Metaheuristics: Computer Decision-Making, pp.

449-464
Manuel A Matos, M. Teresa Ponce de Leao, J. Tome Saraiva, J. Nuno Fidalgo,
Vladimiro Miranda, J. Pes;as Lopes, J. Rui Ferreira, Jorge M. C. Pereira,
L. Miguel Proens;a, and J. Luis Pinto
©2003 Kluwer Academic Publishers B.V.

21 METAHEURISTICS APPLIED TO
POWER SYSTEMS
Manuel A. Matos 1 • 3 ~ M. Teresa Ponce de Leao1 •3 1

J. Tome Saraiva 1 •3 J. Nuno Fidalgo1 •3 Vladimiro Miranda 1•3

I I I

J. Pe~as Lopes 1 • 3 ~ J. Rui Ferreira 1 •3 Jorge M. C. Pereira 2 •3

1 1

L. Miguel Proen~a 3 1 and J. Luis Pinto3

1 FEUP- Faculty of Engineering

University of Porto
Porto, Portugal
mmatos42inescporto.pt

2 FEP- Faculty of Economics

University of Porto
Porto, Portugal
j pereira 42inescporto. pt

3 1NESC
Porto
Power Systems Unit
Porto, Portugal
lproenca42inescporto.pt

Abstract: Most optimization and decision problems in power systems include

integer or binary variables, leading to combinatorial problems. In this paper, sev-
eral approaches using metaheuristics and genetic algorithms are presented that
deal with real problems of the power industry. Most of these methodologies are
now implemented in distribution management systems (OMS) used by several
utilities.
Keywords: Metaheuristics, Power systems, Simulated annealing, Genetic algo-
rithms.
450 METAHEURlSTICS: COMPUTER DECISION-MAKING

21.1 INTRODUCTION
This paper describes a number of real applications of metaheuristics (in this
case, simulated annealing) and genetic algorithms to power system problems.
The research work was developed in the framework of European projects and
industrial contracts and addresses areas like planning and operation of electri-
cal distribution systems, wind park layout, unit commitment of isolated sys-
tems with renewable energy sources and voltage collapse in systems. The
combinatorial nature comes naturally in power systems, since most of the de-
cision variables are binary or integer due to technical reasons. On the other
hand, a common characteristic of these problems is the presence of technical
constraints, which poses difficulties to the application of metaheuristics, lead-
ing to the need of penalty factors in the evaluation functions. The paper also
includes feature selection for security analysis using artificial neural networks,
a related topic, although not really an application of metaheuristics. The paper
is organized as follows. Regarding each topic, the corresponding problem is
briefly described, followed by the presentation of the approach and, in some
cases, a summary of the results. Global conclusions and references complete
the paper.

21.2 DISTRIBUTION SYSTEMS PLANNING

21.2.1 Problem description
A main issue in planning electric distribution systems is to decide what new
branches and injection points (substations) to construct. The decisions must
be taken to meet future demand at minimum cost, without violating technical
constraints (maximum current flow and maximum voltage drop). These net-
works are characterized by a great number of nodes and possible branches.
The nodes are the injection points of the network, consumer points (loads),
substations and generation from independent producers. The problem in-
volves many integer-valued (0-1) variables related to the decisions to build
or not build facilities. In addition, the overall problem is dynamic, since the
decisions about investments must be made over a multi-period horizon. Be-
sides investment cost, other criteria are generally considered (operating cost,
reliability) and uncertainty is present due to forecasting.

21.2.2 Approach
In the simulated annealing strategy, the analogy between the electric distribu-
tion planning problem and the Metropolis algorithm can be stated as follows.
The alternative solutions or configurations of the combinatorial electric distribu-
tion planning problem are equivalent to the physical system states. The net-
work configurations (alternative solutions) (attributes) are equivalent to the
different energy states. The control parameter, which is such that about half the
new configurations found are accepted at the start of the process, is equiva-
 METAHEURISTICS APPLIED TO POWER SYSTEMS 451

lent to the temperature parameter. The SA cooling schedule integrates several

aspects that are detailed in the next paragraphs:
• The choice of the initial temperature was such that about half the config-
urations were accepted;

• The number of iterations during which the temperature is kept constant

was approximately 10 times the average number of possible neighbors.
This value can be obtained from initial temperature calculations where
the possible neighbors are explored;

