Applied Soft Computing 12 (2012) 1180–1186

Contents lists available at SciVerse ScienceDirect

Applied Soft Computing

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

Firefly Algorithm for solving non-convex economic dispatch problems with valve
loading effect
Xin-She Yang a , Seyyed Soheil Sadat Hosseini b , Amir Hossein Gandomi c,∗
a
Department of Engineering, University of Cambridge, Trumpington Street, Cambridge CB2 1PZ, UK
b
Young Researchers Club, Sari Branch, Islamic Azad University, Sari, Iran
c
Young Researchers Club, Central Tehran Branch, Islamic Azad University, Tehran, Iran

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

Article history: The growing costs of fuel and operation of power generating units warrant improvement of optimiza-
Received 5 August 2010 tion methodologies for economic dispatch (ED) problems. The practical ED problems have non-convex
Received in revised form 21 June 2011 objective functions with equality and inequality constraints that make it much harder to find the global
Accepted 22 September 2011
optimum using any mathematical algorithms. Modern optimization algorithms are often meta-heuristic,
Available online 17 November 2011
and they are very promising in solving nonlinear programming problems. This paper presents a novel
approach to determining the feasible optimal solution of the ED problems using the recently devel-
Keywords:
oped Firefly Algorithm (FA). Many nonlinear characteristics of power generators, and their operational
Economic load dispatch
Valve loading effect
constraints, such as generation limitations, prohibited operating zones, ramp rate limits, transmission
Firefly Algorithm loss, and nonlinear cost functions, were all contemplated for practical operation. To demonstrate the
Metaheuristic efficiency and applicability of the proposed method, we study four ED test systems having non-convex
solution spaces and compared with some of the most recently published ED solution methods. The results
of this study show that the proposed FA is able to find more economical loads than those determined by
other methods. This algorithm is considered to be a promising alternative algorithm for solving the ED
problems in practical power systems.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction observed in real input–output characteristics, owing to valve-point

loading in fossil fuel burning plants [5]. Besides, due to physical
Economic dispatch (ED) has become a fundamental function operation limitations (such as faults in the machines themselves)
in operation and control of modern power systems. The ED prob- or the associated auxiliaries (such as boiler, feed pumps), units can
lem can be stated as determining the least cost power generation have prohibited operating regions and generators that operate in
schedule from a set of online generating units to satisfy the load these zones may experience amplification of vibrations in their
demand at a given point of time [1]. Though the core objective of shaft bearing, which should be prevented in practical applications
the problem is to minimize the operating cost satisfying the load [6]. Also due to the change restriction in the unit generation out-
demand, several types of physical and operational constraints make put, the units in the actual operation can have ramp rate limits [7].
ED highly nonlinear constrained optimization problem, especially So, ramp rate limits, prohibited operating zones (POZs) and valve
for larger systems [2]. However, accurate and intelligent scheduling loading effects should be considered to solve a realistic ED problem,
of the units not only can decrease the operating cost significantly which makes the finding of the optimum solution extremely hard.
but also can assure higher reliability with improved security and Several deterministic optimization techniques were proposed
less environmental impact [3]. In traditional ED approaches, the to solve the ED problem, including gradient method [8], lambda
input–output characteristics (or cost function) of a generator is iteration method [9], linear programming [10], quadratic program-
approximately shown by utilizing a single quadratic function. In ming [11], non-linear programming [12], Lagrangian relaxation
practice, operating conditions of many generating units require algorithm [13] and dynamic programming [14]. By modeling of
the cost function to be modeled as a piecewise quadratic function the final cost of generation accurately and taking valve-loading
[4]. However, higher-order nonlinearities and discontinuities are effect into account, the cost function of generators takes a non-
convex form [15]. The theoretical assumptions behind previously
mentioned algorithms (except dynamic programming) make them
∗ Corresponding author. Tel.: +1 234 788 0619. unsuitable for the ED formulation regarding non-convexity and dif-
E-mail addresses: xy227@cam.ac.uk (X.-S. Yang), s.sadathosseini@gmail.com ferentiability. Furthermore, they are local optimizers by nature, i.e.,
(S.S. Sadat Hosseini), a.h.gandomi@gmail.com, ag72@uakron.edu (A.H. Gandomi). they might converge to local solutions instead of global ones if the

1568-4946/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2011.09.017
 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186 1181

initial guess happens to be in the neighborhood of a local solution. This curve contains higher order nonlinearity because of the valve-
Dynamic programming method may cause the dimensions of the point effect, and should be refined by a sine function. Also the
ED problem to become extremely large, thus requiring enormous solution procedure can easily trap in the local minima in the vicinity
computational efforts. of optimal value. To take account for the valve-point effects, sinu-
To overcome these deficiencies, artificial intelligence meth- soidal terms are added to the quadratic cost functions as follows:
ods have been used to solve the ED problem, and these methods ∼  
include Genetic Algorithm (GA) [16], real-coded genetic algo- Fi = ai Pi2 + bi Pi + ci + ei sin(fi (Pimin − Pi ) (3)
rithm (RCGA)[17], Tabu Search (TS) [5], Hopfield neural network
[18], different types of Evolutionary Programming (EP) [19], where ei and fi are constants of the unit with valve-point effects.
biogeography-based optimization (BBO) [20], Evolutionary Strat- The model in Eq. (3) is subject to the following constraints:
egy (ES) [21], Particle Swarm Optimization (PSO) ([22–25]), an
improved coordinated aggregation-based particle swarm opti- 2.1. Power balance constraint
mization (ICA-PSO) [26,27], Bacterial Foraging (BF) [28], harmony
search (HS) [29] and Hopfield Neural Network (HNN) [30]. NG

