You are on page 1of 8

480 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO.

2, MAY 2006

An Initialization Procedure in Solving Optimal


Power Flow by Genetic Algorithm
Mirko Todorovski and Dragoslav Rajičić, Senior Member, IEEE

Abstract—The recently published idea of treating voltage angles ordinal optimization theory [30]. Many of these techniques
at generator-buses as control variables enables to obtain voltages overcome the difficulties in the modeling of complicated cost
at load-buses with less computation. However, application of this functions, discrete control variables, and prohibited unit-op-
approach in solving the optimal power flow problem by genetic al-
gorithms may be ineffective if starting values of voltage angles are erating zones. Some deficiencies in OPF are elaborated on in
selected quite randomly. To overcome these difficulties, a new pro- [31], and many challenges are discussed in [32].
cedure for selection of an initial set of complex voltages at gener- A specific procedure for initialization and treatment of the
ator-buses is proposed in this paper. With this procedure, one can voltage angles at generator-buses as control variables are the
start the optimization process (i.e., genetic algorithm) with a set of main components of the approach for solving OPF proposed in
control variables, causing few or no violations of constraints. The
application of voltage angles at generator-buses as control vari- this paper.
ables and the proposed initialization procedure is illustrated on In order to test the proposed approach, it was applied to
the IEEE test systems. The obtained results are analyzed and com- solve OPF by genetic algorithms (GA-OPF). However, this
pared with the results from the literature. They are competitive, approach could also be applied in other methods based on sim-
with computational time drastically reduced. ilar principles as GAs and evolutionary programming. These
Index Terms—Bus admittance matrix, generator-bus modeling, methods have the ability to handle any type of the objective
genetic algorithms (GAs), load flow analysis, optimal power flow function, variables, and constraints. Computation procedures
(OPF). of these methods offer not one “ideal” solution but rather a set
of applicable near-optimal solutions, and they are suitable for
I. INTRODUCTION parallel computation.

T HE minimization of total fuel costs, referred to as eco-


nomic dispatch [1] or optimal power flow problem (OPF)
[2], is one of the ever-actual power system problems. It has been
II. PROBLEM FORMULATION CONSIDERING POWER FLOW
REQUIREMENTS WITHIN GA
a subject of intense power system research for more than four The OPF is a constrained optimization problem requiring
decades, resulting in many relevant publications. In particular, minimization of an objective function, which is the total power
under the deregulated environment in the electricity industry in generation cost
the past few years, the interest in OPF has become even more
pronounced. (1)
Many optimization techniques have been adapted and used
to solve OPF. A review of selected OPF literature until 1993
can be found in [3] and [4], having the applied techniques subject to
classified as nonlinear programming, quadratic programming,
(2)
Newton-based solution, linear programming, hybrid versions
of linear programming and integer programming, and interior (3)
point method. The continuous research in this field has led
where is the real power output of unit , is the cost function
to many new contributions beyond 1993. Some of the previ-
of unit , and is the total number of units.
ously used approaches have been modified and improved (e.g.,
The equality constraints (2) are the power flow equations,
[5]–[9]), and some other techniques have been used, such as
while the inequality constraints (3) are due to various limita-
simulated annealing [10], genetic algorithms (GAs) [12]–[16],
tions. The limitations include lower and upper limits on gen-
neural networks [17]–[20], dual-type method [21], [22], mean
erator real and reactive powers (respecting possible prohibited
field theory [23], evolutionary programming [24]–[27], tabu
zones as well), limits on voltage magnitudes, line and trans-
search algorithm [28], particle swarm optimization [29], and
former maximum currents, and sets of possible transformer taps
position and shunt admittances.
Manuscript received February 15, 2005; revised October 6, 2005. Paper no. The sets of state and control variables, which should
TPWRS-00088-2005.
M. Todorovski is with the Research Center for Energy, Informatics and Ma- satisfy all constraints, are somewhat different than usual in this
terials, Macedonian Academy of Sciences and Arts, Skopje, Republic of Mace- paper. It is well known that four quantities are assigned to each
donia (e-mail: mirko@manu.edu.mk; t_mirko@yahoo.com). of the buses in a power network. These quantities are injected
D. Rajičić is with the University “Sv. Kiril i Metodij,” Skopje, Republic of
Macedonia (e-mail:dragoslav@ieee.org). real power, injected reactive power, voltage magnitude, and
Digital Object Identifier 10.1109/TPWRS.2006.873120 voltage angle. In the OPF, the only scheduled bus quantities are
0885-8950/$20.00 © 2006 IEEE
TODOROVSKI AND RAJIČIĆ: INITIALIZATION PROCEDURE IN SOLVING OPF BY GA 481