• A stepwise temperature reduction scheme with T(i) = T(i- 1)//3 was

used;

• From our experience an average value for fJ was derived between 0.85
and 0.92. Larger values yield high computation times, and smaller val-
ues, below 0.85, lead to poorer solutions;

• The freezing point is determined where the number of acceptances is

very small. This point is highly dependent on the problem size and must
be determined for each problem.
The initial solution is generated by an auxiliary algorithm that finds a short-
est spanning tree, the weights used for the edges usually being the edges
(branches) costs although other criteria could be used. A structure of neigh-
borhood generation mechanism is then created. To eliminate a great number
of trials only feasible solutions S' from the solution space S are considered.
This procedure consists of choosing all the possible combinations of connected
trees that result from removing a branch to the loop created by each edge of
the co-tree entering the initial configuration. The alternatives are evaluated
according to the criterion to be optimized.

21.2.3 Results
A simulated annealing procedure was combined with the €-constraint search
to form a methodology for solving the multi-objective planning problem. In
practice, the initial multi-objective problem was split into several single objec-
tive problems, successively solved to generate non-dominated solutions. In
Ponce de Leao and Matos (1999), the detailed procedure to generate the set of
efficient solutions is explained in detail.

21.3 DISTRIBUTION SYSTEMS OPERATION - RECONFIGURATION

21.3.1 Problem description

Distribution networks are usually operated radially, but they have a meshed
structure (with open branches) that allows several configurations. The recon-
figuration of these networks by changing the switching devices statuses aims
452 METAHEURISTICS : COMPUTER DECISION-MAKING

at optimizing the network configuration for a given load profile, by minimiz-

ing the power losses. Figure 21.1 shows a typical instance of this problem,
where we can see the selected trees (solid) and the potential branches, not used
in the final solution. A related problem is the service restoration problem, in
the sequence of a faulty situation, when one tries to re-supply at least part
of the loads. Besides the main objective of each problem (minimizing losses
or minimizing load not supplied), the minimization of the total number of
switching operations is also included as a criterion. In either case, the net-
work reconfiguration is a multi-objective optimization problem. An algorithm
to deal with this problem is detailed in Matos and Melo (1999).

Figure 21.1 Distribution network reconfiguration problem.

21.3.2 Approach
The basic simulated annealing algorithm was used for this problem, with the
following characteristics:

• When restoring service, the initial configuration represents the distribu-

tion network with the fault isolated and some unsupplied consumers.
When minimizing losses, the initial configuration is the present status of
the network;

• The neighborhood structure implemented depends on the problem. For

the loss minimization problem, when a branch is Opened another one
is Closed, in such a way that the network remains radial and all the
consumers remain served. For the service restoration problem, the rule
is as follows: Close a branch that close no loops or Open any branch;

• The evaluation includes a cost attribute (losses or power not supplied,

depending on the problem) and penalties related to overload limits and
maximum voltage drop. For the multi-objective case, a penalty rela-
tive to the maximum number of allowed switching operations is also
included. Variation of this limit allows the generation of all the efficient
solutions.
 METAHEURISTICS APPLIED TO POWER SYSTEMS 453

21.3.3 Results
Figure 21.2 shows the efficient solutions for a restoration situation. Note that
similar solutions in the attribute space may be very different in the decision
space. As can be observed, the power not supplied decreases as the number
of switching operations increases. The operator may now choose among the
given set for the most satisfying solution, according to his preferences and
temporary constraints not included in the model.

12

a 10
·JJ
; 8
• •
~
i 6 • •
i'a 4 •
z..
• •
•
2

0
0 2 4 6 8 10
Power Not Supplied (MW)

Figure 21.2 Efficient solutions for the restoration problem

21.4 DISTRIBUTION SYSTEMS OPERATION - REACTIVE POWER

CONTROL
21.4.1 Problem description
In this section, we present a model to identify optimal operation strategies of
distribution networks, considering that one wants to minimize active losses.
This objective can be achieved by adequately selecting transformer taps and
sections of capacitor banks that are in operation if one assumes that the topol-
ogy in operation is fixed. To be more realistic, the model allows the specifica-
tion of voltage ranges for nodes and overload limits for branches.

