POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

Multi-Objective Differential Evolution for Optimal

Power Flow
M. A. Abido N. A. Al-Ali

Electrical Engineering Department Transmission Asset Planning Department

King Fahd university of Petroleum & Minerals Saudi Electricity Comapny
Dhahran 31261, Saudi Arabia Dammam, Saudi Arabia
e-mail: [email protected] e-mails: [email protected]

Abstract— This paper presents a multiobjective differential a true multi-objective optimization problem in which the
evolution (MODE) based approach to solve the optimal power objectives are treated simultaneously and independently [5-9].
flow (OPF) problem. OPF problem has been treated as a true This, however, makes the problem more complicated,
multiobjective constrained optimization problem. Different
whereas, traditional optimization techniques have several
objective functions and different operational constraints have
been considered in the problem formulation. A clustering weakness and drawbacks such as: linearization, continuity,
algorithm is applied to manage the size of the Pareto set. Also, an differentiability, local optima and constraints handling.
algorithm based on fuzzy set theory is used to extract the best Therefore, new optimization techniques such as genetic
compromise solution. Simulation results on IEEE-30 bus test algorithms (GA), particle swarm optimization (PSO),
system show the effectiveness of the proposed approach in solving differential evolution (DE) are recently introduced and also
true multi-objective OPF and also finding well distributed Pareto
applied in the field of power systems with promising success
solutions.
[7-12].
Index Terms— Optimal power flow, multiobjective The literature includes several OPF studies that dealt with
optimization, evolutionary algorithms, differential evolution. multi-objectives and applied evolutionary optimization
techniques. References [4] present a Particle Swarm
I. INTRODUCTION Optimization based OPF incorporating several objectives as a
Nowadays, utilities are facing rapid increase in electricity weighted sum. Reference [5] proposed DE based
demand with slow reinforcement projects due to financial and multiobjective economic environmental power dispatch,
political issues. Proper operation and planning requires solved using the weighted sum approach. In references [7-11],
consideration of different factors such as reduction of a true multi-objective with competing fuel cost and power
generation cost, losses, pollution, security of power system, plants emissions was successfully formulated and produced
and FACT applications, etc. In this regard, the Optimal Power promising results.
Flow (OPF) has taken a wide area in the research field with This paper presents a multiobjective differential evolution
potential applications for both planning and operation of (MODE) based optimization approach to solve the OPF
power systems. The OPF generally seek to optimize a set of problem. A clustering algorithm is applied to manage the size
objectives under certain constraints [1-3]. of the Pareto set. Also, an algorithm based on fuzzy set theory
Previous OPF studies has dealt with single objective cases is used to extract the best compromise solution. The proposed
and applied solution techniques based on traditional MODE based approach has been tested on IEEE 30-bus
optimization methods [1-3]. However, due to the fact that real standard system to demonstrate its effectiveness.
life problems involve several objectives and that the available
traditional optimization techniques have several drawbacks II. PROBLEM STATEMENT
and disadvantages, both multi-objective OPF and application
A. Problem Formulation
of new evolutionary techniques appeared in the recent studies
on OPF. Typically, a multi-objective OPF problem can be
Traditionally, multi-objective OPF were treated as a single mathematically written as follows:
objective optimization problem. In several studies, a new ªf 1 ( x , u )º
objective function is formed as a weighted sum of all « »
objectives using suitable scaling/weighting factors. This min/ max F(x, u) = « # » j=1, 2,……# of objectives
approach has the disadvantage of finding only a single «f j ( x , u ) »
¬ ¼
solution which does not express the trade-off between the
different objectives [4]. Generating multiple solutions using Subject to g(x,u) = 0, (equality constraints)
this approach requires several runs with different weighting h(x,u)  0, (inequality constraints)
factors and hence elongates the running time [5]. As an The variables, x and u represent the controlled and
alternative to this approach, recent studies consider the OPF as uncontrolled system variables. Also, g represents the load flow