the injected real and reactive power at load-buses, and none of 1) Selection of Initial Generator Power Outputs: Assume
the four bus quantities are scheduled at generator-buses. that the operating costs of unit can be represented by
The generator-bus treatment influences, to a great extent, both
speed and robustness of any power flow method, and since the (4)
power flow calculations are dominating within the GA-OPF ap-
proach, appropriate attention should be paid to the subject. As where , , and are cost coefficients. In cases where op-
suggested in [34], complex voltages at generator-buses may be erating costs are not represented by a quadratic function, we
taken as control variables. In such a way, the concept of one could approximate them by such function. The approximation
slack bus is abandoned. This is in accordance with the nature will only be used for getting the initial power outputs of the units
of the OPF, since one cannot know in advance which is the and will not be used in the objective function. In addition, we
best slack bus selection. Therefore, in the proposed approach, assume that all generators are online and connected in one point.
the generating unit’s real and reactive power output are state The total power demand is equal to the sum of loads plus
variables, which should satisfy(2) and (3). Introducing many power losses in the network . For zero-cost units (i.e.,
buses with known complex voltages ( buses), the number hydro power plants), we take corresponding scheduled power
of unknown complex voltages is smaller, and the conditions in outputs. Then, by the subroutine LCONG from the software
the network enable to solve the power flow problem with less package Fortran PowerStation 4.0 [38], which minimizes a gen-
computations. As a result, the whole GA-OPF procedure is less eral objective function subject to linear equality/inequality con-
time-consuming. straints, we obtain generator power output , , (i.e.,
In this paper, the set of state variables includes generator “economic dispatch solution”), as a solution of (1) subject to
outputs (real and reactive), load-bus voltages, line currents, and
transformer currents. The set of control variables consists of (5)
complex voltages at generator-buses, reactive powers of syn-
chronous condensers, transformer tap settings, and shunt de- (6)
vices settings.
where TPP and HPP are sets of the thermal and hydro power
III. GAs plants, respectively. In addition, we calculate weightings
A. Problem Encoding
Each control variable is called a gene, while all control vari- (7)
ables integrated into one vector is called a chromosome.
In this paper, we use real-coded GA where each chromosome
consists of four regions, one for each subset of control variables. (8)
Those subsets are generator-bus voltage magnitudes and angles,
synchronous condensers reactive powers, transformer tap set-
tings, and shunt admittances.
The GA always deals with a set of chromosomes called a where .
population. Transforming chromosomes from a population, we All chromosomes of the population are divided into
obtain a new population, i.e., next generation. To do this, we use three subsets. The first subset contains only one chromosome.
three genetic operators: selection, crossover, and mutation. The generator power outputs obtained by the LCONG optimiza-
tion procedure are assigned to this chromosome. The second
B. Initialization subset contains chromosomes as well as the third.
Usually, at the beginning of the GA optimization process, For each chromosome of the second subset, see the following.
each variable gets a random value from its predefined domain. 1) Randomly select unit using roulette selection method
However, this very simple initialization procedure was found and weightings obtained by (7).1
insufficient for the GA-OPF approach where generator-bus 2) If is equal to lower or upper limit, then repeat the
voltage angles are taken as control variables. In fact, in more selection; otherwise, set power output of the selected unit
complex networks, the GA-OPF procedure with randomly to the lower limit.
initiated voltage magnitudes and angles may not produce a 3) If spinning reserve of the remaining units is lower than a
feasible solution, even in hundreds of generations. Therefore, it certain predefined value (e.g., 10%), then repeat the selec-
is reasonable to make a special initialization procedure in which tion.
the knowledge of the power systems will be incorporated. 4) Calculate power outputs of the other units by the LCONG
Since the generator power outputs have well-defined lower optimization procedure.
and upper limits, while the voltage angles do not, in the initial- For each chromosome of the third subset, see the following.
ization procedure (and nowhere else), we select the real power 1) Randomly select unit using roulette method and weight-
outputs and voltage magnitudes at generator buses. Afterwards, ings obtained by (8).2
the corresponding voltage angles are calculated.
Following this idea, we developed a practical initialization 1Units having higher production costs are more likely to be selected.
procedure, which will be explained in this section. 2Units having lower production costs are more likely to be selected.
482 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