Although several optimization methodologies have been devel- Pi = PD + PL (4)
oped for the ED problem, the complexity of the task reveals the i=1
necessity for development of efficient algorithms to accurately
locate the optimum solution. In this context, the objective of this where PD is the load demand and PL is the total transmission net-
work is to demonstrate a new approach for solving ED problems, work losses of system. To compute network losses, the B-coefficient
aiming to provide a practical alternative for conventional methods. method [7] is commonly utilized by the power utility industry. In
Here, Firefly Algorithm, developed by Yang [31], is used which has the B-coefficient method, the transmission losses are expressed as
been successful to solve mixed variable and constrained engineer- a quadratic function:
ing problems [32]. A more detailed description concerning theoreti- n
n  n
 
cal and implementation feature of the proposed method is provided PL = Pi Bij Pj + B0i Pi + B00 . (5)
in the later sections. To show the efficiency and applicability of the
i=1 j=1 i=1
proposed method, several types of ED problems are analyzed and
results are compared with those available in the literature.
This paper is organized as follows. Section 2 illustrates the ED 2.2. Ramp rate limits
problem formulation considering valve-loading effect, prohibited
operating zone (POZ) constraints and ramp rate limits. Moreover, One of the unrealistic assumptions that prevailed for simplifying
the proposed method for constraints handling is demonstrated in the problem in many of the earlier research is that the adaptations
this section. In Section 3, the Firefly Algorithm is described. In to the power output are instantaneous. However, under practical
Section 4, simulation results are presented that demonstrate the circumstances, ramp rate limit restrains the operating range of all
potential of the proposed algorithm. Finally, Section 5 concludes the online units for tuning the generator operation between two
the paper with discussions operating periods [33]. The generation may increase or decrease
with corresponding upper and lower ramp rate limits. Therefore,
2. Economic load dispatch problems with valve-point units are restricted due to these ramp rate limits as mentioned
loadings problem below.
If power generation increases, we have
The goal of economic dispatch (ED) problem is to find the opti- Pi − Pi0 ≤ URi . (6)
mal combination of power generations that minimizes the total
generation cost, while satisfying an equality constraint and inequal- If power generation decreases, we have
ity constraints. Cost efficiency is the most significant subproblem
of power system operations. Owing to the highly nonlinearity char- Pi0 − Pi ≤ DRi (7)
acteristics of power systems and generators, ED belongs to a class
of nonlinear programming optimization including equality and where Pi0 is the previous power generation of unit i. URi and DRi are
inequality constraints. Practically speaking, while the scheduled the up-ramp and down-ramp limits of the ith generator, respec-
combined units for each specific period of operation are listed from tively. The inclusion of ramp rate limits changes the generator
unit commitment, the ED planning must carry out the optimal gen- operation constraints (5) as follows:
eration dispatch among the operating units in order to meet the
max(Pimax , URi − Pi) ≤ Pi ≤ min(Pimax , Pi0 − DRi ). (8)
load demands and practical operation constraints of generators,
which consist of ramp rate limits, maximum and minimum lim-
its, and prohibited operating zones. In general, the generation cost 2.3. Prohibited operating zones
function is usually stated as a quadratic polynomial. Mathemati-
cally, the problem can be modeled as: A generator with prohibited operating zones has discontinuous
fuel-cost characteristics. The conception of prohibited operating
n
 zones is consisted of the following constraint in the ED:
min f = Fi (Pi ) (1)
⎧ min LB
i=1 P ≤ Pi ≤ Pi,1
⎨ i
⎪
where Fi is the total generation cost for the generator unit i, which UB ≤ P ≤ P LB j = 2, 3, . . . , NP
Pi,j−1 i i . (9)
i,j
is defined by the following equation: ⎪
⎩ UB
Pi,j ≤ Pi ≤ Pimax j = NPi
Fi (Pi ) = ai Pi2 + bi Pi + ci (2)
LB and P UB are the lower and upper boundaries of prohibited
where Pi,j
where ai , bi and ci are coefficients of generator i. i,j
The valve-opening process of multivalve steam turbines pro- operating zone j of generator i, respectively; NPi is the number of
duces a ripple-like effect in the heat rate curve of the generators. prohibited operating zones of generator i.
1182 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186