978-1-4244-2291-3/09/$25.00 ©2009 IEEE 101

POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

equations; while h represents system and equipment initial population of NP candidate solutions is randomly
operational limits. generated over the problem space:
B. Objective Functions xij = xj-lower limit + Random# * (xj-upperr limit - xj-lower limit )
In this work, three objectives corresponding to important where NP: population size and D: problem dimension
concerns in planning and operation are considered [6, 7].
• Step-2 (Evaluate Objectives):
1. Fuel Cost Minimization
The objective of each solution is computed, and in order to
Fuel cost for any machine producing Pi MW is commonly
account for infeasibility, the mismatch of each constraint
modeled in a quadratic model with A, B and C parameters. For
value is multiplied by large value and added to all
a system with n units the total fuel cost can be computed as
objectives.
follows:
n • Step-3 (Find the Pareto Set):
Fuel Cost = ¦ (A + B i * Pi + C i * Pi )
2
$ / Hour
i =1
i With the use of dominance criteria, the population is
2. Minimization of Transmission Losses searched through and the set of non-dominated solutions
are copied to an external Pareto Set. If the size of the
As the power (MW) is transmitted via the transmission
Pareto set exceeded its limit, then clustering is performed.
network, portion of this power is lost in the form of heat. In
this work, the total sum of MW losses throughout the network • Step-4 (Check Stopping Criteria):
is calculated and used as an objective in the OPF. Once the Pareto solutions are found, the stopping criteria
is checked. Maximum number of iterations is normally
3. Enhancement of Voltage Stability
used.
Voltage stability is one of the major concerns in power
• Step-5 (Perform Mutation):
systems which has been said to be a major cause to several
blackout incidents. It relates to loading level and reactive For each vector xi in the population, a mutant vector is
power support. Voltage stability L-index indicator is one of created by randomly selecting three vectors (R1, R2, R3),
the simplest and commonly used indicator for voltage stability different from xi and different from each other.
assessment [13-14]. The value of L index at any load bus x’ = xR1 + F* (xR3 - xR2 )
varies between 0 (no load) and 1 (voltage collapse). In this Where, F is a mutation factor between [1,0] and it controls
work, the maximum L-index will be used as an objective in the solution variation in the mutation stage.
the OPF.
• Step-6 (Perform Crossover):
III. PROPOSED APPROACH To further perturb the mutant vector, a trial vector is
generated by copying parameters from the parent solution
A. Overview Xi and the mutant vector X'i in a probabilistic manner to
To overcome the difficulties in traditional optimization generate a Trail Vector X''i to be used in the selection
techniques, new evolutionary population based searching stage. This is accomplished by comparison between a
techniques were proposed to solve multiobjective optimization randomly generated number and a specified crossover
problems (MOP) with promising success [15-17]. In this factor (CR) between [1, 0], as described in Fig. 1.
paper, an approach based on a technique called "Differential However, in case CR=0, then all parameters are copied
Evolution" is presented and will be used to solve the from parent vector Xi except one, randomly selected from
formulated multi-objective OPF. the mutant vector X'i. Whereas, if CR=1 then all
parameters are copied from mutant vector X'i except one,
B. Differential Evolution randomly selected from the parent vector Xi.
Between the years 1994 to 1996, Rainer Storn and Kenneth
Parent Vector Trial Vector Mutant Vector
Price introduced an optimization technique called differential Xi X''i X''i
Parameter i
evolution (DE). The DE is a population based searching i=1
optimization technique and is characterized by its simplicity, i=2

robustness, few control variables and fast convergence [2518- i=3

19]. Comparison between DE and the well known Particle
i=4
Swarm Optimization technique in Reference [20] indicated
that DE has better performance. The DE has been used for i=5

both single and multi-objective optimization problems [21- i=6

22]. The steps of DE algorithm can be described as follows:
i=7
• Step-1 (Initialization): Random # > CR Random #  CR

As any evolutionary algorithm, the DE works using set of Fig.1: Crossover Process
candidate solutions forming the problem population. An
• Step-7 (Selection)

