Вы находитесь на странице: 1из 10

Electrical Power & Energy Systems, Vol. 20, No. 2, pp.

131–140, 1998
q 1997 Elsevier Science Ltd. All rights reserved
Printed in Great Britain
PII: S0142-0615(97)00036-7 0142-0615/98 $19.00+0.00

Optimization and learning of


load restoration strategies
Tatjana Kostic and Alain J Germond
Laboratoire de Réseaux d’Énergie Électrique,
École Polytechnique Fédérale de Lausanne, Lausanne,
CH-1015, Switzerland

Juan J Alba
Instituto de Investigación Tecnológica, Universidad
Pontificia Comillas de Madrid, Alberto Aguilera, 23,
28015, Madrid, Spain

this task progresses. Simultaneously, transmission capa-


This paper describes an application of optimization and
city is reinforced by energizing additional circuits.
machine learning to load restoration in a generation–
transmission system. An optimization procedure, combining Frequently, service restoration plans are well determined
a genetic algorithm and a power system dynamic simulator, only for the first stage. There are two reasons for this: (a) the
generates the appropriate sequence of operations for each plan is guided by the natural order of control actions (e.g. a
state of the power system. A machine learning technique stand-by unit cannot be synchronized if there is no energized
(induction of decision trees) is applied to extract decision link to it; a link can be energized only if there is voltage on its
criteria that will guide the load restoration after a general- end, given by a black start unit), and (b) during this stage, the
ized black-out. The paper also presents the results of apply- power system is highly vulnerable concerning violation of
ing these techniques to a power system of realistic size. q operating constraints.
1997 Elsevier Science Ltd. On the contrary, the load restoration (LR) stage is con-
sidered less critical, and the operators are left to take
Keywords: restoration, genetic algorithms, machine decisions about control actions according to their own
learning, simulation tools experience. They are likely to reason guided by their intui-
tion and the experience acquired in operating the system in
its normal state. This approach may lead to some actions
I. Introduction (e.g. load connection at an excessive rate) that threaten the
whole state reached until then. Therefore, there is a need for
The stressful conditions in which service restoration is better procedures for the load restoration stage, and even for
carried out after a generalized blackout (limited time, occur- computer-assisted decision making tools.
rence of unexpected events, uncertainties about the There are plenty of works concerning service restoration
behaviour of the power system) would make any decision- at transmission level; a comprehensive bibliographic review
making tool able to assist the operators in this task extremely of this subject can be found in [1]. However, the load
useful. restoration sub-problem has not received much attention.
A restoration procedure can be divided as follows: Although most of the bibliography in this area briefly
• In the first stage, the so-called skeleton of the power discusses it, no solutions are widely accepted.
system is built by synchronizing the available power
plants and energizing the main lines connecting them. I.1 The proposed approach
The voltage set-points of the generators are usually set to The determination of an acceptable sequence of control
values lower than the nominal voltage and some loads are actions to pick up load is a highly combinatorial optimization
picked up, to avoid overvoltages resulting from energiza- problem with multiple constraints. Genetic algorithms have
tion of long unloaded lines. If the system is split into proved quite a successful technique for solving that kind of
islands, the above-mentioned procedure is applied to each problem. In this work, a genetic algorithm has been used to
island, after which the islands are synchronized among generate optimized load restoration (LR) sequences for
themselves. different states of the power system, respecting the oper-
• Most of the load is picked up in the second stage (that will ating constraints. The genetic algorithm has been coupled
be called the load restoration stage). The voltage and with a power system dynamic simulator, whose results are
power set-points of the generators must be increased as used by the evaluation function of the genetic algorithm.

131
132 Load restoration strategies: T. Kostic et al.

