1994 Iba

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

IEEE Transactions on Power Systems, Vol.

9, No, 2, May 1994 685

Reactive Power Optimization by Genetic Algorithm

Kenji Iba
Mitsubishi Electric Corp., Kobe, Japan

Abstract: This paper presents a new approach t o The second problem is the integer problem. Most
optimal reactive power planning based on a genetic control variables ( transformer tap positions, switchable
algorithm. Many outstanding methods to this problem shunt capacitorlreactor banks) should have integer
have been proposed in the past. However, most these values. No matter how precise a continuous solution
approaches have the common defect of being caught to a may be, it is impossible to assign these values directly
local minimum solution. The integer problem which to control valuables. Since modern high voltage
yields integer value solutions for discrete controllers1 systems have high capacity bank units such as 80-
banks still remain as a difficult one. The genetic l2OMVA in their capacitorlreactor banks, simple
algorithm is a kind of search algorithm based on the round-off calculations could lead to significant errors.
mechanics of natural selection and genetics. This Although a mixed integer programming (MIP) is
algorithm can search for a global solution using prepared for this purpose, the process is more
multiple path and treat integer problem naturally. The complicated than conventional continuous approaches.
proposed method was applied t o practical 51-bus and
224-bus systems to show its feasibility and capabilities. Artificial intelligence(A1) technologies, partic-
Although this method is not as fast as sophisticated ularly expert systems are utilized for this OPF
traditional methods, the concept is quite promising and problern[6l7[71. This approach consists of "If-then'' based
useful in the coming computer age. production rules. Conflicting rules and/or rules which
generate endless loops should not be included. If more
Pevwords : Genetic Algorithm, Optimal Power Flow, than two rules are applicable, proper priorities should
Global Minima, Reactive Power Allocation, Integer be selected (conflict resolution). The construction. of
Problem, Expert system such strict rules (meta-rules), however, require
extensive help from skilled knowledge engineers.
1. Introduction Genetic algorithms (GAS) are optimizing methods
based on the mechanics of natural selection and
The optimal power flow problem (OPF) has been a geneticstsl. GAS have already been applied to some
traditional problem in power system controllplanning. power system problems [91-[121. Although GAS are
Extensive studies have been carried out. Nonlinear capable of yielding global minima, they have required
prograaming (NLP) approaches have been presented tremendous computing time.
[11,[21. Successive linear programming (LP) has been
proposed by the present authod31. The Newton approach A new attempt to apply GAS t o reactive power
and the quadratic approach had been successfully planning is presented in this paper. Although the basic
applied [41,[51. Most of these approaches can be idea is based on GAS, the method is quite different from
categorized as constrained optimization techniques. conventional GAS' approach. The principal features
Expert systems based on AI technology have also been and keywords of the method are listed below:
addressed to this problemt61.
Searches multiple paths to reach a global optimal
Although these techniques have been successfully Uses various objective functions simultaneously
implemented in existing power systems, difficulties Treats integeddiscrete variables naturally
remain. The first problem is that it is easy to be caught Uses unique two intentional genetic operations;
by a local minimum solution. Since OPF is not a mathe-
matically convex problem, most techniques might
- Interbreeding: Crossover using network topology
- Manipulation: AI based stochastic "if-then" rules
converge to a local minimum instead of a unique global
minimum. If the starting point (initial condition) is
located near the global optimal point, the solution
obtained may be true. There is, however, no guarantee 2. ConceDts of Proposed Method
of this. Mathematical proof to be the global optimal 2.1 Multide Path Search
may not be indispensable in practical applications if the
obtained solutions are quasi-optimal. The solutions, Most previous approaches, such as LP, NLP and
however, are sometimes infeasible if tight margins Newton approaches, are single path search algorithms.
(constraints) for state variables are applied. Same Starting from an initial condition, they improve the
difficulties can be observed in optimization by neural state variables in every iteration. There is a single path
network using the Hopfield model. from the initial condition through a converged solution.
For example, in the maximizing problem sketched in
Fig.1, the converged goal might be a local maximum
point (i.e., top of lower hills); as well, there is no way t o
judge whether or not it is the global optimal. If such
methods are caught or arrested by local maxima, there
This paper was presented at the 1993 IEEE Power Industry is no way to escape, except applying random noise. The
Computer Application Conference held in Scottsdale, Arizona, same thing occurs in modern artificial neural networks
May 4 - 7, 1993. such as the Hopfield model.
0885-8950/94/$04.00 0 1993 IEEE
686

