st i˙L + (KC L L l i st )T u = 0. (2.34)

Finally, the last subset of equations is:

(KC L E l i st )T u − KC L e sour ce (t ) = 0. (2.35)

 24 2. E LECTRICAL BACKGROUND

The previous methodology is now applied to Figure 2.9 to find the differential alge-
braic equations. As a consequence, the different matrices and vectors for the computa-
tion of the system of equations are:
−1 1 0 0
 

2 • KC L =  0 −1 1 0 ;
0 0 −1 1
1
• G = diag 0 R 0 0 ;
£ ¤

• L l i st = diag 0 0 L 0 ;
£ ¤

• C l i st = diag 0 0 0 C ;
£ ¤

• E l i st = diag 1 0 0 0 ;
£ ¤

¤T
• e sour ce (t ) = E (t ) 0 0 0 ;
£

¤T
• i sour ce (t ) == 0 0 0 0 .
£

When all matrices and vectors found previously are put together in a matrix form, Eq.
2.32 is found. In fact, it would be the same only after performing the following transfor-
mation. If all elements of the i t h row of Eq. 2.26 are equal to 0, the i t h column and i t h
row of the matrices M and A can be deleted, along with the i t h row of the vector g (t ).

2.7.3. C UT- SET METHOD

The cut-set method [17] was popularized by E.S. Kuh and R.A. Rohrer in 1965 [5]. This
method computes the ordinary differential equations of a power system, is based only
on graph theory, and it allows for time-stepping algorithms without restriction (Section
3.3). Nevertheless, this approach is the most complex and computationally expensive
modeling method to arrive at the mathematical expression.

M ATHEMATICAL EXPRESSION
A power system composed of n s sources, n L inductances, and nC capacitances is under
consideration. Thus, the mathematical expression of the cut-set method is:
d x(t )
= ẋ = Ax + B g (t ), (2.36)
dt
R
where matrix A ∈ (nL +nC )×(nL +nC ) is the state matrix, matrix B ∈ R(nn +n
L )×(n )
C s is the in-
R
put matrix, vector x ∈ nL +nC is the state vector, and vector g (t ) ∈ R is the time input
s)

vector.

RLC CIRCUIT EXAMPLE

By applying the methodology developed in [5] to Figure 2.9, the following ordinary dif-
ferential equations are found:
1
u˙c 0 uC 0
· ¸ · ¸· ¸ · ¸
= C + 1 e(t ) (2.37)
i˙L − L1 1
− RL iL L
Equation 2.37 corresponds well to the ordinary differential equations of Figure 2.9,
according to [14].
2.7. M ODELING METHODS 25

G RAPH THEORY
For example, an electrical network is represented by an interconnection of resistors, in-
ductors, capacitors, voltage sources, and current sources. The difference from the other
methods is that the graph is decomposed according to the spanning tree and co-tree, as
developed in Section 2.6. The tree/co-tree decomposition of G(V, E ) is made such that
capacitors and voltage sources are in the tree while the inductors and current sources 2
are in the co-tree. Resistors can appear in both the tree and co-tree. Schematically, the
decomposition is made as:

T = {voltage sources} ∪ {capacitors} ∪ { resistors} (2.38)

L = { resistors} ∪ {inductors} ∪ {current sources} . (2.39)

For brevity, the graph G(V, E ) allows a tree/co-tree decomposition, as prescribed by Eq.
2.38 and Eq. 2.39. Vectors i (t ) and v(t ) are the time-dependent current and voltage
variables associated to the edges E . They can be partitioned into sub-vectors related to
the tree/co-tree decomposition and the variable associated to the edges E , such as:
 
ie
 i
Ct 

¸ 
it  iR 
·  
i= =  t , (2.40)
i`  i R` 
 i L` 
 

iJ

and  
ve

 vC t 

vt v Rt
· ¸  
v= = . (2.41)
 
v`  v R` 
v L`
 
 
vJ
Given the spanning tree T , a maximal number of linearly independent fundamen-
tal loops and cut-sets can be constructed. A fundamental loop consists of a subset of
tree edges and a single co-tree edge. A fundamental cut-set consists of the co-tree edges
and a single tree edge. Let K V L denote the incidence matrix between the edges and the
fundamental cut-sets, and let K V L = [K V L t K V L ` ] = [K V L t I ] be a tree/co-tree partition-
ing of the columns of K V L . A maximal number of linearly independent Kirchhoff’s volt-
age laws can then be expressed as K V L v = 0, or equivalently, as K V L t v t = −v ` . Let Q
denote the incidence matrix between the edges and the fundamental cut-sets, and let
Q = [Q t Q ` ] = [I Q ` ] be a tree/co-tree partitioning of the columns of Q. A maximal num-
ber of linearly independent Kirchhoff current laws can then be expressed as Q i = 0, or
equivalently, i t = −Q ` i ` (t ). The columns of Q and K V L are orthogonal to each other,
and therefore Q T K V L = [K V Lt I ]T [I Q ` ] = 0, from which it follows that K V L t = −Q `T . This
equality can be employed to write a set of linearly independent Kirchhoff’s voltage and
current laws, as
v ` (t ) = Q `T v t and i t = −Q ` i ` (t ) , respectively. (2.42)
 26 2. E LECTRICAL BACKGROUND

By decomposing the tree and co-tree vectors further into their respective components,
it is possible to write:

v R` ve ie i R`
       
 v L  = Q T vC  and i C  = −Q ` i L  respectively. (2.43)
` t t
2 vj
`
v Rt i Rt iJ
`

The first component on the left hand side of both equations can be eliminated using the
constitutive equations Table 2.3, which is re-formulated as:

d i L ` (t ) d vC t (t )
v L` = L ` = L ` i L˙` and i C t = C ` = C ` v˙C t , (2.44)
dt dt
where L ` and C t denotes the inductance of the `-th branch and capacitance of the t -th
branch, respectively. The second component on the right hand side of both equations
can be rewritten using Ohm’s Law (Table 2.3), which is re-formulated as:

v R t = R t i R t and i R` = G ` v R` , (2.45)

With these changes the two vector equations in Eq. 2.43 become:

v R` Ve ie G ` v R`
       
L ` i L˙  = Q  vC  and C t v˙C  = −Q `  i L  respectively.
T
(2.46)
` ` t t `
vJ Rt i Rt i Rt v R`

A system of differential-algebraic equations for the currents through the inductors i L `

and the voltage across the capacitors vC t is thus obtained. By eliminating the algebraic
equations from Eq. 2.46, this system can be reduced to a system of ordinary differential
equations. The voltages associated with the co-tree resistance v R` and the currents asso-
ciated with the tree resistances vi R t (t ), for instance, can be written in terms of the state
variables i L ` and vC t and source term variables by using the second set of equations of
(2.46) and subsequently eliminated. The system of ordinary differential equations that
describes the evolution of the state variables as:

vC t (t )
µ ¶
x(t ) = (2.47)
i L ` (t )

can then be written as:

˙ ) = f (x(t ), t ) = A x(t ) + B g (t ) ,
x(t (2.48)
R R
where matrix A ∈ (nL +nC )×(nL +nC ) , matrix B ∈ (nL +nC )×(ne +ni ) and these matrices
R
are related to the topology of the power system. Vector x ∈ nL +nC contains only differ-
R
ential variables, and vector g (t ) ∈ ne +ni represents the excitation of the power system.

2.8. P OWER SYSTEM ELEMENTS

In this section, some power system elements used in this thesis are presented. These
elements are:

• Switches;
2.8. P OWER SYSTEM ELEMENTS 27

• Generator model;

• Load model;

• PI-section model;

• Double PI-section model;

2
• Arc models.

2.8.1. S WITCHES
A switch can interrupt the current between two nodes. It has two positions, open and
closed. Ideal switches are not recommended due to Kirchhoff’s voltage law, because
a switching device may permit the parallel connection of voltage sources of different
amplitudes. As a consequence, the following model is used.

R sw R sw (α)

Figure 2.14: Non-ideal switch

The resistance R sw represents the internal resistance and α the switch position. In
the case of a closed switch, α is equal to 1. Otherwise, α is equal to 0.

2.8.2. G ENERATOR MODEL

Several generator models are available in the literature. For electromagnetic transients,
the following model is adequate (1-phase model).

R L
iL

e(t ) vC C RC

Figure 2.15: Generator diagram

The state space representation of Figure 2.15 is:

− RL − L1 iL 1
· ¸· ¸ · ¸
ẋ = AG x + BG g (t ) = + L e(t ). (2.49)
1
C − R 1C vC 0
C

where the generator model’s parameters are R, RC , L, C ,and e(t ) is the voltage wave-
form of the voltage source. This generator is usually used for the simulation of transient
recovery voltage. The resistance RC is employed for damping voltage vC .
 28 2. E LECTRICAL BACKGROUND

2.8.3. L OAD MODEL

The load model is represented in the following figure. This model offers the possibility
of tuning the load for being resistive, inductive, or capacitive.

2 R L
iL

vC C

Figure 2.16: Load diagram

The state space representation of Figure 2.16 is:

0 − C1 vC
· ¸· ¸
ẋ = A L x 1 , (2.50)
L − RL iL

where R, L and C are the load model’s parameters.

2.8.4. PI- SECTION MODEL

PI-section models lines and cables from two different points of the network. The PI-
section network symbol is depicted in the following figure.

R L
iL

vC 1 C1 vC 2 C2

Figure 2.17: PI-section diagram

The state space representation of Figure 2.17 is:

− C11
 
0 0 vC 1

1
ẋ = A P I x =  − RL 1 
−L  iL  , (2.51)

L
1
0 C2 0 v C2

where R, L, C 1 and C 2 are the PI-section model’s parameters.

2.8. P OWER SYSTEM ELEMENTS 29

2.8.5. D OUBLE PI- SECTION MODEL

PII-section models consist of two PI-sections in series. It permits the increase of length
in the transmission line, and the accuracy of the model. Its representation is illustrated
in the following electrical diagram.

R L1 R2 L2 2
i L1 i L2

vC 1 C1 vC 2 C2 vC 3 C3

Figure 2.18: Double PI-section diagram

The state space representation of Figure 2.18 is:

0 − C11 0 0 0
 
vC 1
 
− RL 11
 1
− L11 0 0 i L1

 L1  
1
 0 0 − C12 0
  
ẋ = A P I I x =  C2

 vC 2 ,
 (2.52)
 0 0 1
− RL 22 − L12 i L2
  
L2 
0 0 0 1
0 vC 3
C3

where R 1 , R 2 , L 1 , L 2 , C 1 , C 2 and C 3 are the double PI-section model’s parameters.

2.8.6. A RC MODELS
An arc model can replace a circuit breaker, a type of switch (Section 2.8.1). It models the
transition from a closed to an open position of a switch by using non-linear equations.
These equations describe the internal conductivity of the arc. The representation of the
arc model is shown in the following figure.

u ar c

i ar c
R ar c

Figure 2.19: Lumped arc model

The resistance R ar c represents the arc resistance, i ar c is the current through the arc,
and u ar c is the voltage across the arc. In the literature, three classical models are avail-
able (Cassie, Mayr, and Habedank models [18]). The following differential equation per-
forms the change of conductivity described by Cassie’s model:
µ 2
d g c g c u ar
¶
c
= − 1 , (2.53)
dt τc Uc2
 30 R EFERENCES

while the Mayr equation is described by the following equation:

d g m g m u ar c i ar c
µ ¶
= −1 , (2.54)
dt τm Pm
where g c and g m represent the conductivity of Cassie’s and Mayr’s models, respectively.
2 Also, τc and τm are the time constant of the Cassie and Mayr model. Furthermore, Uc is
the steady-state voltage and P m is the steady-state arc power loss.
The series connection of the Cassie and Mayr models is considered as the Habedank
model. As the result, the resistivity of the electrical arc in this model is given by:

1 1
R ar c = = 1
, (2.55)
g ar c gc + g1m
where i ar c is defined as:
u ar c
i ar c = = u ar c g ar c . (2.56)
R ar c

2.9. C ONCLUSION
In this chapter, a brief introduction to circuit theory has been given, and power system
elements are given. The traditional modeling methods are the nodal analysis, the mod-
ified nodal analysis, and the cut-set method. The nodal analysis is mainly used for the
simulation of large-scale power systems. This method is based on the nodal admittance
matrix and a numerical integration method (the Trapezoidal rule). The main disadvan-
tage of this approach is the extensive computation time of an overall simulation due to
the fixed time-step.
The modified nodal analysis applied to the power system under consideration gives
a set of differential algebraic equations. This method can use an adaptive time-stepping
integration method. However, it allows only implicit numerical integration methods,
and it is complicated to solve differential algebraic equations as shown in Chapter 3.
The cut-set method gives the state-space representation of the power system under
consideration. The methodology to compute the set of equations is its primary draw-
back. In fact, this method is used for the investigation of new components. This ap-
proach allows all types of numerical integration methods.
Then, from the various modeling methods, simulations of large-scale power systems
become slow for two reasons:

• They impose a type of numerical integration method;

• They are computationally expensive after a switching action.

R EFERENCES
[1] J. C. Maxwell, A dynamical theory of the electromagnetic field, Philosophical Trans-
actions of the Royal Society of London 155, 459 (1865).

[2] R. DeCarlo and P.-M. Lin, Linear Circuit Analysis, 2nd ed. (OXFORD UNIVERSITY
PRESS, New York, 2001).
R EFERENCES 31

[3] H. W. Dommel, Digital computer solution of electromagnetic transients in single-

and multiphase networks, Power Apparatus and Systems, IEEE Transactions on
PAS-88, 388 (1969).

[4] C. W. Ho, A. E. Ruehli, and P. A. Brennan, The modified nodal approach to network
analysis, Circuits and Systems, IEEE Transactions on 22, 504 (1975). 2
[5] E. Kuh and R. Rohrer, The state-variable approach to network analysis, Proceedings
of the IEEE 53, 672 (1965).

[6] P. R. Amestoy, T. A. Davis, and I. S. Duff, An approximate minimum degree ordering

algorithm, SIAM Journal on Matrix Analysis and Applications 17, 886 (1996).

[7] L. Chua, C. Deseor, and E. Kuh, Linear and nonlinear circuits (McGraw-Hill Book
Company, New York, 1987).

[8] G. Kirchhoff, Ueber den durchgang eines elektrischen stromes durch eine ebene, ins-
besondere durch eine kreisförmige, Annalen der Physik 140, 497 (1845).

[9] E.L.Norton, Design of finite networks for uniform frequency characteristic.Technical

Report TM26-0-1860, Tech. Rep. (Bell Laboratories, 1926).

[10] L. Thévenin, Sur un nouveau théoreme d’électricité dynamique [on a new theorem of
dynamic electricity], CR des Séances de l’Académie des Sciences 97, 159 (1883).

[11] A. Ioinovici, Computer-Aided Analysis of Active Circuits (MARCELL DEKKER, inc,

New York, 1990).

[12] J. Vlach and K. Singhal, Computer methods for circuit analysis and design (Van Nos-
trand Reinhold, New York, 1983).

[13] R. Smeets, L. Van der Sluis, M. Kapetanovic, D. F. Peelo, and A. Janssen, Switching in
electrical transmission and distribution systems (John Wiley & Sons city, Chichester,
2014).

[14] L. van der Sluis, Transients in Power Systems (John Wiley & Sons, Chichester, 2001).

[15] N. Watson and J. Arrillaga, Power systems electromagnetic transients simulation,

Vol. 39 (Institution of Electrical Engineers, London, 2003).

[16] S. Schöps, Multiscale modeling and multirate time-integration of field/circuit cou-

pled problems, Ph.D. thesis, Universitätsbibliothek Wuppertal, Wuppertal (2011).

[17] T. Bashkow, The A matrix, new network description, IRE transactions on circuit The-
ory 4, 117 (1957).

[18] M. Kapetanovic, High voltage circuit breakers (Sarajevo, 2011).

 3
M ATHEMATICAL BACKGROUND

3.1. I NTRODUCTION
T ODAY , mathematics helps to solve a significant amount of problems [1]. This obser-
vation is also true with respect to the simulations of power systems. For example, it
helps find network equations, differential or not (Chapter 2). It also permits the obtain-
ing of the time domain solution, as will be shown in this chapter.
When differential equations model a power system, two possible types of differential
equations exist - differential algebraic and ordinary[2]. For solving them, numerical in-
tegration methods are applied, such as the Euler or Runge-Kutta methods [3–5]. Due to
the stiffness of power systems [6], implicit integration methods are recommended. This
recommendation leads to solving systems of equations during the integration process.
A significant amount of linear and non-linear solvers are available in the literature[7,
8]. One can distinguish direct and iterative linear solvers. In the case of a small linear sys-
tem of equations, direct solvers are recommended. Otherwise, iterative methods are ad-
vised. Iterative methods are typically faster to converge if they are used in combination
with a pre-conditioner. As a consequence, a speed-up can be observed by choosing the
appropriate solver. Finally, to solve a set of non-linear equations, the Newton-Raphson
method is applied.
This chapter is organized as follows: Section 3.2 shows the differences and links be-
tween differential algebraic and ordinary differential equations, and Section 3.3 intro-
duces basic integration methods. A brief description on linear and non-linear solvers is
given in Section 3.4, and finally, Section 3.5 offers the conclusions.

3.2. O RDINARY DIFFERENTIAL EQUATIONS AND DIFFERENTIAL

ALGEBRAIC EQUATIONS
When the modified nodal analysis method (Section 2.7.2) or the cut-set method (Section
2.7.3) models a power system, differential equations are found. It exists in two forms of
differential equations:

33
 34 3. M ATHEMATICAL BACKGROUND

• Ordinary Differential Equations (ODEs);

• Differential Algebraic Equations (DAEs).

The general notation of an ODE is:

dx
ẋ = = Ax + B g (t ), (3.1)
dt
whereas the form of a DEA is:
3 M ẋ + Ax + B g (t ) = 0, (3.2)

where the matrix M is the mass matrix. The matrices M , A and B are affected by the
network topology. The vector x contains only the differential variables in the case of an
ODE. In the case of DAEs, the vector x contains algebraic-variables as well. Finally, the
vector g (t ) excites the power system.
When the matrix M is diagonal, the DAE is an ODE. There are several types of DAEs
that are classified by the so-called differentiation index or simply, index. The index rep-
resents the number of reductions needed to get an ODE. In the example of the modified
nodal analysis method, the index is two for Eq. 2.32. In order to get an ODE, two stages
are necessary. The first stage is to write the voltage across the capacitances in terms of
current through the inductances and algebraic variables. The second stage is to include
the algebraic equations into the differential equation. For example, the demonstration
of this methodology is shown below.

• stage 1 The capacitor is placed between the nodes 3 and 0. As a consequence, u 30

can be renamed uC . Then, the DAE of order 1 can be obtained after reordering Eq.
2.32 as:

  
−C 0 0 0 0 u˙c 0 1 0 0 0 uc
  
 0 −L 0 0 0  i˙L  
  −1 0 1 0 0 
 iL 
1
u˙10 − R1
  
0 0 0 0 0 + 0 0 1 u 10
  
R

  
− R1 1
  
0 0 0 0 0 u˙20   0 −1 0 u 20
   
R
 
0 0 0 0 0 i˙E 0 0 −1 0 0 iE
0
 

 0 

+
 0 =0
 (3.3)
 0 
E (t )

• stage 2 Now, the DAE can be formulated as block matrices of the following form:

M 11 0 ẋ A 11 A 11 x 0
· ¸· ¸ · ¸· ¸ · ¸
+ + = 0, (3.4)
0 0 ẏ A 21 A 22 y g 2 (t )

Then, the ODE representation is:

M 11 ẋ + A 11 + A 11 A −1
22 A 21 x + A 11 A −
22 1A 21 g 2 (t ) = 0. (3.5)
£ ¤£ ¤ £ ¤£ ¤ £ ¤
3.3. N UMERICAL INTEGRATION METHODS 35

Finally, by putting M 11 ẋ at the right hand side of Eq. 3.5, and then, multiplying by
M 11 on the left side of the equation, the ODE form according to Eq. 3.1, Eq. 2.37 is
found.