2) If is equal to lower or upper limit, then repeat the chromosome , and is the penalty term corresponding to con-
selection; otherwise, set power output of the selected unit straint .
to the upper limit. For example, if variable (having upper limit and lower
3) Calculate power outputs of the other units by the LCONG ) is of type , for which the penalty coefficient is , then in
optimization procedure. a case of constraint violation, the corresponding penalty term
2) Selection of Initial Voltage Magnitudes: We split the al- included in (9) will be
lowable generator-bus voltage magnitude interval into number
of levels equal to the number of chromosomes in the population.
Therefore, each voltage level corresponds to one of the chromo- if
(10)
somes. Then, within each of the chromosomes, we assign cor- if
responding voltage level to all generator-buses. This procedure
is referred to as “voltage-grating.” For the set of violated constraints, four different penalty coef-
3) Selection of Initial Taps and Shunts Settings: The initial ficients are used, related to the following state variables: gener-
values of the rest of the control variables are selected at random ator real and reactive powers, voltage magnitudes, and branch
from its predefined domain. MVA flows. As suggested in [2], it is quite effective to start
4) Proposed Initialization Procedure: The initial values of with low values of penalty coefficients and to increase them
voltage angles and magnitudes at generator-buses can be ob- during the optimization process. The penalty increase should be
tained in the following steps. controlled; otherwise, it may be counterproductive and perform
worse than the case with constant penalty coefficients. Conse-
1) Select the initial generator real power outputs.
quently, the following control scheme is applied. After a cer-
2) For each of the chromosomes, assign the corre-
tain number of generations (usually ten), we check whether the
sponding voltage level to generator-buses using the
best chromosome has changed since the last check and whether
“voltage-grating” procedure (introduced in this subsec-
there are violated constraints related to it. If in both cases the an-
tion).
swers are positive, then the corresponding penalty coefficients
3) For each chromosome, do the following.
are multiplied by a certain factor (usually two). In addition, it is
3.1) Calculate voltage angles at generator-buses and
advisable not to let the penalty coefficients increase too much.
complex voltages at load-buses by the fast decou-
Note that in this approach, within each generation of the GA
pled or Newton’s method. In this calculation, we
solution process, for each of the chromosomes, we evaluate
use generator real power outputs from step 1), gen-
complex voltages at generator-buses by GA and then calculate
erator-bus voltage magnitudes from step 2), and
complex voltages at load-buses. With these voltages, we can di-
scheduled real and reactive loads.
rectly calculate injected power at each of the generator-buses,
3.2) Check for violated generator reactive power out-
regardless of the number of generators at the bus. However, two
puts. If violation occurs, reduce/increase the corre-
or more generators can exist at some buses, in which case, the
sponding bus voltage magnitude and go to step 3.1).
share of each generator to the total injected power should be
Otherwise, accept the actual voltage magnitudes and
determined. This can even be done prior to the GA-OPF op-
angles at generator-buses as initial values for the
timization by solving a small problem of a few parallel-con-
chromosome.
nected generators at one point. In such a way, we can construct
At this point, it might seem that step 3.1) is a deviation from a lookup table with parallel generator power outputs offering
the main idea of this paper, which is the use of generator-bus minimal production costs.
voltage angles as control variables. Nevertheless, it should be
emphasized that this is only an intermediate step toward popu-
D. Selection
lation initialization.
Improvement of the average fitness of the population is
C. Chromosome Fitness achieved through selection of individuals as parents from the
Fitness is a quantity related to the chromosome. It serves to completed population. The selection is performed in such a
enable comparison between chromosomes. At the stage when way that chromosomes having higher fitness are more likely to
we calculate the fitness, we have already solved the power flow be selected as parents.
equations, meaning that the constraints (2) are satisfied. How- Bearing in mind that some of the individuals may have signif-
ever, all constraints (3) have to be checked for violations. In this icantly higher fitness than the others, the next generation may be
paper, the penalty method is used, which degrades the fitness in constituted of large number of identical individuals. This poses
cases with violated constraint. The fitness for chromosome is a limitation on the population diversity, meaning that the search
defined by space will be reduced. One can omit such a situation by using
relative fitness [11]. Let and be the minimal and the
maximal fitness in the population, respectively. Then the rela-
(9) tive fitness for chromosome is defined as

where is the objective function related to chromosome


(11)
[as in (1)], is the set of violated constraints associated to
TODOROVSKI AND RAJIČIĆ: INITIALIZATION PROCEDURE IN SOLVING OPF BY GA 483

There are several selection techniques. The roulette selection


method is used here. In this method, we calculate the relative
weight of each chromosome’s fitness as

(12)

Therefore, the likelihood of selecting a chromosome as a


parent is a function of its fitness relative to the total.
To further improve the evolving process, the GA can carry
over the best individuals from the completed population to the
new population set (principle of elitism).