The appropriate load restoration sequence depends on the to select the adequate strategy for a real problem of service
power system state, that consists of: restoration.
• The pre-blackout loads, that are assumed to be the loads to Notice that the most time-consuming tasks (optimization,
be restored. generation and validation of the decision trees) are carried
• The availability of generating units and circuits at the end out off line, while only the application of the criteria coded in
of the first service restoration stage (i.e. when the network the decision tree is performed on line.
skeleton has been created). Decision trees have been chosen instead of other machine
learning techniques since they provide an output that is easy
An ideal use of the GA-based procedure would be the to validate: a decision tree presents a clear organization of
generation of an LR sequence for a given state of the power the relevant attributes, that can easily be analyzed by a
system. However, this approach is not feasible in a real domain expert. Other learning techniques provide black-
environment, because of the required computing time. A box type solutions, that sometimes can give a better perfor-
useful solution to this problem should be the use of simple mance but are more difficult to analyze. Since the work
rules to select the most appropriate sequence to pick up load; reported in this paper is still in a preliminary phase, and its
the simulator could provide some additional refinement to main objective was to evaluate the potential applicability of
the suggestions of the rules, to cope with unexpected situa- this technique, the ‘‘transparency’’ of the decision trees has
tions. Nevertheless, these rules would depend on the parti- been considered an important characteristic.
cular network involved, as there is not enough experience to
write a reliable set of rules for a generic power system. I.2 Overview of the paper
Machine learning techniques can be used to extract this
experience. Induction algorithms are a subset of machine Section II presents a brief introduction to genetic algorithms.
learning techniques that are able to extract useful and The power system dynamic simulator is described in Section
systematic knowledge from examples. Decision trees are a III, to facilitate the detailed discussion of the GA-based
subset of induction techniques that have proven efficient and optimization in Section IV. Section V presents some results.
robust in similar problems. They could be applied to the Section IV introduces the decision tree technique. Section
service restoration problem as follows (see Figure 1): VII describes the combination of the GA-optimization and
the dynamic simulator to generate examples for constructing
• Different power system states are explored off-line. For and validating the decision tree. Some results are presented
each of them, an appropriate sequence for picking up load and discussed in Section VIII. Final remarks and conclusions
is generated, using the GA-based optimization. This follow in Section IX.
information is gathered in the training set.
• An induction algorithm generates a decision tree that
captures rules relating power system pre-blackout data II. Genetic algorithms
and LR sequences. Genetic algorithms [2,3] were developed in the early 70s as
• The decision tree must be validated, using additional data an attempt to model adaptive systems, and are inspired by
(grouped in the so-called test set). The tree construction natural selection and reproduction. They combine the
must be adjusted if needed. random search and hill-climbing techniques with the idea
• The decision tree can be used by an operator support tool of competition. This method can be used for different
purposes, including optimization.

II.1 Principles
GAs manipulate a population of individuals, denoted as
parents, to produce a new population, denoted as children.
The quality of each individual is measured by the fitness.
Children are created through the application of different
operators to the parents. These operators are designed so that
children are expected to have higher fitness than their
parents.
The main GA operators [2] are selection, crossover and
mutation. The selection operator chooses the couple of
parents that will produce a couple of children. The better
the fitness of a parent is, the higher is its chance of being
selected for reproduction. The crossover operator mixes the
genetic material of the selected couple of parents. Finally,
the mutation operator acts on a single child, introducing
random modifications that are independent of the parents’
genetic material. In the next generation, the children become
parents, and a new child population is created. The process
continues during a given number of generations, or until
some other criterion is satisfied. This artificial reproduction
of individuals simulates what happens with chromosomes in
real world reproduction.

II.2 GAs for optimization


Figure 1. Overall structure of the proposed The population of individuals is randomly initialized with a
approach set of feasible solutions of the given optimization problem.
Load restoration strategies: T. Kostic et al. 133