Another particularity of DAEs is that it is possible to increase their index. Finally, the
utilization of DAEs implies the use of a fully implicit numerical integration method [9].

3.3. N UMERICAL INTEGRATION METHODS 3

In this section, the Euler methods and the Trapezoidal rule are presented first [2, 3], fol-
lowed by the Runge-Kutta methods [10]. To study the effect of these various numerical
integration methods, the following electrical diagram is used (Figure 3.1) [11].

i C (t )
C R

Figure 3.1: RC circuit where C = 1F and R = 0.5Ω

According to Figure 3.1, the differential equation of this electrical circuit is:

1
v˙C = f (vC ) = − vc . (3.6)
RC

Eq. 3.6 is defined as an initial value problem. In case of v c (t 0 = 0) = 0, the solution of

this equation is v c (t ) = 0 for t ≥ 0. As a consequence, Eq. 3.6 needs to be initialized with
a value v c (t 0 = 0) different of zero. For this problem, v c is initialized as:

vC (t 0 = 0) = 10V. (3.7)

Moreover, Eq. 3.8 presents the analytical solution of Eq. 3.6:

t t
vC (t ) = vC (t 0 )e − RC = vC (t 0 )e − τ = vC (t 0 )e −λt , (3.8)

where the time constant τ or eigenvalue λ of the electrical circuit depends on the values
of the resistance and of the capacitance. Thus, Figure 3.2 shows the time evolution of the
voltage across the capacitor, according to Eq. 3.8.
 36 3. M ATHEMATICAL BACKGROUND

10

vC [V ]
4

0
3 0 2 4
t [s]

Figure 3.2: Representation of the time domain solution of Eq. 3.6, where the time constant is τ = 0.5s

In general, numerical integration methods are based on the Taylor series of an expo-
nential (Eq. 3.9). Then, the order (p t h ) of the numerical integration method corresponds
to the higher derivative used in this equation. Thus, the local truncation error (error at
each time-step) is calculated from the following equation:

p
∆t i i
A u n + O (∆t )p+1
X
u n+1 = (3.9)
i =0 i !

where the time step is denoted by ∆t , and the time t n at the time iteration n is defined
as t 0 + n∆t for a fixed time-step where t 0 is the initial time. Finally, u n+1 represents
an approximation of the solution of the differential variables at the time t n+1 such that
x(t n+1 ) = u n+1 + O (∆t )p+1 , where O (∆t )p+1 represents the local truncation error.

3.3.1. E ULER METHODS

L. Euler published methods to solve ordinary differential equations numerically [3]. The
explicit and implicit methods are called Euler Forward and Euler Backward respectively.
Furthermore, these two approaches are first order (p = 1). The explicit method uses only
information from the past to arrive at the solution of the next step. On the contrary, the
implicit method also needs the solution to compute it. As a consequence, this method
requires more computations per time-step than explicit methods because a set of equa-
tions has to be solved at each time-step.

E ULER FORWARD METHOD

The numerical formulation of the Euler forward method is:

x n+1 = x n + ∆t f (t n , x n ). (3.10)

Then, this integration method integrates Eq. 3.6 with time-steps of 10ms and 1s. The
numerical solution with these different time-steps is shown in Figure 3.3.
3.3. N UMERICAL INTEGRATION METHODS 37

10 10
Vc (t ) from Eq. 3.10
8 Vc (t ) from Eq. 3.8 5
6
vC [V ]

vC [V ]
0
4
−5
2

0 −10
0 2 4 0 2 4 3
t [s] t [s]
(a) ∆t = 10ms (b) ∆t = 1s

Figure 3.3: Numerical experiments of Euler forward method

Figure 3.3a represents an acceptable approximation of the result according to the

analytical solution (Figure 3.2) because the two curves can be superposed. However,
the curve of Figure 3.3b shows oscillations because the time-step chosen was just at the
stability limit: a larger time-step would have made the time solution divergent. To study
the stability, the test equation used is:

ẋ = λx = f (t , x), (3.11)

So, the convergence is obtained only when:

|1 + ∆t λ| ≤ 1, (3.12)

which yields to the stability region of the Euler forward method. For instance, ∆t is pos-
itive, so λ should be negative to satisfy the following equation:

2
∆t ≤ . (3.13)
−λ
In Figure 3.3b, the requirement of Eq. 3.13 was just at the stability limit because
λ = −2. As a consequence, the numerical integration method was oscillating around the
solution. The stability region could then be drawn in the complex plane (Figure 3.4).

ℑ(λ∆t )

-1 0 ℜ(λ∆t )

Figure 3.4: Stability region: Stable inside of the circle

 38 3. M ATHEMATICAL BACKGROUND

For example, if an RLC circuit is studied, this circuit will have two eigenvalues, and
they could be complex conjugate with the real part negative. The local truncation error
equation (Eq. 3.14) for this numerical method is derived from Eq. 3.9 and 3.10, and it is
defined as:

∆t 2 d 2 x(τn )
e n+1 = , (3.14)
2 dt2

3 where τn is comprised between t n and t n + ∆t , and it is arbitrarily chosen.

E ULER BACKWARD METHOD

The formulation of Euler backward method is:

x n+1 = x n + ∆t f (∆t , x n+1 ). (3.15)

Similar to the Euler forward, time-steps of 10ms, and then, of 1s are applied with this
method to Eq. 3.6. Thus, Figure 3.5 shows the time domain representations.

10 10
Vc (t ) from Eq. 3.15
8 Vc (t ) from Eq. 3.8 8

6 6
vC [V ]

vC [V ]

4 4

2 2

0 0
0 2 4 0 2 4
t [s] t [s]
(a) ∆t = 10ms (b) ∆t = 1s

Figure 3.5: Numerical experiments of Euler backward

According to Figure 3.5, the time domain solution of Eq. 3.6 converges for a time-
step of 1s. To understand this difference with regard to the previous method, the region
of convergence is studied (Eq. 3.16).

1
≤ 1. (3.16)
|1 − ∆t λ|

From Eq. 3.16, the stability region can be plotted as seen in Figure 3.6.
3.3. N UMERICAL INTEGRATION METHODS 39

ℑ(λ∆t )

0 1 ℜ(λ∆t )

Figure 3.6: Stability region: Stable outside of the circle

3

In the case of Euler backward, there is convergence when ℜ(λ) ≤ 0 for any ∆t
(no mathematical restriction on the time-step).For this reason, although the Euler for-
ward method diverged to a suitable solution with a large time-step, the Euler backward
method converged to 0V. However, it requires more computational power than the Euler
forward, but can allow more significant time-steps and still converge. The local trunca-
tion error equation of this approach is the same as that for Euler forward (Eq. 3.14).

3.3.2. T RAPEZOIDAL RULE

The trapezoidal rule averages the Euler forward and backward methods [2]. The mathe-
matical formulation of this numerical integration method is:

∆t ¡
x n+1 = x n + f (t , x n ) + f (t + ∆t , x n+1 ) . (3.17)
¢
2

Also, this method is more accurate than the Euler methods because it is a second or-
der integration method (p = 2), and is implicit too. That means that the local truncation
error is smaller than with the Euler methods, and it is defined as:

∆t 2 d 3 x(τn )
en = , (3.18)
12 dt3

where τn is located between t n and t n + ∆t .

Figure 3.7 shows the time domain solution of Eq. 3.6 when this equation is integrated
by the aim of the trapezoidal rule, with time-steps of 10ms and then of 1s.
 40 3. M ATHEMATICAL BACKGROUND

10 10
Vc (t ) from Eq. 3.17
8 Vc (t ) from Eq. 3.8 8

6 6
vC [V ]

vC [V ]
4 4

2 2

0 0
3 0 2 4 0 2 4
t [s] t [s]
(a) ∆t = 10ms (b) ∆t = 1s

Figure 3.7: Numerical experiments of the Trapezoidal rule

As expected, the time domain solutions converge for these time-steps. This conver-
gence can also be seen in Figure 3.8, which shows the stability region.

ℑ(λ∆t )

0 ℜ(λ∆t )

Figure 3.8: Stability region: Stable in the left half plane

According to this figure, the trapezoidal rule allows any time-step when ℜ(λ) ≤ 0.
Moreover, the trapezoidal rule is used during the nodal analysis method (Section 2.7.1).
For example, an inductance L is placed between a node k and m. According to the dif-
ferential equation of an inductance (Table 2.3), the current injection formulation is:

∆t
i k,m (t + ∆t ) = (v k (t + ∆t ) − v m (t + ∆t )) + i k,m (t ), (3.19)
2L
where:

∆t
i k,m (t ) = i k,m (t ) + (v k (t ) − v m (t )), (3.20)
2L
When the inductance L is replaced by a capacitance C , it holds:

2C
i k,m (t + ∆t ) = (v k (t + ∆t ) − v m (t + ∆t )) + i k,m (t ), (3.21)
∆t
where:
3.3. N UMERICAL INTEGRATION METHODS 41

2C
i k,m (t ) = −i k,m (t ) − (v k (t ) − v m (t )), (3.22)
∆t
where i k,m represents the historical current. The equivalent representations of Eq. 3.19
to 3.22 are shown in Figure 2.10.

3.3.3. RUNGE -K UTTA METHODS

One of the most famous numerical integration method schemes is the Runge-Kutta
methods (RK methods). They had been developed a century ago, and are the results
of research by C. Runge and M. Kutta. The standard mathematical formulation of this
3
method is:

k i = f (t + c i ∆t , x n + ∆t sj =1 a i j k j ) for 1 ≤ i ≤ s
½ P
Ps , (3.23)
x n+1 = x n + ∆t i =1 b i k i )
where s represents the number of stages, and a i j , b i and c i are the coefficients of the
method according to the Butcher tableau (Table 3.1) [12]. The Butcher coefficients a i j
and b i are computed according to the Taylor expansion in Eq. 3.9.

c a
b

Table 3.1: Butcher tableau

For example, the following tables represent different the Butcher tableau with three
stages. From these representations, explicit and implicit methods can be extracted.

c1 a 11 0 0
0 0 0 0 c1 a 11 a 12 a 13
c2 a 21 a 22 0
c2 a 21 0 0 c2 a 21 a 22 a 23
c3 a 31 a 32 a 32
c3 a 31 a 32 0 c3 a 31 a 32 a 32
b1 b2 b3
b1 b2 b3 b1 b2 b3
(b) Diagonally Implicit
(a) Explicit Runge-Kutta (c) Implicit Runge-Kutta
Runge-Kutta

Table 3.2: Rung-Kutta Butcher tableau type

For example, Table 3.2a represents an explicit method or Explicit Runge-Kutta (ERK)
because the matrix formed by the entries a i j is strictly lower diagonal. In addition, the
vector k i is only dependent on the other vector k with an index inferior to i . Meanwhile,
Table 3.2c represents an implicit method or Implicit Rung-Kutta (IRK) because all vectors
k i are needed to compute all of them. Therefore, a linear system of equations has to
be solved with a size proportional to the size of the matrix A and the number of stages
of the method. The last category of Runge-Kutta method is represented by Table 3.2b.
This method is also an implicit Runge-Kutta method; however, it a Diagonally Implicit
Runge-Kutta (DIRK) [13]. The difference with the IRK methods is that there is now s
linear system of equations of size of the matrix A. A further reduction in computational
 42 3. M ATHEMATICAL BACKGROUND

cost is provided by single diagonal implicit methods (SDIRK) characterized by a lower

diagonal matrix a i j with a constant diagonal. The value of the diagonal will be denoted
by γ. Finally, Explicit Single Diagonal Implicit Runge-Kutta method (ESDIRK) is a sub-
method of a SDIRK where a 11 = 0 or the vector k 1 is found explicitly.
For example, Table 3.3 shows the Butcher tableau of the Euler methods and the
Trapezoidal rule.

0 0 1 1
1 1
3
(a) Euler forward (b) Euler backward
0 0 0
1 0 1
0.5 0.5
(c) Trapezoidal rule

Table 3.3: Butcher tableau of basic numerical integration methods

The RK method formulation (Eq. 3.23) is now applied to solve Eq. 3.6, when the
Butcher tableau of the trapezoidal rule, which is an ESDIRK method, is used as shown in
Eq. 3.24.

 k 1 = λuC


(1 − ∆t λ)k 2 = λ(uC ) . (3.24)

uC n+1 = uC n + ∆t [0.5k 1 + 0.5k 2 ]


According to Eq. 3.24, a linear system of equation has to be solved, and it concords
with the implicit definition of the method.
In addition, some Runge-Kutta methods allow the adaptive change of the time-step
from the local truncation error. The Butcher representation of adaptive ESDIRK methods
is

0 0 0 ··· 0
c2 a 21 a 22 ··· 0
.. .. .. ..
. . . . 0
c6 a 61 a 62 ··· a 66
b1 b2 ··· b6
b̂ 1 b̂ 2 ··· b̂ 6

Table 3.4: Butcher tableau of adaptive ESDIRK methods

Table 3.4 allows an adaptive time-stepping strategy based on an approximation of

the local truncation error e n+1 because the b̂ coefficients correspond to solution of the
order p − 1. Then, e n+1 is defined as:
s ¡
X
e n+1 = ∆t b i − b̂ i k i . (3.25)
¢
i =1
3.4. S OLVERS 43

The new time ∆t new is then computed as either an increase or a decrease of the old one
∆t ol d , according to the rule [14]:
v
Tol
u
=t θ∆t ol d , (3.26)
u
∆t new p
||e n+1 ||∞
||u n+1 ||∞

where θ ≤ 1 is a safety factor and Tol is a user-specified absolute tolerance.

In this thesis, the implicit part of the ARK4(3)6L[2]SA-ESDIRK (p = 4, s = 6) is used
[15]. This method is an ESDIRK method, and is depicted as follows. 3
0 0 0 0 0 0 0
0.50 0.25 0.25 0 0 0 0
0.33 0.14 −0.06 0.25 0 0 0
0.62 0.14 −0.22 0.45 0.25 0 0
0.85 0.10 −0.59 0.81 0.28 0.25 0
1 0.16 0 0.19 0.68 −0.28 0.25
0.16 0 0.19 0.68 −0.28 0.25
0.15 0 0.19 0.70 −0.32 0.27

Table 3.5: Butcher tableau of ARK4 method

As a consequence, only one factorization is necessary for each time-step. Finally,

according to Eq. 3.23 and Table 3.5, the system of equations for each stage at each time-
step is defined as solving for Eq. 3.1:
³ ´
(I − γ∆t A)k i = f u n + ij−1
=1 a i j ∆t k j , t n + c i ∆t for 2 ≤ i ≤ 6, (3.27)
P

where the matrix I is the identity matrix.

This ESDIRK method has been chosen for several reasons. This method is A-stable,
stiffly accurate, and has a stage order equal to two. The stage order strongly influences
the accuracy of the method when applied to stiff problems. Besides, this approach is L-
stable, which is beneficial when handling the stiffness of the problem. Finally, adaptive
time-stepping strategies can be applied.

3.4. S OLVERS
As developed before, systems of equations have to be solved to obtain the time domain
solution. In the literature, two types of systems of equations exist:

• linear;

• non-linear.

Linear systems of equations are easy to solve, and many solvers are available[7, 8]. How-
ever, non-linear systems of equations require more computations than linear systems.
In general, the Newton-Raphson method is applied in such cases.
 44 3. M ATHEMATICAL BACKGROUND

3.4.1. L INEAR SOLVERS

There are two types of linear solvers - direct and iterative. Direct methods are mostly
used for small sets of equations, and iterative ones are used for larger systems of equa-
tions. In this section, the following equation is considered:

Ax = b, (3.28)
where the matrix A represents the coefficient of the set of equations, the vector b is the
input vector, and the vector x is the unknown vector. Also, re-ordering methods can be
3 applied to the matrix A to reduce the number of non-zero elements of the decomposi-
tion.

D IRECT METHODS
The first direct solver studied in this thesis is the LU factorization or Gaussian elimi-
nation [16]. This method decomposes the matrix A into a lower and upper triangular
matrices, L and U . The process to solve the Eq. 3.28 by the LU factorization is:

• Computing L and U such as:

LU = A; (3.29)

• Solving y from: Ly = b (Forward substitution);

• Solving x from: Ux = y (Backward substitution).

In the case of the Cholesky factorization, the principle is to compute the decomposition
of a matrix A which should be symmetric and definite positive, such as:

LL T = A. (3.30)
It appears that the Cholesky decomposition is more efficient regarding memory us-
age and CPU time by a factor of two. However, the forward and backward substitution
computation times remain the same as that of the LU factorization. For example, the
following matrix A is symmetric definite positive:

2 −1 0
 

A =  −1 2 −1  , (3.31)
0 −1 2
Then, after applying the LU factorization to matrix A, the matrices L and U are:

1 0 0 2 −1 0
   
3
L =  0.5 1 0  U = 0 2 −1  . (3.32)
−2 4
0 3 1 0 0 3
In the case of Cholesky factorization, the matrix L is:
 p
2 0 0

q
 1 3
p 0 

L= (3.33)
 2 q2 
2 p2
0 3 3
3.4. S OLVERS 45

Another factorization used for time domain simulations of a power system is the KLU
factorization [17]. Its principle is to decompose the matrix A in matrices L and U . The
difference with the LU factorization is that a re-ordering method is applied to obtain a
block triangular form. This action on the matrix, if block triangular blocks are found, per-
mits the use of the Schur complement. As a consequence, it might decrease the compu-
tation time if triangular blocks are found during the factorization process. The Cholesky
and KLU factorizations are mainly used to solve the set of equations of the nodal analysis
method (Section 2.7.1).

I TERATIVE METHODS 3
A basic iterative method for solving Eq.3.28 is defined as follows [7]:

x k+1 = x k + M −1 r k , (3.34)
where M ≈ A and the vector r is the residual vector, and it is defined as:

r k = b − Ax k . (3.35)

Basic iterative methods efficiently converge to a suitable solution if:

• the matrix M is sufficiently close to the matrix A;

• M −1 r k is easy to compute.

The simplest method is the Richardson iteration method[7]. This approach defines the
matrix M as the identity matrix. To approximate the speed of convergence, the spectral
radius is defined by:

ρ = max|ei g I − M −1 A |, (3.36)
¡ ¢

Then, ρ measures the speed of convergence. When ρ << 1, the convergence is fast.
When ρ > 1, there is no convergence.
Advanced iterative methods, such as GMRES or Bi-CGSTAB [18, 19], are based on the
Krylov Subspace, defined as:

K i (A, r 0 ) = span(r 0 , Ar 0 , ..., A i −1 r 0 ), (3.37)

Thus, it is possible to generate the iterate such as:

x i ∈ x 0 + K i (A, r i ). (3.38)
To terminate the iteration process, stopping criteria should be applied. These crite-
ria are the maximum number of iterations, or when the residual error becomes insignif-
icant. Furthermore, the choice of the starting point (vector x 0 ) of the iteration impacts
the convergence. For example, when x 0 is close to the solution, then the convergence is
usually fast.
Advanced iterative methods use a preconditioner in their algorithm. The precondi-
tioner typically accelerates convergence. For example, the incomplete LU factorization
(iLU(`))[7] is used as a preconditioner for them. The variable ` defines the level of fill-in
 46 3. M ATHEMATICAL BACKGROUND

of the decomposition (` > 0 and when ` = 0, there is no fill-in). The principle of this
factorization is to obtain a good approximation of the matrix A of the form:

A = LU + R = M + R, (3.39)

where the matrix R is the residual matrix. Thus, the forward and backward substitutions
are used to solve a system of equations as with the LU factorization. However, the com-
puted solution is an approximation.

3 R E - ORDERING METHODS
It is typically useful to re-order the matrix A to obtain less non-zero elements during the
decomposition, and then to be faster during the forward and backward substitutions. In
the literature, several re-ordering methods are available - such as the Approximate Mini-
mum Degree (AMD), the Quotient Minimum Degree (QMD), and the Nested Dissection
(ND) [20–22]. The idea is to solve the following system of equations:

P AP T P x = P b, (3.40)

where the matrix P is the permutation matrix, and is the result of the re-ordering
method. Moreover, the matrix P is defined as

PPT = I. (3.41)

As a result, the solution of the system of equation is:

x = P T (P AP T )−1 P b. (3.42)

Reordering methods are useful for reducing the fill-in of the iLU(l ) factorization and
the number of non-zero elements of any factorization.

