Energy Conversion and Management 89 (2015) 43–62

Contents lists available at ScienceDirect

Energy Conversion and Management

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

A new modified artificial bee colony algorithm for the economic dispatch
problem
Dinu Calin Secui ⇑
Department of Energy Engineering, University of Oradea, Universitatii, 1, 410087 Oradea, Romania

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 new modified artificial bee colony algorithm (MABC) is proposed to solve the economic
Received 2 June 2014 dispatch problem by taking into account the valve-point effects, the emission pollutions and various
Accepted 11 September 2014 operating constraints of the generating units. The MABC algorithm introduces a new relation to update
the solutions within the search space, in order to increase the algorithm ability to avoid premature con-
vergence and to find stable and high quality solutions. Moreover, to strengthen the MABC algorithm per-
Keywords: formance, it is endowed with a chaotic sequence generated by both a cat map and a logistic map. The
Economic/emission dispatch
MABC algorithm behavior is investigated for several combinations resulting from three generating
Artificial bee colony algorithm
Chaotic maps
modalities of the chaotic sequences and two selection schemes of the solutions. The performance of
Valve-point effects the MABC variants is tested on four systems having six units, thirteen units, forty units and fifty-two ther-
Transmission losses mal generating units. The comparison of the results shows that the MABC variants have a better perfor-
mance than the classical ABC algorithm and other optimization techniques.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction by the optimal planning of the load dispatched among the system’s
generating units.
Economic dispatch (EcD) is an important optimizing problem in Solving EcD/EmD involves formulating a mathematical model of
the electric power system operating. The aim of the EcD problem is optimization and then selecting and/or developing an appropriate
to determine the power outputs of all generating units from a sys- optimization technique. The simplest model for the EcD (EmD)
tem, for a given time interval (one hour), in order to have mini- problem is one in which the fuel cost (or the emissions level) of
mum fuel cost, and to meet the required constrains. the generating units is defined by a quadratic function, and the
Another problem that plays an important role in the operating constraints are limited to only two: the equality between the pow-
of the power systems is the emission dispatch (EmD). The EmD ers generated and demanded in the system, respectively the gener-
problem is similar to the EcD problem, but in this case it aims ating units operation between the minimum and maximum limits
determining the power outputs of the generating units that mini- of power. However, in order to take into account more practical
mize the emissions’ quantity – nitrogen oxides (NOx), carbon diox- features of the operating units, the mathematical model has been
ide (COx), sulfur oxides (SOx) – released in the environment (caused improved by considering also the valve-point effects and by intro-
by the burning of fossil fuels within thermal power plants), in cer- ducing the constraints regarding the ramp rate limits, prohibited
tain imposed operating conditions. Searching for an optimal solu- zones of the units, the transmission losses. These additions deter-
tion in the dispatching of the power outputs of the generators mine a non-linear, non-smooth and non-continuous mathematical
taking into consideration both fuel cost and the environmental model of optimization. To solve the EcD/EmD or CEED problem,
emissions level, is known as Combined Economic and Emission several methods, classic or based on artificial intelligence, have
Dispatch (CEED) problem. Basically, EcD and EmD problems aim been used over time. Among the classical methods the following
to find a way to reduce operating costs and emissions levels for a can be mentioned: quadratic programming [1], non-linear pro-
power system without making investments in new equipment gramming [2,3]. The classical methods may have difficulties in
and/or technologies. The cost and emissions reduction is achieved finding the optimal solution due to the non-linear and non-contin-
uous nature of the optimization model [4,5].
As a result, several methods based on artificial intelligence were
⇑ Tel.: +40 751912527; fax: +40 0259414626. developed and applied for the above mentioned problems. These
E-mail address: [email protected] methods have the ability to identify higher quality solutions

http://dx.doi.org/10.1016/j.enconman.2014.09.034
0196-8904/Ó 2014 Elsevier Ltd. All rights reserved.
44 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

[4,6,7] and can be grouped into three categories. The first category constraints [51]. Also, in [41] is concluded that ABC algorithm is
consists of methods applied in their original version, the second good at exploration but poor at exploitation.
refers to methods/algorithms derived from the original version To overcome these problems several variants of ABC algorithm
by changing the relations to updating the solutions or by adapting have recently been proposed in order to enhance the performance
some of the algorithm’s parameters, and the third includes the of the original version.
hybrid methods. Some of the ways used to achieve this purpose are: the introduc-
From the first category several methods to solve the EcD/EmD tion of new relations (inspired by other algorithms such as DE or
or CEED problems were applied, such as: Genetic Algorithm (GA) PSO) to update the solutions in the search space [39–43,51], ABC
[8,9], Particle Swarm Optimization (PSO) [4,8], Differential Evolu- algorithm hybridization with other classical algorithms (derivative
tion (DE) [5,8], Virus Optimization Algorithm (VOA) [8], Harmony methods [52]) or metaheuristics (PSO, DE) [53,54], using chaotic
Search (HS) [8], Seeker Optimization Algorithm (SOA) [8], Ant Col- maps [55], fuzzy theory [56], contraction strategies of the solutions’
ony Optimization (ACO) [10], Artificial Bee Colony (ABC) [11], Bac- search space [57], multi-objective ABC [58], self-adaptation of
terial Foraging Optimization (BFO) [12], Gravitational Search mutation step size [59].
Algorithm (GSA) [8,13], Firefly Algorithm (FA) [8,14], Imperialist In this paper, a modified ABC algorithm (called MABC) is pro-
Competitive Algorithm (ICA) [15]. posed, to solve the economic dispatch problem, taking into account
In the second category of methods the following are included: systems having various characteristics (different operating con-
Modified HS algorithm (MHSA) [16], Self-Adaptive Real Coded straints, the emission pollutions, the valve-point effects). The MABC
Genetic Algorithm (SARGA) [17], PSO with Time Varying Accelera- algorithm has the structure of the original ABC algorithm, but relies
tion Coefficients (PSO_TVAC) [18], Improved Bacterial Foraging on a new relation for the update of the solutions in the search space,
Algorithm (IBFA) [19], Modified Bacterial Foraging Algorithm based on the relations presented in [42] and [51]. In [42] an ABC
(MBFA) [20], Improved Harmony Search with Wavelet Mutation variant it is proposed – Global best artificial bee colony algorithm
(IHSWM) [21], Modified ABC (MABC) [22], Incremental ABC (IABC), – that updates the solutions using the information from the neigh-
ABC with Dynamic Population (ABCDP), ABCDP with Local Search borhood of the best solution. To improve the convergence rate of
(ABCDP-LC) [23], Modified Shuffled Frog Leaping algorithm (MSFL) the basic ABC algorithm, in [51] two parameters are introduced to
[24]. control the frequency and magnitude of the perturbation in the
Among the hybrid methods a few are mentioned: Hybrid Differ- update process of the solutions.
ential Evolution algorithm (HDE) [25], Hybrid Genetic Algorithm The MABC algorithm behavior is also investigated under two
based on differential evolution (HGA) [26], hybrid PSO with conditions: (i) incorporation of chaotic components (generated
Sequential Quadratic Programming (PSO-SQP), Evolutionary Pro- by cat and logistic maps) in the relation used for the solutions
gramming-SQP (EP-SQP) [27], hybrid GA–Pattern Search–SQP update; (ii) the introduction of two schemes for the solutions
(GA–PS–SQP) [28], Fuzzy Adaptive PSO with Variable Differential selection (disruptive selection and classical proportional selection).
Evolution (FAPSO-VDE) [29], Hybrid Multi-Agent based PSO (HMA- It must be mentioned that the cat map and disruptive potential
PSO) [30], hybrid Chaotic PSO algorithm and SQP technique (CPSO– selection scheme have not been investigated when solving the eco-
SQP) [31], hybrid that combines Shuffled frog leaping algorithm nomic/emission dispatch problem. The performance of the MABC
and Differential Evolution (SDE) [32], Hybrid swarm intelligence method was tested on systems with different characteristics and
based Harmony Search (HHS) [33], hybrid Simulated Annealing then compared with several other optimization techniques from
with PSO (SA-PSO) [34]. the solutions stability and quality point of view.
Each of the methods above have their advantages and disadvan-
tages regarding the quality and stability of the solution, the con- 2. Formulation of the problems (EcD, EmD, CEED)
vergence of iterative process, the calculus efficiency or the
simplicity of implementation. This paper proposes a method from 2.1. The economic dispatch (EcD) problem
the second category, being a modified ABC method.
The ABC algorithm is a biological-inspired optimization tech- The mathematic optimization model for the EcD problem
nique, based on population which imitates the foraging behavior includes three elements – optimizing variables, the objective func-
of the real honey bee, being introduced by Karaboga in 2005 tion and constraint relations – listed below.
[35]. Initially, the ABC algorithm has been used in solving uni- The optimizing variables are the real output powers
modal and multi-modal numerical optimization problems on a Pj, j = 1,2, . . . n of the generating units, presented as a solution vec-
limited set of test functions [35]. Due to its many positive features tor P = [P1,P2, . . . Pj, . . . , Pn], where n is the total number of units from
– simple in concept and easy to implement, flexible, the possibility the analyzed system.
of using chaotic maps and of developing hybrids from combina- The objective function F[P] aims to minimize the fuel total cost
tions with other techniques – the ABC algorithm has been success- for the entire system:
fully applied to the optimization of complex mathematical
functions with or without constraints [36–43], or solving engineer- X
n
min F ¼ F j ðPj Þ ð1Þ
ing problems, such as: the optimization of truss structures [44],
j¼1
engineering design optimization [45], automatic voltage regulator
system [46], design and economic optimization of shell and tube where Fj(Pj) is the fuel cost, in $/h, for the jth unit, which is modeled
heat exchangers [47], fault section estimation in power systems by a quadratic polynomial function such as:
[48], optimal reactive power flow [49], economic dispatch [50].
Several studies show that the ABC algorithm performs better or F j ðPj Þ ¼ cj P2j þ bj Pj þ aj ð2Þ
just as good as other biological-inspired algorithms, such as genetic
If the valve-point effects are taken into consideration, then the cost
algorithm, differential evolution and particle swarm optimization
function for unit j includes a sine factor, besides the quadratic poly-
[36–38]. Besides the already mentioned advantages, ABC algorithm
nomial function [19,27,60]:
may encounter a number of challenges (sub-optimal solutions, low
robustness, slow convergence) in the optimization of composite F j ðPj Þ ¼ cj P2j þ bj Pj þ aj þ jej sinðf j ðPj;min  Pj ÞÞj ð3Þ
functions, non-separable functions or problems that require
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 45

where aj, bj and cj are the fuel cost coefficients of unit j, and ej and fj generating unit j, can be modeled through a quadratic function to
are the coefficients of unit j reflecting the valve-point effects. which an exponential term is added. The total quantity of emis-
The EcD problem constrains can be expressed by the relations sions (E) released in the environment by the operation of n units
(4)–(9) [4,5]: is described by relation [23]:
X
n X
n
(i) Minimum and maximum operating power limits are: min E ¼ Ej ðP j Þ ¼ ðcj þ bj Pj þ aj P2j þ nj expðkj Pj ÞÞ ð13Þ
j¼1 j¼1
Pj;min  Pj  Pj;max ; j ¼ 1; 2; . . . ; n ð4Þ
where Pj,min and Pj,max represent the minimum and the max- where aj, bj, cj, nj, kj are the emission coefficients of the generating
imum operating power limits of unit j. unit j.
(ii) Generator ramp-rate limits: The constraints of the EmD problem are the ones taken into
consideration in the EcD problem in the relations (4)–(9).
Pj  P0j þ URj ; if the output power increases ð5Þ
Pj  P0j  DRj ; if the output power decreases ð6Þ 2.3. The combined economic and emission dispatch (CEED) problem

where P0j is the output power of unit j in the previous hour. CEED is an optimizing problem with two objectives: minimizing
DRj and URj are the down-ramp and up-ramp limits of the j the total fuel cost (F) and the quantity of polluting emissions (E).
unit (in MW/time-period). One way to solving the CEED problem is changing it into an opti-
mizing problem with a single objective, using relation [16]:
Combining the relations (4)–(6) the following output power (Pj)
X
n X
n
limits for the unit j are obtained: FE ¼ w  Ej ðP j Þ þ ð1  wÞ  ðFðPj;max Þ=EðPj;max ÞÞ  Ej ðPj Þ ð14Þ
j¼1 j¼1
POj;min  Pj  POj;max ð7Þ
where POj,min = max(Pj,min, P0j  DRj) and POj,max = min(Pj,max, where w is a weighting factor with values in the range of [0, 1]. The
P0j + URj). ratio F(Pj,max)/E(Pj,max) is a scaling factor specific to each generating
(iii) Unit’s j operating zones, considering the prohibited zones unit, that makes possible summing up the Fj(Pj) and Ej(Pj) terms. If
are: relation (14) considers w = 1 or w = 0, then two particular forms
8 of the CEED problem are obtained: the EcD problem (when w = 1)
>
> P 6 Pj 6 Plj;1 and the EmD problem (when w = 0). The two values taken by w
< j;min
determine two points in the objectives space (Fuel cost, Emission),
Pj;z1 6 Pj 6 Plj;z ; z ¼ 2; 3; . . . ; NZ j
u
ð8Þ
>
> located on the Pareto-front. In order to obtain other points (solu-
: Pu 6 P 6 P
j;NZ j j j;max tions) located on the Pareto-front, the weight factor varies from 0
value to 1, using a step of 0.1 [16]. In practice, the system’s operator
where NZj is the number of prohibited zones of unit j. P lj;z and
must chose a single operating point as a compromise between the
Puj;z are the lower and upper boundary of the z prohibited
two opposing objectives (Fuel cost and Emission). Following this
operating zone for the unit j.
the question for the solution ranking from Pareto front is put and
(iv) Real power balance constraint is:
determine the best compromise solution (BCS), that helps the sys-
PG  PL  PD ¼ 0 ð9Þ tem’s operator in taking a decision regarding the current function-
ality of the system (based on a priority list). Next two ways in
where PD is the load demand of the system, in MW. PL repre-
choosing a BCS are shown, based on a fuzzy mechanism. Consider-
sents the transmission line losses, in MW.
ing the Pareto-front was obtained following the previously
described process (changing the w factor in relation (14)), the first
The transmission line losses PL at the entire system level are
method consists of the following steps [20]:
quadratic functions in relation to variables Pj and they are calcu-
(i) For each objective i and each solution k from the Pareto-
lated using the constant B coefficient formula:
front a value using a linear membership function is
X
n X
n X
n calculated:
PL ¼ Pi Bij Pj þ B0i Pi þ B00 ð10Þ 8
i¼1 j¼1 i¼1 >
< 1 f i;k 6 f i;min
f i;min f i;k
where Bij is an element of the loss coefficient matrix of size nxn, B0i li;k ¼ f i;max f i;min f i;min < f i;k < f i;max ð15Þ
>
:
is i element of the loss coefficient vector of size n and B00 is the loss 0 f i;k P f i;max
coefficient constant.
where fi,min and fi,max represent the minimum and maximum
The total generated power (PG) is:
value of ith objective function.
X
n
(ii) For each non-dominated solution k, the normalized linear
PG ¼ Pj ð11Þ
membership function l⁄k is calculated as follows:
j¼1