Each individual is a coded representation of a feasible commands (e.g. put a branch in service while it is already
solution (for instance, an acceptable sequence for picking up in service), are skipped.
load). The choice of the coding strongly depends on the
problem that is to be solved. The information contained in III.2 Execution flow
the individual must be later decoded to evaluate its fitness;
When a command is entered, the dynamic part of the DS
the evaluation function captures the objective function (in
computes the response of the power system. The evolution of
this case, maximize the picked up load) and the constraints of
the average system frequency and the active power output of
the problem (frequency, voltage and current limits, etc.).
generators are simulated until the steady state is reached.
The search for the solution is carried out through the
This state is quasi-steady, since it has been obtained with
reproduction process described above. GA optimization
power losses corresponding to the power system state before
explores a problem’s space of solutions by random search,
not across the whole solution space, but only through the the execution of the command. The load flow is then used to
compute the system voltages, reactive power of units and
regions that are chosen considering the value of the objective
actual power losses. Finally, the dynamic part is executed
function.
again to take into account the actual power losses, computed
by the load flow. When the actual steady state is reached, the
DS is ready to process the next command.
III. The dynamic simulator The power system is considered to be in steady state if the
The selection of LR sequences must take into account the frequency is constant during ten consecutive integration
power system dynamic response to avoid violations of the cycles, and in unstable state if the frequency is out by
operating constraints (frequency, voltage, power flow) 62% of the nominal frequency (50 Hz). These two condi-
during the load restoration. This response is used to evaluate tions are tested after each integration cycle to control the
the sequences and select the most appropriate ones. A power simulation. As said above, when the steady state is reached,
system dynamic simulator [4,5] is used for these purposes. either the load flow is run, or the next command is expected.
The dynamic simulator (DS) is able to analyze the power If the state is found to be unstable, the execution of the DS is
system in different operating states, particularly during interrupted.
service restoration. It can simulate the response of the
whole network, or its islands, to different LR actions.
Given the need to simulate an important number of restora- IV. Optimization of load restoration sequences
tion scenarios, speed has been the main design requirement, GA optimization has been applied to generate load restora-
leading to the use of simplified models of the power system tion sequences [5]. Each individual in the population is a
components. coded representation of a possible solution, i.e. an LR
Mechanical aspects of production units (prime mover, sequence. The ‘‘quality’’ of a sequence depends on the
speed governor) are dynamically simulated thanks to a evolution of the power system during its execution, that is
model of the long-term dynamics. The electrical model of obtained by the DS. Therefore, the DS decodes each indivi-
production units (alternator, exciter, governor) has not been dual to obtain a solution (the evolution of the power system
implemented, but the steady state values of system voltages when a load restoration sequence is executed) that can be
and reactive power of units are obtained, in discrete time evaluated. The results from the DS are thus used by the
intervals, from the load flow computation. Thanks to this evaluation function to compute the fitness for each indivi-
decoupled modelling of mechanical and electrical aspects of dual. The execution flow of the GA-based optimization is
the power system, significant savings of processing time are shown in Figure 2.
obtained, while keeping a satisfactory precision.

III.1 Commands
Commands represent LR actions. From the complete set of
available commands, only those that are relevant to the
present work will be briefly described.
• gen set unit: Increase the power set-point of unit to the next
set-point level. For a thermal unit, the power set-point
levels are [20; 50; 90], expressed as a percentage of its
rated power P n; the load increment rate is 0.05 P n/min. For
a hydraulic unit, the power set-point levels are
[15; 30; 45; 60; 75; 90], expressed as a percentage of
its rated power P n; load increment is assumed to be
instantaneous.
• gen vol unit v: The voltage set-point of unit will be
increased by v [p.u.] if v . 0, or decreased by v [p.u.] if
v , 0.
• loa set load p: Pick up p [MW] of load; whenever p MW
are picked up, q MVAR are also connected at the same
bus, being q ¼ (Q i/P i)p, where Qi and Pi are the pre-
blackout loads at the bus.
• bra on branch: Put circuit branch in service.
Figure 2. The GA-based optimization generates an
Before the execution of a command by the DS, different optimized sequence for a given state of the power
tests are performed to determine its validity. Invalid system
134 Load restoration strategies: T. Kostic et al.

IV.1 Initial population [unstab ¼ true] OR [Nfre . 0] OR [Ndiv . 0]