3.4.2. N EWTON -R APHSON METHOD

In the case of a non-linear system of equations, the Newton-Raphson method solves the
following problem:

f (x) = 0, (3.43)

where f : R → R and x ∈ R. To obtain the solution to this problem, the following equation
is iterated as:

f (x n−1 )
x n = x n−1 − , (3.44)
f 0 (x n−1 )
d f (x)
where the function f 0 (x) = d x . For converging, it is necessary to take into account
several factors. The primary factor for a fast convergence is to start close to the solution.
Another factor is the value of the derivative. When the derivative is large, the conver-
gence is faster. However, if the derivation is null, another starting point is chosen to
re-start the iterative process.
For example, the function f (x) is defined by:
3.4. S OLVERS 47

f (x) = e x − 4. (3.45)
Then, the Newton-Raphson iteration is applied to obtain the solution of Eq. 3.45
when x 0 = 3 and x 0 = 30. The results after several iterations are shown in Table 3.6.

x0 = 3 x 0 = 30
x 1 = 2.1991 x 1 = 29
x 2 = 1.6427 x 2 = 28
x 3 = 1.416
..
x 3 = 27
..
3
. .
x 10 = 1.3863 x 10 = 20

Table 3.6: Solution of the iterate for two starting points

The solution of this problem is x = l n(4) ≈ 1.386. As shown in Table 3.6, the starting
point has an influence on the speed of convergence, and so, on the computation time.
To compute the solution of a non-linear system of equations, the following equation
is considered:

f (x) = 0, (3.46)
where f : R n → R n and x ∈ R n . In order to solve a set of equations, the iterative process
of Eq. 3.44 becomes:

x n = x n−1 − J (x)−1 f (x), (3.47)

where the matrix J (x) is defined as:
∂ f (x)
J (x) = J = , (3.48)
∂x
and it is the Jacobian matrix. The same problems as Eq. 3.45 can be encountered, and
so a nearby starting point for the solution is advised. In the case of a singular Jacobian
matrix, a new starting point is also recommended. To solve the following linear system
of equations:
J (x)y = f (x), (3.49)
a linear system of equations needs to be solved. Either a direct or iterative solution
method can be employed.
In the case of non-linear power systems, the same principle for solving non-linear
equations is applied during the time integration method. Moreover, according to the
Butcher tableaux of an ESDIRK, for example (Table 3.4), Eq. 3.50 shows the ESDIRK
formulation derived from the Newton-Raphson iteration method at each stage if non-
linear equations are used:

Pi −1
(I − ∆t γJ )δki = k ik − f (t + c i ∆t , x n + ∆t
(
j =1 a i j k j )
f or 2 ≤ i ≤ s, (3.50)
k ik+1 = k ik − δki
 48 R EFERENCES

where the vector δki represents the error on the iterate. One of the stopping criteria of this
methodology is to stop it when kδki k becomes smaller than a specified defined tolerance.
Also, the Jacobian matrix is usually computed once at each time-step.

3.5. C ONCLUSION
This chapter demonstrates that the mathematical properties of the power system under
consideration depend on the modeling method used (Section 2.7) and the user spec-
ifications. For these reasons, short introductions to differential equations, numerical
3 integration methods, and solvers are offered.
A power system is usually stiff. Then, an implicit numerical method is advised. In
this thesis, the ESDIRK ARK4 is mainly used to integrate the ODE expression of the power
system under consideration. Then, the focus of this thesis in Chapters 6 and 7 is about
the use of linear solvers - especially when the power system becomes large and its inter-
action with the numerical integration method is used.

R EFERENCES
[1] P. Tipler, Physics for Scientists and Engineers (W. H. Freeman Worth Publishers, New
York, 1999).

[2] U. M. Ascher and L. R. Petzold, Computer Methods for Ordinary Differential Equa-
tions and Differential-Algebraic Equations, 1st ed. (Society for Industrial and Ap-
plied Mathematics, Philadelphia, 1998).

[3] L. Euler, Institutionum calculi integralis, 3rd ed., Institutionum calculi integralis
(Impensis Academiae Imperialis Scientiarum Petropolitanae, Petropoli, 1824).

[4] C. Runge, Uber die numerische auflosing von differentialgleichungen, 46, 167
(1895).

[5] W. Kutta, Beitrag zur naherungsweisen integration totaler differentialgleichungen,

46, 435 (1901).

[6] L. van der Sluis, Transients in Power Systems (John Wiley & Sons, Chichester, 2001).

[7] Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd ed. (Society for Industrial
and Applied Mathematics, Philadelphia, 2003).

[8] J. M. Ortega and W. C. Rheinboldt, Iterative solution of nonlinear equations in sev-

eral variables, Vol. 30 (Society for Industrial and Applied Mathematics, New York,
1970).

[9] K. Brenan, S. Campbell, and L. Petzold, Numerical Solution of Initial-Value Prob-

lems in Differential-Algebraic Equation (North-holland, Amsterdam, 1989).

[10] J. Butcher, A history of Runge-Kutta methods, Applied Numerical Mathematics 20,

247 (1996).
R EFERENCES 49

[11] R. DeCarlo and P.-M. Lin, Linear Circuit Analysis, 2nd ed. (OXFORD UNIVERSITY
PRESS, New York, 2001).

[12] J. Butcher, On the implementation of implicit Runge-Kutta methods, BIT Numerical

Mathematics 16, 237 (1976).

[13] C. A. Kennedy and M. H. Carpenter, Diagonally implicit Runge-Kutta methods for

ordinary differential equations. a review, (2016).

[14] W. Hundsdorfer and J. Verwer, Numerical Solution of Time-Dependent Advection-

Diffusion-Reaction Equations (Springer, New York, 2003). 3
[15] C. A. Kennedy and M. H. Carpenter, Additive Runge-Kutta schemes for convection-
diffusion-reaction equations, Applied Numerical Mathematics 44, 139 (2003).

[16] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to algorithms,

Vol. 6 (MIT press Cambridge, Cambridge, 2001).

[17] T. A. Davis and E. Palamadai Natarajan, Algorithm 907: KLU, a direct sparse solver for
circuit simulation problems, ACM Transactions on Mathematical Software (TOMS)
37, 36 (2010).

[18] Y. Saad and M. H. Schultz, GMRES: A generalized minimal residual algorithm for
solving nonsymmetric linear systems, SIAM Journal on scientific and statistical com-
puting 7, 856 (1986).

[19] H. A. Van der Vorst, Bi-CGSTAB: A fast and smoothly converging variant of bi-cg for
the solution of nonsymmetric linear systems, SIAM Journal on scientific and Statis-
tical Computing 13, 631 (1992).

[20] P. R. Amestoy, T. A. Davis, and I. S. Duff, An approximate minimum degree ordering

algorithm, SIAM Journal on Matrix Analysis and Applications 17, 886 (1996).

[21] A. George and J. W. Liu, A fast implementation of the minimum degree algorithm
using quotient graphs, ACM Transactions on Mathematical Software (TOMS) 6, 337
(1980).

[22] B. Hendrickson and E. Rothberg, Improving the run time and quality of nested dis-
section ordering, SIAM Journal on Scientific Computing 20, 468 (1998).
 4
S WITCHING ACTION NETWORK
CALCULATION

4.1. I NTRODUCTION
A B asic power system is composed of generators, transmission lines and loads - as ex-
plained in Chapter 1. They all have their mathematical representation independent
of each other. Also, switching devices connect these power system components for the
configuration or the reliability of the electrical network, as shown in Section 2.8.1.
Switching action can occur - for exmaple - due to a change of topology, a short-
circuit, or when switching a transmission line off. Then, fast transients are visible [1, 2].
These transients are composed of high-frequency oscillations until the power system
reaches a new steady state value.
The literature on transient calculations is vast [3–5]. A significant amount of com-
puter programs is also available [6–10]. However, their main drawback is the full re-
computation of the power system equations or the introduction of non-linearities in the
power system equations. In both cases, the overall computation of the time domain sim-
ulation could be less expensive if an optimization method is applied to re-compute the
set of equations after a switching action.
The idea behind this new approach is to only update the necessary elements of the
differential equations after a switching action. Also, this approach should not introduce
unnecessary non-linearities [10]. Finally, the power system should be modeled via a
state-space representation or Ordinary Differential Equations (ODEs) [11], and the use
of non-ideal switches [2, 4, 12].
The choice of a state-space representation has been made for several reasons. The
first one is that there are a significant number of methods to obtain the time domain so-
lution of ODEs numerically, as shown in Chapter 3 [11, 13–16]. Secondly, it is possible to
have an adaptive time-stepping algorithm to allow a small time-step during transients
and a large one during steady-states [16]. Finally, the studies of power system compo-
nents are usually made in a state-space representation [17–19].

51
 52 4. S WITCHING ACTION NETWORK CALCULATION

In the literature review of Chapter 2, three main modeling methods are available. The
first one is the nodal analysis method [20]. H. Dommel proposed this approach in the
late 1960s. This method is appreciated, for example, for the simulation of large-scale
power systems [21]. The second method is the modified nodal analysis method [22].
Chung-Wen Ho et al. published this method in 1975. Several studies of large-scale power
systems are made with this modeling method [23]. The final one is the cut-set method
[24]. E. S. Kuh and R. A. Rohrer popularized this in 1969 [25]. However, this modeling
method is not popular for the simulation of large-scale power systems.
As discussed previously, the nodal analysis method is mainly applied to simulate
large-scale power systems, and it is based on Kirchhoff’s current law. It relies on the
nodal admittance matrix in combination with a numerical integration method (the
Trapezoidal rule)[2, 3]. The main problems of this modeling approach are the fixed time-
steps imposed by the numerical integration method, and also the difficulty in introduc-
4 ing non-linear elements. Finally, this approach is not applicable if an ODE representa-
tion is used. However, the fast calculation of the admittance matrix is the main advan-
tage of this method - especially if only resistances are taken into account. Besides, the
admittance matrix does not vary with the time-step chosen [26].
The modified nodal analysis is based on Kirchhoff’s current and voltage laws. This
modeling approach then gives a Differential Algebraic Equation (DAE). One of the major
computer programs for this method is the Simulation Program with Integrated Circuit
Emphasis (SPICE) [8, 27]. It is used for the simulation of electronic devices or some-
times, power electronic devices. In the literature, people studied the switching of capac-
itors with this modeling approach [28, 29]. Also, these studies are mostly for electronic
and signal applications.
The cut-set method is rarely used for the simulation of large-scale power systems.
However, this approach gives a state space representation. The reason for this is its
methodology to compute the set of equations, from the cut-set of its graph represen-
tation of its electrical diagram [4]. After a switching action, the re-computation of the set
of equations necessitates the restarting of the cut-set method. Another approach is to
consider a switch as a controlled voltage source [10]. However, this approach includes
non-necessary non-linear equations.
The ODE representation can be obtained from the DAE formulation, as shown in
Chapter 3 [30, 31]. However, after a switching action, this transformation is necessary
and that made this method expensive. Converter studies have shown that it is possible
to write their state space representation according to the position of the switch [32–34].
However, for large-scale power systems with n switches, it is useless to compute and
store n 2 state space representations. In paper [19], the authors studied a methodology
for coupling state representation of a power system. They couple the differential equa-
tions via the nodal voltages and currents at each component terminal.
As shown before, the studies of power system components are usually made in a
state-space representation. Then, the idea is to combine these mathematical descrip-
tions together in a larger state space representation, which represents the power system.
Finally, the link to couple those differential equations is to use non-ideal switches to
connect the different power system components.
These switching devices have stray capacitances to ground, as it is the case in prac-
4.2. B LOCK MODELING METHOD METHODOLOGY 53

tice [1]. Then, differential variables are introduced, and they represent the voltage at
each terminal of each power system component in the state space representation. Af-
ter that, the computation of the switching admittance matrix (which was derived from
the admittance matrix of a resistive network) helps to calculate the current leaving each
component terminal. The global mathematical description can then be easily updated
after a switching action via the new switching admittance matrix. The difference with
[19] is that this methodology updates the state space representation of only a few equa-
tions according to the switch topology for interconnected power system components.
From the review of problems and advantages of the different modeling methods, this
approach overcomes the considerable computer time required for building the set of
equations needed by the cut-set method. In addition, it is as simple as establishing the
nodal admittance matrix after a switching action. At the same time, this approach gives
a state space representation. The novelty is the methodology being used to connect dif-
ferent components of a power system by using the switches’ position and updating only 4
necessary equations. Also, this method can easily handle non-linear power system ele-
ments such as arc models or variable lumped elements.
This chapter is organized as follows. Section 4.2 provides a first methodology to use
this new approach. For more complex connection, the methodology is demonstrated in
Section 4.3. Section 4.4 shows the application of the block modeling methodology with
an analytical example. Then, Section 4.5 shows the utilization of time-dependent and
non-linear electrical power systems. Finally, Section 4.6 offers the conclusions.

4.2. B LOCK MODELING METHOD METHODOLOGY

4.2.1. I NTRODUCTION
A power system contains various types of components - like generators, transformers,
loads, transmission lines, and cables (Section 2.8) [35]. Each of them can be represented
by a network of lumped elements (e.g. resistors, inductors, capacitors, etc.) and a set of
differential equations can then be derived (Table 2.3 ).
Other important devices, which are essential for power systems operation and sus-
tainability, are switching devices [2]. In general, each terminal of a power system com-
ponent is connected to one or more switches. A switch has two terminals, and it can
be either open or closed. The open position isolates its terminals. When the switching
device is closed, the connection between the two terminals is made. In this chapter, a
closed one is represented by a resistor with a small value of resistance. Finally, every
switch terminal has a parasitic capacitance.
The idea is now to connect the power system components, called block models, with
each other via their terminals. The connection between them is made via an intercon-
nection of switching devices. After a switching action, traditional methods require the
re-computation of the complete set of equations. This approach updates only a few
numbers of differential equations. To use the method, the block models’ conditions are:
• Each block model has its own ODE representation when it is uncoupled from other
blocks;
• Each block model (except when the block represents a voltage source or a capaci-
tor) consists of a loop of lumped elements;
 54 4. S WITCHING ACTION NETWORK CALCULATION

• Each block model has at least one terminal;

• Each terminal has a capacitance (real or parasitic from switches) called a link ca-
pacitor or a voltage source connected in parallel.

According to these conditions, the voltage at each terminal, called link voltage, is present
as a differential or input variable. Then, these voltages assist in the computation of the
current flowing between block models by taking into account the switches’ position. Fi-
nally, the updation of the equations is done from the calculation of the admittance ma-
trix between block models.

4.2.2. M ATHEMATICAL FORMULATION

Now, a power system with n block models is considered as an example. The mathemati-
4 cal formulation of the general state space representation is:

dx
= ẋ = Ax + B g (t ). (4.1)
dt
Then, it is beneficial to write Eq. 4.1 as:

ẋ = ( Â + Ã)x + (B̂ + B̃ )g (t ), (4.2)

where matrices  and B̂ are constant during the simulation, and matrices à and B̃ are
related to switching actions. As a consequence, some elements of matrices à and B̃ will
change after a switching action in the electrical network.
In addition, matrices  and B̂ are block diagonal. These block diagonal matrices
are composed of the various  ∗ and B̂ ∗ block matrices of the different block models
that form the electrical network. Also, matrices  and B̂ , and vectors x and g (t ) of an
electrical
 diagram of n block models can be expressed  as:
 1,p 1 ×p 1 0p 1 ×p 2 ··· 0p 1 ×p n
.. x1
 
 .. 
 0p 2 ×p 1

 2,p 2 ×p 2 . .

 . 
 =  , x =  .. ,

.. .. ..
. . . 0p n−1 ×p n 
 
 xn
0p ×p 1
n ··· 0p ×p n−1 Â n,p n ×p n
n

B̂ 1,p 1 ×s1 0p 1 ×s2 0p 1 ×sn

 
···
.. g 1 (t )
 
 .. 
 0p 2 ×s1

B̂ 2,p 2 ×s2 . .

..
B̂ =  , and g (t ) =  .
  
.. .. .. .
. . . 0p n−1 ×sn 
 
 g n (t )
0p n ×s1 ··· 0p n ×sn−1 B̂ n,,p n ×sn
The state space representation after applying the cut-set method is computationally
expensive, especially for large power systems. This occurs because large linear systems
of equations need to be built during the computational process when the topology of
the power system changes. For example, an electrical network composed of E voltage
sources, R l resistors, R t resistors, C capacitors, and L inductors requires to solve 5C lin-
ear systems of equations of size C , 5L systems of equations of size L, C + R l + E linear
systems of equations of size R l and L + R t + E linear systems of equations of size R t . This
task is rather time-consuming.
4.2. B LOCK MODELING METHOD METHODOLOGY 55

The same approach can also be applied for the computation of the ODE representa-
tion of the network from a DAE representation of index one. A power system composed
of E voltage sources, R resistors, C capacitors, and L inductors requires to solve one lin-
ear system of equations of size R by applying the Schur complement [36]. This approach
is less time-consuming than the cut-set method.

The block modeling method gives the same state space representation as the cut-set
method or the DAE transformation. The difference is that the block modeling approach
only requires the differential equations of the block models and how they are connected
with each other. Thus, it is computationally cheap because it needs to solve only small
set of equations in some cases, as explained in Section 4.3. Meanwhile, in the case of
unknown block equations, these two approaches can be applied and also be perfomed
in parallel because each block is independent of each other. 4
When a switching action takes place in the electrical network, the whole process has
to be repeated for the cut-set method or DAE transformation, bus this is not necessary
for the block modeling method. After a switching action, only few differential equations
change in general. As a result, a limited number of elements of the matrices à or B̃ have
to be updated.

4.2.3. B LOCK MODEL REPRESENTATION AND PARAMETERS

In this section, the block model representation of linear power system elements - which
are presented in Section 2.8 - are developed according to the block modeling mathemat-
ical approach in Table sec-BMM:MF:tab:PSC and Table 4.1. In addition, a voltage source
and capacitor representation are given too.

Power
Number of
system Link voltages Link capacitors
terminals
components
Generator 1 Vc C
Load 1 Vc C
PI section 2 Vc1 ,Vc2 C 1 ,C 2
Double PI
2 Vc1 ,Vc2 C 1 ,C 2
section
Voltage
1 e(t )
source
Capacitance 1 Vc C

Table 4.1: Block model link voltage and link capacitor representation
 56 4. S WITCHING ACTION NETWORK CALCULATION

Power
system Symbol  B̂ x g(t)
components
1
Generator G ÂG = AG B̂G = BG x̂G = xG ĝ G (t ) = g G (t )

1
Load L Â L = A L B̂ L2×0 x̂ L = x L ĝ L (t )0×1

1 2
PI section PI Â P I = A P I B̂ P3×0
I x̂ P I = x P I ĝ P I (t )0×1

Double PI 1 2
PII Â P I I = A P I I B̂ P5×0
II x̂ P I I = x P I I ĝ P I I (t )0×1
section
4 e(t )
Voltage 1 ÂV0×0 B̂V5×0 x̂V0×1 ĝ V (t ) = e(t )
source

C
Capacitance 1 ÂC = 0 B̂C1×0 x̂C = VC ĝ C (t )0×1

Table 4.2: Block model symbol and state space representation

4.2.4. F IRST EXAMPLE

The following sample network diagram and block diagram of Figure 4.1 illustrates the
block modelling method.

