Multi-objective Reactive Power Compensation

Benjamín Barán, Member, IEEE, José Vallejos, Rodrigo Ramos and Ubaldo Fernández, Member, IEEE. 1

Abstract  Reactive Power Compensation in Electric Systems To solve the Reactive Power Compensation Problem, this
is usually studied as a constrained Single-objective Optimization paper presents a new approach based on the Strength Pareto
Problem where an objective function is a linear combination of Evolutionary Algorithm (SPEA) [7], which is a MOEA with
several factors, such as, investment and transmission losses. At an external population of Pareto Optimal solutions that best
the same time, constrains limit other parameters as reliability
and voltage profile.
conform a Pareto Front, provided by a clustering process that
This paper presents a new approach using Multi-objective saves the most representative solutions.
Optimization Evolutionary Algorithms. It proposes a variant of
the Strength Pareto Evolutionary Algorithm (SPEA) that
independently optimizes several parameters, turning most II. MATHEMATICAL FORMULATION
traditional constraints into new objective functions. That way, a
wide set of optimal solutions, known as Pareto set, is found For the purposes of this paper, the following assumptions
before deciding which solution best combines different features. where considered in the formulation of the problem:
Several sets of solutions calculated by different methods are
compared to a Pareto set found with the proposed approach • shunt-capacitor/reactor bank cost per MVAr is the same
using appropriate test suite metrics. Comparison results for all busbars of the power system;
emphasize outstanding advantages of the proposed
computational approach, such as: ease of calculation, better • power system is considered only at peak load.
defined Pareto Front and a larger number of Pareto solutions. Based on these considerations, four objective functions Fi
(to be minimized) have been identified [4, 8]: F1 and F2 are
Index Terms  Reactive Power Compensation, Multi-
related to investment and transmission losses, while F3 and F4
objective Optimization, Evolutionary Algorithms.
are related to quality of service. The objective functions to be
considered are:
I. INTRODUCTION

R EACTIVE Power Compensation is commonly addressed as

a constrained Single-objective Optimization Problem [1-
F1: Investment in reactive compensation devices
n
3]. It basically consists in determining an adequate location
and size of shunt capacitor/reactor banks. In this context, the
F1 = ∑kB
i =1
i
(1)

objective function is a linear combination of several factors,

such as: investment and transmission losses, subject to α if 0 ≤ B i < B m
operational constrains such as reliability and voltage profile k =
 β if - B m < B i < 0
[4]. Traditional Single-objective Optimization Algorithms
usually provide a unique optimal solution. On the contrary, s.t.: F1 ” F1m ;
Multi-objective Optimization Evolutionary Algorithms
(MOEA) independently and simultaneously optimize several where: F1 is the total required investment; F1m is the
parameters turning most traditional constraints into new maximum amount available for investment; Bi is the
objective functions. This seems more natural for real world compensation at busbar i measured in MVAr; Bm is the
problems where choosing a threshold may seem arbitrary [5]. absolute value of the maximum amount of compensation in
As a result, a wide set of optimal solutions (Pareto set) may be MVAr allowed at a single busbar of the system; α is the cost
found. Therefore, an engineer may have a whole set of per MVAr of a capacitor bank; β is the cost per MVAr of a
optimal alternatives before deciding which solution is the best reactor bank and n is the number of busbars in the electric
compromise of different (and sometimes contradictory) power system.
features. This approach has already been treated as a Multi-
objective Optimization Problem (MOP) with two conflicting F2: Active power losses
objective functions [6].
F2 = Pg − Pl ≥ 0 (2)
This work was supported by the National Computing Center (CNC) of the
National University of Asuncion (UNA) – Paraguay.
B. Barán, J. Vallejos and R. Ramos are with the National Computing Center where: F2 is the total transmission active losses of the
of the National University of Asuncion – Paraguay. P.O.Box 1439, San Lorenzo power system in MW; Pg is the total active power generated
– Paraguay. (e.mail: {bbaran, jvallejo, rramos}@cnc.una.py). in MW and Pl is the total load of the system in MW.
U. Fernández is with the National Electricity Administration of Paraguay –
ANDE. (e.mail: [email protected]).