E. Crossover
After the selection, the GA picks a pair of selected chromo-
somes in order to create two new chromosomes. The GA ap-
plies a random generator to cut the strings at any position (the
crossover point) and exchanges the substrings between the two
chromosomes. After the crossover is performed, the new chro-
mosomes are added to the new population set.

F. Mutation
The mutation is specifically applied to increase population
diversity. Mutation involves randomly selecting genes within
the chromosomes and assigning them random values within the
corresponding predefined interval. In order not to destroy good
genetic code, nonuniform mutation has to be applied. In such
a manner, in later stages of GA optimization process, the in-
terval for random selection of genes’ values is narrowed [11]. In
the course of mutation, gene will take a new value depending
on generation number , maximum number of generations ,
nonuniform mutation parameter , and two random numbers
and as follows: Fig. 1. Flowchart of the proposed optimization procedure.

if number of generations, for which we assume that the search


(13) process has covered sufficient search space.
Other parameters, such as crossover probability, mutation
if rate, selection, and crossover mechanisms, seem to affect the
GA process less significantly when evaluated over a larger
where and are gene’s minimal and maximal values (from number of generations.
the permissible domain). The flowchart of the proposed optimization procedure is
The probability of mutation is normally kept very low, as high shown in Fig. 1.
mutation rates could degrade the evolving process into a random
search process. IV. RESULTS AND DISCUSSION
The proposed approach was applied on the following test sys-
G. GA Parameters tems: IEEE 30 [25], [27], [39], IEEE 118 [26], [39], 1-area IEEE
GA requires definition of a number of parameters, which can RTS96 [15], [39], and 3-area IEEE RTS96 [15], [39]. In some
affect the efficiency of the search process in several ways. cases, the results obtained by the proposed o influence on the
The population size should be large enough to create final result. approach are compared with the results from the
sufficient diversity covering the possible solution space. Gen- literature. In all those cases, the crossover and mutation proba-
erally, one cannot know the optimal value of in advance. bility, as well as the population size and number of generations,
Clearly, a more complex problem domain requires a larger were taken from the corresponding papers. The nonuniform mu-
due to larger possible combination of variables. tation parameter in (13) is set to 5. The results were used to
Another user-defined criterion is the point at which the opti- investigate evolution of the objective function, final objective
mization process terminates. In this paper, we use GA with fixed function values, computation times, and method robustness.
484 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

TABLE I
SOLUTION FOR THE IEEE 30-BUS TEST SYSTEM

Fig. 2. Objective function evolution—IEEE 30.

TABLE II
FINAL RESULTS OF THE TESTS

Fig. 3. Objective function evolution—1-area IEEE RTS96.

In some of the figures, we can see intervals in which the objec-


tive function increases. They appear when there are overloaded
lines as a consequence of the GA activity to eliminate the vio-
lations.

B. Final Results
Table I presents one solution for the IEEE 30-bus test system,
containing the genes’ values (voltage magnitudes and angles)
along with the generators’ real and reactive power outputs.
Fig. 4. Objective function evolution—IEEE 118. From the OPF solution, the power system operator can get the
injected real power and voltage magnitude for every generator,
as well as transformer taps, and shunt admittances settings. In
other words, he can set the system in optimal state by adjusting
the real power output through the governor loop and voltage
magnitude through the exciter loop. Also, Table I contains the
production costs for all generators.
In Table II, we present the final values of the objective func-
tion obtained by the proposed approach (20 runs), along with
the results reported in the corresponding papers. The table con-
tains the best solutions, as well as the average values of all 20
runs. A good concordance in the results is evident.
Fig. 5. Objective function evolution—3-area IEEE RTS96. C. Computation Times
In Table III, the time consumption (measured on AMD
A. Evolution of the Objective Function Athlon 1,833 MHz) for the whole GA-OPF run for two dif-
Figs. 2–5 illustrate the evolution of the best objective func- ferent cases is presented. In both cases, we used the same
tion values through generations. In these figures, there are two GA-OPF approach explained in this paper, but we calculated
lines representing two different initialization procedures: solid voltages at load-buses by different methods. In case 1, we
line for the initialization proposed in this paper (see Section III) applied the power flow method from [34] (see the Appendix),
and dotted line for the standard initialization procedure (random while in case 2, we applied the fast decoupled power flow
selection of real powers and voltage magnitudes). These figures method from [33]. In both cases, we used the real and reactive
show that appropriate reduction of the number of generations power mismatches as a termination criterion for the voltage cal-
has almost n culation procedure. In addition, the same seed for the random
TODOROVSKI AND RAJIČIĆ: INITIALIZATION PROCEDURE IN SOLVING OPF BY GA 485