RG LG S w (α) RL LL
i LG i LL

e(t ) v CG CG RC G vC L CL

(a) Lumped representation of a generator (left) connected to a load (right) via a switching device
(S w )

G L
S w (α)
(RG , LG , RCG , (R L , L L ,C L )
CG , e(t ))

(b) Block representation of a generator (left) connected to a load (right) via switching device (S w )

Figure 4.1: Lumped elements and a block model representation of a generator (left) connected to a load (right)
via a switching device (S w )
4.2. B LOCK MODELING METHOD METHODOLOGY 57

The state space representation of this network, when the switch is open, matrices A
and B are equal to  and B̂ , respectively, and given by:

ÂG,2×2 02×2 xG B̂G,2×1 02×0 g G (t )

· ¸· ¸ · ¸· ¸
ẋ = + , (4.3)
02×2 Â L,2×2 xL 02×1 B̂ L,2×0 g L (t )

which leads to:

R
− L1
 
− LG 0 0   1 
i

G G L G LG
1 1
−R 0 0   vC  
 
0

CG C G CG

G
ẋ =    vC  +   e(t ). (4.4)
 
− C1
 
 0 0 0 L  0 
L
 
0 0 1
− RL L i LL 0
LL L

Now, the interconnection of the generator and the load is made via a resistor R sw 4
which represents the closed switching device. Then, a current flows from the genera-
tor to the load, and the differential equations of the link capacitors C g and C L change.
Besides, the current flow is a function of R sw and of the link voltages vCG and vC L . The
differential variable equations of the generator then become:

RG 1 1
i˙LG = − i LG − v CG + e(t ), (4.5)
LG LG LG
v CG − v C L
¡ ¢
1 1
v̇CG = iL − vC − , (4.6)
C G G RC G C G G R sw CG

and the load block differential variable equations are now:

v CG − v C L
¡ ¢
1
v̇C L = − iL , (4.7)
R sw CG CL L
1 RL
i˙LG = vC − iL . (4.8)
LL L LL L

Then, the new state space representation of the block model representation, when
the switch is closed, becomes:
 R
− L1

− LG 0 0 
i LG
  1 
 1G G LG
− R 1C − R sw1C 1
0   v CG  

 0 
 C 
G CG G G R sw CG
ẋ =  1 + e(t ). (4.9)
− R sw1C − C1 
   
 vC L
  0 
 0
 
R sw C L

L L
0 0 1
− RL i LL 0
LL LL

By subtracting Eq. 4.10 to 4.4, the matrices à and B̃ of this block model diagram are:

0 0 0 0
 
0
 
 0 − R swαC α
R sw CG 0   0 
à =  G
 and B̃ = 
 0 ,
 
α
− R swαC

0 R sw C L 0
L
 
0 0 0 0 0
 58 4. S WITCHING ACTION NETWORK CALCULATION

where α represents the position of the switch. When α = 1, the switching device is
closed. Otherwise, α = 0 and the switch is open. As mentioned before, the calculation
of the two states does not require solving a linear system of equations. From the matrix
Â, a pseudo admittance matrix of the connection between the two block models can be
observed. The nodal equations of the connection between the block models are:

α
− Rαsw
" #·
v CG i CG
¸ · ¸
R sw
− = . (4.10)
− Rαsw α
R sw vC L iCL

In addition, the value and the position in the state variable of the link capacitors are
necessary.

4
4.3. C ONNECTING BLOCK MODEL MATRICES
The following two block representations are now studied.

G L
S sw1
(RG , LG , RCG , (R L , L L ,C L )
CG , e(t ))

G L
S sw2
(RG , LG , RCG , (R L , L L ,C L )
CG , e(t ))

(a) Block representation of two electrical networks, composed of a generator (left) connected to a
load (right) via switching devices

G L
S sw1 S sw2

(RG , LG , RCG , (R L , L L ,C L )
CG , e(t )) S sw4 S sw3

G L

(RG , LG , RCG , (R L , L L ,C L )
CG , e(t ))

(b) Block representation of one network, composed of two generators (left) connected to a loads
(right) via an interconnection of the switching devices S sw1 to S sw4 to an additional electrical
node

Figure 4.2: Two examples of possible connections between block models

4.3. C ONNECTING BLOCK MODEL MATRICES 59

The state space representation of Figure 4.2a is:

ẋ =
R
− LG − L1
 
0 0 0 0 0 0
G G
1 1  i LG

A 2,2 0 0 0 0 0

CG R sw CG
 
vCG
 
1
− R 1C − C1

0 0 0 0 0
  
R sw C L sw L vC L
 
L 
R
 
1
− LL

0 0 0 0 0 0 i LL
 
LL

L
  
R
 
− LG − L1 i LG

0 0 0 0 0 0
  
G G
 
vCG

1 1
 
0 0 0 0 A 6,6 0
  
CG R sw CG

vC L
 
1
− R 1C − C1

0 0 0 0 0
 
R sw C L iLL
 
 sw L L 
1 R
0 0 0 0 0 0 LL − LL
L
1
 
0
 LG
 0 0

 0

0


 4
 0 0  e(t )
 · ¸
+ 1  e(t ) , (4.11)

 0
 LG 
 0 0
 

 0 0
 

0 0
1
where A 2,2 = A 4,4 = − R − R sw1C and the ODE representation of Figure 4.2b is:
C G CG G

ẋ =
R
− LG − L1 0 0 0 0 0 0
 
G G
1 (G 1 ∗G 2 ) (G ∗G ) (G ∗G )
− G 1 C3 − G 1 C4

A 2,2 0 0 0

i LG

CG G t ot CG
 
 t ot G t ot G 
(G 1 ∗G 2 ) G 2 −G 2 vCG
 
− G C2 − C1 − GG2∗G3 − GG2∗G4

0 0 0
  
G t ot C L t ot L C
t ot G C
t ot G vC L
 
L 
R
 
1
− LL

0 0 0 0 0 0 iLL
 
LL

L
  
R
 
− LG − L1 i LG

0 0 0 0 0 0
  
G G
 
vCG

G ∗G G ∗G 1 G 3 ∗G 4
 
0 −R1 C3 −R2 C3 0 A 6,6 0
  
CG G t ot CG

sw G sw G vC L
  
G 4 −G 2
 
G ∗G G ∗G G 3 ∗G 4
− G 1 C4 − G 2 C4 − G C4 − C1 i LL
 
 0 0 0 G t ot C L

 t ot G t ot G t ot L L 
1 R
0 0 0 0 0 0 LL − LL
L
1
 
LG 0
0 0
 
 
0 0
 
 
0 0  e(t )
 · ¸

+ 1  e(t ) , (4.12)


 0 LG 
0 0
 
 
0 0
 
 
0 0

G t ot = G 1 +G 2 +G 3 +G 4 if α1 + α2 + α3 + α4 6= 0,
½
G 1 −G 2
where , A 2,2 = − R 1C − G t ot C 1 and
G t ot = 1 else C G G G

1 G 3 −G 2
A 6,6 = − R − G t ot C 3 .
C G CG G
From Eq. 4.11 and Figure 4.2a, there is no connection possible between the two sub-
networks. Then, when one of the switches changes position, it will not affect the differ-
ential equations of the other sub-network. For this reason, there are two switch blocks.
On Figure 4.2b, there is an extra-node to connect the block models. In Eq. 4.12, a con-
 60 4. S WITCHING ACTION NETWORK CALCULATION

nection exists between all block models. Then, there is one switch block. In addition,
such cases with an extra-node require the solving of a linear set of equations of size one.
From the previous electrical networks, a switch block is an interconnection of nbs
switches between two or more block models. Each switch block has nbn nodes that
correspond to the nbb block models connected via the switch block, and to k additional
nodes from the topology of switching elements (nbn = nbb + k). Finally, a complete
block diagram may contain more than one switch block.
Then, the incidence matrix KC L of every switch block is defined as:


 +1 if the current through the switch j
is connected to and oriented towards the node m




if the current through the switch j

−1

KC L (m, j ) = ,
4 
 is connected and not oriented towards the node m
0 if switch j




is not connected to node m



to satisfy the following equation:

KC L i s = −i , (4.13)

where i s = [i s1 · · · i snbs ]T and i = [i t 0]T . The vector i t corresponds with the current leav-
ing each block model terminal connected via the switch block. Also, the vector i t is or-
ganized such that the first components of the vector correspond to the currents through
the nbs link capacitors, and next, the current from the nbe voltage sources.
During the initialization, the incidence matrix of each block switch is computed and
stored. These matrices will be used for the initial state space representation, and reused
when one of their switching devices changes position. Also, each block switch is as-
sociated with a diagonal conductance matrix G, where each element of the diagonal
is composed of the admittance of corresponding switching devices (S w n − > G(n, n) =
αn /R sw n = G n for example).
The incidence matrix and the conductivity of the switches are the inputs to calcu-
late the admittance matrix by using the same rules as those to compute the admittance
matrix of the nodal analysis method. Then, from this matrix it is possible to update the
matrices à and B̃ . However, a block model can be connected to only one additional extra
node via switches, but it can also be connected with many block models. Finally, three
steps are used for each switch block during the initialization, and when a switching ac-
tion takes place in this particular switch block. These steps are specified in the following
subsections.

4.3.1. F IRST STEP : F ULL ADMITTANCE MATRIX

The full admittance matrix (Y f ul l ) is computed as:

Y f ul l v = KC L GKCTL v = −i , (4.14)

where G = d i ag ([G sw 1 (α1 ) · · ·G sw nbs (αnbs )]) and v = [v t v n ]T .

4.3. C ONNECTING BLOCK MODEL MATRICES 61

4.3.2. S ECOND STEP : R EDUCTION OF THE FULL ADMITTANCE MATRIX

This step is not necessary if the number of nodes of the block switch is equal to the num-
ber of terminals (nbb = nbn). In that case, the admittance matrix Y is simply defined as:

Y = −Y f ul l . (4.15)

Now, when the number of nodes of the block switch is different from the number
of terminals (nbb¬nbn), the Schur complement method [36] is applied to compute the
matrix Y as:

Y f ul l 1 Y f ul l 2 vt it
· ¸· ¸ · ¸
Y f ul l v = =− , (4.16)
Y f ul l 3 Y f ul l 4 vn 0

which gives: 4
Y v t = −i t (4.17)

where:
Y = −Y f ul l 1 + Y f ul l 2 Y f−1
ul l 4 Y f ul l 3 . (4.18)

It is, however, not possible to form the Schur complement when:

• A switch is open between a terminal and an extra node, or between two additional
nodes. Then, zeros may appear on the diagonal of the matrix Y f ul l 4 ;

• A switching device is closed between two extra nodes, and neither of these addi-
tional nodes is connected via a closed switch to a terminal. Then, a singularity may
appear in the matrix Y f ul l 4 .

To circumvent these problems, the following actions are taken:

• For the first problem, zeros on the diagonal of the matrix Y f ul l 4 can be replaced by
ones. This action will not affect the resulting matrix Y , and it is then possible to
form the Schur complement.

• For the second problem, this particular switch is then considered open because
there is no current flowing by it. Then, it does not play a role for the computation
of the differential equations, and the first issue is present, and has already been
tackled.

Another way to overcome these problems is to connect a small capacitance to each ex-
tra node, and the Schur complement does not have to be formed any more (Y f ul l = Y ).
However, with this approach, the number of differential variables increases and the ma-
trix A can be bad conditioning and lead to problems during the integration process. The
Schur complement is only used when the number of nodes is larger than the number of
block models connected through a switch block. Finally, when a switching action takes
place, the Schur complement is applied on a small scale because only the switch block
containing the switching action is taken into account.
 62 4. S WITCHING ACTION NETWORK CALCULATION

4.3.3. T HIRD STEP : U PDATING THE CONNECTING BLOCK MATRICES

From the admittance matrix Y , the matrices à and B̃ can be updated where nbb block
models are connected via the switch block - which includes nbe voltage source block
models like:
1
à I 1 (m,l )I 1 (m, j ) = Yl j , (4.19)
C (m, l )
for 1 ≤ l , j ≤ nbb −nbe where I 1 (m, l ) gives the position of the link capacitor of the block
model l , at the switch block m, in state vector x, and C (m, l ) contains the value of the
link capacitor of block model l at the switch block m.

1
B̃ I 1 (m,l )I 2 (m, j ) = Yl ( j +nbb−1) (4.20)
C (m, l )
4 for nbb − nbe ≤ j ≤ nbb where I 2 (m, j ) gives the position of the voltage source in the
time-input vector, at the switch block m.

4.4. E XAMPLE OF AN ANALYTICAL SOLUTION OF A SWITCH

BLOCK
In this section, a block model diagram composed of one switch block is investigated. The
results achieved by the block modeling method are identical to the results from Kirch-
hoff’s laws. As explained before, all switches are non-ideal. For this demonstration, the
block diagram of Figure (4.3) is considered.

C1 C2
1 2
Sw2 (α2 ) Sw3 (α3 )
5

Sw6 (α6 ) Sw5 (α5 )

Sw1 (α1 ) 7 8 Sw4 (α4 )

Sw7 (α7 ) Sw8 (α8 )

C3
6
Sw10 (α10 ) Sw9 (α9 )
3 4
e(t )
Figure 4.3: Electrical block diagram of a switch block with eight switching devices
4.4. E XAMPLE OF AN ANALYTICAL SOLUTION OF A SWITCH BLOCK 63

4.4.1. A NALYTICAL SOLUTION

From the block diagram of Figure (4.3), it is possible to write x = [vC 1 vC 2 vC 3 ]T , Â = 03×3 ,
B̂ = 03×3 and g (t ) = e(t ). The analytical solution for the current through each capacitor is
related to the position (closed or open) of each switching device. This results in a rather
complicated formula. For the demonstration, α1 = α2 = α3 = α5 = α7 = 1 is considered
and the other values of αi are zero. The solution with Kirchhoff’s laws is:
G 2 VC 1 +G 3 VC 2
 ³ ´
 v̇C 1 = G 1 (vC 3 − vC 1 ) −G 2 vC 1 +G 2 G +G /C 1

2 3

G 2 VC 1 +G 3 VC 2 . (4.21)
³ ´
v̇C 2 = −G 3 vC 2 +G 3 /C 2

 ¢ G 2 +G 3
v̇C 3 = G 1 (vC 1 − vC 3 ) /C 3
 ¡

4.4.2. B LOCK - MODELING SOLUTION

According to Figure (4.3), the block switch parameters are nbb = 4, nbe = 1, nbs = 10, 4
nbn = 8, and m = 1. In addition, the state vector employed defines I 1 (1, 1) = 1, I 1 (1, 2) =
2, I 1 (1, 3) = 3, I 2 (1, 1) = 1, C (1, 1) = C 1 , C (1, 2) = C 2 and C (1, 3) = C 3 . Initially, the matrix
KC L is computed, as defined in Section 4.3, as:

1 1 0 0 0 0 0 0 0 0
 
 0 0 1 1 0 0 0 0 0 0 
 
 −1 0 0 0 0 0 0 0 0 1 
 
 0 0 0 −1 0 0 0 0 1 0 
 
KC L =  , (4.22)
 0 −1 −1 0 1 1 0 0 0 0 
 0 0 0 0 0 0 −1 1 −1 −1 
 
 0 0 0 0 0 −1 1 0 0 0 
 

0 0 0 0 −1 0 0 −1 0 0

and the matrix G(α) is defined as:

G(α) = diag G1 G2 G3 G4 G5 G6 G7 G8 G9 G 10 . (4.23)

¡£ ¤¢

¤T
In addition, the following vector is considered i t = i c1 i c2 i c3 i e . Now, the
£

three steps necessary to obtain the updated matrices à and B̃ according to Section 4.3
are applied as shown below.

F IRST STEP : F ULL ADMITTANCE MATRIX

By applying Eq. 4.14, Eq.4.24 is found:

G 1 +G 2 0 −G 1 0 −G 2 0 0 0
 

 0 G3 0 0 −G 3 0 0 0 


 −G 1 0 G1 0 0 0 0 0 

0 0 0 0 0 0 0 0
 
Y f ul l = (4.24)
 
−G 2 −G 3 0 0 G 2 +G 3 +G 5 0 0 −G 5

 
0 0 0 0 0 G7 −G 7 0
 
 
0 0 0 0 0 G7 0
 
 −G 7 
0 0 0 0 −G 5 0 0 G5
 64 4. S WITCHING ACTION NETWORK CALCULATION

S ECOND STEP : R EDUCTION OF THE FULL ADMITTANCE MATRIX

In this case, nbn is different from nbb. Then, the Schur complement is used according
to the following sub-matrices and to Eq. 4.16:

G 1 +G 2 0 −G 1 0
 
 0 G3 0 0 
Y f ul l 1 =  , (4.25)
 −G 1 0 G1 0 
0 0 0 0
−G 2 −G 3 0 0
 
0 0 0 0 
Y f ul l 3 = Y fTul l 2 = 

, (4.26)
 0 0 0 0 
0 0 0 0
4 
G 2 +G 3 +G 5 0 0 −G 5

 0 G7 −G 7 0 
Y f ul l 4 =  . (4.27)
 0 −G 7 G7 0 
−G 5 0 0 G5

The matrix Y f ul l 4 is not invertible. Its determinant equals zero because the switch
labelled Sw 7 in Figure 4.3 is closed. Since this switch is connected to two extra nodes
and none of these nodes are linked to a terminal via a closed switch, it can be considered
as an open switch. Finally, Y f ul l 4 becomes:

G 2 +G 3 +G 5 0 0 −G 5
 
 0 0 0 0 
Y f ul l 4 =  . (4.28)
 0 0 0 0 
−G 5 0 0 G5

Now, there are several zeros on the diagonal. In line with the methodology, the zeros are
replaced by the number one:

G 2 +G 3 +G 5 0 0 −G 5
 
 0 1 0 0 
Y f ul l 4 =  . (4.29)
 0 0 1 0 
−G 5 0 0 G5

Finally by applying Eq. 4.18, the following admittance matrix Y is found:

G 22
 
G 2G 3
G 2 +G 3 −G 2 −G 1 G 2 +G 3 G1 0

G3 2 
G 2G 3
G 2 +G 3 −G 3 0 0 .
 
Y =
 G 2 +G 3  (4.30)
 G1 0 −G 1 0 
0 0 0 0
4.5. T IME - DEPENDENT AND NON - LINEAR ELECTRICAL NETWORKS 65

T HIRD STEP : U PDATING THE CONNECTING BLOCK MATRICES

Now by combining Eq. 4.19 and Eq. 4.20, the following state space representation is
found:
G2
 
2 G2G3
G 2 +G 3 −G 2 −G 1 G 2 +G 3 G1
C1 C1 C1
 
 v  
0

G2 C1

 G2G3 3

ẋ =  G 2 +G 3 G 2 +G 3 −G 3  vC 2  +  0  e(t ). (4.31)
 
C2 C2 0
 vC 3 0
 
G1 −G 1
0

C3 C3
 
0 0 0
The state space representation corresponds to the analytical solution found from
Kirchhoff’s laws (Eq. 4.21).

4.5. T IME - DEPENDENT AND NON - LINEAR ELECTRICAL NET- 4

WORKS
4.5.1. M ATHEMATICAL FORMULATION
Power systems are non-linear. For this reason, Eq. 4.2 can be developed, for example, as:

ẋ = ( Â + Ã)x + A non (x)x + (B̂ + B̃ )g (t ) + v(x), (4.32)

where the time dependent lumped element and non-linear functions are included via
the functions A non (x) and v(x).

