LTH Iea 1060 PDF

Automatic Power System
Restoration
Application of a Search Algorithm
Lars Lindgren
Licentiate Thesis
Department of Measurement Technology and
Industrial Electrical Engineering
2009
Department of Measurement Technology and Industrial Electrical Engineering
Faculty of Engineering
Lund University
Box 118
SE-221 00 LUND
SWEDEN
http://www.iea.lth.se
ISBN 978-91-88934-52-9
CODEN:LUTEDX/(TEIE-1060)/1-144/(2009)
c Lars Lindgren, 2009

°
Printed in Sweden by Media-Tryck
Lund University
Lund, 2009
ii
School of Technology
Malmö University
iv
Abstract
Large blackouts are rare but can have a high impact on society and the econ-
omy. This impact strongly depends on their duration and therefore speedy
restoration is important.
More distributed and renewable generation and more interconnected power
systems that are operated close to their limits will lead to new challenges in
power system restoration. At the same time society is becoming ever more
dependent on electric power supply. The ability to restore the power supply
after blackouts is therefore vital and means to improve it should be investi-
gated.
For the operators in a control centre, restoration of a power system in-
volves taking a number of actions. If possible these follow written instruc-
tions, but often also needs to be improvised. This thesis proposes that au-
tomation is also applied to this task. An algorithm that finds a suitable se-
quence of actions to restore the power supply after a blackout can be used
both to assess different strategies or potential restoration difficulties before
any real event and to support the operators during an actual restoration situ-
ation.
This thesis shows that the restoration problem is NP-hard and proposes
a simple heuristic restoration algorithm based on the well known tree search
algorithm A∗ . The proposed algorithm is applied to a simplified power sys-
tem model using static load flow analysis.
NORDIC32, which is used as test system, is a fictitious power system
model with 32 substations and 36 generators. It attempts to broadly represent
the synchronized Nordel transmission system.
The algorithm has been found to be robust against variations in both the
search parameters and the power system model. For black start in the base
case the algorithm typically finds a sequence of around 750 actions that re-
v
store all loads after the algorithm has tested about 2000 actions. The actions
include network switching, startup and setpoint changes of generators and
load reconnections. On a 2.2 GHz AMD Opteron processor this computa-
tion takes less than 30 seconds.
The fact that the computation time is as short as 30 seconds indicates that
the algorithm realistically could handle more complex systems either larger
or with more detailed modelling.
Keywords
Power system restoration, Load flow analysis, Tree searching, A∗ -search,

Bulk power system, Black start, Power system control, Power systems, Search
methods, Automation, Predictive control, NP-hard.
vi
Sammanfattning
Omfattande strömavbrott är ovanliga men kan få stora konsekvenser för sam-

hället och ekonomin. Eftersom konsekvenserna är starkt beroende av avbrot-
tens längd är snabb återstart av kraftsystemet viktig.
Fler distribuerade och förnybara energikällor och mer sammanlänkade
elkraftsystem som drivs närmare sina gränser kommer att leda till nya utma-
ningar vid återstart av elkraftsystemet. Samtidigt har samhället blivit allt mer
beroende av tillförlitlig elförsörjning. Förmågan att återställa elförsörjningen
efter strömavbrott är därför viktig och möjligheterna att förbättra denna bör
undersökas.
För kontrollrumspersonalen innebär återstart av kraftsystemet att ett stort
antal åtgärder vidtas. Detta följer om möjligt en skriven instruktion men krä-
ver vanligen mycket improvisation. Denna avhandling föreslår att även auto-
matisering utnyttjas. En algoritm som finner en lämplig sekvens av åtgärder
för att återställa elförsörjningen efter ett strömavbrott kan användas både för
att utvärdera olika strategier eller bedöma eventuella svårigheterna vid en
återstart innan någon verklig händelse inträffat och för att stödja operatörer-
na under en verklig återstart.
Denna avhandling visar att återstartsproblemet är NP-hårt och föreslår
en enkel heuristisk återstartsalgoritm baserad på den välkända trädsöknings-
algoritmen A*. Den föreslagna algoritmen arbetar med en förenklad modell
av kraftsystemet baserad på statiska effektflödesberäkningar.
Som testsystem har NORDIC32 använts , vilket är en fiktiv kraftsystemsmo-
dell med 32 kopplingsstationer och 36 generatorer. NORDIC32 är ett försök
att återspegla huvuddragen hos det synkroniserade Nordel-systemet.
Algoritmen har visats vara robust mot variationer både i sökparametrar
och i kraftsystemsmodellen. För dödnätsstart med de grundparametrar som
använts genererar algoritmen typiskt en sekvens med omkring 750 åtgärder
vii
för att återställa elförsörjningen till alla förbrukare. För att finna dessa åtgär-
der testar algoritmen ca 2000 åtgärder. Dessa åtgärder innefattar kopplingar
i nätet, start och ändring av driftläge för kraftverksenheter samt belastnings-
tillkopplingar. På en 2,2 GHz AMD Opteron processor tar denna beräkning
mindre än 30 sekunder.
Det faktum att beräkningstiden är så kort som 30 sekunder visar att al-
goritmen bör kunna hantera mer komplexa system, antingen större eller med
mer detaljerade modellering.
viii
Acknowledgement
This project was initiated by Dr. Bo Eliasson who also supervised the begin-
ning of this project and my B.Sc. project.
This thesis had not been possible without the support from people around
me. I want to thank my supervisors at Lund University, especially:
Assoc. Prof. Olof Samuelsson, my main supervisor for his strong sup-
port even in periods when the work has progressed slowly and for his help
with both content and language.
Prof. Sture Lindahl, for using his expertise to give me many new views
and insights.
Prof. Gustaf Olsson and Magnus Akke for their enthusiasm and insights.
My colleagues at Malmö University and at the Department of Measure-
ment Technology and Industrial Electrical Engineering at Lund University
have been very important in the work with this thesis. I want to especially
thank Prof. Bengt J. Nilsson for his help with the computer science part of
this work and fellow postgraduate student Tech. Lic. Christian Andersson
for our close cooperation in the beginning of the project. I also want to thank
John Niubo at Malmö University for his support.
Energinet.dk, previously Eltra, has financed this project which is grate-
fully acknowledged. Two persons at Energinet.dk have been especially im-
portant, they are Assoc. Prof. John Eli Nielsen and Dr. Per Lund.
Of course I also want to thank my family and others for their support
during this work.
I have learned very much during this work.
ix
x
Terminology
Active power The average power over a cycle in the AC voltage.
Bottom up restoration Restoration from several points and then synchro-

nizing the different islands.
Circuit breaker A mechanical switch that can operate fast and is able to
interrupt fault current at short-circuits without being damaged.
Disconnector A mechanical switch that only can be opened and closed

when the current or the voltage difference is very low. Disconnectors
are used to change the substation configuration, often after a circuit
breaker has been used to turn off the current. They are also used to
safely isolate parts for repair and maintenance.
Distributed slack bus A slack bus in a power flow calculation is the bus
where the active power is adjusted to get the active power balance
right. In a power flow calculation with distributed slack bus this ad-
justment is distributed to several buses according to some weighting
factor.
Distribution network The low voltage and medium voltage networks that
distribute the power to small and medium sized customers. It is com-
mon that distribution networks are operated in a radial configuration
since that simplifies the protection system.
Distributed generation Small generating units that are connected to the

distribution network.
Electrical island A part of a power system that is disconnected from the rest
but remains energized. The main energized part of the power system
xi
can be seen as special case of an island.
Energy not delivered The energy that customers wanted to buy but could
not because of a power outage. Normally calculated as the predicted
consumption in the area that is without power.
Evaluation function A function that is used to evaluate and compare how

good states or solutions are.
Flat start Flat start means that all voltages are set to a fixed voltage, usually
the nominal value (1 p.u.) and all voltage phase angles are set to 0
before the first iteration in a power flow calculation.
Generating unit A power plant can have one or more generating units, each
mainly consisting of a turbine and a generator. Each generating unit
can be started independent of the other generating units.
Heuristics Method based on experience such as “rule of thumb”.

In a search algorithm context, heuristics cannot be guaranteed to gen-
erate the correct answer but are useful in some cases.
Load flow analysis See Power flow analysis.
N-1 criterion A rule in power system operation that demands that the sys-
tem shall be able to withstand the failure of any component without a
collapse.
NP-complete A class of computation problems that are considered hard to

solve efficiently.
NP-hard The class of all computation problems that is at least as hard to

solve as any NP-complete problem, see section 5.1.
Power flow analysis A calculation of the active and reactive power flows
in a power system given the generation and consumption in the sys-
tem. Can include different regulators such as voltage regulators and
tap changers. The most demanding step is to calculate the voltage
magnitude and phase angle for all buses, which is done iteratively. Af-
ter that the power flows on the lines and in the transformers are trivial
to calculate.
xii
Power plant A power plant consists of one or more co-located generating
units.
Reactive power Power that over a cycle in the AC voltage moves in both
directions so that on average no power is transferred but it still gener-
ates losses. In a system with inductive power lines the voltage can be
controlled by controlling the reactive power.
Renewable generation Power generation that use renewable energy sources

such as wind, sunshine or biomass. Often confused with distributed
generation since renewable generation is often distributed.
Substation A location for switching equipment and transformers used to

connect power lines to each other.
Switch A device that can close or open an electrical circuit, see circuit
breaker and disconnector.
Synchronization The act of closing a circuit breaker between two electrical

islands. Before the synchronization it is important that the difference
between the voltage magnitudes, frequencies and voltage phase angles
in the two parts are small enough. Synchronization is rather much like
engaging a mechanical clutch with no slip in that the axes in the syn-
chronous machines in the whole synchronized system will be spinning
synchronously.
Three phase symmetry A situation where the currents and voltages in the
three phases are identical except for a phase shift of 120 degree be-
tween the phases. In this case it is sufficient to calculate the power,
voltage and current in one phase and then multiply the power by three
to get the three phase power. High voltage AC systems are normally
operated close to three phase symmetry.
Top down restoration Restoration from one island in the system and then
expanding that island by energizing more and more components.
Topology calculation In power system analysis topology calculation is the

calculation of how the system is connected given the state of the switches.
It answers two questions:
xiii
• Which components are connected directly to each other and thus
have the same voltage?
• Which electrical islands exists in the system?
It’s a necessary pre-processing step before a power flow calculation

can be performed.
Transmission system The high voltage part of a power system that trans-
mits power over long distances.
xiv
Contents
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Scope of the thesis . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . 7
2 Overview 9
2.1 The operational states of the power system . . . . . . . . . . 9
2.2 Restoration . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Power plant characteristics . . . . . . . . . . . . . . . . . . 15
2.4 Historical development of power system restoration . . . . . 18
2.5 Problem formulation . . . . . . . . . . . . . . . . . . . . . 19
3 Modeling and simulator 21

3.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Power system components . . . . . . . . . . . . . . . . . . 24
3.3 Action sequence and states . . . . . . . . . . . . . . . . . . 27
3.4 Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4 Test system 31
4.1 Adaptations and details . . . . . . . . . . . . . . . . . . . . 31
4.2 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
xv
5 Characterization of the problem 41
5.1 NP-hardness . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6 Search algorithms 47
6.1 The A∗ algorithm . . . . . . . . . . . . . . . . . . . . . . . 49
6.2 Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.4 Power system restoration . . . . . . . . . . . . . . . . . . . 54
6.5 The fast and simple evaluation function, f1 (N, A) . . . . . . 58
6.6 The full evaluation function, f2 (N ) . . . . . . . . . . . . . 59
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7 Simulation results 63
7.1 Resulting sequence . . . . . . . . . . . . . . . . . . . . . . 63
7.2 The search . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
8 Algorithm parameter sensitivity analysis 79

8.1 The fast and simple evaluation function, f1 (N, A) . . . . . . 79
8.2 The full evaluation function, f2 (N ) . . . . . . . . . . . . . 81
8.3 Other parameters . . . . . . . . . . . . . . . . . . . . . . . 85
8.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9 Model parameter sensitivity analysis 89

9.1 Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2 Random faults . . . . . . . . . . . . . . . . . . . . . . . . . 92
9.3 Scaling of the load . . . . . . . . . . . . . . . . . . . . . . 92
9.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
10 Conclusions 97
10.1 Algorithm and test system . . . . . . . . . . . . . . . . . . 97
10.2 Computation time . . . . . . . . . . . . . . . . . . . . . . . 99
11 Future work 101

11.1 Time dependent and dynamic models . . . . . . . . . . . . 101
11.2 Other model improvements . . . . . . . . . . . . . . . . . . 102
11.3 Algorithmic improvements . . . . . . . . . . . . . . . . . . 102
11.4 Distributed and hierarchical restoration . . . . . . . . . . . . 103
xvi
11.5 Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . 103
References 105
A Implementation 111
B Net data 115
xvii
xviii
Chapter 1
Introduction
The reliability of a power system depends on the frequency and the duration
of power outages. Much of the research about power systems focuses on
reducing the frequency of power outages.
The focus in this thesis is instead to reduce the duration of large power
outages or blackouts by improving the restoration process.
1.1 Background
This research project deals with the electric power system and specifically
how to manage a restoration situation in which large parts of the power
system have lost voltage after a fault or a collapse. In such situations, the
start-up and regulation of generating units must be accurately coordinated
with the connection of loads and power lines, to eliminate the risk of another
collapse.
This project was motivated by the growing amount of distributed power
generation which gives new questions about how restoration should be per-
formed. The large number of small generating units also raises questions
about if restoration can be performed automatically at a lower level in the
power system.
Five approaches to the study of these issues have been identified:
1. Theoretical arguments.
2. Analysis of actual sequences of events.
1
2 Chapter 1. Introduction
3. Simulation studies with sequences of actions produced manually by a

person, without real-time requirements.
4. Realistic real-time simulation exercises in which a team of operators

attempts to manage prepared scenarios unknown to them, to determine
how well they manage the situation, given different strategies for the
work.
5. To use automatic search algorithms to find suitable sequences of ac-

tions in different restoration situations and assessment of how difficult
it is for the algorithm to start the power system in each situation. By
varying algorithm parameters and comparing the results, conclusions
as to which strategy should be applied can be reached.
It is difficult to develop theoretical arguments in such a way that they are
applicable in practice and generally applicable to such complex problems.
Analyzes of actual sequences of events have already been done; see [4,
20], for example. The problem is that one can only investigate what actually
happened; to determine whether alternative strategies would have been better
is hard.
Simulation studies with sequences of actions produced manually by a per-
son with or without real-time requirements must be structured in some way;
otherwise they will lead to results that depend entirely on the person’s skill
and expectations in different situations. If the production of the sequence
is structured according to a particular set of rules, it would bee far better to
automate the sequence generation. It then falls into category 5 above.
Realistic real-time simulation exercises as mentioned in category 4 above
would have been interesting, but would have required access to operators
to act as experimental subjects and set up realistic scenarios in which all
significant aspects of the situation were simulated. It was felt that this was
not possible within the scope of the project.
This licentiate-thesis deals with the development of an algorithm to be
used in approach number five above.
1.2 Scope of the thesis
In order to compare different power system restoration strategies or situa-

tions in terms of energy not delivered, it would be helpful to have an algo-
1.3. Previous work 3
rithm for finding a sequence of actions that restores a power system.

The idea is that while not necessarily giving the optimal solution the al-
gorithm can be used for finding a feasible solution and this can be used to
estimate the energy not delivered1 . The aim of this thesis is to investigate
such an algorithm. The algorithm is tested on a fictive power system.
The power system model is simplified by not including any time depen-
dent or dynamic models and only modelling deterministic behaviour.
It is assumed that the protection system makes sure that the system state
after the collapse contains no equipment that operates outside its specifica-
tion. It is also assumed that the state of the system is fully known.
After all loads are restored some process is needed to reach a more normal
operating point that fulfils the N-1 criterion and other constraints for normal
operation this is not discussed in this thesis.
1.3 Previous work
Power system restoration has of course been a necessary part power system
operation since the first power systems were built. The traditional method
is to have written instructions with general rules that are executed manually
[15]. Since the restoration situation can often not be known beforehand this
includes more or less improvisation. Since the use of computers became
common there have been attempts to automate the whole process or part of
the process.
The book “Power System Restoration methodologies & implementation
strategies“ by M. M. Adibi collects many of the key papers in the area and
complements these with comments [8]. The book was published year 2000.
The power system restoration research can be categorized according to
several different criteria:
Power System type Power system restoration needs to be done on different

types of power systems and at different levels, such as transmission
systems [18, 25, 36], distribution systems [19, 33, 36] and shipboard
power systems [37].
Disturbance Some research focuses on the common case that only a small
part of the system is without power [18], other research focuses on
1
Defined as the energy that the customers are prevented from consuming by the blackout.
black start after a complete blackout [27, 25, 33].