0-7803-7285-9/01/$17.00 (C) 2001 IEEE 97

0-7803-7287-5/01/$17.00 (C) 2001 IEEE
 F3: Average voltage deviation problem consists of all decision vectors B for which the
corresponding objective vectors F can not be improved in any
n
dimension without degradation in another. This set of
∑ V −V i i
*

decision vectors are known as Pareto Optimal, represented as

F3 = i=1 (3)
n P. The corresponding set of objective vectors F calculated
using equations (1) to (4) conform a set known as Optimal
where: F3 is the per unit (pu) average voltage difference; Vi is Pareto Front, denoted PF [7].
the actual voltage at busbar i (pu) and Vi * is the desired Because the true Pareto Optimal Set (termed Ptrue), with its
voltage at busbar i (pu). corresponding PFtrue, are not completely known in practice
without extensive calculation (computationally not feasible in
F4: Maximum voltage deviation most situations), it would be normally enough for practical
purposes to find a known Pareto Optimal Set, termed Pknown,
F4 = max (Vi − V i * ) = V − V ∗ ∞
≥0 (4) with its corresponding Pareto Front PFknown, close enough to
i
the true optimal solution [5].
where F4 is the maximum voltage deviation from the desired
value (pu); V ∈ ℜ n is the voltage vector (unknown) and III. PROPOSED METHOD
∗
V ∈ ℜ is the desired voltage vector.
n
A new approach based on the Strength Pareto Evolutionary
Algorithm was developed for this work. This method, closely
In summary, the optimization problem to be solved is the related to Genetic Algorithms [10], is based on generating a
following: stored External Population composed by the best known
individuals B of a general evolutionary population. This
min F = [F1 F2 F3 F4 ] (5) external group of solutions conforms Pknown, available at each
moment of the computation, i.e., the best known
approximation to Ptrue. The original SPEA evaluates an
where
individual’s fitness depending on the number of decision
vectors it dominates in an evolutionary population, i.e.,
 n

 n ∑V i − Vi* 
decision vectors that are not better in any objective function
F = ∑ Bi Pg − Pl i =1
V − V* ∞ Fi , but with a worse objective function Fi for at least one
 i =1 n  value of i.
  SPEA preserves population diversity using Pareto
is known as objective vector, dominance relationship and incorporating a clustering
subject to F1 ≤ F1m and the load flow equations [9]: procedure in order to reduce the nondominated set without
destroying its characteristics. In general, cluster analysis
partitions a collection of m elements into g groups of
n
Pk = V k ∑ YkiVi cos (δ k − δ i − θ ki ) relatively homogeneous elements, where g < m, selecting a
i =1 representative individual for each of the g clusters. That way,
(6) a fixed number of g individuals may be maintained in the
n
Qk = V k ∑ YkiVi sin (δ k − δ i − θ ki ) external population preserving the main characteristics of the
i =1 Pareto Front [7].
An important issue with SPEA is its converge property,
where: Vk is the voltage magnitude at node k; Yki is the assured by Theorem 4 proved in [5], a characteristic not
admitance matrix entry corresponding to nodes k and i; δk is always present in other MOEAs. Consequently, the algorithm
the voltage phase angle at node k; θki is the phase admitance implemented for this work is based on the original SPEA [7],
matrix entry corresponding to nodes k and i; Pk is the active but differs from it in the following aspects:
power injected at node k; Qk is the reactive power injected at
node k. • Heuristic Initialization. A special heuristic method is used
To represent the amount of reactive compensation to be to generate the initial population in order to obtain
allocated at each busbar i, an unknown vector B, known as individuals electrically well compensated. The proposed
decision vector [7], is used to indicate the size of each
heuristic is based on encouraging compensation at busbars
reactive bank in the power system, i.e.:
with large number of branches and voltage profile far
from the desired value. This is done by using a method
B = [B1 B2  Bn ], Bi ∈ ℜ , Bi ≤ Bm (7)
summarized as follows:
a. Choose a total amount of compensation Btot.
The set of solutions of a multi-objective optimization b. For each busbar i of the system, calculate a factor Ki