4.5.2. I NTRODUCTION TO SOME TIME - DEPENDENT AND NON - LINEAR

ELECTRICAL NETWORKS
For example, the inductance of the load of Figure 4.1 is time-dependent (L L = l (t )).
Then, the mathematical equations of the load block model are now:

x˙L = AˆL x L + BˆL g L (t ) (4.33)

− C1
" #·
0 vC L
¸
= 1 R d l (t ) , (4.34)
l (t ) − l (t ) − l (t )d t i LL

and then, the new state space representation is:

AˆG 0 xG BˆG 0 g G (t )
·· ¸ ¸· ¸ ·· ¸ ¸· ¸
ẋ = + Ã + + B̃ (4.35)
0 AˆL xL 0 BˆL g L (t )
 R
1
   
G
− −L 0 0 0 0 0 0 i LG

 1LG G
− R 1C 0 0   0 − R αC α
0  vC
  
 C R sw C G  
G CG G G
= 
1
+ sw G
α α   
 vC L
 0 0 0 −   0 − 0   
 CL  R sw C L R sw C L
0 0 1
− lR(tL) − ld(tl)d
(t ) 0 0 0 0 i LL
l (t ) t
 1 
 LG 
 0 
+  e(t ). (4.36)
 0 
0
 66 4. S WITCHING ACTION NETWORK CALCULATION

In the second example, the switching device is now replaced by an arc model called
the Habedank model (Section 2.8.6)[37]. This arc model simulates the interruption of the
current through it by changing its conductivity. Its mathematical expressions are shown
from Eq. 2.53 to Eq. 2.56. From these equations, the mathematical representation of this
electrical network is:

R
− LG − L1 0 0 0 0
    1 
i L˙G
 
G G i LG L
vC˙ G
1
−R 1
0 0 0 0  v CG   G 
  0 
  
CG C G CG

   
vC˙ L − C1 vC L   0 
 
0 0 0 0 0 
    
 
 = +  e(t )
 
L
i L˙ 1
− RL L i LL   0 
 
 L
 
  0 0 LL 0 0 
L
 g˙c  gc   0 
    
  0 0 0 0 0 0 
g˙m 0 0 0 0 0 0 gm 0
4  
0
0 0 0 0 0 0
  
i LG
g ar c g ar c
 0 
 0 − C1 C1 0 0 0  v
  CG  
  
g ar c g 0
 
0 − Car2c 0 0 0   vC L  
    
+
 C2 + 0 ,

0 0 0 0 0 0   i LL  

  α (g ar c u ar c )2 − g
 ³ ´ 
0 0 0 0 0 0   g c   τc c
  
2
³ Uc g c 2
 
0 0 0 0 0 0 gm α (g ar c u ar c )
 ´ 
τm −g
Pm m
(4.37)

where α represents the state of the arc model. When α = 0, the arc model is inactive, else,
it is activated.
During the integration process, non-linear lumped elements will lead to time-
dependent or non-linear system of equations. Then, the Jacobian matrix is computed
to solve them with a Newton-Raphson iteration. The definition of the Jacobian matrix is:

∂ f (x)
J (x) = J = (4.38)
∂x
 ∂ Pp A non (1, j )x j ∂
Pp
A non (1, j )x j
 
j =1 j =1 ∂v(1) ∂v(1)

∂x 1 ··· ∂x p ∂x 1 ··· ∂x p
  
.. .. .. .. .. ..
   
= A + .
+ .

 . .   . . 
Pp Pp ∂v(p) ∂v(p)

∂ A non (p, j )x j ∂ A non (p, j )x j
 
j =1
···
j =1
∂x p ··· ∂x p
∂x p ∂x p

= A + A non + J non + J v . (4.39)

At each time-step, the Jacobian matrix has to be computed. Two methods are avail-
able to compute the equations - the numerical method and the analytical method. The
numerical method will require a larger amount of computation time than the analytical
one. For example, the Jacobian matrix of Figure 4.1, when the non-linear inductance is
used, is:
4.5. T IME - DEPENDENT AND NON - LINEAR ELECTRICAL NETWORKS 67

R
− L1
   
− LG 0 0 0 0 0 0
G G
1 1
−R 0 0   0 − R swαC α
0
   
CG C G CG R sw CG
 
J = + α
G .
− C1 − R swαC
 
 0 0 0   0 0 
 L
 R sw C L L 
0 0 1
l (t ) − lR(tL) − d l (t )
l (t )d t
0 0 0 0
(4.40)
In addition, the Jacobian matrix, when the arc model is present, is:

R
− L1
 
− LG 0 0 0 0
G G
1
− R 1C 0 0 0 0 
 
CG

 CG G 
0 0 0 − C1 0 0 
 
J =


0 0 1
− LL
R
L
0

0 
 4
 LL L
0 0 0 0 0 0 
 

0 0 0 0 0 0
0 0 0 0 0 0
 
αu ar c g2 αu ar c g2
0 − R αC α
0 − −
 

 ar c G R ar c CG CG (1,1)g c2 2
CG g m


α αu ar c g 2 αu ar c g 2
− R αC
 
+
 0 R ar c C L 0 
ar c L C L g c2 2
C L (1,2)g m 
0 0 0 0 0 0
 
 
 
 0 0 0 0 0 0 
0 0 0 0 0 0
0 0 0 0 0 0
 

 0 0 0 0 0 0 


 0 0 0 0 0 0 


+ 0 0 0 0 0 0 

2 g3
2g 2 u a r c 2g h2 u ar c 2
µ ¶

α ua r c gh
³
2g
´ α2u ar c h

 0 α −α 0 1− g −1
 
τc Uc2 g c τc Uc2 g h τ c 2 2 Uc g c m τµc Uc2 g c g m
2 
 ¶ 
2 g3 2 g3
 2g h2 u ar c 2g h2 u ar c α2u ar c h 2u ar c h

0 α τ P −α τ P 0 − τα 2 −1
m m m m τm P m g c2 m Pm g m
(4.41)

4.5.3. D ISCUSSION
Now, the illustration of time-dependent or non-linear elements has been shown in this
section. Also, the Jacobian computation is possible for such simple electrical networks.
However, when the lumped element value changes in time, the Jacobian matrix is the
same as the state space representation. In the case of the arc models, the state space
representation can be computed easily from the same information as that for a switching
device. Meanwhile, it is recommended to put the arc model between two block termi-
nals. Then, a methodology can be extrapolated from the previous example to compute
the state space representation and the Jacobian matrix. In fact, in order to compute the
Jacobian matrix, and the place of the state variables (which represent the block model
terminal), are necessary along with the position of the differential variables, which rep-
resent the different conductivities of the arc models.
 68 R EFERENCES

4.6. C ONCLUSION
In this chapter, a new approach to model power systems containing switching devices
is studied. The method couples the differential equations of the various components
of a power system via switches. This results in local updates of the system of equations
when switching actions occur. Besides, this approach allows for the parallelization of
computationnal effort of the cut-set method or the modified nodal analysis method, by
using these modeling approaches only on power system components, in oder to obtain
an ODE representation when block models are not known. Then, this methodology can
be applied to couple these block models. The conclusions of this approach are:

1. The method can be applied for most configurations of switches among power sys-
tem components;

4 2. This approach can be used to model power electronic converters;

3. Non-linear power systems can be handled.

Finally, on a supercomputer, this modeling method can also be implemented to increase

the size of the power system under consideration and to decrease the computation time.

R EFERENCES
[1] R. Smeets, L. Van der Sluis, M. Kapetanovic, D. F. Peelo, and A. Janssen, Switching in
electrical transmission and distribution systems (John Wiley & Sons city, Chichester,
2014).

[2] L. van der Sluis, Transients in Power Systems (John Wiley & Sons, Chichester, 2001).

[3] J. Vlach and K. Singhal, Computer methods for circuit analysis and design (Van Nos-
trand Reinhold, New York, 1983).

[4] N. Watson and J. Arrillaga, Power systems electromagnetic transients simulation,

Vol. 39 (Institution of Electrical Engineers, London, 2003).

[5] P. R. Adby, Applied circuit theory: matrix and computer methods, JOHN WILEY &
SONS, N. Y., 1980, 250 (1980).

[6] PSCAD/EMTDC: Electromagnetic transients program including dc systems (1994).

[7] Electromagnetic Transient Program EMTP RV (2007).

[8] A. Vladimirescu, The SPICE book (John Wiley & Sons, Inc., New York, 1994).

[9] N. Bijl and L. D. Van Sluis, New approach to the calculation of electrical transients,
part i: Theory, European transactions on electrical power 8, 175 (1998).

[10] MATLAB SimPowerSystems for Use with Simulink User’s Guide, Version 4.1.1.

[11] U. M. Ascher and L. R. Petzold, Computer Methods for Ordinary Differential Equa-
tions and Differential-Algebraic Equations, 1st ed. (Society for Industrial and Ap-
plied Mathematics, Philadelphia, 1998).
R EFERENCES 69

[12] R. DeCarlo and P.-M. Lin, Linear Circuit Analysis, 2nd ed. (OXFORD UNIVERSITY
PRESS, New York, 2001).

[13] K. E. Brenan, S. L. Campbell, and L. R. Petzold, Numerical solution of initial-value

problems in differential-algebraic equations, Vol. 14 (Society for Industrial and Ap-
plied Mathematics, Philadelphia, 1996).

[14] G. Wanner and E. Hairer, Solving ordinary differential equations II : Stiff and
Differential-Algebraic Problems, Vol. 1 (Springer Berlin Heidelberg, Berlin Heidel-
berg, 2010).

[15] C. A. Kennedy and M. H. Carpenter, Diagonally implicit Runge-Kutta methods for

ordinary differential equations. a review, (2016).

[16] W. Hundsdorfer and J. Verwer, Numerical Solution of Time-Dependent Advection- 4

Diffusion-Reaction Equations (Springer, New York, 2003).

[17] M. Vaezi and A. Izadian, Piecewise affine system identification of a hydraulic wind
power transfer system, IEEE Transactions on Control Systems Technology 23, 2077
(2015).

[18] A. Luna, J. Rocabert, J. I. Candela, J. R. Hermoso, R. Teodorescu, F. Blaabjerg, and

P. Rodríguez, Grid voltage synchronization for distributed generation systems under
grid fault conditions, IEEE Transactions on Industry Applications 51, 3414 (2015).

[19] M. Brucoli, F. Torelli, and M. Trovato, State space representation of interconnected

power systems for dynamic interaction studies, Electric Power Systems Research 5,
315 (1982).

[20] H. W. Dommel, Digital computer solution of electromagnetic transients in single-

and multiphase networks, Power Apparatus and Systems, IEEE Transactions on
PAS-88, 388 (1969).

[21] J. K. Debnath, A. M. Gole, and W.-K. Fung, Graphics processing unit based accelera-
tion of electromagnetic transients simulation, (2015).

[22] C. W. Ho, A. E. Ruehli, and P. A. Brennan, The modified nodal approach to network
analysis, Circuits and Systems, IEEE Transactions on 22, 504 (1975).

[23] Y. Oshima, H. Oguma, N. Ishihara, and K. Masu, Simulation and evaluation of pv

power generation for energy management system using spice, in 2016 International
Conference on Information and Communication Technology Convergence (ICTC)
(2016) pp. 74–77.

[24] T. Bashkow, The A matrix, new network description, IRE transactions on circuit The-
ory 4, 117 (1957).

[25] E. Kuh and R. Rohrer, The state-variable approach to network analysis, Proceedings
of the IEEE 53, 672 (1965).
 70 R EFERENCES

[26] A. Ioinovici, Computer-Aided Analysis of Active Circuits (MARCELL DEKKER, inc,

New York, 1990).

[27] F. Bizzarri, A. Brambilla, G. S. Gajani, and S. Banerjee, Simulation of real world

circuits: Extending conventional analysis methods to circuits described by hetero-
geneous languages, IEEE Circuits and Systems Magazine 14, 51 (2014).

[28] J. Vandewalle, H. De Man, and J. Rabaey, Time, frequency, and z-domain modified
nodal analysis of switched-capacitor networks, IEEE Transactions on circuits and
systems 28, 186 (1981).

[29] J. Xu and J. Yu, Equivalent circuit models of switches for SPICE simulation, Electron-
ics letters 24, 437 (1988).

4 [30] G. Verghese, B. Lévy, and T. Kailath, A generalized state-space for singular systems,
IEEE Transactions on Automatic Control 26, 811 (1981).

[31] G. M. Huang, K. Men, and X. Song, A new remodeling technique for power system
dynamic analysis, in 2005 IEEE/PES Transmission & Distribution Conference & Ex-
position: Asia and Pacific (IEEE, 2005) pp. 1–6.

[32] B. De Kelper, L. A. Dessaint, K. Al-Haddad, and H. Nakra, A comprehensive approach

to fixed-step simulation of switched circuits, IEEE Transactions on Power Electronics
17, 216 (2002).

[33] R. J. Dirkman, Generalized state space averaging, in Power Electronics Specialists

Conference, 1983 IEEE (IEEE, 1983) pp. 283–294.

[34] P. C.-K. Luk, S. Aldhaher, W. Fei, and J. F. Whidborne, State-space modeling of a

class converter for inductive links, IEEE Transactions on Power Electronics 30, 3242
(2015).

[35] P. Schavemaker and L. van der Sluis, Electrical Power System Essentials (John Wiley
and Sons, Chichester, 2008).

[36] F. Zhang, The Schur complement and its applications, Vol. 4 (Springer Science &
Business Media, 2006).

[37] M. Kapetanovic, High voltage circuit breakers (Falculty of Electrotechnical Engi-

neering, Sarajevo, Sarajevo, 2011).
 5
N UMERICAL ANALYSIS OF THE
BLOCK MODELING METHOD

5.1. I NTRODUCTION
I N this chapter, various networks are studied to demonstrate the block modeling
method. As presented in Chapter 4, two type of power systems are considered - linear
and non-linear. Then, several sub-cases of different sizes are investigated as well. To
compare the results of the block modeling method in terms of speed-up and accuracy,
the software PSCAD is used for linear networks and Matlab/SimPowerSystem is used
for non-linear systems [1, 2]. In the case of a linear network, the simulation of these
test cases with the block modeling method is implemented in C/C++ with the help of
the toolbox PETSc [3, 4]. To illustrate the non-linear network, this modeling method has
been performed in Matlab. Finally, for the same precision in the larger test cases, a speed
improvement factor 11 is achieved for the linear case, and factor 20 for the non-linear
system.
All computations were performed on a machine with an Intel(R) Core(TM)2 Duo
3GHz CPU and 4GB memory. Two cores were used to obtain the results with PSCAD
and Matlab simulation. Other simulations have been computed on only one computer
core.
This chapter is organized as follows. Section 5.2 presents and discusses the different
linear networks. Section 5.3 gives the numerical results when arc models are used in-
stead of switching devices. Finally, Section 5.4 offers the conclusion of the utilization of
the block modeling method.

5.2. L INEAR POWER SYSTEMS

5.2.1. D ESCRIPTION OF TEST CASES
Two networks, modeled with the nodal analysis method and with the block modeling
method, are considered. The computation time of a complete simulation (from the cal-

71
 72 5. N UMERICAL ANALYSIS OF THE BLOCK MODELING METHOD

culation of the first set of equations till the end of the simulation) and the accuracy of
the time-domain solution are taken for comparison.
The first test network consists of 40 block models (five generators, eleven loads,
fourteen PI-sections and nine double PI-sections) and sixteen switch blocks (sixty-three
switches), as depicted in Figure 5.1. The block models’ parameters in accordance with
Table 4.2 are:
• G: R = 0.1Ω, L = 0.2mH , C = 1µF and RC = 100Ω;
• L: R = 1Ω, L = 0.4mH and C = 1µF ;
• PI-section: R = 0.1mΩ, L = 0.2mH and C 1 = C 2 = 0.1µF ;
• Double PI-section: R 1 = R 2 = 0.1mΩ, L 1 = L 2 = 0.2mH and C 1 = C 2 = C 3 = 0.1µF ;
• Switch: R sw = 0.1Ω.

Figure 5.1: Block diagram representation of the network composed of 40 block-models and 16 switch blocks

The initial conditions of the network are x(t 0 ) = 0. All switches are in the closed
position, except for the switch with label 1 in Figure 5.1 that is being closed at t = 0.06s.
The second test network is created by copying the block diagram of Figure 5.1 five
times. Different switching times are applied to illustrate the effect of updating the set of
equations several times. For this reason, there are no links among the five networks.
The systems are simulated with the nodal analysis program PSCAD and the block
modeling approach, with the Trapezoidal rule as integration method. A time step of
1µs was chosen to capture the different transient oscillations because of the initial
conditions and switching actions. To get a better understanding of the method, the
ARK4(3)6L[2]SA-ESDIRK [5] shown in Section 3.3.3 is employed as a numerical integra-
tion method to replace the Trapezoidal rule. This approach and the various numerical
integration processes are implemented in C/C++ by using the PETSc toolbox.
5.2. L INEAR POWER SYSTEMS 73

5.2.2. R ESULTS AND DISCUSSIONS

From the PSCAD software and this new approach, results are given in Table 5.1 and illus-
trated by Figures 5.2 to 5.4. Table 5.1 shows the computation time for different integra-
tion methods. Then, Figures 5.2 to 5.4 present the time domain solution of the voltage U
(close to switch 1; see Figure 5.1) for the initialization and around the time of a switch-
ing event. The error between the computed and the analytical solution is also shown in
these figures. In these figures, only the smaller network is considered.

Modeling Integration Network

method method 1 2
PSCAD TR (1µs) 25.4 139
BMM TR (1µs) 30 48
BMM ARK4 (tol=0.001) 5.71 12
5
Table 5.1: Computation times in seconds of the Nodal Analysis program PSCAD and the Block Modeling
Method (BMM) with a time-step of 1µs (Trapezoidal rule), and the ARK4 method with a tolerance of 0.001

Table 5.1 shows the computation time, in seconds, of the overall simulation of the
various test networks. For the larger test case, PSCAD was almost 11 times slower than
the block modeling method with adaptive time-stepping. The speedup was of the order
22 when both were running on a single core.

2 40 1 6
|Uc −UcLU | [mV ]

|Uc −UcLU | [V ]

1.5 30
4
Uc [V ]

Uc [V ]

1 20 0.5
2
0.5 10

0 0 0 0
0 0.5 1 1.5 2 56 58 60 62 64
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(a) Initialization (b) Around the switching action

Figure 5.2: Computation of signal U and its error with the sofware PSCAD, and a time-step of 1µs
 74 5. N UMERICAL ANALYSIS OF THE BLOCK MODELING METHOD

2 40 1 6

|Uc −UcLU | [mV ]

|Uc −UcLU | [V ]
1.5 30
4
Uc [V ]

Uc [V ]
1 20 0.5
2
0.5 10

0 0 0 0
0 0.5 1 1.5 2 56 58 60 62 64
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(a) Initialization (b) Around the switching action

Figure 5.3: Computation of signal U and its error with the block modeling method, the Trapezoidal rule as
integration method, and a time-step of 1µs

5
2 40 1 6
|Uc −UcLU | [mV ]

|Uc −UcLU | [V ]
1.5 30
4
Uc [V ]

Uc [V ]

1 20 0.5
2
0.5 10

0 0 0 0
0 0.5 1 1.5 2 56 58 60 62 64
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(a) Initialization (b) Around the switching action

Figure 5.4: Computation of signal U and its error with the block modeling method, the ARK4(3)6L[2]SA-
ESDIRK method as integration method, and an adaptive time-stepping algorithm

From Figures 5.2 to 5.4, the time domain solutions of the signal U were almost iden-
tical regarding the waveforms at the initialization, and around the switching action with
the various methods. Also, error waveforms were similar with the different methods. Fi-
nally when the ARK4(3)6L[2]SA-ESDIRK methd is used, the error waveform was smaller
than the waveforms for other methods around the switching action, because the order
of the ARK4 method is higher than the order of the Trapezoidal rule, and it has a time-
stepping algorithm based on the error.

5.3. N ON - LINEAR POWER SYSTEMS

In this part, two power systems were considered. They were both modeled via the block
modeling method implemented in Matlab for the different arc models. The integration
5.3. N ON - LINEAR POWER SYSTEMS 75

method chosen is the numerical integration called ode23tb with a relative tolerance of
10−4 .

5.3.1. T EST CASES PRESENTATION

In this part, two test cases are studied. The first one is composed of 22 block models, 10
switch blocks, 22 switches and two Habedank models - as shown in Figure 5.5.

U Arc 1 U Arc 2

G Arc 1 PI Arc 2 PI PII PII G

PII PII
Sc
PI G PI G

PII PII
U1
i1
L L L

L L L

Figure 5.5: Block representation of the synthetic network 2, composed of 22 block models, 10 node-blocks and 5
including 2 Habedank models

This network consists of 66 differential variables. Then, the second test case under
consideration is a 3-phase power system where each phase is composed of two arc mod-
els, 36 block models, 16 switch blocks and 42 switches. Figure 5.6 represents a one-line
representation of the network under consideration.