N P
Relation (9) is satisfied with a certain tolerance (TOL), calculated by li;k
the relation (12):
lk ¼ PP i¼1
PN ð16Þ
k¼1 i¼1 li;k
TOL ¼ PG  PL  PD ð12Þ
where P is the number of non-dominated solutions, and N is
the number of the objective functions.
2.2. The emission dispatch (EmD) problem (iii) The best compromise solution corresponds to the maximum
value of lk : lmax = max(lk , k = 1,2, . . . ,P).
The EmD problem is similar to the EcD problem, but instead of
minimizing the fuel cost it aims minimizing the quantity of emis- The second method is applied in a similar manner with the first
sions released in the environment. The quantity of emissions Ej(Pj), one, with the difference that step 2 is replace by the relation (17)
expressed in ton/h, released in the environment by the operation of for the quantity calculus l1k [19]:
46 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

l1k ¼ minfli;k ; i ¼ 1; 2; ::; Ng; k ¼ 1; 2; . . . ; P ð17Þ The new position of the food source Vi = [vi1,vi2, . . . ,vij, . . . vin],
i = 1,2 . . . SN, j = 1,2 . . . n, determined through the relation (20), is
Then, BCS is chosen using a similar relation with the one in step 3:
evaluated and then compared to the old position Xi. If the new food
lmax = max(l1k, k = 1,2, . . . ,P). source is better than the old one, then the position of the new
Details regarding the non-dominated solutions and the Pareto-
source is memorized by the bee. Otherwise, the old position is
front notions can be followed in [19,58].
retained in the memory. After all employed bees SN conducted a
search, they share the information regarding the quality of food
3. The original ABC algorithm sources and their positions, with the onlooker bees in a waggle
dance.
In the ABC algorithm, the swarm includes three categories of An onlooker bee evaluates the information from all employed
bees (scout bees, employed bees and onlooker bees), each of them bees, and then chooses a food source with a probability based on
carrying on various activities to sustain the hive life. Scout bees its quality. The onlooker bee makes the change, in its memory, of
perform a random search for food sources in a certain area. the food source’s position, described by relation (20). Then, as in
Employed bees have the role of exploiting the identified food the case of the employed bees, the quality of the new food source
sources and then sharing various pieces of information (such as (solution) is evaluated and compared with the quality of the old
orientation and quality of the food source) with other bees waiting food source. If the new food source is better than the old one,
in the hive (called onlooker bees) to make an as good as possible the bee memorizes the new position. Otherwise, the old position
decision in choosing a food source. In the ABC algorithm the fol- is maintained. An onlooker bee chooses a food source with a prob-
lowing associations are made: food sources positions represent ability pi, determined by relation (21):
possible solutions to the optimizing problem and the amount of
food (nectar) in the sources represents the value of fitness function.
fitnessi
The number of the employed bees (SN) is equal to the number of pi ¼ PSN ð21Þ
the onlooker bees and to the number of food sources from the u¼1 fitnessu
search area.
Both in the employed bees phase, and in the onlooker bees phase, if
The ABC algorithm consists of the following phases [35,37]: ini-
the new position Vi established by relation (20) is better than the
tialization, employed bee phase, computation of the probabilities
old position Xi, then Triali counter is set to zero. Otherwise Triali
for onlooker bees, onlooker bee phase and scout bee phase, as
counter is incremented by 1.
described bellow.
In the ABC algorithm, if a food source position cannot be
In the initialization process SN scout bees randomly search and
improved after a predetermined number of trials (called Limit),
identify SN food sources in a limited area, and then they become
then the food source is considered abandoned. The abandoned
employed bees. The random positions of the food sources, which
source is replaced with a new food source discovered by a scout
also represent the initial solutions, can be determined with relation
bee. The position of the new source is randomly generated using
(18):
relation (18). Other details of the ABC algorithm can be seen in
xij ¼ xj;min þ ðxj;max  xj;min Þ randð0; 1Þ ð18Þ [35–38].
The ABC algorithm uses three control parameters: the number
where xij is the jth component of i solution vector, at iteration zero;
of employed bees or onlooker bees (SN), the value of the predeter-
xj,min, xj,max are the minimum and maximum values of j components
mined parameter (Limit) and the maximum number of iterations
(j = 1,2 . . . n, where n represents the dimension of solution vector)
(kmax).
for the Xi vector; i = 1,2 . . . SN; rand(0, 1) is uniformly distributed
The main steps of the algorithm are given below [36,38]:
random number in the range of (0, 1).
After initializing the SN solutions Xi = [xi1, xi2, . . . xij, . . . xin],
i = 1,2 . . . SN, they are evaluated using the objective function Initialization of the population;
F(Xi) = Fi. k=1;
For a minimization problem the fitness function (fitness(Xi) = fit- Repeat
nessi) corresponding to Xi solution is determined by the relation: Employed bees phase;
 Computation of the probabilities for onlooker bees;
1=ð1 þ F i Þ; if F i P 0 Onlooker bees phase;
fitnessi ¼ ð19Þ
1 þ absðF i Þ; otherwise Scout bees phase;
k = k + 1; (k is the current number of the iterations)
In the case of a maximization problem the fitness function is repre-
Until k = Maximum iterations number (kmax).
sented by the objective function (fitnessi = Fi).
Memorizing the best solution achieved.
Also, in this phase, a counter Triali (Triali = 0, i = 1,2 . . . SN) is
initialized, memorizing the number of trials in which a solution
vector Xi could not be improved.
After the initialization phase, an iterative search process is con- 4. Several strategies for solutions update and selection schemes
ducted to determine the new positions of food sources, including
employed bees, onlooker bees and scout bees. 4.1. Strategies for solutions update
In the ABC algorithm, the employed bees make a change (in
their memory) on the positions of food sources, based on the local In this section three strategies are presented, aiming to improve
information. The process of updating the solutions is described by the performance of ABC algorithm, that are based on modifications
the relation: of the solution update relations within employed bees phase and
onlooker bees phase.
v i;j ¼ xi;j þ Ui;j ðxi;j  xk;j Þ ð20Þ These strategies were presented in [42,43,51] and underlay the
where, vij, xij represent the new, respective the old position (solu- modified ABC (MABC) proposed in this paper. The relations
tion) of food source i for the jth component, i = 1,2 . . . SN, describing the strategies are:
j = 1,2 . . . n; k = 1,2 . . . SN is a randomly chosen index and k – i; Uij
is a random number in the range [1, 1]. xi;j ¼ xi;j þ Ui;j ðxr1;j  xr2;j Þ ð22Þ
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 47

xi;j ¼ xbest;j þ Ui;j ðxr1;j  xr2;j Þ ð23Þ Table 1

Chaotic maps used.

xi;j þ Ui;j ðxi;j  xk;j Þ; if r i;j < MR Chaotic maps Equation
xi;j ¼ ð24Þ
xi;j ; otherwise Cat map [61] xk+1 = (xk + yk) mod 1;
yk+1 = (xk + 2yk) mod 1;
where xbest,j is the jth component of the best solution; r1 and r2 are where x mod 1 = x  [x];
random indexes chosen from {1,2, . . . ,SN}, with r1 – r2 – i. MR is a x0, y0 2 (0, 1), xk, yk 2 (0, 1).
control parameter, and ri,j is a random number in the range [0, 1]. Logistic map [62] xk+1 = 4xk(1  xk); xk 2 (0, 1);
x0 2 (0, 1) x0 – {0.25,0.5,0.75}.
The significance of other parameters was presented in Section 3.
The first strategy is defined by (22) and is inspired by the origi- x0, y0 represent the initial values used to generate the chaotic sequences.
nal ABC algorithm [37], but the vectors (xi,j, xk,j) from the difference {xk}k=1. . .1, {yk}sk=1. . .1 represent the numbers generated through a chaotic map, at
the kth iteration.
(xi,j  xk,j) are replaced by two other vectors (xr1,j, xr2,j) randomly
selected from the current population. The second strategy is based
on the relation (23), proposed in [42], in which the search for new
solutions is done around the best solutions (xbest,j). The last strategy characteristics were modified. These modifications were imple-
uses the relation (24), proposed in [51], in which some parts of the mented in employed bees phase and the onlooker bees phase
current solution (xij) are altered if the condition rij < MR is satisfied. and are presented below.
Basically, the relations (22)–(24) replace the relation (20) of the
original ABC algorithm. (i) relation (20) of the ABC algorithm for solutions update was
The algorithms that are based on the relations (23) and (24) replaced with relation (26):
may have better performance than the original ABC algorithm 
xi;j þ ð2cv ij  1Þðxr1;j  xr2;j Þ; if ri;j < SF
[42,43,51] or other population-based algorithms, such as differen- xi;j ¼ ð26Þ
tial evolution, dynamic multi-swarm particle swarm optimizer xbest;j þ ð2cv ij  1Þðxr1;j  xr2;j Þ; otherwise
with local search, covariance matrix adaptation evolution strategy where xi,j represents the jth component of the current solu-
with increasing population size [51]. These aspects recommend the tion i; xbest,j is the jth component of the best solution; r1
use of relations (22)–(24) in testing new combinations of strategies and r2 are random indexes chosen from {1, 2, . . . , SN}, with
to update the solutions, which could lead to improve their quality r1 – r2 – i; cvij are chaotic variables whose values are gener-
and stability. ated by chaotic maps; ri,j is a uniform random number in the
range [0, 1]; SF is a selection factor which takes values in
4.2. Schemes for solutions selection range [0, 1].

Besides the classical proportional scheme of solutions selection Relation (26) follows combining two solutions update relations
(called ‘‘Proportional’’) – commonly used the ABC algorithm, the inspired by [37] and respectively [42], through the selection factor
potential of disruptive selection scheme (Disruptive) is analyzed. (SF). Since the first equation described by relation (26) performs a
Proportional and Disruptive schemes are based on the roulette search around the current solution xij, and the second equation
wheel selection method and the probabilistic selection (pi) is deter- around the best solution of the previous iteration (xbest,j), relation
mined using the same relation in both schemes – relation (21). The (26) has the capacity of searching for solutions at least as good
difference between the two selection schemes consists in the com- as the two equations taken individually. The new solution gener-
putation of the fitness function’s value. Thus, in the Proportional ated by relation (26) differs both from current solution (xij), and
scheme (commonly used in the ABC algorithm) the fitness function the best solution xbest,j. This way of defining relation (26) eliminates
value is calculated with relation (19) and in the Disruptive scheme the drawback set by using relation (24) [51], in which the new
the fitness function value is calculated with relation (25) [63,64]: solution generated may coincide with the current solution.
fitnessi ¼ jF i  meanðF i Þj; i ¼ f1; 2; . . . ; SNg ð25Þ
(ii) in the MABC algorithm the values of the variables (cvij), from
where Fi = Fi(Xi) is the objective function for Xi solution, and relation (26), are generated using two chaotic maps (Table 1).
mean(Fi) is the average value of the Fi functions from the current Chaotic maps have been successfully used in several algo-
iteration. rithms, in order to improve the population diversity and to
Classical proportional selection (Proportional), presents some obtain better solutions in the end [42,43,55,57].
disadvantages, one of them being the possibility of creating a
‘‘super-individual’’ towards which the entire population converges With the same purpose, this paper investigates the behavior of
gradually, leading to population’s diversity reduction and conse- two chaotic maps (cat map [61] and logistic map [62]), whose
quently the algorithm’s stagnation in a local minimum. By its def- characteristics are shown in Table 1. In Fig. 1 the frequencies of
inition, the disruptive selection tends to favor solutions that lead to the values generated by the two maps are presented. Logic map
extreme values (minimum/maximum) for the functions Fi, the
other solutions having a lower probability of being selected in 3500
the next iteration. Also, in the case of onlooker bees, disruptive Cat map
3000
selection could improve the diversity of the population and main- Logistic map
tain the balance between exploration and exploitation processes 2500
Frequence

[63,64]. In this paper disruptive selection is compared to classical 2000

proportional selection of solutions (commonly used the ABC 1500
algorithm).
1000
500
5. Modified ABC (MABC) algorithm xk
0
0 0.2 0.4 0.6 0.8 1
The MABC proposed algorithm has the general structure of the
original ABC algorithm, presented in Section 3, but several of its Fig. 1. The frequency of the values generated by cat and logistic maps.
48 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