3. Firefly Algorithm problem. The POZ constraints (9) are utilized as follows. If the gen-
eration of unit i is settled in its jth POZ, i.e.:
The Firefly Algorithm was developed by Yang ([34,35]), and it
LB UB
was based on the following idealized behavior of the flashing char- Pi,j ≺ Pi ≺ Pi,j (13)
acteristics of fireflies:
then the amount of generation is cut to the nearest boundary of the
• All fireflies are unisex so that one firefly is attracted to other jth POZ as follows:
fireflies regardless of their sex;
LB + P UB
Pi,j
• attractiveness is proportional to their brightness, thus for any ave i,j
Pi,j = (14)
two flashing fireflies, the less bright one will move towards the 2
brighter one. The attractiveness is proportional to the brightness
and they both decrease as their distance increases. If no one is LB LB ≺ P ≤ P ave
Pi,j if Pi,j i i,j
brighter than a particular firefly, it moves randomly; Pi = . (15)
• the brightness or light intensity of a firefly is affected or deter- UB
Pi,j ave ≺ P ≺ P UB
if Pi,j i i,j
mined by the landscape of the objective function to be optimized.
For a nonlinear optimization problem with equality and inequal-
For a maximization problem, the brightness can simply be ity constraints, a common method is the penalty method. The idea
proportional to the objective function. Other forms of brightness is to define a penalty function so that the constrained problem is
can be defined in a similar way to the fitness function in genetic transformed into an unconstrained problem. Now we can define
algorithms or the bacterial foraging algorithm (BFA) ([36]).
M
 N

(x, i , vj ) = f (x) + i ϕi2 (x) + vj 2
j
(x) (16)
Firefly Algorithm
i=1 j=1
Objective function f(x), x = (x1 , . . ., xd )T
Initialize a population of fireflies xi (i = 1, 2, . . ., n) where i ≥ 1 and vj ≥ 0 which should be large enough, depending
Define light absorption coefficient on the solution quality needed. As we can see, when an equality
while (t < MaxGeneration) constraint it met, its effect or contribution to is zero. How-
for i = 1: n all n fireflies ever, when it is violated, it is penalized heavily as it increases
for j = 1: i all n fireflies significantly. Similarly, it is true when inequality constraints
Light intensity Ii at xi is determined by f(xi ) becomes tight or exactly. It should be mentioned that generation
if (Ij > Ii ) and ramp rate limits are similar type of constraints. These con-
Move firefly i towards j in all d dimensions straints together state the overall upper/lower generation limits of
end if the units.
Attractiveness varies with distance r via exp [− r2 ]
Evaluate new solutions and update light intensity
end for j 4. Implementation and numerical experiments
end for i
Rank the fireflies and find the current best All the EAs for the ED problems are implemented using
end while MATLABTM 7.0 on a PC with a Pentium IV, Intel Dual core 2.2 GHz,
Postprocess results and visualization 1 GB RAM. Owing to the random nature of the FA (and in fact all
metaheuristic algorithms), their performance cannot be judged by
the result of a single run. Many trials with independent popula-
The movement of a firefly i is attracted to another more attrac- tion initializations should be made to obtain a useful conclusion of
tive (brighter) firefly j is determined by the performance of the approach. Therefore, the results should be
− r2 analyzed using statistic measures such as mean and standard devi-
xit+1 = xit + ˇ0 e ij (xit − xit ) + ˛εti (12) ation. The best, worst and mean obtained in 100 trials are used to
where ˇ0 is the attractiveness at r = 0, the second term is due to the compare the performances of different EAs. To find the effective-
attraction, while the third term is randomization with the vector ness of the proposed FA, the test results are also compared with the
of random variables εi being drawn from a Gaussian distribution. results already reported by the most recently published methods
The distance between any for solving the ED problem.
 two  fireflies i and j at xi and xj can be the
There are 4 important parameters in the Firefly Algorithm: ˛0 ,
Cartesian distance rij = xi − xj  or the l2 -norm. For other applica-
2 ˇ0 , , and the population size n. In order to obtain the right param-
tions such as scheduling, the distance can be time delay or any suit-
eters, we have carried out a detailed parametric study by varying
able forms, not necessarily the Cartesian distance. For most cases
these parameters. More specifically, we varied ˛0, ˇ0 from 0.1 to 1.0
in our implementation, we can take ˇ0 = 1, ˛ ∈ [0, 1] , and = 1.
with a step increase of 0.1, from 0.01 to 100 with a step increase
In addition, if the scales vary significantly in different dimensions
of 0.01 up to 1, and then 5 up to 100. We also varied n from 5 to
such as −105 to 105 in one dimension while, say, −10−3 to 103 along
100 with an interval of 5. By analyzing the optimal solutions for
others, it is a good idea to replace ␣ by ˛Sk where the scaling param-
a wide range of test functions, we found that the best values or
eters Sk (k = 1, . . ., d) in the d dimensions should be determined by
ranges of the parameters are: ˛0 = 0.4–0.9, ˇ0 = 0.5–1.0, = 0.1–10
the actual scales of the problem of interest. In essence, the param-
and n = 25–50. For most problems, we can use the fixed value of
eter characterizes the variation of the attractiveness, and partly
ˇ0 = 1. Therefore, the parameters of the proposed FA to solve the
controls how the algorithm behaves. It is also possible to adjust
ED problem in the four test cases are: ˇ0 = 1, = 1/L (where L is
so that multiple optima can be found at the same during iterations.
the characteristic scale of the problem), ˛ = 0.5 and then is gradu-
3.1. Constraint handling ally reduced to 0.01 as iterations proceed. n varies from 25 to 50,
depending on the complexity of the problem. In addition, the max-
A significant factor in the application of optimization tech- imum number of function evaluations varies from 5000 to 50,000,
niques is how the algorithm handles the constraints concerning the depending on the size of the ED problem.
 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186 1183

Table 1
Comparison of FA and global optimal on the three benchmark constrained problems.