Since this work is focused on the load restoration stage, it
will be assumed that the network skeleton has been con- OR [Nvol . Nldf ] OR [Nloa . Nldf ] ⇒ fitness ¼ 0 ð1Þ
structed with the available units and circuits, (except for the Otherwise, the fitness is computed as:
second circuits in double circuit lines). Voltage set-point of
units is 0.95 U nom, and the power set-point is 0 MW. There- Load
fitness ¼ Fstat 3 Fdyn 3 (2)
fore, the actions to be executed (LR actions) consist in 10
picking up load, increasing the voltage and power set- where
points of units and energizing those circuits that were not (A 3 Nvol þ B 3 Nloa )
connected in the skeleton. The list of possible actions stores Fstat ¼ 1 ¹ , AþB¼1 (3)
Nidf
all the LR actions. The initial population is composed of
sequences, each of which is randomly filled with actions Ifre þ þ Ifre ¹
from the list of possible actions. Fdyn ¼ 1 ¹
Ifre
For each load bus L, there are three actions: loa set L 100,
loa set L 75, loa set L 50. For each unit G, there are two Load being the percentage of the pre-blackout load that has
actions: gen set G, gen vol G 0.02. For each double circuit B, been picked up, and A and B representing the penalty factors
there is one action: bra on B. Each action has an associated with respect to voltage limit violations or overloads. In this
probability of being included in an individual of the initial work, A ¼ 0.5, B ¼ 0.5.
population. This allows higher priorities to those actions that F stat and F dyn can vary from 0 to 1, and Load from 0 to
are known to be more frequently found in successful LR 100%. The factor ten is introduced for the sake of readability
sequences, thus reducing the expected duration of the GA during simulations. Therefore, the maximal possible fitness
search. is ten, corresponding to the case where there are no voltage,
The maximal length of an individual is greater that the load and frequency violations, and the total load is restored.
expected length of an acceptable sequence. Redundancy of The load restoration time is not considered: it has been
actions that compose an individual is needed to reach 100% noticed that, for similar power system states, all the opti-
of restored load. In all cases, it is up to the DS to reject an mized LR sequences have quite similar duration.
action if it is not possible (e.g. if load L1 is completely
restored, any eventual action concerning this load will be IV.3 Strategy and operators
rejected). The strategy applied is reproduction without duplicates.
Only individuals with positive fitness value (see next In addition, a certain number (E ¼ 2) of best ranked
section) are accepted in a population, to speed up the GA individuals are kept and copied in the next generation
search. This constraint can make the creation of the initial (elitism). All the remaining individuals are replaced with
population from the elementary actions a time consuming children created by using the genetic operators: selection,
task; this problem can be solved by randomly selecting the crossover and mutation. The selection operator used here is
initial population from a previously created set of individuals the standard roulette wheel operator (see reference [2]).
with positive fitness. This set must contain acceptable The crossover operator is the two-point crossover with
sequences for any power system state. variable length of the individual. For each parent (see
Figure 3), two points are chosen, elements between the two
points (e 1 for p 1 and e 2 for p 2) are exchanged to obtain two
IV.2 Evaluation children. This kind of crossover keeps a certain level of
Each individual in the population represents an LR redundancy in the elements of the individuals (LR actions),
sequence. Its fitness is computed from the results of simulat- needed to increase the likelihood of obtaining a successful
ing the application of the sequence to the power system. The sequence. However, the maximal length of the individual is
DS results used as inputs by the evaluation function are the defined and limited by the evaluation function.
following (these values are obtained during the simulation of The mutation operator replaces a single element of a child
a complete sequence): with a randomly chosen element from the list of possible
• unstab: this flag is true if the unstable state has been actions (see Section IV.1).
encountered. When two children are created, their fitness values are
• N fre: number of times the frequency went out of the computed. The children are copied in the next generation
transient limits (62% of f nom). only if their fitness values are greater than zero. Otherwise,
• N ldf: total number of load flows. another couple of parents is selected and crossed over, and
• N div: number of load flows that did not converge. their children evaluated.
• N loa: number of overloaded lines and transformers.
• N vol: number of buses with voltage limit violations (both
N loa and N vol can be greater than N ldf, since several over-
V. Results of the GA-optimization
loads and voltage violations can be found in a single load A large number of optimizations have been carried out to
flow). adjust the GA parameters as follows:
• I freþ: integral of those parts of the frequency that are above
its upper steady state limit (50.5 Hz)
• I fre¹: integral of those parts of the frequency that are below
its lower steady state limit (49.8 Hz)
• I fre: total integral of absolute frequency deviation with
respect to 50 Hz
A first test is performed to reject a sequence if significant Figure 3. Two-point crossover with variable length
violations are found: of the individual
Load restoration strategies: T. Kostic et al. 135

• The number of individuals in a generation is 30.


• The maximal length of an individual is 500.
• The probability of crossover, P c, is set to one.
• The mutation probability P m is a function of the present
generation gen and the maximal number of generations
MaxGen:
gen
¹4
Pm ¼ 0:1 þ 0:4e MaxGen (4)

• A nearly maximum value of the fitness (10 . fitness .


threshold) is usually reached during the first 5–15 gen-
erations, except for certain initial states (near operating
limits). Therefore, the maximum number of generations is
fixed at 20.
• The optimization process is stopped if the fitness is greater
than a threshold (9.5) and if it has not changed by more
than 0.05 during three generations.