shows higher frequencies at the margins of the (0, 1) range, and cat {Compare the old solution Xi with the new solution X00 i};
map shows a relatively uniform frequency of values within the 25: If F(X00 i) < F(Xi) Then begin Xi = X00 i, Triali = 0 end Else
(0, 1) range. The logistic map was commonly used in solving vari- Triali = Triali + 1;
ous problems, having good results. The cat map has not been pre- {Update the best solution Xbest and function Fbest = F(Xbest)};
viously used for EcD/EmD problem solving therefore with this 26: If F(Xi) < F(Xbest) Then begin Xbest = Xi; Fbest = F(Xbest)
paper the aim is to investigate its potential in EcD/EmD problem end
by MABC algorithm. A detailed procedure for the MABC algorithm 27: h = h + 1 (end If)
is presented below. 28: Else begin i = i + 1; If i = SN + 1 Then i = 1 end
29: Until h 6 SN (end Repeat)
{Phase 1: Initialization phase} {Phase 5: Scout bee phase}
1: The parameters for the MABC algorithm are set: SN, n, Limit, 30: The index of the abandoned Xix solution is identified from
SF, kmax; the condition ix = arg{max(Triali), i=1,2, . . . SN}.
2: Initialize: the iteration counter k (k = 0); the counters 31: If Trialix P Limit Then begin
Triali = 0, i=1,2 . . . SN; the initial values for the chaotic 32: Replace solution Xix with a new solution randomly
sequences cx0, cy0 (for Cat map) or cx0 (for Logistic map); generated with relation (18);
3: Initialize the SN solutions (Xi, i = 1,2, . . . SN) using relation 33: Check the imposed constraints on the problem
(18) (for each solutions the imposed constraints shown in (Section 6);
Section 6, are checked); 34: Triali=0; end (end If)
4: Assess the value of the functions Fi=F(Xi) assigned to the 35: k = k + 1 {increment the iterations counter}
initial solutions (Xi, i = 1,2, . . . SN), during iteration k = 0; 36: Until k 6 kmax (end Repeat)
5: Select the index corresponding to the best solution Xbest, 37: Print the optimal solution Xbest and the best value of the
best = Arg(Min(Fi, i=1,2 . . . SN)) and calculate Fbest = F(Xbest). objective function Fbest = F(Xbest).
6: Repeat
{Phase 2: Employed bee phase}
7: For i = 1 To SN Do begin
In the MABC algorithm four parameters are used: the number of
8: Randomly select an index j to be updated,
employed bees or onlooker bees (SN), the number of cycles
j 2 {1,2, . . . n};
required (Limit), the maximum iteration number (kmax) and
9: Randomly select the indexes r1 and r2 out of the
selection factor (SF). Given the possibility of improving the ABC
current population, where r1 – r2 – i, r1, r2 2 {1,2, . . . SN};
algorithm characteristics by introducing the relation (26) for the
10: Update the values of the chaotic sequences for Cat
solutions update, the sequences generated by chaotic maps, and
map (cx, cy) or Logistic map (cx), using the equations
also the selection schemes included, the MABC variants could have
displayed in Table 1;
high performance in terms of solutions’ quality and stability.
11: Update the xij component for the Xi solution using
relation (26); returns a new solution X0 i;
6. Constraints handling mechanism (CHM)
12: Check the imposed constraints on the problem,
shown in Section 6;
In order to create the connection between the notations used
{Compare the old solution Xi with the new solution X0 i};
inside the MABC algorithm (described in Section 5) and the specific
13:If F(X0 i) < F(Xi) Then begin Xi = X0 i, Triali = 0 end Else
notations used by the mathematical model of the stated issue (EcD,
Triali = Triali + 1;
EmD or CEED), we assume that all SN solutions, generated during
{Update the best solution Xbest and function Fbest=F(Xbest)};
iteration k, are symbolized as follows: Xi  Pi or [xi1, xi2, . . . xij,
14:If F(Xi) < F(Xbest) Then begin Xbest = Xi; Fbest = F(Xbest);
. . . ,xin]  [Pi1, Pi2, . . . Pij, . . . ,Pin], i = 1,2 . . . SN; j = 1,2 . . . n.
end (end For)
In order to meet the inequality and equality constraints, the fol-
{Phase 3: Computation of the probabilities for onlooker
lowing mechanism is applied:
bees}
15: Compute the values for the fitnessi functions, i = 1,2 . . . SN
(i) In order to satisfy relation (4) (if the ramp-rate limits are not
using relation (25) (for a Disruptive selection scheme) or
imposed) proceed as follows: if Pij < Pj,min then Pij = Pj,min; if
relation (19) (for a Proportional selection scheme);
Pij > Pj,max then Pij = Pj,max, otherwise Pij remains unchanged.
16: For i = 1 To SN Do Compute the probabilities pi using
(ii) In order to satisfy relation (7) proceed similarly to case (i): if
relation (21);
Pij < POj,min then Pij = POj,min; if Pij > POj,max then Pij = POj,max,
{Phase 4: Onlooker bee phase}
otherwise Pij remains unchanged.
17: i = 1, h = 0
(iii) In order to satisfy relation (8) proceed as follows: if Pij
18: Repeat
belongs to prohibition zone z (having the limits Plj,z and Puj,z),
19: If rand(0, 1) < pi Then
then Pij is set to the closest limit (Plj,z or Puj,z);
20: Randomly select an index j to be updated,
(iv) To ensure the relation (9) complies with the tolerance TOL,
j 2 {1,2, . . . n};
an equality constraint handling mechanism similar with
21: Randomly select indexes r1 and r2 out of the current
the one proposed by [65] is used.
population, where r1 – r2 – i, r1, r2 2 {1,2, . . . SN};
22: Update the values of the chaotic sequences for Cat
7. Case study
map (cx, cy) or Logistic map (cx) using the equations shown
in Table 1;
The efficiency of the proposed method is shown through the
23: Update the xij component of the Xi solution using
analysis of four different test systems consisting of 6, 13, 40 and
relation (26); returns a new solution X00 i;
52 generating units. Depending on their basic characteristics (load
24: Check the imposed constraints on the problem
demand, power losses or a valve-point effect consideration) seven
(Section 6);
cases are analyzed, presented in Table 2.
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 49

Table 2
Basic features of the analyzed systems and the parameters setting for MABC variants.

Parameters PD (MW) kmax Tested values for kmax SN Tested values for SN Limit SF Valve-point effect (VP)
Cases
C1:6-units, with power losses (PL) 1263 400 300, 400, 500 5 4, 5, 6 0.02  kmax 0 No
C2:13-units, 1800 MW 1800 18,000 1.2  104, 1.5  104, 1.8  104 12 10, 12, 14 0.02  kmax 0.9 Yes
C3:13-units, 2520 MW 2520 15,000 1.2  104, 1.5  104, 1.8  104 12 10, 12, 14 0.02  kmax 0.4 Yes
C4:13-units, 1800 MW, with PL 1800 20,000 1.2  104, 1.5  104, 2.0  104 20 10, 15, 20 0.02  kmax 0.9 Yes
C5:13-units, 2520 MW with PL 2520 15,000 1.2  104, 1.5  104, 1.8  104 12 10, 12, 15 0.02  kmax 0.5 Yes
C6:52-units (4  13-units) 7200 25,000 2.0  104, 2.2  104, 2.5  104 20 16, 18, 20 0.02  kmax 0.9 Yes
C7:52-units (4  13-units) 10,080 20,000 1.5  104, 1.8  104, 2.0  104 15 10, 12, 15 0.02  kmax 0.7 Yes
C8:40-units, (EcD problem) 10,500 15,000 1.0  104, 1.2  104, 1.5  104 15 10, 12, 15 0.02  kmax 0.2 Yes
C9:40-units, (EmD problem) 10,500 2500 1500, 2000, 2500 10 6, 8, 10 0.02  kmax 0.2 Yes

The quality of the solutions is evaluated based on a sequence of solutions update) is better than the individual relations that
100 trials. For this sequence the values of the following items are underlie it.
kept: best total fuel cost F (B), average total fuel cost F (A), worst
total fuel cost F (W), standard deviation (SD) and the success rate The MABC algorithm was applied in six variants resulting from
(SR). The SR is defined as follows [66]: SR = NS100/NTrials, where a combination of three modalities of generating random sequences
NTrials represents the total number of trials (NTrials = 100), and NS (cat map, logistic map and ‘‘Random’’) with two selection schemes
is the number of trials for which the condition (F(P)-B)8760 6 dF (Disruptive selection and Proportional selection). The third modal-
is satisfied. F(P) represents the fuel total cost F corresponding to ity (Random) uses the random number Uij from relation (20). The
a feasible solution P, obtained by a certain method, and B is the MABC variants are presented in Table 4 and were applied consider-
best fuel total cost F obtained by the same method. dF quantity is ing that parameters kmax, SN, Limit and SF have the values shown in
an assumed cost, considered insignificant for the analyzed systems. Table 2. All case studies were implemented in MathCAD, on a per-
For this paper an assumption that dF = 1000 $/year (0.1141 $/h) sonal computer having a 1.79 GHz processor and 896 MB of RAM.
was made.
Setting the parameters of the MABC variants (SN, kmax, Limit and 7.1. Test system 1: 6-unit with power losses (case C1)
SF) was done using the MABC/D/Cat algorithm, following a proce-
dure that is based on the design of experiments. This procedure A six-unit test system that takes into consideration the trans-
allows testing the combinations between the different parameters mission losses, ramp rate limits and prohibited operating zones
and the values selected for them [8]. The mentioned parameters of the units is analyzed in solving the EcD problem with the MABC
were set for each individual case (C1–C9) following three stages. variants. The data related to the cost coefficients (a, b, c), power
In each stage the targeted parameters (SN and kmax for the first operating limits, ramp-rate limits, prohibited operating zones of
stage; Limit for the second stage and SF for the last stage) have been the units, and also the loss coefficient B are taken from [4], and cor-
selected in accordance with the best value of item A, and for equal rected after [67] in regards to the B00 factor. The load demand is
values of items A the tiebreaker was based on the item’s SD value. PD = 1263 MW.
In the first stage the SN and kmax parameters were set, for which
three test values were taken into consideration (pointed out in 7.1.1. Solution’s quality and robustness
Table 2). These values were combined among themselves generat- The best solutions obtained by six MABC variants are presented
ing 32 pairs of (SN, kmax). For each pair (SN, kmax) 30 trials were per- in Table 5. In order to analyze the robustness of MABC variants 100
formed with MABC/D/Cat algorithm, followed by the computation independent trials were performed.
of items A and SD. In the second stage for the Limit parameter three The values obtained for items B, A, W and SD using MABC vari-
test values were taken into consideration (0.01kmax, 0.02kmax, ants are presented in Table 6, case C1. Following the values for the
0.03kmax) and the parameters SN and kmax were set with the values B, A, W and SD items, a very good stability for all obtained solu-
obtained during the first stage. With these settings 30 trials were tions is noticed. Also satisfying the power balance is achieved with
performed for each value of the Limit parameter. In the last stage, a very low tolerance (TOLM < 1010 MW, Table 5). However, from
to identify the best values of the SF, the MABC/D/Cat algorithm was Table 6, it can be noted that the MABC/D/Cat variant has the best
run for 11 values in the [0, 1] range using a step of 0.1 (the param- value for SD item, the other items (B, A, W) having very close
eters SN, kmax and Limit are set to the values established in the first values.
two stages). For each value of the SF parameter the items A and SD
(shown in Table 3) are calculated, this time performing 100 trials. 7.1.2. Comparing MABC with other methods
For each case the best value of the SF parameter was recorded in Table 7 shows a comparison between the results reached by the
Table 2. MABC/D/Cat variant and the results obtained by other methods
Analyzing the values of items A and SD from Table 3 for all the recently published. The methods presented in Table 7 are
nine cases (C1–C9), the following can be mentioned: M = {MHSA [16], MSFL [24], IABC, IABC-LS [68], h-PSO [69], DHS
[70]}.
(i) for the cases C2, C4, C6, C7 and C8 SF parameter has an influ- From Table 7 it is found that TOLM tolerance differs from one
ence over A and SD; for the cases C3, C5 and C9 the influence method to another. This difference arises because some references
of SF parameter over A and SD is small; for the case C1 the [16,24,68,69] used data from [4], but in [70] data from [4] were
parameter SF has a low influence over the values of the men- used, corrected on the basis of observations reached from [67]. In
tioned items; order to compare the MABC/D/Cat variant to methods M, MABC/
(ii) for most cases (C2–C9) the best values of SF parameter D/Cat is applied for each method M, imposing the same tolerance
(Table 2) differs from SF = 0 or SF = 1, in relation to items A TOLM (indicated in Table 7). The results are then presented in
and SD. This means that relation (26) (proposed for the Table 8. MHSA methods [16], IABC and IABC-LS [68] do not provide
 50 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

Table 4
Different variants of the MABC algorithm implementation.

121433.6668

176682.2647
15449.8995

17971.8415

24169.9177

18134.4222

24514.8756

96611.1548
71813.2240
6.04  108

3.57  106

3.85  105
1.06  106
MABC variants Selection scheme Modalities of generating
random sequences (GRS)

15.69

11.35

12.03

22.60
2.84
M1: MABC/D/Cat Disruptive (D) Cat map
0

M2: MABC/D/Log Disruptive (D) Logistic map

121434.1985

176682.2647
M3: MABC/D/Rand Disruptive (D) Random (Rand)
15449.8995

17971.2979

24169.9177

18131.9616

24514.8756

96613.8557
71812.9504
1.95  106

7.92  107

3.84  105
1.80  107

M4: MABC/P/Cat Proportional (P) Cat map

M5: MABC/P/Log Proportional (P) Logistic map
16.58

13.10

20.16
3.36

9.78
M6: MABC/P/Rand Proportional (P) Random (Rand)
0.1

M7: ABC Classic Proportional (P) Random (Rand)

121431.7793

176682.2646
15449.8995

24169.9177

18133.6854

24514.8756

71811.6858
17971.0172

96611.9507

3.82  105
2.14  107

8.98  107
9.20  107

sufficient data for conducting the comparisons using statistical

19.16
15.89

11.78
9.78
3.60
0.2

tests. However, a qualitative analysis based on items B, A, W and

SD was achieved below (Table 7 and Table 8), finding the follow-
121434.2588

176682.2647
ing: (i) the MABC/D/Cat has a very good stability (item SD has very
15449.8995

24169.9177

24514.8756
17969.7580

18142.1707

96609.6732
71809.2706
1.36  107

7.48  107

3.48  107

4.11  105
low values), (ii) the MABC/D/Cat is superior to the five methods, in
15.26

11.42

21.92
10.04

consideration to the values of items B, A, W and SD. Furthermore,

4.27
0.3

the W item obtained by the MABC/D/Cat (WMABC/D/Cat) is smaller

than the B item obtained with any other M method (BM) from
121435.2361

176682.2647
24169.9177
15449.8995

24514.8756
17969.5088

18143.6280

71810.6142

96610.9146

Table 7 (WMABC/D/Cat < BM, M = {MHSA, MSFL, IABC, IABC-LS,

6.31  107
2.21  107

8.37  107

8.81  105

The values in bold correspond to SF factor that determines the lowest average cost (A) or the lowest standard deviation (SD), for a particular case (C1–C9).

h-PSO, DHS}); (iii) In order to compare the results obtained by

16.28

11.55

21.52
4.33

9.80

MABC/D/Cat and MSFL/DHS/h-PSO, a parametric statistical test

0.4

(t-test) was applied, assuming a normal distribution for data col-

lections. In all cases the specific value for t-test (tvalue) is greater
176682.2647
121436.6106
24514.8756
15449.8995

17968.4123

24169.9177

18137.5887

71805.5548

96607.3674
3.50  107
5.67  107

6.83  107

1.13  104

than its critical value (tcritic = 2.576 for a significance level of 1%):
for MABC/D/Cat – MSFL comparison the obtained tvalue is 13.04,
12.52

21.58
11.07
4.51

9.52

for MABC/D/Cat and h-PSO tvalue is 12.41, and for MABC/D/Cat

0.5

and DHS tvalue is 13.23. Therefore the null hypothesis (there are
121435.5695

176682.2648

not significant statistic differences between the results obtained

15449.8995

24169.9177

18134.6542

24514.8756
17967.1120

71801.9949

96607.2097
8.19  107

2.12  104
4.06  107

1.09  106

using the two methods) is rejected for a significance level of 1%.

19.57
10.43

11.60

The average computation time for the case C1 is 0.62 s.

4.32

9.00
0.6
The values of items A and SD for the analyzed cases (C1–C9) and SF parameter, obtained by MABC/D/Cat algorithm.