Sub problems Much research focus on different sub problems in power sys-
tem restoration such as generator start-up sequence [26], standing volt-
age phase angles [38], target switch configuration without finding a se-
quence of actions [18, 19, 27] and selecting suitable islands to restart
[30].
Modelling The resolution of the model that is used when analyzing power
system restoration is a compromise between many details to capture
the behaviour of the real system or few details to keep down the im-
plementation complexity and the computation time. Both static power
flow calculations [18, 27] and dynamic electromechanical models [39,
22, 33] are used. Some aspects not normally included in power system
models such as generating unit start-up times and ramping are impor-
tant to consider in restoration studies.
Automation level Much research is on the fundamental issues in power sys-

tem restoration and applies to manual restoration as well as automated
restoration [9, 11]. Other research treats how to support the operators
during restoration using simple prepared rules [15] or how to automate
power system restoration [25, 33, 36] or parts of the process.
Algorithm Proposed algorithms for automatic restoration or operator sup-
port use many different algorithms such as expert systems [29, 22, 36],
petri-nets [25], tree search [25, 33], genetic algorithms [19, 36], dis-
crete particle-swarm optimization [27] and combined tabu search and
a multi-agent system [18].
The most relevant articles found are listed below. The following criteria
have been used when selecting the articles:
• The used algorithm can find a sequence of actions for restoration from
black start or near black start of bulk power systems.
• Describes the used algorithm in enough details so the inner workings
of the system can be understood.
• Needs no system specific manually prepared restoration plans or guide-
lines.
1.3. Previous work 5
• Uses a significantly different approach from the other shown articles.
The found articles are:
• ”Backtracking based algorithm in hierarchical time-extended petri net

model for power system restoration“ by Dong Liu et al. from 2005
[25] uses a petri net model and heuristic rules to choose which actions
to test and backtracking when it does not find a way forward. The
presented implementation does not perform a power flow calculation
and can therefore not check voltage levels and so on.
• ”An agent approach to bulk power system restoration“ by T. Nagata et

al. from 2005 [31] uses a decentralized multi-agent system where the
different parts of the power system has associated computer programs
that negotiate with each other. The presented implementation has no
constraints on reactive power or voltage and does not perform a power
flow calculation. An overview of an agent based system that includes
power flow calculation is given in [24].
• ”Development of an interactive rule-based system for bulk power sys-

tem restoration“ by Teo C.Y. and Wei Shen from 2000 [13]. This arti-
cle is interesting in that it compare the values from the simulation with
measured values and updates the model accordingly and then updates
the model. It uses very fixed rules for how to restore the system for
example ”IF any two subsystem have met 50% of their local load de-
manding, THEN these two subsystems are ready to synchronize.“ and
the active power generation is only controlled by the local frequency.
This probably makes the system unsuitable for comparing restoration
in different situations.
• ”Generation capability dispatch for bulk power system restoration: a

knowledge based approach“ by Liu C.-C. et al. from 1993 [40] fo-
cuses on finding a generator startup sequence given the capabilities of
the production units. The A∗ -algorithm is used to find paths in the
power system directly. It does not check the power flow solution for
the proposed solution steps.
• ”Development of a guidance method for power system restoration“

by Kojima Y. et al. from 1989 [12] uses an expert system to guide
operators during restoration. First a target state is defined for the power
system a few hours later that the operators can modify if they wish.
Then the expert system tries to find actions that take the system to the
target state.
An interesting article on restoration at the distribution level is ”New chal-

lenges in power system restoration with large scale of dispersed generation
insertion“ by Thi Thu Ha Pham et al. from 2009 [33]. The focus in this
article is on restoration at the distribution level by using distributed genera-
tion while waiting on the restoration at the transmission level. This article
uses a depth first tree search to enumerate switching sequences that fulfils
the power balance in each step. They are then ordered by how much load
they restore and a dynamic simulation is preformed on each until a valid se-
quence is found. This method is interesting but probably results in too many
switching sequences to test if applied on the transmission level. The article
also describes solutions to a number of related problems.
None of these methods have gained widespread use. Furthermore there
seems to be no commonly used test-cases to test algorithms against.
Finally it has been stated without proof that the power system restoration
problem is NP-hard [21].
1.4 Contribution
In this work, a new tree search algorithm for power system restoration has
been developed. It is based on the well-known A∗ -algorithm. The algorithm
has been shown to be able to do a full restoration of a rather complex power
system with a simplified model. This is done without specifying any overall
restoration plan or specific goal state beforehand.
The work has been published in the papers:
Lindgren, L. and Eliasson, B. Automated power system restoration, Nordic

Wind Power Conference, November 1-2, 2007, Risø DTU, Denmark.
Samuelsson, O., Lindgren, L. and Eliasson, B. Simulated power system

restoration, 43rd International Universities Power Engineering Conference,
IEEE, september 1-4, 2008, Padova, Italy.
1.5. Applications 7
The performance of the algorithm under different conditions has been in-
vestigated.
The power system restoration problem has been shown to be NP-hard.
This was done by reduction from Steiner tree in graphs.
1.5 Applications
An automatic power system restoration algorithm can be used in different

ways:
• Automatic power system restoration.
• Operator support, suggesting actions to the operator in real time.
• Off-line planning for restoration in advance of the blackout. This in-

cludes both comparing different restoration plans/strategies’ and com-
paring restoration sequences after modifications in the net.
The focus in this thesis is planning for restoration in advance.
1.6 Outline of the thesis
Chapter 2 gives an overview of the power system restoration problem from a

practical power system operation view point. Readers with previous knowl-
edge of these issues can skip this chapter.
Chapter 3 and 4 describe the modelling of the power system components
and the used test system. Especially the sections 3.4 about time calculation
and 4.2 about power flow visualization are important in order to interpret the
results.
Chapter 5 places the power system restoration problem in a computer sci-
ence context and shows that the problem is NP-hard. It is rather independent
from the rest of the thesis.
Chapter 6 introduces search algorithms, explains the used algorithm and
the adaptation of the algorithm to the power system restoration problem.
Chapter 7 presents restoration sequences that are generated by the algo-
rithm when applied to the test system.
Chapter 8 and 9 on sensitivity analysis show the effect of changing the
search and model parameters respectively.
Chapter 10 summarizes and evaluates the results.

Chapter 11 proposes possible improvements to the presented algorithm.
Appendix A gives a brief overview of how the software for this project
was developed.
Appendix B contains the model file that describes all the details of the
used test system and the test cases.
Chapter 2
Overview
2.1 The operational states of the power system
In order to describe operating situations in a power system they can be cat-

egorized in five groups called operational states of the power system, see
Figure 2.1 [16]. Each of these states has its own operational rules.
In the normal operational state the objective is to run the power system as
economical as possible with enough security margins. Typically it is required
that the system should be able to withstand any component failure at any
time without any power interruption to the consumers. This is called the N-1
criterion.
When component failures and/or unexpected conditions lead to a state
where the N-1 criterion is no longer fulfilled the power system is said to be
in the alert state. In this state the goal for the operators is to bring the system
back into the normal operating state by preventive control actions. In the
alert state all or almost all load is still supplied.
Additional disturbances in the alert state can cause the system to enter the
emergency state. This state is characterized by problems such as overloads,
over/under-voltages, over/under-frequency and instability. Most actions in
this state are taken by the protection system. There is not enough time for
manual operations. The goal in this state is to reach a stable operating point
where all equipment operates within its specifications.
The state in extremis means a partial or total blackout, the goal in this
state is to stop the deterioration of the operation state and to reach a stable
operating point from which to start restoration.
9
10 Chapter 2. Overview
Figure 2.1: The operating states and state transitions of a power system,
adapted from [16]. White arrows represents foreseen or unfore-
seen disturbance. Gray arrows represent control and/or protective
actions.
2.2. Restoration 11
When a stable operating point with partial or total blackout is reached the
system enters the restoration state. The goal in this state is to reconnect all
loads, when this is achieved the system either enters the alert or the normal
operating state. This thesis focuses on the process of going from a stable
operating point to the alert state.
2.2 Restoration
Power system restoration is a very complex task. The restoration is here

described as seen from the perspective of a TSO (Transmission System Op-
erator) such as Energinet.dk or Svenska Kraftnät.
The first problem facing the operators is to establish an overall picture of
the state of the power system. Often the cause of the blackout is not fully
known and it would take too long to analyze it before starting the restoration.
The list of alarms will be very long and the clocks are often not synchronous
so it can be hard to see what event that caused the other. It is likely that there
are unknown faults in the system.
It is also not uncommon that there are problems with the SCADA (Su-
pervisory Control And Data Acquisition) system so there can be problems
both with reading remote measurements and taking remote actions. This is
because there will be very many alarms in the SCADA system and also the
backup power in substations could fail.
The operators are left with a situation where they may not know what has
happened and why. They also do not know if there are any faults in the power
system and the SCADA system may not be fully operational.
The operators often also need to communicate with control centres in ad-
jacent transmission systems and with the power plant operators in order to
get a picture of to what extent they can help or need help. There will also
be many requests from media to get information of what happened and esti-
mates of when the power supply will normalize.
Small blackouts, involving only a few substations, can often be handled
rather easily since they occur more often than the larger blackouts so the op-
erators has more experience with them and the load can often be reconnected
as soon as the operational reserves are able to meet the demand.
Larger blackouts, affecting the whole or significant parts of the power
system, are much harder to restore since large parts of the power system
need to be energized and a significant part of the generating units will not be
available.
In order to handle restoration situations the operators have prepared restora-
tion plans and guidelines for some typical blackout situations. The operators
will try to apply these to the current situation.
Two major strategies in power system restoration can be defined, bottom
up vs. top down. Bottom up means that several smaller electrical islands
are started in parallel; these are then used to energize the transmission sys-
tem. The need to synchronize the islands can slow down the process. Top
down means that the transmission system is energized from one point and all
the lower voltage levels are energized from the transmission system and the
whole energized power system is kept synchronized.
In Sweden the general plan is to restore the system from the North and
southwards, which is an example of top down. The reason for this is that the
hydropower in the North is normally much easier to use during the restora-
tion and faster to start. Also it is very likely that some part of the Northern
part of the system will not be affected by the blackout. Energinet.dk is work-
ing towards an ambitious bottom up strategy in their cell-project [28]. In
power systems that are dominated by large thermal power plants it is com-
mon to have separate plans for the restart of each large steam turbine plant,
which is a form of bottom up restoration.
The first priority is normally to energize all substations. This has several
reasons. The remote controls of substations operate on battery power and
are normally dimensioned for powering the substation control equipment a
specified time, Svenska Kraftnät has for example chosen 12 hours in new
substations [6], but this can be uncertain due to aging of the batteries. If the
batteries run out before the substations are energized the restoration is further
delayed due to the need to send service crew with mobile diesel generators
to the substation to power the control system and recharge the batteries.
It is also important to get power to the power plants so they can initiate
their preparation to start as soon as possible. If a thermal power plant is
without power for some time it will cool down and it will take a consider-
able amount of time to restart it. Also as larger parts of the system become
energized more resources become available to the operators. If there is some
other prioritized loads in addition to substation control and power plant aux-
iliaries it is also necessary to energize the substations in order to be able to
supply the prioritized loads. Another advantage of energizing substations
2.2. Restoration 13
early is that permanent faults in disconnectors, current breakers and trans-

mission lines will often be found early.
During the energization the line currents will be very low so the reactive
generation in the lines will give problems with very high voltages. In order
to avoid this some loads may be reconnected during the energization.
Another way to handle this is to synchronize generating units as soon as
possible, as many generating units cannot operate without load this is another
reason to connect load during the energization phase. In order to limit the
reactive generation in the lines, some lines may not be energized in this phase
hence the system will have fewer redundant paths.
When the transmission system is energized the focus is shifted towards
connecting the remaining load as soon as the power plants are able to supply
it given the available transmission capacity. The behaviour of the load is very
uncertain during situations like this. One reason is that it is very hard to get
statistics that are valid for the outage duration and weather, time of day and
so on.
The level of load as a function of the time since reconnection is called
cold load pickup [9]. Thermostatically controlled loads such as electrical
heating, air condition and refrigerators will have higher consumption after a
blackout than before. The reason for this is that a much larger percentage of
the thermostats will be on than normal since the diversity in switching cycles
has been lost. On the other hand many machines and industrial processes
will need manual interventions before they can be restarted so they will have
lower consumption when reconnected compared to what they normally have.
When closing a mesh in the grid it is necessary to check so the voltage
phase angle difference over the circuit breaker is not too large. Otherwise
the transient could initiate a new collapse and damage some equipment, es-
pecially the axis in generating units could be damaged due to high torque.
In order to avoid this; the power flow often need to be changed somehow
since the voltage phase angle difference is mainly due to the active power
flow through inductive power lines [38].
Some degree of automation of the power system restoration task is al-
ready in common use. For example Svenska kraftnät uses automatic zero
voltage detection and automatic restoration equipment (In Swedish: Drift-
UppByggnadsAutomatik, DUBA).
The automatic zero voltage detection opens all circuit breakers in parts of
the grid that are not energized [3]. This simplifies the restoration since one
object can be energized at a time without manually opening all the circuit
breakers first.
The DUBA system of each circuit breaker tries to re-close the breaker if
it was closed before the event and some predefined local conditions are met
[2]. The operations performed by the DUBA operation are delayed between
10 and 300 seconds after the conditions are fulfilled and after the last DUBA
initiated action in the substation. After large blackouts the DUBA will help
with the energization of the first substations but much of the restoration need
to be done manually [4]. This is due to the fact that the DUBA mostly tries
to restore the state before the event directly with very limited ability to adapt
to the current situation.
2.2.1 Geographical aspects

The type of power plants and their location relative to the loads will give
different power system different characteristics.
Hydro power units are often located far away from the load centers since
they need to be placed were the water flows are. Thermal power plants are
often located closer to the load since there is more freedom when placing
them even if then need for cooling water can give some restrictions.
Industry that consumes very much power is often located in areas with
good power supply while the power supply is a very minor concern when
placing most other loads such as homes and offices.
The load will be very weather dependent in areas with much electrical
heating or air conditioning.
In large cities where it is not feasible to use overhead lines high voltage AC
cables are used. These have very high capacitance so the reactive generation
can be a major concern. One other concern is that they often need power in
order to maintain the oil pressure that is needed for the insulation.
Neighbouring power systems are often interconnected with EHV AC (Ex-
tremely High Voltage Alternating Current) lines and then form one AC power
system from a physical point of view even if they sometimes are regarded as
separate systems from an administrative view point.
If the systems are separated by a significant distance of water, have dif-
ferent frequencies or there are some other reason to not connect them by AC
lines then they can be connected by HVDC (High Voltage Direct Current).
This gives a much looser coupling between the systems.
2.3. Power plant characteristics 15
In some cases it is also economical to use HVDC in parallel to an AC

system as reinforcement. Unlike AC-lines the active power flow and the
reactive power can easily be controlled by HVDC-lines.
All these aspects will give different power systems very different char-
acteristics and therefore the restoration strategy needs to be adapted to the
specific case.
2.3 Power plant characteristics
Power plants have an important role in bulk power system restoration. Differ-
ent types of power plants have very different characteristics, the most impor-
tant difference is that steam turbine power plants have much longer starting
times than for example hydro power and wind power.
2.3.1 Hydro power
Hydro power has very short starting time (a few minutes) and the active
power can be ramped up and down very fast (100 %/minute). It also runs
with high efficiency on partial load, and for this reason it is often used as
regulating power. A drawback when using hydro power as fast frequency
regulating power is that the power output has a non minimum phase charac-
teristic, e.g. when opening the guide vanes to give more power, the power
will temporarily decrease due to the acceleration of the water in the water-
ways. This is especially true in power stations with long waterways.
It has very low marginal cost but since a river has a limited amount of
water and the water can be stored in dams it is common to assign a price to
the water according to the expected electricity prices at later times. If the
water flow is too large it can be necessary to spill the water. If there are more
hydro power stations after each other the production must be coordinated so
that they use each others’ water.
There are also legal restrictions on maximum and minimum water flows
and levels in order to not disturb the ecology too much and not flood ar-
eas downstream. Hydro power generators are also very good at regulating
voltage.
2.3.2 Wind power
Different types of wind turbine generators have very different characteristics.

Since power system restoration needs to be done even if there is no wind,
restoration plans cannot depend on wind but when it is available it can be
used.
In systems with much wind power there will be less conventional produc-
tion ready to start (in a hot state) after a blackout when large wind production
is forecasted than otherwise.
Early wind turbines were stall regulated with directly connected induction
generators. These could not control active and reactive power, which make
them harder to use in a power system restoration situation. Newer wind tur-
bines often use doubly-fed induction generators (DFIG) controlled by power
electronics or synchronous generators with electronic frequency converters
which make it much easier to use them in a power system restoration situa-
tion. Modern wind turbines can be used for regulating both frequency and
voltage by controlling the power electronics; which can be done in the order
of milliseconds.
Since the marginal cost is very low it is very seldom economical to run
on less than their maximum output at the current wind speed, but this can
be done in a power system restoration situation. Most wind turbines nor-
mally start automatically about 10 minutes after they got stable voltage after
a blackout. Some wind turbines can use the power electronics to act as a
STATCOM1 even when there is no wind this can help regulating the voltage
during restoration.
2.3.3 Diesel engine
Diesel engines are mostly used as backup power, they are very fast to start
and regulate and also cheap to build but expensive to run. In sizes up to about
1 MW they can be mobile. They often have black start capability based on a
battery.
1
A power electronic device that can continuously control its generation or absorption of
reactive power.
2.3. Power plant characteristics 17
2.3.4 Gas turbine
Gas turbines are used for larger units than diesel engines and take 5-30 min-
utes to start. They are fast to control and are therefore often used as regulat-
ing power and emergency power. They can often be started in a black net by
using compressed air or battery. The marginal cost is high but they are cheap
to build.
2.3.5 Steam turbine
Steam turbines are used in most nuclear and coal fired power plants and pro-
duce about 80 % of the world’s electricity. They often have very long starting
times since it takes long time to heat the boiler. A steam turbine can be in
different states of readiness, cold (starting time 4-48 h), hot (starting time
30 minutes - 2 h), house load (starting time 5 minutes). The thermal iner-
tia makes steam turbines rather slow at regulating their active power output,
especially when the active power is increased.
Power plants based on steam turbines need rather much external power in
order to start.
When disconnected from the net due to faults in the net many units try to
continue supplying the internal power consumption (house load). In order
to achieve this, the power to the turbine must be decreased rapidly, which
is often done by letting steam bypass the turbine and go directly to the con-
denser. This is a sensitive process and in about 50% of the cases the unit will
trip when transition to house load is attempted [4].
A hot restart is only possible if the unit gets outside power within a speci-
fied time after the blackout.
Nuclear power plants also have to follow strict safety regulations and
sometimes need to get permission before starting. In the Swedish black-
out in 1983 the first nuclear generating unit was synchronized after 10 h and
all nuclear production was completely restored after about 80 h [1, pages
45-50].
Often a gas turbine is combined with a steam turbine in a combined cycle
power plant. In a combined cycle power plant the heat in the exhaust gases
from the gas turbine is used to produce steam for the steam turbine, this gives
higher overall efficiency.
2.3.6 Combined heat and power

Gas turbines, steam turbines and diesel engines generate large amount of
heat. Therefore it is often economical to use the heat for e.g. district heating.
In industrial processes it is common to use steam, often it is more economical
to produce it at higher pressure and temperature than needed and let it pass a
steam turbine in order to generate electric power before it is used.
These power plants are called CHP, combined heat and power. They are
often dispatched based on the heat demand and the electricity is a by-product.
They can often not run if there is no way to use up the heat.
Buildings without electricity cannot absorb the heat even if it is cold
weather. The pipe network and heat storage tanks in district heating can
often absorb some hours of heat production.
2.3.7 Properties of distributed generation