0-7803-7285-9/01/$17.00 (C) 2001 IEEE 98

0-7803-7287-5/01/$17.00 (C) 2001 IEEE
 using the following expression: with a maximum number g of nondominated individuals,
fixed by clustering, that participates in the ordinary
(Vi − Vi ∗ ) l i if Vi < Vi ∗ evolutionary process. That way, the external population
Ki = 
0 if Vi ≥ Vi ∗ used in the evolutionary process does not diminish the
influence of the evolutionary population and no optimal
where l i is the number of branches connected to node i. solution is lost. Note that this second external population
Ki = 0 indicates that no reactive compensation is may be stored on disk, because it does not participate in
heuristically assigned to busbar i. the evolutionary process.
c. Normalize Ki using:
• Freezing. Inspired in Simulated Annealing technique,
Ki probabilities (of mutation Pm, crossover Pc and for using
K i′ = n the local optimization Plo) change with the number of
∑K
i =1
i generations and fitness value, freezing at the end of the
computation to improve convergence [11].
d. Compensate each busbar i with Bi calculated as
follows: The proposed method may be summarized as follows:
1. Generate an initial population Pop using the heuristic
Bi = K i′Btot
method previously exposed and create two empty
external nondominated sets Pknown and SPknown (stored
• Local Optimization. A special heuristic technique is external population).
implemented to improve individuals based on determining
an adequate search direction using the power flow 2. Copy nondominated members of Pop to Pknown and
mismatch expression [4]: SPknown.
3. Remove individuals within SPknown which are covered
ǻP   J1 J 2  ǻį  (dominated) by any member of SPknown.
ǻQ = J J  ǻV  (8)
4. Remove solutions within Pknown which are covered by
   3 4  
any member of SPknown.
From (8) and neglecting J 3 as well as the non-diagonal 5. If the number of externally nondominated solutions in
{ }
elements of J 4 = J 4 ij , the following expression is
Pknown exceeds a given maximum g, clustering is applied
in order to reduce the external population to a size g.
derived:
6. Calculate the fitness of each individual in Pop as well as
in Pknown using standard SPEA fitness assignment
∆Qi ≈ J 4 ∆Vi = J 4 (Vi ∗ − Vi )
ii ii
(9) procedure.
7. Select individual from Pop + Pknown (multiset union)
where ∆Qi is the amount of reactive compensation to be until the mating pool is filled. In this study, roulette
added at busbar i. wheel selection is used.
8. Apply Plo, Pc and Pm to determine whether and
• Stop criterion. Computation is halted when no new individual is locally optimized or selected for crossover
nondominated solution is found to dominate an individual and mutation, in which case, standard genetic operators
of the external population for a given number Nstop of are applied.
successive generations.
9. Go to step 2 if stop criterion is not verified.
• Two External Populations. If only one external population
is used, it is possible:
IV. EXPERIMENTAL ENVIRONMENT
a. to save all found Pareto solutions, but this population
may become too large and the evolutionary population As a study case, the IEEE 118 Bus Power Flow Test Case
looses genetic importance in the search process; or has been selected [12]. In order to stress the original system,
its active and reactive loads were incremented by 40%,
b. to loose found solutions using clustering to maintain a
turning the power network in an adequate candidate for
given number g of external solutions (original SPEA
reactive power compensation.
approach).
For comparison purposes, the Pareto set generated by the
In this new proposal, two external populations are stored,
proposed method has been compared to Pareto sets obtained
one with all found nondominated solutions and another
using four different methods:

0-7803-7285-9/01/$17.00 (C) 2001 IEEE 99

0-7803-7287-5/01/$17.00 (C) 2001 IEEE
 1· Compensation schemes elaborated by a team of 4) Generational Distance (G)
specialized engineers using standard computational
programs (Specialist).
G=
(∑ d )
∆
N
i =1 i
2
1/ 2

2· Original SPEA implementation, with random initial N

population [7] (SPEA).
where di is the Euclidean distance (in objective space)
3· Original SPEA implementation with heuristic
initialization (SPEA+). between each objective vector F in PFknown and its nearest
member in PFtrue. A large value of G indicates PFknown is
4· A special SPEA with heuristic initialization and the far from PFtrue, being G = 0 the ideal situation.
mutation genetic operator replaced by the previous
exposed local optimization (SPEAlo). 5) Maximum Pareto Front Error (ME)
For the experimental results presented in the following
section, it has been assumed that α = β , i.e., capacitor and
∆
(
ME = max min F i − F j
j i ∞
)
reactor banks have the same cost per MVAr. At the same F i ∈ PFtrue ; F j ∈ PFknown
time, Nstop = 100 was experimentally chosen.
To evaluate the experimental results using all five methods, It indicates the maximum error band that, when
an appropriate test suite metrics is used [5], because no single considered with respect to PFknown, encompasses every
metric can entirely capture total MOEA performance, vector in PFtrue. Ideally, ME = 0 is desired.
effectiveness and efficiency. The test suit comprises the Since most of these metrics reflect the likeness between the
following metrics: true Pareto Front Optimal set PFtrue and a computed Pareto
Front set PFknown, a good approximation of PFtrue is built by
1) Overall Nondominated Vector Generation (N) gathering all nondominated individuals from all five sets. In
∆ other words, for the following results, PFtrue is approximated
N = PFknown c
by the best known solutions of all our experiments.