102
POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

Selection is performed using the dominance criteria. For Finally, the best compromise solution is the one achieving
the case of MOP and in order not to lose non-dominated the maximum member ship function (uk).
solutions and hence improve the diversity and The complete procedure of the proposed multi-objective
convergence, the non-dominated solutions are added to the Differential Evolution (MODE) is shown in Fig. 2.
main population and used later in the process of finding
Pareto solutions. This has a big impact on the diversity of
problem. INITIALIZE THE CONTROL VARIABLES

C. Clustering The Pareto Set

CALCULATE OBJECTIVES
At the end of each iteration the Pareto Set is updated by (Run Load Flow)
storing new or replacing the dominated solutions. During this
process, the size of this set may end up with accumulating FIND PARETO SOLUTIONS
large number of solutions. Therefore, in order to limit the
number of solutions up to the satisfaction of the results, CLUSTERING
clustering techniques are adopted to perform this task.
However, it should also maintain the diversity of the solutions. Yes
In this study, an approach based on the calculation of the STOPPING
CRITERIA
distances between the solutions is applied. The objectives
VARIATIO No
space is search through to find the nearest two solutions. Then, FURTHER
CLUSTERING
the solution that is closer to its neighborhood is excluded from MUTATION AND AND FIND BEST
CROSSOVER COMPROMISE
the Pareto Set. This process continue until the number of
SOLUTION
Pareto solutions is less than or equal to a predefined maximum
number of Pareto solutions. RUN LOAD FLOW
To enhance the diversity of the problem, the size of the CALCULATE OBJECTIVE STOP
SELLECTION
Pareto set is given a large number during the optimization
process. However, the final Pareto set can be further clustered
as required.
Fig. 2: Procedure of Proposed MODE
D. Best Compromise Solution
IV. SIMULATION RESULTS AND DISCUSSION
For the purpose of decision making, a best compromise
solution is computed. In this work, a technique based on fuzzy
set theory [8] was applied to extract the best compromise This section present simulation results performed on IEEE-
solution. The procedure of this technique can be explained as 30 bus test system using the proposed approach. The complete
follows: data for equipment, rating and operational constrains as
• Search through all solutions, to find Fmax and Fmin necessary for performing OPF study are given in references
corresponding to each objective function. [23]. In this study, three objectives are considered. These are
• Use the following linear membership function to calculate fuel cost, transmission losses, and L-index.
a membership function for each objective. A. Single Objective Optimization
­ 1 Fi = Fi min
Initially, single objective OPF is solved considering one of
° max the objectives at a time. This helps to study the range of each
° Fi − Fi
ui = ® max Fi min < Fi ≤ Fi max objective and also identify any inverse relation between the
° Fi − Fi
min
objectives. These cases were performed using the proposed
°0 Fi = Fi max
¯ DE based approach with NP=100, F=0.5, CR=0.98 and a
maximum of 300 iterations. The results and settings of the
The above equation gives a measure of the degree of
control variables for the three single objective cases are shown
satisfaction for each objective function for a particular
in Table 1. Also, the convergence of the objectives is shown in
solution and also map the objectives into the rage 1 ~ 0.
Figs. 3, 4 and 6.
• The corresponding membership function for the non-
dominated solution k, is calculated as follows:
• The initial settings of the system provided in [23]
NO corresponds to a fuel cost of 901.96 $/Hour. However,
¦u k
i after optimization, this value reached approximately
uk = M
i =1
NO
800.56 $/Hour which is comparable to the value achieved
¦¦ u
k =1 i =1
k
i
in reference [4]. The optimized value indicates an inverse
relation between the two objectives.
where, M : # of Pareto solutions; NO: # of objectives

103
POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

0.18
910

900
0.17
890

880
0.16
870
Fuel Cost ($/Hour)

L-index
860
0.15

850

840 0.14

830

820 0.13

810

800 0.12

0 50 100 150 200 250 300 0 50 100 150 200 250 300
Iterations Iterations

Fig. 3: Fuel Cost Minimization

Fig. 5: L-index Optimization

• The results in Table-1 and Fig. 4 show a reduction in the 10

losses from 5.83 MW to 3.24 MW. Also, the optimized

MW loss corresponds to an adverse impact on the fuel 9

cost.
• Fig. 5 shows the convergence of L-index from 0.172 to 8

0.125, and hence improving the voltage stability. The

optimized value for the L-index is nearly equal to the 7
L o sse s (M W)

value reported in [4] using PSO.