TABLE III V. CONCLUSION


TIME CONSUMPTIONS USING DIFFERENT POWER FLOW METHODS
In this paper, a novel approach in solving the OPF problem
by GA has been proposed. It is based on the application of
the new initialization procedure and recently published idea of
using voltage angles at generator-buses as control variables.
The application of the proposed approach to several IEEE test
systems shows that the proposed initialization procedure im-
proves the performance of the whole GA-OPF procedure. Nu-
merous tests illustrate that the proposed approach gives results
TABLE IV competitive to the results obtained by corresponding methods
RESULTS WHEN CHROMOSOMES INCLUDE GENERATOR from the literature. In addition, the computational time is dras-
REAL POWERS INSTEAD OF VOLTAGE ANGLES
tically reduced.

APPENDIX

A. Review of the Power Flow Method From [34]


In this approach, it is assumed that complex voltages at all
generator-buses are known. The complex bus admittance ma-
trix is formed taking into account initial values of the
TABLE V transformer turn ratios and shunt admittances.
RESULTS FOR THE IEEE 30-BUS TEST SYSTEM Using the elements of , for load-bus , we can write

(14)

(15)

(16)

number generator was applied, so we are certain that the initial where is the complex load at bus , is the dif-
populations are identical, as well as the evolution path. In such ference between the actual and initial values of complex shunt
a way, we produced exactly the same final results. admittance at bus , is the complex voltage at bus ,
In Table IV, we present the results (time consumption as an is the sum of currents modeling tap changes [as in (23)] at all
average of all 20 runs and the best generation costs) for the transformers connected to bus , is the element of in
cases where the generator real powers and voltage magnitudes row and column , and and are sets of indexes of
are selected as control variables (i.e., as genes). In case 1, the load-buses and generator-buses directly connected to bus , re-
initial generator real power outputs and voltage magnitudes are spectively. Asterisk as a superscript denotes complex conjugate.
selected randomly, while in case 2, the procedure proposed in Writing (14) for every load-bus of the system, a set of simul-
this paper was applied. In both cases, the fast decoupled power taneous equations is obtained. It can be expressed in matrix form
flow method was employed. The final results are pretty similar as
to the results of Table II and, in the case when the proposed
initialization was used, are slightly better. As it can be seen, the (17)
difference between the running times from Table III, case 1, and
Table IV is obvious. where elements of vectors and are calculated from
(15) and (16), respectively. At the beginning of the voltage cal-
D. Robustness culation procedure, can be obtained from by ex-
cluding rows and columns related to generator-buses.
In order to investigate the robustness of the proposed ap- Load-bus voltages can be calculated from (17) by an iterative
proach, it is applied to the IEEE 30-bus test system using the procedure, since the elements of depend on the load-bus
following three types of generator cost curves (as in [25]): voltages that are not known at the beginning. On the other hand,
quadratic, piecewise quadratic, and quadratic with a sine the elements of do not change from iteration to iteration,
component superimposed upon it. In the last type, the sine and we calculate them only once. In addition, application of the
component is used to represent the valve-point loading effect transformer model explained in Subsection B of this Appendix
[12]. allows keeping constant in spite of changes in transformer
The obtained results are presented in Table V, where a good tap settings. As a result, should be formed and factorized
match with the results from [25] is obvious. only once.
486 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

For and in two successive iterations, we can define

(18)
Fig. 6. Transformer equivalent circuit.
and

(19)