where ⋅ c denotes cardinality. V. EXPERIMENTAL RESULTS

This metric indicates the number of solutions in PFknown. Tables I and II present experimental results using the
A good PFknown set is expected to have a large number of IEEE-118 study case, showing the figures obtained by all five
individuals, in order to offer a wide variety of options to methods. For the methods used to compare the proposed
the engineer. approach, the best results obtained by a single run have been
selected, having the SPEA implementations reached a
stagnant population, i.e., no new solutions are obtained with
2) Overall Nondominated Vector Generation Ratio
new generations for Nstop = 100 generations. On the other
(ONVGR) hand, the proposed method has been stopped using a
∆
N maximum number of generation criterion, since it continues
ONVGR = generating new solutions reaching more than 2000 stored
PFtrue c solutions (SPknown). This is an important advantage since it
gives the user a wider variety of alternative solutions.
It denotes the ratio between the number of solutions in
PFknown to the number of solutions in PFtrue. Since the TABLE I
objective is to obtain a PFknown set as similar as possible to EXPERIMENTAL RESULTS: 60 GENERATION RUN OF THE PROPOSED METHOD
PFtrue, a value near to 1 is desired. Specia- Proposed
Metrics SPEA SPEA+ SPEAlo
list Method
3) Error Ratio (E) N 170 100 150 172 181
ONVGR 0.4315 0.2538 0.3807 0.4365 0.4594
∑
N
∆ ei
E = i =1
;
N E 0.2353 0.9500 0.9533 0.2558 0.2431

 0 if a vector in PF is also in PF G 0.5702 0.7325 0.6060 0.5814 0.5431

ei = 
known true

1 otherwise ME 0.0852 0.1040 0.2332 0.0324 0.0554

This ratio reports the proportion of objective vectors in
PFknown that are not members of PFtrue. Therefore; an Table I presents experimental results running only 60
generation of our proposed approach while the other methods
Error Ratio E close to 1 indicates a poor correspondence
run until convergence. Table II presents figures when the
between PFknown and PFtrue, i.e., E = 0 is desired.