B. Multiobjective Optimization 6

To perform multi-objective OPF, the following cases were

5
performed, considering NP = 100 and a maximum of 100
Pareto solutions.
• Fuel Cost – Losses 4

This case was run with F=0.5, CR=0.98, and maximum

3
number of 500 iterations. The final Pareto set contained 790 800 810 820 830 840 850 860 870 880 890 900 910 920 930 940
Fuel Cost ($/Hour)
100 Pareto solutions which are well distributed, as
Fig 6: Fuel Cost –Losses Pareto Set
shown in Fig. 6.
• Fuel Cost – L-index
This case was run with CR=0.98 and maximum number
of 1000 iterations. F was changed from 0.5 to 0.1 during
0.1252
the optimization to help increase the number of Pareto
solutions. The final Pareto set is plotted in Fig. 7, and it
includes 74 solutions which are fairly well distributed.
0.1251

6.0
L-index

0.1250

5.0
Losses (MW)

0.1249

4.0

0.1248
800.57 800.58 800.59 800.60 800.61 800.62 800.63
Fuel Cost ($/Hour)

3.0 Fig 7: Fuel Cost & L-index Pareto Set

0 50 100 150 200 250 300
Iterations

Fig. 4: Losses Optimization

104
POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

TABLE I
SIMULATIONS RESULTS
Limits Single Objective Fuel Cost & Losses Fuel Cost and L-index
Before
Variables Best Best Best Best Best L- Best
Min Max OPF Fuel Cost Losses L-index
Cost Losses Comp. Cost index Comp.
P1 50 200 99.23 175.39 54.00 127.12 175.77 71.73 123.16 177.21 177.29 177.22
Generators Voltage Generators Output

P2 20 80 80.00 48.82 79.93 38.85 46.09 62.91 53.70 48.20 48.54 48.36
P5 15 50 50.00 21.61 49.97 44.76 21.31 49.67 31.32 21.50 21.18 21.44
(MW)

P8 10 35 20.00 22.21 34.49 34.00 23.89 34.85 34.58 21.27 20.93 21.10
P11 10 30 20.00 12.28 28.42 26.69 13.10 29.85 24.83 12.25 12.52 12.33
P13 12 40 20.00 12.00 39.83 17.38 12.30 37.96 21.50 12.01 12.00 12.00
V1 0.95 1.10 1.050 1.085 1.064 1.070 1.089 1.055 1.073 1.085 1.084 1.084
V2 0.95 1.10 1.040 1.065 1.059 1.052 1.064 1.049 1.061 1.064 1.063 1.064
V5
( p.u. )

0.95 1.10 1.010 1.033 1.044 1.061 1.029 1.041 1.034 1.031 1.029 1.030
V8 0.95 1.10 1.010 1.037 1.047 1.040 1.030 1.037 1.040 1.038 1.037 1.038
V11 0.95 1.10 1.050 1.087 1.099 1.098 1.086 1.054 1.074 1.085 1.082 1.086
V13 0.95 1.10 1.050 1.047 1.031 1.052 1.001 1.061 1.059 1.045 1.048 1.045
T6-9 0.90 1.10 1.078 1.038 1.011 1.039 1.034 1.002 1.008 1.023 1.022 1.026
Position

T6-10 0.90 1.10 1.069 0.943 1.048 0.959 1.066 0.922 0.914 0.943 0.940 0.939
Tap

T4-12 0.90 1.10 1.032 0.977 0.982 0.996 1.025 1.012 1.000 0.978 0.986 0.980
T28-27 0.90 1.10 1.068 0.970 0.978 0.982 0.975 0.990 0.992 0.979 0.979 0.979
Qc10 0.0 5.0 0.00 4.05 1.65 4.05 2.28 0.31 1.97 3.82 4.00 3.88
Qc12 0.0 5.0 0.00 4.68 4.87 4.42 2.11 1.77 2.68 4.51 4.99 4.97
Qc15 0.0 5.0 0.00 3.95 4.29 4.19 4.36 2.60 3.57 4.68 5.00 4.89
Shunt Elements