21.4.2 Approach
The problem of minimizing active power losses can be formulated by (21.1-
21.2). In this formulation, h() represents the AC power flow equations, t 1
stands for the available values of transformer taps, ec represents the available
reactive power steps of capacitor banks, V are the voltage magnitudes, I are
the branch currents, and gij corresponds to the conductance of branch ij. Tra-
454 METAHEURISTICS: COMPUTER DECISION-MAKING

ditionally, this problem has been solved assuming that transformer taps and
sections of capacitor banks are represented by continuous variables. Such for-
mulation consists of a nonlinear optimization problem which can be solved
using, for instance, gradient methods. The solution obtained this way must be
rounded to the closest discrete solution, which may, or may not, be the optimal
solution.

Minimize Z = Enr 9ij · (V? + 'V;2 - 2 · Vi · Vj · cos (Jij) (21.1)

subject to h(V, 9, tf, ec) = 0 (21.2)
Vmin ~ V ~ Vmaz (21.3)
J.. I >.: IJ-IJ.. ,maz
I-IJ ~
(21.4)
t1 E {tJl,tf2,tJa, ... } (21.5)
BeE {Scl!Sc2,Sca, ... } (21.6)

The application detailed in Pereira et al. (1999) uses simulated annealing

to identify a strategy to minimize the active power losses. The neighborhood
of a solution, considering all the possible combinations of transformer taps
and capacitor sections, includes all the configurations that differ from the cur-
rent one by a tap increase or decrease. The evaluation function integrates the
active losses for the present configuration and penalties related to violations
of the maximum and minimum voltage magnitudes in each node and of the
maximum current flow in each branch.

21.4.3 Results
Tests with the IEEE24 system (24 nodes, 31lines, 5 transformers) were con-
ducted considering 11 taps in each transformer. Each tap corresponds to a
0.01 increment within the interval [0.95,1.05]. Initially, the transformer taps
were considered at the nominal values (position 6). The selected simulated
annealing parameters were:
• the number of iterations for the same temperature level is 25;
• the initial temperature level was set to 1.0. The lowering step determin-
ing the cooling scheme (parameter (3) is 95% of previous temperature;
• the maximum number of iterations without improvement of the evalua-
tion function is set to 75;
With these parameters, the algorithm converged in 312 iterations and the tem-
perature was reduced till 0.54. Figure 21.3 shows the evolution, along the
search process, of the present evaluation function value (continuous line) and
for the best solution found till the current iteration (dotted line).
To evaluate the efficiency of the search process, we simulated all the pos-
sible combinations for the 5 transformer taps. This enumeration process re-
quired 161051 power flow studies, to obtain the true optimal solution, that
 METAHEURISTICS APPLIED TO POWER SYSTEMS 455

138000

0 100 200 300

Iteration number

Figure 21.3 Evolution of the evaluation function for the current solution (continuous line)
and for the best-identified solution (dotted line) for the 24 node network.

happened to be the same that was obtained with our approach, with only 312
power flow studies. Additional tests with a larger distribution system (645
nodes, 59 million solutions) showed also a good performance, with conver-
gence obtained after only 256 iterations.

21.5 WIND PARK LAYOUT

21.5.1 Problem description
In a wind park the location of wind generators is defined having in mind the
wind and terrain characteristics. The internal electrical network of the wind
park is a problem that involves the network topology determination and the
definition of the cable sections to be used. The solution of this problem can
be obtained by solving a global optimization problem that involves minimiza-
tion of investments and minimization of operation costs. If reliability is taken
into account in the design procedure, an evaluation of the possible benefits of
supplementary connections that provide alternative paths for the delivery of
the energy is performed.
This optimization problem has combinatorial characteristics, due to the
possibilities of connecting the several wind generators in the field and the type
of discrete cable sections available (Resende, 1999). The identification of solu-
456 METAHEURISTICS: COMPUTER DECISION-MAKING

tions for the problem described was performed using a genetic algorithm that
implements the concept of geographic isolation, based on a niching scheme
(Mahfoud, 1995). The chromosome is a string of bits, each bit having a binary
value (0 or 1). In our problem, this string has 8 bits for each generator, that
represent:
• L 1 to L 4 - type of connections in terms of topology (£1 and L 3 are
mandatory lines, L 2 and L 4 alternative lines, used for reliability pur-
pose);
• Wgui - contains information associated with situations where generator
i is the last in the feeder;