Small scale distributed generation could in theory start and respond faster
than large plants since small systems have faster thermal and mechanical
dynamics in general. But often it is too expensive to equip small power plants
with special control equipment for unusual situations such as restoration.
If the power plant has remote control it often will not work during power
outages and also the remote control will often only be available to the plant
owner and not the network owner or TSO. This can lead to a situation where
distributed generating units will go online faster than desired in a sensitive
stage of the restoration.
Also the large number of plants with different characteristics will make
the planning problem harder. This will require that they are aggregated in
the same way that loads are.
2.4 Historical development of power system restoration
Restoration was straightforward in the first power systems as they were built
around a power plant that of course had black start capability. Since then
the power systems have become more and more interconnected over larger
and larger areas. Many newer power plants are not built to be able to start
independently of external power.
The way the power system is operated has also changed. The first systems
2.5. Problem formulation 19
were to a large extent operated by direct manual control of the primary com-
ponents. Then the operation was moved to local control rooms and these
have been more and more aggregated. Now most national power systems
are operated by computerized remote control (SCADA) from one or a few
centralized control rooms.
This means that the number of people that operates the system has de-
creased. Even if the operation of the system has been heavily rationalized
there will be a very high work load on the operators in a restoration situ-
ation. It also means that the operators do not have the same level of local
knowledge that was common before. On the other side they will have much
better overview of the system.
In Sweden the operators of the local control rooms had prepared rules for
what to do in the event of a blackout. When these control rooms were closed
some of these rules have been implemented in automatic systems and this is
what we now call DUBA.
Another important change is the deregulation of the electric market in
large parts of the world. This has led to the legal and administrative sep-
aration of network and production. This makes it harder to coordinate the
operations in uncommon situations where the market cannot operate as in-
tended.
2.5 Problem formulation
The research problem in this thesis is to design an algorithm that generates

a complete sequence of actions to restore a power system after a collapse.
This will be done for a static power system model. The initial state is not
predetermined but a result of an unknown power system incident. The algo-
rithm should not rely on prepared plans or rules that are specific to the power
system or the restoration situation. This is important since the algorithm
is intended for comparison of different power system configurations from a
restoration perspective.
Chapter 3
Modeling and simulator
This chapter describes the models of the different system components and
how the power system is simulated. The objective of these models is that they
are easy to implement and computationally cheap while still giving results
that are relevant from a power system perspective. The model is used by the
restoration algorithm to test different actions and to get the resulting values
such as frequency, voltages and currents. The model also generates a list
of available actions in each state. The model contains checks to determine
if a state is feasible which means that all components operate within their
specifications.
3.1 Simulation
The power system is simulated with a power flow calculation after every
action in the search algorithm. Much computation is saved by avoiding dy-
namic simulations for every step. If this algorithm should be used for real
restorations the most promising actions should be checked with a dynamic
simulation, which has not been implemented yet. Only the positive sequence
is modelled; in other words, three-phase-symmetry is assumed. For a trans-
mission system without faults this should be a reasonable assumption. The
power flow calculation includes frequency calculation and limits on active
and reactive power at generating units. This is necessary in order to model
the production distribution during the restoration. The frequency calculation
is developed from the concept of distributed slack bus. The frequency calcu-
lation assumes that all power units can be assigned a frequency response in
21
22 Chapter 3. Modeling and simulator
Figure 3.1: An example that shows how the frequency control efforts of sev-
eral generating units are merged. All generating units in the dia-
gram have a maximum power of 300 MW and a minimum power
of 50 MW. Generating units 1 and 2 have a frequency response of
833 MW/Hz and power unit 3 has a frequency response of 1250
MW/Hz. Every Generating unit has a different set point for ac-
tive power. The set point is the power obtained at 50 Hz. The
frequency is calculated so that a balance exists between the pro-
duction and the consumption including the active power losses.
MW/Hz that applies over the entire control range. Figure 3.1 illustrates how
the MW/Hz characteristics of several generating units are combined.
The program also contains a complete topology calculation and handles
several electrical islands running at the same time by performing a separate
power flow calculation for each of them. Since only static power flows are
calculated no proper synchronization checks can be done. The only check
that is done when islands are synchronized is that the resulting power flow
should represent a feasible state. The impedances of the components are
independent of the deviation from nominal frequency. The power flow equa-
tions are solved with the Newton-Raphson method [23].
3.2. Power system components 23
Figure 3.2: The model of a line, a π-link, so named because of the similarity of
the diagram to the Greek letter π (pi). The model has series induc-
tance (L), series resistance (R) and capacitance (C) to earth. Since
three-phase symmetry is assumed, only one phase is modelled.
   ∂P ∂P ∂P   
∆P ∂θθ ∂V ∂f ∆θθ
 ∆Q  =  ∂Q ∂Q
(0...0)T  ∆V (3.1)
P ∂θθ ∂V
∆ θ 1...1 0...0 0 ∆f
Each power flow iteration solves the equation 3.1 to update angles by ∆θθ ,
voltages by ∆V and frequency by ∆f . ∆P and ∆QP are the mismatches in
active and reactive power at each bus. Specifying ∆ θ drives the average
of the bus voltage phase angels to zero, which is an alternative to selecting
a reference bus. This is easier to handle when the topology changes than
a reference bus. All power flow calculations are calculated from flat start.1
The number of iterations in the power flow calculation has been limited to
50 and the tolerance for convergence are 10 kW active power error and 10
kvar reactive power error in each node. Incremental power flow calcula-
tions would decrease the time to calculate a power flow but would increase
the memory requirements. Tripping constraints for protections etc. are not
modelled. The nominal frequency is assumed to be the same throughout the
power system. All connections and disconnections are made with explicitly
modelled switches.
3.2 Power system components
3.2.1 Transmission lines

Transmission lines are modelled with a π-link with series resistance and re-
actance and only capacitance to earth. The capacitance is assumed to be
symmetrically shared between the legs of the π-link; see Figure 3.2. The
transmission lines have a highest permitted current. Transmission lines in-
clude both overhead lines and cables.
3.2.2 Transformers
Transformers are modelled as an ideal transformer in series with a π-link
with series resistance and series reactance and with shunt reactance. Only
real transformer ratios are modelled (no phase-shifting transformers). Trans-
formers with more than two windings are not modelled. In this thesis no
tap changers have been used. Transformers have a highest permitted current.
It is assumed that the transformers can be energized from either side but in
practice energization from the low voltage side is often avoided due to high
inrush currents.
3.2.3 Shunt capacitors and reactors

Shunt capacitors and reactors are modelled as pure admittances. Any auto-
matic control is assumed to be replaced by the presented algorithm during
the restoration.
3.2.4 Loads
All the power consuming equipment connected to a power system is aggre-
gated into larger objects called loads in simulations and calculations. This is
done both since the net owner lacks information about details of the customer
power consuming equipment and to simplify the calculations. At transmis-
sion level the load objects also include subtransmission and distribution nets
and tap changing transformers. In reality the loads are the most complex part
1
This means that all voltages are set to 1 p.u. and all voltage phase angles is set to 0 before
the first iteration. The reason for this is that it is easier to handle for the search algorithm since
values from all earlier net states need not to be stored.
3.2. Power system components 25
of a power system. Transmission system operators normally do not recon-

nect loads themselves. Rather they tell the subtransmission system operators
how much load they can connect in different areas. The power consumption
after a blackout will be different from what it would have been if the blackout
had not occurred since some equipment takes time to start and other equip-
ment is controlled by thermostats and will consume more than normal. This
is called cold load pickup [9].
In order to model cold load pickup some form of dynamic simulation
would be needed. In this model loads are modelled only as constant active
and reactive power. The full amount of load at each point in the transmis-
sion system is divided into a number of small parts that are reconnected by
closing of switches.
3.2.5 Generating units
Generating units in the model can be on or off; black start units and plants
that survived the blackout are on in the initial state of the model. When a
generating unit that is off is energized from the net it will switch to the on-
state. Generating units that are on will not be turned off even if they are
disconnected. The model parameters and state variables can be seen in table
3.1 respectively 3.2. Active power is adjusted between Pmin and Pmax for
frequency control. Pset is the active power at nominal frequency. See Figure
3.1. Reactive power is adjusted between Qmin and Qmax by the automatic
voltage controller. Vset is the voltage with zero reactive power. Pset can be
changed between Pmin and Pmax in 10 steps. Vset can be changed in steps
of 0.01 p.u. as long as the reactive power is within its limits. When reactive
power is above Qmax , Vset cannot be increased and when reactive power is
below Qmin , Vset cannot be decreased. No changes can be made when the
generating unit is off.
3.2.6 Power system node
A power system node is any location where two or more components are
connected. All nodes have a highest and a lowest permitted voltage and
frequency. A base voltage has also been defined for each node.
Table 3.1: Parameters in the generating unit model

Pmax Highest possible active output power from the unit
Pmin Lowest possible active output power from the unit
Qmax Highest possible reactive output power from the unit
Qmin Lowest possible reactive output power from the unit
dP/df Frequency control characteristic, MW/Hz. Valid as long
as the active power is between Pmax and Pmin . This is
also known as the frequency response of the frequency con-
troller. See Figure 3.1.
dQ/dV Voltage control characteristic. Valid as long as the reactive
power is between Qmax and Qmin .
3.2.7 Switch
A switch can be placed between two nodes. Switches have two states: on
(closed) and off (open). Switches are modelled as capable of switching all
occurring currents on and off. Open switches can only be closed if one side
is energized; this simplifies the algorithm by avoiding to trying to close all
switches that are not affecting the current power flow. All closed switches
can be opened.
3.2.8 Faults
All components are assumed to be working properly. Any faulty components

are assumed to be removed from the model before the simulation is started.
3.2.9 Substations
Substations are defined only as those components that can be connected to

each other without impedance between them. This means that two groups of
switchgear with transformer/transformers between them are regarded as two
substations with different names. This has no relevance for the power flow
calculation but is used when generating statistics for a net state.
3.3. Action sequence and states 27
Table 3.2: Variables in the generating unit model

Pset Set point for active power, the power delivered by the
power unit at a nominal frequency. Pset must be between
Pmax and Pmin and can be adjusted during operation. The
algorithm makes the adjustments in steps of 10 percent of
the difference between Pmax and Pmin .
Vset Set point for the voltage on the connection node of the
power unit. This is the voltage that gives a reactive power
of 0 Mvar and can be adjusted during operation. The al-
gorithm makes the adjustments in steps of 1 percent of the
nominal voltage.
On Indicates whether or not the unit is in operation. Only is-
lands that contain operating generating units need power
flow calculation. Generating units that are or has been con-
nected to a “live” island are assumed to be in operation.
3.3 Action sequence and states
The starting point is a power system with a specific (operating) state. From
this state, the algorithm must find a sequence of actions that finally connects
all loads. All the intermediate states in this sequence must be feasible. There
may be several energized subsystems (islands). A feasible state is a state in
which the voltage and frequency at all energized nodes are within set limits
and the currents in all lines and transformers are below their thermal limits.
In order to simplify the algorithm only discrete actions are modelled, even
when changing continuous values such as active power set points. Available
actions are shown in Table 3.3.
Switches in non-energized parts of the power system are assumed to be off
(open) in their initial state. If automatic zero voltage protection is installed,
as it is in Sweden, this is a very reasonable assumption. Since only actions on
closed switches and energized components are considered, the search only
needs to cover parts of the power system that have been energized at some
time during the restoration. For a sequence of actions to be a valid solution
to the problem, all intermediate states must be feasible.
A state is described by:
Table 3.3: Actions available to the algorithm.

Type of Type of action When applicable Description
compo-
nent
Switch Switch_open_action If the switch is Open
closed the switch
Switch Switch_close_action If the switch is Close
open and at least the switch
one side of it is en-
ergized
Generating PV_inc_P10 If the generating Increase the
units unit is energized set point for
and Pset is not at active power
the upper limit at the gener-
ating unit by
10 % of the
control range
Generating PV_dec_P10 If the generating Decrease the
unit unit is energized set point for
and Pset is not at active power
the lower limit at the generat-
ing unit by 10
% of the con-
trol range
Generating PV_inc_V1 If the generating Increase the
unit unit is energized voltage set
and Q is not at the point by 1 %
upper limit
Generating PV_dec_V1 If the generating Decrease the
unit unit is energized voltage set
and Q is not at the point by 1 %
lower limit
3.4. Time 29
• Switch states
• Voltage set points of generating units.
• Active power set points of generating units.
• The set of generating units that are in the On state.
Since the load in this model is determined only by the switch states, no extra
state is needed for the load.
3.4 Time
In order to calculate the energy not delivered and plot results against restora-
tion time an estimate of the time for each action is needed. This is arranged
in the following approximate way. Increasing and decreasing active power
set point by 10% on thermal non black start generating units are estimated
to take 5 minutes. All other actions are estimated to take 1 minute. Since in
real situations several actions can be taken in parallel by different operators,
these estimated times are divided by 3 before they are used in the calcula-
tion. This is of course a very coarse model of the time consumption during
the restoration. A more refined model would be needed to explicitly model
the used parallelism and the time dependent behaviour of the components.
Chapter 4
Test system
Nordic32 shown in Figure 4.1 is selected as test system. Nordic32 is a fic-

titious power system model with 32 substations. It has been designed to
broadly represent the Nordel system. Nordic32 is chosen because it is a well
known test system so the data and results can be published freely and has
enough complexity to give interesting restoration plans without being too
complex. As a comparison this model is considered to be about one-third
the complexity of the real power system in western Denmark (Jutland and
Funen) down to and including 150 kV.
4.1 Adaptations and details
Nordic32 was originally proposed as a test power system by Svenska Kraft-

nät in a CIGRE report on Long term dynamics [7]. In that version it is
a dynamic net model without station topology, but Svenska Kraftnät later
added station topology, among other things, so that it could serve as an ex-
ample for the ARISTO real-time simulator [14]. The ARISTO version of
Nordic32 has been used in this project, with the following minor changes:
The substation topology has been simplified to include only switches that are
normally closed. Double busbar double breaker substations have been con-
verted to single busbar substations. These changes will make the algorithm
simpler. All the transformers have been fixed at the nominal tap ratio. The
transmission system loads have been divided into several smaller loads be-
tween 8 and 60 MW. Since the loads can only be connected or disconnected
it is necessary to be able to connect load in smaller steps. An extra 540 MW
31
32 Chapter 4. Test system
Figure 4.1: Nordic 32, a fictitious power system model with certain similari-
ties with the “Nordel” power system. The diagram shows the sta-
tions with their designations and all lines and transformers with
approximate fictive geographic positions.
4.2. Visualization 33
generating unit with black start capability has been added in the South1 . This
shall represent gas turbines and other peaking generating units that are not
included in Nordic 32. The model has 545 switches, 10 900 MW of load
and 36 generating units. There are 17 256 MW generating capacity of which
10 866 MW are hydro power in the North the rest are thermal power in the
South. Only the hydropower and the black start unit in the South are used
for frequency control. The voltage regulators have almost perfect voltage
regulation, the full reactive control range will be used for a voltage deviation
of only 0.01 %.
The complete model file for this system can be seen in appendix B.
4.2 Visualization
To be able to easily present an overview of interesting power flows, a graphic

representation of power flows has been developed. Given the coordinates of
the substations, transformers and lines the visualization of a power flow is
generated automatically. The visualization was developed to be suitable for
viewing in black and white such as in this thesis. The visualization has also
been of great value when running restoration searches interactively and when
debugging the program. Table 4.1 and Figure 4.2 describes how the various
parts of the visualization are interpreted. The normal state for the Nordic32
power system can be seen in Figure 4.3. In order to simplify the view the
connections inside a substation are not shown. The shown production and
consumption are the sum for the substation. Another example of a visualiza-
tion of power flows can be seen in the commercial power flow solver Power
World [32].
4.3 Test cases
The three different starting states shown in the Figures 4.4 - 4.6 are chosen
to test the algorithm under different conditions.
• Restoration from the North after a complete blackout. Only one gen-
1
This was done in order to be able to do black start in the South of the model. All
the existing generating units in the Southern part of Nordic32 has a minimum active power
generation and can therefore not be black started in the model. Another solution had been to
modify one of the existing generators, this has not been tested.
Table 4.1: Representation of different variables in the oneline schematics. The

graphical representation can be seen in 4.2.
Nr Variable Graphic representation
1 Active power flow Size of black arrowhead in the middle of line.
2 Reactive power flow Size of grey arrowhead in the middle of line.
3 Line or transformer cur- Thickness of the line is a constant plus a term
rent in p.u. proportional to the current.
4 Total load, active power Area of unfilled circle
5 Current load, active Area of filled circle
power
6 Total generating capac- Area of unfilled square
ity
7 Synchronized maximum Area of outer grey filled square
generating capacity
8 Current generation Area of black filled square
9 Synchronized minimum Area of inner grey filled square. This is zero
generation for hydro power.
10 Voltage Arrow in the half circle to the right, the arc
represents 1 p.u. and deviations are amplified
five times. When all parts of a substation are
not connected there are more than one arrow.
The relative voltage phase angles are directly
given by the angle of the arrows.
11 Maximum reactive sup- Length of upper part of the vertical grey line
port, capacitors and gen- to the right in each substation.
erators.
12 Current reactive support Length and direction of the vertical black line
or absorption, capaci- to the right in each substation.
tors, reactors and gener-
ators.
13 Maximum reactive ab- Length of lower part of the vertical grey line
sorption, reactors and to the right in each substation.
generators.
4.3. Test cases 35
Figure 4.2: Power flow visualisation elements. For interpretation see table 4.1.
erating unit in the equivalent of Northern Finland (CT72-G1) is black

started. This is an extreme case of the traditional restoration strategy
in the Swedish system with restoration from the North. It is intended
for testing of full restoration from one point.
• Restoration from the North and the South after a complete blackout.
The extra black starting generating unit in the South (FT63-G11) is
started in addition to the generating unit used in the previous example.
This is intended for testing of the algorithm in a situation with multiple
islands.
• Restoration after a partial blackout. This case is intended to be some-

what more realistic in that it is only a partial blackout. This case was
created by turning off the Southern part of the system and three gen-
erating units in CT72, otherwise it is identical to the normal state in
Figure 4.3.
4.3.1 Model parameters