Qc17 0.0 5.0 0.00 4.84 3.62 4.98 1.44 2.21 0.90 4.97 5.00 4.96
(Mvar)

Qc20 0.0 5.0 0.00 3.76 4.36 4.86 0.83 0.95 2.30 4.77 4.99 4.96
Qc21 0.0 5.0 0.00 4.98 4.81 4.90 3.36 1.96 0.26 4.94 5.00 5.00
Qc23 0.0 5.0 0.00 4.23 3.98 4.96 3.51 1.80 3.43 4.91 5.00 4.99
Qc24 0.0 5.0 0.00 4.38 4.39 4.90 4.39 3.75 4.95 3.64 3.82 3.67
Qc29 0.0 5.0 0.00 1.91 3.76 4.90 1.03 0.39 2.88 4.83 5.00 4.92
OBJECTIVE VALUES
Fuel Cost 901.96 800.56 963.27 856.90 801.54 931.13 828.59 800.58 800.62 800.59
Losses (MW) 5.83 8.91 3.24 5.40 9.05 3.57 5.69 9.04 9.07 9.05
L-index 0.1720 0.1275 0.1273 0.1246 0.1380 0.1399 0.1324 0.1251 0.1248 0.1249

REFERENCES
V. CONCLUSIONS [1] J.A Momoh, M.E. El-Hawary and R. Adapa, “A review of selected
optimal power flow literature to 1993, Part I: non-linear and quadratic
In this paper, a true multi-objective optimal power flow was programming approaches” IEEE Trans. on Power Systems, vol. 14, no.
formulated. The problem was solved using a proposed 1 pp. 96-104, Feb. 1999.
multiobjective differential evolution based approach. A source [2] J.A Momoh, M.E. El-Hawary and R. Adapa, “A review of selected
code in Fortran was developed. Then, both single objective optimal power flow literature to 1993, part ii: newton, linear
programming and interior point methods” IEEE Trans. on Power
and multi-objective OPF cases were simulated using the Systems, vol. 14, No. 1, pp. 105-111, Feb. 1999.
standard IEEE-30 bus test system. Also, the program was [3] M. Huneault and F.D. Galiana, “A survey of the optimal power flow
supported with subroutine to perform clustering and extract literature”, IEEE Trans. on Power Systems, vol. 6, no. 2, pp. 762-770,
May 1991.
the best compromise solution to assist the decision maker. [4] M. Abido, “Optimal power flow using particle swarm optimization”
Comparison with the available results in the literatures showed International Journal of Electrical Power and Energy Systems, vol. 24,
that the proposed differential evolution has shown fast no. 7, , pp:563-571, 2002.
[5] R. E. Perez-Guerrero and J.R Cedeno-Maldonado, “Differential
convergence and ability to capture large number of Pareto evolution based economic environmental power dispatch", Proceedings
solutions. The results shows promising performance indicating of the 37th Annual North American Power Symposium, pp. 191-197, 23-
its effectiveness in solving true multi-objective OPF and also 25 Oct. 2005.
[6] M. Abido, "Environmental/economic power dispatch using
finding well distrusted Pareto solutions. multiobjective evolutionary algorithms", IEEE Trans. on Power
Systems, vol. 18 , issue 4, pp 1529 – 1537, Nov. 2003.
[7] M. Abido, “Multiobjective optimal power flow using strength pareto
evolutionary algorithm”, 39th International Universities Power
ACKNOWLEDGMENT Engineering Conference, 2004. UPEC 2004, vol. 1, pp:457 – 461, 6-8
The authors acknowledge the support and encourage of Sept. 2004.
[8] M. A. Abido, "Multiobjective evolutionary algorithms for electric power
King Fahd University of Petroleum & Mineral. dispatch problem", IEEE Trans. on Evolutionary Computation, vol. 10,
issue 3, pp. 315 – 329, June 2006.
[9] M.A Abido and J. M. Bakhashwain, “A novel multiobjective
evolutionary algorithm for optimal reactive power dispatch problem”,

105
POWERENG 2009 Lisbon, Portugal, March 18-20, 2009

Proceedings of the 2003 10th IEEE International Conference on

Electronics, Circuits and Systems, 2003. ICECS, vol. 3, 14-17 Dec.
2003, pp. 1054-1057.
[10] P.X. Zhang, B. Zhao, Y.J. Cao and S.J. Cheng, “A novel multi-objective
genetic algorithm for economic power dispatch”, 39th International
Universities Power Engineering Conference, vol. 1, 6-8 Sept. 2004 pp.
422 - 426.
[11] Bo ZHAO, Yi-jia CAO, “Multiple objective particle swarm optimization
technique for economic load dispatch”, Journal of Zhejiang University
SCIENCE, vol. 6A, no. 5, pp. 420-427, 2005.
[12] R. C. Bansal, “Optimization Methods for Electric Power Systems: An
Overview”, International Journal of Emerging Electric Power Systems,
vol. 2, no. 1, Article 1021.
[13] I. Dobson, H. Glavitsch, C.-C Liu, Y. Tamura and K. Vu, “Voltage
collapse in power systems”, IEEE Circuits and Devices Magazine, vol.
8, issue 3, pp. 40 – 45 , May 1992.
[14] C.A. Belhadj and M. A. Abido, “An optimized fast voltage stability
indicator”, International Conference on Electric Power Engineering,
1999. PowerTech Budapest 99. on 29 Aug.-2 Sept. 1999, pp. 79.
[15] E. Zitzler, M. Laumannus and S. Bleuler, “A tutorial on evolutionary
multi-objective optimization”, Swiss Federal Institute of Technology
(ETH) Zurich, Computer Engineering and Networks Laboratory (TIK)
Golriastrasse 35, CH-8092 Zurich, Switzerland. Available at:
(www.cs.cinvestav.mx/~emoowork group/zitzler04.pdf).
[16] C. Fonseca and P. Fleming, “An overview of evolutionary algorithms in
multiobjective optimization”, Evolutionary Computation, 3(1):1-16,
Spring 1995. Available at: (http://www.lania.mx/~ccoello/EMOO
/fonseca95.ps.gz).
[17] A. Carlos., Coello Coello, “A comprehensive survey of evolutionary-
based multiobjective optimization techniques”, Knowledge and
Information Systems. An International Journal, 1(3): pp. 269-308, Aug.
1999. Available at
(http://www.lania.mx/~ccoello/EMOO/informationfinal. ps.gz).
[18] R. Storn and K. Price, “Differential evolution – a simple and efficient
adaptive scheme for global optimization over continuous spaces” ,
Technical Report TR-95-012, Available at: (ftp://ftp.icsi.berkeley.edu/
pub/techreports /1995/tr-95-012.pdf).
[19] R. Storn and K Price, “Minimizing the real functions of the icec'96
contest by differential evolution”, Proceedings of IEEE International
Conference on Evolutionary Computation, 20-22 May 1996, pp. 842–
844.
[20] J Vesterstrom and R Thomsen, “A comparative study of differential
evolution, particle swarm optimization and evolutionary algorithms on
numerical benchmark problems”, Congress on Evolutionary
Computation, 2004. CEC2004, vol. 2, 19-23 June 2004, pp 1980-1987.
[21] N. Madavan, “Multiobjective optimization using a pareto differential
evolution approach”, Proceedings of the 2002 Congress on Evolutionary
Computation, 2002. CEC '02. vol. 2, 12-17 May 2002, pp. 1145–1150.
[22] F. Xue, A. C. Sanderson and R. J. Graves, “Pareto-based multi-objective
differential evolution”, The 2003 Congress on Evolutionary
Computation, 2003. CEC '03, vol. 2, 8-12 Dec. 2003, pp. 862-869..
[23] K. Lee, Y Park and J Ortiz, “A united approach to optimal real and
reactive power dispatch”, IEEE trans. on Power apparatus and systems,
vol. PAS-104, No 5 May 1985, pp 1147-1153.

106