• Seci - 3 bits used to encode the section type of the cable to be used;
The length of each chromosome is therefore 8 · n, where n is the number of
wind generators in the park. Figure 21.4 describes the adopted chromosome
coding.

l•jojoj•jojoj•j•joj•joj•j•jojojoj ....... j•j•joj•j•jojoj•

Wmd •aurat.or 0 Wind.aenentorl Wind ameratorn-1

Figure 21.4 Chromosome coding

The fitness function corresponds to the total cost, but it includes penalty
terms related to the technical constraints, as shown in (21.7), where nf is the
number of feeders, ngu in the number of generators in the end of the feeder,
LNc is the number of lines not connected correctly, Vc and VT are the number
of violations of the voltage profile and of the thermal limits, ec is the cost of
cables (including protections), eJL is the cost of energy losses, eND is the cost
of energy not delivered to the grid, due to internal faults on the wind park,
and kl ... k4 are penalty factors.

FIT = k1·nf+k2·LNc+k3lngu-nj]+ (21.7)

k4(VT + Vc) + ec + eJL +eND·
This approach was tested in the electrical design of some wind parks, namely
in the case of a 10 MW wind farm with twenty 500kW wind generators. This
corresponds to a real wind park presently in operation in Portugal. In the
genetic algorithm approach, the values adopted for the crossover probability
and the mutation probability were respectively 0.8 and 0.04, and the process
was found to converge after 1500 generations. The optimal solutions obtained
are, for the tests performed, systematically better than the ones provided by
the traditional engineering judgment and rules. Figure 21.5 shows two of the
best solutions for this problem.
 METAHEURISTICS APPLIED TO POWER SYSTEMS 457

Figure 21.5 Two of the best possible layouts for the wind park

21.6 PREVENTIVE CONTROL PROCEDURE TO AVOID VOLTAGE

COLLAPSE

21.6.1 Problem description

Most power systems are presently operated close to their operating limits, due
to liberalization and increase in consumption, while economic and environ-
mental constraints have limited the construction of new generation facilities
and lines. Therefore, it is important to evaluate the distance to voltage collapse
(see Figure 21.6) and to identify preventive measures that avoid this problem
(Lemaitre et al., 1989).
The model includes the possibility of modifying the current positions of
transformer taps as well as the steps of capacitor banks or performing load
shedding. It is obvious that load shedding is understood as an end of line re-
source, in the sense that it will only be used if there is no other way of supply-
ing loads even changing transformer taps or capacitor steps. Most of the con-
458 METAHEURISTICS: COMPUTER DECISION-MAKING

i" Distance to voltage ,.,

Voltage collapse

Critical po lnt

Load

Figure 21.6 Voltage collapse situation

trol measures (changes in tap positions, capacitor steps, feeders to be discon-

nected) correspond to discrete variables that lead to a combinatorial problem.

21.6.2 Approach
The mathematical formulation of this problem corresponds to the optimiza-
tion problem (21.8-21.12). The objective function (21.8) integrates a term mea-
suring the distance to voltage collapse (given by the total increase of load ad-
mitted by the system) and a second term where J.t assumes a sufficiently large
value so that load shedding is used only if necessary. In this formulation, the
equality constraints f(X, u) represent the power flow equations, X is the state
vector and u is the vector of control variables. Constraint (21.10) puts a hard
limit on the distance to the collapse, and (21.11) and (21.12) constrain trans-
former taps and capacitor steps to their feasible (discrete) values.

Maximize C = DisLcolp - J.t · L Pshed (21.8)

subject to f(X, u) = 0, (21.9)
DisLcolp ~ E (21.10)
tmin ~ ti ~ tmax (21.11)
Ymin ~ Yi ~ Ymax (21.12)

To avoid local minimum solutions, a deterministic crowding genetic algo-

rithm was used to solve the optimization problem described. The fitness func-
tion corresponds to the distance to collapse (in MVA) plus a penalty factor
proportional to the summation of the load shed in each bus (MVA). Details of
the methodology can be seen in Ferreira (1999).
 METAHEURISTICS APPLIED TO POWER SYSTEMS 459

21.6.3 Results
To evaluate the performance of the previously described approach, we con-
ducted several numerical tests considering the Reduced Mato Grosso system
(Granville et al., 1996). The results obtained were excellent in quality, match-
ing with the ones obtained when the optimization problem in a continuous
space was solved using the interior point method. When curtailment of loads
was needed, the approach adopted was able to provide better results than the
traditional methods, due to the fact that it was dealing with the real speci-
ficity and characteristics of the amount of load supplied by each feeder in a
substation.