The test system has a number of global parameters that affect the behaviour
of the system. Changing these will be used to test the sensitivity to changes
in different parameters. All the parameters have a default value that will be
used when not specified otherwise. These parameters are listed in table 4.2.
Figure 4.3: Power flow solution for the base case in Nordic 32. For interpre-
tation see table 4.1. 400 kV substations have names with 2 digits.
130 kV substations have names with three digits and the first are 1.
220 kV substations have names with three digits and the first are 2.
Lines between substations represent power lines if the substations
have the same voltage, otherwise they represent transformers.
4.3. Test cases 37
Figure 4.4: Initial state with one island in the North consisting of only the
substation CT72.
Figure 4.5: Initial state with one island in the North and one in the South with
the substations CT72 and FT63.
4.3. Test cases 39
Figure 4.6: Initial state with the Northern half of the system running, some-
what similar to the situation after the collapse 1983.
Table 4.2: Parameters of the test system
Parameter Description
Allowed fre- If the frequency in any part of the system deviates more than
quency deviation this amount from nominal frequency (50 Hz) the state is not
considered feasible. Default value is 0.5 Hz.
Highest allowed If the voltage in any part of the system exceeds this value the
voltage state is not considered feasible. Default value is 1.1 p.u.
Lowest allowed If the voltage in any part of the system is lower than this value
voltage the state is not considered feasible. Default value is 0.9 p.u.
Generation factor Maximum and minimum active generation is scaled by this
P value for all generators. Default value is 1.
Generation factor Maximum and minimum reactive generation is scaled by this
Q value for all generators. Default value is 1.
Generation Pinit The initial active power set point of the generating units as
fraction of Pmax − Pmin , Pinit = 0 ⇒ Pset = Pmin ,
Pinit = 1 ⇒ Pset = Pmax . If this does not match the
load there will be a frequency deviation. Default value is 0
Generation Vset Default initial voltage set point in per unit for the generating
units. Default value is 1
Load factor P All active load is multiplied by this factor.
Default value is 1
Load factor Q All reactive load is multiplied by this factor.
Default value is 1
Load step size All loads are divided into load steps of at maximum this value
factor multiplied by the default step size for the load.
Default value is 1
I limit factor All current limits on transmission lines and transformers are
multiplied by this value. Default value is 5. This almost
means that current limits are disregarded by default.
Load on factor The percentage of the loads on each busbar that will be con-
nected to that busbar from the start. This can be regarded as
loads without zero voltage protection. Default value is 0
Random lines This number of lines will be chosen randomly for removal.
missing This can be used for simulating random known line faults in
the system. Default value is 0.
Chapter 5
Characterization of the problem
The restoration problem is a planning problem1 . What characterizes such

problems is that, given an initial state, a sequence of actions is sought such
that a more desirable state is reached.
In the restoration problem there is no specific target state. The aim is to
arrive at any state that satisfies certain requirements, for instance that every
customer’s electricity supply has been restored and that the state must fulfil
certain requirements on the voltages and frequency.
The main difference between this problem and the planning problems,
normally studied in the field of artificial intelligence, is that the conditions
that govern which actions are permitted cannot be expressed as explicit log-
ical expressions. Instead a simulation (static or dynamic) needs to be done,
after which it is possible to check whether variables, such as voltages and
frequency, are within permitted limits.
Consequently, traditional planning algorithms such as “partial order plan-
ning” [35, pages 387-395] cannot be used. It is also difficult to use the
technology to search backwards from the target state to the initial state [35,
pages 384-386], since there is no fixed target state. For these reasons, a more
direct method is used, namely to search for a sequence of actions using the
initial state.
1
Planning problems are here used in a computer science context. In a power system
context it is common to differ between operational planning and expansion planning. The
difference is that the operational planning deals with how to run the system the coming hours
and days, while the expansion planning deals with how to add new components to meet future
demands over the coming years and decades.
41
42 Chapter 5. Characterization of the problem
5.1 NP-hardness
The power system restoration problem can be shown to be NP-hard. This

has been asserted in papers without proof or citations [21]. That the power
system restoration problem is NP-hard means that it belongs to a category of
problems for which it is believed that no efficient solution methods exists. A
well known example of a NP-hard problem is the travelling salesman prob-
lem.2 However, it has also not been shown that efficient solution methods do
not exist.
The expression efficient solution methods means methods whose calcula-
tion time is guaranteed not to increase asymptotically faster than a polyno-
mial in the size of the problem, for example expressed in number of compo-
nents n in the electric power system, as the size of the problem increases. For
example, if the computation time for a solution method grows with n1000 it is
considered an efficient solution method while if the computation time grows
with 2n is not considered to be an efficient solution method.
One consequence of this is that it is unrealistic to have as a goal an algo-
rithm that is guaranteed to solve the problem in reasonable time, especially
when searching for a solution that is in any sense optimal. It is therefore
necessary to focus on finding algorithms which, in as large a proportion of
realistic cases as possible, find solutions that are as good as possible in ac-
ceptable time.
Since the power system restoration problem is so complex, only a simpli-
fied version is used in the proofs below. This does in fact make the proofs
stronger since any reasonable power system restoration model can model
this simplified version. It is possible to define subsets of the power system
restoration problem that can be guaranteed to be solvable in polynomial time,
but it is hard to see how such a subset could include all or most of the non
trivial practical cases in any useful way3 .
2
The travelling salesman problem is the problem facing a travelling salesman that need to
visit each city in a set of cities and want to find the order he/she should visit them in order
to minimize the travelled distance. The distance between each pair of cities is assumed to be
known.
3
It is always possible to state that all practical power system restoration problems have
less than some large number of components and then the computation time can be shown to
be less than some very large constant but this gives no useful limit.
5.1. NP-hardness 43
Figure 5.1: Example of a Steiner tree, the nodes 3, 5, 8 is required to be con-

nected and the minimal Steiner tree are shown in bold.
5.1.1 Steiner tree in graphs

The normal way of proving that a problem X is NP-hard is by proving that
any instance of a known NP-hard problem could be solved in polynomial
time if X could be solved in polynomial time.
Since this would contradict the assumption that NP-hard problems can
not be solved in guaranteed polynomial time it is assumed that X can not be
solved in polynomial time.
Minimum Steiner tree in graphs is an NP-hard problem [10, page 395].
Given an arbitrary graph4 where each edge has an associated cost, the prob-
lem is to decide if there exists a subset of the edges and vertices that form a
tree5 that connects all vertices in a given subset of the vertices with a total
cost of the edges in the tree below a given value c. For an example see Figure
5.1.
Given an arbitrary instance of the Minimum Steiner tree in graphs problem
there are different ways to create an instance of the power system restoration
problem such that the solution of this problem gives the answer to the Mini-
mum Steiner tree in graphs problem.
Two such transformations will be given. In both cases the graph is directly
4
A graph in a computer science is a number of vertices connected by edges.
5
A tree in a computer science is a number of vertices connected by edges in such way that
there are no loops.
mapped to the power system, nodes and edges correspond directly to busbars
and power lines. There is only one generator in the system. The set of
nodes that need to be connected, correspond to the busbars with loads and
the generator busbar.
Reactive limits
In this case the associated cost of an edge is translated into the reactive
power generation of the corresponding power line and the maximum cost
c corresponds to the maximum reactive absorption of the generator. All line
impedances are assumed to be zero and all loads are assumed to be purely
active power. If all limits and conditions other than the reactive limit on the
generator are assumed to be set with large margins, then this system can be
restored if and only if there exists a Steiner tree with lower cost than c. This
proves that the power system restoration problem is NP-hard.
Since real power systems are built to be able to absorb the reactive power
from the lines this can seem to be unrealistic but in a real situation there can
be component failures. This could also be thought of as whether the load can
be restored before all generating units have been started.
Restoration time
In this case it is assumed that there are no stability problems in the system
so only the topology is interesting. Each power line has an associated time
needed to connect it6 . This time corresponds to the cost of the corresponding
edge.
If an optimal7 power system restoration sequence can be found, the cor-
responding minimum Steiner tree in graphs problem can be solved by com-
paring the restoration time with the maximum Steiner tree cost c. This again
proves that the power system restoration problem is NP-hard.
6
In the model used in this thesis where the cost of all switching operations is constant this
could be modelled by replacing the power line by a series of power lines that have switches
between them.
7
Optimal here means shortest possible restoration time.
5.1. NP-hardness 45
Discussion
There exist polynomial time algorithms that can calculate an approximation
of the minimum Steiner tree in graphs so that it is guaranteed to find a Steiner
tree that has a cost that is no more than 1.55 times the minimum cost [34].
Since the full power system restoration problem is much more complex
it is not likely that such approximation algorithms can be found for power
system restoration. The reason the full power system restoration problem is
more complex is that it need to take into account factors such as voltage drop
and the interaction between active and reactive power flows.
Chapter 6
Search algorithms
In order to find a sequence of actions that restores the power supply, a search
algorithm is used. The algorithm is designed so that, given an arbitrary state,
it attempts to find a sequence of actions that leads to a better state. Search
algorithms are an important part of computer science and artificial intelli-
gence.
A search algorithm is used to test different solution candidates to a prob-
lem in a systematic way. This is useful when it is easy to test if a proposed
solution is valid but it is hard to generate a valid solution directly. Since the
goal is to find a sequence of actions or in other words a path, a path finding
algorithm1 is needed. The environment in which the path exists is the state
space graph, where every possible state in the power system is represented
by a vertex in the graph and all actions in each state are represented by a di-
rected edge to the resulting state. This graph is enormously large; fortunately
it need not to be stored, only the relevant parts are generated when needed.
Every action has an associated cost; the algorithm tries to minimize the sum
of the cost of the actions in the action sequence. The search proceeds as a
tree where every choice in each node generates a new node; see Figure 6.1.
In this chapter the following terminology is used:
State A state in the state space graph, in the power system restoration this
means a state of the power system. Denoted by S.
1
The same algorithms can be used both when finding a physical path such as planning a
tour using public transportation and in more abstract state spaces such as in power system
restoration.
47
48 Chapter 6. Search algorithms
Figure 6.1: Example of a tree structure of nodes and actions, the arcs repre-
sents the actions. Examples of actions: closing a switch or increas-
ing the set point for active power production by 10 %.
Node A node in the search tree, each node has a unique path or sequence
of actions from the initial state to the state corresponding to this node.
Several nodes can represent the same state if there are several paths
to that state. Denoted by N . When referring to a node in the power
system this will be stated explicitly to avoid confusion.
Action An action can be applied to a state and this will lead to a new state.
Denoted by A.
Path A sequence of states and actions that goes from one state to another
state.
The search algorithm works much like a chess computer; the difference
is that in this case there is no opponent and because of this different search
algorithms are used.
Search algorithms can be categorized as uninformed and informed. Un-
informed search algorithms do not use any application specific information;
they only need to know what actions are available in each state and the cost
of the actions. Therefore, they need to test many paths until a path to a
goal is found. For any reasonably large power system restoration problem
this would be extremely slow. Informed search algorithms use some sort of
heuristics to find the most promising of the tested paths and extend them.
The heuristics are often in the form of an evaluation function that generates
6.1. The A∗ algorithm 49
a value for each path. This value indicates which path that is most promis-
ing. With a good evaluation function the search time can be reduced by
several orders of magnitude. Search algorithms can also be divided in online
algorithms and offline algorithms. Offline algorithms calculate the whole
sequence of actions from the start while online algorithms use information
collected during the execution of the actions to decide what to do next. On-
line algorithms are useful for handling uncertainty. Since no uncertainty is
modelled here an offline algorithm is used. For real power system restoration
an online algorithm would be needed. An online algorithm can be created by
rerunning an offline algorithm with the current state as the initial state during
the execution of the action sequence.
6.1 The A∗ algorithm
The algorithm used for the power system restoration in this project is based
on an informed search algorithm called A∗ [17]. In order to solve a path find-
ing problem with A∗ an initial state and four application specific functions
are needed. These are:
Available actions Generates a list of available actions in a given state.
Resulting State Calculates the resulting state after a given action is applied
in a given state. It also calculates the cost of the action.
Estimated cost to reach the goal from the current state, h(S) Returns an
estimate of the cost for the shortest path from the state S to a goal state.
Can also be referred to as h(N ) where N is a node in the search tree
that corresponds to the state S.
Is goal Test if a given state is a goal state, in some cases this is equivalent to
h(S) = 0.
The sum of costs for all actions in the path to the node N is denoted g(N ).
A∗ uses f (N ) = g(N ) + h(N ) to guide the search, f (N ) is an estimate of
the cost of the shortest path from the initial state to a goal state such that the
path begins with the path to N .
A∗ begins with one set of nodes called OPEN that only contains the node
for the initial state and one empty set called CLOSED. The OPEN set will
contain nodes that have untested actions or actions that need to be re-evaluated.
The CLOSED set will contain nodes where all actions have been evaluated.
The following steps are repeated until a path to a goal is found or the OPEN
set is empty which indicates that no path from the initial state to a goal exists:
1. Select the node with the lowest value of f (N ) from the OPEN set, this
node will be called Nn and the corresponding state Sn . If the OPEN
set is empty then there does not exists any path from the initial state to
any goal.
2. If Sn is a goal state then exit and return the path to Nn .
3. For each available action in node Nn do:
(a) Calculate the resulting state after applying the action. This new
state is represented by a new node in the search tree.2
(b) If a node with the same state as the new node is not already in
the OPEN or CLOSED set then add it to the OPEN set.
(c) If a node with the same state as the new node is found in the
OPEN or CLOSED set and the new path gives a lower cost than
the previous then remove the old node and add the new node to
the OPEN set.
4. Move the state Sn from the OPEN set to the CLOSED set.
5. Go to step 1.
Generally an h(S) function that gives larger values will result in a nar-
rower faster search. If h(S) is exactly the cost of reaching a goal given any
state S and all values of f (N ) are unique only states on the optimal path will
be expanded.
6.2 Optimality
It can be shown that A∗ will find the shortest path to a goal if the heuristic
function h(S) never overestimates the cost of the cheapest path to the goal
2
If a node with the same state has previously been expanded then this could be replaced
by looking up the resulting state in a hash table in order to save calculations.
6.3. Example 51
[17]. This is called an admissible heuristic. It can also be shown that if

h(S) is consistent3 , then no algorithm that is guaranteed to find the shortest
path can expand fewer nodes if it is not given any extra application specific
information [17]. One problem with A∗ is that the whole search tree needs
to be stored; in many applications this requires too much memory. In order
to handle this there are modified variants of A∗ such as SM A∗ that works
like A∗ until the memory is full and then starts deleting the worst leaf nodes,
if needed these can be regenerated later so the algorithm will always4 return
the same path as A∗ but can save memory at the cost of more computation.
6.3 Example
The small state space graph shown in Figure 6.2 will be used to illustrate
how A∗ works. The heuristic in Figure 6.2 is an example of an admissible
heuristic. The search tree generated by A∗ for this graph can be seen in
Figure 6.3. The arrows represent actions and the numbers on the arrows
represent the cost of the action. Nodes that represent the same states but
are on different paths are connected by a dotted line. The name of the state
is given at the upper row in each node. On the lower row are three fields
separated by /, these are:
1. Node number or number of nodes found before this node during the
search.
2. The value of f (N ) = g(N ) + h(N ).
3. The status of the node when the search is finished. The abbreviations
are:
O The node is in the OPEN set.

C The node is in the CLOSED set.
W The node was worse than an earlier found node in the OPEN or
CLOSED sets for the same state.
RO The node was removed from the OPEN set when a better path to
the same state was found. (Step 3c in the algorithm)
3
This means that f (N ) never decreases along any path.
4
Given of course that the path to be found fits in the given memory.
Figure 6.2: Example of a state space graph, the arrows represent actions and
the numbers on the arrows represent the cost of the action. The
name of the state and the value of h(S), the estimated cost of
reaching a goal from that node, are given in each state. The short-
est path from start to a goal is Start – A – B – Goal1 and has a cost
of 11.
RC The node was removed from the CLOSED set when a better path
to the same state was found. (Step 3c in the algorithm)
The table 6.1 shows the progress of the search for each iteration. The
algorithm does not need to visit the states E and F since they cannot be on
an optimal path unless h(C) is an overestimate5 . Three paths to Goal1 are
found and one to Goal2, the search continues until the optimal path is found.
It finds a path back to the start state (Start -> A -> Start) but correctly detects
that it has already a better path (The empty path) and stops searching that
path. All the directly following actions are always tested in sequence. The
cost of going from A to B is 3, this is less than h(A) − h(B) = 5, this means
that h(S) is not consistent and this is the reason that the suboptimal path to
B is expanded before A on the optimal path. When the new path to the state
B is found in node 7 the “children” Goal1 and C get updated so g(N ) and
f (N ) take the new path in to account.
5
The cost of reaching C is 6 and the estimated cost of reaching a goal from C is 10 so the
estimated cost of reaching a goal through any path that includes C is 16 while the cheapest
found path costs 11.
6.3. Example 53
Figure 6.3: Example of a search tree generated by A∗ for the graph in Figure
6.2. The arrows represent actions and the numbers on the arrows
represent the cost of the action. Nodes that represent the same
states but are on different paths are connected by a doted line.
Table 6.1: The content of the OPEN and CLOSED set in each step in the iter-
ation. Each node is represented by the name of the state, the value
of f (N ) and the value of g(N )
OPEN set CLOSED set
(Start, 6, 0)
(B, 10, 6) (A, 11, 2) (Goal1, 13, 13) (Start, 6, 0)
(A, 11, 2) (Goal1, 12, 12) (C, 17, 7) (Start, 6, 0) (B, 10, 6)
(B, 9, 5) (Goal2, 12, 12) (Goal1, 12, (Start, 6, 0) (A, 11, 2)
12) (C, 17, 7)
(Goal1, 11, 11) (Goal2, 12, 12) (C, 16, (Start, 6, 0) (A, 11, 2) (B,
6) 9, 5)
6.4 Power system restoration
Since the power system restoration problem requires so much computation