The proposed method uses similar weighted


summation as an objective function to evaluate total
power system performances. The method, however,
does not use linearized sensitivities at all. Therefore,
we do not have to adjust each weight t o get good
convergent characteristics.
Although the goal of optimization is to minimize
Equation(l1, total and subsystems are independently
evaluated with various objective functions in this
method. These local and various informations are used
for assembling genes to create preferable solutions.
2.3 Discrete State Variables (Chromosomes)
Fig.1 Global and Local Solutions Control variables, such as transformer tap posi-
tions, number of switchable shunt capacitorlreactor
banks, and generator reference voltages, are repre-
sented as an integer stringlvector X, as shown in
Equation(2). These string are called "chromosomes" in
GAS. Unlike conventional GAS which utilize 110 bits
(binary digit), this approach utilizes integer values. The
generator voltages might be treated as continuous
variables in some applications. The reference voltage
specified by a control center or a digital controller has
discrete values in many cases. It seems to be quite
practical to treat them as discrete variables.

0 initial Condition 0 First Generation


0 intermediate Point 0 Second Generation
0 Optimal point QD Third Generation
0 Final Generation where
Fig.2 Single Search and Multiple Search by GAS TPi : tap position at i-th transformer
SCi : No. of shunt compensator banks at i-th bus
VGi : position of generator voltage at i-th gen.
The proposed method has various conditions o r (capital letter implies integer values)
states in every iteration to search using multiple paths.
In GAS, the iteration and number of these states are Control variables in per unit can be obtained from
called "generation" and "population" respectively. Equation(3).
Every intermediate point in the same generation can
interchange information with each other point. The taPi=l.O+Tpi Atapi
larger the population, the greater the possibility of
convergence with the global optimal. Although there is sc i = sc-fixedi + SC i ASC i ----- (3)
no guarantee that it will be global, it is more likely to be
so. Fig.2 illustrates single and multiple path search. Vgi = 1.0 + VGi AVg i