Figure 5. Evolution of the frequency during the


V.1 Test power system
application of an optimized load restoration
Tests have been executed on the EPRI New England system: sequence
39 buses (with 19 load buses), ten generators (three hydraulic,
seven thermal, total installed capacity is 7860 MW), 35
lines, 11 transformers and a total load of 6150 MW. It must sequences can be obtained for the same power system state.
be pointed out that service restoration plans are frequently The generated optimal sequences have been analyzed and
prepared at regional level, involving power systems of a compared, and it has been concluded that it is almost
similar size. impossible to deduce generic rules that could describe the
strategy for the second restoration stage (i.e. load pickup) for
V.2 Results
the sequences as they are defined at present. In addition,
A single run of the optimization, with the above defined since rules cannot be extracted, it is impossible to give any
parameters but without the stop criterion, is shown in explanation about the order of actions in the sequence, even
Figure 4. The evolution of the frequency when the optimized when the sequence leads to a satisfactory result. To over-
sequence is executed is depicted in Figure 5. come this problem, the following sections propose a method
The duration of the optimization process is about 1 h to classify the optimized sequences, generated by the GA,
(although it strongly depends on the state of the power according to different power system variables to obtain a
system). The simulation of an LR sequence by the DS decision tree.
takes about 5 s. The experiments have been carried out on
an HP 735/125 workstation.
VI. Decision trees
V.3 Conclusions
The procedures described in the previous sections can be
After many runs of the GA it has been observed that, given used to generate a set of power system states (defined by
the random nature of the GA, different equally acceptable variables such as loads, availability of generators and lines,
etc., that will be called attributes), with their associated LR
sequences (that will be called classes). Provided that this set
is statistically representative of the space of possible power
system conditions, an induction procedure (such as an
algorithm for decision tree generation) can be used to
identify which attributes of the power system state (i.e.
which of the variables) are relevant to the classification
(i.e. to the selection of an LR sequence).
VI.1 Overview of decision trees
A decision tree (DT) is a set of ordered tests that, when
performed on the attributes of a new item, allow one to
assign a class to it. Reference [6] provides an excellent
survey of decision trees and other machine learning techni-
ques. The application of this technique to security assess-
ment is described in [6,7]. There has also been some work on
the application of DTs to learning sequences of events in
network operation [8].
The core of the decision tree approach is the organization
Figure 4. Evolution of the fitness value for a GA- of the attributes in a tree structure in which the more relevant
based optimization. There are 30 LR sequences in an attribute is, the closest to the root it is placed; the tree will
each generation. The stop criteria is the maximal be used to determine which attributes must be checked to
number of generations classify a new situation.
136 Load restoration strategies: T. Kostic et al.

VI.2 Decision tree construction procedures


Three procedures for DT generation have been used in this
work [9]. The main difference between them is the measure
used to select the most relevant attribute.
All of them keep on expanding nodes until a given purity
threshold is reached. Ideally, the process of selecting
adequate attributes and splitting nodes should continue
until, for every leaf of the tree, all the examples belong to
the same class. Nevertheless, in most realistic applications,
this approach would lead to enormous trees with a few
examples in each leaf and poor performance. This phenom-
enon is similar to the overfitting problem that is found
whenever a polynomial of an excessive degree is used for
curve fitting. Therefore a stop criterion or termination
condition must be used to decide whether decomposing a
node is worthwhile.
• The ID3 algorithm [10] uses entropy as a measurement of
the impurity I(N) of a node N:
X n n
Ni
I(N) ¼ ¹ log Ni (5)
i ¼ 1, M n N nN
Figure 6. Generation of a decision tree
where M is the number of classes, n Ni is the number of
elements of N belonging to class i, and n N is the total
The construction of a DT is based on determining which is number of elements of N. The separator that provides
the more relevant attribute for the classification process. the maximum decrease of impurity will be selected. The
The process starts with an initial training set where there is a decrease of impurity provided by a separator S will be
mixture of items belonging to different classes (i.e. the computed as:
power system states with different associated sequences). X nNj
The relevance of an attribute is related to its capacity to DI(N, S) ¼ I(N) ¹ I(Nj) (6)
purify this set, i.e. to separate examples belonging to j ¼ 1, H nN
different classes (therefore, the relevant attributes are also where the separator S splits the set N into subsets
called separators). An example of a separator could be N1,...,Nj,...,NH.
[TotalLoad , 6000 MW]. This separator would split the set The stop criterion that has been used in the ID3 imple-
of examples into two subsets: one for heavily loaded pre- mentations is the comparison of the percentage of the most
blackout conditions and the other for lightly loaded situa- numerous class with a threshold.
tions. • The ULg method [6,7] defines a normalized version of the
A DT can be seen as a procedure to split the initial set of purity increment equation (5):
examples (a mixture of individuals belonging to different
classes) into subsets of increasing purity. 2DI(N, S)
DI9(N, S) ¼ X nNj nNj (7)
The most widely used procedure for the construction of I(N) ¹ log
decision trees is the following (see Figure 6): j ¼ 1, NH nN nN
(1) The initial set of examples is associated to a node (root The ULg method makes use of two stop criteria. The first
node). one is a threshold value for the entropy I(N). Besides, it
(2) A node must be expanded unless it is a leaf node (it has uses a statistical test (x 2) to decide whether the expansion
an admissible purity grade or is not considered to be of a node is expected to increase the tree reliability. The
statistically significant). test is controlled by the a parameter, that measures the
(3) The best separator is selected and applied to separate probability of expanding a non-significant node. Low
the set of examples associated to the node, into two or values of a will provide very cautious expansion proce-
more new nodes, which will be eventually expanded. dures and simpler trees, while high values will lead to
The purity of the subsets that result from each split is the more complex trees, where more nodes will be expanded.
criterion to select a separator. The purity can be eval- • The Fayyad algorithm [11] uses a different family of
uated using different measures, that will be introduced in measures, aiming at selecting separators that are able to
Section VI.2. Usually, there exist different possible isolate a class as early as possible during the tree con-
separators for a set. struction. It has been implemented with the same termina-
(4) Steps 2 and 3 are repeated until all nodes have been tion condition used by ID3.
expanded.
A set of different examples (known as the test set) must be
created to evaluate the quality of the DT (i.e. its ability to VII. Application of decision trees to load
correctly classify ‘‘new’’ examples, that have not been restoration
considered during the tree generation). The application of As stated in Section V.3, many different sequences for
the tree consists in carrying out the requested tests on the picking up load, leading to equally acceptable results
attribute values according to the order stated in the tree, until (since they have high fitness values), can be obtained for
a leaf is reached. The leaf will have an associated class (LR the same pre-blackout conditions. This high number of
sequence). different, good solutions makes the direct application of
Load restoration strategies: T. Kostic et al. 137