in order to evaluate each node, the limiting factor is the computation time,
not memory so in this application SM A∗ is not needed.
No heuristic function that give estimates high enough to limit the num-
ber of expanded nodes to a reasonable number while still being admissible
is known. Therefore a non admissible heuristic is used and it can not be
guaranteed that an optimal path will be found.
An example of an admissible heuristic function is the total cost of closing
a switch to every load that is not connected to anything. This could be im-
proved by adding the cost of connecting a minimum set of generating units
and the cost of increasing the active power set point. But such heuristics
will not give any guidance in how to avoid voltage collapses and limits in
e.g. voltage. Therefore such problems will be handled only by trial and error
which will make the progress very slow.
In most states there are several hundred actions, this makes it too expen-
sive to do a full power flow calculation in order to evaluate the result of all
actions in each expanded node. Therefore the evaluation is divided into two
steps. When a new node is reached, the node in itself is evaluated by an
evaluation function f2 (N ) and all the actions of that node are evaluated by
a very fast and simple function f 1(N, A), where N is the node and A is the
action. When selecting a node to expand, the node with the lowest value of
f2 (N ) is selected. Rather than calculating the result of each action, only one
action is tested. The action to test is chosen randomly6 from the best actions
according to f1 (N, A).
By testing it was found that choosing the action from the best 3 % of the
actions gives good results but the exact value is not critical.
The resulting node of this action is evaluated together with its actions
as described above. The evaluation function f2 (N ) includes calculating a
power flow for the state. This means that for each iteration in the algorithm
only one action is tested so the parent node must remain in the open set until
it has no untested actions.
If the search continues until the node with the lowest value of f2 (N ) cor-
6
The random choices are of course not truly random but are the result from a pseudo ran-
dom number generator. The Mersenne Twister algorithm is used as pseudo random number
generator.
6.4. Power system restoration 55
responds to a goal state it will take too long time so instead the search stops
as soon as a goal state is found.
Since there is no guarantee that there exists a path from the initial state to
a goal or that it will be found within reasonable time even if it exists, a limit
on the number of nodes to evaluate is needed. Since little improvement is
seen in the test runs after about 2000 nodes we set the maximum number of
nodes to 3000 after which the search will be aborted. The maximum number
of nodes to evaluate depends of course on the power system that shall be
restored.
The reason for choosing the action to test randomly from the most promis-
ing is that if no path to a goal is found in a search then a new search can be
run with a new and hopefully better result.
The high number of states and available actions makes the probability of
finding multiple paths to identical states low. Therefore the check for iden-
tical states in step 3 in the algorithm description has not been implemented.
In step 3 the resulting state is just added to the OPEN set. This saves some
computation but could on the other hand lead to recalculation of already cal-
culated states in rare cases. If the power system model is extended with
time dependent models of loads and power plants, the probability of finding
multiple paths to identical states will become even lower since the timing of
actions will affect the state.
Figure 6.4 and table 6.2 shows the search for a path in the example graph
in Figure 6.2 without checking for duplicate states.
The evaluation functions have a number of parameters that must be given
a value. The parameters are denoted pn where n is an integer.
In order to have some starting point when studying variations in these pa-
rameters, in chapter 8, all parameters are given a default value. These values
have been found by estimating the size of each term and the relative impor-
tance of the term. The chosen values were then refined by testing variations
around these values. The default parameters are chosen to contain only one
digit out of 1, 2, 3 or 5 in different decimal position or to be zero in order to
show that the exact value are irrelevant.
Figure 6.4: Identical to Figure 6.3 except that the check for duplicated states
has been removed. When the start state is reached in node 6 parts
of the search tree are duplicated.
6.4. Power system restoration 57
Table 6.2: The content of the OPEN and CLOSED set in each step in the it-
eration of A∗ with the check for duplicated states removed. Each
node is represented by the name of the state, the value of f (N ) and
the value of g(N ). Some states are represented by several paths but
only the cheapest path is of interest.
OPEN set CLOSED set
(Start, 6, 0)
(B, 10, 6) (A, 11, 2) (Goal1, 13, 13) (Start, 6, 0)
(A, 11, 2) (Goal1, 12, 12) (Goal1, 13, (Start, 6, 0) (B, 10, 6)
13) (C, 17, 7)
(B, 9, 5) (Start, 10, 4) (Goal2, 12, 12) (Start, 6, 0) (B, 10, 6) (A,
(Goal1, 12, 12) (Goal1, 13, 13) (C, 17, 11, 2)
7)
(Start, 10, 4) (Goal1, 11, 11) (Goal1, (Start, 6, 0) (B, 10, 6) (A,
12, 12) (Goal2, 12, 12) (Goal1, 13, 13) 11, 2) (B, 9, 5)
(C, 16, 6) (C, 17, 7)
(Goal1, 11, 11) (Goal1, 12, 12) (Goal2, (Start, 6, 0) (B, 10, 6) (A,
12, 12) (Goal1, 13, 13) (B, 14, 10) (A, 11, 2) (B, 9, 5) (Start, 10,
15, 6) (C, 16, 6) (C, 17, 7) (Goal1, 17, 4)
17)
6.5 The fast and simple evaluation function, f1 (N, A)
f1 (N, A) is mainly used to avoid repeatedly testing bad actions in all nodes.
The function f1 (N, A) is shown in equation 6.1, lower values mean better
action.
f1 (S, A) = p1 nb − p2 ng + p3 r (6.1)
The function has three terms:
p1 nb nb is the number of times the action has been tested on other nodes in
the search tree before reaching the current node and the resulting node
had worse evaluation than the parent node. This statistic is collected
for each individual action e.g. closing a specific switch. This is in-
tended to avoid actions that have been bad in other nodes. The default
value of p1 is 1.
p2 ng ng is the number of times the action has been tested on other nodes
in the search tree before reaching the current node and the resulting
node had better evaluation than the parent node. This is intended to
prioritize actions that have been good in other nodes, this is most rel-
evant for actions that can be applied several times e.g. increasing or
decreasing set points. The default value of p2 is 0.
p3 r r is a random number between 0 and 1. This is intended to mix actions
that otherwise should have had the same evaluation. Since the other
terms are small integers it is likely that many actions otherwise would
have received the same evaluation. The default value of p3 is 0.01.
Theoretically any value strictly between 0 and 1 would give exactly
the same result.
If the action opens a switch and does not only disconnect a reactor or
capacitor then an additional value p4 is added to the evaluation function.
The default value of p4 is 10. The intention with this parameter is to give
lower priority to actions that open switches since the number of switches
that needs to be opened during a restoration is much lower than the number
of other actions.
This evaluation function could be improved in many ways, for example
the prioritization when connecting and disconnecting capacitors and reactors
could depend on the voltage in the power system node in the parent state.
6.6. The full evaluation function, f2 (N ) 59
The use of statistics from use of actions in other states gives the algorithm
a limited form of learning. Observe that the function is not recalculated for
actions in a given node when ng and nb changes.
Since actions that have been available in many nodes generally collect a
number of bad usages, newly available actions such as closing switches on
newly energized busbars will be given higher priority.
6.6 The full evaluation function, f2 (N )
The full evaluation function is calculated after the topology and power flow
of the new state is calculated. Therefore there is much more information
available in order to give a better evaluation. This function is used to select
the node to expand next. The function can be seen in eq. 6.2.
f2 (N ) = g(N ) + p5 h(S) (6.2)
S is the state corresponding to the node N . f2 (N ) has two terms:
g(N ) is the number of steps in the path from the initial state to the node N .
p5 h(S) is an estimate of the remaining cost to a goal (a state were all loads
are supplied). p5 is used to adjust how optimistic the prediction will
be, the default value is 2.
The function h(S) is shown in eq. 6.3.
Pload Pgen max − Pgen min

h(S) = p6 nopen switches − p7 − p8 +
Ptotal load Ptotal load
nlive
X nodes nlive
X nodes
(6.3)
(fi − fbase )2 (|Vi | − 1)2
i=1 i=1
+p9 + p10
nplive
11
nodes nplive
11
nodes
h(S) has five terms:

p6 nopen switches : This term gives priority to states where many switches are
closed. nopen switches is the number of open switches in the system.
Since only normally closed switches and switches to shunt capacitors
and reactors are modelled most switches will be closed in a goal state.
The default value of p6 is 1.
Pload
−p7 Ptotal load
: This term gives priority to states that already supply a large
fraction of the load. Pload is the active power of the currently con-
nected loads. Ptotal load is the total amount of load in the system. The
default value of p7 is 500.
Pgen max −Pgen min
−p8 Ptotal load : This term gives priority to states that have much reg-
ulating power. Pgen max is the sum of the maximum active power of
all the energized generating units. Pgen min is the sum of the minimum
active power of all the energized generating units. The default value
of p8 is 50.
nlive
X nodes
(fi − fbase )2
i=1
p9 p 11
nlive
: This term gives priority to states where the fre-
nodes
quency in the power system nodes are near the base frequency 7 This
is the sum of the squared deviation in frequency normalized according
to the total number of energized power system nodes. The normaliza-
tion is done by dividing by nplive
11
nodes , the default value of p11 is 0.5.
This is a compromise between the sum of squared deviations (p11 = 0)
and the mean of squared deviations (p11 = 1). All power system nodes
in the same island will have the same frequency so the frequency of
a large island will have greater impact than the frequency of a small
island. The default value of p9 is 5.
nlive
X nodes
(|Vi | − 1)2
i=1
p10 p11
nlive
: This term is similar to the previous but gives pri-
nodes
ority to states with voltages near 1 p.u. instead of states with good
frequency. The same normalization exponent p11 is used. The default
value of p10 is 500. This value needs to be larger since voltages are
calculated in p.u. while frequencies are calculated in Hz and therefore
can have larger numerical deviations.
For states with violated limits in voltage, frequency or current a large value
is added to h(S) in order to guarantee that it has worse evaluation than any
state without violated limits.
7
fbase = 50Hz in the test system.
6.7. Summary 61
6.7 Summary
The power system restoration problem is a path finding problem in the state
space graph. A power system restoration algorithm based on the A∗ -algorithm
has been described. A∗ has a number of good theoretical properties if the
evaluation function fulfils some constraints.
No good evaluation function that fulfils these constraints has been found
for the power system restoration problem so the algorithm can not be guar-
anteed to find a path if one exists or to find the shortest path. According to
section 6.1 it is very likely that every evaluation function that fulfils these
constraints will give exponential computation time in some cases and thus
no useful guarantees could be made anyhow. If this is not true an efficient
algorithm to solve all NP-complete problems would exist.
The algorithm will hopefully find a good path in most of the cases.
The proposed algorithm has two evaluation functions. f1 (N, A) is a very
simple and fast evaluation function to compare different actions in the same
search tree node. It is based on statistics for the same action in other search
tree nodes. f2 (N ) calculates a power flow for the state corresponding to
search tree node N in order to compare it with other nodes, therefore it takes
longer time to calculate.
The algorithm is randomized so if it fails to find a good solution a new
attempt at running the algorithm can be made with a new result.
Chapter 7
Simulation results
The algorithm has been tested with the three cases presented for the test
system. Each of these has been tested with six different initiations of the
pseudo random generator. For each of the test cases one of these six test runs
has been selected for closer examination1 .
7.1 Resulting sequence
The state of the system at different stages during the restoration from both
the North and the South for the selected test run can be seen in the Figures
7.1, 7.2, 7.3, 7.4 and 7.5.
The Southern island grows much faster than the Northern island. The
lines from CT71 to CT11 and CT12 are energized very late which prevents
the Northern island from growing. Larger islands have larger number of
available actions and therefore the algorithm will test more actions on large
islands than on smaller. This will result in that large islands grow faster than
small. In reality the longer starting times for thermal power plants would
have slowed down the Southern island so the Northern island had grown
more before synchronization. The reason that the synchronization happens
so late is probably that the algorithm does no special handling of synchro-
nization and it is hard to find net states in which synchronization is immedi-
1
The random seed for the selected test run for start in the North and for start after a partial
blackout are 0. This initiation gave atypical behaviour for start from both the North and the
South at the same time so 1 was chosen instead. The initiation for the other 5 test runs was
10,11,12,13 and 14 for all three cases.
63
64 Chapter 7. Simulation results
Figure 7.1: The state of the power system after 10 % of the actions in the
restoration sequence when the restoration is started both from the
North and the South. Five substations are energized and some load
and generation has been connected.
7.1. Resulting sequence 65
North and the South. The Southern island has grown to 19 substa-
tions while the Northern island is just two substations. Large parts
of the system are energized but very lightly loaded.
North and the South. 28 substations are energized. The voltage
angle differences in the system have increased. The islands have
met in substation CT71 but are not yet synchronized.
North and the South. All substations and generating units are en-
ergized. Some power lines are not energized or only connected in
one end. The islands are still separated in substation CT71.
Figure 7.5: The state of the power system after the complete restoration se-
quence when the restoration is started both from the North and the
South, compare with the normal state in Figure 4.3.
Figure 7.6: Number of actions taken as function of the simulated restoration

time. The simulated restoration time are calculated according to
section 3.4. The black lines represent the selected test run for each
test case and the grey lines represent the other test runs.
ately advantageous.
The total number of actions taken during the restoration sequence can be
seen in Figure 7.6. The graphs for restoration from the North and the South
and from the North are very similar and the restoration from a partial black-
out is similar to the later half of the other graphs. Two test runs for the partial
blackout that finds very few actions and then get stuck (Not visible in Figure
7.6). The sequences starts with almost the maximum 180 actions per hour,
during the latest 3 hour of the sequence the number of actions falls down
to about 100 actions per hour due to more adjustments of active power set
points is needed and takes longer time. This time estimate is of course very
rough and does not take into account for example that the number of actions
that can be taken in parallel increases with the size of the energized system.
The total available generation never limits the restoration in this case as
can be seen in Figure 7.7. In a real restoration it is possible that the rate of
load restoration would increase with the amount of synchronized generation
since larger load steps can be connected and the active power can be ramped
up faster. Since these aspects are not in the model it cannot be seen here.
Figure 7.7: Upper line, connected active generating capacity. Lower line, con-
nected active load. Both as function of the simulated time for start
from both the North and the South at the same time. The black
lines are for the selected test run and the grey are for the other test
runs.
Figure 7.8: The frequency’s in the system as function of the simulated time for
start from both the North and the South. The black lines show the
selected test run and the grey lines show the other five test runs.
The frequencies in the system during the restoration from both the North
and the South can be seen in Figure 7.8. In the selected test run the synchro-
nization of the islands can be seen after ca. 4.5 hours. Frequencies between
49.5 and 50.5 are allowed, the frequencies are mostly in the lower part of
that band. One reason for this is that there are many more actions that lower
the frequency than raise it so the frequency will tend to be adjusted down
when too high faster than it is adjusted up when too low. One other reason
is that the evaluation function has terms that prioritize states with much load
connected and many switches closed.
The highest and lowest voltage in the system during the restoration from
both the North and the South can be seen in Figure 7.9. It also shows the
voltage after actions that have been tested but not used in the selected test
run. Voltages between 0.90 and 1.10 p.u. are allowed. There are some
tested actions that give too low voltages after about one hour, otherwise most
voltage related difficulties seem to be with too high voltages, probably due
to attempts to connect long lines.
The maximum voltage angle difference for the island with the largest an-
gle difference can be seen in Figure 7.10. After about one hour there are
Figure 7.9: Highest and lowest voltage as function of the simulated time for
start from both the North and the South. The black lines show the
selected test run and the grey lines show the other five test runs.
Doted lines represent actions tested in the selected test run but not
used in the final sequence.
7.2. The search 73
Figure 7.10: Maximum voltage angle difference in any of the islands as func-
tion of the simulated time for start from both the North and the
South. The black lines show the selected test run and the grey
lines show the other five test runs. Doted lines represent actions
tested in the first test run but not used in the final sequence.
some actions that result in large increases in the maximum voltage angle
difference, this indicates some stability problems and is probably related to
the low voltages at the same time in Figure 7.9. The synchronization of the
island cannot be seen since none of the nodes in the Northern island nei-
ther get the highest or the lowest angles. The large differences in maximum
voltage angle difference between the different test runs are probable because
the voltage angles are not directly used in the evaluation function so it only
effects the evaluation indirectly. At the end of the restoration all the test runs
converge between 65 and 95 degrees.
7.2 The search
The percentage of the test runs that succeeds in restoring the entire load and
the average not restored load can be seen in table 7.1. The average restora-
tion time for the different test cases can be seen in table 7.2. This is calcu-
lated as energy not delivered divided by the total load in the system. This
Table 7.1: Restoration results for test cases with base parameters. Average
over 50 test runs.
Test case Complete Average load
restorations [%] not restored [%]
Restoration from the 86 0.13
North and the South
Restoration from the 92 1.7
North
Restoration after 32 22.0
partial blackout
Table 7.2: Average restoration times for test cases with base parameters. Av-
erage over 50 test runs.
Test case During Compensated
sequence [h] for not restored
load [h]
Restoration from the 2.83 2.85
North and the South
Restoration from the 2.91 3.07
North
Restoration after 1.25 3.45
partial blackout
is similar to the reliable index ASIDI [5]. If not all load is restored in the
found restoration sequence it is assumed that the remaining load will be re-
connected manually 10 hours after the end of the sequence when calculating
the energy not delivered. The best found complete restoration sequences for
the three test cases with respect to average restoration time have an average
restoration time of 2.53 hours for start from both the North and the South,
2.58 hours for start from the North only and 1.45 hours after a partial black-
out.
In Figure 7.11 the number of found actions that will be part of the final
sequence are plotted against the number of tested actions. For the total black-
7.2. The search 75
Figure 7.11: Length of sequence as function of number of investigated states.

The black lines represent the selected test run for each test case
and the grey lines represent the other tests.
outs about 50 % of the first 1200 actions tested are used in the final sequence,
then the search becomes more difficult. The test runs for restoration from a
partial blackout are much more diverse and only about 30 % of the tested
actions is used in the final sequence. In Figure 7.12 a small part of a search
tree can be seen in detail. The algorithm very seldom backs up and tests
actions on other branches of the search tree but it has been observed in some
test runs. The reason for this is that there are so many actions to test on each
state.
The value of the evaluation function f2 (N ) for the search tree nodes along
the found sequence (path) can be seen in Figure 7.13. This value should
represent an estimate of the number of steps in the shortest sequence that
restores all loads and that begins with the current action sequence. In this
context it can seems strange to have negative values but it does not mater
since these values are only compared to each other. As stated in the algorithm
chapter this heuristic function is not admissible, this can be seen by the fact
that the value of f2 (N ) is higher for states along the path than for the found
goal state. The algorithm has very hard to find sequences that pass states
with higher evaluation than earlier states due to the high number of available
Figure 7.12: Length of sequence as function of number of investigated states

for start in both the North and the South. Zoomed in at the be-
ginning of the sequence. Grey lines represent actions that are not
used in the final sequence.
actions that need to be tested.
7.3 Summary
The algorithm has been tested with the base parameter set presented in the
algorithm chapter and found to work well. The algorithm has some diffi-
culties with the partial blackout; this is probably because the algorithm has
difficulties redistributing the power production. Another reason for this can
be that the search parameters were mainly selected by trial and error on the
test case with start from both the North and the South. The algorithm does
very little backtracking. If no better state is found all the actions in the search
tree node with best evaluation will be tested before actions is tested in other
states. This takes long time so if the algorithm reaches a dead end the search
will be very slow.
7.3. Summary 77
Figure 7.13: The value of the evaluation function f2 (N ) for the search tree
nodes along the found sequence (path). The black lines represent
the selected test run for each test case and the grey lines represent
the other tests.
Chapter 8
Algorithm parameter sensitivity

analysis
It is important that the need for tuning the search parameters is as low as
possible and that one set of search parameters should work well in many
different cases. Therefore a sensitivity analysis on the search parameters has
been done.
One parameter at a time has been changed; the default values have been
used for the other parameters. For each investigated value of each parameter
50 test runs have been performed. The average restoration time as defined in
7.2 has been used in order to compare how good different parameter values
are.
8.1 The fast and simple evaluation function, f1 (N, A)
Equation 6.1:
f1 (S, A) = p1 nb − p2 ng + p3 r
The most important parameter in f1 (N, A) is p1 that reduces the priority

of actions that are tested in other states with bad result. If this is set to 0 then
14% and 25% of the load will not be restored for start from the North and the
South and start from only the North respectively. This is probably because
the search needs more than the maximum 3000 states due to a higher number
of bad actions. As long as p1 is greater than the randomization parameter p3 ,
79
80 Chapter 8. Algorithm parameter sensitivity analysis
Figure 8.1: Average restoration time as a function of p4 the correction added

to the evaluation of actions that opens switches other than switches
to shunt reactors and shunt capacitors. The default value is 10. The
lines with circles represent restoration from the North and the lines
with x:s represents restoration from both the North and the South.
the value of p1 and p3 do not to make any difference. The average restoration
time becomes about 6 minutes longer if p3 is 0.
p2 is used to increase the priority of actions that have been tested in other
states and given good results. p2 makes very little difference, probably be-
cause most actions are only useful once in each sequence and the search tree
is very dominated by one sequence.
The parameter p4 reduces the priority of actions that open switches other
than switches to shunt reactors and shunt capacitors. The impact of different
values of the parameter p4 can be seen in Figure 8.1.
All figures in this chapter will have the same lines but with different pa-
rameters on the horizontal axis.
The total restoration time including the assumed manual restoration of the
remaining load in 10 h after the end of the automatic restoration sequence is
shown. Below this line is a line that shows the part of the average restoration
time that is due to the manual reconnection of the remaining load. The sensi-
tivity to changes in the parameters has been tested both against the test case
with start from the North and the test case with start from both the North and
the South. As can be seen lower values than about 4 gives less successful
restoration. Values above that seem to give no difference, probably could
these actions be removed totally since they are not needed in most cases.
The default value is 10.
The fraction of the available untested actions that the action to test is cho-
sen from seems to make no difference as long as it is below 5% but it is
important to allow for good randomization so the algorithm can be rerun on
the same case until a good restoration sequence is found.
8.2 The full evaluation function, f2 (N )
Equation 6.2 and 6.3:
f2 (N ) = g(N ) + p5 h(S)
Pload Pgen max − Pgen min

h(S) = p6 nopen switches − p7 − p8 +
Ptotal load Ptotal load
nlive
X nodes nlive
X nodes
(fi − fbase )2 (|Vi | − 1)2