Starting from various initial conditions, the tradi- where :A implies step size or b a d u n i t capacity in pu
tional approach seems t o trace multiple paths. Each
path, however, will not interchange information with It should be noted that the larger the A*,step size,
other paths. The multiple path search is completely the smaller the possible integer number a t each
different from the repetition of a single path search. controller. The state space i.e., total number of
Finding the global optimal will be essential in long combinations can be reduced for large step sizes. This
range planning for power systems where the systems' is one of excellent feature of GA based techniques.
configuration is significantly changed.
2.4 Genetic ODerationg
2.2
. .
Various Obiectwe Functions A text bookM gives following GA features;
.............................
There are many aims in reactive power planning,
such as, loss andlor invest cost minimization. Some GA is a blind search using stochastic operations, not
constraints might be treated as penalty functions. In deterministic rules. No information or knowledge
many applications, a weighted summation of various concerning to objects is required. Operators are simple
objective functions is used to evaluate total performance itself, involving nothing more complex than random
as shown in Equation(1). This becomes a quite number generation (mutation), string copying (repro-
duction) and partial string exchanging (crossover).
complicated function which may have steep cliffs or .............................
walls in state-space. This fact makes it more difficult to These features are powerful in some applications
find a global optimal solution in linear-sensitivity based
such as information processing. The proposed method,
approaches. however, violates these criteria. Instead of applying
flx) = C wi gi(x) -->minimize ---- (1) mutation (random noise), simple AI-based rules are
applied to improve solutions. This operation might be
where gi(x) : objective function of i-th objective called "manipulation". In addition to applying random
string exchanging (crossover), topolo@al information
wi : weight of i-th objective
687
and various objective functions are utilized for string 3. Detail of ProDosed ADD roach
exchanging. This operation might be called “inter-
breeding“. These intentional operations, which may 3.1 Remesentab’on of Control Variables
not be GAS in the strict sense, dramatically accelerate The details of the proposed approach are described
the convergent speed. This approach, however, fails to in this chapter. Fig. 5 shows a flowchart of this
find the global optimal point if these operations are approach. Discrete control variables are represented as
applied deterministically. a chromosome or string X in Equation(2). If an X is
specified, a unique power flow solution can be obtained
2AI-based”if-then”f using traditional power flow calculations ( in which
The difficulties in constructing production-rules multiple power flow solutions[l3] are ignored). The
has been described previously. However, there are combination of strings, i.e, the size of the state-space, is
many simple rules which may not be strict but may be defined in Equation(6), where Xmin and Xmax denote
valid in many cases. The proposed approach applies the upper and lower limits in integer values.
such if-then rules stochastically. The priority or validity
of these rules can be taken into account by introducing a r”r (xy““- x p
i=l
+ 1) ---_---- (6)
kind of probabilitc factor. Equations (4) and ( 5 ) show the 3.2 DecomDositiodClassification into Subsvstem8
relationships.
A network is decomposed into subsystems using
AI( expert system): regiodarea information. Considering network topology
If (condition)i then (action)i ---- (4) and connections, one can divide a system manually.
Proposed approach Every controller and state variable are categorized and
Generate uniform random value ‘rnd ( 0 < rnd < 1) labeled in each subsystem. Various objective functions
If ( rnd < (thresholdh and (condition)i ) then (action)i are calculated independently in each subsystems. The
----- (5) purpose of these classification and evaluation is to get
8 some information for effective gene operations. The
power flow (PF) calculation is applied to entire systems
Considering regiodarea information, we can ( not to each subsystems 1.
decompose or classify a large power system into some
subsystems. We do not have to apply a mathematical 3.3 Creation of First Generation
strict approach for this division, rough and manual
approach is s a c i e n t . Suppose a system is decomposed The first generation of chromosomes is created
into four subsystems ( area-A,B,C and D) as illustrated using uniform random variables, as shown in Equation
in Fig.3, each subsystems are partially evaluated with (7).Generated chromosomes which do not converge to a
an objective function. Fig. 4 shows an example of such solution during PF calculations are eliminated from the
evaluations for three individuals or members in a list. It is difficult to create a first generation for large
generation. The present proposed approach applies a systems, since chromosomes generated by Equation(7)
genetic operation: partial string exchanging (crossover) seldom converge to a solution. Another scheme should
using this information. Selecting one of excellent be applied to large systems. If some initial conditions
subsystems, this approach constructs a new system as are located near the optimal point, it is a good idea to
shown in Fig.3. This selection should be done according add these conditions to the first generation, though this
to the “roulette rule“, a stochastic selection technique, may lessen the capability of finding the global optimal
so as to maintain the flexibility and global sense. The point. Although creation of the first generation seems
idea of this operation is similar to agricultural plant- to be a trivial process, this is in fact one of the most
breeding, however, is not a conventional GA technique. important aspects.
Xi = INT (RND . ( X Y -~ XP + I)) + XP ---- (7)
where:
RND : Generate uniform random value 0 < RND < 1
INT(*) : Return the next lowest whole number of *