7.2. Test system 2: 13-unit with and without power losses (cases
121441.9779

176682.2650
96604.5720
15449.8995

24169.9177

18134.1873

24514.8756
17964.8207

71801.3616
3.63  107

2.34  106

7.35  104
1.10  106

C2–C5)
22.26
10.32

10.69

5.89
2.82

The MABC variants are used to solve the EcD problem for a 13
0.7

thermal generating unit system with valve-point loading effects,

121444.2366

176682.2662

with and without taking into consideration the transmission line

15449.8995

17963.9194

24169.9177

18134.4222

24514.8756

71797.9524

96606.1833
2.37  106

8.53  106

3.42  103
3.10  106

losses. The data of the system (a, b, c, e, f coefficients and limits

of generated powers) are taken from [60], and the loss coefficient
23.34
10.88

5.67
9.60
0.90
0.8

B from [32]. Two values are considered for the load demand:
PD = 1800 MW, respectively PD = 2520 MW.
121456.6659

176682.2706
17963.8293

18131.9616

71795.9105
15449.8995

24169.9194

24514.8756

96608.6093
2.26  104

2.33  103

2.28  105

1.61  102
3.07  105

7.2.1. Solution’s quality and robustness

11.25

28.16
6.20

5.74

The best solutions obtained by MABC variants are shown in

0.9

Table 9. The solutions are presented for MABC variants that have
121474.3266

the best items A and SD. The values of items B, A, W, SD and SR

176682.3025
15449.8995

24169.9444

18133.6854

24514.8763

71815.8549

96643.6510
17964.2100
3.73  105

3.35  102

9.61  102
1.60  103

for the cases C2–C5, obtained by MABC variants are presented in

Table 6. It can be seen that, generally, the values of the items B,
28.26

29.74
15.04
1.65

3.34

A, W, SD and SR are very close, except for the case C4 (1800 MW

1

with PL) in which there are differences between items. Also, from
SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

SD ($/h)

Table 9 it can be noticed that tolerance values (TOLM) are very

A ($/h)

A ($/h)

A ($/h)

A ($/h)

A ($/h)

A ($/h)

A ($/h)

A ($/h)

A ($/h)

small.
For cases C2, C3 and C5 the success rate is maximum (SR = 100%)
C5: 13-units with PL, 2520 MW
C4: 13-units with PL, 1800 MW

for all MABC variants. This means that the cost of generating
C9:40-units, (EmD problem)

achieved through a MABC variant does not differ from the cost
C8:40-units, (EcD problem)
C7: 52-units, 10,080 MW

obtained by another variant with more than dF = 1000 $/year

C3: 13-units, 2520 MW
C2: 13-units, 1800 MW

C6: 52-units, 7200 MW

(assumed limit). Also, for C4, the best MABC variant (MABC/P/
Log) selected based on item A is the same as the one selected based
on SR criterion.
C1: 6-units

7.2.2. Comparing MABC variants with other methods

Cases
Table 3

Tables 10 and 11 display the results obtained by various optimi-

SF

zation techniques/methods for the case C2 (1800 MW without PL)

 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 51

Table 5
The best solutions obtained through MABC variants, for the case C1 with 6-units (PD = 1263 MW, 100 trials).

Algorithms MABC/D/Cat MABC/D/Log MABC/D/Rand MABC/P/Cat MABC/P/Log MABC/P/Rand

Output
P1 (MW) 447.50320837 447.50446888 447.50397527 447.50362341 447.50316334 447.5039396
P2 (MW) 173.31776370 173.31850746 173.31785688 173.31843533 173.31876029 173.3187842
P3 (MW) 263.46315194 263.46250373 263.46298445 263.46344721 263.46322108 263.4622817
P4 (MW) 139.06507749 139.06536865 139.06521482 139.06581038 139.06543237 139.0650065
P5 (MW) 165.47352386 165.47329094 165.47347683 165.47291888 165.47283280 165.4731297
P6 (MW) 87.13552025 87.13410047 87.13473649 87.13399580 87.13482349 87.13509942
PG (MW) 1275.95824564 1275.95824017 1275.95824476 1275.95823104 1275.95823337 1275.95824111
PL (MW) 12.95824564 12.95824017 12.95824476 12.95823104 12.95823337 12.95824111
BM ($/h) 15449.899525 15449.899525 15449.899525 15449.899525 15449.899525 15449.899525
TOLM (MW) 7.208  1011 3.183  1012 3.865  1012 2.092  1011 7.094  1011 1.590  1012

Table 6
The values of items B, A, W, SD and SR for different analyzed cases (C1–C9) and MABC variants.

Cases System Algorithm B ($/h) A ($/h) W ($/h) SD ($/h) SR (%)

C1 6-units 1263 MW MABC/D/Cat (M1) 15449.8995 15449.8995 15449.8995 6.04  108 100
MABC/D/Log (M2) 15449.8995 15449.8995 15449.8995 1.47  107 100
MABC/D/Rand (M3) 15449.8995 15449.8995 15449.8995 1.30  107 100
MABC/P/Cat (M4) 15449.8995 15449.8995 15449.8995 6.39  107 100
MABC/P/Log (M5) 15449.8995 15449.8995 15449.8995 2.34  106 100
MABC/P/Rand (M6) 15449.8995 15449.8995 15449.8995 1.27  106 100
ABC Classic (M7) 15449.8998 15449.9091 15449.9496 1.01  102 100
C2 13-units 1800 MW MABC/D/Cat (M1) 17963.8292 17963.8293 17963.8305 2.26  104 100
MABC/D/Log (M2) 17963.8292 17963.8294 17963.8317 3.84  104 100
MABC/D/Rand (M3) 17963.8292 17963.8296 17963.8372 9.53  104 100
MABC/P/Cat (M4) 17963.8292 17963.8296 17963.8374 9.24  104 100
MABC/P/Log (M5) 17963.8292 17963.8296 17963.8339 6.80  104 100
MABC/P/Rand (M6) 17963.8292 17963.8301 17963.8560 3.21  103 100
ABC Classic (M7) 17963.8292 17965.2669 17973.2375 3.20 82
C3 13-units 2520 MW MABC/D/Cat (M1) 24169.9177 24169.9177 24169.9177 6.31  107 100
MABC/D/Log (M2) 24169.9177 24169.9177 24169.9177 5.77  107 100
MABC/D/Rand (M3) 24169.9177 24169.9177 24169.9177 6.33  107 100
MABC/P/Cat (M4) 24169.9177 24169.9177 24169.9177 9.19  107 100
MABC/P/Log (M5) 24169.9177 24169.9177 24169.9177 7.13  107 100
MABC/P/Rand (M6) 24169.9177 24169.9177 24169.9177 1.00  106 100
ABC Classic (M7) 24169.9177 24170.0202 24171.8733 2.49  101 78
C4 13-units 1800 MW with PL MABC/D/Cat (M1) 18127.7821 18131.9616 18158.8463 6.20 51
MABC/D/Log (M2) 18127.7821 18130.4507 18134.3784 3.21 59
MABC/D/Rand (M3) 18127.7821 18131.7758 18165.8342 6.61 56
MABC/P/Cat (M4) 18127.7821 18130.0015 18149.8375 3.59 66
MABC/P/Log (M5) 18127.7821 18129.7010 18134.3131 2.95 69
MABC/P/Rand (M6) 18127.7821 18132.3448 18165.8319 7.47 52
ABC Classic (M7) 18127.7821 18134.1370 18150.4282 3.74 10
C5 13-units 2520 MW with PL MABC/D/Cat (M1) 24514.8756 24514.8756 24514.8756 3.50  107 100
MABC/D/Log (M2) 24514.8756 24514.8756 24514.8756 5.09  107 100
MABC/D/Rand (M3) 24514.8756 24514.8756 24514.8756 1.05  106 100
MABC/P/Cat (M4) 24514.8756 24514.8756 24514.8756 1.85  106 100
MABC/P/Log (M5) 24514.8756 24514.8756 24514.8756 1.80  106 100
MABC/P/Rand (M6) 24514.8756 24514.8756 24514.8756 1.37  106 100
ABC Classic (M7) 24514.8756 24515.1370 24540.9439 2.61 99
C6 52-units 7200 MW MABC/D/Cat (M1) 71770.3207 71795.9105 71821.3679 11.25 1
MABC/D/Log (M2) 71770.2823 71788.5302 71808.8760 10.52 8
MABC/D/Rand (M3) 71770.2892 71798.8556 71820.8519 10.09 2
MABC/P/Cat (M4) 71771.0620 71797.4292 71823.8676 11.26 0
MABC/P/Log (M5) 71770.2738 71788.5467 71808.5627 10.61 16
MABC/P/Rand (M6) 71770.2871 71798.6050 71821.7501 11.99 1
ABC Classic (M7) 71785.6965 71825.8070 71886.9981 18.61 0
C7 52-units 10,080 MW MABC/D/Cat (M1) 96601.8504 96604.5720 96629.7631 5.89 20
MABC/D/Log (M2) 96601.8517 96605.2397 96628.3065 6.94 10
MABC/D/Rand (M3) 96601.8509 96606.1761 96630.9067 7.32 11
MABC/P/Cat (M4) 96601.8556 96607.3925 96631.9366 8.25 10
MABC/P/Log (M5) 96601.8549 96605.3712 96629.3396 6.86 16
MABC/P/Rand (M6) 96601.8529 96605.3877 96628.6297 6.07 12
ABC Classic (M7) 96601.8531 96632.1958 96706.5650 20.19 9
C8 40-units, 10500 MW (EcD problem, w = 1) MABC/D/Cat (M1) 121412.5409 121431.7793 121503.7552 19.16 2
MABC/D/Log (M2) 121412.6013 121434.8539 121502.8264 22.37 2
MABC/D/Rand (M3) 121412.8738 121436.7163 121494.5488 19.60 0

(continued on next page)

52 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

Table 6 (continued)

Cases System Algorithm B ($/h) A ($/h) W ($/h) SD ($/h) SR (%)

MABC/P/Cat (M4) 121412.5796 121432.3026 121502.8316 18.54 1
MABC/P/Log (M5) 121412.5918 121431.5763 121493.1885 18.16 1
MABC/P/Rand (M6) 121412.5390 121435.5004 121502.8826 22.78 1
ABC Classic (M7) 121414.1619 121474.2957 121558.5607 27.00 0
C9 40-units, 10,500 MW (EmD problem, w = 0) MABC/D/Cat (M1) 176682.2646 176682.2647 176682.2648 3.82  105 100
MABC/D/Log (M2) 176682.2646 176682.2647 176682.2655 1.14  104 100
MABC/D/Rand (M3) 176682.2646 176682.2647 176682.2650 5.17  105 100
MABC/P/Cat (M4) 176682.2646 176682.2647 176682.2654 1.00  104 100
MABC/P/Log (M5) 176682.2646 176682.2647 176682.2673 2.70  104 100
MABC/P/Rand (M6) 176682.2646 176682.2647 176682.2649 5.72  105 100
ABC Classic (M7) 176682.2646 176682.2723 176682.4185 2.21  102 99

The values in bold represents the best performing MABC variants in relation to items A and SD, for the cases (C1–C9).

Table 7
The results obtained by different optimization techniques for the case C1: 6 units, PD = 1263 MW).

Method MHSA [16] MSFL [24] IABC, IABC-LS [68] h-PSO [69] DHS [70]
Output
P1 (MW) 446.7273 445.0140 451.5204 447.1045 447.5285
P2 (MW) 173.4889 175.5156 172.1750 173.1123 173.2791
P3 (MW) 263.7642 264.2614 258.4186 263.6503 263.4772
P4 (MW) 138.8321 137.3012 140.6441 139.1516 139.0291
P5 (MW) 165.6496 162.7899 162.0797 165.9343 165.4864
P6 (MW) 86.9463 90.4992 90.3415 86.5037 87.1587
BM ($/h) 15442.52 15442.5911 15441.108 15443.1830 15449.8996
AM ($/h) 15443.594 15447.60 15441.108 15443.2117 15449.9264
WM ($/h) 15444.25 15460.29 15441.108 15443.3548 15449.9884
SDM ($/h) – 4.07 – 0.0291 0.0204
TOLM (MW)⁄ 0.5450142341 0.562145394 0.69359931601 0.49658658671 0.00003723189
⁄
TOLM was calculated for each method M with the relation (12), using the solutions indicated in the mentioned references [16,24,68–70].

Table 8
The best solution obtained by MABC/D/Cat algorithm considering the same tolerance TOLM as the methods M shown in Table 7 (the case C1: 6 units, 100 trials).

Algorithms MABC/D/Cat MABC/D/Cat MABC/D/Cat MABC/D/Cat MABC/D/Cat

Output for TOLMHSA [16] for TOLMSFL [24] for TOLIABC, IABC-LS [68] for TOLh-PSO [69] for TOLDHS [70]
P1 (MW) 447.3913243019 447.3871680920 447.3599709396 447.4005436386 447.5037628412
P2 (MW) 173.2346721727 173.2321144459 173.2120654879 173.2417784405 173.3179818758
P3 (MW) 263.3751759646 263.3724113955 263.3512011770 263.3829635426 263.4628031149
P4 (MW) 138.9724439133 138.9692633733 138.9470009752 138.9809886902 139.0650293142
P5 (MW) 165.3848697672 165.3822824272 165.3608248035 165.3934157108 165.4732723149
P6 (MW) 87.04459336632 87.0423939604 87.0206659678 87.05271960856 87.1353562784
BMABC/D/Cat ($/h) 15442.51985534 15442.28790885 15440.50812541 15443.17554399 15449.89902070
AMABC/D/Cat ($/h) 15442.51985542 15442.28790891 15440.50812547 15443.17554406 15449.89902075
WMABC/D/Cat ($/h) 15442.51985622 15442.28790949 15440.50812608 15443.17554540 15449.89902164
SDMABC/D/Cat ($/h) 1.6152  107 9.7913  108 8.2998  108 1.6725  107 9.9131  108
PL (MW) 12.94809372002 12.9477790883 12.9453286670 12.94899621794 12.9582429712
TOLMABC/D/Cat (MW) 0.54501423400 0.56214539394 0.693599316010 0.49658658668 0.00003723182

and C3 (2520 MW without PL). Among these the following are as the value of items A and/or SD. For this reason only a qualitative
mentioned: PSO varieties and hybrids (PSO [27,72], CLPSO [80], analysis is performed, based on items B, A, W and SD.
FAPSO (Fuzzy Adaptive PSO) [72], FAPSO-NM (FAPSO-with According with items (B, A, W and SD) it is noticed that the
Nelder–Mead) [72], PSO-SQP [27], DSPSO-TSA [74], CPSO (Chaotic behavior of the MABC variants (Table 6, the case C2: 1800 MW
PSO), CPSO-SQP [31] etc); Harmony Search varieties and hybrids without PL) is better than most of the algorithms shown in
(HS, HHS(Hybrid HS) [6]; evolutionary algorithms (IGA_MU Table 10, and is just as good as FAMPSO [7] (algorithms FAPSO-
(Improved Genetic Algorithm with Multiplier Updating) [9], GA VDE [29] and MsEBBO [76] obtain high quality solutions (by item
[74], EP-SQP [27], HGA (Hybrid GA) [26]; Ant Swarm Optimization B), but do not respect the imposed constrains). Similarly for the
varieties (ACO [10], CASO – Chaotic Ant Swarm Optimization [73], case C3 (2520 MW without PL), where the behavior of MABC
FCASO-SQP – Fuzzy adaptive CASO [73]); ABC algorithm [11]; Fire- variants (Table 6, the case C3) is better than most of the algorithms
fly algorithm (FA) [14]; Biogeography-Based Optimization (BBO) shown in Table 11, and is just as good as DE [5], FAMPSO [7] or HHS
[76,78]. [6]. Needed to be mentioned is that the DE, FAMPSO or HHS
A statistical comparison between MABC variants and the most algorithms either have a standard deviation (SD) higher than
competitive algorithms presented in Table 10 or Table 11 is not 5.77  107 $/h (obtained by MABC/D/Log), or the SD value was
possible to be conducted because it lacks certain information, such not specified by the authors.
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 53

Table 9
The best solutions obtained through MABC variants for the cases C2–C5 with 13 units (100 trials).