(2) A new power system state is randomly chosen. All


sequences in the list of adequate sequences are tested
on this new state, and the resulting fitness values are
computed.
(3) If none of the sequences provide a high enough fitness,
the GA is applied to the power system state. The
obtained sequence is added to the list of adequate
sequences. If the GA-optimization is not able to return
a solution, it will be assumed that the power system state
is out of the operation region that has been considered of
interest.
(4) Steps 2 and 3 are repeated until the list of adequate
sequences includes enough sequences to (roughly)
cover all the feasible states of the power system. This
condition holds when
of new sequences found in last N cycles
,e (8)
N

(5) A new power system state is randomly chosen. All the


sequences in the list of adequate sequences are tested
on this new state, and the resulting fitness values are
computed.
(6) If any of the sequences provides a fitness greater than a
given threshold, it is selected as an adequate solution for
the new state of the power system. The variables defin-
ing the power system state and the sequence are added to
the training set.
(7) Steps 5 and 6 are repeated until the desired number of
elements in the training set is reached.
The purpose of this procedure is to identify regions in the
Figure 7. The use of the GA-optimization and the space of power system states in which the same LR sequence
dynamic simulator in the generation of the training is adequate. These regions should be as big as possible, to
set minimize the number of different sequences to be used in the
learning procedure.
The power system states are defined in terms of a set of
the learning algorithm impossible, since it could be dealing variables: availability of generators, lines and transformers
with as many different classes as examples. (in the network skeleton), pre-blackout load at each bus and
Regarding now a single LR sequence, it has been con- total pre-blackout load.
cluded that the sequence might be adequate for several The population of power system states will be randomly
power system states, thanks to redundancy of the actions it generated. Each of the variables will have an associated
contains (see Section IV.1). For instance, an action (e.g. pick distribution probability. Nevertheless, the use of actual
up 50 MW more of a load) which is invalid for one power probabilities would lead to an insufficient presence of
system state (lightly loaded) may be valid for another one ‘‘critical’’ power system states (those with unavailable
(heavily loaded). generators, since these problems have low probability of
The fact that the same LR sequence can be associated to occurring) in the training set, and would decrease the ability
different power system states is the basis of the procedure of the DT to treat such situations. Thus, the used probabilities
which generates the training set. must ensure that an adequate coverage of the ‘‘critical’’
states is provided.
VII.1 Generation of the training set
Different LR sequences can be proved adequate for the same VII.2 Validation of the decision tree
power system state. Moreover, the GA-based optimization
The validation of the DT must be carried out using a second
procedure does not guarantee the generation of all the
set of power system states, the test set, that is not used for
adequate sequences, but only a few of them. Nevertheless,
training purposes (Figure 8):
a sequence obtained for a given power system state could be
appropriate for another condition of the power system. This • For each example in the test set the DT is used to
fact is the basis of the training set generation procedure. determine a sequence for picking up load.
The approach tries to minimize the number of executions • The fitness value corresponding to each power system
of the GA-based optimization (since it is the most time state and its associated sequence is computed, using the
consuming task), while ensuring a maximum coverage of the dynamic simulator.
space of power system states. This goal is achieved as • If the fitness value is greater than a given threshold, the
follows (see Figure 7): classification is successful.
• The percentage of successful classifications provides a
(1) A power system state is randomly selected. The GA-
measure of the quality of the decision tree.
optimization generates an LR sequence, that is included
in a list of adequate sequences. It must be noticed that this validation approach is slightly
138 Load restoration strategies: T. Kostic et al.