No. Decision var. no. Constraint no. Global optimal Predicted by FA

Continues Discrete [Xmin ; Fmin ] [Xmin ; Fmin ]

1 1 1 1 [1.375, 1; 2.124] [1.375, 1; 2.124]

2 2 1 2 [0.94194, 2.1, 1; 1.07654] [0.94194, 2.1, 1; 1.07654]
3 3 4 8 [0.2, 1.280624, 1.954483, 1, 0, 0, 1; 3.557463] [0.2, 1.280624, 1.954483, 1, 0, 0, 1; 3.557463]

Table 2
The best, average and worst results of different ED solution methods for the 3 unit test system.

Methods Generation cost ($/h)

Best Average Worst Standard deviation No. of evaluation

GAB [19] 8234.08 NA NA NA 10,000

GAF [19] 8234.07 NA NA NA 10,000
CEP [19] 8234.07 8235.97 8241.83 NA 1000
FEP [19] 8234.07 8234.24 8241.78 NA 1000
MFEB [19] 8234.08 8234.71 8241.8 NA 1000
IFEP [19] 8234.07 8234.16 8234.54 NA 1000
FA 8234.07 8234.08 8241.23 3.63 5000

NA: not available.

4.1. Validation Table 3

Output power of generators in the best result of the proposed FA for the 3 unit test
system.
Before solving economic dispatch problems, FA was bench-
marked using three mixed-variable numerical examples which are Unit Power (MW)
given as follows in detail. The example 1 is chosen from Floudas [37] 1 300.267
and the examples 2 and 3 are selected from Costa and Oliveria [38]. 2 149.733
The results obtained by FA and the global optimal are presented in 3 400.000

Table 1. As it can be seen from this table the FA obtained the global Total generation (MW) 850
optimal solution correctly in all the three examples. Generation cost ($/h) 8234.074

Example 1.
4.2. Economic dispatch problems
x
Minimize : f (x, y) = −y + 2x − ln
2
4.2.1. Case 1: 3 generating units
x
This test case includes three generating units. The expected load
Subject to : g1 = −x − ln +y≤0
2 demand to be met by all the three generating units is 850 MW.
  The description of the system can be found from [19]. It has
where 0.5 ≤ x ≤ 1.5 and y ∈ 0, 1 .
been reported in Ref. [5] that the global minimum found for the
Example 2. three-generator system is 8234.07. Based on the above mentioned
parameters, the Firefly Algorithm has been executed for 100 trials
Minimize : f (x, y) = −0.7y + 5(x1 − 0.5)2 + 0.8 with various starting points to verify its performance and efficiency.
The best, average and worst of cost functions achieved by various
g1 = − exp(x1 − 0.2) − x2 ≤ 0 methods are shown in Table 2. All methods give a similar ‘best solu-
Subject to : g2 = x2 + 1.1y + 1.0 ≤ 0 tion’, whereas ‘average’ and ‘worst’ costs differ. Table 3 shows that
g3 = x1 − 1.2y − 0.2 ≤ 0 the proposed method has accomplished in finding the global opti-
  mal solution presented in Ref. [5]. The average execution time of
where 0.2 ≤ x1 ≤ 1.0, −2.22554 ≤ x2 ≤ − 1.0 and y ∈ 0, 1 .
the FA for this test system is 2.07 s.
Example 3.

Minimize : f (x, y) = (y1 − 1)2 + (y2 − 1)2 + (y3 − 1)2 4.2.2. Case 2: 13 generating units
This test case consists of thirteen generating units; the complex-
− ln(y4 + 1) + (x1 − 1)2 + (x2 − 2)2 + (x3 − 3)2
ity to the solution process has significantly increased. Inasmuch as
this is a larger system with higher non-linearity, it has more local
g1 = y1 + y2 + y3 + x1 + x2 + x3 − 5.0 ≤ 0 minima and thus it is difficult to attain the global solution. To be
g2 = y32 + x12 + x22 + x32 − 5.5 ≤ 0 able to deal with more complicated case with highly non-linearity
g3 = y1 + x1 − 1.2 ≤ 0 is one of the main goals of FA applications. The load demand of this
g4 = y2 + x2 − 1.8 ≤ 0 test system is 1800 MW. Exactly the same data of all units as given
g5 = y3 + x3 − 2.5 ≤ 0 in Ref. [19] will be utilized in this case. Table 4 shows the best, aver-
age and worst results of different ED solution methods among 100
Subject to : g6 = y4 + x1 − 1.2 ≤ 0
trial runs in the same way as listed in Table 2. The outcomes of the
g7 = y22 + x22 − 1.64 ≤ 0
other approaches shown in Table 4 have been directly quoted from
g8 = y32 + x32 − 4.25 ≤ 0 their corresponding references (‘NA’ means the related result is not
g9 = y22 + x32 − 4.64 ≤ 0 available in the corresponding reference). The average execution
  time of the FA for this test system is 10.36956 s. The computa-
where 0 ≤ x1 , x2 , x3 and y1 , y2 , y3 , y4 ∈ 0, 1 .
tion times of the FA are not compared with the other ED solution
1184 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186

Table 4
The best, average and worst results of different ED solution methods for the 13 unit test system.