3.4 Evaluation of Each Chr omos0me


After translating the information of chromosomes
to discrete control variables, power system profiles can
be obtained by PF calculations. Every profile is evaluated
as shown in Fig. 6, where each subsystem is evaluated
with various objective functions. Even if a total profile is
bad, its subsystem may have good features for a certain
objective. Using the table in Fig.6, we can easily select
Fig.3 InterbreedinglCrossing Operation such subsystems. Objective functions used in this paper
are voltage violation, generator VAR violation, power
loss and weighted summation of these three functions.
The total objective function which evaluates total
performance is the OBJtotal in Equation(8). The goal of
our optimization is to minimize this value. If there is no
violation of constraints, the OBJtotal becomes the power
losses.
OBJtotal = 10 C ( % of VoltageViolation)2
Subsystems + 5 C ( pu of Gen-vAR Violation) 2
Fig.4 Local Evaluation ( Example ) + Power Losses in pu _____
(8)
688

.
ReadPFData
I_.- ...-..-....... .r-.:.~.1 --::.:.~:
.......................................
I....-----.....----- .............................
i Represent Discrete Control Variables
! and Create Chromosome Template

---"-.--:-*===:iDecompose into Subsystems .-.... _-.,


Create First Generation From Random Value Fig.6 Table of Objective Functions ( Example)
1

."Y_............... and Calculate


Y.."
PF
..............--....and
.........Evaluate
.. ............".............. ,".....
""

of objective
unction
Fig.7 Roulette Used in Stochastic Selection

3.6 ODeration of Mutation and M a n i d a tion


According t o evolutionism, animalhegetable life
undergoes physical change due to mutation. Recent
biochemical and genetic engineering, however, make it
possible to manipulate or recombine genes intentionally
Fig. 5 Flow Chart of the Proposed Approach and artificially. In the proposed approach, this
manipulationhecombination is applied using AI-based
Convergent characteristic is also evaluated using this "If-then" rules to improve power system profiles. The
OBJtotal. Individual objective functions ( voltage and sequence of this process is described below:
gen-VAR violation and losses ) are used independently to
select good subsystems in various sense. The weighting 1. Using a random value, select one manipulation
parameters for penalty functions in Equation(8) can be from the following six candidates, which are
arbitrary selected. Since these parameters do not affect described in a later section:
convergent characteristics in the proposed method, no Tap manipulation
adjustment is required. Shunt compensator manipulation
Generator voltage manipulation
&5 InterbreedindCrossinp between Subvstemq Remote controller manipulation
Mutation (random noise operation)
Applying gene operation to the chromosomes in No operation (path-through)
first generation, we try to make next generation. Based
on "Crossover" which is a conventional GA operation, a 2. Generate uniform random value 'rnd (0 < rnd < 1)
more intentional operation "Interbreeding" is proposed. and apply the AI-based rule in Equation(5).
The interbreeding is a kind of crossover using partial 3. Repeat 2 with i=lto Nc (all related controllers)
evaluation of each subsystems. The sequence of this 4. If more than 10 identical members exist, apply
process is described below: "mutation" t o avoid intermarriage
5. Apply PF calculation to the new chromosome to
1. Select one objective function using random value check convergence. If not converge then go to 1.
from various objectives. 6. Repeat operations 1through 5 to all members created
2. Make a roulette according to the values of the selected by interbreeding, until the total population reaches
object functions evaluated in the i-th subsystem, as the specified number.
shown in Fig. 7. Objective functions which should be
minimized are converted to maximizing functions.
3. Select one individuallmember using this roulette. 3.7 Utilized "If-then" Rules
The greater the objective function value, the more
likely its selection. "If-then" Rules utilized in the method are listed
4. Extract that part of this selected chromosome which in this subsection. These rules and their actions should
contains the information on control variables in the be applied stochastically to many members. If we
i-th subsystem. assign small values to the thresholds in Equation(51,the
5. Reproducelcopy that part of chromosome to create a possibility of applying this rule is limited. Some rules
new chromosome. for loss minimization have no condition. These rules
6. Repeat 1through 5 with i=l to n (No. of subsystems) will be applied occasionally according t o random
to form one complete chromosome. values. The "nearest" controllers are assigned in
7. Check whether new chromosome is identical to the advance using networks' topological tree-search.
others in the new generation. If more than 10 exist,
apply "mutation" to avoid intermarriages.
8. If this new chromosome converges to a solution in Rules for Tap Manipulation:
the PF calculation, add it to the next generation. Check whether both sides a transformer's voltage
9.Repeat 1through 8 until the total population of new match to the condition illustrated in Fig.8. Improve the
chromosomes reaches the specified number. tap position according to the table.
689