Then, from (17)–(19), we obtain

(20)
Fig. 7. Representation of the transformer from Fig. 6.
Since is sparse, the use of sparse matrix factorization
is recommended when solving (17) and (20). In addition, it is voltages can be calculated by using methods effective for radial
beneficial to apply the method proposed by Tinney and Walker and weakly meshed networks.
for structurally symmetric matrices [35], as well as Markowitz’s
strategy and Duff’s search technique [36]. In this procedure, C. Handling Changes of Transformer Taps
the application of row-column permutation contributes to fill-ins
minimization, while the threshold pivoting strategy improves Let transformer have complex turns ratio and admittance
the numerical stability. , and connect buses and . Corresponding equivalent circuit
At the beginning of the GA-OPF procedure, the matrix contains an ideal transformer with complex turns ratio :1 in
is created and factorized. The iterative procedure for calculation series with admittance (see Fig. 6). The bus admittance ma-
of load-bus voltages consists of the following steps. trix of the transformer from Fig. 6 is (e.g., [37])
1) Set all load-bus voltages at per unit (flat start).
2) Calculate the elements of [using (15) and (23)] and
(21)
the elements of [using (16)].
3) Solve (17) and get new load-bus voltages.
4) Calculate new elements of [using (15) and (23)] and
According to (21), three out of four elements of are
the elements of [using (19)].
not constant, since ratio keeps changing as the GA searches
5) Solve (20) for and update .
for the optimal solution. This could pose serious drawbacks on
6) If magnitudes of all elements of are not less than
the voltage calculation procedure, because if one goes straight-
the specified voltage tolerance, go to step 4); otherwise,
forward, many factorizations of will be required.
the iterative process is finished.3
This problem can be bypassed following a different calcu-
lation path, in which the influence of the ratio changes can be
B. Network-Model
modeled by current injections. To do this, for every transformer
It should be noted that the selection of voltage angles at , we first calculate the transformer admittance matrix
generator-buses as control variables enables the construction using the initial value of its turns ratio and put it into .
of a network-model, which can be used in calculation of volt- Then, in cases when turns ratio of the transformer changes to
ages at load-buses. The network-model does not contain the , corresponding admittance matrix can be repre-
generator-buses. It can be obtained from the original network sented as a sum of and the matrix increment
making the following changes: each branch connecting a
generator-bus with a load-bus should be substituted by a shunt
branch and shunt current generator at the load-bus. The shunt
branch impedance is equal to the impedance of the substituted (22)
branch. The current of the current generator is equal to the
quotient of the voltage at the generator-bus and the impedance
of the substituted branch. Then, can be obtained as the In order to keep constant (and avoid additional factor-
bus admittance matrix of the network-model. izations), we do not put into , but we simulate it in
It should be noted that in some cases, network-models consist (20) by injected currents
of several parts. Voltage calculations in each of those parts can
be done separately. In addition, some of the parts can have radial
or weakly meshed topological structure. In those cases, load-bus

(23)
3In order to have the same termination criterion for both cases in Table III,
the procedure does not stop after step 6) but rather continues with step 7):
1 1
7) Calculate power mismatches ( P and Q ). If magnitude of any power
that should be subtracted from and , respectively (see
mismatches is greater than the specified power tolerance, then continue as in
step 4); otherwise, the iterative process is finished. Fig. 7).
TODOROVSKI AND RAJIČIĆ: INITIALIZATION PROCEDURE IN SOLVING OPF BY GA 487