Methods Generation cost ($/h)

Best Average Worst Standard deviation No. of evaluation

CEP [19] 18,048.21 18,190.32 18,404.04 NA NA

PSO [2] 18,030.72 18,205.78 NA NA 10,000
MFEP [19] 18,028.09 18,192 18,416.89 NA NA
FEP [19] 18,018 18,200.79 18,453.82 NA NA
IFEP [19] 17,994.07 18,127.06 18,267.42 NA NA
EP–SQP [2] 17,991.03 18,106.93 NA NA 10,000
HDE [39] 17,975.73 18,134.8 NA NA 100
CGA MU [40] 17,975.34 NA NA NA NA
PSO–SQP [2] 17,969.93 18,029.99 NA NA 10,000
HS [29] 17,965.62 17,986.563 18,070.176 26.3702 22,500
IGA MU [40] 17,963.98 NA NA NA NA
DEC(1)-SQP(1) [41] 17,963.94 17,973.13 17,984.81 NA NA
ST-HDE [39] 17,963.89 18,046.38 NA NA 100
FA 17,963.83 18,029.16 18,168.80 148.542 25,000

NA: not available.

methods, since the computation times of each ED method are com- results of the optimal solution of the proposed method, includ-
puted on a different hardware. Output power of the generators of ing generation output of each unit for this test system, are shown
the 13 unit test system in the minimum solution of the FA is shown in Table 7.
in Table 5.

4.2.3. Case 3: 40 generating units 4.2.4. Case 4: 15 generating units

The test system has forty generating units with non-convex fuel In this case study, all mentioned practical constraints and non-
cost function incorporating valve loading effects. The required load linear characteristics of the ED problem are included. The valve
demand to be met by all the forty generating units is 10,500 MW. loading effects, ramp rate limits and POZs are considered for the
The detailed information of generating units of test system is units of this test system, whose data is given in Ref. [22]. The
given in Ref. [19]. This case study has a larger and more complex prohibited operating zones embedded in the 4 units, units 2, 5,
solution space than all the previous case studies, and so any dif- 6, and 12. This problem proves particularly challenging because
ference between different ED solution techniques can be better these zones result in a non-convex decision space where 192
revealed in this test case. The Firefly Algorithm has been executed convex subspaces can be constituted for the dispatch problem.
for a hundred times with various starting points. The obtained The remaining units have simple operational zone. This challeng-
results of the proposed FA to resolve the ED problem for this test ing problem not only requires the proper implementation of the
system are shown in Table 5. In this table, the detailed compar- constraints, but also employs an efficient search in different sub-
isons of the best, average and worst solutions of the proposed regions without wasting too much time on the prohibited regions.
FA and most recently published ED solution methods are shown. This means that number of fireflies should be sufficient enough
As seen from Table 6, the best solution of the proposed method so that they can distribute in most promising regions with high
is better than those of all other methods, indicating FA’s higher probability, while they can also fly or jump to different regions
efficiency to solve the ED problem comparing with the other meth- when necessary. Therefore, a fine balance between solution quality
ods. Hence, for power system ED problems of greater size with and computational effort is required. To ensure the global opti-
higher non-linearities, the proposed method is proved to be the mum solution is reachable, we have tried to do a few test runs
best approach among all the methods. The average execution time by varying the number of fireflies and the total number of func-
of the FA for this test system is 4.72801 s, and such a computation tional evaluations. From this initial learning experience, we can
time to solve the ED problem is reasonable and practical. Detailed get a crude estimate what parameters are appropriate for this
problem. We then use these as a basis for the more intensive
100 runs.
Table 5
Output power of generators in the best result of the proposed FA for the 13 unit test The comparison of the best, average and worst solutions of the
system. proposed FA and most recently published ED solution methods, is
shown in Table 8. Again, the FA offers an improved generation cost
Unit Power (MW)
over the other methods, clearly showing the proposed approach of
1 628.31852
locating better solutions is superior to others. Detailed results of the
2 149.59952
3 222.74912
optimal solution of the proposed method, discovered through the
4 109.86655 proposed method are shown in Table 9. All mentioned constraints
5 109.86655 were satisfied. The average execution time of the FA for this test
6 109.86655 system is 16.05 s, which is again a good computation time.
7 109.86655
From our simulations, we observed that Firefly Algorithms have
8 60.00000
9 109.86655 some advantages over other algorithms such as particle swarm
10 40.00000 optimization. By comparing with PSO and other algorithms, Firefly
11 40.00000 Algorithm has two superiorities: automatic subdivision and ran-
12 55.00000
dom reduction. The fireflies in the FA can automatically divide into
13 55.00009
subgroups so that these subgroups swarm around the multimodal
Total generation (MW) 1800 optima. This makes it possible for the algorithm to find all global
Generation cost ($/h) 17963.83080
optima simultaneously. Thus, the algorithm is particularly suitable
 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186 1185

Table 6
The best, average and worst results of different ED solution methods for the 40 unit test system.