elapsed CPU time were 27 seconds with 17 iterations; 32


seconds with 21 iterations; and 93 seconds with 80
iterations by HP9000/730 (76MIPS).
The same problem was solved by a successive LP
methodC31 within 22 seconds. Although this SLP method
is based on precise linearization, the computing speed
is inferior to recent sophisticated methods. The objective
ni = 1 + ( TPi + rn ) Atapi function value of a continuous optimal solution by this
SLP method was 0.1952. However, a discrete solution
Fig. 8 Tap Manipulation Rule obtained by simple round-off calculation shows poor
Rules for Shunt Compensator Manipulation: performance (OBJtotal was 881.5, since violations of
If Vk-low then add 1SC bank at bus-k constraints exist 1.
If Vk-high then remove 1SC bank at bus-k
If Vk-low then remove 1Sh-R bank at bus-k It should be noted that the proposed multiple
If Vk-high then add 1Sh-R bank at bus-k search can prepare some alternatives. Table 2 shows
Add 1SC bank ( rule for loss minimize) some quasi-optimal solutions which remains in the last
Remove 1Sh-R bank ( rule for loss minimize) --- (9) generations. A great merit of this method is that it
presents these candidates in addition t o global
Rules for Generator Voltage Manipulation: solutions. Since the difference of controller positions is
If Qgk > Qgmaxk then decrease Vgk setting 1step small, objective function values are almost same as the
If Qgk c Qgmink then increase Vgk setting 1step best solutions in these candidates.
--- (10)
Rules for Remote Controller Manipulation: Regarding t o the global optimal solution, it is
If vk-low then move 1tap a t nearest difficult to declare that we succeeded to find the global
transformer to increase Vk optimal, since many quasi-optimal solutions were
If Vk-high then move 1tap at nearest found from various seeds. Fig. 11 shows control
transformer to decrease Vk positions of two solutions. A statistical analysis for
If vk-low then add 1SC or remove 1Sh-R bank globability check is shown in Fig.12. This figure
at nearest upper bus indicates that the proposed method does not necessarily
If Vk-high then remove 1 SC or add 1Sh-R bank obtain the global solution in strict sense. However, the
at nearest upper bus method can always get quasi-optimals stably and
If Vk-low then increase Vgk setting 1step surely.
at nearest generator
If Vk-high then decrease vgk setting 1 step
at nearest generator ---411)
Rules for Mutation:
xi= INT (RND . ( x y - x p + 1))+ x p ---- (7)
Gen Bus 6 J T a p 10 (lOx8,20x2) 0.01~~
3.8 Alternation of Generations and Converpence Check Load Bus 45 I SC 6 (4,22,30,24,6,6) 40 MVA
Branch 53 J Vg 6 (10 x 6 ) 0.01 pu
The best members for each objective function Population 50 1 Rate of mutation 0.05
evaluated in the total system are left for the next Subsystem 5 1 Rate of manipulation 0.05
generation. Therefore, the objective functions decrease ..................................................
uniformly. Identical members in the same generation
should be limited i n number, since homogeneous
generations can not improve themselves, and tend t o
converge local minima. 1000 n\
Convergence judgement is not easy, since GAS
sometimes form a plateau in the improvement of a
solution. In this paper, "If no improvement is achieved
through three generations and there is no violation of 100
=
a
constraints, then terminate the iteration" is applied. c
0
5
m