Figure 5.6: Block representation of one phase of the synthetic network 3, composed of 36 block models, 16
node blocks and two Habedank models

The total number of differential variables was 324 for this network. The parameters
for each block model type of the previous different block diagrams are:

• G: R =½0.001Ω, L = 3.52mH , C = 1.98µF , RC = 100Ω, and

591960t cos(100πt ) if t<0.1
e(t ) = ;
59196 cos(100πt ) else
 76 5. N UMERICAL ANALYSIS OF THE BLOCK MODELING METHOD

• L: R = 1Ω, L = 35.2mH and C = 1.98µF ;

• PI-section: R = 10µΩ, L = 20µH and C 1 = C 2 = 20µF ;

• Double PII-section: R 1 = R 2 = 10µΩ, L 1 = L 2 = 20µH and C 1 = C 2 = C 3 = 20µF ;

• Arc: τc = 12µs, Uc = 5kV , τm = 4µs and P m = 2MW .

• Switch: R sw = 0.1mΩ.

The scenario of these test cases is the following:

5
• From t = 0s to t < 0.345s: All currents and voltages of the block model diagrams
reach their steady state;

• At t = 0.345s: A short-circuit is initiated by closing the switching device SC;

• At t = 0.5s: Arc models are activated;

• From t > 0.5s to t = 1s: Arc models either interrupt or do not interrupt the short-
circuit current in accordance with the parameters of the block diagrams.

In the first instance, the computation time of the previous scenario is studied for the
network of Figure 5.5. The numerical and the analytical Jacobian during the simulation
have then been investigated.
Secondly, the block diagram shown in Figure 5.6 is examined. Now, the short-circuit
is only present in the first phase. Then, the computation time of the overall simulation
is investigated when the analytical and numerical Jacobian are used.

5.3.2. P RESENTATION OF THE RESULTS AND DISCUSSION

Figure 5.7a and 5.7b present the time domain solution of the current and voltage related
to the arc model, denoted 1 and 2 in Figure 5.5. Then, the voltage and current denoted
U and I are drawn in Figure 5.7c. Finally, Table 5.2 shows the computation of the overall
simulation for the analytical and numerical calculation of the Jacobian matrix.
5.3. N ON - LINEAR POWER SYSTEMS 77

50 50
u ar c1 [kV ]

i ar c1 [k A]
0 0

−50 −50

−100 −100
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t [s]
u ar c1 i ar c1

(a) Voltage across and through the arc denoted ar c 1

120 300

60 150
u ar c2 [kV ]

i ar c2 [k A]
5
0 0

−60 −150
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t [s]
u ar c2 i ar c2

(b) Voltage across and through the arc denoted ar c 2

100 20

50 10
u 1 [kV ]

i 1 [k A]

0 0

−50 −10

−100 −20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t [ms]
u1 i1

(c) Voltage at the node 1 and the current flowing through it

According to Figures 5.7a and 5.7b, the interruption of the current to both sides of
the PI-section, which contains the short-circuit, was achieved. Although, once the arc
model number two interrupted the short-circuit current, the voltage and the current of
the node 1 on Figure 5.7c reached a new steady state because the topology of the network
 78 R EFERENCES

changed.

Numerical Jacobian Analytical Jacobian

90.3 20.3

Table 5.2: Computation time in seconds of the network shown in Figure 5.5

According to Table 5.2, the computation time was four times faster with the analytical
Jacobian than with the numerical Jacobian. This result was expected because usually,
it is computationally less expensive to compute the analytical Jacobian instead of the
numerical Jacobian.
Now, the results of the second part are shown. Table 5.3 gives an indication of the
computation time for the overall simulation.

Numerical Jacobian Analytical Jacobian

>3600 211
5
Table 5.3: Computation time in second of the network shown in Figure 5.6

Table 5.3 shows that for such a relatively small 3-phase network, the application of
several arc models made it necessary to compute the Jacobian matrix analytically to ob-
tain an acceptable computation time.

5.4. C ONCLUSION
In this chapter, the speed-up in getting the time domain solution by using the block
modeling method is shown. The simulation of linear networks was 11 times faster when
using a combination of this new approach to model power systems and advanced nu-
merical integration methods. Then, the simulation of an electrical network including
several arc models can be sped up by giving the mathematical expression of the Jaco-
bian matrix.
Finally, the mathematical expression of a network when a switching action appears
has been improved by using this modeling approach, but the computation of the time
domain solution can still be improved. The numerical integration methods applied were
implicit, and the overall computation time can be improved, as explained and shown in
the next chapters.

R EFERENCES
[1] PSCAD/EMTDC: Electromagnetic transients program including dc systems (1994).

[2] MATLAB SimPowerSystems for Use with Simulink User’s Guide, Version 4.1.1.

[3] S. Balay, S. Abhyankar, M. F. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin,

V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, L. C. McInnes, K. Rupp, B. F.
R EFERENCES 79

Smith, S. Zampini, H. Zhang, and H. Zhang, PETSc Users Manual, Tech. Rep. ANL-
95/11 - Revision 3.7 (Argonne National Laboratory, 2016).

[4] S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith, Efficient management of paral-

lelism in object oriented numerical software libraries, in Modern Software Tools in Sci-
entific Computing, edited by E. Arge, A. M. Bruaset, and H. P. Langtangen (Birkhäuser
Press, 1997) pp. 163–202.

[5] C. A. Kennedy and M. H. Carpenter, Additive Runge-Kutta schemes for convection-

diffusion-reaction equations, Applied Numerical Mathematics 44, 139 (2003).

5
 6
I NACCURATE SOLVER FOR THE
SIMULATION OF POWER SYSTEMS

6.1. I NTRODUCTION
T HE design and daily operation of electrical networks are currently undergoing rapid
changes. New projects of HVDC link, large offshore wind farms, and the intercon-
nection of power systems make the electrical network increasingly larger and complex.
In addition, all these changes rely on the use of mathematical modeling and numerical
simulation tools. The development of new numerical techniques is, therefore, manda-
tory.
In this chapter, the efficiency of computing the time variation of the current and volt-
age in an electrical network is studied [1–3]. In addition, the block modeling method
is applied to arrive at the state space representation of the electrical network, with the
properties shown in Chapter 4. The resulting system of ordinary differential equations
is typically stiff due to the parameters of the different network elements. Also, only the
linear block model shown in Table 4.1 is employed. The network topology is assumed to
be fixed throughout the simulation. Moreover, to simplify the studies, the exclusion of
switching action is made during the simulation. Additionally, zero initial conditions are
imposed. These initial conditions give rise to a set of initial transients to appear. Once
these transients have disappeared, the system oscillates in its steady state regime.
The resulting system of ordinary differential equations can be integrated, as shown
in Chapter 3 [4, 5]. The stiffness present in the problem motivates the use of implicit
methods. The requirement of solving systems of linear equations renders these meth-
ods computationally expensive. Two directions of research to cope with this issue have
been developed in the literature. The first resulted in diagonally implicit Runge-Kutta
that retains good stability properties while being more efficient than their fully implicit
counterpart. The second gives rise to a body of research showing how to efficiently de-
ploy iterative linear and non-linear solvers in the time-step loop. The development of
fast linear and non-linear solvers embedded in backward difference methods was con-

81
 82 6. I NACCURATE SOLVER FOR THE SIMULATION OF POWER SYSTEMS

sidered, for example, [6–9]. For fully implicit Runge-Kutta methods, preconditioners
exploiting the Kronecker product were considered in [10, 11]. Domain decomposition
iterative solvers for transients in power systems have recently been studied in [12].
The ARK4(3)6L[2]SA-ESDIRK [13] method shown in Section 3.3.3 is used for this
study. In addition, this integration method can handle either a fixed or adaptively con-
structed time-step. This method requires solving a sequence of five linear systems at
each time-step. The solution to each of these systems provides an intermediate stage
in updating the current numerical approximation to the next time-step. The size of the
linear system is equal to the number of differential equations. To alleviate the compu-
tational bottleneck of the linear system’s solving process, iterative methods can be used.
Both stationary and non-stationary iterative methods for linear systems exist [14–16]. An
incomplete LU (iLU) factorization is used with partial fill-in after re-ordering to split the
coefficient matrix. This splitting is then used to construct a stationary iterative method.
Also, a limited number of iterations are employed, and the result obtained in this way is
denoted as an inexact linear solver. The accuracy of this solver can be varied by altering
the amount of fill-in or by combining it with an outer GMRES iteration. The inexact lin-
ear solver is embedded as an inner iteration in a Runge-Kutta time integration method
that acts as an outer iteration. The speed of convergence of the inner iteration then de-
pends on factors such as the size of the time-step and the stiffness of the problem. The
interest of this chapter is to theoretically and computationally investigate the efficiency
of the inner-outer scheme.
6 This chapter is organized as follows. Section 6.2 states the problem’s definition.
Then, a definition of approximate solvers is given in Section 6.3. It is followed by the
conclusion in Section 6.4.

6.2. D EFINITION OF PROBLEM

In this chapter, the numerical integration method used is the ARK4(3)6L[2]SA-ESDIRK
method shown in Section 3.3.3. This numerical approach proposes to solve five depen-
dent linear systems of equations (Eq. 3.27), which can be summarized as:

A (∆t )k i = q i for 2 ≤ i ≤ 6, (6.1)

³ ´
where A (∆t ) = I − γ∆t A and q i = f u n + ij−1 =1 a i j ∆t k j , t n + c i ∆t .
P

For calculating the k i vectors, the LU factorization or the GMRES method, with the
incomplete LU factorization (iLU) as preconditioner, are usually used to solve Eq. 6.1. In
this chapter, more efficent alternatives are considered .

6.3. A PPROXIMATE LINEAR SOLVER DEFINITION

In this section, instead of using an LU factorization or GMRES method, inaccurate
solvers are studied. In particular, this section studies the interaction between the inac-
curate solutions, the stages of the numerical integration method, and the time-stepping
algorithm.
First, the qualities of an approximate linear solver are:

• Provides a good approximation of the solution of the system of equations;

6.3. A PPROXIMATE LINEAR SOLVER DEFINITION 83

• Faster than Gaussian elimination.

Second, a Basic Iterative Method (BIM) - where the stopping criterion has been modified
- could be an inexact linear solver. Although the solution is approximated, it is only one
component of the time domain solution. Here, one iteration of iLU(`) factorization as
BIM is used as an inexact linear solver. Finally to construct an iterative scheme, the
decomposition of the matrix A (∆t ) can be made as:

A (∆t ) = M (∆t ) − N (∆t ), (6.2)

Then, the iLU factorization of the matrix A (∆t ) of Eq. 3.27 can be defined as:

A (∆t ) = L(∆t )U (∆t ) − R(∆t ), (6.3)

where the matrix L(∆t ) is the lower triangular matrix and the matrix U (∆t ) is the upper
triangular matrix of the iLU factorization. Moreover, the matrix R(∆t ) is the residual
matrix of the iLU factorization. This matrix is null in the case of the LU factorization. In
addition, at each time-step, only one factorization is effectuated due to the definition of
the matrix A (∆t ) (Eq. 6.1). Moreover, according to Eq. 6.2, matrices M (∆t ) and N (∆t )
are defined as:
M (∆t ) = L(∆t )U (∆t ), (6.4)
N (∆t ) = R(∆t ). (6.5)
6
The relation between the residual vector of the stage i at the iteration m (r im ) is de-
fined as:
r im = q i − A (∆t )k̃ im , (6.6)
and the error vector is expressed as:

²m m
i = k i − k̃ i , (6.7)

where k̃ im is the approximate solution of the stage i after m iterations. The norm of r im
informs the quality of the approximate linear solver. The residual vector is usually easier
to compute than the error vector because the vector k i is unknown. In fact, if the norm
of the residual error becomes small, the norm of the error vector is also small. Finally,
the normalized residual at each stage (z i ) is defined as:

kr im k2
zi = , (6.8)
kq i k2

for 2 ≤ i ≤ s. In order to judge the inexact solver, the spectral radius can be employed as:

ρ(I − M (∆t )−1 A (∆t )) (6.9)

From the definition of the spectral radius, the solution of the linear system of equations
converges only when ρ is less than one. Moreover, an asymptotic convergence or di-
vergence can be observed after some iterations. Finally, a super linear convergence can
be seen after one iteration, when the residual error values are plotted in function of the
number of iterations.
 84 R EFERENCES

Here, one iteration is employed. Then, ρ needs to be smaller than one. Besides, as
shown in Eq. 6.1, the time-step plays an important role in the converge of the BIM. Also,
two possible strategies to control the time-step can be applied: a fixed time-step method
and an adaptive time-step. In addition, as shown in Eq. 6.3, the level of fill-in ` affects
the spectral radius.

6.4. C ONCLUSION
For a fixed level of fill-in ` of the iLU factorization and a fixed time-step, several cases
are possible. In the case where ρ is strictly superior to one, the time domain solution will
diverge and the simulation will stop due to the accumulation of the error after each time-
step, because the inexact solver diverges. Then, for being able to simulate the electrical
network under consideration, two actions can be taken. The first one is to decrease the
time-step. The system of equations will then become mainly diagonally dominant, and
the spectral radius will be reduced. The second action is to increase the level of fill-
in ` of the iLU factorization. Then, the matrix decomposition will be closer to the LU
factorization. Finally, when the level of fill-in becomes too large, the iLU factorization
becomes the LU factorization and it will not lead to a speed-up.
For a fixed level of fill-in ` of the iLU factorization and an adaptive time-stepping al-
gorithm, the time-step selection procedure will reduce or increase as a function of error
between the order p and p-1 of the numerical integration method. For example, if the
6 inaccurate solutions of the k i vectors have large residual errors, then the error between
between the two orders of the method will be considerable. Then, the time-step will be
rejected and a smaller time-step will be employed. In this way, the non-convergence is
avoided at the expensive of a larger number of time-steps.
Finally during a simulation, the variation of the level of fill-in ` is not considered.
This non-variation comes from the fact that a symbolic factorization and numerical fac-
torization is employed, because it will speed-up the computation time. Then, if there
is a variation over time of the level of fill-in ` of the iLU factorization, a new symbolic
factorization would have to be computed, and will slow down the required computation
time.

R EFERENCES
[1] L. van der Sluis, Transients in Power Systems (John Wiley & Sons, Chichester, 2001).

[2] J. Grainger and W. Stevenson, Power System Analysis, Electrical Engineering Series
(McGraw-Hill, New York, 1994).

[3] N. Watson and J. Arrillaga, Power systems electromagnetic transients simulation,

Vol. 39 (Institution of Electrical Engineers, London, 2003).

[4] G. Wanner and E. Hairer, Solving ordinary differential equations II : Stiff and
Differential-Algebraic Problems, Vol. 1 (Springer Berlin Heidelberg, Berlin Heidel-
berg, 2010).

[5] W. Hundsdorfer and J. Verwer, Numerical Solution of Time-Dependent Advection-

Diffusion-Reaction Equations (Springer, New York, 2003).
R EFERENCES 85

[6] T. F. Chan and K. R. Jackson, The use of iterative linear-equation solvers in codes for
large systems of stiff IVPs for ODEs, SIAM journal on scientific and statistical com-
puting 7, 378 (1986).

[7] P. N. Brown, A. C. Hindmarsh, and L. R. Petzold, Using Krylov methods in the so-
lution of large-scale differential-algebraic systems, SIAM Journal on Scientific Com-
puting 15, 1467 (1994).

[8] F. Perini, E. Galligani, and R. D. Reitz, A study of direct and Krylov iterative sparse
solver techniques to approach linear scaling of the integration of chemical kinetics
with detailed combustion mechanisms, Combustion and Flame 161, 1180 (2014).

[9] J. E. Pessanha, A. A. Paz, R. Prada, and C. P. Poma, Making use of BDF-GMRES meth-
ods for solving short and long-term dynamics in power systems, International Jour-
nal of Electrical Power & Energy Systems 45, 293 (2013).

[10] L. O. Jay, Inexact simplified Newton iterations for implicit Runge-Kutta methods,
SIAM Journal on Numerical Analysis 38, 1369 (2000).

[11] K. Dekker, Partitioned Krylov subspace iteration in implicit Runge-Kutta methods,

Linear Algebra and its Applications 431, 488 (2009).

[12] P. Aristidou, S. Lebeau, and T. Van Cutsem, Power system dynamic simulations using
a parallel two-level Schur-complement decomposition, IEEE Transactions on Power 6
Systems 31, 3984 (2016).

[13] C. A. Kennedy and M. H. Carpenter, Additive Runge-Kutta schemes for convection-

diffusion-reaction equations, Applied Numerical Mathematics 44, 139 (2003).

[14] A. Greenbaum, Iterative Methods for Solving Linear Systems, Frontiers in Applied
Mathematics, Vol. 17 (SIAM, Philadelphia, 1997).

[15] Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd ed. (Society for Industrial
and Applied Mathematics, Philadelphia, 2003).

[16] H. A. van der Vorst, Iterative Krylov Methods for Large Linear Systems, Cambridge
Monographs on Applied and Computational Mathematics, Vol. 13 (Cambridge Uni-
versity Press, New York, 2003).
 7
A PPROXIMATE SOLVER FOR THE
SIMULATIONS OF POWER SYSTEM

7.1. I NTRODUCTION
I N this chapter, various power systems are studied to demostrate the use of approxi-
mate linear solvers during the numerical integration process. As presented in Chapter
6, two sizes of power systems are considered - small and large. In order to describe the
relationship between approximate solvers, time-step, size of the power system, and stiff-
ness, this chapter focuses first on the use of a fixed time-step, and then second, on the
utilization of an adaptive time-stepping algorithm. Finally, for the same precision for the
larger test cases, a speed-up of five can be noticed in the case of the stiffer large power
system.
The PETSc software library is employed for the implementation. At each time-step,
a linear solver [1, 2], which is computed by the Krylov Subspace (KSP) and the precondi-
tioner (PC) methods of this library, is used. Simulations are performed on a Linux system
with an Intel (R) Core(TM) 2 Duo 2.7GHz CPU processor, with 4GB RAM.
This chapter is organized as follows: Section 7.2 presents and discusses the various
power systems and experiments; Section 7.3 offers the results and explains them; and,
Section 7.4 states the conclusion of the utilization of approximate solvers.

7.2. T EST CASE PRESENTATION

In this section, details of the test cases are given, in particular, on the time-stepping
procedures and the linear solution methods used in the numerical experiments.
The construction of four synthetic power systems is made so that there are two small
and two large ones. The small and large systems contain d = 1, 034 and d = 20, 999 dif-
ferential variables, respectively. Each system is created by interconnecting - via trans-
mission lines - a set of generators and loads. The block modeling method is applied to
generate the matrices A and B in the linear state space representation Eq. 3.1 of the
power system (Chapter 4). The smallest time constant, and therefore the stiffness of the

87
 88 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

system of ODEs, is determined by the largest eigenvalue of the matrix A. The eigenvalues
of the matrix A are determined by the topology of the network and the lumped elements
properties of the different components. These properties are the resistance, capacitance
and inductance values. The construction is made such that the spectral radius of the
matrix A is more sensitive to the changes in the values of the lumped elements than to
the changes in the topology of the network. This fact allows for the increase in stiffness of
a power system, by changing the range of the lumped elements values without changing
the topology of the system - and therefore - the non-zero structure of the matrices A and
B . It also allows for the increase in size of the power system, while simultaneously pre-
serving its smallest time constant. The ratio between the largest and smallest eigenvalue
is 3×108 for the non-stiff power systems, and 7×1012 and 1×1013 for the small and large
stiff power system, respectively. In all four cases, the input function g (t ) has a frequency
of 50 Hz.
The system of ODEs (Eq. 3.1) is integrated with zero initial conditions (x 0 = 0) from
t = 0 to t = 50 ms using the fourth order SDIRK (ARK4) method with either a fixed (Alg.
1) or adaptive time-step (Alg. 2). In the former case, various time steps are used. In
the latter case, two absolute tolerance values will be imposed. To illustrate the adaptive
selection of the time-step, Figure 7.1 shows the evolution of the time-step over time in
integrating the four different power systems. Also, two absolute tolerances Tol are used,
and they are defined in Eq. 3.26, namely Tol = 10−4 and Tol = 10−6 .