Algorithm MABC/D/Cat MABC/D/Log MABC/P/Log MABC/D/Cat

Output (C2:1800 MW, without PL) (C3:2520 MW, without PL) (C4:1800 MW, with PL) (C5:2520 MW, with PL)
P1 (MW) 628.3185307138 628.3185307163 448.7989505145 628.3185307131
P2 (MW) 149.5996501669 299.1993002422 299.1993059732 299.1993003418
P3 (MW) 222.7490698600 299.1992999811 224.3994800693 299.1993003372
P4 (MW) 109.8665500538 159.7331000954 109.8665500881 159.7331001093
P5 (MW) 109.8665500567 159.7331001135 109.8665501037 159.7331001144
P6 (MW) 109.8665500401 159.7331001020 109.8665502275 159.7331001111
P7 (MW) 109.8665500472 159.7331000836 109.8665500941 159.7331001030
P8 (MW) 109.8665500487 159.7331000765 109.8665500558 159.7331001114
P9 (MW) 60.0000000000 159.7331001088 109.8665500980 159.7331001144
P10 (MW) 40.0000000000 77.3999122700 40.0000000000 77.3999124208
P11 (MW) 40.0000000000 77.3999120533 40.5518593925 113.1111504227
P12 (MW) 55.0000000000 87.6845308618 55.0000000013 92.3999125176
P13 (MW) 55.0000000000 92.3999123012 55.0000000000 92.3999124897
PL (MW) – – 22.14889661794 40.426619906476
SD ($/h) 2.26  104 5.77  107 2.95 3.50  107
TOLM (MW) 9.87  107 9.94  107 6.02  1011 2.91  1011
BM ($/h) 17963.829200 24169.917696 18127.782085 24514.875629

Table 10
Results of the comparison between MABC/D/Cat and other published algorithms (the
maintained another nine sequences were subsequently added,
case C2 with 13-units, PD = 1800 MW without PL).
each consisting in 100 trials. The results obtained with MABC/D/
Items B ($/h) A ($/h) W ($/h) SD ($/h) Log (for the case C3) and MABC/D/Log (for the case C3) are shown
Algorithms
in Fig. 2 respectively Fig. 3. A very good stability of the solutions
DPSO [30] 17976.31 18084.99 18310.43 – (especially for C3 case, Fig. 3) can be noticed.
HMAPSO [30] 17969.31 17969.31 17969.31 0.00
Table 12 shows a comparison of results obtained by MABC/P/
NDS [71] 17976.9512 17976.9512 17976.9512 0.0000
SA [71] – 18299.2550 – 123.8335 Log, MABC/D/Log and other algorithms (STHDE [25], SDE [32],
PSO_TVAC [18] 17963.879 18154.562 18358.310 – BBO, DE/BBO, ORCCRO [83]), for C4 and C5 cases (1800 MW and
NEW PSO [18] 18120.594 18227.052 18427.631 – 2520 MW with PL). For PD = 2520 MW, the MABC/D/Cat algorithm
IFEP [60] 17994.07 18127.06 18267.42 – is more stable than the algorithms previously presented (SD item
DE [5] 17963.83 17965.48 17975.36 –
PSO-SQP [27] 17969.93 18029.99 – –
has a very small value). It must be mentioned that the value of item
EP-SQP [27] 17991.03 18106.93 – – BORCCRO = 24513.91 $/h, from [83], is obtained by working with a
PSO [27] 18030.72 18205.78 – – higher tolerance (|TOLORCCRO| = 0.98 MW > |TOLMABC/D/Cat| = 2.91
ED-DE [81] 17963.86 17972.70 17975.89 –  1011 MW). TOLORCCRO was calculated based on the solution pre-
ST-HDE [25] 17963.89 18046.38 – –
sented in [83]. In the case of PD = 1800 MW, it can be noticed in
HDE [25] 17975.73 18134.80 – –
FAPSO [72] 17963.84 17969.9187 17976.35 – Table 12 that MABC/P/Log algorithm obtains better values for B
FAPSO-NM [72] 17963.84 17963.9577 17964.21 – and A items, comparing to SDE algorithm [32] and therefore better
PSO [72] 18030.72 18205.9247 18401.35 – quality solutions. The best solutions for cases C4 and C5 are shown
FAMPSO [7] 17963.82 17963.83041 17963.835 – in Table 9. The average computation time is 38.2 s. (for the case C2),
FAPSO-VDE 17963.82 17963.82484 17963.832 –
32.7 s. (for the case C3), 4.07 min. (for the case C4) and 1.96 min (for
[29]
HGA [26] 17963.83 17988.04 – – the case C5).
CASO [73] 17965.15 18022.04 – –
FCASO-SQP [73] 17964.08 18001.96 – –
GSA [13] 17963.84 18041.21 18910.31 –
7.3. Test system 3: 52-unit without power losses (the cases C6 and C7)
CLPSO [80] 17970.67 18019.41 18071.57 22.67055
SQP-CLPSO [80] 17973.12 18005.05 18069.35 23.81023
IPSO [82] 17998.44 18176.95 – – This case presents a system consisting of 52 units, resulting by
HHS [33] 17963.8293 17972.4822 – 2.4185 multiplying by four the 13-units system described in paragraph 7.2
DEC-SQP [79] 17963.9401 17973.1339 17984.8105 1.9735 (test system 2). The characteristics of this system (cost coefficients,
BFO [12] 17974.48 17997.12 18018.75 –
GA-PS-SQP [28] 17964.25 – 18,199 –
power operating limits) are the same as for the system with
TSARGA [17] 17963.94 17974.31 18264.23 3.18 13-units [60]. Two values are considered for the load demand:
ABC [11] 17963.86 17987.22 17995.11 – PD = 7200 MW, respectively PD = 10,080 MW. The transmission
aBBOmDE [78] 17963.8521 17967.3560 17975.0552 – losses are neglected in this case.
IHSWM [21] 17963.83 17976.475 18041.345 25.6079
NPSO [75] 17976.015 – – –
MsEBBO/sin 17963.8292 17967.0705 17972.8105 4.1876
[76] 7.3.1. Solution’s quality and robustness
MsEBBO [76] 17963.8292 17964.0468 17969.0323 1.9215 The values of items B, A, W, SD and SR obtained through MABC
h-PSO [69] 17963.8297 17965.2055 17980.2030 4.3807 variants for cases C6 and C7 (52-units with 7200 MW and
FA [14] 17963.83 18029.16 18168.80 148.542
SDE [32] 17963.83 – – –
10,080 MW) are shown in Table 6. In relation to item A (from
MABC/D/Cat 17963.82920 17963.82933 17963.83045 2.26  104 Table 6) the best performing variants are MABC/D/Log (for the case
C6) and MABC/D/Cat (for the case C7). Also, item B value is the
approximately the same for all MABC variants (for both C6 and
For C2 and C3 cases (1800 MW and 2520 MW without PL) the C7). Following the values for the B, A, W and SD items, a good sta-
stability of solutions obtained by MABC variants was tested using bility of the obtained solutions using MABC/D/Log, respectively
a sequence of 100 trials (Table 6). To check if the stability is MABC/D/Cat, is noticed.
54 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

Table 11
Results of the comparison between MABC/D/Log and other published algorithms (the case C3 with 13-units, PD = 2520 MW without PL).

Items B ($/h) A ($/h) W ($/h) SD ($/h)

Algorithms
IGA_MU [9] 24169.9790 24385.4113 24754.1450 –
DE [5] 24169.9177 – 24169.9180 4.45  105
FAMPSO [7] 24169.9176 24169.9176 24169.9176 –
FAPSO-NM [72] 24169.92 24170.0017 24170.5 –
FAPSO [72] 24170.93 24173.0069 24176.4 –
PSO [72] 24262.73 24271.9231 24277.81 –
HHS [6] 24169.9 24169.9 24169.9 –
HS [6] 24208.7 24323.2 24503.7 –
ACO [10] 24174.39 24211.09 24243.90 21.10
PSO-SQP [27] 24261.05 – – –
EP-SQP [27] 24266.44 – – –
HGA [26] 24169.92 – – –
DSPSO-TSA [74] 24169.923 24173.137 24230.803 7.72
TSA [74] 24171.211 24184.055 24392.203 41
GA [74] 24170.804 24188.394 24567.974 59.53
PSO [74] 24170.167 24184.849 24377.890 38.86
SA-PSO [34] 24171.395 – – –
CPSO [31] 24211.56 – – –
CPSO-SQP [31] 24190.97 – – –
CASO [73] 24212.93 – – –
FCASO-SQP [73] 24190.63 – – –
SDE [32] 24169.92 – – –
SDE [77] 24169.9176968035 24170.7459 24171.4402 –
MABC [22] 24208.8330 – – –
MABC/D/Log 24169.917696 24169.917698 24169.917700 5.77  107

‘‘–’’ Data not available.

The parameters specific to each algorithm were obtained by

17964.05 performing several experimental tests. Parameters values which
led to the best value for the item A are given in Table 13. In Table 14
17964.00 we present the items B, W, A and SD obtained through the algo-
rithms: PSO, HS, DE, ABC, MABC/D/Log and MABC/D/Cat, using a
Cost F ($/h)

17963.95 sequence of 100 trials.

The results of applying the Mann–Whitney U statistic test
17963.90 (MW-U) for comparing MABC/D/Log algorithm for the case C6 (or
MABC/D/Cat for the case C7) and competitor algorithms (PSO, HS,
17963.85
DE and ABC) are shown in Table 14. For both cases (C6 and C7)
No. of trials the test MW-U shows that MABC variants are superior to the
17963.80
0 100 200 300 400 500 600 700 800 900 1000
competitor algorithms (p-values are less than the significance level
of 1%).
Fig. 2. The best total cost F obtained with MABC/D/Cat, in the case C2 (13-units Details regarding the MW-U test are presented in Section 7.5.1.
1800 MW without PL). The best solutions obtained through the MABC/D/Log and MABC/D/
Cat algorithms are presented in Table 15. Also, from Table 14 it can
be noticed that algorithms HS, PSO, ABC and DE cannot find a solu-
24169.917703
tion as good as that obtained by MABC variants (BMABC < BM, where
M = {HS, PSO, ABC, DE}). The average computation time is 5.47 min.
24169.917701 (for the case C6) and 3.51 min (for the case C7).
Cost F ($/h)

24169.917699 7.4. Test system 4: 40-unit without power losses (the cases C8 and C9)

24169.917697 In this case a large practical system with 40 thermal units is

considered to solving the EcD problem combined with the EmD
No. of trials problem using the MABC variants. The fuel cost coefficients (a, b,
24169.917695
0 200 400 600 800 1000 c, e, f) for the system’s units are provided by [20,60,85]. The emis-
sions coefficients (a, b, c, f, k) for the system’s units are provided by
Fig. 3. The best total cost F obtained with MABC/D/Log, in the case C3 (13-units [20,23,85,86,88]. The transmission losses are neglected in this case.
2520 MW without PL).
The load demand is PD = 10,500 MW.
For this system three situations are analyzed: (i) minimizing
only the fuel cost (w = 1); (ii) minimizing only the emissions
7.3.2. Comparing MABC variants to other methods (w = 0); (iii) minimizing both the fuel cost and the emissions (w
The results obtained through the best performing MABC vari- varies between 0 and 1 using a step of 0.1).
ants (MABC/D/Log and MABC/D/Cat) are compared with the results
obtained through other well known algorithms (PSO, HS, DE and 7.4.1. Solution’s quality and robustness
ABC), whose performance has been proven for various mathemat- Table 6 (the cases C8 and C9) shows the results obtained using
ical functions or technical systems. the MABC variants (M1–M6) and ABC Classic algorithm (M7) for
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 55

Table 12
Results of the comparison between MABC/D/Cat and other published algorithms (the cases C4 and C5, 13-units with PL).

Items PD (MW) B ($/h) A ($/h) W ($/h) SD ($/h) PL (MW)

Algorithms
SDE [32] 2520 24514.88 24516.31 – – 40.43
STHDE [25] 2520 24560.08 – – – –
ORCCRO [83] 2520 24513.91 24513.91 24513.91 – 39.43
BBO [83] 2520 24515.21 24515.32 24516.09 – –
DE/BBO [83] 2520 24514.97 24515.05 24515.98 – –
MABC/D/Cat 2520 24514.8756 24514.8756 24514.8756 3.50  107 40.426
SDE [32] 1800 18134.49 18138.56 – – 19.13
MABC/P/Log 1800 18127.7821 18129.7010 18134.3131 2.95 22.14

Table 13
The setting of parameters specific to algorithms ABC, PSO, DE and HS.

Algorithm ABC PSO DE HS

Parameters
PD (MW) 7200 10,080 7200 10,080 7200 10,080 7200 10,080
kmax 25,000 20,000 5000 20,000 10,000 3000 25,000 20,000
NP/Limit 20/0.02 15/0.02 100/– 15/– 50/– 100/– 20/– 15/–

For PSO: c1 = 1.5, c2 = 2, wmin = 0.3, wmax = 0.9 (for the cases C6 and C7).
For DE: CR = 0.7, AF = 0.1 (for the case C6) and CR = 0.2, AF = 0.1 (for the case C7).
For HS: HMCR = 0.99, PAR = 0.1, BW = 0.01 (for the case C6) and HMCR = 0.98, PAR = 0.8, BW = 0.01 (for the case C7).

Table 14
Results of comparison between MABC/D/Log (for C6) and MABC/D/Cat (for C7) with other well-known algorithms HS, PSO, ABC and DE (the system with 52- units, 100 trials).