ACKNOWLEDGMENT [22] C.-H. Lin, S.-Y. Lin, and S.-S. Lin, “Improvements on the duality based
method used in solving optimal power flow problem,” IEEE Trans.
The authors would like to thank Dr. N. Markovska and Ms. S. Power Syst., vol. 17, no. 2, pp. 315–323, May 2002.
Secrest for proofreading the manuscript. Also, the authors spe- [23] L. Chen, H. Suzuki, and K. Katou, “Mean field theory for optimal power
flow,” IEEE Trans. Power Syst., vol. 12, no. 4, pp. 1481–1486, Nov.
cially acknowledge the data provisions for IEEE test systems by 1997.
Messrs. M. A. Abido, P. N. Biskas and P. Venkatesh via private [24] H.-T. Yang, P.-C. Yang, and C.-L. Huang, “Evolutionary programming
e-mail communication. based economic dispatch for units with nonsmooth fuel cost function,”
IEEE Trans. Power Syst., vol. 11, no. 1, pp. 112–118, Feb. 1996.
[25] J. Yuryevich and K. P. Wong, “Evolutionary programming based op-
REFERENCES timal power flow algorithm,” IEEE Trans. Power Syst., vol. 14, no. 4,
pp. 1245–1250, Nov. 1999.
[1] J. Carpentier, Contribution a l’étude du dispatching économique, ser. 8: [26] P. Venkatesh, R. Gnanadass, and N. P. Padhy, “Comparison and appli-
Bulletin de la Société Française des Électriciens, Août 1962, vol. III, pp. cation of evolutionary techniques to combined economic emission dis-
431–447. patch with line flow constraints,” IEEE Trans. Power Syst., vol. 18, no.
[2] H. W. Dommel and W. F. Tinney, “Optimal power flow solution,” IEEE 2, pp. 688–697, May 2003.
Trans. Power App. Syst., vol. PAS-87, no. 10, pp. 1866–1876, Oct. 1968. [27] M. A. Abido, “Environmental/economic power dispatch using multiob-
[3] J. A. Momoh, M. E. El-Hawary, and R. Adapa, “A review of selected jective evolutionary algorithm,” IEEE Trans. Power Syst., vol. 18, no. 4,
optimal power flow literature to 1993, Part 1: Nonlinear and quadratic pp. 1529–1537, Nov. 2003.
programming approaches,” IEEE Trans. Power Syst., vol. 14, no. 1, pp. [28] T. Kulworawanichpong and S. Sujitjorn, “Optimal power flow using tabu
96–104, Feb. 1999. search,” IEEE Power Eng. Rev., vol. 22, no. 6, pp. 37–40, Jun. 2002.
[4] , “A review of selected optimal power flow literature to 1993, Part 2: [29] Z. L. Gaing, “Particle swarm optimization to solving the economic dis-
Newton, linear programming and interior point methods,” IEEE Trans. patch considering the generator constraints,” IEEE Trans. Power Syst.,
Power Syst., vol. 14, no. 1, pp. 105–111, Feb. 1999. vol. 18, no. 3, pp. 1187–1195, Aug. 2003.
[5] Y.-C. Wu, A. S. Debs, and R. E. Marsten, “A direct nonlinear predictor- [30] S.-Y. Lin, Y.-C. Ho, and C.-H. Lin, “An ordinal optimization theory-
corrector primal-dual interior point algorithm for optimal power flows,” based algorithm for solving the optimal power flow problem with dis-
IEEE Trans. Power Syst., vol. 9, no. 2, pp. 876–883, May 1994. crete control variables,” IEEE Trans. Power Syst., vol. 19, no. 1, pp.
[6] G. Torres and V. Quintana, “On a nonlinear multiple-centrality-correc- 276–286, Feb. 2004.
tions interior-point method for optimal power flows,” IEEE Trans. Power [31] W. F. Tinney, J. M. Bright, K. D. Demeree, and B. A. Hughes, “Some
Syst., vol. 16, no. 2, pp. 222–228, May 2001. deficiencies in optimal power flow,” IEEE Trans. Power Syst., vol. 3, no.
[7] Y.-C. Wu, “Fuzzy second correction on complementary condition for op- 2, pp. 676–683, May 1988.
timal power flow,” IEEE Trans. Power Syst., vol. 16, no. 3, pp. 360–366, [32] J. A. Momoh, R. J. Koessler, M. S. Bond, B. Stott, D. Sun, A. Papalex-
Aug. 2001. opoulos, and P. Ristanović, “Challenges to optimal power flow,” IEEE
[8] V. Miranda and J. T. Saraiva, “Fuzzy modeling of power system optimal Trans. Power Syst., vol. 12, no. 1, pp. 444–455, Feb. 1997.
load flow,” IEEE Trans. Power Syst., vol. 7, no. 2, pp. 843–849, May [33] B. Stott and O. Alsaç, “Fast decoupled load flow,” IEEE Trans. Power
1992. App. Syst., vol. PAS-93, no. 3, pp. 859–869, May/Jun. 1974.
[9] K. H. Abdul-Rahman and S. M. Shahidehpour, “Static security in power [34] M. Todorovski and D. Rajičić, “A power flow method suitable for