4. Test Results
910
4.1 5l-Bus SVste~
The proposed method was applied to a practical 51-
bus system. Details of this system and parameters are
given in Table 1. Since this stochastic method depends
on generated random values, its performance was
checked using various seeds (i.e., initial values for
random generator). Fig. 9 shows good convergent 5 1 A B C D E F G
characteristics of the proposed approach. Although
case-G are terminated before its convergence, the 0 10 20 30 40 50
convergent characteristics are quite good in many Iteration
cases. Fig.10 shows a statistical analysis for the 51-bus Fig. 9. Convergent Characteristics (51-Bus System)
system. The possibility of converging within 70
iterations is 90%, within 26 iterations is 50%. Required
690
50 r ~ . ~ ~
: : l1 : . .........................
: ._ .. _ ~
: _:
: ..............., 4.2 224-BUSS ~ tme
The proposed method was applied to a large scale
224-bus system. Details of this system are shown in
Table 3. The convergent characteristics are shown in
Fig. 13. Required elapsed CPU time and iterations were
11:19 with 33 iteration and 1510 with 41 iterations.
There are many quasi-optimal solutions in this system,
too. This behavior is a little similar to multiple load
flow problem[W. Difference of control variables are
shown in Fig. 14. Although the differences in power-
loss is slight, the difference in controllers are
0 10 20 30 40 50 60 70 80 significant. The reason why there are so many
Converged iterations solutions seems to be in the simple objective function.
Fig.10 Required Iteration to Converge Applying more objectives, such as security constrains,
investment costs, we may reduce the number of quasi-
optimal solutions. Some statistical analysis are shown
20 4 in Fig. 15 and Fig. 16. OBJtotal of a continuous optimal
n Case-A
solution by SLP was 0.562. However, OBJtotal of the
10 discrete solution obtained by simple round-off was
2117.8 with violations.
0
4.3 Effect of Parameter%
There are three important parameters in the
f---T A P -*
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
S C _jf_ V G + method threshold for rules; population; and number of
subsystems. The most sensitive and important factor is
Controllers threshold in Equation(5). If we assign high value to the
Fig.11 Comparison of Two Solutions in 51-Bus System threshold (e.g. 0.21, the convergent speed is improved.
However, the possibility converging to a local solutions
becomes higher. If we assign low value to the threshold
such as 0.05, the convergent speed becomes slow.
However, most of converged solution are close to the
global optimal. This threshold is similar t o t h e .
"Temperature" in simulated annealing.
The population is also important. The greater the
number, the higher possibility to get a global optimal.
The greater population, however, requires the
computing time. After a lot of iterations, generations
0.196 0.198 0.2 0.202 0.204 0.206 0.208 tend to become homogeneous. Even if the population is
Objective Function Value ( Power Losses) high, the variety will be lost in such cases. Keeping
Fig.12 Distribution of Objective Function Value fresh members in generations is more essential.
Although interbreeding is the most important
operation in the proposed method, the number of
subsystems are not sensitive. The decomposition seems
to be easy for the reactive power optimization in radius
practical systems.