Systems Algorithms ABC PSO DE HS MABC/D/Log (for C6)

Items MABC/D/Cat (for C7)
52-units 7200 MW (C6) B ($/h) 71785.6965 71990.9487 71807.6321 71806.4399 71770.2823
A ($/h) 71825.8070 72316.8523 71889.5475 71848.1929 71788.5302
W ($/h) 71886.9981 72734.9164 71997.5088 71929.3871 71808.8760
SD ($/h) 18.61 162.2506 37.3469 28.0473 10.5221
p-value 0.000 0.000 0.000 0.000 –
52-units 10,080 MW (C7) B ($/h) 96601.8531 96851.7648 96609.7373 96686.8987 96601.8504
A ($/h) 96632.1958 97350.5705 96619.0013 96780.1785 96604.5720
W ($/h) 96706.5650 97807.2548 96633.6183 96918.8149 96629.7631
SD ($/h) 20.19 192.4468 5.2046 74.5547 5.8908
p-value 0.000 0.000 0.000 0.000 –

Table 15
Best solutions obtained with MABC/D/Log and MABC/D/Cat (the system with 52-units).

Algorithm MABC/D/Log, PD = 7200 MW (for the case C6) MABC/D/Cat, PD = 10,080 MW (for the case C7)
Output (MW)
P1/P14/P27/P40 448.798950 538.558741 448.798951 448.798950 628.318531 628.318531 628.318531 628.318531
P2/P15/P28/P41 299.143951 297.746123 224.390149 299.180044 299.199296 299.199299 299.199298 299.199300
P3/P16/P29/P42 224.379100 224.383194 224.379956 224.378884 299.199300 299.199292 299.199299 299.199300
P4/P17/P30/P43 109.866550 109.866480 109.866548 109.866550 159.733100 159.733100 159.733100 159.733100
P5/P18/P31/P44 109.866550 109.866527 60.000000 109.866549 159.733100 159.733100 159.733100 159.733100
P6/P19/P32/P45 109.866550 109.866550 109.866532 109.866550 109.866550 159.733100 159.733100 159.733100
P7/P20/P33/P46 109.865942 109.866543 109.866550 109.866550 159.733100 159.733100 159.733100 159.733100
P8/P21/P34/P47 109.866540 109.866203 60.000000 109.866550 159.733100 159.733100 159.733100 159.733100
P9/P22/P35/P48 109.866547 109.866546 109.866550 109.866549 159.733100 159.733100 159.733100 159.733100
P10/P23/P36/P49 40.000000 40.000000 40.000000 40.000000 77.399909 77.399911 77.399896 77.399905
P11/P24/P37/P50 40.000000 40.000000 40.000000 40.000000 77.399912 77.399912 108.405810 77.399910
P12/P25/P38/P51 55.000000 55.000000 55.000000 55.000000 92.399823 92.399869 92.399775 92.399781
P13/P26/P39/P52 55.000000 55.000000 55.000000 55.000000 92.399623 92.399859 92.399888 92.399859
BM ($/h) 71770.2823 96601.8504
TOLM (MW) 2.01  1011 2.00  1011

two situations: considering only the total fuel cost (case C8), close results for the items B, A, W and SD. In C9 case all MABC vari-
respectively considering only the total emissions (case C9). In rela- ants achieve very close results and the success rate is SR = 100%
tion to items A and SD, the best variants are MABC/P/Long and (considering dF = 1000 $/year = 0.1141 $/h).
MABC/D/Cat (for case C8), respectively MABC/D/Cat and MABC/D/ The best solutions reached by applying the high performing
Rand (for case C9). In C8 case the MABC variants achieve relatively variants – MABC/P/Log and MABC/D/Cat for EcD problem (w = 1,
56 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

case C8) and MABC/D/Cat for EmD problem (w = 0, case C9) – are both Fuel cost, as also for the Emission quantity. For Fuel cost
shown in Table 16. Also, in Table 16, the best compromise solutions the comparison is done based on the B, A, W, SD items, and for
(BCS) for the MABC/D/Cat and MABC/P/Log variants are shown. The Emission based on Best Emission item. It must be specified that
BCS were selected from the Pareto-front solutions using two rank- the value of the TOLM item was calculated considering the solutions
ing fuzzy-based mechanism methods described in Section 2.3. The obtained with the algorithms mentioned in Table 18.
Pareto-front solutions were calculated using relation (14), where The Best Emission achieved with the MABC variants (for exam-
the value of the weight factor w was increased from 0 to 1 with ple, 176682.26460 ton/h for MABC/P/Log and 176682.264590 ton/
a step of 0.1. In order to have a Pareto-front as complete as possible h for MABC/D/Cat) have a better performance than MBFA [20],
in the promising zone found in the range w 2 (0.6, 0.7), two addi- MODE, NSGA-II [87]. Making the comparison with the results
tional values were considered for the weight factor (w = 0.64 and obtained with other algorithms/methods (like HBMO and IHBMO
w = 0.68). For the w values within the range (0.6–0.7) the factor [85], CIABC [86], DE [88], IABC [23] etc) raises a series of issues
l⁄ has the highest values in the range (0.0828, 0.0856), as can be described below:
noticed in Table 17. Thus, the 13 points resulted in the objectives
space (Fuel cost, Emission) are shown in Table 17. From Table 17 (i) the HBMO and IHBMO [85] algorithms do not properly cal-
the BCS is obtained for a w = 0.68 (l⁄w=0.68 = 0.0872) with a total culate the Best emission value. Performing a recalculation,
fuel cost of 124491.1616 $/h and an emissions value of the correct values for the emissions are: 281016.0597 ton/
256551.5725 ton/h (with MABC/P/Log algorithm). Also, it can be h instead of 161305.061 ton/h (for HBMO [85]),
noticed in Table 17 that both methods applied in order to rank 280133.0596 ton/h instead of 160807.648 ton/h (for IHBMO
the solutions indicate the same BCS. [85]). These values are bigger that the values obtained with
MABC variants (item B).
7.4.2. Comparing MABC variants to other methods (ii) in case of the CIABC [86] algorithm the constraint regarding
In Table 18 the results obtained with the highest performance the maximum power operating limit for unit 30 is broken
MABC variants (MABC/P/Log and MABC/D/Cat) are compared with (P30 [86] = 99 MW > P30,max = 97 MW – see the solution from
those obtained by other optimizing algorithms/techniques, for Table 16 for CIABC [86]). If MABC/P/Log algorithm is

Table 16
The solutions corresponding to the Best cost, Best emission and BCS when the MABC/D/Cat, MABC/P/Log and CIABC [86] algorithms are applied (40-units).

Algorithm Best cost Best emission Best compromise solution

Output MABC/D/Cat MABC/P/Log CIABC [86] MABC/D/Cat CIABC [86] MABC/D/Cat MABC/P/Log
P1 (MW) 110.8000081 110.7998263 110.7704 114.0000000 113.9998 110.7998216 110.799823
P2 (MW) 110.7998478 110.8029523 110.7705 114.0000000 114.0000 110.7998237 110.7998243
P3 (MW) 97.3999212 97.3999110 97.3707 120.0000000 120.0000 97.39991059 97.39991167
P4 (MW) 179.7330987 179.7330900 179.7038 169.3680972 171.0044 174.5504493 174.5486362
P5 (MW) 87.7999294 87.7999044 87.7705 97.0000000 96.9995 87.79991926 97
P6 (MW) 139.9999999 139.9999990 140 124.2571469 127.0000 105.3999112 105.3999124
P7 (MW) 259.5996696 259.5996492 259.5699 299.7111594 298.2855 259.5996498 259.5996487
P8 (MW) 284.5996495 284.5996495 284.5704 297.9148326 298.5914 284.5996451 284.5996462
P9 (MW) 284.5996507 284.5996498 284.5706 297.2603479 296.9671 284.5996474 284.5996449
P10 (MW) 130.0000000 130.0000000 130 130.0000000 134.0000 130 130
P11 (MW) 94.0000000 94.0000000 94 298.4102742 297.8076 318.1921565 318.2129706
P12 (MW) 94.0000000 94.0000000 94 298.0262812 296.1682 243.5996499 243.5996501
P13 (MW) 214.7597894 214.7597902 214.7298 433.5562147 434.7314 394.2793698 394.2793677
P14 (MW) 394.2793664 394.2793479 394.2497 421.7283602 421.0079 394.2793686 394.2793686
P15 (MW) 394.2793704 394.2793149 394.2496 422.7800257 423.0154 394.2793694 394.2793696
P16 (MW) 394.2793703 394.2793506 394.2497 422.7802015 420.5107 394.2793694 394.2793702
P17 (MW) 489.2793697 489.2793695 489.2499 439.4119664 441.3810 399.5195798 399.5195794
P18 (MW) 489.2793694 489.2793695 489.2499 439.4030913 436.7767 399.5195802 399.5195805
P19 (MW) 511.2793697 511.2793691 511.2497 439.4132875 436.4718 506.1985408 506.1716929
P20 (MW) 511.2793693 511.2793693 511.2708 439.4134585 439.4165 506.1801105 506.2206802
P21 (MW) 523.2793702 523.2793703 523.2729 439.4466782 436.9569 514.1472578 514.1105547
P22 (MW) 523.2793702 523.2793702 523.2737 439.4468932 441.8603 514.1455791 514.1472783
P23 (MW) 523.2793702 523.2793703 523.2747 439.7724577 438.5185 514.5237739 514.5664453
P24 (MW) 523.2793702 523.2793702 523.2758 439.7716054 438.3634 514.5386386 514.4868056
P25 (MW) 523.2793702 523.2793703 523.2758 440.1117982 442.1713 433.5196115 433.5195803
P26 (MW) 523.2793702 523.2793702 523.2697 440.1112710 441.3550 433.5195807 433.5196074
P27 (MW) 10.0000000 10.0000000 9.9912 28.9933236 29.0661 10 10
P28 (MW) 10.0000000 10.0000000 9.9915 28.9936888 27.5502 10 10
P29 (MW) 10.0000000 10.0000000 9.9909 28.9938865 31.3780 10 10
P30 (MW) 87.8004645 87.8067375 87.7723 97.0000000 99.0000 97 87.80420249
P31 (MW) 189.9999996 190.0000000 189.9824 172.3318327 170.6259 159.7330999 159.7330995
P32 (MW) 190.0000000 189.9999959 190 172.3316979 173.7795 159.7330999 159.7331057
P33 (MW) 189.9999985 189.9999995 190 172.3318840 173.2124 159.7330994 159.7330992
P34 (MW) 164.7999830 164.7998247 164.7731 200.0000000 200.0000 200 200
P35 (MW) 200.0000000 194.3884987 194.3642 200.0000000 200.0000 200 200
P36 (MW) 194.3968155 199.9999982 200 200.0000000 200.0000 200 200
P37 (MW) 110.0000000 109.9994763 110 100.8384255 100.0000 89.11413552 89.11413643
P38 (MW) 109.9999983 109.9999695 110 100.8384327 100.0000 89.11413645 89.11413656
P39 (MW) 109.9999999 109.9999948 109.938 100.8386375 100.0000 89.1141381 89.11413522
P40 (MW) 511.2793691 511.2793702 511.3501 439.4127410 438.0276 506.1879753 506.1951354
Cost ($/h) 121412.540947 121412.591816 121412.4 129995.274021 129996.6 124490.903193 124491.161616
Emission (ton/h) 359901.397571 359901.503351 359842.0 176682.264590 176633.0 256560.266940 256551.572517
TOL (MW) 9.01  107 7.18  107 0.607800 8.97  107 1.81  1012 9.80  107 7.28  107
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 57

Table 17 (iv) the DE [88] algorithm obtains for Best emission a value of
The fuel cost and the emission obtained by the MABC/P/Log algorithm, for different 176,680 ton/h, but the exact value is 176683.2612 ton/
values of the factor w (40-units).
h > 176682.2646 ton/h (obtained by MABC/P/Log). There-
w MABC/P/Log fore, the MABC variants provide higher quality solutions
Fuel cost ($/h) Emission (ton/h) l⁄ (method 1) l1 (method 2) comparing to DE [88].
0.00 129995.2691 176682.2646 0.0724 0.0000
0.10 129516.6708 203839.7691 0.0657 0.0558 Analyzing the results from Table 18, related to the ‘‘Fuel cost’’,
0.20 129095.4231 204937.2152 0.0688 0.1048 the following was found:
0.30 128814.0623 205947.6104 0.0708 0.1376
0.40 128401.8957 208129.6635 0.0734 0.1856
(i) the MABC/P/Log and MABC/D/Cat variants have better perfor-
0.50 127711.2165 213534.9781 0.0771 0.2661
0.60 126445.8876 226033.5918 0.0828 0.4136 mance than all other algorithms mentioned in Table 18 regard-
0.64 125435.8396 238936.8995 0.0862 0.5312 ing B, A, W, SD items (exception makes the IABC-LS algorithm
0.68 124491.1616 256551.5725 0.0872 (max) 0.5641 (max) [50] which has better values in regards to item W). It must be
0.70 124042.7577 270197.0390 0.0856 0.4896
specified that the CIABC algorithm [86] achieves the best cost
0.80 123054.0402 306960.4929 0.0794 0.2889
0.90 122047.9086 331426.7288 0.0783 0.1554
given that it breaks two types of constraints: minimum power
1.00 121412.5918 359901.5033 0.0724 0.0000 operating limits for units 27, 28 and 29 (P27 = 9.9912 MW
< P27,min = 10 MW, P28 = 9.9915 MW < P28,min = 10 MW and
The values in bold correspond to BCS.
P29 = 9.9909 MW < P29,min = 10 MW – see Table 16) and real
power balance constraint (TOLCIABC [86] = 0.6078 MW). If
an MABC variant, for example, MABC/P/Log algorithm is
compared to CIABC [86], enforcing the same limit for unit 30
applied considering the obtained tolerance by using the CIABC
(P30,max = 99 MW), then the best emission obtained with
algorithm [86] (TOLCIABC [86] = 0.6078 MW) and for units 27,
MABC/P/Log is 176504.613635 ton/h (value that is better
28 and 29 the minimum limits from above are imposed, then
than the one obtained with CIABC [86] of 176633.0 ton/h).
the Best cost obtained with MABC/P/Log is 121405.7417 $/h
(iii) IABC, IABC-LS, ABCD, ABCDP-LS [23] algorithms use a rela-
(value that is the better than the one obtained with CIABC
tively high tolerance (|TOL| > 0.0130 MW – see the TOL value
[86] of 121412.4 $/h);
from Table 18). If, for example, MABC/P/Log algorithm is
(ii) the MABC/P/Log and MABC/D/Cat algorithms have better
applied considering the obtained tolerance by applying the
performances than the well known and frequently used
IABC algorithm (TOLIABC [23] = 0.0130 MW), then the Best
algorithms in solving different mathematical and engineer-
emission is 176680.8888 ton/h (value which is better than
ing problems (like: MBFA [20], DE [88], EP [27], ABC
that obtained by IABC [23], so MABC/P/Log algorithm is
[11,86], GA [10], PSO [10,27,89], ACO [10], CSO [89], BBO
more efficient than the IABC algorithm [23]).

Table 18
The values of the items (B, A, W, SD, TOL) for the Fuel cost, Best emission and TOL for emission.