7
Algorithm 1: Fixed time-step
Data: A, B , g (t ), x(0), t f i nal and ∆t
begin
t0 = 0
u 0 = x(0)
n=0
AMD re-ordering method on A (∆t )
Factorization of the re-ordered matrix A (∆t )
while t < t f i nal do
Solve the linear systems of equations
Update u n+1
t n+1 = t n + ∆t
n = n +1
7.2. T EST CASE PRESENTATION 89

Algorithm 2: Adaptive time-step

Data: A, B , g (t ), x(0), t f i nal and Tol
begin
tn = 0
u 0 = x(0)
n=0
∆t = 10−3
AMD re-ordering method on A (∆t )
Symbolic factorization of the re-ordered matrix A (∆t )
while t < t f i nal do
Numerical factorization
Solve the linear systems of equations
Calculate e n+1
if e n+1 ≤ Tol then
Update u n+1
t n+1 = t n + ∆t
n = n +1
Calculate new ∆t

1.5 1.5
∆t [ms]

∆t [ms]

1 1
0.5 0.5
0
0 20 40
0
0 20 40
7
t [ms] t [ms]
Tol=10−4 Tol=10−6 Tol=10−4 Tol=10−6

(a) Small non-stiff power system (b) Small stiff power system
1.5 30
∆t [ms]

∆t [µs]

1 20
0.5 10
0 0
0 20 40 0 20 40
t [ms] t [ms]
Tol=10−4 Tol=10−6 Tol=10−4 Tol=10−6

(c) Large non-stiff power system (d) Large stiff power system

Figure 7.1: Time evolution of the time-step in cases where adaptive time-stepping with a tolerance Tol = 10−4
and Tol = 10−6 is used with a direct solver for the four power systems.

Figure 7.1 shows that - as expected - the more stringent tolerance results in a smaller
time-step. The figure also indicates that for the small and large non-stiff system, the
time-step initially grows to oscillate subsequently between a minimum and maximum
 90 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

value. As the stiffness of the small and large systems is approximately the same, the mini-
mum and maximum values of the time-step are roughly equal. This oscillation indicates
that the system is in a steady state. The consecutive peaks in the time-steps are sepa-
rately by approximately 10 ms, the half period of the 50 Hz excitation. For the small stiff
power system, the growth of time-step and reaching the steady state requires a longer
period. Once the steady state is obtained, the time-step oscillates between a minimum
and maximum value, as in the small non-stiff system. For the large stiff system, the time-
step remains uniformly inadequate in time. To obtain a larger time-step, the simulation
needs a final time superior of 9.6 seconds with a tolerance of Tol = 10−4 . The ARK4 time-
integration method involves six stages and requires the solution of five linear systems of
size d × d with the same coefficient matrix A (∆t ) = I − ∆t γ A - as shown in Section 6.2.
The results are obtained by both iterative and direct solution methods. In the case of a
fixed time-step, the matrix A (∆t ) is AMD reordered and LU-factored before the time-
stepping loop (Alg. 1). In the case of a variable time-step, only the reordering, and the
symbolic factorization can be performed before the time-stepping loop (Alg. 2). As iter-
ative solution method, iLU(`) as a BIM and as a preconditioner for GMRES for various
values of ` is used. By using iLU(`) as a BIM, a single iteration is always employed. The
accuracy of this approximate solver is increased by increasing `. By using iLU(`) as a
preconditioner for GMRES, one or five iterations are performed. As different stages have
no immediate relation, a zero initial guess is made.
The efficiency of iLU(`) is determined by the balance of its computational cost and
its speed of convergence. The computational cost is determined by the number of non-
zeros in the L(∆t ) and U (∆t ) factors, as shown in Figure 7.2.
nnz(L(∆t )+U (∆t ))

nnz(L(∆t )+U (∆t ))

7 1.5 5
nnz(A (∆t ))

nnz(A (∆t ))

1.4 4
1.3 3
1.2
1.1 2
1 1
iL (1 )
iL (2 )
iL (3 )
U 0)
iL (5 )
iL (6 )
U 0)
0)

iL (1 )
iL (2 )
iL (3 )
iL (4 )
iL (5 )
iL (6 )
U 0)
0)
LU

LU
U 0
U 0
U 0

U 0
U 0

U 0
U 0
U 0
U 0
U 0
U 0
iL U(

iL (4

(7

iL U(

(7
iL

iL

Factorization Factorization
(a) Small power system (b) Large power system

Figure 7.2: Normalized number of non-zeros in the incomplete factorization of A (∆t ) for various levels of
fill-in, and the LU factorization of A (∆t ) for the small (a) and large (b) power systems.

Figure 7.2 shows a number scaled by the number of non-zeros in the matrix A for
various levels of fill-in ` after AMD reordering for both the small and large power system.
The number of non-zeros of I − ∆t γA is equal to the number of non-zeros of A (∆t ) for
all possible values of ∆t , and does not vary with the stiffness of the power system. Figure
7.2a shows that, for the small system, the fill-in ` = 20 results in many non-zeros of L(∆t )
and U (∆t ) that is almost as large as the number of non-zeros of A (∆t ). For the large
system, the fill-in ` = 40 results in a number of non-zeros of L(∆t ) and U (∆t ) - is roughly
half the number of non-zeros of the LU decomposition of A (∆t ). Now, the speed of
7.3. R ESULTS AND DISCUSSIONS 91

converge of iLU(`) is determined by the spectral radius of the error propagation matrix
(Eq. 6.9), and it is shown in Figure 7.3 for the small power systems.
ρ(I − M (∆t )−1 A (∆t ))

ρ(I − M (∆t )−1 A (∆t ))

102 102

10−3 10−3

10−8 −10 10−8 −10

10 10−7 10−4 10−1 10 10−7 10−4 10−1
∆t [s] ∆t [s]

iLU(0) iLU(10) iLU(30) iLU(0) iLU(10) iLU(30)

(a) Small non-stiff power system (b) Small stiff power system

Figure 7.3: Spectral radius of the incomplete factorization used as a BIM for various levels of fill-in as a function
of the time-step for the small non-stiff (a) and stiff (b) power systems.

Figure 7.3 shows the spectral radius as a function of the time-step ∆t for the small
non-stiff and stiff power system for various levels of fill-in `. This figure shows a decrease
of the spectral radius for decreasing ∆t . By reducing ∆t , the matrix A (∆t ) increasingly
approximates the identity matrix better, and the iLU approximation becomes closer to
the complete LU factorization that - in the limit of ∆t → 0 - results in a spectral radius 7
equal to zero. A larger level of fill-in leads to a smaller spectral radius. A smaller one is
easier to obtain for the non-stiff power system. In the case of the stiff power system, a
low level of fill-in and a large time-step results in a non-convergent solver.

7.3. R ESULTS AND DISCUSSIONS

In this section, numerical results obtained are shown in two subsections. While the first
one discusses the fixed time-step, the second one investigates the adaptive time-step.

7.3.1. F IXED TIME - STEP

In this section, the results obtained using a fixed time-step for the small non-stiff and the
small stiff power system are discussed first. Then, the results obtained using a fixed time-
step again for the large non-stiff, and large stiff power systems are investigated. This
order displays the relationship between the stiffness and parameters in the approximate
linear solver.
Tables 7.1 and 7.2 show the computation time of Alg. 1 with various linear
solvers. The "NC" character represents the non-convergence of a simulation. The non-
convergence can appear for two reasons: first, the simulation may have ended before the
final time was reached, due to the accumulation of error; second, the output may have
given an incorrect result, such as an impossible voltage or current, and this simulation
 7
92

Linear solver Small non-stiff Small stiff

∆t 1ms 0.1ms 10µs 1µs 1ms 0.1ms 10µs 1µs
LU 0.039 0.194 1.752 17.287 0.039 0.194 1.743 17.280
ILU(0) NC NC 1.425 14.071 NC NC NC NC
ILU(10) 0.036 0.168 1.495 14.716 NC NC NC NC
ILU(30) 0.036 0.170 1.521 14.979 NC 0.171 1.518 15.090
GMRES ILU(0) 1it NC NC 2.510 24.973 NC NC NC NC
GMRES ILU(10) 1it NC 0.284 2.655 26.257 NC NC NC NC
GMRES ILU(30) 1it 0.048 0.290 2.699 26.828 NC 0.289 2.697 26.783
GMRES ILU(0) 5it NC 0.665 4.373 34.133 NC NC NC NC
GMRES ILU(10) 5it 0.090 0.385 2.653 26.439 NC NC NC 68.471
GMRES ILU(30) 5it 0.058 0.290 2.719 26.834 0.069 0.402 3.705 26.999

various iterative linear solvers, in case of fixed time-steps with various values of ∆t
could have stopped after a few more time-steps due to the accumulation of error.
7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

Table 7.1: CPU time in seconds for the small non-stiff and the small stiff power systems, for the direct and
7.3. R ESULTS AND DISCUSSIONS 93

1436.65
765.45

310.03
330.91

590.15
646.49

784.61
653.57
1µs

NC
NC

NC
NC

NC

155.89
117.75
95.24

51.47

83.72
10µs

NC
NC
NC

NC
NC
NC

NC
NC
Large stiff
0.1ms
27.01

36.19
NC
NC
NC
NC
NC
NC
NC
NC
NC
NC
NC
20.70

22.88
1ms

NC
NC
NC
NC
NC
NC
NC
NC
NC
NC
NC
770.30
276.81
293.06
311.25
332.45
503.15
539.88
539.88
643.08
700.00
572.19
592.11
645.35
1µs

107.48
94.33
45.44
47.32
49.12
50.84
68.23
72.39
72.39
82.18

72.83
88.47
82.47
10µs
lLarge non-stiff
0.1ms
26.97

22.18
22.22
23.02

25.06
25.06
26.24
33.00
27.10
25.99
26.15
NC

NC
20.17

20.01
19.74
20.16

20.29

21.37
21.63
20.60
1ms

NC

NC
NC
NC

NC

7
GMRES ILU(0)L 1it
GMRES ILU(10) 1it
GMRES ILU(30) 1it
GMRES ILU(40) 1it

GMRES ILU(10) 5it

GMRES ILU(30) 5it
GMRES ILU(40) 5it
GMRES ILU(0) 5it
Linear solver

ILU(10)
ILU(30)
ILU(40)
ILU(0)
LU
∆t

Table 7.2: Total CPU time (in seconds) for the large non-stiff and the large stiff power systems, for both direct
and various iterative linear solvers, in case of fixed time-steps with various values of ∆t (τ = 10ms)

Tables 7.1 and 7.2 show that when the ∆t decreases, the number of time-steps in-
creases, and then the computation time increases. The computation effort for comput-
ing u n is the same for any ∆t when LU or one iteration of iLU(`) factorization is applied.
Although the solution converged for the non-stiff power systems, the solution diverged
for the stiff power systems for the same ` and ∆t . This remark is expected from Figure
7.3 because the spectral radius of these stiff power systems for those simulations is al-
most one for such ∆t . When the size of the power systems increases, the level of fill-in
required increases, or a smaller ∆t has to be employed. These increases are understand-
 94 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

able from the fact that the iLU`) factorization depends on the time-step, and the size is
also important.
In terms of computation time, the use of an inexact linear solver for the small system
permits a gain of 2s with ` = 30 and ∆t = 1µs for a simulation of 50ms. However, for
the larger test cases, a speed-up of a factor two can be observed with ` = 30 and ∆t =
1µs. This speed-up almost corresponds to the difference of the ratio of the number of
non-zero elements of the factorization, divided by the number of non-zero elements of
A (∆t) (Figure 7.2). Finally, the use of GMRES with iLU(`) as preconditioner requires
more computation time than LU or one iteration of iLU(`) factorization. Meanwhile, for
∆t = 1ms and the large stiff power system, the time domain solution converges after five
iterations. However, the utilization of such a time-step (∆t = 1ms) is not recommended
in case of transient simulations because it cannot capture the fast transient oscillations.
Finally, the use of an inexact solver can improve the computation time.
Figures 7.4 and 7.5 represent the normalized residual of the sixth stage, which is char-
acteristic of the other stages, of ARK4(3)6L[2]SA-ESDIRK method for various level of fill-
in `, stiffnesses, and power system sizes.

kr 6 k2 /kq 6 k2
kr 6 k2 /kq 6 k2

0.04 100
0.02 50
0 0
0 20 40 0 0.2 0.4 0.6 0.8 1
t [ms] t [ms]
(a) ` = 0 and the small non-stiff power system (b) ` = 0 and the small stiff power system
7 3
kr 6 k2 /kq 6 k2

kr 6 k2 /kq 6 k2

0.04
2
0.02 1
0 0
0 20 40 0 0.2 0.4 0.6 0.8 1
t [ms] t [ms]
(c) ` = 10 and the small non-stiff power system) (d) ` = 10 and the small stiff power system
kr 6 k2 /kq 6 k2

kr 6 k2 /kq 6 k2

0.04 0.04
0.02 0.02
0 0
0 20 40 0 20 40
t [ms] t [ms]
(e) ` = 30 and the small non-stiff power system (f) ` = 30 and the small stiff power system

Figure 7.4: Scaled residual norm after the sixth stage of the ARK4(3)6L[2]SA-ESDIRK method, using ILU(`) as a
BIM with ` = 0 (a and b), ` = 10 (c and d) and ` = 30 (e and f) in case of a fixed time-step of ∆t = 10µs. The left
and right columns correspond to the small non-stiff and small stiff power system, respectively
7.3. R ESULTS AND DISCUSSIONS 95

kr 6 k2 /kq 6 k2

kr 6 k2 /kq 6 k2
0.04 0.04
0.02 0.02
0 0
0 20 40 0 20 40
t [ms] t [ms]
(a) ` = 0 and the large non-stiff system (b) ` = 40 and the large stiff system

Figure 7.5: Scaled residual norm after the sixth stage of the ARK4(3)6L[2]SA-ESDIRK method, using ILU(`) as a
BIM with ` = 0 for the large non-stiff power system (a) and ` = 40 for the large stiff power system (b) in case of
a fixed time-step of ∆t = 10µs

Figure 7.4 shows that the non-stiff system was able to go to the end of the simulation
with a maximal normalized residual of 0.045 for iLU(0). Also, as expected, it decreases
when the level of fill-in ` increases. For iLU(0), the first peak appears due to the transient
oscillations when the power system passed from the initial condition to the steady-state.
Then every 10ms, a pick occurred when the time domain solution of the different signal
crossed zero. The zero-crossing implies, in this case, the higher derivative of the system.
When the stiffness increased, the simulation stopped after few time-steps if the level of
fill-in ` was too small. In fact, the normalized error was significant in such cases. The
simulation stopped because the error realized at one time-step impacts the succeeding 7
time-steps. For this reason, the simulation stopped at an earlier time with a level of fill-
in ` = 0 than with ` = 10. The normalized residual was 100 on average for ` = 0, and
2 on average for ` = 10. The non-convergence to a small residual can be extrapolated
from Figure 7.3 because ρ > 1 for these levels of fill-in `. Also, these simulations stopped
because the accumulation of the error at each time-step finally approached a higher
number than the computer could handle. Finally, with ` = 30, the same remark as for
the non-stiff power system can be concluded. That time, the duration of the transient
part was almost 10ms. Although the peaks occur every 10ms for the non-stiff power
systems, for the stiff power systems, a depression appeared every 10ms.
For the large power system, the same observations can be made. For this reason,
the non-stiff power system with ` = 0 and the stiff power system with ` = 40 (Figure
7.5) is shown. The normalized residual has the same shape as that of the small non-stiff
system, and same observations can be made between both non-stiff systems. For the
stiff system, the maximal normalized residual is 0.5. With such a value, an error could
be expected in the time domain solution. Meanwhile, it was supposed to have a lower
normalized residual based on the results for the non-stiff power system.
Figures 7.6 and 7.7 represent the time domain solution of a random variable Uc for
various level of fill-in `, stiffnesses, and power system sizes. Also, the error between
this solution and the solution obtained, when the LU factorization is used as a solver, is
shown as well.
 96 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

|Uc −UcLU | [mV ]

|Uc −UcLU | [mV ]

·10131 ·10134
1 2 4 4
Uc [V ]

Uc [V ]
0 1 2 2
−1 0 0 0
0 5 10 15 0 0.2 0.4 0.6 0.8 1
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(a) ` = 0 and the small non-stiff system |Uc −UcLU | [mV ] (b) ` = 0 and the small stiff system

|Uc −UcLU | [mV ]

·10135 ·10138
1 2 0 8
−2 6
Uc [V ]

Uc [V ]
0 1 −4 4
−6 2
−1 0 −8 0
0 5 10 15 0 0.2 0.4 0.6 0.8 1
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

7 (c) ` = 10 and the small non-stiff system (d) ` = 10 and the small stiff system
|Uc −UcLU | [mV ]

|Uc −UcLU | [mV ]

1 2 1.6 2
0.8
Uc [V ]

Uc [V ]

0 1 0 1
−0.8
−1 0 −1.6 0
0 5 10 15 0 5 10 15
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(e) ` = 30 and the small non-stiff system (f) ` = 30 and the small stiff system

Figure 7.6: Component solution of the voltage signal through one of the capacitors Uc (t ) using an approximate
linear solver and the absolute value of the difference between this solution and the solution UcLU (t ), computed
using a direct solver in case of a fixed time-step of ∆t = 10µs. As an approximate linear solver, i LU (`) is used
as a BIM with ` = 0 (a and b), ` = 10 (middle row) and ` = 30 (bottom row). The left and right columns of
subfigures correspond to the small non-stiff and small stiff power systems, respectively
7.3. R ESULTS AND DISCUSSIONS 97

|Uc −UcLU | [mV ]

|Uc −UcLU | [mV ]

160 2 1.6 2
Uc [mV ]

80 0.8

Uc [V ]
0 1 0 1
−80 −0.8
−160 0 −1.6 0
0 5 10 15 0 5 10 15
t [ms] t [ms]
Uc |Uc −UcLU | Uc |Uc −UcLU |

(a) ` = 0 and the large non-stiff power system (b) ` = 40 and the large stiff power system

Figure 7.7: Component solution of the voltage signal through one of the capacitors Uc (t ) using an approximate
linear solver and the absolute value of the difference between this solution and the solution UcLU (t ), computed
using a direct solver in case of a fixed time-step of ∆t = 10µs. As an approximate linear solver, I LU (`) is used
as a BIM with ` = 0 for the large non-stiff problem (a) and ` = 40 for the large stiff problem (b)

Figure 7.6 reveals that the errors between the approximate linear solvers and LU on
the signal Uc were equal to zero. These results are expected according to Figure 7.4 for
` = 10 and ` = 30. For ` = 0, an error is observed. Meanwhile, this error is small (less
than 1mV for a signal of 1V). This error is understandable because the choice of this ∆t
multiplies the residual error, which helps to get an excellent approximation of the time-
domain solution. For the stiff case, the signal Uc reached an impossible voltage and then
stopped suddenly. In fact, the simulation ended because some signals were higher than
the maximal number accepted by the computer. Finally, Uc for a fill-in of ` = 30 for the
stiff power system had almost an error of 0mV. The same remarks can be applied to the
non-stiff with ` = 0. 7
For the large power system, the same observations can be done. For this reason, the
non-stiff power system with ` = 0 and the stiff power system with ` = 40 are studied (Fig-
ure 7.7). A peak of error occurred during the transient part for the non-stiff system. How-
ever, the error (2mV) is relatively small in comparison with the amplitude of Uc (160mV).
Meanwhile, the computation time is 100s faster with the iLU(`) factorization than with
the LU factorization. In the case of the large stiff power system, the error is less impor-
tant than with the non-stiff case. Finally, one iteration of iLU(40) was 1.63 times faster
than using the LU factorization for solving the systems of equations.
The fixed time-step simulations offer an understanding of the properties of inexact
solvers. According to Tables 7.1 and 7.2, there is a strong connection between the level
of fill-in ` and the time-step for performing a simulation. Also from Figures 7.4 and 7.5,
the level of fill-in depends on the size of the system of equations and the stiffness, as
well, to get a small normalized residual. Moreover from Figures 7.6 and 7.7, when the
time domain solution of the power system converges with the utilization of approximate
solvers, the error is almost negligible. Finally, the choice of approximate solvers depends
on the size of the system of equations, and its stiffness according to these results.