21.7 UNIT COMMITMENT

21. 7.1 Problem description
The unit commitment problem is related to the hourly scheduling of the gen-
erators that will be running to satisfy the entire load, in a horizon of sev-
eral hours (at least 48 hours). The problem is not separable in the intervals,
since there are transition costs (due to start-up and shut-down costs) and
constraints, and has been addressed traditionally by dynamic programming
(small problems) and Lagrangian relaxation.

21. 7.2 Appro~ch

The genetic algorithm used to solve the unit commitment problem is based on
the general scheme of genetic algorithm, but includes several enhancements
(Hang, 1999). It replaces the standard selection scheme with a dedicated nich-
ing selection scheme- deterministic crowding selection. In addition, a dy-
namic mutation rate technique, chromosome repair and neighborhood dig-
ging schemes are also included, in order to accelerate convergence and enforce
robustness. Encoding the chromosome plays an important role in the genetic
algorithm. In the UC module, an efficient way is proposed to compress the
length of the chromosome. The main idea comes from the observation that
the generators will not switch on/ off frequently and thus some of the vari-
ables can be reduced. With this method, the length of the chromosome can be
reduced a lot. For example, in a problem with 17 generators and an hourly
scheduling for 48 hours, the direct encoding will lead to 816 bits. On the other
hand, if divide the 48 hours into 16 sections, the length of the chromosome
will only be 368 bits.
The fitness function includes the generation and start-up and the shutdown
costs of generators. To identify an infeasible chromosome a penalty cost is
added when the schedule violates the constraints. To avoid trapping in local
optima, dynamic mutation rate adjustment is used. Deterministic crowding
(DC) selection scheme is a niching method, which is used to preserve the di-
versity of the genetic algorithm. Chromosome repair is a corrective algorithm.
The algorithm will correct slightly an infeasible chromosome, which is near
460 METAHEURISTICS: COMPUTER DECISION-MAKING

. . . . . .. . . . . . . . . ,. • .at
,.rMIItiiA(t • • t *UIIffllf

-·
. . ...... til • • ,. •
. . . . ,.W'I ......... • • • •lt•a
. . . . . . . . . . . . . 11'1 . . . .. . . . . ;
... .,. . ,~ . . . . 111 .

L...-....11'--•J...I_•...._•....1'--___..1_•'-1_...._._,_..1-- -- · I 1·1 ...... 1-- ·--I 1••1 ....... I .... 1- --1 ....

Alll 4ihllllloiM:r.tW!I•

...................,
····~- tdl tl
flwrlt~tttlt&~diU
Plltriufl

Figure 21.7 Compression scheme for the Unit Commitment problem

the feasible region. Afterwards, the corrected chromosome will place it back
to the population and enrich the genetic pool. Neighborhood digging con-
centrates mainly in feasible chromosomes. By searching the neighbor of the
chromosome, the algorithm will determine a better chromosome to replace
the original. By experience, we find that this genetic operator not only accel-
erates the convergence of the process, but also gives better results regarding
both the economic optimization and the execution time, which was an impor-
tant project constraint.

21. 7.3 Results

The implementation of this method lead to very good results regarding both
the economic optimization and the execution time, which was an important
limitation in the project where it was developed. Figure 21.8 shows a typical
solution for a 48 hour problem.

21.8 FEATURE SELECTION USING ANN

21.8.1 Problem description
Feature subset selection (FSS) is a central issue in a vast diversity of prob-
lems including classification, function approximation, machine learning, and
adaptive control. In this study, the selection of adequate variable subset is ac-
complished by sensitivity analysis of trained artificial neural networks (ANN).
Data sets concern to power systems of islands of Crete (Greece) and Madeira
(Portugal). The problem under analysis refers to the dynamic security clas-
sification. This section describes how FSS and ANN tools may be used for
enhancing classification performance. Results attained so far support the va-
lidity of the developed approach.
 METAHEURISTICS APPLIED TO POWER SYSTEMS 461

----......·
300 - - ·-- --·--··- --··- ---· ----- --------- - - . - •.•.