i=1 i=1
+p9 + p10
nplive
11
nodes nplive
11
nodes
The full evaluation function, f2 (N ) consists of two parts, the number of

steps to the current state and the estimate of the remaining number of steps
to a goal. The estimate can be scaled with p5 to adjust how optimistic it
is over all, this being equivalent with scaling all the parameters p6 through
p10 . The average restoration time for different scaling of the estimate can
be seen in Figure 8.2. The test runs were done by scaling the real cost so
far in the sequence, g(N ). From a conceptual perspective it makes more
sense to scale the estimated cost since the real cost must be assumed to be
known. Since the values of f2 (N ) are only compared with other values of
the same function this makes no difference so the equivalent values for p5
are shown on the horizontal axis. If the estimate is too optimistic about
the remaining cost then most tested actions will look bad so a large number
of actions are tested before one that improves the evaluation is found. If
Figure 8.2: Average restoration time as a function of p5 , that scales the es-
timated remaining number of steps (cost). Observe the scale on
the horizontal axis. The default value of p5 is 2. The lines with
circles represent restoration from the North and the lines with x:s
represents restoration from both the North and the South.
the value of p5 is lower than about 1.2 then the algorithm does not find a
complete restoration sequence before the maximum number of states, 3000
is reached. If the estimate is too pessimistic more actions will look good so
in reality the found actions will not go as straight towards the goal; this is
seen as the faster restoration at 1.43 than at ∞. The default value of p5 is 2.
The parameter p6 is used to prioritize states with many closed switches,
the effect of this parameter can be seen in Figure 8.3. If this is below 0.5 the
action to close a switch with no other impact on the state such as energizing
an unconnected bus bar will look bad to the algorithm. Since a number of
such steps needs to be taken the algorithm will not find a complete restora-
tion sequence before it reaches the maximum number of tested actions. The
location of this transition is strongly dependent on the value of p5 , the tran-
sition in Figure 8.2 is to some extent a mirror image of this transition. A too
high value of p6 generates somewhat worse restoration sequences since there
will be too much focus on closing switches.
The evaluation of connected load, p7 is rather uncritical as can be seen
Figure 8.3: Average restoration time as a function of the parameter p6 scaling

the evaluation of open switches. The default value of p6 is 1. The

the evaluation of connected load. The default value is 500. The

the evaluation of frequency deviations. The default value is 5. The
in Figure 8.4. This is probably because p6 has a similar function. A value

between 200 and 500 seems to give the best results. The default value is 500.
The evaluation of regulating power range p8 makes no difference and can
be omitted, this is probably because the regulating range of the generating
units is never an issue and it is almost always good to start all available
generating units as soon as possible anyhow due to the voltage and frequency
regulation.
As can be seen in Figure 8.5 the amount of automatically restored load
decreases drastically if the prioritization of frequency deviation in the eval-
uation are below about 1, this is probably because then it does not seem
advantageous to increase the active power set point of generating units even
if the frequency is low. The default value is 5.
The impact of different prioritizing of the deviations in voltages can be
seen in Figure 8.6. The default value is 500. The impact of lower values is
surprisingly low.
The normalization exponent p11 was tested with the values 0.3, 0.4, 0.5,
0.6 and 0.7 The differences between the restoration times for the different
8.3. Other parameters 85
the evaluation of voltage deviations. The default value is 500. The
values are less than 20 minutes. The best result was fond for the values 0.5
and 0.6. The default value of p11 is 0.5.
8.3 Other parameters
The maximum number of tested actions must of course be large enough in

order to find a complete restoration sequence, on the other hand a too large
number will mean that when the search gets stuck it will use much compu-
tation time without doing significant progress. As can be seen in Figure 8.7
the restoration of the complete blackouts needs to test about 2 000 actions
and more actions makes no further improvement while the restoration of a
partial blackout continues to improve at least until 15 000 tested actions. The
default is to test maximum 3000 actions.
If the maximum number of iterations in the power flow algorithm is less
than 30 then at average more than 5% of the load will not be reconnected due
to more states with non-converging power flows. A small improvement can
be seen if the number of iterations is increased to 40. Larger values give no
Figure 8.7: Average restoration time as a function of the maximum number of

tested actions. The lines with triangles represent restoration from
a partial blackout; the lines with circles represent restoration from
the North and the lines with x:s represents restoration from both
the North and the South. The default is to test maximum 3000
actions.
8.4. Summary 87
improvement but increases the computation time for non-converging states.

The default value is 50.
In addition to the parameters above a number of additional parameters
have been tested but found to be of little or no value.
Pnlive For example has it been
Pnlive
tested to use i=1 nodes
|fi − fbase | in steed of i=1 nodes (fi − fbase )2 .
8.4 Summary
Most parameters have a wide range where the restoration algorithm gives
acceptable performance. There exists some critical transitions when the dif-
ferent parts of the evaluation function interact with each other, for example
p5 and p6 .
Chapter 9
Model parameter sensitivity analysis
Variations in the model parameters in table 4.2 have been tested in the same
way that variations in the search parameters were tested in the previous chap-
ter. This shows which conditions in the system that makes it hard to restore
the system. The ability to compare the difficulties of restoration in different
restoration situations was one of the main goals with automating the proce-
dure. The original goal was to be able to compare the restoration difficulties
in system with different amount of distributed generation of different types
but more development is needed before the algorithm can be used to compare
restoration with different production mixes in a useful way.
9.1 Limits
In order to make sure that all equipment works as intended the voltages, cur-
rents and frequencies need to be kept within certain limits. The narrower
these limits are the cheaper equipment can be used and the probability that
the equipment works as intended increases. On the other hand too narrow
limits can slow down the restoration as more actions will be rejected and
more set point adjustments need to be done. Figure 9.1 shows how the av-
erage restoration time depends on the allowed frequency deviation. If the
allowed frequency deviation is below ca 0.2 Hz then the restoration algo-
rithm gets difficulties. No improvement can be seen if the allowed frequency
deviation is above 0.8 Hz.
The default allowed voltage range is 0.9 to 1.1 p.u. the effect of changing
these limits can be seen in Figure 9.2. The voltage limits seem to have almost
89
90 Chapter 9. Model parameter sensitivity analysis
Figure 9.1: Average restoration time as a function of the maximum allowed

deviation from nominal frequency. The default value is 0.5 Hz.
The lines with circles represent restoration from the North and the
lines with x:s represents restoration from both the North and the
South.
9.1. Limits 91
Figure 9.2: Average restoration time as a function of the highest and lowest
allowed voltage. The default values are 0.9 and 1.1 p.u. Grey
lines represent the lower voltage limit and black lines represent the
higher voltage limit. The lines with circles represent restoration
from the North and the lines with x:s represents restoration from
both the North and the South. The voltage limits have been tested
independent of each other with the other limit fixed to its default
value.
no impact if they are outside 0.95 and 1.1 p.u. respectively. The restoration
seems to need to be allowed more over voltage than under voltage, this is
probably in order to be able to energize long lines. The algorithm will try to
keep the voltage as close to 1 p.u. as possible and will in most cases not find
a way to lower the voltage before connecting the line. These difficulties are
bigger when energizing from the North only.
As can be seen in Figure 9.3 the maximum allowed currents in the trans-
mission lines in this model are rather high compared to the currents in the
normal power flow in Figure 4.3.
The effect of scaling the maximum allowed currents can be seen in Figure
9.4. The restorations are only marginally affected as long as the current
limits are at least 50 % of their nominal values. By default the scale factor is
set to 5 which are almost equivalent to turning off the current limits.
9.2 Random faults
Often during a restoration some components are out of service due to known
faults and maintenances work. Unknown faulty components are an extra
complication that is not handled here. In order to test the ability to restore
the system in such situations, restorations with different number of random
transmission lines removed have been tested1 . The result can be seen in
Figure 9.5. As expected restoration from the North are more sensitive to
missing lines than restoration from both the North and the South. Restoration
from the North is almost unaffected for up to 3 missing lines. Restoration
from both the North and the South is almost unaffected up to 6 missing lines.
9.3 Scaling of the load
The effect of uniformly scale the active and reactive load can be seen in
Figure 9.6. For scaling factors below 0.5 the restoration gives difficulties,
probably due to high reactive generation in lightly loaded lines, this mostly
affects restoration from the North only. For scaling factors between 0.5 and
1
In order to get as consistent result as possible 50 random permutations of the 52 trans-
mission lines where created. For each of the permutations and each value of n from 0 to 52
a restoration was attempted with the n first transmission lines in the sequence removed from
the model.
9.3. Scaling of the load 93
Figure 9.3: Schematic showing the current limits as the width of the transmis-
sion lines in the same scale as the earlier power flow visualizations.
Figure 9.4: Average restoration time as a function of the scale factor for the
highest allowed current on transmission lines. The default values
are 5. The lines with circles represent restoration from the North
and the lines with x:s represents restoration from both the North
and the South.
9.3. Scaling of the load 95
Figure 9.5: Average restoration time as a function of the number of removed

transmission lines from the model. The lines with circles repre-
sent restoration from the North and the lines with x:s represents
restoration from both the North and the South.
Figure 9.6: Average restoration time as a function of the scaling of active and
reactive load. The lines with circles represent restoration from the
North and the lines with x:s represents restoration from both the
North and the South.
1.0 the average restoration time is almost proportional to the load. This can
be explained by the proportionality between the number of load steps and the
amount of load. For scaling factors above 1.0 it is difficult to energize the
entire load even if the generating capacity is large enough due to difficulties
transmitting the power to some loads.
9.4 Summary
The restoration algorithm has been tested in different power system restora-
tion situations. Much of the features in the obtained result can be explained
from a basic understanding of power system restoration.
Chapter 10
Conclusions
In this work, the search algorithm A∗ has been adapted to the power system
restoration problem.
10.1 Algorithm and test system
A power system restoration algorithm can be utilized in a number of different

ways for example:
• To help operators in real time by suggesting relevant actions.
• To completely automate the power system restoration. Because of lim-
ited personnel on duty this is especially interesting in small power
systems or independent restoration of many small islands in a larger
power system.
• To make comparative off line studies of power system restoration in
different situations such as:
– Addition of new components such as power lines or generating
units.
– Restoration with some non operational components.
– Variations in load patterns.
The proposed algorithm makes few direct assumptions about the restora-
tion sequence which makes it suitable to compare different restoration situa-
tions and priorities during the restoration.
97
98 Chapter 10. Conclusions
The proposed algorithm is relatively simple and can easily be extended

with new rules and priorities.
The algorithm checks each state with a power flow calculation. The cal-
culated power flow solution includes a frequency for each island, calculated
according to the generating units frequency regulation in steady state and
their limits in active power. The power system model contains no dynamic
components so all changes are the direct result of an action.
The algorithm is randomized; the advantage of this is that the algorithm
can be rerun several times with a new resulting restoration sequence each
time until a satisfying sequence is found.
The algorithm has been tested on a modified variant of the NORDIC32
test system. NORDIC32 is standardised by CIGRE for studying long-term
dynamics. It has 32 substations, 36 generating units and 545 switches. Most
of the switches are due to the fact that all loads have been divided in to load
steps of between 8 and 60 MW, each with its own switch. The total load in
the system is 10 900 MW.
Three different initial states have been tested for the NORDIC32 test sys-
tem:
• Total blackout with start in the North and the South of the test system.
• Total blackout with start in the North.
• Partial blackout somewhat similar to the blackout in year 1983.
The first two had high success rates, 92% and 86% of the test runs restores
all loads. The start after a partial black out had a lower success rate, 32% of
the test runs restores all load. One reason to why the partial blackout seems
harder than the other cases is probably that the algorithm has difficulties to
re-dispatch the production from the North to the newly energized Southern
part.
Restoration with random power lines removed is tested with good results
up to a point where about 20 % of the lines are removed.
The algorithm has been shown to be robust against variations in both
search parameters and model parameters. This has been shown by changing
one parameter at a time, run 50 test runs for each value of each parameter
with different random number sequences and calculate the average restora-
tion time for each value. In total more than 20 000 restorations have been
tested with a total number of power flow calculations over 40 000 000.
10.2. Computation time 99
The most important approximations made in the model are that it does not
contain:
• Cold load pickup.
• Generating units starting times and ramping limits.
• Uncertainty and unpredicted events in the power system model.
When used in a real situation each action needs to be checked against a

dynamic simulator. The proposed algorithm does not attempt to reach a state
that fulfils the N-1 criterion after all loads are reconnected.
10.2 Computation time
This thesis also shows that any reasonable formulation of the power system
restoration problem is NP-hard. Hence it is unlikely that an algorithm with
guaranteed success in a reasonable time exists.
The algorithm implementation needs less than 3000 power flows and 30
seconds on a 2.2 GHz AMD OpteronTM in order to find a sequence of about
750 actions that restores the power after an almost complete blackout in
NORDIC32.
No direct tests are done on how the computation time increases with the
complexity of the power system. Since the problem is NP-hard the compu-
tation time increases exponentially with the complexity of the power system
in the worst case. The current results suggest that the algorithm scales al-
most linearly in terms of number of power flow calculations except when it
reaches a state where the evaluation function does not give enough guidance,
then the algorithm gets exponential running time.
100 Chapter 10. Conclusions
Chapter 11
Future work
The presented algorithm shows promising results but the modelling needs
to be complemented in order to reflect more important aspects of the power
system restorations. The most obvious features to add are uncertainty in how
the power system behaves and time dependent and dynamic models for load
and generation.
11.1 Time dependent and dynamic models
The present algorithm only handles static systems, i.e. systems in which
every change is due to an action. The algorithm needs to be upgraded
to handle dynamic phenomena such as generating unit start up times, lim-
ited rate of change of the active power of generating units, reconnection of
thermostatically-controlled loads, and so on.
The evaluation of the result of an action in such a model should not be
based only at the state of the system immediately after the action but should
take in consideration the future development of the power system state. This
is easiest done if the evaluation assumes that every evaluated action is the
last in the sequence.
A dynamic electromechanical simulation of the result of each action prob-
ably takes too much computation time. Instead repeated power flow calcu-
lations, for example at the times 1, 2, 4, 8, 16, 32, 64, 128, 256 and 512
minutes from each action is taken can be used.
Other forms of time dependence that need to be considered are the limited
rate at which actions can be executed and the delay from decision to execu-
101
102 Chapter 11. Future work
tion. This is especially important for actions that are executed manually.
11.2 Other model improvements
There are a lot of details in the modelling of the power system components
that could be added such as:
• Tap changers.
• Protective relays.
• Switches with different ability to make and break current, this is the
difference between circuit breakers and disconnectors.
• Distribution nets that shall not be run in a meshed configuration due to

protection issues.
• More detailed power plant models.
11.3 Algorithmic improvements
The introduction of uncertainty and time dependent models increases the

computation time needed. This combined with the general goal of being able
to handle as many power system restoration situations as possible motivates
improvements in the algorithm.
The algorithm can be improved at several levels:
• Improvements in the general algorithm that are not directly related to

the details of power system restoration. For example:
– Change the order in which actions are tested so that not all ac-
tions in the state with the best evaluation needs to be tested before
actions in other states are tested.
– Divide the evaluation of states into more steps where faster cal-
culations are used to determine which states to evaluate further.
The last step when it is almost decided that the action will be
used could be a full dynamic simulation of the action.
11.4. Distributed and hierarchical restoration 103
• Improve the evaluation functions for example by adding terms for new
variables such as voltage angle differences, short circuit power and so
on.
• Speed improvements such as using incremental power flow calculation

were appropriate and improving details in the implementation of the
algorithm.
11.4 Distributed and hierarchical restoration
In order to plan restoration in large power systems or power systems mod-

elled in great detail with reasonable computation times it is necessary to
divide the model into loosely coupled sub systems.
This can be done with systems at the same level such as transmission
systems with few tie lines between each other.
It can also be done in a hierarchical manner that separates the planning
of the transmission system restoration from the restoration planning for each
of the subtransmission systems at lower levels. The planning of the trans-
mission system restoration then sets the limits for the exchange of power
for each of the subtransmission systems. In this manner the complete power
flow for the entire system need not to be recalculated each time a small action
such as connecting a small load at low voltage level is considered.
This is similar to the restoration planning in the planned cell concept of
Energinet.dk [28]. The cell concept is an attempt to redefine the distribu-
tion of responsibility between local grid companies and the transmission grid
company responsible for the system.
11.5 Uncertainty
Restoration under uncertainty can be simulated by using two power system

models. One to represent the real system, actions taken on this model can
not be undone just as in reality. The second one is used to represent the
algorithm knowledge about the real system, the algorithm can try different
actions in a search tree on this model as described earlier. The algorithm
searches for an action sequence in the second model and then applies one
action at a time on the first model. The algorithm reads values from the
first model corresponding to the measurements done in a real system. These
104 Chapter 11. Future work
values are compared against what is expected from the second model. If the
difference is larger than some threshold then the second model is updated by
some estimation algorithm so that it predicts values more similar to the one
from the first model and a new action sequence is computed from the current
state.
By introducing errors in the second model and in the measurements the
effect of such errors on the restoration can be studied
When doing a real restoration online the first model is of course replaced
by the real system.
This extension of the algorithm will increase the computation time sig-
nificantly by requiring a recalculation of the action sequence each time the
second model needs to be updated. By not computing a complete restoration
sequence each time the second model is updated this increase in restoration
time can be limited.
References
[1] Elavbrottet den 27 december 1983 – rapport från Vattenfalls störn-

ingskommission. Technical report, Vattenfall, 1984. In Swedish.
[2] Tekniska riktlinjer för kontrollanläggningar, automatiksystem, drift-

uppbyggnadsautomatik duba. kats 6.1 rev. 96-11-15 b. Technical re-
port, Svenska Kraftnät, 1996. In Swedish.
[3] Tekniska riktlinjer för kontrollanläggningar, automatiksystem, noll-

spänningsautomatik (uo-aut.). kats 6.4 rev. 96-11-15 b. Technical re-
port, Svenska Kraftnät, 1996. In Swedish.
[4] Elavbrottet 23 september 2003 – händelser och åtgärder. Technical

report, Svenska Kraftnät, 2003. In Swedish.
[5] Ieee guide for electric power distribution reliabilty indices. IEEE Std
1366-2003 (Revision of IEEE Std 1366-1998), 2004.
[6] Tekniska riktlinjer för kontrollanläggningar, hjälpkraftsystem, lik-

strömssystem. tr2-09-1 rev. 06-02-02 d. Technical report, Svenska
Kraftnät, 2006. In Swedish.
[7] CIGRE TF 38.02.08. Long term dynamics. phase ii. final report. Tech-
nical report, CE/SC 38 GT/WG 02, Ref. No. 102, 1995.
[8] M.M. Adibi. Power System Restoration, methodologies & implemen-

tation strategies. IEEE Press, 2000.
[9] Evert Agneholm. Cold Load Pick-up. PhD thesis, Chalmers university
of Technology, 1999.
105
106 References
[10] Giorgio Ausiello, M. Protasi, A. Marchetti-Spaccamela, G. Gambosi,