Items Fuel cost Emission

Algorithm B ($/h) A ($/h) W ($/h) SD ($/h) TOLM* (MW) Best emission TOLM* (MW)
MODE [87] 121836.9839 – – – 9.99  105 176683.2718 1.00  104
NSGA-II [87] 124963.5028 – – – 1.00  104 176691.9677 1.00  104
HBMO [85] 121416.03 122019.65 – – 2.50  103 161305.061 4.70  103
IHBMO [85] 121412.7533 121875.58 – – 0 160807.648 0
MBFA [20] 121415.653 – – – 2.00  104 176682.269 1.00  104
DE [88] 121,840 – – – 9.99  105 176,680 1.00  104
IABC [23] 121414.80 – – – 1.45  102 176682.25 1.30  102
IABC-LS [23] 121412.72 – – – 1.09  102 176682.25 1.99  102
ABCDP [23] 121412.82 – – – 1.80  102 176682.25 1.80  102
ABCDP-LS [23] 121412.74 – – – 1.19  102 176682.25 1.49  102
CIABC [86] 121412.4 121623.15 – – 0.6078 176633.0 1.81  1012
EP [27] 122624.35 123382.00 – – – – –
CPSO-SQP [31] 121458.54 122028.16 – – – – –
FCASO-SQP [73] 121456.98 122026.21 – – – – –
h-PSO [69] 121420.9027 121509.8423 121852.4249 – 3.63  1012
PSO [27] 123930.45 124154.49 – – – – –
PSO [89] 122588.5093 123544.8853 124733.6795 – 1.81  1012 – –
PSO [10] 121800.13 121899.57 122000.80 84.21 – – –
ACO [10] 121532.41 121606.45 121679.64 45.58 0.16  102 – –
GA [10] 121996.40 122919.77 123807.97 320.31 – – –
CSO [89] 121461.6707 121936.1926 122844.5391 – 2.99  104 – –
BBO [90] 121479.5029 121512.0576 121688.6634 – 0.277  102 – –
ABC [11] 121441.03 121995.82 122123.77 – 1.82  1012
HDE [25] 121698.51 122304.30 – – 7.00  105 – –
IABC [50] 121412.75 – 121503.58 – 1.006  104  –
IABC-LS [50] 121412.73 – 121471.61 – 9.99  105 – –
ABC [86] 121515.1 124827.34 – – – – –
ABCLogistic [86] 121440.2 123314.12 – – – – –
ABCTend [86] 121418.51 122831.22 – – – – –
MABC/P/Log 121412.591816 121431.576266 121493.188471 18.16 7.18  107 176682.264600 9.42  107
MABC/D/Cat 121412.540947 121431.779282 121503.755217 19.16 9.01  107 176682.264590 8.97  107
*
TOLM was calculated for each method M with the relation (12), using the solutions indicated in the mentioned references. ‘‘–‘‘ data not available.
58 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

[90]) or different variants of the ABC algorithm recently pub- generating very stable and high quality solutions – Table 6).
lished (IABC, IABC-LS, ABCDP, ABCDP-LS [23], CIABC [86], However, from a mathematical point of view, between the
IABC, IABC-LC [50], ABCLogistic, ABCTend [86]); MABC variants applied in the five cases (C1, C2, C3, C5 and
(iii) The tolerance (TOL) achieved for the solutions obtained with C9) statistically significant differences can exist. Therefore,
the MABC/P/Log and MABC/D/Cat algorithms are smaller these cases will not be taken into consideration in the rank-
than 106 MW. It must be mentioned that the tolerance ing process of the MABC variants, as not to modify their final
TOLM which is relatively high – (0.001, 0.02) MW range – standings.
for some methods from Table 18, can be partially explained (iii) in contrast, in four cases (C4, C6, C7 and C8), there are differ-
by the fact that the solutions presented in the corresponding ences between MABC variants for B, A, W, SD and SR items
references were rounded to two or three digits. The average (these differences may determine values greater than the
computation time is 1.92 min. (for the case C8) and 12.03 s. assumed limit of dF = 1000 $/year). Therefore, only these
(for the case C9). cases will be taken into consideration in the MABC variants
ranking.
7.5. Comparing the results
7.5.1. Statistical test
Analyzing the results in Table 6, obtained through the MABC The performance comparison for MABC variants, selection
variants for the nine studied cases (C1–C9), the following can be schemes (P and D) and modalities (Cat, Log and Random) of gener-
noticed: ating random sequences (GRS) is performed using Mann–Whitney
U statistical test with Bonferroni–Holm correction.
(i) considering the items A and SD the best performing MABC Mann–Whitney U statistical test is a non-parametric test based
variants are MABC/D/Cat (for the cases C1, C2, C5, C7 and on ranks, which shows if there are significant differences between
C9), MABC/D/Log (for the cases C3 and C6) and MABC/P/Log two independent samples. Bonferroni–Holm correction is applied
(for the case C4 and C8); when performing repeated comparisons between the same collec-
(ii) there are five cases (C1, C2, C3, C5 and C9) in which the MABC tions of data and is based on the adjustment of p-value obtained by
variants lead to have very close values for B, A, W and SR applying Mann–Whitney U test [84].
items. The differences between W and B items, for cases The null hypothesis (H0) for applying Mann–Whitney U test is:
C1, C2, C3, C5 and C9, are lower than the assumed cost of there is no significant difference between the results obtained by
dF = 1000 $/year. Therefore, in such cases, the economic two methods Mx and My; the alternate hypothesis (H1) is: there
effect obtained through a certain MABC variant, compared is a significant difference between the results of Mx and My meth-
to another MABC variant, is considered insignificant (this ods. If the adjusted p-value (using Bonferroni–Holm corrections) is
means that all MABC variants behave just as good, less than or equal to the level a, then the null hypothesis is rejected

Table 19
Results of comparison between MABC variants (M1–M6), for the cases (C4, C6, C7, C8), using Mann–Whitney U test.

Cases C4 C6 C7 C8 Indicator (RC4,C6,C7,C8)Mx

Algorithm
M1 vs. M2 Idem M2 Idem Idem (RC4,C6,C7,C8)M1 = 5 (Rank 3)
M3 Idem M1 (0.039) M1 (0.030) M1 (0.012)
M4 Idem Idem M1(0.000) Idem
M5 Idem M5 Idem Idem
M6 Idem Idem M1 (0.042) Idem
M2 vs. M1 Idem M2 (0.000) Idem Idem (RC4,C6,C7,C8)M2 = 6 (Rank 1.5)
M3 Idem M2 (0.000) Idem Idem
M4 Idem M2 (0.000) M2 (0.005) Idem
M5 Idem Idem Idem Idem
M6 M2 (0.029) M2 (0.000) Idem Idem
M3 vs. M1 Idem M1 M1 M1 (RC4,C6,C7,C8)M3 = 0 (Rank 6)
M2 Idem M2 Idem Idem
M4 Idem Idem Idem Idem
M5 Idem M5 Idem Idem
M6 Idem Idem Idem Idem
M4 vs. M1 Idem Idem M1 Idem (RC4,C6,C7,C8)M4 = 1 (Rank 4.5)
M2 Idem M2 M2 Idem
M3 Idem Idem Idem Idem
M5 Idem M5 M5 Idem
M6 M4 (0.001) Idem M6 Idem
M5 vs. M1 Idem M5(0.000) Idem Idem (RC4,C6,C7,C8)M5 = 6 (Rank 1.5)
M2 Idem Idem Idem Idem
M3 Idem M5 (0.000) Idem Idem
M4 Idem M5 (0.000) M5 (0.008) Idem
M6 M5 (0.005) M5 (0.000) Idem Idem
M6 vs. M1 Idem Idem M1 Idem (RC4,C6,C7,C8)M6 = 1 (Rank 4.5)
M2 M2 M2 Idem Idem
M3 Idem Idem Idem Idem
M4 M4 Idem M6 (0.037) Idem
M5 M5 M5 Idem Idem

The p-value resulting from Mann–Whitney U test and Bonferroni–Holm correction is presented between round brackets (); Items marked ‘‘Idem’’ indicate that between two
compared methods (Mx, My) there is no statistically significant difference; (RC4,C6,C7,C8)Mx represents the number of cases in which the method Mx is superior to other
methods, considering the cases which produce more important economic effects (C4, C6, C7 and C8).
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 59

Table 20 and Bonferroni–Holm correction. For each method Mx the number

Results of comparison between MABC variants (M1–M6) and ABC algorithm (M7), for of times when the method was superior to other methods was
each case (C1–C9), using Mann–Whitney U test.
added up, leading to item (RC4,C6,C7,C8)Mx to be calculated. Then a
Cases C1, C2, C4 C9 C3 (RC1–C9)Mx hierarchy of all MABC variants was set. Analyzing the behavior of
Algorithm the MABC variants (M1–M6) for the analyzed cases (C4, C6, C7, C8),
M1 vs. M7 M1 (0.000) M1 (0.048) (RC1–C9)M1 = 9 the following can be seen from Table 19: (i) based on RC4,C6,C7,C8
M2 vs. M7 M2 (0.000) M2 (0.048) (RC1–C9)M2 = 9 indicator, the MABC variants can be ranked as follows: M2 = M5 >
M3 vs. M7 M3 (0.000) M3 (0.026) (RC1–C9)M3 = 9
M4 vs. M7 M4 (0.000) M4 (0.036) (RC1–C9)M4 = 9
M1 > M4 = M6 > M3; (ii) M2 (MABC/P/Log) and M5 (MABC/P/Log)
M5 vs. M7 M5 (0.000) M5 (0.014) (RC1–C9)M5 = 9 have a better behavior than other variants, both on large systems
M6 vs. M7 M6 (0.000) M6 (0.045) (RC1–C9)M6 = 9 (52-units system) and small ones (13 units system); (iii) M3 variant
(MABC/D/Rand) has the worst behavior in the cases C4, C6, C7 and C8.

and the alternative hypothesis is accepted. The significance level 7.5.3. Comparison of the MABC variants and ABC algorithm
for the null hypothesis testing is a = 5%. In Table 20 can be noticed that all the MABC variants (M1–M6)
Mann–Whitney U test was performed using specialized are superior to ABC algorithm (M7) – both when each case was
software called SPSSÒ (Statistical Package for the Social Sciences). taken under consideration individually, and for all cases
To test the performance of methods, selection schemes or GRS (RC1–C9)Mx = 9 analyzed as a whole. Also MABC variants behave
modalities the data resulting following 100 simulations (the Cost better than ABC algorithm (M7) for both large systems (40-units
F) is grouped in three ways: (i) by cases; (ii) by both cases and and 52-units systems) and for smaller ones (6-units and 13-units
selection schemes; (iii) by both cases and GRS modalities grouping. systems).
The results from applying Mann–Whitney U test and Bonferron-
i–Holm correction are shown in Tables 19–22. When comparing 7.5.4. Comparison of the selection schemes performance
two methods ((Mx vs. My, x – y), in Tables 19–22 the better This paragraph refers to the results of the comparisons between
method was recorded. If between the compared methods there is the performance of the selection schemes (Disruptive(D)/Propor-
no significant difference, the table entry is marked with ‘‘Idem’’. tional(P)), with which the MABC variants were endowed. It is noted
Also, alongside Mx or My method, the p-value resulting from that when two selection schemes were compared, the GRS
Mann–Whitney U test with Bonferroni–Holm correction was modality remains the same (Cat, Log or Random). The results of
specified in round brackets. the comparisons (using Mann–Whitney U test) are shown in
Table 21, and it can be seen that the Disruptive scheme (D) is better
7.5.2. Comparison of the results obtained by MABC variants than the classical Proportional scheme (P), when map cat was used.
Table 19 shows the pairwise comparisons of the MABC variants In all the other situations there are no statistically significant dif-
(M1–M6), for the cases C4, C6, C7 and C8, using Mann–Whitney U test ferences between the compared variants.

Table 21
The results of comparing the selection schemes (Disruptive (D)/Proportional (P)), using Mann–Whitney U test, when the random sequence generating is maintained the same
(Cat, Logistics or Random).

Cases C4 C6 C7 C8 Indicator RC4,C6,C7,C8

Selection schemes (D/P) Map
Disruptive (D) vs. Proportional (P) Cat Idem Idem D (0.000) Idem (RC4,C6,C7,C8)D = 1 (Rank 1)
Disruptive (D) vs. Proportional (P) Log Idem Idem Idem Idem
Disruptive (D) vs. Proportional (P) Random Idem Idem Idem Idem
Proportional (P) vs. Disruptive (D) Cat Idem Idem D Idem (RC4,C6,C7,C8)P = 0 (Rank 2)
Proportional (P) vs. Disruptive (D) Log Idem Idem Idem Idem
Proportional (P) vs. Disruptive (D) Random Idem Idem Idem Idem

Items marked ‘‘Idem’’ indicate that between the compared schemes there is no statistically significant difference.

Table 22
The results of comparing the GRS modalities (Cat/Log/Rand), using Mann–Whitney U test, when the selection scheme is maintained the same (Disruptive (D) or Proportional (P)).