Methods Generation cost ($/h)

Best Average Worst Standard deviation No. of evaluation

HGPSO [42] 124,797.13 126,855.70 NA 1160.91 NA

SPSO [42] 124,350.40 126,074.40 NA 1153.11 NA
PSO [2] 123,930.45 124,154.49 NA NA 10,000
CEP [19] 123,488.29 124,793.48 126,902.89 NA NA
HGAPSO [42] 122,780.00 124,575.70 NA 906.04 NA
FEP [19] 122,679.71 124,119.37 127,245.59 NA NA
MFEP [19] 122,647.57 123,489.74 124,356.47 NA NA
IFEP [19] 122,624.35 123,382.00 125,740.63 NA NA
TM [43] 122,477.78 123,078.21 124,693.81 NA 4050
EP–SQP [2] 122,323.97 122,379.63 NA NA 10,000
MPSO [44] 122,252.26 NA NA NA NA
ESO [21] 122,122.16 122,558.45 123,143.07 NA 75,000
HPSOM [42] 122,112.40 124,350.87 NA 978.75 NA
PSO–SQP [2] 122,094.67 122,245.25 NA NA 10,000
PSO-LRS [23] 122,035.79 122,558.45 123,461.67 NA 20,000
Improved GA [45] 121,915.93 122,811.41 123,334.00 NA 100,000
HPSOWM [42] 121,915.30 122,844.40 NA 497.44 NA
IGAMU [16] 121,819.25 NA NA NA NA
HDE [39] 121,813.26 122,705.66 NA NA 100
DEC(2)-SQP(1) [46] 121,741.97 122,295.12 122,839.29 386.181 18,000
PSO [24] 121,735.47 122,513.91 123,467.40 NA 20,000
APSO(1) [24] 121,704.73 122,221.36 122,995.09 NA 20,000
ST-HDE [39] 121,698.51 122,304.30 NA NA 100
NPSO-LRS [23] 121,664.43 122,209.31 122,981.59 NA 20,000
APSO(2) [24] 121,663.52 122,153.67 122,912.39 NA 20,000
SOHPSO [22] 121,501.14 121,853.57 122,446.30 NA 62,500
BBO [20] 121,479.50 121,512.06 121,688.66 NA 50,000
BF [28] 121,423.63 121,814.94 NA 124.876 10,000
GA–PS–SQP [47] 121,458.00 122,039.00 NA NA 1000
PS [48] 121,415.14 122,332.65 125,486.29 NA 1000
FA 121,415.05 121,416.57 121,424.56 1.784 25,000

NA: not available.

Table 7 Table 8
Output power of generators in the best result of the proposed FA for the 40 unit test The best, average and worst result of different ED solution methods for the 15 unit
system. test system including POZ constraints, ramp rate limits and transmission losses.

Unit Power (MW) Unit Power (MW) Methods Generation cost ($/h)

1 110.8099 21 523.2793 Best Average Worst Standard No. of

2 110.8059 22 523.2793 deviation evaluation
3 97.40230 23 523.2832
4 179.7332 24 523.2832 PSO [49] 32,858 33,039 33,331 NA 20,000
5 92.70700 25 523.2793 GA [49] 33,113 33,228 33,337 NA 20,000
6 140.0000 26 523.2793 SOH PSO [22] 32,751 32,878 32,945 NA 62,500
7 259.6004 27 10.0000 CPSO1 [50] 32,835 33,021 33,318 NA 8000
8 284.6004 28 10.0000 CPSO2 [50] 32,834 33,021 33,318 NA 8000
9 284.6004 29 10.0000 BF [28] 32,784.5 32,796.8 NA 85.7743 10,000
10 130.0028 30 87.8008 FA 32,704.5 32,856.1 33,175.0 147.17022 50,000
11 168.8008 31 189.9989 NA: not available.
12 168.8008 32 189.9989
13 214.7606 33 189.9989 Table 9
14 304.5204 34 164.8036 Output power of generators and transmission losses in the best result of the pro-
15 394.2801 35 164.8036 posed FA for the 15 unit test system.
16 394.2801 36 164.8036
Unit Power (MW)
17 489.2801 37 110.0000
18 489.2801 38 110.0000 1 455.0000
19 511.2817 39 110.0000 2 380.0000
20 511.2817 40 511.2794 3 130.0000
4 130.0000
Total 10500
5 170.0000
generation
6 460.0000
(MW)
7 430.0000
Generation 121415.0522
8 71.7450
cost ($/h)
9 58.9164
10 160.0000
11 80.0000
12 80.0000
13 25.0000
14 15.0000
for nonlinear, multimodal optimization problems. In addition, the
15 15.0000
randomness is reduced gradually by using a similar strategy as that
used in simulated annealing, and this will speed up convergence Losses (MW) 30.6614
Generation cost ($/h) 32,704.4501
when the global optimality is approaching.
1186 X.-S. Yang et al. / Applied Soft Computing 12 (2012) 1180–1186