P. Crescenzi, and V. Kann. Complexity and Approximation: Combi-
natorial Optimization Problems and Their Approximability Properties.
Springer-Verlag New York, Inc., Secaucus, NJ, USA, 1999.
[11] S. Barsali, R. Salvati, and R. Zaottini. Use of hvdc links for power
system restoration. Electric Power Systems Research, 79(6):973–983,
2009.
[12] Liu C.-C., Liou K.-L., Chu R.F., and Holen A.T. Generation capa-
bility dispatch for bulk power system restoration: a knowledge-based
approach. IEEE Transactions on Power Systems, 8(1):316–325, 1993.
[13] Teo C.Y. and Wei Shen. Development of an interactive rule-based sys-
tem for bulk power system restoration. IEEE Transactions on Power
Systems, 15(2):646–653, 2000.
[14] A. Edström and K. Walve. The training simulator aristo - design and
experiences. In IEEE Power Engineering Society Winter Meeting, New
York, NY, USA, 1999.
[15] J. Figueiredo, J. Sauve, P. Nicolletti, E. Rocha, S. Araujo, F. Amorim,

A. Feitosa, R. Agra, and W. Ribeiro. Smart action: A tool to help power
system restoration. In Proceedings of the 41st Annual Hawaii Interna-
tional Conference on System Sciences (HICSS 2008), pages 167–167,
2008.
[16] L. H. Flink and K. Carlsen. Operating under stress and strain. In IEEE
Spectrum 15, 1978.
[17] P.E. Hart, N.J. Nilsson, and B. Raphael. A formal basis for the heuristic
determination of minimum cost paths. IEEE Transactions on Systems
Science and Cybernetics, 4(2):100–107, July 1968.
[18] Xiangzhen He, Zhiwei Liao, Wenxin Guo, Fushuan Wen, Junhui Liang,
and Jian Fu. A multi-objective model for transmission network restora-
tion based on multi-agent and tabu search. In 2008 Third Interna-
tional Conference on Electric Utility Deregulation and Restructuring
and Power Technologies, pages 2392–2397, Nanjing, China, 2008.
References 107
[19] Keisuke Iwasaki and Hidenori Aoki. Service restoration problem in

distribution power system using improved ga. Electrical Engineering
in Japan, 166(1):10–19, 2009.
[20] Roger Kearsley. Restoration in sweden and experience gained from the
blackout of 1983. IEEE Transactions on Power Systems, 2(2):422–428,
May 1987.
[21] S. Khushalani, J.M. Solanki, and N.N. Schulz. Optimized restoration of

unbalanced distribution systems. IEEE Transactions on Power Systems,
22(2):624–630, 2007.
[22] T. Kostic, R. Cherkaoui, A. Germond, and P. Pruvot. Decision aid

function for restoration of transmission powersystems: conceptual de-
sign and real time considerations. IEEE Transactions on Power Sys-
tems, 13(3):923–929, 1998. Lab. de Reseaux Electr., Ecole Polytech.
Federale de Lausanne.
[23] P. Kundur. Power system stability and control. McGraw-Hill, Inc, 1993.
ISBN 0-07-035958-X.
[24] Dong Liu, Yunping Chen, Guang Shen, and Youping Fan. A multi-
agent based approach for modeling and simulation of bulk power sys-
tem restoration. 2005 IEEE/PES Transmission & Distribution Confer-
ence & Exposition: Asia and Pacific, 2005.
[25] Dong Liu, Yunping Chen, Rongxiang Yuan, and Guang Shen. Back-
tracking based algorithm in hierarchical time-extended petri net model
for power system restoration. In 2005 IPEC The 7th International
Power Engineering Conference, Singapore, 2005.
[26] Qiang Liu, Libao Shi, Ming Zhou, Gengyin Li, and Yixin Ni. A new
solution to generators start-up sequence during power system restora-
tion. In 2008 Third International Conference on Electric Utility Dereg-
ulation and Restructuring and Power Technologies, pages 2845–2849,
Nanjing, China, 2008.
[27] Yan Liu and Xueping Gu. Reconfiguration of network skeleton based
on discrete particle-swarm optimization for black-start restoration. In
108 References
2006 IEEE Power Engineering Society General Meeting, Montreal,

Quebec, Canada, 2006.
[28] Per Lund. The danish cell project - part 1: Background and general
approach. In 2007 IEEE Power Engineering Society General Meeting,
Tampa, Florida, USA, 2007.
[29] K. Matsumoto, T. Sakaguchi, R.J. Kafka, and M.M. Adibi. Knowledge-

based systems as operational aids in power systemrestoration. Proceed-
ings of the IEEE, 80(5):689–697, 1992.
[30] A.A. Mota, L. Toledo, M. Mota, and A. Morelato. Restoration build-

ing blocks identification using a heuristic search approach. In 2006
IEEE Power Engineering Society General Meeting, Montreal, Quebec,
Canada, 2006.
[31] T. Nagata, Y. Tahara, and H. Fujita. An agent approach to bulk power

system restoration. Power Engineering Society General Meeting, 2005.
IEEE, 3:599–604 Vol. 1, 2005.
[32] T.J. Overbye and J.D. Weber. Visualization of power system data. In
Proceedings of the 33rd Annual Hawaii International Conference on
System Sciences, 2000., volume vol.1, 2000.
[33] Thi Thu Ha Pham, Y. Besanger, and N. Hadjsaid. New challenges

in power system restoration with large scale of dispersed generation
insertion. IEEE Transactions on Power Systems, 24(1):398–406, 2009.
[34] Gabriel Robins and Alexander Zelikovsky. Tighter bounds for graph
steiner tree approximation. SIAM Journal on Discrete Mathematics,
19:122–134, 2005.
[35] Stuart J. Russell and Peter Norvig. Artificial Intelligence: A Modern

Approach (2nd Edition). Prentice Hall, 2003.
[36] K. Salek, U. Spanel, and G. Krost. Flexible support for operators in

restoring bulk power systems. 2003 CIGRE/IEEE PES International
Symposium - Quality and Security of Electric Power Delivery Systems,
pages 187–192, 2003.
References 109
[37] S. Srivastava and K. L. Butler-Burry. Expert-system method for auto-

matic reconfiguration for restoration of shipboard power systems. "IEE
Proceedings – Generation, Transmission & Distribution", 153(3):253–
261, 2006.
[38] System Restoration Subcommittee. An approach to standing phase an-

gle reduction. IEEE Transactions on Power Systems, 9(1):470–478,
1994.
[39] K. Walve, A. Edstrom, and L. Karlstrom. Advanced dynamic simula-

tions as an important aid to restoration studies and training. In 2008
IEEE Power and Energy Society General Meeting - Conversion and
Delivery of Electrical Energy in the 21st Century, Pittsburgh, PA, USA,
2008.
[40] Kojima Y., Warashina S., Nakamura S., and Matsumoto K. Develop-
ment of a guidance method for power system restoration. IEEE Trans-
actions on Power Systems, 4(3):1219–1227, 1989.
110 References
Appendix A
Implementation
The first prototype of the system was developed in Matlab1 and used PSAT2
for power flow calculations. This was to slow and limited, it needed ca 30
minutes to find a restoration sequence.
Therefore the system was rewritten in C++3 . It worked well but lacked a
proper user interface and it was inconvenient to write simple scripting tasks
in C++.
This was solved by writing a Python4 program that encapsulates the com-
putation core that remains in C++.
The connection between the parts in C++ and Python is created by SWIG.
SWIG is an acronym for Simplified Wrapper Interface Generator and is a
program for generating interface code between code written in C/C++ and a
dynamic language such as Python5 . An overview of the software implemen-
tation can be seen in Figure A.1
1
See http://www.mathworks.com, 20 May 2009, the website of Mathworks, the company
that sells Matlab.
2
See F. Milano An Open Source Power System Analysis Toolbox, IEEE Trans.
on Power Systems, Vol. 20, No. 3, August 2005, pp. 1199-1206. and
http://www.power.uwaterloo.ca/ fmilano/psat.htm, 20 May 2009. The PSAT website of
PSAT’s creator Dr. Federico Milano.
3
GCC is used to compile the C++ code see http://gcc.gnu.org/, 20 May 2009, the GNU
Compiler Collection (GCC) website.
4
Python is a high-level, general purpose, interactive and object oriented language with a
large standard library. More info on: http://www.python.org/, 20 May 2009
5
See http://www.swig.org/, 20 May 2009, The SWIG website.
111
112 Appendix A. Implementation
The Python code contains the following parts:
• The overall search algorithm.
• Functions to load and save power system models including the state of
the system.
• The web interface and the scripting interface.
• Functions for generating plots and other figures.
The C++ code contains these parts:
• The topology calculation that deduces the islands and nodes to cal-
culate power flow for from the system topology and the status of the
switches.
• The power flow solver. This is a Newton Raphson solver that uses the
library Sparse 1.46 to solve the sparse linear equation systems of each
iteration, see section 3.1.
• The evaluation functions for both states and actions.
• The function for choosing an action to apply on a given state. This

uses the library RandomLib in order to randomize the algorithm. 7
The power system model file, the model parameter file and the search
parameter file are python programs that are run by the main program in order
to set up the power system model and the parameters. Appendix B contains
the power system model file for the variant of NORDIC32 that is used in this
thesis.
The program can be used in two modes, by the web interface and by the
scripting interface. The web interface is used for testing restorations interac-
tively. A screenshot of the web interface can be seen in Figure A.2.
The scripting interface is used by the example script that generates the
data for the figures in chapter 7.
The batch system also uses the scripting interface.
6
See http://sparse.sourceforge.net, 20 May 2009. The website of the Sparse linear algebra
library.
7
See http://charles.karney.info/random/, 20 May 2009. The website for RandomLib
113
Figure A.1: Overview of the software implementation.

114 Appendix A. Implementation
Figure A.2: A screenshot of the web interface. The plot shows the maximum
voltage angle difference during a restoration sequence.
The batch system is a set of Python scripts that are used to run large num-
bers of restoration simulations with variations in the parameters on a com-
puter cluster. It is used to generate the results in chapter 8 and 9.
Appendix B
Net data
1 #Import some modules

2 import restore # The API for the C++ −power flow and restoration code
3 from math import ∗ #Elementary math functions
4 from model_helper import ∗ #Helper functions used to build the net model
5 import copy
6 r = restor #Just a shorter name
7
8 #Define defaults for net components.
9 #This makes the declaration of each component shorter and easier to
understand .
10 # It is also necessary in order to be able to easy change global
properties of the model.
11 #These defaults can be overridden by defining rw. defaults before loading
the model.
12 #These defaults can also be overridden for each component by
13 #modifying the values before sending them to the component declaration .
See FT63−G11
14 defaults =Base_struct ()
15 defaults . Initial_state ="NS" #other states are "N", " half " and " full "
16 defaults . f_base=50.0 # Hz
17 defaults .S_base=100.0 # MVA
18 defaults .Vmax=1.1 #Maximum allowed voltage on any node
19 defaults .Vmin=0.9 #Minimum allowed voltage on any node
20 defaults .fmax=50.5 #Maximum allowed frequency on any node
21 defaults .fmin=49.5 #Minimum allowed frequency on any node
22 defaults .Load_factor_P=1 #provides a method of scaling all active
load at once.
115
116 Appendix B. Net data
23 defaults . Load_factor_Q=1 #provides a method of scaling all

reactive loads at once.
24 defaults . Load_Step_Size_factor=1 #provides a method of scaling the
number of load steps in all loads .
25 defaults . Load_on_factor=0 # fraction of all load switches
that are closed .
26 defaults . Generation_P_factor =1 #provides a method of scaling all
active generation at once
27 defaults . Generation_Q_factor=1 #provides a method of scaling all
reactive generation at once
28 defaults . Generation_Pinit =0 #The initial active power set
point of the generating units as fraction of Pmax−Pmin, Pinit=0 −>
pset=Pmin, Pinit=1 −> Pset=Pmax
29 defaults . Generation_On=False #The default is that the
generation is off in the initial state .
30 defaults . Generation_Vset=1.000 #Default voltage set point in per
unit for generating units .
31 defaults . Generation_dQ_per_dV=10000.0 # Multiplied by (Qmax−Qmin),
delta V in p.u.
32 defaults . Generation_dP_per_df_hydro=−0.1512 #Multiplied by (Pmax−Pmin),
delta f in Hz
33 defaults . Generation_dP_per_df_thermal=0.0
34
35 #Has the defaults been overridden?
36 if rw.__dict__ .has_key(" defaults ") :
37 d= copy.deepcopy(rw. defaults ) #Just a shorter name
38 defaults =d
39 else :
40 d= defaults #Just a shorter name
41 rw. defaults_used =d #For use when printing key figures etcetera .
42
43 #Create an empty power net .
44 n=r.Net(d. f_base , d.S_base)
45 #Add components
46 #Switches to connect the objects to the busbars are automatically
generated by the helper functions .
47 #The nodes between the switches and the objects are also automatically
created .
48 #All data are given in per unit .
49
50 #Declare nodes that should be considered busbars and the nominal voltages
117
for these .
51 Busbars(n,d,Vn=400,
52 Names=["CT11B","CT71","CT11A","CT21","CT22B","CT12A","CT12B","
CT72A","CT72B","CT22A","CT31","CT32A","FT44B","CT32B","FT41A",
"FT41B","FT42","FT44A","FT43B","FT43A","FT45","FT47A","FT61","
FT50A","FT50B","FT47B","FT62","FT63A","FT63B","FT51"])
53
54 Busbars(n,d,Vn=220, Names=["CT231","AT241"])
55
56 Busbars(n,d,Vn=130,Names=["AT111","CT111","AT121","CT112","AT131","
CT122","RT132B","FT142A","RT132A","FT142B","RT131B","RT131A","
RT133","FT150A","FT150B"])
57
58 #Declare the substation names and one node in each.
59 #The program finds the rest of the nodes by following all
60 #switches .
61 Substations (n, d, [
62 ("AT111_swy","AT111" ),
63 ("CT111_swy","CT111" ),
64 ("AT121_swy","AT121" ),
65 ("CT112_swy","CT112" ),
66 ("AT131_swy","AT131" ),
67 ("CT122_swy","CT122" ),
68 ("CT231_swy","CT231" ),
69 ("AT241_swy","AT241" ),
70 ("CT71_swy" ,"CT71" ) ,
71 ("CT11_swy" ,"CT11A" ),
72 ("CT21_swy" ,"CT21" ) ,
73 ("CT12_swy" ,"CT12A" ),
74 ("CT72_swy" ,"CT72A" ),
75 ("CT22_swy" ,"CT22A" ),
76 ("CT31_swy" ,"CT31" ) ,
77 ("CT32_swy" ,"CT32A" ),
78 ("FT41_swy" ,"FT41A" ),
79 ("FT42_swy" ,"FT42" ) ,
80 ("FT44_swy" ,"FT44A" ),
81 ("FT43_swy" ,"FT43A" ),
82 ("FT45_swy" ,"FT45" ) ,
83 ("FT47_swy" ,"FT47A" ),
84 ("FT61_swy" ,"FT61" ) ,
85 ("FT50_swy" ,"FT50A" ),
86 ("FT62_swy" ,"FT62" ) ,
87 ("FT63_swy" ,"FT63A" ),
88 ("FT51_swy" ,"FT51" ) ,
89 ("FT142_swy","FT142A"),
90 ("RT132_swy","RT132A"),
91 ("RT131_swy","RT131A"),
92 ("RT133_swy","RT133" ),
93 ("FT150_swy","FT150A")])
94
95 #Loads(net, defaults , load_point , Busbar,P , Q , step_size_p ) ;
96 #Each row generates as many loads as necessary so
97 #that P of each load will not be bigger than step_size_p .
98 #P and Q are distributed even between the loads generated on one row.
99 Loads(n, d, "AT111−T" , "AT111" , 1.0 , 0.4 , 0.2 ) ;
100 Loads(n, d, "CT111−T" , "CT111" , 2.0 , 0.8 , 0.2 ) ;
101 Loads(n, d, "CT112−T" , "CT112" , 3.0 , 1.0 , 0.2 ) ;
102 Loads(n, d, "CT122−T" , "CT122" , 2.8 , 0.95, 0.4 ) ;
103 Loads(n, d, "FT142−T1", "FT142A", 4.0 , 1.5 , 0.4 ) ;
104 Loads(n, d, "FT142−T2", "FT142B", 4.0 , 1.5 , 0.3 ) ;
105 Loads(n, d, "FT150−T1", "FT150A", 1.5 , 0.5 , 0.5 ) ;
106 Loads(n, d, "FT150−T2", "FT150B", 5.5 , 2.0 , 0.3 ) ;
107 Loads(n, d, "RT131−T1", "RT131A", 1.725, 0.03, 0.2 ) ;
108 Loads(n, d, "RT131−T2", "RT131B", 0.575, 0.01, 0.2 ) ;
109 Loads(n, d, "RT132−T1", "RT132A", 1.5 , 0.4 , 0.3 ) ;
110 Loads(n, d, "RT132−T2", "RT132B", 1.5 , 0.4 , 0.3 ) ;
111 Loads(n, d, "RT133−T" , "RT133" , 6.0 , 0.2 , 0.3 ) ;
112 Loads(n, d, "CT231−T" , "CT231" , 1.0 , 0.3 , 0.2 ) ;
113 Loads(n, d, "AT241−T" , "AT241" , 2.0 , 0.5 , 0.2 ) ;
114 Loads(n, d, "CT71−T" , "CT71" , 3.0 , 1.0 , 0.6 ) ;
115 Loads(n, d, "CT72−T1" , "CT72A" , 10.0 , 2.5 , 0.4 ) ;
116 Loads(n, d, "CT72−T2" , "CT72B" , 10.0 , 2.5 , 0.4 ) ;
117 Loads(n, d, "FT41A−T1", "FT41A" , 2.85 , 0.9 , 0.5 ) ;
118 Loads(n, d, "FT41−T2" , "FT41B" , 2.55 , 0.7 , 0.3 ) ;
119 Loads(n, d, "FT43−T1" , "FT43A" , 4.75 , 1.11, 0.3 ) ;
120 Loads(n, d, "FT43−T2" , "FT43B" , 4.25 , 1.92, 0.3 ) ;
121 Loads(n, d, "FT44−T1" , "FT44A" , 0.75 , 0.4 , 0.3 ) ;
122 Loads(n, d, "FT44−T2" , "FT44B" , 3.25 , 1.1 , 0.3 ) ;
123 Loads(n, d, "FT45−T" , "FT45" , 7.0 , 2.5 , 0.4 ) ;
124 Loads(n, d, "FT47−T1" , "FT47A" , 0.75 , 0.375,0.3 ) ;
125 Loads(n, d, "FT47−T2" , "FT47B" , 0.25 , 0.125,0.1 ) ;
126 Loads(n, d, "FT51−T" , "FT51" , 8.0 , 3.0 , 0.4 ) ;
119
127 Loads(n, d, "FT61−T" , "FT61" , 5.0 , 1.5 , 0.5 );