appropriate sequence for each power system state). Never-


theless, since in this problem many different classes
(sequences) are roughly equally appropriate, the previous
class assignment makes no sense: its results would probably
never agree with the output of the DT. The only valid method
is based on analyzing the DT results using the same evalua-
tion function (the dynamic simulator) that was used to
generate the examples.
The results of the testing process can be used to adjust
several parameters:
• The number of elements in the training set.
• The fitness thresholds beyond which a given sequence is
considered to be acceptable.
• The procedure for decision tree construction and its
associated parameters (end node purity, a, etc.).

VIII. Results
The GA–DT approach has been tested on the EPRI New
Figure 8. The validation of the decision tree is England network (see Section V.1):
based on analyzing its outcome with the dynamic • 1000 power system states were analyzed. Total pre-black-
simulator out loads were between 5280 and 7050 MW. The unavail-
ability of two units was simulated. Circuit unavailability
different from the one usually reported in the literature. In was not considered. The e parameter (see equation (8) in
the usual approach, the validation consists of comparing the Section VII.1) was equal to 0.03. 19 adequate LR
DT results with a previous classification of the test set sequences were found.
elements. This approach is possible when there is only one • Since most of the power system states have more than one
possible class for each element in the test set (i.e. only one associated sequence, the training set has 7181 elements.

Table 1. Comparison of several DTs


Algorithm ID3 ULg a ¼ 0.0005 ULg a ¼ 0.04 Fayyad
Purity threshold 20% 0.6 0.6 20%
Number of nodes 1313 53 119 1595
Fitness [ [0] 2.90% 0.00% 0.00% 2.60%
Fitness [ [3,8] 1.00% 0.00% 0.40% 3.30%
Fitness [ [8,9] 6.50% 3.40% 3.80% 7.50%
Fitness [ [9,10] 89.60% 96.60% 95.80% 86.60%

Figure 9. This DT has been created using the ULg algorithm, with a ¼ 0.0005 and an entropy threshold of 4.0.
The size of each node corresponds to the number of associated elements of the training set. The shading of a
node corresponds to its associated LR sequences (0, 1,...,18). Separators (section 6.1) are shown under their
nodes (g9 is the availability of unit nine). The numbers close to each node (2294, 4887, etc.) are the numbers of
associated training set examples. ND:2 and ND:3 are node identifiers
Load restoration strategies: T. Kostic et al. 139

• The test set included 1000 elements.