0-7803-7285-9/01/$17.00 (C) 2001 IEEE 100

0-7803-7287-5/01/$17.00 (C) 2001 IEEE
 proposed method run for 200 generations. In both tables, best methods or as the best set of solutions calculated by a team of
figures are indicated by a shadowed cell. specialists.
Experimental results using the proposed approach
TABLE II demonstrated several advantages when using the proposed
EXPERIMENTAL RESULTS: 200 GENERATION RUN OF THE PROPOSED METHOD
method, such as a set of solutions closer to the True Pareto
Specia- Proposed
Metrics SPEA SPEA+ SPEAlo Set outperforming other methods in every studied figure of
list Method
merits, and a wider variety of options. This last feature is of
N 170 100 150 172 222 special importance, since a richer set of alternatives are
ONVGR 0.4208 0.2475 0.3713 0.4257 0.5495 offered to the network planners. In order to select sub-sets of
solutions which best fit the interests of the user, an adaptive
E 0.3235 0.9800 0.9867 0.4128 0.1486 constrain philosophy is suggested. That way, the network
G 0.6138 0.7635 0.6221 0.6022 0.5315 engineer may restrict the constraints to reduce the number of
solutions after having a good idea of the whole Pareto
ME 0.0852 0.0948 0.2294 0.0462 0.0342 solutions, searching forward only in the redefined domain.
This process may continue iteratively until a good solution
For the first two metrics N and ONVGR, it is clear that the with an acceptable compromise among objective functions is
proposed method has the best performace, since it generates found.
the widest variety of solutions. This fact is even more evident As future work, new specialized genetic operators are being
when the stored external population is considered. developed to locally improve reactive compensation of a given
In Table I, the compensation set elaborated by the individual. At the same time, other objective functions (such
specialists has the smallest Error Ratio E, closely followed by as voltage stability margin) are going to be considered.
the proposed method, what gives an idea of the good job Finally, parallel asynchronous computation using a network
being done by the engineers; however, with a greater number of computers are considered for larger networks with more
of generations, the proposed method improves sensitively its objective functions.
performance, reducing in nearly 40% its error ratio,
outperforming other methods, as shown in Table II.
REFERENCES
The values obtained for the Generational Distance G show
that the proposed method has the best results in both tables, [1] K. Miu, H. Chiang and G. Darling, " Capacitor Placement, Replacement
offering a PFknown set closer to PFtrue. and Control in Large-Scale Distribution Systems by a A-Based Two-Stage
Algorithm," IEEE Trans. Power Systems, vol. 12, nº 3, pp. 1160-1166,
Table I indicates that the proposed method occupies the Aug. 1997.
second place in the Maximum Pareto Front Error (ME) [2] J. Carlisle, A. El-Keib, D. Boyd and K. Nolan, " A Review of Capacitor
ranking for a lower number of generations, beaten by the Placement Techniques on Distribution Feeders," in Proc. 1997 IEEE 29th
SPEAlo implementation. However, with a larger number of Southeastern Symposium on System Theory (SSST '97).
[3] M. Delfanti, G. Granelli, P. Marannino and M. Montagna, " Optimal
generations, the proposed method have the smallest ME value Capacitor Placement Using Deterministic and Genetic Algorithms," IEEE
(see Table II). Trans. Power Systems, vol. 15, nº 3, pp. 1041-1046, Aug. 2000.
A final fact to be emphasized is that the proposed method [4] P. Kundur, Power System Stability and Control. New York: Mc Graw-
Hill, 1993.
improves its performance for most metrics with a larger [5] D. Van Veldhuizen, "Multiobjective Evolutionary Algorithms:
number of generations, while the other methods have Classifications, Analysis, and New Innovations," Ph.D. dissertation,
converged before whole exploration of search space, what is Faculty of the Graduate School of Engineering , Air Force Institute of
Technology, 1997.
known as premature convergence. [6] B. Venkatesh, G. Sadasivam and M. Abdullah Khan, " A New Optimal
Power Schedulling Method for Loss Minimization and Voltage Stability
Margin Maximization Using Successive Multiobjective Fuzzy LP
VI. CONCLUDING REMARKS Technique," IEEE Trans. Power Systems, vol. 15, nº 2, pp. 844-851, May.
2000.
In this paper, Reactive Compensation Problem is first [7] E. Zitzler and L. Thiele, " Multiobjective Evolutionary Algorithms: A
treated as a Multi-objective Optimization Problem with 4 comparative Case Study and the Strength Pareto Approach," IEEE Trans.
Evolutionary Computation, vol. 3, nº 4, pp. 257-271, Nov. 1999.
conflicting objective functions: (i) investment in reactive [8] H. Dommel and W. Tinney, " Optimal Power Flow Solutions, " IEEE
compensation devices, (ii) active power losses, (iii) average Trans. Power Apparatus and Systems, vol. PAS-87, nº 10, pp. 1866-
voltage deviation and (iv) maximum voltage deviation. 1876, Oct.. 1968.
To solve the problem, a new approach based on SPEA is [9] J. D. Glover and M. Sarma, Power System Analysis and Design. Boston:
PWS Publishing Company, 1994.
proposed. This new approach introduces several proposals as: [10] D. Goldberg, Genetic Algorithm is Search, Optimization & Machine
(i) heuristic initialization, (ii) a local optimization technique, Learning. New York: Addison Wesley, 1989.
(iii) a stop criterion, (iv) two external populations and (v) a [11] Z. Dong, Y. Wang, D. Hill and Y. Makerov, " A New Apporach to Power
System VAR Planning Aimed at Voltage Stability Enhancement with
freezing feature. Feedback Control," in Proc. 1999 IEEE Powertech Conference, paper
For comparison purposes, the solution set obtained in a BPT-212-24.
single run of the proposed method is compared to four sets of [12] R. Christie, " Power Systems Test Case Archive," Electrical Engineering
dept., University of Washington, Apr. 2000. [Online]. Available:
solutions calculated as the best of several runs using other http://www.ee.washington.edu/research/pstca.

0-7803-7285-9/01/$17.00 (C) 2001 IEEE 101

0-7803-7287-5/01/$17.00 (C) 2001 IEEE