5. Conclusions [22] K.T. Chaturvedi, M. Pandit, L. Srivastava, Self-organizing hierarchical particle

swarm optimization for nonconvex economic dispatch, IEEE Trans. Power Syst.
23 (3) (2008) 1079–1087.
In this paper, we have presented a new approach to non-convex [23] A.I. Selvakumar, K. Thanushkodi, A new particle swarm optimization solution to
ED problems based on the FA. Many of the nonlinear characteris- nonconvex economic dispatch problems, IEEE Trans. Power Syst. 22 (1) (2007)
tics of power systems such as valve-point loadings, ramp rate limits, 42–51.
[24] A.I. Selvakumar, K. Thanushkodi, Anti-predatory particle swarm optimization:
and prohibited operating zones are considered for practical gener- solution to nonconvex economic dispatch problems, Electr. Power Syst. Res. 78
ator operation in the proposed method. Four test cases have been (2008) 2–10.
studied and comparisons of the quality of the solution and per- [25] K.T. Chaturvedi, M. Pandit, L. Srivastava, Particle swarm optimization with
crazy particles for nonconvex economic dispatch, Appl. Soft Comput. 9 (2009)
formance have been conducted against several of most recently
962–969.
published ED solution methods. Based on simulation results, the [26] J.G. Vlachogiannis, K.Y. Lee, Economic load dispatch—a comparative study on
solution quality and reliability show the superiority of the FA over heuristic optimization techniques with an improved coordinated aggregation-
based PSO, IEEE Trans. Power Syst. 24 (2) (2009) 991–1001.
other approaches. The proposed algorithm capability and robust-
[27] S.S. Sadat Hosseini, A.H. Gandomi, Discussion of economic load dispatch—a
ness make it suitable to solve complex optimization problems like comparative study on heuristic optimization techniques with an improved
non-convex ED. Future studies can focus on the inclusion of more coordinated aggregation-based PSO, IEEE Trans. Power Syst. 25 (1) (2010) 590.
realistic constraints to the problem structure. Large-scale, realistic [28] B.K. Panigrahi, V.R. Pandi, Bacterial foraging optimization: Nelder–Mead hybrid
algorithm for economic load dispatch, IET Gener. Transm. Distrib. 2 (4) (2008)
ED problems can be attempted by the proposed methodology. In 556–565.
addition, it would be a good research topic to extend the proposed [29] L.S. Coelho, V.C. Mariani, An improved harmony search algorithm for power
approach to solve mixed integer programming problems which are economic load dispatch, Energy Convers. Manage. 50 (2009) 2522–2526.
[30] H. Altun, T. Yalcinoz, Implementing soft computing techniques to solve eco-
often NP-hard. A detailed parametric study of the algorithm may nomic dispatch problem in power systems, Expert Syst. Appl. 35 (2008)
also prove fruitful. 1668–1678.
[31] X.S. Yang, Firefly algorithm, stochastic test functions and design optimization,
Int. J. Bio-Inspired Comput. 2 (2) (2010) 78–84.
References [32] A.H. Gandomi, X.S. Yang, A.H. Alavi, Mixed variable structural optimization
using Firefly Algorithm, Comput. Struct. 89 (23–24) (2011) 2325–2336.
[1] A.J. Wood, B.F. Wollenberg, Power Generation, Operation and Control, John [33] H. MA, A.A. El-Keib, R.E. Smith, A genetic algorithm-based approach to eco-
Wiley and Sons, New York, 1996. nomic dispatch of power systems, in: Proc. 1994 IEEE Conf., Miami, FL, April,
[2] T.A.A. Victoire, A.E. Jeyakumar, Hybrid PSO-SQP for economic dispatch with 1994, pp. 212–216.
valve-point effect, Electr. Power Syst. Res. 71 (1) (2004) 51–59. [34] X.S. Yang, Nature-Inspired Metaheuristic Algorithms, Luniver Press, UK, 2008.
[3] M.A. Abido, Multiobjective evolutionary algorithms for electric power dispatch [35] X.S. Yang, Firefly algorithms for multimodal optimization, in: O. Watanabe, T.
problem, IEEE Trans. Evol. Comput. 10 (3) (2006) 315–329. Zeugmann (Eds.), Stochastic Algorithms: Foundations and Appplications, SAGA
[4] K.Y. Lee, A. Sode-Yome, J.H. Park, Adaptive hopfield neural networks for eco- 2009, Lecture Notes in Computer Science, vol. 5792, Springer-Verlag, Berlin,
nomic load dispatch, IEEE Trans. Power Syst. 13 (2) (1998) 519–526. 2009, pp. 169–178.
[5] W.-M. Lin, F.-S. Cheng, M.-T. Tsay, An improved tabu search for economic dis- [36] K. Gazi, K.M. Passino, Stability analysis of social foraging swarms, IEEE Trans.
patch with multiple minima, IEEE Trans. Power Syst. 17 (1) (2002) 108–112. Syst. Man Cybern. Part B – Cybern. 34 (2004) 539–555.
[6] F.N. Lee, A.M. Breipohl, Reserve constrained economic dispatch with prohibited [37] C.A. Floudas, Nonlinear Mixed-Integer Optimization. Fundamentals and Appli-
operating zones, IEEE Trans. Power Syst. 8 (1) (1993) 246–254. cations, Oxford University Press, New York, USA, 1995.