Gen Bus 27 I Tap 89 (av. 12 steps ) 0.01 pu


Load Bus 197 I SC 43 (total 9 GMVA ) 40 MVA
Branch 237 I Vg 27 (9%=5steps ) 0.01 pu
Population 100 I Rate of mutation 0.02
Subsystem 10 I Rate of manipulation 0.08

-4:
.
-6 I
-8 I I

0 40 80 I60
Controllers 120

Fig. 14 Difference Between Two Solutions in 224-Bus System (Case-C minus Case-A)
69 1

10000 5. Conclusions
This paper presents a new approach t o the
1000 reactive power allocation planning. The GA-based
method utilizes unique intentional operations. The first
is "interbreeding" which is a kind of crossover using
decomposed subsystems. This idea is a little similar to
agricultural plant-breeding, since it assembles a whole
system using good parts with various features.
The second is "gene-recombination'' or "manip-
ulation" which improves power system profiles using
AI-based stochastic "If-then'' rules. Rough rules can be
A B C D applied stochastically to many chromosomes. This idea
comes from recent biochemicallgenetic engineering.
. , I . I . I . I .
0 40 60 20 80 This approach is successfully applied to practical
iterations 51-bus and 224-bus systems. Multiple search can find
Fig. 13. Convergent Characteristics (224-Bus System) out many quasi-optimal solutions in discrete control
values. Although there is no mathematical proof or
guarantee that the solution is global, it is most likely to
......I ........:.................
" ~
: ...................:.. ~
: . .................,100
,: be so. Test results show good convergent characteristics
95 8 and feasible computing speed.
10 80 '5
50 5 Although this method is not a s fast as
30 2 sophisticated traditional methods for large systems,
10 2
remarkable advance in computer technology will soon
, .-g
U)
eliminate this defect. It might be a paradox, however,
0 creating time consuming methods is important t o
n utilize the all CPU power of recent computers.
"10 20 30 40 50 60 70 80
Converged Iterations
Fig.15 Required Iterations to Converge Referencess
........I ....................,.........I ................. ~ ...............
: : : : ; j ; : i 1
[ll R.C. Burchett, et. al.,"Large Scale Optimal Power
Flow," IEEE Trans. on PAS Sep 1983, pp 2995
[2] R.A. Fernamdes, et. al., "Large Scale Reactive
Planning," IEEE Trans. on PAS May 1983,pp 1083
[31 K.Iba, et.al.,"Practical Reactive Power Allocation
Using SLP," IEEE Trans. on PS v01.3,No.2,May 1988.
0.559 0.561 0.563 0.565 0.567 0.569 0.571 0.573 [4] D.I.Sun et.al.,"Optimal Power Flow by Newton
Objective Function Value ( Power Losses) Approach," IEEE Trans. on PAS. Oct 1984,pp 2864
Fig. 16 Distribution of Objective Function [51 R.C.Burchett et.al., "uadratically Convergent Opti-
mal Power Flow," IEEE Trans. on PAS Nov 1984
[6] C.C. Liu, et.al.,"An Expert System Assisting Decision
4.4 Considerations for bulk Dower svstemg Making of VQ Control," IEEE P m - l , n o . 3 . Aug. 1986
[a F.D. Galiana, et. al. , "Expert Systems in Transmission
The proposed method is successfully applied for Planning," IEEE Proc., vol 80, no. 5 May 1992
practical large systems. However, some difficulties will [8] D.E.Goldberg,"Genetic Algorithms in Search"
arise for bulk power systems which may require large Addison Wesley 1989. ISBN 0-201-15767-5
number of population and excessive CPU time. Two [9] T.Haida,Y .Akimoto,"Voltage Optimization by
possible ways of overcoming these difficulties are Using Genetic Algorithms", ESAP'91 Japan
considered. The first idea is population control. Staring [lo] K.Nara, et. al. ,"Implementation of Genetic Algo-
from large number of populations, we may reduce it rithm for Distribution System Loss Minimum Re-
gradually in every generation. This scheme will reduce Configuration,"PES SM 1991,9lSM467-1PWSR
CPU time without losing globability. The second idea is [ll] H Mori et. al. ,"A Genetic Approach to Power System
resolution control. Applying larger step size (Atap, Topological Observability,"Proc, IEEE 1991 ISCAS pp.1141
Asc,Avg), we can find some promising regions, first. [12] D.C.Walters et.al.,"Genetic Algorithm Solution of
Resetting the step size to smaller values, we can search Economic Dispatch with Value Point Loading,"
such regions to find precise solutions. This scheme will IEEE PES Summer Meeting 1992,92SM414-3PWSR
reduce searching area. These two improvements seem [13] K.Iba et.al."A Method for Finding a Pair of Multiple Load
to be effective, but they have not yet been tested. Flow Solutions," IEEE, Trans. on PS vo1.5, No. 2, May 1990.