,_
- ·-
•oo
_ ..
=>Oooo.O

- ·· ....
.........
c:::::::::.~~-~

. ~--

"...
c.... "'" "'")

Figure 21.8 Unit commitment for a 48 hour horizon

21.8.2 Approach
The FSS methodology applied in Fidalgo (2001) consists basically of three
steps:

1. Use of correlation analysis to discard linearly correlated features;

2. Train an ANN with the remaining ones, compute sensitivities of its out-
put (dynamic security index) with respect to all inputs and discard in-
puts with lowest sensitivity Si indexes;

3. Train a second ANN using only the remaining variables.

21.8.3 Results
Tests with a data set of the power system of Crete, initially with N1 = 60 at-
tributes, lead to N2 = 20 after correlation analysis. An ANN was then trained
and si calculus were performed, leading to further elimination of 14 variables.
A new ANN was finally trained with the remaining N3 = 6 features. Perfor-
mance was computed and results were compared to the ones reported in Pe<;as
Lopes et al. (1999), were 22 attributes were used, as shown in the Table 21.8.3.
Results show that some real data applications may be amazingly simplified
without loss of performance.
462 METAHEURISTICS: COMPUTER DECISION-MAKING

Table 21.1 Error comparison.

Approximation (MAPE) Classification(%)

old perf. new perf. old perf. new perf.
train test train test train test train test
0.044 0.043 0.033 0.039 99.8 99.7 100 99.8

21.9 CONCLUSIONS
Many power system optimization or decision problems are of combinatorial
nature, due to the intrinsic presence of integer and binary variables. The use of
simulated annealing and genetic algorithms has proved to be successful even
on difficult situations, related to real planning and operation problems. The
examples shown are by no means complete, since many other applications
exist, but are sufficient to demonstrate the practical interest of these method-
ologies to a demanding engineering field.
Bibliography

J. R. Ferreira. Voltage stability evaluation considering fuzzy injection scenarios and

identification of control procedures (in Portuguese). PhD thesis, University of
Porto, 1999.
J. N. Fidalgo. Feature selection based on ANN sensitivity analysis- A prac-
tical study. In International Conference on Neural Networks and Applications
WSES2001, February 2001.
S. Granville, J. C. 0. Mello, and A. C. G. Melo. Application of interior point
methods to power flow unsolvability. IEEE Transactions on PWRS, 11(1),
February 1996.
P. S. Hang. Evolutionary computing and fuzzy logic applied to generation
scheduling in systems with renewables. Master's thesis, University of Porto,
December 1999.
C. Lemaitre, J.P. Paul, J. M. Tesseron, Y. Harmand, andY. S. Zhao. An indicator
of the risk of voltage profile instability for real-time control applications. In
IEEE/PES Summer Meeting. IEEE/PES, 1989.
S. Mahfoud. Niching methods for genetic algorithms. PhD thesis, University of
Illinois at Urbana-Champaign, 1995.

M. Matos and P. Melo. Multi-objective reconfiguration for loss reduction and

service restoration using simulated annealing. In Proceedings of IEEE Bu-
dapest Power Tech'99. IEEE, August 1999.
J. A. Pe~as Lopes et al. On-line dynamic security assessment of isolated net-
works integrating large wind power production. Wind Engineering Review,
23(2), 1999.

Jorge Pereira, J. Tome Saraiva, and M. T. Ponce de Leao. Identification of op-

eration strategies of distribution networks using a simulated annealing ap-
proach. In Proceedings of the IEEE Budapest Power Tech'99. IEEE, August 1999.
463
464 METAHEURISTICS: COMPUTER DECISION-MAKING

M. T. Ponce de Leao and M. Matos. Multicriteria distribution network plan-

ning using simulated annealing. International Transactions on Operational Re-
search, 6:377-391, 1999.
F. Resende. Optimization of the electrical internal network of wind farms.
Master's thesis, University of Porto, December 1999. In Portuguese.