VIII.1 Validation of the trees


The three decision tree construction algorithms described in
Section VI.2 were used. All of them provided results of
similar quality: the ID3 and ULg trees were quite similar
(since both algorithms use similar criteria for attribute
selection); Fayyad gives completely different trees, since it
is based on a different approach.
The number of nodes is a measure of the complexity of the
DT. Both ID3 and Fayyad are very sensitive to changes in the
threshold that is used as stop condition; for instance, ID3
trees can range between 3 and 1313 nodes for a small
variation of the threshold (between 11 and 20%) with a
significant deterioration in performance for trees that are too
small or too large. On the other hand, ULg (because of its
additional termination condition) consistently provides trees
with a smaller size, tipically between 40 and 70 nodes.
Table 1 shows the main characteristics of some decision
trees that have been obtained. The table also presents the Figure 11. The use of the decision tree in a realistic
fitness distribution of the LR sequences suggested by the DT load restoration environment
for the elements of the test set. The purity measure used as a
stop condition has been implemented in different ways for
each algorithm: for ULg, the entropy is used, while for ID3 IX. Conclusions
and Fayyad a simplified purity measure (defined as the The first tests confirm the potential applicability of this
percentage of the most numerous class) has been selected. approach to the selection of LR sequences. Decision trees
Figure 9 shows an example of decision tree. constructed with training sets of reasonable size are able to
select high quality load restoration sequences for a wide
VIII.2 Analysis of results range of power system states. However, these preliminary
More detailed results are presented in [9]. Several conclu- results are to be confirmed with more extensive analysis,
sions can be drawn from the analysis of the experiments involving the complete simulation of unit and circuit una-
carried out so far: vailability.
• A small number of LR sequences cover a wide region of IX.1 Possible use in a real environment
power system states. The DT would be constructed and tested off-line, using
• A relatively small training set provides a satisfactory real data from the power system. It should be updated
accuracy. whenever a significant change occurs in the topology,
• However, the problem is too complex for a single attribute installed capacity or load profile. When a blackout occurs,
to be sufficient to select the appropriate sequence. For the decision tree should use the available data to select an
instance, see the frequency distribution of LR sequences LR sequence.
for two different availability states of unit nine (Figure The sequence is used to propose actions to the operator.
10): although this is the most important attribute for ID3 Before suggesting each operation, a final verification is
and ULg, it is not sufficient for selecting a sequence. performed using the available data and the DS (Figure 11).
• The DTs provide useful criteria: for the best obtained The application of the sequence is modified according to the
decision tree, 96.70% of its suggestions have a fitness results of the verification. For instance, if the simulation of a
beyond nine, while 100% of the solutions have a fitness ‘‘connect load’’ operation results in an unacceptably low
beyond eight. On the other hand, the random selection of a value of the frequency (because some load has been
sequence for each power system state gives only 64% of unexpectedly connected), the operation is delayed. The
results beyond fitness nine, and 9.5% of results with first ‘‘change generator set-point’’ in the sequence is imme-
fitness 0. diately suggested instead of the ‘‘connect load’’ manoeuvre.
These tests are intended to ensure that unexpected changes in
the power system conditions are taken into account, and the
selected sequences are not blindly applied.

IX.2 Further work


The application of this approach requires additional research
in several principal directions, such as:
• The fitness function must be improved, to capture actual
operators’ criteria.
• The decision tree must be updated to deal with the
continuous evolution of the power system (new com-
ponents added, new generators and loads, etc.).
• The LR selection criteria can be arranged in a hierarchy of
Figure 10. Frequency distribution of LR sequences DTs, to deal with more complex power systems without
for two availability states of g9 building a single, huge tree.
140 Load restoration strategies: T. Kostic et al.

X. Acknowledgements system restoration. Faculty of Electrotechnics, Belgrade University,


Belgrade, 1994.
The authors want to express their thanks to Rachid 5. Kostic, T., Decision aid function for restoration of transmission
Cherkaoui, for his extensive support in the development power systems after a blackout. Département d’électricité, Laboratoire
and testing of the simulator, as well as to Eugenio Sánchez, de réseaux d’énergie électrique, École Polytechnique Fédérale de
who developed the software for decision tree construction Lausanne, Lausanne, 1997.
6. Wehenkel, L., Machine learning approaches to power system security
and assisted in its use. This work has been supported by assessment. Faculté des Sciences Appliquées, Université de Liège,
the Swiss National Foundation for Energy Research (NEFF), Liège, 1994.
Électricité de France (EDF–DER) and the Electric Energy 7. Wehenkel, L. and Pavella, M., Decision tree approach to power system
Systems UETP. security assessment. International Journal of Electrical Power and
Energy Systems, 1993, 15, 13–36.
8. Ypsilantis, J. and Yee, H., Machine learning of sequences of events:
application to power distribution. Electric Power Systems Research,
1992, 23, 115–122.
XI. References 9. Alba, J. J., Switching planning in electric power systems: application of
1. Ancona, J. J., A framework for power system restoration following a knowledge engineering techniques. ETSII, Departamento de Electro-
major power failure. IEEE Transactions on Power Systems, 1995, 10, tecnia y Sistemas. Universidad Pontificia Comillas, Madrid, 1996, p. 172.
1480–1485. 10. Quinlan, J. R., Induction of decision trees, In Machine learning, An
2. Goldberg, D. E., Genetic algorithms. In Search, Optimization and Artificial Intelligence Approach, Vol. 1, ed. R. S. Michalski, J. G.
Machine Learning. Addison-Wesley, Reading, MA, 1989. Carbonell and T. M. Mitchell. Morgan Kaufmann Publishers, Los Altos,
3. Davis, L., Handbook of Genetic Algorithms. Van Nostrand Reinhold, CA, 1986, pp. 81–106.
New York, 1991, p. 385. 11. Fayyad, U. M. and Irani, K., The attribute selection problem in decision
4. Kostic, T., Dynamic simulator as a part of an expert system for power tree generation. Presented at AAAI, San Jose, CA, 1993.

Вам также может понравиться