[7] P.H. Chen, H.C. Chang, Large-scale economic dispatch by genetic algorithm, IEEE [38] L. Costa, P. Oliveria, Evolutionary algorithms approach to the solution jof mixed
Trans. Power Syst. 10 (4) (1995) 1919–1926. integer non-linear programming problems, Comput. Chem. Eng. 21 (2001)
[8] J.C. Dodu, P. Martin, A. Merlin, J. Pouget, An optimal formulation and solution 257–266.
of short-range operating problems for a power system with flow constraints, [39] S.-K. Wang, J.-P. Chiou, C.-W. Liu, Nonsmooth/non-convex economic dispatch
Proc. IEEE 60 (1) (1972) 54–63. by a novel hybrid differential evolution algorithm, IET Gener. Transm. Distrib.
[9] C.L. Chen, S.C. Wang, Branch-and bound scheduling for thermal generating 1 (5) (2007) 793–803.
units, IEEE Trans. Energy Convers. 8 (2) (1993) 184–189. [40] C.L. Chiang, Improved genetic algorithm for power economic dispatch of units
[10] J. Parikh, D. Chattopadhyay, A multi-area linear programming approach for with valve-point effects and multiple fuels, IEEE Trans. Power Syst. 20 (4) (2005)
analysis of economic operation of the Indian power system, IEEE Trans. Power 1690–1699.
Syst. 11 (February (1)) (1996) 52–58. [41] L.D.S. Coelho, V.C. Mariani, Erratum correlation to combining of chaotic
[11] J.I-Y. Fan, L. Zhang, Real-time economic dispatch with line flow and emission differential evolution and quadratic programming for economic dispatch opti-
constraints using quadratic programming, IEEE Trans. Power Syst. 13 (May (2)) mization with valve point effect, IEEE Trans. Power Syst. 21 (3) (2006) 1465.
(1998) 320–325. [42] S.H. Ling, H.H.C. Iu, K.Y. Chan, H.K. Lam, B.C.W. Yeung, F.H. Leung, Hybrid particle
[12] J. Nanda, L. Hari, M.L. Kothari, Economic emission load dispatch with line flow swarm optimization with wavelet mutation and its industrial applications, IEEE
constraints using a classical technique, IEE Proc. Gener. Transm. Distrib. 141 (1) Trans. Syst. Man Cybern. Part B – Cybern. 38 (3) (2008) 743–763.
(1994) 1–10. [43] D. Liu, Y. Cai, Taguchi method for solving the economic dispatch problem with
[13] J.F. Bard, Short-term scheduling of thermal-electric generators using nonsmooth cost functions, IET Gener. Transm. Distrib. 1 (5) (2007) 793–803.
Lagrangian relaxation, Oper. Res. 36 (5) (1988) 756–766. [44] J.B. Park, K.S. Lee, J.R. Shin, K.Y. Lee, A particle swarm optimization for economic
[14] P.G. Lowery, Generating unit commitment by dynamic programming, IEEE dispatch with nonsmooth cost functions, IEEE Trans. Power Syst. 20 (1) (2005)
Trans. Power Apparat. Syst. PAS-85 (5) (1996) 422–426. 34–42.
[15] I.G. Damousis, A.G. Bakirtzis, P.S. Dokopoulos, Network constrained economic [45] S.H. Ling, F.H.F. Leung, An improved genetic algorithm with average-bound
dispatch using real-coded genetic algorithm, IEEE Trans. Power Syst. 18 (1) crossover and wavelet mutation operation, Soft Comput. 11 (1) (2007)
(2003) 198–205. 7–31.
[16] C.L. Chiang, Genetic-based algorithm for power economic load dispatch, IEE [46] L.D.S. Coelho, V.C. Mariani, Combining of chaotic differential evolution and
Proc. Gener. Transm. Distrib. 1 (2) (2007) 261–269. quadratic programming for economic dispatch optimization with valve-point
[17] S. Kumar, R. Naresh, Non-convex economic load dispatch using an efficient effect, IEEE Trans. Power Syst. 21 (2) (2006) 989–996.
real-coded genetic algorithm, Appl. Soft Comput. 9 (1) (2009) 321–329. [47] J.S. Alsumait, J.K. Sykulski, A.K. Al-Othman, A hybrid GA–PS–SQP method to
[18] S. Ching-Tzong, L. Chien-Tung, New approach with a Hopfield modeling frame- solve power system valve-point economic dispatch problems, Appl. Energy 87
work to economic dispatch, IEEE Trans. Power Syst. 15 (2) (2000) 541–545. (2010) 1773–1781.
[19] N. Sinha, R. Chakrabati, P.K. Chattopadhyay, Evolutionary programming tech- [48] J.S. Al-Sumait, A.K. Al-Othmam, J.K. Sykulski, Application of pattern search
niques for economic load dispatch, IEEE Trans. Evol. Comput. 7 (1) (2003) method to power system valve-point economic load dispatch, Electr. Power
83–94. Energy Syst. 29 (10) (2007) 720–730.
[20] A. Bhattacharya, P.K. Chattopadhyay, Solving complex economic load dispatch [49] Z.L. Gaing, Particle swarm optimization to solving the economic dispatch
problems using biogeography-based optimization, Expert Syst. Appl. 37 (2010) considering generator constraints, IEEE Trans. Power Syst. 18 (3) (2003)
3605–3615. 1718–1727.
[21] A. Pereira-Neto, C. Unsihuay, O.R. Saavedra, Efficient evolutionary strategy opti- [50] C. Jiejin, M. Xiaqqian, L. Lixiang, P. Haipeng, Chaotic particle swarm opti-
mization procedure to solve the nonconvex economic dispatch problem with mization for economic dispatch considering the generator constraints, Energy
generator constraints, IEE Proc. Gener. Transm. Distrib. 152 (5) (2005) 653–660. Convers. Manage. 48 (2007) 645–653.