system operation with fuzzy real load conditions,” IEEE Trans. Power solving OPF problems using genetic algorithms,” in Proc. IEEE Region
Systems, vol. 10, no. 1, pp. 77–87, Feb. 1995. 8 EUROCON, vol. 2, 2003, pp. 215–219.
[10] K. P. Wong and C. C. Fung, “Simulated annealing based economic dis- [35] W. F. Tinney and J. W. Walker, “Direct solution of sparse network equa-
patch algorithm,” Proc. Inst. Elect. Eng., Gen., Transm., Distrib., vol. tions by optimally ordered triangular factorization,” Proc. IEEE, vol. 55,
140, no. 6, pp. 509–515, Nov. 1993. no. 11, pp. 1801–1809, Nov. 1967.
[11] Z. Michalewicz, Genetic Algorithms+Data Structures=Evolution Pro- [36] I. S. Duff, A. M. Erisman, and J. K. Reid, Direct Methods for Sparse
grams, 3rd ed. Berlin, Germany: Springer-Verlag, 1999, pp. 111–112. Matrices. London, U.K.: Oxford Univ. Press, 1986.
[12] D. C. Walters and G. B. Sheblé, “Genetic algorithm solution of economic [37] G. W. Stagg and A. H. El-Abiad, Computer Methods in Power Systems
dispatch with valve point loading,” IEEE Trans. Power Syst., vol. 8, no. Analysis. New York: McGraw-Hill, 1968.
3, pp. 1325–1332, Aug. 1993. [38] “Fortran PowerStation 4.0,” Microsoft Developer Studio, Microsoft Cor-
[13] G. B. Sheblé and K. Brittig, “Refined genetic algorithm—economic dis- poration, 1994–1995.
patch example,” IEEE Trans. Power Syst., vol. 10, no. 1, pp. 117–124, [39] University of Washington, Department of Electrical Engineering. Power
Feb. 1995. Systems Test Case Archive. [Online]Available: http://www.ee.wash-
[14] P.-H. Chen and H.-C. Chang, “Large-scale economic dispatch by genetic ington.edu/research/pstca.
algorithm,” IEEE Trans. Power Syst., vol. 10, no. 4, pp. 1919–1926, Nov. [40] [Online] Available: http://cuaerospace.com/carroll/whats_new_ga.html.
1995.
[15] A. G. Bakirtzis, P. N. Biskas, C. E. Zoumas, and V. Petridis, “Optimal
power flow by enhanced genetic algorithm,” IEEE Trans. Power Syst.,
vol. 17, no. 2, pp. 229–236, May 2002.
[16] T. Yalcinoz, H. Altun, and M. Uzam, “Economic dispatch solution using Mirko Todorovski received the B.Sc., M.Sc., and D.Sc. degrees from Univer-
a genetic algorithm based on arithmetic crossover,” in Proc. IEEE Porto sity “Sv. Kiril i Metodij,” Skopje, Republic of Macedonia, in 1995, 1998, and
Power Tech. Conf., Porto, Portugal, Sep. 2001. 2004, respectively.
[17] J. H. Park, Y. S. Kim, I. K. Eom, and K. Y. Lee, “Economic load dispatch From 1997–2005, he was with the Research Center for Energy, Informatics,
for piecewise quadratic cost function using Hopfield neural network,” and Materials of the Macedonian Academy of Arts and Sciences, Skopje, Re-
IEEE Trans. Power Syst., vol. 8, no. 3, pp. 1030–1038, Aug. 1993. public of Macedonia. In 2006, he joined the Faculty of Electrical Engineering
[18] J. Kumar and G. B. Sheblé, “Clamped state solution of artificial neural in Skopje, and presently, he is a Teaching Assistant in the Department of Power
network for real-time economic dispatch,” IEEE Trans. Power Syst., vol. Systems. His research interests are related to computer applications in power
10, no. 2, pp. 925–931, May 1995. system analysis and planning.
[19] C.-T. Su and G.-J. Chiou, “A fast-computation Hopfield method to eco-
nomic dispatch of power systems,” IEEE Trans. Power Syst., vol. 12, no.
4, pp. 1759–1764, Nov. 1997.
[20] T. Yalcinoz and M. J. Short, “Neural networks approach for solving eco- Dragoslav Rajičić (SM’97) received the B.Sc. degree from University “Sv.
nomic dispatch problem with transmission capacity constraints,” IEEE Kiril i Metodij,” Skopje, Republic of Macedonia, and the M.Sc. and D.Sc. de-
Trans. Power Syst., vol. 13, no. 2, pp. 307–313, May 1998. grees from the University of Belgrade, Serbia and Montenegro, in 1963, 1970,
[21] C.-H. Lin and S.-Y. Lin, “A new dual-type method used in solving op- and 1978, respectively.
timal power flow problems,” IEEE Trans. Power Syst., vol. 12, no. 4, pp. He is a retired Professor from the Faculty of Electrical Engineering, Univer-
1667–1675, Nov. 1997. sity “Sv. Kiril i Metodij.”

You might also like