It is true that the complexity of the method grows Kenii Iba (M87) was born in Tokyo, Japan in 1955. He
exponentially with system size. However, the ability of received the B.S, the M.S. and the Ph.D. degrees in
recent computers, has been improved exponentially, electrical engineering from Waseda University in 1978,
too. Required memory for the method will not be a in 1980 and in 1990 respectively. He joined Mitsubishi
serious problem for recent computers. Since the Electric Corp. in 1980. Dr. Iba is a member of IEEE and
proposed method has not been tuned up very well, it IEEJ. His office address is "1-1-2 Wadasaki-cho, Hyogo-
have possibility to be improved more by some ideas or ku, Kobe 652, Japan" ( E-mail [email protected])
careful investigations. GA based methods will be quite
promising in coming new computer age.
692

Discussion K. Iba, Mitsubishi Electric Corp. : The author would


like to thank the discussers for their interests in my
paper and their valuable comments. The first comment
D.S. Kirschen, B.F. Wollenberg, Empros Power is related to globability and performance. Theoretical
Systems Control and University of Minnesota, Minneapolis, concepts of GAS have been developed using generalized
MN: Genetic algorithms have not yet been applied problems. It is a strict blind search. However, even to
successfully to power systems analysis problems because they this original GA, there is no mathematical proof that it
could not produce results in an acceptable amount of time for always converges to an unique global optimal. Most of
models of a realistic size. The author of this paper shows that recent GA-based applications, which show successful
performance in practical problems, use problem-
the performance of this type of algorithm can be considerably specific informations. This kind of modification seems
improved if the “evolution” of the solution is not left entirely to be essential to get good performance. It is also true
to chance but is guided towards the optimum. As the author that we have to apply intentional operations carefully as
correctly points out, this approach goes somewhat against the the discussers pointed out.
underlying philosophy of genetic algorithms. If appropriate
precautions are not taken (such as the use by the author of The second comment is related to discretelinteger
random variables in the selection of the heuristic rules and in problems. One of most significant advantage of the
proposed method is that it can deal with discrete nature
the interbreeding process), the final solution may actually be of the problem. There are so many discrete problems in
quite far from the global optimum. power industries, i.e., switching of circuit breakers,
unit commitment, etc.. The GA-based methods are
The comparison between the results obtained using the genetic quite promising to these fields.
algorithm and those obtained using a successive LP method
suggest that a very important advantage of the proposed The third comment is a valuable suggestion. Although
method may be that it deals with the discrete nature of the troubles caused by “masking“ are not significant yet,
problem intrinsically and not as a nuisance which must be this phenomena are observed in some test cases. This
suggestion will be helpful for future improvements. As
dealt with by post-processing. the discussers pointed out, the method can deal with
any kind of objective functions. We do not care whether
The second degree terms used in the objective function to it is differentiable or not. This feature is a great
minimize voltage and VAR constraint violations could cause advantage in optimizations.
“masking”, i.e. a situation where one large violation is hidden
by the sum of several smaller violations. Using a larger The fourth comment is concerning to the combination
exponent would reduce the likelihood of masking and would with OPF. Although the author has no experience
not increase the complexity of this method because it does not using such formulations, it seems to work well. If the
step size of transformer taps and generator voltages are
rely on the evaluation of derivatives of the objective function. relatively small, they might be treated as continuous
value, rather than disadvantageous discrete value.
We would like to know if the author has considered a different Anyway, the high performancdspeed of OPF package is
problem formulation where the only control variables would indispensable for this combination.
be the location and size of the capacitors and the conventional
power flow would be replaced by an optimal power flow which Manuscript received August 25, 1993.
would adjust the transformer taps and the generator voltages to
minimize the objective function. This formulation might
facilitate the consideration of contingencies in the planning
problem.

Manuscript received May 28, 1993.

You might also like