Cases C4 C6 C7 C8 Indicator RC4,C6,C7,C8

GRS modality Scheme (D/P)
Cat vs. Log Disruptive (D) Idem Log Idem Idem (RC4,C6,C7,C8)Cat = 4 (Rank 2)
Cat vs. Rand Disruptive (D) Idem Cat (0.037) Cat (0.015) Cat (0.012)
Cat vs. Log Proportional (P) Idem Log Log Idem
Cat vs. Rand Proportional (P) Cat (0.001) Idem Rand Idem
Log vs. Cat Disruptive (D) Idem Log(0.000) Idem Idem (RC4,C6,C7,C8)Log = 6 (Rank 1)
Log vs. Rand Disruptive (D) Idem Log(0.000) Idem Idem
Log vs. Cat Proportional (P) Idem Log(0.000) Log (0.004) Idem
Log vs. Rand Proportional (P) Log (0.005) Log(0.000) Idem Idem
Rand vs. Cat Disruptive (D) Idem Cat Cat Cat (RC4,C6,C7,C8)Rand = 1 (Rank 3)
Rand vs. Log Disruptive (D) Idem Log Idem Idem
Rand vs. Cat Proportional (P) Cat Idem Rand(0.037) Idem
Rand vs. Log Proportional (P) Log Log Idem Idem

Items marked ‘‘Idem’’ indicate that between the compared GRS modalities there is no statistically significant difference.
60 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

0.5 C1: 6-units, 1263 MW C2: 13-units, 1800 MW variants are superior to classical ABC algorithm for all the analyzed
C3: 13-units, 2520 MW C4: 13-units, 1800 MW, with PL cases.
0.4 C5: 13-units, 2520 MW, with PL C6: 52units, 4x1800 MW The best performing MABC variants – MABC/D/Cat (for cases C1,
C7: 52-units, 4x2520 MW C8:40-units, 10500 MW
C2, C5, C7 and C9), MABC/D/Log (for cases C3, C6) and MABC/P/Log
Cost F ($/h)

0.1
0.3 (for the case C4 and C8) – behave better than the vast majority of
0.05 optimization methods with which they were compared (methods
0.2 relatively recently reported in the scientific literature and also well
0
0 200 400 600 800 1000
known methods (PSO, HS, ABC, DE) applied by the author). The best
0.1 solutions obtained meet all of the equality and inequality con-
No. of iterations straints with high accuracy. The MABC variants have the ability
0 to offer high-quality solutions for both small systems (6-units
0 5000 10000 15000 20000 25000
and 13 units) and large systems (40-units and 52-units). Also,
Fig. 4. The convergence process for MABC/D/Cat algorithm, cases C1–C8. the solutions obtained have very good stability for the 6-units
and 13- units systems, and good stability for the 40-units and
52-units systems.
Following the MABC algorithm structure it is noticed that it
7.5.5. Comparison of the random sequences
retains the positive characteristics of the original ABC algorithm:
The comparison of the performance for GRS modalities (Cat, Log
simple in concept, easy to implement, allows the hybridization
and Random), with which the MABC variants were endowed, are
with other classical or artificial intelligence based algorithms, flex-
presented in Table 22. It needs to be mentioned that when two
ible, can include the chaos element either in the solution update
GRS modalities were compared, the selection scheme (D or P)
relations, or as a local search procedures, uses a moderate number
remains the same. Based on Table 22 the following can be seen:
of specific parameters. Compared to other metaheuristic algo-
(i) for all the cases C4, C6, C7 and C8, logistic map ((RC4,C6,C7,C8)Log = 6)
rithms (ABC, PSO, HS, DE, BBO, FA, GA etc) the MABC algorithm
is superior to the other two (cat map or Random). The hierarchy of
keeps the exploration–exploitation balance better, having the
the GRS modalities is Logistic > Cat > Random; (ii) logistic map has
capacity to obtain stable and high quality solutions.
a good behavior in combination with Proportional scheme
((RC4,C6,C7,C8)Log = 4 for the scheme Proportional); (iii) logistic and
cat maps have a close performance if Disruptive scheme is used;
(iv) ‘‘Random’’ has poor performance for both D scheme and P References
scheme ((RC4,C6,C7,C8)Rand = 1).
[1] Nanda J, Hari L, Kothari ML. Economic emission load dispatch with line flow
constraints using a classical technique. IEE Proc Gener Trans Distrib
1994;141(1):1–10.
7.6. Convergence process [2] Papageorgiou L, Fraga E. A mixed integer quadratic programming formulation
for the economic dispatch of generators with prohibited operating zones.
Fig. 4 shows the Cost F variation throughout the optimizing pro- Electr Power Syst Res 2007;77(10):1292–6.
[3] Chen SD, Chen JF. A direct Newton–Raphson economic emission dispatch. Int J
cess by MABC/D/Cat algorithm for the eight cases (C1–C8). It is Electr Power Energy Syst 2003;25(5):411–7.
noticed that convergence is good in all the situations. For a single [4] Gaing Z-L. Particle swarm optimization to solving the economic dispatch
graphical representation of Cost F values, these are normalized considering the generator constraints. IEEE Trans Power Syst
2003;18(3):1187–95.
within the [0, 1] range for each case. The value zero (0) corresponds [5] Noman N, Iba H. Differential evolution for economic load dispatch problems.
to the best value of the Cost F (item B), and the value one (1) cor- Electr Power Syst Res 2008;78(3):1322–31.
responds to Cost F maximum (item W). The number of iterations [6] Fesanghary M, Ardehali MM. A novel meta-heuristic optimization
methodology for solving various types of economic dispatch problem.
for the analyzed systems is different. Therefore, to get a clearer
Energy 2009;34(6):757–66.
picture on the convergence, the optimization process for the four [7] Niknam T, Mojarrad HD, Nayeripour M. A new fuzzy adaptive particle swarm
cases (C1: 6-units, C2 and C3: 13-units with 1800 MW and optimization for non-smooth economic dispatch. Energy 2010;35(4):1764–78.
2520 MW, C4: 13-units, 1800 MW with PL and C8: 40-units, [8] Liang Y-C, Juarez JRC. A normalization method for solving the combined
economic and emission dispatch problem with meta-heuristic algorithms. Int J
10,500 MW) is detailed in Fig. 4. Electr Power Energy Syst 2014;54(1):163–86.
[9] Chiang C-L. Genetic-based algorithm for power economic load dispatch. IET
Gener Trans Distrib 2007;1(2):261–9.
8. Conclusions [10] Pothiya S, Ngamroo I, Kongprawechnon W. Ant colony optimisation for
economic dispatch problem with non-smooth cost functions. Int J Electr Power
Energy Syst 2010;32(5):478–87.
This paper proposes a modified ABC algorithm (MABC), which is [11] Hemamalini S, Simon SP. Artificial bee colony algorithm for economic load
successfully applied for the economic dispatch problem solving, dispatch problem with non-smooth cost functions. Electr Power Compo Syst
2010;38(7):786–803.
taking into account inequality and equality constraints, the [12] Vijay R. Intelligent bacterial foraging optimization technique to economic load
valve-point effects, the emission pollutions and the transmission dispatch problem. Int J Soft Comput Eng 2012;2(2):55–9.
line loses. The MABC algorithm proposes a new relation for the [13] Swain RK, Sahu NC, Hota PK. Gravitational search algorithm for optimal
economic dispatch. Proc Technol 2012;6:411–9.
solutions update, applied both in employed bee phase, and in [14] Yang X-S, Hosseini SSS, Gandomi AH. Firefly algorithm for solving non-convex
onlooker bees phase. Four test systems having 6, 13, 40 and 52 economic dispatch problems with valve loading effect. Appl Soft Comput
units were analyzed. The results indicate that the proposed 2012;12(3):1180–6.
[15] Bijami E, Jadidoleslam M, Ebrahimi A, Askari J, Farsangi MM. Implementation
relation is effective, especially for 13 units, 40 units and 52 units of imperialist competitive algorithm to solve non-convex economic dispatch
systems. The behavior of MABC algorithm is also assessed when problem. J Chin Inst Eng 2014;37(2):232–42.
it is endowed with three modalities for generating random [16] Jeddi B, Vahidinasab V. A modified harmony search method for environmental/
economic load dispatch of real-world power systems. Energy Convers Manage
sequences (Cat, Log and Random) and two selection schemes of
2014;78(2):661–75.
the solutions (disruptive selection and classical proportional [17] Subbaraj P, Rengaraj R, Salivahanan S. Enhancement of self-adaptive real
selection). The combination between the GRS modalities and the coded genetic algorithm using Taguchi method for economic dispatch
selection schemes determine six variants of MABC, which are problem. Appl Soft Comput 2011;11(1):83–92.
[18] Chaturvedi KT, Pandit M, Srivastava L. Particle swarm optimization with time
applied to nine cases (C1–C9) and then compared in pairs through varying acceleration coefficients for non-convex economic power dispatch. Int
Mann–Whitney U test. Mann–Whitney U test shows that the MABC J Electr Power Energy Syst 2009;31(6):249–57.
 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62 61

[19] Pandit N, Tripathi A, Tapaswi S, Pandit M. An improved bacterial foraging [52] Eslami M, Shareef H, Khajehzadeh M. Optimal design of damping controllers
algorithm for combined static/dynamic environmental economic dispatch. using a new hybrid artificial bee colony algorithm. Int J Electr Power Energy
Appl Soft Comput 2012;12(11):3500–13. Syst 2013;52(11):42–54.
[20] Hota PK, Barisal AK, Chakrabarti R. Economic emission load dispatch through [53] Li Y, Wang Y, Bin L. A hybrid artificial bee colony assisted differential evolution
fuzzy based bacterial foraging algorithm. Int J Electr Power Energy Syst algorithm for optimal reactive power flow. Int J Electr Power Energy Syst
2010;32(7):794–803. 2013;52(11):25–33.
[21] Pandi VR, Panigrahi BK, Mohapatra A, Mallick MK. Economic load dispatch [54] Kıran MS, Gündüz M. A recombination-based hybridization of particle swarm
solution by improved harmony search with wavelet mutation. Int J Comput Sci optimization and artificial bee colony algorithm for continuous optimization
Eng 2011;6(1):122–31. problems. Appl Soft Comput 2013;13(4):2188–203.
[22] Hardiansyah. Solving economic dispatch problem with valve-point effect using [55] Liao X, Zhou J, Ouyang S, Zhang R, Zhang Y. An adaptive chaotic artificial bee
a modified ABC algorithm. Int J Electr Comput Eng 2013;3(3):377–85. colony algorithm for short-term hydrothermal generation scheduling. Int J
[23] Aydin D, Ozyon S, Yasar C, Liao T. Artificial bee colony algorithm with dynamic Electr Power Energy Syst 2013;53(12):34–42.
population size to combined economic and emission dispatch problem. Int J [56] Khorsandi A, Hosseinian SH, Ghazanfari A. Modified artificial bee colony
Electr Power Energy Syst 2014;54(1):144–53. algorithm based on fuzzy multi-objective technique for optimal power flow
[24] Sayedi E, Farsangi MM, Barati M, Lee KY. A modified shuffled frog leaping problem. Electr Power Syst Res 2013;95(2):206–13.
algorithm for nonconvex economic dispatch problem. Power Energy Soc [57] Gao F, Fei F-X, Xu Q, Deng Y-F, Qi Y-B, Balasingham I. A novel artificial bee
General Meeting, 2012 IEEE, San Diego, CA, p. 1–8. colony algorithm with space contraction for unknown parameters
[25] Wang S-K, Chiou J-P, Liu C-W. Non-smooth/non-convex economic dispatch by identification and time-delays of chaotic systems. Appl Math Comput
a novel hybrid differential evolution algorithm. IET Gener Trans Distrib 2012;219:552–68.
2007;1(5):793–803. [58] Akbari R, Hedayatzadeh R, Ziarati K, Hassanizadeh B. A multi-objective
[26] He Da-K, Wang FL, Mao ZZ. A hybrid genetic algorithm approach based on artificial bee colony algorithm. Swarm Evol Comput 2012;2(2):39–52.
differential evolution for economic dispatch with valve-point effect. Int J Electr [59] Gu W, Yin M, Wang C. Self adaptive artificial bee colony for global numerical
Power Energy Syst 2008;30(1):31–8. optimization. IERI Proc 2012;1:59–65.
[27] Victoire TAA, Jeyakumar AE. Hybrid PSO–SQP for economic dispatch with [60] Sinha N, Chakrabarti R, Chattopadhyay PK. Evolutionary programming
valve-point effect. Electr Power Syst Res 2004;71(1):51–9. techniques for economic load dispatch. IEEE Trans Evol Comput
[28] Alsumait JS, Sykulski JK, Al-Othman AK. A hybrid GA–PS–SQP method to solve 2003;7(1):83–94.
power system valve-point economic dispatch problems. Appl Energy [61] Cat Map, Wikipedia Web Site, <http://en.wikipedia. org/wiki/Arnold%27s cat
2010;87(5):1773–81. map> [retrieved 25.02.14].
[29] Niknam T, Mojarrad HD, Meymand HZ. A novel hybrid particle swarm [62] Logistic Map, Wikipedia Web Site, <http://en.wikipedia.org/wiki/Logistic
optimization for economic dispatch with valve-point loading effects. Energy map> [retrieved 25.02.14].
Convers Manage 2011;52(4):1800–9. [63] Alzaqebah M, Abdullah S. Hybrid artificial bee colony search algorithm based
[30] Kumar R, Sharma D, Sadu A. A hybrid multi-agent based particle swarm on disruptive selection for examination timetabling problems, Lecture Notes
optimization algorithm for economic power dispatch. Int J Electr Power Energy in Computer Science (including subseries Lecture Notes in Artificial
Syst 2011;33(1):115–23. Intelligence and Lecture Notes in Bioinformatics) 6831 LNCS 2011; p. 31–45.
[31] Cai J, Li Q, Li L, Peng H, Yang Y. A hybrid CPSO–SQP method for economic [64] Kuo T, Hwang S-Y. Using disruptive selection to maintain diversity in genetic
dispatch considering the valve-point effects. Energy Convers Manage algorithms. Appl Intel 1997;7(3):257–67.
2012;53(1):175–81. [65] Park J-B, Jeong Y-W, Shin J-R, Lee KY. An improved particle swarm
[32] Reddy AS, Vaisakh K. Shuffled differential evolution for large scale economic optimization for nonconvex economic dispatch problems. IEEE Trans Power
dispatch. Electr Power Syst Res 2013;96(3):237–45. Syst 2010;25(1):155–66.
[33] Pandi RV, Panigrahi BK, Bansal RC, Das S, Mohapatra A. Economic load dispatch [66] Alatas B. Chaotic harmony search algorithms. Appl Math Comput
using hybrid swarm intelligence based harmony search algorithm. Electr 2010;216(9):2687–99.
Power Compo Syst 2011;39(8):751–67. [67] Gaing Z-L. Closure to discussion of ‘‘particle swarm optimization to solving the
[34] Kuo CC. A novel coding scheme for practical economic dispatch by modified economic dispatch considering the generator constraints’’. IEEE Trans Power
particle swarm approach. IEEE Trans Power Syst 2008;23(4):1825–35. Syst 2004;19(4):2122–3.
[35] Karaboga D, An idea based on honey bee swarm for numerical optimization. [68] Özyön S, Aydin D. Incremental artificial bee colony with local search to
Erciyes University, Kayseri, Turkey, Technical Report-TR06; 2005. economic dispatch problem with ramp rate limits and prohibited operating
[36] Karaboga D, Akay B. A comparative study of artificial bee colony algorithm. zones. Energy Convers Manage 2013;65(1):397–407.
Appl Math Comput 2009;214(1):108–32. [69] Hosseinnezhad V, Babaei E. Economic load dispatch using h-PSO. Int J Electr
[37] Karaboga D, Basturk B. A powerful and efficient algorithm for numerical Power Energy Syst 2013;49(7):160–9.
function optimization: artificial bee colony (ABC) algorithm. J Global Optim [70] Wang L, Li LP. An effective differential harmony search algorithm for the
2007;39(3):459–71. solving non-convex economic load dispatch problems. Int J Electr Power
[38] Karaboga D, Akay B. A modified Artificial Bee Colony (ABC) algorithm for Energy Syst 2013;44(1):832–43.
constrained optimization problems. Appl Soft Comput 2011;11(3): [71] Lin W-M, Gow H-J, Tsai M-T. Combining of direct search and signal-to-noise
3021–31. ratio for economic dispatch optimization. Energy Convers Manage
[39] Luo J, Wang Q, Xiao X. A modified artificial bee colony algorithm based on 2011;52(1):487–93.
converge-onlookers approach for global optimization. