Vous êtes sur la page 1sur 11

Energy Conversion and Management 77 (2014) 597607

Contents lists available at ScienceDirect

Energy Conversion and Management


journal homepage: www.elsevier.com/locate/enconman

A modied teachinglearning based optimization for multi-objective


optimal power ow problem
Amin Shabanpour-Haghighi a, Ali Reza Sei a,, Taher Niknam b
a
School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran
b
Department of Electronic and Electrical Engineering, Shiraz University of Technology, Shiraz, Iran

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a modied teachinglearning based optimization algorithm is analyzed to solve the multi-
Received 6 July 2013 objective optimal power ow problem considering the total fuel cost and total emission of the units. The
Accepted 12 September 2013 modied phase of the optimization algorithm utilizes a self-adapting wavelet mutation strategy. More-
over, a fuzzy clustering technique is proposed to avoid extremely large repository size besides a smart
population selection for the next iteration. These techniques make the algorithm searching a larger space
Keywords: to nd the optimal solutions while speed of the convergence remains good. The IEEE 30-Bus and 57-Bus
Optimal power ow
systems are used to illustrate performance of the proposed algorithm and results are compared with
Multi-objective problem
Modied teachinglearning based
those in literatures. It is veried that the proposed approach has better performance over other
optimization techniques.
Pareto-optimal set 2013 Elsevier Ltd. All rights reserved.

1. Introduction eters such as those in GA, PSO and ACO. [11]. This algorithm is used
to solve the economic dispatch problem [12] but there is no re-
Optimal power ow (OPF) is a non-linear programming prob- search on solving the OPF problem by this method.
lem that species the optimal control points of a power network The OPF problem may have various objective functions. The
to minimize a desired objective, commonly cost of generation, sub- most commonly used objective is the minimization of the overall
ject to a set of certain system constraints [1]. Generally, the OPF fuel cost of generators. However, other traditional objectives are
problem is a large-scale highly constrained non-linear non-convex minimization of active power loss, bus voltage deviation, emission
optimization problem [2]. There are some mathematical tech- of units, number of control actions, and load shedding [13]. A mul-
niques to solve this problem such as linear and non-linear pro- ti-objective optimal power ow (MO-OPF) problem is made if more
gramming [36], quadratic programming [7], and the interior than one of these objective functions should be optimized [14].
point method [8,9]. All of these methods may be trapped in local Several techniques are used to solve a multi-objective optimiza-
minima that prevent the algorithm reaching the true optimal solu- tion problem (MOP). One method is combining all objective func-
tion. Other disadvantages of these techniques are enormous com- tions into one objective function using weighting factors. It nds
putational effort and time consumption. Recent methods to deal just one solution for the problem that is very dependent on the
with the shortcomings of these mathematical approaches are evo- weighting factors and this is the main disadvantage of this tech-
lutionary algorithms such as genetic algorithm (GA), tabu search nique. Another way to deal with a multi-objective problem is solv-
(TS), particle swarm optimization (PSO) and ant colony optimiza- ing all objective functions simultaneously using evolutionary
tion (ACO). methods. Since these algorithms are population-based techniques,
One of the recently proposed techniques to solve the optimiza- multiple Pareto-optimal solutions can be found in one program run
tion problems is the teachinglearning based optimization (TLBO) [15].
[10]. This technique is a new efcient optimization algorithm that The aim of this paper is to solve the MO-OPF problem using the
has been inspired by learning mechanism in a class. The main modied TLBO algorithm. A self-adaptive mutation wavelet tech-
advantage of this method over the other evolutionary algorithms nique is proposed to deal with the search capability, population
is that TLBO is an algorithm-parameter-free technique and the diversity, and convergence speed. An external repository is utilized
effectiveness of the method is not affected by the algorithm param- to save all non-dominated optimal solutions during the process.
Then a fuzzy decision making method is applied to sort these solu-
tions according to their importance and decision makers can
Corresponding author.
choose the desired solution between the Pareto-optimal solutions
E-mail addresses: shabanpour.amin@gmail.com (A. Shabanpour-Haghighi),
sei@shirazu.ac.ir (A.R. Sei), niknam@sutech.ac.ir (T. Niknam).
by applying this fuzzy decision making mechanism. Furthermore,

0196-8904/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.enconman.2013.09.028
598 A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607

a fuzzy clustering approach is employed to decrease size of the where F2(X) is the total emission of units (ton/h), ai, bi, ci, ni, and ki
repository without losing its characteristics. Also, a smart popula- are the emission coefcients of the ith generator.
tion selection is used to choose population of the next iteration of
the algorithm efciently. Simulations are done on the standard 2.3. Equality constraints
IEEE 30-Bus and 57-Bus systems and results are compared with
the conventional TLBO and other evolutionary methods. The effec- The OPF equality constraints show the power ow feasibility
tiveness and advantages of the proposed method is veried by var- that can be expressed as:
ious criterions. Nb
X  
The rest of the paper is organized as follows. In Section 2, math- Pgi  Pdi  V i V j Gij cos hij Bij sin hij 0 9
ematical formulation of the OPF objective functions are reviewed. j1
In Section 3, the TLBO algorithm and its proposed modication
are described in details, while Section 4 consists of some overview Nb
X
on the multi-objective problems. Moreover, the utilized fuzzy Q gi  Q di  V i V j Gij sin hij  Bij cos hij 0 10
method and the smart population selection are described in this j1

section. Application of the proposed algorithm for the MO-OPF is


where i = 1,2,. . .,Nb and Nb is the number of buses. Pgi and Qgi are the
determined in Section 5 followed by some case studies and simu-
generated active and reactive power at the ith bus. Also Pdi and Qdi
lations in Section 6. Finally, Section 7 concludes the discussion.
are the active and reactive power demand at the ith bus. Vi and Vj
are the voltage magnitude at the ith and jth buses and hij = hi  hj
2. Mathematical formulation is the difference between the voltage angle of these buses. The
mechanism of handling the equality constraint is shown in Fig. 1.
In this section a brief review on the objective functions of the It is worthwhile to note that this constraint should satisfy the
OPF problem is presented. The total cost of generation and the total equality of generation level with the load level plus losses in the
emission from units are two objective functions that are used in system. The process of the algorithm is started by choosing gener-
this paper. ator output powers randomly within their limits. The difference of
the total power generation from the total demand plus losses is
2.1. Generation cost function generated by the slack. If slack is reached to its limits, the algorithm
chooses another generator randomly to help the slack satisfying the
It is necessary to consider the valve-point effect of generators to equality constraint. For each population member of the optimiza-
obtain a more accurate model [1618]. Thus, the ripple curve cre- tion algorithm, his process is repeated until the equality constraint
ated by the valve-point effect is modeled in this paper and the total is satised.
cost of generation can be formulated as follows:
0 2 1
XNg ai Pgi bi P gi ci 2.4. Inequality constraints
F 1 X @  h   i A 1
 
i1 di sin ei Pmin gi  P gi  Limits of the OPF variables are expressed by inequality con-
  straints as follows:
X P g ; V g ; T; Q c 1N 2
Pmin
gi 6 Pgi 6 P max
gi ; i 1; 2; . . . ; Ng 11
P g Pg1 ; Pg1 ; . . . ; P gNg 1 1Ng 1 3
Q min
gi 6 Q gi 6 Q max
gi ; i 1; 2; . . . ; Ng 12
V g V g1 ; V g1 ; . . . ; V gNg 1N 4
g
V min
i 6 V i 6 V max
i ; i 1; 2; . . . ; NL 13
T T 1 ; T 2 ; . . . ; T Nt 1Nt 5
T min
i 6 T i 6 T max
i ; i 1; 2; . . . ; Nt 14
Q c Q c1 ; Q c2 ; . . . ; Q cNc 1Nc 6
Q min
ci 6 Q ci 6 Q max
ci ; i 1; 2; . . . ; Nc 15
N Ng  1 Ng Nt Nc 7
jP ij j 6 P max
ij 16
where F1(X) is the total cost of generation ($/h), ai, bi, ci are the
fuel cost coefcients of the ith generator, di and ei shows the where NL is the total number of load buses and Pij is the active
valve-point effect of generators. Pgi is the generated active power power ows between bus i and bus j. The superscript min and
of the ith unit and Pmingi is its lower margin. Control variables of max shows the lower and upper limits of their respected variables.
the optimization problem is named as X that includes the gener-
ated active power vector, Pg, the voltage magnitude vector of 3. Teachinglearning based optimization
units, Vg, the tap of transformers vector, T, and the reactive power
vector of shunt capacitances, Qc. Furthermore Ng, Nt, and Nc are the TLBO is a new population based optimization algorithm with
total number of generators, the total number of tap transformers, efcient calculation demand. It is inspired by the inuence of a tea-
and the total number of shunt capacitances in the power system, cher on his learners in a class [10]. The conventional algorithm has
respectively. two modes: (1) teacher phase and (2) learner phase. However a
third phase is added to the algorithm in the modied TLBO
2.2. Emission function (MTLBO) method.

The emission objective function can be represented as below: 3.1. Teacher phase
Ng
X   
F 2 X ai bi Pgi ci Pgi n exp ki P gi 8 The teacher is considered as the most knowledgeable person in
i1 a class who shares his knowledge with the students to improve the
A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607 599

Fig. 1. Flowchart of satisfying equality constraint.

output (i.e. grades or marks) of the class [12]. The teacher improves 3.3. Modied phase
the mean of the class up to his level. Let Mi and Ti be the mean and
the teacher at the ith iteration. As it is mentioned above, Ti tries to The interactions in the learner phase may lead to inappropriate
enhance Mi to his own level. So the teacher phase of the algorithm knowledge exchange between learners in the way that the method
can be formulated as follows: can be trapped in local minima. Therefore, another learning modi-
cation phase is applied to overcome this shortcoming. This ap-
DMi randi  T i  T F Mi 17 proach consists of a self-adapting wavelet mutation (SAWM)
strategy to improve the performance of the original TLBO algo-
T F round1 rand0; 1 18 rithm that makes faster convergence and avoids getting trapped
in local minima. Using wavelet theory, the SAWM technique can
dynamically change the mutation space along the convergence
X new
i X old
i DM i 19 procedure. This technique is based on the fact that the students
are tends to move in the similar way with the teacher and avoid
where randi is a random number between [0, 1] and TF is the teach- moving to the worst way. A mutation probability, PX, is dened
ing factor that decides the value of the mean to be changed. The for each learner that its value is between [0, 1]. Then a random
new value of Xi is accepted if its tness function value is better than number between 0 and 1 is generated and compared with PX.
the old one. The mutation takes happened if that random number is equal or
less than PX. The new position of the learner can be calculated as
follows:
3.2. Learner phase 8  
>
< X Old x  T i  X Old if x > 0
i i
Each learner interacts with other learners randomly via group X New   20
i
discussions, presentations, formal communications, etc. to increase >
: X Old x  W i  X Old if x 6 0
i i
his knowledge [19]. The process of this phase is as follows:
where Wi is the worst student in each iteration. Again the accep-
1. Randomly select two learners Xi and Xj where i j. tance of the new Xi is dependent on whether its tness value is im-
2. According to the value of their tness function, if F(Xi) < - proved or not. The value of x is calculated by the Morlet wavelet
F(Xj) then the new learner is X new
i randi  X i  X j else function as below:
the new learner is X new randi  X j  X i .
 2
i
1 1 u  u
3. The replacement procedure can be implemented if the t- x p exp  cos xc 21
ness value of the new Xi is better than the old one. h 2 h h
600 A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607

where xc is the central frequency of the wavelet. In this paper, 4.2. External repository
xc = 5 is selected. A larger value of |x| makes a larger change in
the mutation and vice versa. Furthermore, a positive value of x The proposed MTLBO uses an external archive called repository
makes the mutant student moves toward the teacher. Conversely, to save the non-dominated solutions found so far. In each itera-
if x is negative the mutant learner recedes from position of the tions of the MTLBO algorithm, besides the optimization process,
worst learner. Due to the fact that 99% of the total energy of the solutions found by each phase of the algorithm are compared with
mother wavelet function is located between the intervals the repository members. The new non-dominated solutions are
[2.5, +2.5], the value of parameter u can be randomly generated stored in the repository. Also the algorithm removes the domi-
between [2.5h, +2.5h]. The algorithm uses a larger search space nated members of the repository. The repository size of most opti-
at the beginning to nd the nearly global best solution and then mization problems could be increased extremely large during the
the search space is restricted around this global solution to increase optimization algorithm. It is obvious that the large number of indi-
the accuracy of the nal solution. In this context, the dilation viduals in the repository may leads to more computation burden
parameter, h, is changed iteration by iteration in order to reach to and even the memory constraints. So limiting size of the repository
a ne-tuned value. It is set to a small value at the beginning that without losing the characteristic and quality of the POF is neces-
makes the value of |x| being large enough to produce a larger sary. In this paper, a repository with a determined size is used
searching space and then it increases in each iteration to make a and the new solution would be added to it if one of the following
smaller value of |x|, and hence a smaller searching space. As a re- conditions is satised:
sult, h can be calculated as:
   The repository is empty.
r
h exp  lng  1  k=kmax lng 22  The new solution dominates one of the individuals of the
repository.
where k and kmax are the current iteration and the total number of  The repository is not full and the new solution is not dominated
iterations, respectively. The upper limit and shape format of by any individuals of the repository.
increasing function of h can be dened by g and r, respectively.  The repository is full but the new solution is not dominated by
The value of r critically affects the performance of the algorithm. any individuals of the repository and it is in a less crowded area
In order to achieve the exploration capability of the algorithm and of the POF than at least one of the existing members of the
the accuracy of the nal results, the value of r should be dened repository.
correctly. To do so, it is set to a small value at the beginning and in-
creased iteration by iteration as follows: The detailed process of limiting size of the repository is

explained in the next subsection.
r  rmin
r rmin max k 23
kmax
4.3. Fuzzy decision making method
where the upper and lower limits of r are dened by rmax and rmin,
respectively.
4.3.1. Fuzzy based clustering
Whenever the repository is full, in order to nd out whether the
4. Multi-objective problem new non-dominated solution should be replaced by one of the
repository members or not, a fuzzy decision making strategy is
4.1. Principles of MOP used. Since the objective functions have conicting behaviors in
the human judgment, decision maker can determine his prefer-
The aim of solving a multi-objective optimization problem is to ences for corresponding fuzzy objectives in fuzzy set to derive a
optimize several incomparable and conicting objectives simulta- Pareto-optimal solution efciently [21]. In a multi-objective prob-
neously. The algorithm to solve the MOP nds a set of non- lem, a fuzzy membership function is given to each objective. Mem-
dominated solutions known as Pareto-optimal front (POF) that bers with better tness value are given a higher value of the fuzzy
generally each solution has no priority over the others. Suppose membership function and vice versa. Assume that the ith objective
X1 and X2 are two different solutions of the system. X1 dominates function Fi(X) has the lower and upper margins called F min
i and F max
i ,
X2 and X1 is called the non-dominated solution when the following respectively. These margins are obtained by considering only this
conditions are satised [20]: objective function. In other word, other objective functions of the
8 MO-OPF are not considered to obtain these boundaries. Then the
< 8i 2 f1; 2; . . . ; ng; F i X 1 6 F i X 2
> fuzzy membership function of Fi (X) can be expressed by:
9j 2 f1; 2; . . . ; ng; F j X 1 < F j X 2 24
>
: 8
>
> 1 ; F i X 6 F min
i
< max
Fi F i X
li X F max F min
; F min
i 6 F i X 6 F max
i 26
where n is the number of objective functions. Therefore, the POF is >
>
: i i

obtained by mapping these non-dominated solutions into the 0 ; F i X P F max


i
objective space as follows:

POF fF 1 X; F 2 X; . . . ; F n XjX 2 Sg 25 In the proposed MO-OPF problem, the obtained POF may be
enormously large. In this regards, each individual of the repository
where S is the set of non-dominated solutions found by the constitutes a cluster with denite radius. Adjacent clusters are
algorithm. merged until the required repository size is obtained. In the merg-
This paper uses the capability of the MTLBO algorithm in ing process, the member with higher membership value of joining
searching the feasible space and develops it to solve the MO-OPF clusters is selected to store in the repository. The whole procedure
problem. Some modications are suggested in the proposed of proposed fuzzy based clustering is shown in Fig. 2. Note that in
MTLBO method to improve the algorithm which is discussed in this owchart, S is the number of non-dominated solutions, Nmax is
the next subsections. the maximum size of the repository, Ni and Nj are the numbers of
A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607 601

4.3.3. Choosing teacher of the MTLBO


As it is mentioned before, the teacher in the TLBO algorithm is
the one that its tness function is better that all the other learners.
Therefore, a fuzzy decision maker should be implemented to nd
out the best solution, and hence the teacher, in each iteration. To
do so, a decision making function is described as [22]:
nP
ri lki
lk PM i1
Pn 28
i1 ri li
k
k1

where M is the number of non-dominated solutions (the number of


Pareto-optimal solutions) and lk is the membership value of the
non-dominated solution k. Then the Pareto-optimal solution that
obtains the maximum value of lk is selected as the teacher for
the next iteration. Note that ri is the desired weighting factor of
objective function i that can be specied based on the expert expe-
P
rience or the trial and error method. It is obvious that ni1 ri 1 .

4.3.4. Smart population


As it is mentioned before, the MO-OPF is a non-linear, non-con-
vex, non-smooth, and high-dimension optimization problem.
Therefore, nding a POF with uniformly distribution for this prob-
lem is so difcult.
In this paper, a heuristic method is used to select population of
the following iteration from the solutions obtained by the previous
iterations. This approach consists of two phases.

4.3.4.1. Phase 1. When the repository is not lled, the MTLBO is


going to nd new non-dominated solutions. These non-dominated
individuals are better for the next iteration to encourage the stu-
dents of the MTLBO to search for more Pareto-optimal solutions.
So the population is sorted as follows: Set S0: members of the cur-
rent repository; Set S1: individuals that are dominated one time;
Set S2: individuals that are dominated two times; Set S3: individu-
als that are dominated three times; etc. Thus the population of the
next iteration is determined by this priority sets where it consists
the repository members S0 followed by the individuals of set S1, S2,
S3, . . . until the maximum size of the population is reached for the
next iteration.

4.3.4.2. Phase 2. Whenever the repository is lled, the algorithm


puts its effort on nding the near global solutions. To do so, %R
of the repository members is chosen and the rest of the population
of the next iteration is selected randomly from the current popula-
tion. This procedure makes the evolutionary algorithm nding the
global and near global solutions in a non-convex and high-dimen-
sional problem %R can be dened as below:
!
pkmax  k 
%R sin   100 29
P kmax  krep;filled

Fig. 2. Flowchart of fuzzy based clustering. where k, kmax, and krep,lled are the current iteration, the maximum
number of iterations, and the iteration that the repository is lled,
respectively. The value of constant P denes the performance of R
solutions in clusters Ci and Cj, respectively. The function that calcu- from an approximately linear characteristic to a periodic one.
lates the distance between solutions X and Y is called d (X, Y). Fig. 3 shows this characteristic for different values of P. Note that
in this gure krep,lled = 50 and kmax = 300. In this paper P = 2 is
4.3.2. Best compromise solution selected.
To nd out the best compromise solution among the nal
repository members, the fuzzy membership functions of all objec- 5. Application of the MTLBO to MO-OPF
tives are extracted separately and the fuzzy solution can be calcu-
lated as follows: The following steps should be done to solve the MO-OPF using
MTLBO.
FX minl1 X; . . . ; ln X 27

where n is the number of objective functions. The best compromise Step 1: Dene the input data such as line data, load parameters,
solution is the maximum value of F(X) among other values. fuel cost coefcients and emission coefcients of generators,
602 A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607

Table 1. The initial population of the MTLBO algorithm and the


maximum number of iterations are set to 100 and 200, respec-
tively. The lower and upper voltage magnitude limits of buses
and tap transformers are set between [0.95, 1.05] and [0.9, 1.1]
per-unit, respectively. The objective functions that are used in this
paper are the total cost of generation and total emission of the
units. Two different test cases are analyzed. Accuracy of the pro-
posed method is shown in the rst case. The second case surveys
the multi-objective approach to solve the MO-OPF using the
MTLBO algorithm.

6.1.1. Single-objective OPF


In this case, a comparison between various algorithms that can
be used to solve the OPF problem is done to verify the accuracy of
Fig. 3. Variations of %R for different P.
the proposed method. All objective functions are optimized indi-
vidually to obtain the lower and upper limits of each objective
initial value of generator active powers, initial value of the tap for the MO-OPF problem which is analyzed in the next test case.
position of transformers, initial value of the reactive power gen- Tables 2 and 3 show the obtained results by the TLBO, the MTLBO,
erated by shunt capacitances, and lower and upper limits of Pgi, and other evolutionary algorithms proposed in literatures. In Ta-
Qgi, Vi, Ti and Qci. Moreover the marginal values of all objective ble 2, best fuel cost of generators that are achieved by TLBO and
functions, F min and F max , should be specied. MTLBO algorithms are shown in comparison with other algorithms
i i
Step 2: Initialize the learners within the feasible range. include tabu search (TS) [24], evolutionary programming (EP) [25],
Step 3: Check for the equality constraints as shown in Fig. 1. improved evolutionary programming (IEP) [26], differential evolu-
Step 4: Calculate the tness value of objective functions for each tion (DE-OPF) [27], modied differential evolution (MDE-OPF) [27],
learner and nd out the fuzzy membership value of each one. stochastic search algorithm (SSA) [28], enhanced genetic algorithm
Step 5: Check for the non-dominated solutions and identify the (EGA) [29], ant colony optimization (ACO) [30], fuzzy genetic algo-
teacher. rithm (FGA) [31], particle swarm optimization (PSO) [18], honey
Step 6: Calculate the mean value of the learners. bee mating optimization (HBMO) [16], and modied honey bee
Step 7: Run the teacher phase of the MTLBO algorithm using mating optimization (MHBMO) [16]. It is clear that the MTLBO
(17) to (19). Check for any improvement in the situation of has better performance and better results can be achieved by this
learners and apply the required modications. method in comparison with other evolutionary algorithms. Also,
Step 8: Check for the non-dominated solutions and update the this method is better than the original TLBO. These values are
repository. not negligible because of the continuous operations of power dis-
Step 9: Run the learner phase of the algorithm and modify the patch throughout the years as well as the numerous power plants
position of learners if they have been improved. in the entire world [15]. Best emissions of units by different algo-
Step 10: Check for the non-dominated individuals and update rithms are shown in Table 3. Again, result of the MTLBO method is
the repository. better than the other algorithms. In order to verify the effective-
Step 11: Run the modied phase of the algorithm according to ness and robustness of the proposed MTLBO algorithm to solve
(20) to (23). Update the learner positions if any enhancement the OPF problem, the mentioned 30-Bus system is solved for 30 tri-
is happened. als. The best, the average and the worst result for the total fuel cost
Step 12: Check for the new non-dominated solutions and update and the emission objective functions are shown in Figs. 4 and 5,
the repository. respectively.
Step 13: Determine the population for the next iteration as Table 4 shows the mean value and the standard deviation (SD)
described in subsection 4.4. of these trials and the superiority of the modied algorithm is obvi-
Step 14: Go to step 4 until the maximum number of iteration is ous compared with the original one.
reached. The convergence property of the original TLBO method along
with the proposed MTLBO algorithm is shown in Fig. 6. It is obvious
that in MTLBO, the included modied phase improves performance
6. Case studies of the algorithm, increases the convergence speed, and obtains bet-
ter nal results.
6.1. The standard IEEE 30-Bus system The successful percentage of the implemented MTLBO is shown
in Table 5 to verify robustness of the technique. This parameter can
The MTLBO algorithm is applied to the IEEE 30-Bus system to be dened as the number of successful runs which converge to the
verify its performance. This system has six generators and four best solution divided by the total number of runs [32]. Results
tap transformers. The parameters of this system are given in show that the MTLBO method is more successful in nding satis-
[16,23]. The emission coefcients of generators are shown in factory solution compared with the original TLBO.
Population size of an optimization algorithm has a great impact
on the convergence speed and accuracy of the nal results. It is re-
Table 1 lated to the search space dimension and complexity of the prob-
Emission coefcients of 30-Bus IEEE system.
lem. However, very large or too small populations may not be
G1 G2 G3 G4 G5 G6 reached the optimal solution, especially in the non-convex multi-
a 0.04091 0.02543 0.04258 0.05326 0.04258 0.06131 dimensional problems. In this study, different population sizes
b 0.05554 0.06047 0.05094 0.0355 0.05094 0.05555 are used to carry out the OPF problem in 30 independent trials
c 0.0649 0.05638 0.04586 0.0338 0.04586 0.05151 and the effect of each population size on the performance of the
n 0.0002 0.0005 0.000001 0.002 0.000001 0.00001
MTLBO is evaluated. Table 6 shows the results. The problem is
k 2.857 3.333 8 2 8 6.667
solved by 20, 40, 70, 100, 150, and 200 population sizes. The
A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607 603

Table 2
Best fuel cost of generators obtained by various optimization algorithms.

PG1 PG2 PG3 PG4 PG5 PG6 Total Generation Total Cost
TS [24] 176.04 48.76 21.56 22.05 12.44 12 292.85 802.29
EP [25] 173.848 49.998 21.386 22.63 12.928 12 292.79 802.62
IEP [26] 176.2358 49.0093 21.5023 21.8115 12.3387 12.0129 292.9105 802.465
DE-OPF [27] 176.009 48.801 21.334 22.262 12.46 12 292.866 802.394
MDE-OPF [27] 175.974 48.884 21.51 22.24 12.251 12 292.859 802.376
SSA [28] 192.5105 48.3951 19.5506 11.6204 10 12 294.0766 804.1072
EGA [29] 176.2 48.75 21.44 21.95 12.42 12.02 292.78 802.06
ACO [30] 181.945 47.001 21.4596 21.446 13.207 12.0134 297.072 802.578
FGA [31] 175.137 50.353 21.451 21.176 12.667 12.11 292.894 802.0003
PSO [18] 175.6915 48.639 21.4494 22.72 12.2302 12 292.7301 802.0136
HBMO [16] 176.4646 46.274 21.4596 21.446 13.207 12.0134 292.8646 802.211
MHBMO [16] 177.0431 49.209 21.5135 22.648 10.4146 12 292.8242 801.985
TLBO 177.3986 48.0701 21.7722 21.89616 12.08228 11.61326 292.8326 801.9908
MTLBO 177.2561 48.0762 21.1925 22.1182 12.1124 11.821 292.5764 801.8925

Table 3
Best emission of generators obtained by various optimization algorithms.

PG1 PG2 PG3 PG4 PG5 PG6 Total emission


GA [15] 69.73 67.84 49.73 34.42 29.15 39.29 0.20723
PSO [15] 67.13 68.94 49.86 34.89 29.67 39.94 0.2063
IPSO [15] 67.04 68.14 50 35 30 40 0.2058
TLBO 63.5221 68.7345 49.9931 34.9894 29.9824 39.9801 0.205
MTLBO 64.2924 67.625 50 35 30 40 0.20493

802.6
Best
802.4
Average
802.2 Worst

802

801.8

801.6

801.4
TLBO MTLBO

Fig. 4. Best, average and worst result of 30 trails for the total fuel cost of generators.

0.216
Fig. 6. Convergence graphs of different TLBO methods.
Best
0.212 Average
Worst
0.208 Table 5
Successful percentage of various methods to reach the best result in 30 trials.
0.204
Total fuel cost ($/h)

0.2 801.8802 802802.2 802.2802.4 802.4802.6


(%) (%) (%) (%)
0.196 TLBO 6.66 26.66 53.33 13.33
TLBO MTLBO MTLBO 100 0 0 0

Fig. 5. Best, average and worst result of 30 trails for the total emission of
generators.

Table 6
Effect of population size for 30 trails.
Table 4
The mean and the standard deviation of TLBO and MTLBO for 30 trails. Population size No. of hits to 801.8802 Average CPU time (min)

Total Total 20 8 0.325


Fuel cost Emission 40 17 0.417
70 23 0.592
Mean SD Mean SD
100 30 0.717
TLBO 802.12 0.61 0.21 0.39 150 26 1.021
MTLBO 801.95 0.23 0.2078 0.19 200 18 1.463
604 A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607

Table 7 Table 8
Control variables related to multi-objective problem for the TLBO and MTLBO C-Metric values in 30 trials.
methods.
C(TLBO, MTLBO) C(MTLBO, TLBO)
Total fuel cost Total emission
Best 0.005 1
TLBO MTLBO TLBO MTLBO Average 0.002 0.998
Worst 0 0.995
PG1 177.3986 177.2561 63.5221 64.2924
PG2 48.0701 48.0762 68.7345 67.625
PG3 21.7722 21.1925 49.9931 50
ja2 2 A2 ; 9a1 2 A1 : a1 < a2 j
PG4 21.89616 22.1182 34.9894 35 CA1 ; A2 30
PG5 12.08228 12.1124 29.9824 30 jA2 j
PG6 11.61326 11.821 39.9801 40
VG1 1.0498 1.049 1.0094 1.05
VG2 1.0381 1.0391 1.0014 1.0459 C(A1, A2) = 1 if All the members of A2 are dominated by the
VG3 1.0102 1.0127 0.97698 1.0284 members of A1 and C(A1, A2) = 0 if none of the members of A2 are
VG4 1.0173 1.0179 0.98951 1.0361 dominated by the members of A1.
VG5 1.05 1.05 1.0435 1.05
VG6 1.0499 1.0499 1.0487 1.04
T6-9 1.05 1.06 0.98 1.06
T6-10 0.91 0.9 0.95 0.91 6.1.2.3. D-Metric. This criterion shows the spread of the obtained
T4-12 1.05 1.06 1.02 1.04 Pareto-optimal set [34]. It is dened as follows:
T27-28 0.98 0.99 0.95 0.98
Cost 801.9908 801.8925 947.4392 945.1965 PM1 
df dt jdi  dj
i1
Emission 0.3668 0.3665 0.20503 0.20493 D  31
df dt M  1d

maximum iteration size is set to 200 for all of them. The average where di is the Euclidean distance between neighboring solutions in
CPU time is also reported in Table 6. It can be seen that the popu- the obtained non-dominated solutions set. df and dt are the Euclid-
lation size of 100 has been reached to more consistent solution. ean distance between the extreme solution and the boundary solu-
tion of the obtained Pareto-optimal set and d is mean value of all d .
The obtained operation points by considering each objective i

functions separately are shown in Table 7. Results of the cost-based A smaller value of D-Metric indicates a better distribution of the
OPF cause the emission to increase by 78.89% of the emission- Pareto-optimal solutions [34].
based OPF. Moreover, the emission-based OPF makes the cost of The MTLBO algorithm is compared with the original TLBO
generation increasing by 17.88% compared to the cost-based OPF. method in Tables 8 and 9. The MTLBO method utilizes a modied
It is clear that the best operating point of an objective function phase that makes the algorithm searching a larger space to nd
may not satisfy other objective functions. Hence a multi-objective the optimal solutions. So its nal solution set would be better than
optimization should be done to achieve the optimal operating the original TLBO method. The average, the best and the worst C-
point of the power system that satises all of the objective func- Metric values that are calculated in 30 trails are shown in Table 8.
tions simultaneously. Results show that the average value of C(MTLBO, TLBO) is larger
than the average of C(TLBO, MTLBO). It means that the Pareto-opti-
mal set obtained by MTLBO can dominates more members of the
6.1.2. Multi-objective OPF
original TLBO solution sets. In a similar way, the lower value of
The MO-OPF problem is solved by the MTLBO as described in
C(TLBO, MTLBO) shows that fewer members in the Pareto-optimal
Section 5. Fig. 7 shows the Pareto-optimal front of the objective
set achieved by the MTLBO can be dominated by the members of
functions for the TLBO and MTLBO techniques. To verify whether
solution sets obtained by the TLBO method.
the achieved result is better than the results of other algorithm
Table 9 summarizes the best, average, and the worst D-Metric
or not, two criterions is used for comparison.
values of the mentioned algorithms in 30 trails. It is clear that
the MTLBO can reach to a lower value of D-Metric compared with
6.1.2.2. C-Metric. This parameter indicates how much the outcome
the original method. This shows that the Pareto-optimal set ob-
of an algorithm dominates the outcome of another algorithm. Sup-
tained by this algorithm has a better distribution in the solution
pose that A1 and A2 are the Pareto-optimal sets obtained by two
space.
different algorithms. The C-Metric between these two sets is
dened as below [33]:
Table 9
D-Metric values in 30 trails.

TLBO MTLBO
Best 0.591 0.402
Average 0.675 0.416
Worst 0.697 0.459

Table 10
Emission coefcients of 57-Bus IEEE system.

G1 G2 G3 G6 G8 G9 G12
a 0.04 0.03 0.04 0.035 0.05 0.045 0.06
b 0.05 0.06 0.05 0.03 0.05 0.04 0.05
c 0.06 0.05 0.04 0.035 0.045 0.05 0.05
n 0.00002 0.00005 0.00001 0.00002 0.00004 0.00001 0.00001
k 0.5 1.5 1 0.5 2 2 1.5
Fig. 7. Two-dimensional Pareto-optimal front.
A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607 605

Table 11
Best fuel cost of generators obtained by various optimization algorithms.

PG1 PG2 PG3 PG6 PG8 PG9 PG12 Total Generation Total Cost
GSA [35] 142.369 92.630 45.318 72.355 464.743 84.999 363.951 1266.365 41695.8717
EADDE [36] 143.15 95.29 45.32 73.60 464.85 83.44 361.24 1266.89 41713.62
HAPSO [37] 143.21 87.84 44.97 72.88 461.13 96.81 359.97 1266.81 41713.8868
TLBO 141.4505 85.6498 45.4567 75.0154 463.8289 93.0827 364.6708 1269.1548 41688.7431
MTLBO 142.6409 88.7977 45.0144 72.4740 459.9777 96.1366 359.9570 1264.9983 41638.3822

Table 12
Best fuel cost of generators obtained by TLBO and MTLBO algorithms.

PG1 PG2 PG3 PG6 PG8 PG9 PG12 Total Emission


TLBO 235.8670 99.9952 139.9940 99.9954 292.5465 99.9983 297.9174 1.0781
MTLBO 236.0420 100 140 100 292.1019 100 297.7938 1.0772

41720
Best
41700
Average
41680 Worst

41660

41640

41620

41600
TLBO MTLBO

Fig. 8. Best, average and worst result of 30 trails for the total fuel cost of generators.

Fig. 10. Convergence graphs of different TLBO methods.


Best
1.09 Average
the MTLBO algorithm and the maximum number of iterations are
Worst
set to 100 and 200, respectively. The lower and upper voltage mag-
nitude limits of buses and tap transformers are set between
1.08 [0.95, 1.05] and [0.9, 1.1] per-unit, respectively.

1.07 6.2.1. Single-objective OPF


TLBO MTLBO The proposed technique is used for this case study and results
Fig. 9. Best, average and worst result of 30 trails for the total emission of
are shown in Table 11 in comparison with other methods found
generators. in literatures include gravitational search algorithm (GSA) [35],
evolving ant direction differential evolution (EADDE) [36], and
hierarchical adaptive particle swarm optimization (HA-PSO) [37].
Table 13 It is clear that the proposed method reaches to a better optimal
The mean and the standard deviation of TLBO and MTLBO for 30 trails. solution. Table 12 shows the best emissions of units by the TLBO
Total Total and MTLBO algorithms. It shows that the modied algorithm can
Fuel Cost Emission reach to a better solution.
Mean SD Mean SD The procedure of solving the OPF is done for 30 trials to verify
TLBO 41701.65 12.52 1.0802 0.018 the effectiveness and robustness of the proposed method. The best,
MTLBO 41651.24 9.87 1.0781 0.011 the average and the worst result for the total fuel cost and the
emission objective functions are shown in Figs. 8 and 9, respec-
tively. Table 13 shows the mean value and the standard deviation
(SD) of these trials for the MTLBO technique in comparison with
the unmodied one. The convergence property of the original TLBO
6.2. The standard IEEE 57-Bus system method along with the proposed MTLBO algorithm is shown in
Fig. 10 and the superiority of the MTLBO is obvious. The obtained
The proposed algorithm is also applied to the IEEE 57-Bus sys- results using the TLBO and MTLBO methods are shown in Table 14.
tem to verify its performance. The parameters of this system are gi- It is clear that a multi-objective procedure should be done to nd
ven in [35]. Table 10 shows the emission coefcients of the the optimal solution whenever more than one objective function
generators that are used in this paper. The initial population of should be optimized simultaneously.
606 A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607

Table 14 Table 16
Control variables related to multi-objective problem for the TLBO and MTLBO D-Metric values in 30 trails.
methods.
TLBO MTLBO
Total fuel cost Total emission
Best 0.691 0.512
TLBO MTLBO TLBO MTLBO Average 0.701 0.546
Worst 0.762 0.603
PG1 141.4505 142.6409 235.867 236.042
PG2 85.6498 88.7977 99.9952 100
PG3 45.4567 45.0144 139.994 140
PG6 75.0154 72.4740 99.9954 100
PG8 463.8289 459.9777 292.5465 292.1019
6.2.2. Multi-objective OPF
PG9 93.0827 96.1366 99.9983 100
PG12 364.6708 359.9570 297.9174 297.7938 The Pareto-optimal front of the objective functions for the TLBO
VG1 1.0342 1.058 1.0592 1.06 and MTLBO methods are shown in Fig. 11. The mentioned criteri-
VG2 1.0356 1.06 1.06 1.06 ons for the MO-OPF are calculated and shown in Table 15 and
VG3 1.0299 1.0515 1.0538 1.06
Table 16. It is veried that the proposed MTLBO method has higher
VG6 1.0463 1.0599 1.0444 1.0572
VG8 1.06 1.06 1.047 1.06 C-Metric value and lower D-Metric value compared to the TLBO
VG9 1.0457 1.0573 1.0378 1.0505 algorithm that shows the proposed one has better performance
VG12 1.0251 1.0462 1.0291 1.0408 over the original technique.
T418 0.9556 0.9 0.9917 0.9
T418 1.0461 0.9 0.9023 0.9
T2021 1.0438 0.9828 1.0216 0.9823
T2425 0.9615 0.9001 1.0771 0.9 7. Conclusion
T2425 1.0081 0.9 1.0773 0.9
T2426 1.0332 1.0015 0.9822 0.9859
In this paper, a modied TLBO algorithm is analyzed to solve the
T729 0.9928 0.9 0.9011 0.9
T3234 0.9835 0.9 1.089 0.9 MO-OPF problem. As the TLBO is an algorithm-parameter-free, the
T1141 0.936 0.9 0.9002 0.9 performance of this technique is not affected by the algorithm
T1545 0.9416 0.9 0.915 0.9 parameters. A modied phase is added to the algorithm that in-
T1446 0.9338 0.9 0.9177 0.9 cludes a self-adaptive wavelet mutation strategy. Moreover, the
T1051 0.957 0.9057 0.9236 0.9106
proposed algorithm includes a fuzzy clustering technique to avoid
T1349 0.9079 0.9 0.9007 0.9
T1143 0.9191 0.9 0.9063 0.9 increasing repository size and a smart population approach is
T4056 1.0113 1.0118 1.0279 1.0115 utilized to efciently select the population required for the next
T3957 1.0344 0.9861 0.9709 0.9836 iteration of the algorithm. All of these modications bring out a
T955 1.0318 0.9076 0.9027 0.9
high convergence speed with great accuracy. The standard IEEE
QC18 0.1534 0.1524 0.1624 0.1569
QC25 0.1461 0.1505 0.1598 0.1681
30-Bus and 57-Bus systems are analyzed by the proposed method.
QC53 0.1572 0.1621 0.1432 0.1652 Two different case studies are used in this paper. At rst, the sin-
Cost 41688.7431 41638.3822 45606.52 45597.92
gle-objective OPF problem is solved by the suggested algorithm
Emission 1.9716 1.9152 1.0781 1.0772 to verify the performance of this method in comparison with other
optimization algorithms proposed in literatures. Accuracy, robust-
ness, and effect of population size on the proposed algorithm are
veried. In the next case study, the MTLBO is successfully used
to solve the MO-OPF. Two different criterions are used to compare
the optimal solutions obtained by different techniques. Results
verify that the implemented self-adaptive strategy has superior
performance and effectiveness over the original method.

References

[1] Huneault M, Galiana F. A survey of the optimal power ow literature. Power


Syst IEEE Trans 1991;6:76270.
[2] Zhao B, Guo C, Cao Y. Improved particle swam optimization algorithm for OPF
problems. Power Syst Conference and Exposition. 2004 IEEE PES. IEEE2004. p.
2338.
[3] Mukherjee SK, Recio A, Douligeris C. Optimal power ow by linear
programming based optimization. Southeastcon92, Proceedings IEEE.
IEEE1992. p. 5279.
[4] Olofsson M, Andersson G, Soder L. Linear programming based optimal power
ow using second order sensitivities. Power Syst, IEEE Trans 1995;10:16917.
[5] Habibollahzadeh H, Luo G-X, Semlyen A. Hydrothermal optimal power ow
based on a combined linear and nonlinear programming methodology. Power
Fig. 11. Two-dimensional Pareto-optimal front.
Syst IEEE Trans 1989;4:5307.
[6] Zehar K, Sayah S. Optimal power ow with environmental constraint using a
fast successive linear programming algorithm: Application to the Algerian
power system. Energy Convers Manage 2008;49:33626.
[7] Burchett R, Happ H, Vierath D. Quadratically convergent optimal power ow.
Power Apparatus Syst IEEE Trans 1984;5:326775.
Table 15 [8] Wei H, Sasaki H, Kubokawa J, Yokoyama R. An interior point nonlinear
C-Metric values in 30 trials. programming for optimal power ow problems with a novel data structure.
Power Syst IEEE Trans 1998;13:8707.
C(TLBO, MTLBO) C(MTLBO, TLBO)
[9] Momoh JA, Zhu J. Improved interior point method for OPF problems. Power
Best 0.008 1 Syst IEEE Trans 1999;14:111420.
Average 0.0255 0.9325 [10] Rao R, Savsani V, Vakharia D. Teachinglearning-based optimization: a novel
Worst 0 0.8992 method for constrained mechanical design optimization problems. Comput
Aided Des 2011;43:30315.
A. Shabanpour-Haghighi et al. / Energy Conversion and Management 77 (2014) 597607 607

[11] Rao R, Savsani V, Vakharia D. Teachinglearning-based optimization: an [24] Abido M. Optimal power ow using tabu search algorithm. Electr Power
optimization method for continuous non-linear large scale problems. Inform Compon Syst 2002;30:46983.
Sci 2012;183:115. [25] Yuryevich J, Wong KP. Evolutionary programming based optimal power ow
[12] Niknam T, Golestaneh F, Sadeghi MS. H-Multiobjective teachinglearning- algorithm. Power Syst IEEE Trans 1999;14:124550.
based optimization for dynamic economic emission dispatch. Syst J IEEE [26] Costa AL, Costa AS. Energy and ancillary service dispatch through dynamic
2012;6:34152. optimal power ow. Electric Power Syst Res 2007;77:104755.
[13] AlRashidi M, El-Hawary M. Hybrid particle swarm optimization approach for [27] Sayah S, Zehar K. Modied differential evolution algorithm for optimal power
solving the discrete OPF problem considering the valve loading effects. Power ow with non-smooth cost functions. Energy Convers Manage
Syst IEEE Trans 2007;22:20308. 2008;49:303642.
[14] Basu M. Multi-objective optimal power ow with FACTS devices. Energy [28] Bouktir T, Slimani L, Mahdad B. Optimal power dispatch for large scale power
Convers Manage 2011;52:90310. system using stochastic search algorithms. Int J Power Energy Syst
[15] Niknam T, Narimani M, Aghaei J, Azizipanah-Abarghooee R. Improved particle 2008;28:118.
swarm optimization for multi-objective optimal power ow considering the [29] Bakirtzis AG, Biskas PN, Zoumas CE, Petridis V. Optimal power ow by
cost, loss, emission and voltage stability index. IET Gener Transm Distrib enhanced genetic algorithm. Power Syst IEEE Trans 2002;17:22936.
2012;6:51527. [30] Slimani L, Bouktir T. Economic power dispatch of power system with pollution
[16] Niknam T, Narimani M, Aghaei J, Tabatabaei S, Nayeripour M. Modied Honey control using multiobjective ant colony optimization. Int J Comput Intelligence
Bee Mating Optimization to solve dynamic optimal power ow considering Res (IJCIR). 2007;3:14553.
generator constraints. IET Gener Transm Distrib 2011;5:9891002. [31] Saini A, Chaturvedi DK, Saxena A. Optimal power ow solution: a GA-Fuzzy
[17] Niknam T, Narimani MR, Azizipanah-Abarghooee R. A new hybrid algorithm system approach. Int J Emer Elec Power Syst 2006;5:5.
for optimal power ow considering prohibited zones and valve point effect. [32] Niknam T, Azizipanah-Abarghooee R, Zare M, Bahmani-Firouzi B. Reserve
Energy Convers Manage 2012;58:197206. constrained dynamic environmental/economic dispatch: a new multiobjective
[18] Ben Attous D, Labbi Y. Particle swarm optimization based optimal power ow self-adaptive learning bat algorithm. Syst J IEEE 2013.
for units with non-smooth fuel cost functions. Electrical and Electronics [33] Agrawal S, Dashora Y, Tiwari MK, Son Y-J. Interactive particle swarm: a pareto-
Engineering. ELECO 2009 International Conference on. IEEE 2009. p. I-377I- adaptive metaheuristic to multiobjective optimization systems, man and
81. cybernetics. Part A: systems and humans. IEEE Trans 2008;38:25877.
[19] Niknam T, Azizipanah-Abarghooee R, Aghaei J. A new modied teaching [34] Wu L, Wang Y, Yuan X, Zhou S. Environmental/economic power dispatch
learning algorithm for reserve constrained dynamic economic dispatch. Power problem using multi-objective differential evolution algorithm. Electric Power
Syst IEEE Trans 2012. Syst Res 2010;80:117181.
[20] Niknam T, Jabbari M, Malekpour AR. A modied shufe frog leaping algorithm [35] Duman S, Gven U, Snmez Y, Yrkeren N. Optimal power ow using
for multi-objective optimal power ow. Energy 2011;36:642032. gravitational search algorithm. Energy Convers Manage 2012;59:8695.
[21] Agrawal S, Panigrahi B, Tiwari MK. Multiobjective particle swarm algorithm [36] Vaisakh K, Srinivas L. Evolving ant direction differential evolution for OPF with
with fuzzy clustering for electrical power dispatch. Evol Comput IEEE Trans non-smooth cost functions. Eng Appl Articial Intelligence 2011;24:42636.
2008;12:52941. [37] Mahdad B, Srairi K. Hierarchical adaptive PSO for multi-objective OPF
[22] Hazra J, Sinha A. A multi-objective optimal power ow using particle swarm considering emissions based shunt FACTS. IECON 2012-38th Annual
optimization. Eur Trans Electr Power 2011;21:102845. Conference on IEEE Industrial Electronics Society. IEEE2012. p. 133743.
[23] Alsac O, Stott B. Optimal load ow with steady-state security. Power
Apparatus Syst IEEE Trans 1974:74551.

Vous aimerez peut-être aussi