128 Loads(n, d, "FT62−T" , "FT62" , 3.0 , 0.85, 0.5 );
129 Loads(n, d, "FT63−T1" , "FT63A" , 2.92 , 1.5 , 0.5 );
130 Loads(n, d, "FT63−T2" , "FT63B" , 2.92 , 1.5 , 0.5 );
131
132 #Generation_unit ( net , defaults , Type, Name, Busbar, Pmax, Pmin, Qmax,
Qmin)
133
134 Generation_unit (n, d, "hydro" , "AT111−G" , "AT111" , 5.7 , 0.0, 3.0
,−1.0 )
135 Generation_unit (n, d, "hydro" , "AT121−G" , "AT121" , 6.65, 0.0, 3.5
,−1.0 )
136 Generation_unit (n, d, "hydro" , "AT131−G" , "AT131" , 5.7 , 0.0, 3.0
,−0.6 )
137 Generation_unit (n, d, "hydro" , "CT112−G1", "CT112" , 2.85, 0.0, 1.5
,−0.2 )
138 Generation_unit (n, d, "hydro" , "CT112−G2", "CT112" , 4.75, 0.0, 2.5
,−0.4 )
139 Generation_unit (n, d, "hydro" , "CT122−G" , "CT122" , 2.38, 0.0,
1.75,−0.25)
140 Generation_unit (n, d, "hydro" , "RT131−G" , "RT131A" , 1.8 , 0.6, 2.0 ,
0.0 )
141 Generation_unit (n, d, "thermal" , "RT132−G1", "RT132A", 3.6 , 1.2, 2.0 ,
0.0 )
142 Generation_unit (n, d, "thermal" , "RT132−G2", "RT132B", 1.8 , 0.6, 1.0 ,
0.0 )
143 Generation_unit (n, d, "hydro" , "AT241−G1", "AT241" , 4.75, 0.0, 2.5
,−1.0 )
144 Generation_unit (n, d, "hydro" , "AT241−G2", "AT241" , 4.75, 0.0,
1.75,−0.8 )
145 Generation_unit (n, d, "hydro" , "CT11−G1" , "CT11A" , 3.8 , 0.0, 2.0
,−0.5 )
146 Generation_unit (n, d, "hydro" , "CT11−G2" , "CT11B" , 5.7 , 0.0, 3.0
,−0.5 )
147 Generation_unit (n, d, "hydro" , "CT12−G1" , "CT12A" , 4.75, 0.0, 2.5
,−0.3 )
148 Generation_unit (n, d, "hydro" , "CT12−G2" , "CT12B" , 2.85, 0.0, 1.5
,−0.3 )
149 Generation_unit (n, d, "hydro" , "CT21−G" , "CT21" , 2.85, 0.0, 1.5
,−0.3 )
150 Generation_unit (n, d, "hydro" , "CT31−G" , "CT31" , 3.33, 0.0,
1.75,−0.4 )
151 Generation_unit (n, d, "hydro" , "CT71−G" , "CT71" , 4.75, 0.0, 2.5
,−1.0 )
,−3.0 )
,−3.0 )
,−1.5 )
,−1.5 )
156 Generation_unit (n, d, "thermal" , "FT41−G" , "FT41A" , 0.0 , 0.0, 3.0
,−2.0 )
157 Generation_unit (n, d, "thermal" , "FT44−G1" , "FT44B" , 6.3 , 1.8, 3.5 ,
0.0 )
158 Generation_unit (n, d, "thermal" , "FT44−G2" , "FT44A" , 1.8 , 0.6, 1.0 ,
0.0 )
0.0 )
160 Generation_unit (n, d, "thermal" , "FT47−G2" , "FT47B" , 5.4 , 1.0, 3.0 ,
0.0 )
161 Generation_unit (n, d, "thermal" , "FT51−G1" , "FT51" , 6.3 , 1.5, 3.5 ,
0.0 )
0.0 )
163 Generation_unit (n, d, "thermal" , "FT61−G" , "FT61" , 1.8 , 0.6, 1.0 ,
0.0 )
0.0 )
0.0 )
0.0 )
0.0 )
0.0 )
169 #Define a black start unit in the South
170 # represents an aggregation of Gas turbines and hydro power in Scania and
Zealand
171 #that has black start capability , not part of the original Nordic 32
121
model.
172 #Give some modified defaults .
173 d_black_start = copy.deepcopy(d)
174 d_black_start . Generation_Pset =1.5/(5.4+1.5) #To get 50 Hz at 0 MW
175 d_black_start . Generation_dP_per_df_thermal=−4.0/7.0
176 Generation_unit (n, d_black_start , "thermal" , "FT63−G11", "FT63A",
5.4,−1.5, 3.0,−3.0 )
177
178 #Shunt_reactance( net , defaults , Name, bus, Qnom)
179 Shunt_reactance (n, d, "CT122−EK", "CT122" , 0.5)
180 Shunt_reactance (n, d, "FT142−EK", "FT142A", 2.0)
181 Shunt_reactance (n, d, "FT150−EK", "FT150A", 2.0)
182 Shunt_reactance (n, d, "RT131−EK", "RT131B", 1.0)
183 Shunt_reactance (n, d, "RT133−EK", "RT133" , 0.7)
184 Shunt_reactance (n, d, "CT11−X1" , "CT11A" ,−1.0)
185 Shunt_reactance (n, d, "CT11−X2" , "CT11B" ,−1.0)
192 Shunt_reactance (n, d, "CT21−X1" , "CT21" ,−1.5)
202 Shunt_reactance (n, d, "CT71−EK" , "CT71" , 4.0)
207 Shunt_reactance (n, d, "FT41−EK" , "FT41B" , 2.0)
208 Shunt_reactance (n, d, "FT41−X" , "FT41B" ,−1.0)
209 Shunt_reactance (n, d, "FT42−X1" , "FT42" ,−1.0)

212 Shunt_reactance (n, d, "FT43−EK" , "FT43A" , 2.0)
213 Shunt_reactance (n, d, "FT44−X1" , "FT44A" ,−1.5)
214 Shunt_reactance (n, d, "FT44−X2" , "FT44B" ,−1.0)
216 Shunt_reactance (n, d, "FT45−EK" , "FT45" , 1.0)
218 Shunt_reactance (n, d, "FT50−X2" , "FT50B" ,−1.0)
219 Shunt_reactance (n, d, "FT51−EK" , "FT51" , 1.0)
220 Shunt_reactance (n, d, "FT51−X" , "FT51" ,−1.0)
224
225 Ibase130=d.S_base /130.0/(3.0∗∗0.5) #base current in kA for 130, 220 and
400 kV
226 Ibase220=d.S_base /220.0/(3.0∗∗0.5)
227 Ibase400=d.S_base /400.0/(3.0∗∗0.5)
228
229 #Line(n, d, Name, busbar1, busbar2, R, X, B, Imax)
230 Line(n, d, "AL1", "AT111", "CT111", 0.010, 0.070, 0.014, 3.12/ Ibase130 )
232 Line(n, d, "AL3", "AT111", "AT121", 0.007, 0.050, 0.010, 4.37/ Ibase130 )
233 Line(n, d, "AL4", "AT111", "AT121", 0.007, 0.050, 0.010, 4.37/ Ibase130 )
234 Line(n, d, "AL5", "CT112", "AT121", 0.014, 0.090, 0.018, 2.25/ Ibase130 )
239 Line(n, d, "AL10","CT231", "AT241", 0.012, 0.090, 0.015, 1.50/ Ibase220 )
240 Line(n, d, "CL1", "CT11B", "CT71" , 0.005, 0.045, 1.400, 1.50/ Ibase400 )
241 Line(n, d, "CL2", "CT11A", "CT21" , 0.006, 0.060, 1.800, 2.25/ Ibase400 )
242 Line(n, d, "CL3", "CT11B", "CT22B", 0.004, 0.040, 1.200, 2.25/ Ibase400 )
243 Line(n, d, "CL4", "CT11A", "CT12A", 0.001, 0.008, 0.200, 6.25/ Ibase400 )
245 Line(n, d, "CL6", "CT71" , "CT72A", 0.003, 0.030, 3.000, 2.25/ Ibase400 )
246 Line(n, d, "CL7", "CT71" , "CT72B", 0.003, 0.030, 3.000, 2.25/ Ibase400 )
247 Line(n, d, "CL8", "CT12A", "CT22A", 0.004, 0.035, 1.050, 2.25/ Ibase400 )
249 Line(n, d, "CL10","CT22A", "CT31" , 0.004, 0.040, 1.200, 2.25/ Ibase400 )
250 Line(n, d, "CL11","CT21" , "CT32A", 0.004, 0.040, 1.200, 2.25/ Ibase400 )
123
251 Line(n, d, "CL12","CT21" , "FT44B", 0.010, 0.060, 3.000, 2.25/ Ibase400 )

252 Line(n, d, "CL13","CT31" , "CT32B", 0.001, 0.010, 0.300, 6.25/ Ibase400 )
253 Line(n, d, "CL14","CT31" , "FT41A", 0.006, 0.040, 2.400, 2.25/ Ibase400 )
254 Line(n, d, "CL15","CT31" , "FT41B", 0.006, 0.040, 2.400, 2.25/ Ibase400 )
255 Line(n, d, "CL16","CT32A", "FT42" , 0.006, 0.050, 2.400, 2.25/ Ibase400 )
256 Line(n, d, "CL17","CT32B", "FT44A", 0.010, 0.040, 2.000, 2.25/ Ibase400 )
257 Line(n, d, "FL1" ,"FT43B", "FT44A", 0.002, 0.015, 0.500, 4.37/ Ibase400 )
258 Line(n, d, "FL2" ,"FT41A", "FT42" , 0.003, 0.030, 0.900, 2.25/ Ibase400 )
259 Line(n, d, "FL3" ,"FT42" , "FT43A", 0.001, 0.010, 0.300, 6.25/ Ibase400 )
264 Line(n, d, "FL8" ,"FT42" , "FT50B", 0.002, 0.020, 0.600, 4.37/ Ibase400 )
265 Line(n, d, "FL9" ,"FT43B", "FT47B", 0.002, 0.020, 0.600, 4.37/ Ibase400 )
266 Line(n, d, "FL10","FT61" , "FT62" , 0.002, 0.020, 0.600, 4.37/ Ibase400 )
267 Line(n, d, "FL11","FT50A", "FT62" , 0.011, 0.080, 2.400, 2.00/ Ibase400 )
268 Line(n, d, "FL12","FT62" , "FT63A", 0.003, 0.030, 0.900, 2.25/ Ibase400 )
269 Line(n, d, "FL13","FT62" , "FT63B", 0.003, 0.030, 0.900, 2.25/ Ibase400 )
270 Line(n, d, "FL14","FT50A", "FT51" , 0.004, 0.040, 1.200, 2.25/ Ibase400 )
271 Line(n, d, "FL15","FT50B", "FT51" , 0.004, 0.040, 1.200, 2.25/ Ibase400 )
272 Line(n, d, "FL16","FT42" , "FT44B", 0.002, 0.020, 0.600, 4.37/ Ibase400 )
273 Line(n, d, "RL1" ,"RT132B","FT142A",0.038, 0.280, 0.060, 1.50/ Ibase130 )
274 Line(n, d, "RL2" ,"RT132A","FT142B",0.038, 0.280, 0.060, 1.50/ Ibase130 )
276 Line(n, d, "RL4" ,"RT131A","FT142B",0.010, 0.080, 0.016, 3.12/ Ibase130 )
277 Line(n, d, "RL5" ,"RT131A","RT133", 0.010, 0.060, 0.012, 4.37/ Ibase130 )
278 Line(n, d, "RL6" ,"RT131B","RT133", 0.010, 0.060, 0.012, 4.37/ Ibase130 )
280 Line(n, d, "RL8" ,"RT133", "FT150A",0.015, 0.120, 0.025, 2.25/ Ibase130 )
281 Line(n, d, "RL9" ,"RT133", "FT150B",0.015, 0.120, 0.025, 2.25/ Ibase130 )
282
283 #Transformer(net , defaults , Name, busbar1, busbar2, R, X, B, Imax,
tap_changer_steps , current_TC_step)
284
285 Transformer(n, d, "TRAFO1", "CT11A", "CT111" , 0.0, 0.008, 0.0, 1e+200,
(1.0,) , 1)
286 Transformer(n, d, "TRAFO2", "CT12B", "CT112" , 0.0, 0.008, 0.0, 1e+200,
(1.0,) , 1)
287 Transformer(n, d, "TRAFO3", "CT22A", "CT122" , 0.0, 0.012, 0.0, 1e+200,
(1.0,) , 1)
288 Transformer(n, d, "TRAFO4", "FT42" , "FT142A", 0.0, 0.010, 0.0, 1e+200,

(1.0,) , 1)
289 Transformer(n, d, "TRAFO5", "FT42" , "FT142B", 0.0, 0.010, 0.0, 1e+200,
(1.0,) , 1)
290 Transformer(n, d, "TRAFO6", "FT50A", "FT150A", 0.0, 0.010, 0.0, 1e+200,
(1.0,) , 1)
291 Transformer(n, d, "TRAFO7", "FT50B", "FT150B", 0.0, 0.010, 0.0, 1e+200,
(1.0,) , 1)
292 Transformer(n, d, "TRAFO8", "CT31" , "CT231" , 0.0, 0.010, 0.0, 1e+200,
(1.0,) , 1)
293
294 #switches between busbars need to be defined manually.
295 #Switch(n, d, Name, node1, node2, state )
296
297 Switch(n, d, "CB_CT11" , "CT11A" , "CT11B" , False)
302 Switch(n, d, "CB_FT41" , "FT41A" , "FT41B" , False )
308 Switch(n, d, "CB_FT142", "FT142A", "FT142B", False)
311 Switch(n, d, "CB_RT131", "RT131A", "RT131B", False)
314
315 if d. Initial_state =="NS":
316 # Generation_state ( net , Name,on,Vset,Pset)
317 Generation_state (n,"FT63−G11pv",True,1.0,0.0)
318 Generation_state (n,"CT72−G1pv",True,1.0,0.0)
319
320 elif d. Initial_state =="N":
321 # Generation_state ( net , Name,on,Vset,Pset)
323
125
324 elif d. Initial_state =="half" :

325 # Generation_state ( net ,Name,on,Vset,Pset)
326 Generation_state (n,"AT111−Gpv",True,1.060000,3.000000)
331 Generation_state (n,"CT122−Gpv",True,1.060000,2.000000)
332 Generation_state (n,"AT241−G1pv",True,1.030000,0.000000)
342 Generation_state (n,"CT72−G2pv",False,1.030000,7.500000)
345 #close switches
346 Set_switch_state_re (n, True, [".∗CT.∗", ".∗AT.∗"])
347 #open some switches
348 Set_switch_state_re (n, False , [".∗−EK.∗", ".∗−X.∗"])
349 Set_switch_state_re (n, False , [".∗CL12.∗", ".∗CL14.∗", ".∗CL15.∗"
, ".∗CL16.∗", ".∗CL17.∗"])
350 Set_switch_state_re (n, False , ["CB_CT72−G2","CB_CT72−G3","
CB_CT72−G4"])
351 #close some switches
352 Set_switch_state_re (n, True, ["CB_CT32−X1", "CB_CT21−X1"])
353
354
355 elif d. Initial_state ==" full " :

356 # Generation_state ( net ,Name,on,Vset,Pset)
363 Generation_state (n,"RT131−Gpv",True,1.060000,1.800000)

364 Generation_state (n,"RT132−G1pv",True,1.060000,3.000000)
365 Generation_state (n,"RT132−G2pv",True,1.060000,0.600000)
379 Generation_state (n,"FT41−Gpv",True,1.030000,0.000000)
386 Generation_state (n,"FT61−Gpv",True,1.030000,1.700000)
392
393 #close all switches
394 Set_switch_state_re (n, True, [".∗" ])
395 #open a few switches to reactive compensation
396 Set_switch_state_re (n, False , [".∗−EK.∗",".∗−X.∗"])
397 Set_switch_state_re (n, True, ["CB_CT32−X1", "CB_CT21−X1", "
CB_FT62−X1"])
398
399 else :
400 print " error unknown initial state "

LTH Iea 1060 PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LTH Iea 1060 PDF

Uploaded by

Copyright:

Available Formats

Automatic Power System

c Lars Lindgren, 2009

Power system restoration, Load flow analysis, Tree searching, A∗ -search,

Omfattande strömavbrott är ovanliga men kan få stora konsekvenser för sam-

Active power The average power over a cycle in the AC voltage.

Bottom up restoration Restoration from several points and then synchro-

Disconnector A mechanical switch that only can be opened and closed

Distributed generation Small generating units that are connected to the

Evaluation function A function that is used to evaluate and compare how

Heuristics Method based on experience such as “rule of thumb”.

Load flow analysis See Power flow analysis.

NP-complete A class of computation problems that are considered hard to

NP-hard The class of all computation problems that is at least as hard to

Renewable generation Power generation that use renewable energy sources

Substation A location for switching equipment and transformers used to

Synchronization The act of closing a circuit breaker between two electrical

Topology calculation In power system analysis topology calculation is the

It’s a necessary pre-processing step before a power flow calculation

3 Modeling and simulator 21

8 Algorithm parameter sensitivity analysis 79

9 Model parameter sensitivity analysis 89

11 Future work 101

B Net data 115

2. Analysis of actual sequences of events.

3. Simulation studies with sequences of actions produced manually by a

4. Realistic real-time simulation exercises in which a team of operators

5. To use automatic search algorithms to find suitable sequences of ac-

1.2 Scope of the thesis

In order to compare different power system restoration strategies or situa-

rithm for finding a sequence of actions that restores a power system.

1.3 Previous work

Power System type Power system restoration needs to be done on different

black start after a complete blackout [27, 25, 33].

Automation level Much research is on the fundamental issues in power sys-

• Uses a significantly different approach from the other shown articles.

The found articles are:

• ”Backtracking based algorithm in hierarchical time-extended petri net

• ”An agent approach to bulk power system restoration“ by T. Nagata et

• ”Development of an interactive rule-based system for bulk power sys-

• ”Generation capability dispatch for bulk power system restoration: a

• ”Development of a guidance method for power system restoration“

An interesting article on restoration at the distribution level is ”New chal-

Lindgren, L. and Eliasson, B. Automated power system restoration, Nordic

Samuelsson, O., Lindgren, L. and Eliasson, B. Simulated power system

An automatic power system restoration algorithm can be used in different

• Automatic power system restoration.

• Operator support, suggesting actions to the operator in real time.

• Off-line planning for restoration in advance of the blackout. This in-

The focus in this thesis is planning for restoration in advance.

1.6 Outline of the thesis

Chapter 2 gives an overview of the power system restoration problem from a

Chapter 10 summarizes and evaluates the results.

2.1 The operational states of the power system

In order to describe operating situations in a power system they can be cat-

Power system restoration is a very complex task. The restoration is here

early is that permanent faults in disconnectors, current breakers and trans-

2.2.1 Geographical aspects

In some cases it is also economical to use HVDC in parallel to an AC

2.3 Power plant characteristics

2.3.1 Hydro power

2.3.2 Wind power

Different types of wind turbine generators have very different characteristics.