7.3.2. A DAPTIVE TIME - STEPPING STRATEGY

In this section, results obtained using an adaptive time-step algorithm (Alg. 1) are shown
and discussed. Firstly, the presentation of the results of the non-stiff power system is
 98 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

made. Secondly, stiff power systems are studied. This ordering differs from the one in
the previous section. It allows for the highlighting of the effects of problem size on the
performance of the approximate linear solvers. The results mainly show the number
of time-steps, and the overall CPU time varies as a function of the level of accuracy of
the various approximate solvers, the tolerance imposed on the time-stepping procedure,
and the stiffness of the problem. .
Figure 7.8 shows the evolution of the number of time-steps as a function of the level
of fill-in ` of iLU(`) as BIM, or as preconditioner for GMRES for 0 ≤ t ≤ 50 µs with an
imposed absolute tolerance Tol = 10−4 as a function of the level of fill-in ` using various
approximate linear solvers. The number of time-steps using the direct solver is clearly
independent of `, and is represented as a vertical line as a reference.
Nb time-steps

300

200

100
0 5 10 15 20 25 30
Level of fill-in `
ILU(`) GMRES+ILU(`) 1 it GMRES+ILU(`) 5 it LU

(a) Small non-stiff power system

7
Nb time-steps

300

200

100
0 5 10 15 20 25 30
Level of fill-in `
ILU(`) GMRES+ILU(`) 1 it GMRES+ILU(`) 5 it LU

(b) Large non-stiff power system

Figure 7.8: Number of time-steps for three approximate linear solvers and a direct solver versus the level of
fill-in time integrating the small (a) and large (b) non-stiff power system with an adaptive time-step with a
tolerance of 10−4

Figure 7.8 shows that the number of time steps using the direct solver is - therefore
- the same for both the small non-stiff and large non-stiff power systems because they
have the same smallest time constant. Then, the number of iterations decreases with
the increase of the level of fill-in ` of the iLU factorization used as BIM or as precondi-
7.3. R ESULTS AND DISCUSSIONS 99

tioner for GMRES. In addition, several iterations of GMRES with iLU(`) as preconditioner
helps to reduce the number of time-steps with regard to one iteration of iLU(`) or of GM-
RES with the same preconditioner. Also, five iterations of GMRES reach the number of
time-steps required by the LU factorization faster than the other solvers. Then, when the
number of time-steps are the same, increasing the level of fill-in ` of the iLU factoriza-
tion will only increase the computation time. The same remark is valid when the other
solvers increase their level of fill-in `, whereas they have already reached the same num-
ber of time-steps of the direct solver. For example, a significant amount of time-steps is
required for iLU(0) in comparison with LU. In fact, this remark is expected from the fixed
time-step results (Section 7.3.2) and the ∆t evolution (Figure 7.1). The iLU(0) does not
converge with relatively large ∆t (∆t < 10µs). However, the number of time-steps might
positively impact the overall computation time. For this reason, Figure 7.9 shows the
computation time of the overall simulation of the non-stiff power systems and various
solvers, according to Alg. 2. In addition, the value for the direct solver is now different
for small and large non-stiff systems, and is included in the figures as a reference.

1.5
CPU time [s]

0.5

0
0 5 10 15 20 25 30
7
Level of fill-in `
ILU(`) GMRES+ILU(`) 1 it GMRES+ILU(`) 5 it LU

(a) Small non-stiff system

120
CPU time [s]

100

80

60

40
0 5 10 15 20 25 30
Level of fill-in `
ILU(`) GMRES+ILU(`) 1 it GMRES+ILU(`) 5 it LU

(b) Large non-stiff system

Figure 7.9: Overall CPU time in seconds for three approximate linear solvers versus the level of fill-in and a
direct solver in time integrating the small (a) and large (b) non-stiff power system with a tolerance of 10−4
 100 7. A PPROXIMATE SOLVER FOR THE SIMULATIONS OF POWER SYSTEM

From Figure 7.9, a distinction between the small and large non-stiff power systems
can be observed. In the case of the former power systems, the computation decreased
when the level of fill-in ` of the factorization increased to reach the same computation
time as when the LU factorization was used. For this reason, direct solvers are recom-
mended for the computation of a small power system. In the case of the large non-stiff
power systems, the computation time is always timed faster for the different levels of
fill-in ` than with the LU factorization. Although the number of time-steps is more sig-
nificant with iLU(0) than with LU, the computation is already faster with iLU(0) than
with LU. This time, the size of the power system starts to be significant, and then, the
number of non-zero elements is two times smaller than with LU. Besides, the numerical
factorization consumes less computation time because there are fewer factors to com-
pute. Also, the smallest computation time was observed when the number of time-steps
is higher than with the direct solver. Furthermore, a small increase just after the min-
imum computation time is seen, and when the fill-in increases, the computation time
will reach the same computation time as the direct solver. Finally, the utilization of one
or five iterations of GMRES with iLU(`) was slower than one iteration of iLU(`) as BIM.
Table 7.3 and Table 8.1 show the overall computation time of Alg. 2 for the stiff power
systems. Also, the number of time-steps accepted and rejected are shown too. As seen in
the previous results of this section, the same time domain accuracy is found due to the
adaptive method.

7 Linear solver Tol = 10−4 Tol = 10−6

CPU in secs. (accepted/rejected)
LU 0.10 ( 116 / 2 ) 0.27 ( 368 / 6 )
iLU(30) 0.28 ( 492 / 12 ) 1.21 ( 2235 / 99 )
iLU(35) 0.13 ( 202 / 3 ) 0.47 ( 850 / 20 )
iLU(40) 0.08 ( 116 / 2 ) 0.21 ( 368 / 6 )
GMRES iLU(30) 1 it 0.55 ( 619 / 88 ) 1.75 ( 2240 / 71 )
GMRES iLU(35) 1 it 0.23 ( 239 / 17 ) 0.69 ( 866 / 27 )
GMRES iLU(40) 1 it 0.11 ( 116 / 2 ) 0.30 ( 368 / 6 )
GMRES iLU(30) 5 it 0.15 ( 116 / 2 ) 0.41 ( 368 / 6 )
GMRES iLU(35) 5 it 0.14 ( 116 / 2 ) 0.37 ( 368 / 6 )
GMRES iLU(40) 5 it 0.11 ( 116 / 2 ) 0.30 ( 368 / 6 )

Table 7.3: CPU time (in seconds) for the small stiff power system, for various iterative solvers and the direct
solver and two absolute tolerances in case of an adaptive time-step selection. The first number inside the
brakes is the total number of factorizations and the second one is the number of time-steps accepted
7.4. C ONCLUSIONS 101

Linear solver Tol = 10−4 Tol = 10−6

CPU in secs. (accepted/rejected)
LU 945.70 ( 3817 / 267 ) 2599.47 ( 11804 / 115 )
iLU(30) 326.88 ( 23729 / 207 ) 563.46 ( 41938 / 771 )
iLU(35) 172.69 ( 9328 / 15 ) 412.46 ( 23181 / 79 )
iLU(40) 142.32 ( 5804 / 5 ) 416.75 ( 18767 / 5 )
GMRES iLU(30) 1it 455.86 ( 23684 / 230 ) 793.66 ( 41839 / 759 )
GMRES iLU(35) 1it 230.00 ( 9672 / 10 ) 546.47 ( 24402 / 86 )
GMRES iLU(40) 1it 178.20 ( 5799 / 8 ) 533.13 ( 18774 / 6 )
GMRES iLU(30) 5it 171.96 ( 3817 / 267 ) 350.79 ( 11803 / 81 )
GMRES iLU(35) 5it 145.62 ( 3817 / 267 ) 333.92 ( 11803 / 111 )
GMRES iLU(40) 5it 165.69 ( 3817 / 267 ) 415.21 ( 11804 / 111 )

Table 7.4: CPU time (in seconds) for the large stiff power system, for various iterative solvers and the direct
solver and two absolute tolerances in case of an adaptive time-step selection. The first number inside the
brakes is the total number of factorizations, and the second one is the number of time-steps accepted.

From Table 7.3, the same remarks for the small stiff power systems can be made for
the small non-stiff power systems too. A direct solver can be recommended. However,
the results for the large power systems are a bit different. GMRES method is seen to
reduce the number of time-steps, but does not bring any gain in CPU time. Table 8.1
shows that for the large stiff system, the use of iLU(40) as a BIM results in a more-than-
five fold reduction of the CPU-time employed by the direct solver, for both tolerances
employed. This gain is achieved despite the fact that ILU(40) requires roughly one and
half times more time-steps. The GMRES-based solvers are seen to be more competitive
than the direct solver in terms of CPU time. For both tolerances, the use of five outer
7
iterations makes these solvers accurate to the degree that the number of time-steps re-
quired equals the one used by the direct solver. The CPU time required, however, is not
less than ILU(40) method as a BIM.
From all these results, the choice of a linear solver during the integration process
depends on three factors - the size and the stiffness of the power system under consid-
eration, and the accuracy desired. The size of the linear system of equations affects the
time to resolve it. For small systems of equations, LU factorization is recommended.
When the size increases, a new factor is to be taken into account - the stiffness. The stiff-
ness plays a major role in solving the linear system of equations using iterative method,
due to the eigenvalues of the matrix A. For example, a medium size power system can be
simulated with LU factorization for non-stiff systems, and with adequate basic iterative
methods for stiff systems. The last factor is the accuracy for a time-stepping algorithm.
When the required accuracy is high, it may impose a very small time-step. For stiff large
power systems, GMRES method with a suitable preconditioner is recommended.

7.4. C ONCLUSIONS
The sequence of linear systems for the stage vectors at each time-step were solved inac-
curately using an incomplete factorization, with various levels of fill-in. A limited num-
ber of iterations of this solver, either as a basic iterative method or as preconditioner for
 102 R EFERENCES

GMRES, are used. This inner solver was combined with an outer time-stepping loop,
with either a fixed or an adaptively selected time-step. In the case of a fixed time-step,
the time-stepping procedure fails to converge in case the linear solver is not sufficiently
accurate. This non-convergence is more likely to occur in cases where the power system
modeled is stiff. With the adaptive time-step strategy, the non-convergence is compen-
sated by reverting to a smaller time-step, with linear systems that are easier to solve. Nu-
merical results show a speed-up of computations that increases with the problem size.
For the largest problem considered in this chapter, a speed-up by a factor of more than
five was observed.

R EFERENCES
[1] S. Balay, S. Abhyankar, M. F. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin,
V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, L. C. McInnes, K. Rupp, B. F.
Smith, S. Zampini, H. Zhang, and H. Zhang, PETSc Users Manual, Tech. Rep. ANL-
95/11 - Revision 3.7 (Argonne National Laboratory, 2016).

[2] S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith, Efficient management of paral-

lelism in object oriented numerical software libraries, in Modern Software Tools in Sci-
entific Computing, edited by E. Arge, A. M. Bruaset, and H. P. Langtangen (Birkhäuser
Press, 1997) pp. 163–202.

7
 8
C ONCLUSION

8.1. I NTRODUCTION
C HAPTER 1 defined several research questions and some research objectives. Then, in
Chapters 2 and 3, a short introduction to the different parts of the thesis was given
for the electrical and mathematical backgrounds. Chapters 4 and 5, and Chapter 6 and 7
display the contributions to electrical engineering and numerical mathematics, respec-
tively. Finally, this concluding chapter summarizes the answers to the research ques-
tions, and the main contributions of the presented research.
This chapter is organized as follows: Section 8.2 gives the answers to the research
questions; the main contributions are given in Section 8.3; and, Section 8.4 provides
some recommendations for future research.

8.2. A NSWER TO THE RESEARCH QUESTION

Let first consider the first research question, stated in Section 1.2:
Is it possible to design a new methodology to model a power system, which allows
for the use of adaptive time-stepping strategy with a relatively small computation
time for re-computing or updating the network equations in the case of a switching
event within an ordinary differential equation formulation?
This can be answered in three parts:

• As shown in Chapter 4, the block modeling method permits the achievement of

power system equations in a state space representation. As defined in Chapter
3, this mathematical formulation allows for the utilization of an adaptive time-
steping algorithm during the numerical integration;

• After a switching action, the re-computation of the network equations - in some

particular cases - requires the utilization of the Schur complement (Chapter 4).
However, the Schur complement is usually applied on a small system of equations.
Also, the user could implement the change of topology analytically: for example,
for the utilization of power electronics. Finally, the utilization of capacitance on

103
 104 8. C ONCLUSION

an extra node allows one to avoid the Schur complement. This approach increases
the number of differential variables. However, the matrix A can then be bad con-
ditioning; this can lead to iterative methods requiring more iterations to solve the
systems of equations during the integration process (Chapter 3);

• In Chapter 5, the use of arc models was shown with this new approach to model
power systems. The use of arc models results in non-linear differential equations.
The Jacobian matrix can be computed analytically. The use of numerical differen-
tiation is thus avoided. This is a distinct advantage of the block modeling method
for transient simulations.

The second research question stated in Section 1.2 is:

How strong is the interaction between the time loop of the numerical implicit in-
tegration method of an ordinary differential equation and the solver for obtaining the
time domain solution of a power system?
The interaction between the time loop and the solver of the systems of equations is
hard to quantify. This interaction depends on the size of the power system, the stiffness
of the model, the time-step, and the time-stepping strategy used. Then, the answer to
this question can be split into three segments:

• As shown in Chapters 6 and 7, the use of an inexact solver can reduce the compu-
tation time. However, the accuracy of the time domain solution is affected. Mean-
while, two others parameters can affect the accuracy of the solution: the electri-
cal network models and the integration method. Finally, a compromise has to be
reached between computation time and accuracy of the results;

• As shown in Chapter 7, the use of the adaptive time-stepping strategy adds sev-
eral advantages. The first one is that by reducing the time-step, the approximate
solver converges to a sufficiently accurate solution. However, the number of time-
steps increases. The second one is that several iterations may not help decrease
8 the computation of the time domain solution. The last advantage is the relative
speed-up observed when an appropriate approximate solver is chosen;

• From the results of Chapter 7, the following table summarizes the type of linear
solvers to use, according to the size and the stiffness of the power system, and the
accuracy required during the numerical integration method.

nb 1000
Low accuracy
Non-stiff LU iLU(`) iLU(`)
Stiff LU iLU(`) GMRES+iLU(`)
High accuracy
Non-stiff LU iLU(`) GMRES+iLU(`)
Stiff LU GMRES+iLU(`) GMRES+iLU(`)

Table 8.1: Summary of the type of linear solver to use, according the number of differential variables nb and
the stiffness of the power system, and the accuracy required
8.3. C ONTRIBUTION 105

8.3. C ONTRIBUTION
The first contribution of this thesis is to give a new view on power system modeling. In-
stead of considering it as a connection of lumped elements, it is possible to see each
element that comprises it as a set of equations. Those components are then connected
- and therefore, those sets of equations have to be interconnected too. This new ap-
proach to viewing power systems permits the linking of different sets of equationsusing
the properties of the power system by itself.
The second contribution of this thesis is the study of inaccurate solvers to reduce
the computing time during the numerical integration method. As seen in the previous
chapter, the computation time can be reduced by at least half, in the case of a large power
system.

8.4. F UTURE WORK

Based on this study, three conjoint directions can be considered for future research stud-
ies in the electrical engineering or numerical analysis fields.

• The use of power electronics in power system

Today, the utilization of converters in power systems is increasing. According
to the results of this thesis - and particularly, the results of the block modeling
method - the simulation of power electronics in a large electrical network could
be improved. For example, considering a converter as a block model simular to
the arc model, where the conductivity could change, could be one such simula-
tion.

• Modeling transformers
Another optic for improving the block modeling method would be the ability to
model transformers. Transformers are usually non-linear because of the relation-
ship between magnetic and current excitation. Then, if the non-linearity would be
handled, similiar to how the arc models are handled, especially for the calculation 8
of the Jacobian matrix, a significant speed-up may appear.

• Inexact Newton methods for transients in power systems

Approximate solvers allow a gain of relative computation time for the large power
system cases in this thesis. Also, approximate solvers gave good results because of
the time-step employed. Besides, when a non-linearity appears, a small time-steps
in necessary; then, the use of approximate solvers could help reduce the compu-
tation time without compromising the quality of the time domain solution.
 A CKNOWLEDGEMENTS
First of all, I thank Prof. L. van der Sluis and Prof. C. Vuik for offering me the opportunity
to work on the simulation of electrical networks within the Intelligent Electrical Power
Grids and the Numerical Analysis groups at the faculty of Electrical Engineering, Math-
ematics and Computer Science of the Delft University of Technology, The Netherlands.,
and the initiative Power Web. Also, I thank Dr. D. Lahaye my daily supervisor.
I would like to thank Prof. L. van der Sluis for his expertise in the history and his over-
look of electrical engineering. I also liked our conversations, which is as important as our
field of study, baking cake. Finally, I am thankful for his help to visit RTDS Technologies
Inc. for 3 months.
I would like to acknowledge Prof. C. Vuik for his support, wisdom and way to manage
the communication between the electrical engineers and mathematicians. Also, I am
thankful for allowing me to go for 3 months in Canada.
I would like to be grateful to Dr. D. Lahaye for all discussions we had during the last
4 years at least one every week. I learned with him the importance of being clear. Also, I
would like to thank you for the 4 years of co-operation. I would like to thank for having
challenging me and make me aware of how scientific world works.
I express gratitude to Paul Forsyth for accepting to supervise me during 3 months at
RTDS Technologies Inc., Winnipeg, Canada. I had a nice time there where I discovered
what I wanted to after studying. I met amazing persons there, Eshan, Gregory, Melanie,
Christian and all people of the company.
I am thankful to IEPG department where I worked. I would like first to thanks all my
officemate Laura, Freek, Gerben, Shahab, Arun, Claudio, Rishabh and Zong. Without
all our conversations, studying with you all, would not have passed so fast. Also, I don’t
forget Matija, Ilya, Peter, Mart, Kaikai, Jose, Hossein, Lian, Marjan, Ellen, Mart, Martjin,
Sharmila, Arcadio and Arjen for our scientific and unscientific talks. I would like to ac-
knowledge Mario, Bart and Swasti for the losing hair equation and all talks we had about
that primordial equation.
Ann-Sophie and Heiko, I am so happy that we built a tower of glasses. Without you,
my 4 years in Delft would have not been the same.
3 years ago, I started to play Ultimate Frisbee and that was legend. . . wait for it . . .
dary and still it is. I met a lot of exceptional people around the world, thanks Rienk,
Mario, Baline, Alex, Michiel, Annika, . . . and Spike.
I would like to thank my new colleagues and DigSILENT Gmbh for their support.
Finally I would like to thank my family and friends from France who supported me
along this amazing experience.
22nd October 2017, Nehren, Germany,
Romain Thomas

107
 C URRICULUM V ITÆ
Romain T HOMAS

09-04-1989 Born in Quimper, France.

2004–2007 Hight school

Lycée Yves-Thépot, Quimper, France

2007–2009 DUT Génie électrique et informatique industrielle

Universitée de Bretagne Occidental, Brest, France

2009–2012 Master of Engineering in Électronique, Énergie Électrique & Automatique

INP-ENSEEIHT, Toulouse, France

2012–2016 PhD. in Electrical Engineering and Numerical mathematics

Technische Universiteit Delft, Delft, The Netherlands

2017 Software developer

DIgSILENT, Gomaringen, Germany

109
 L IST OF P UBLICATIONS

L IST OF J OURNAL P UBLICATIONS

• R. Thomas, L. van der Sluis, D. Lahaye and C. Vuik, Network Calculations with
Switching Devices, under review for IET Electric Power Application.

L IST OF C ONFERENCE P UBLICATIONS

• R. Thomas, D. Lahaye, C. Vuik and L. van der Sluis, Inexact Linear Solvers for Tran-
sients in Power System Networks, Proceedings of NUMELEC 2017, Paris, France. (In
preparation, abstract accepted)

• R. Thomas, D. Lahaye, C. Vuik and L. van der Sluis, Simulation of Arc Models with
the Block Modelling Method, Proceedings of IPST 2015, Cavtat, Croatia.

L IST OF B OOK CONTRIBUTIONS

• R. Smeets and al. , Switching in electrical transmission and distribution systems,
John Wiley & Sons city, Chichester, 2014.

• P. Schavemaker and L. van der Sluis, Electrical Power System Essentials,John Wiley
& Sons city, Chichester, 2014.

111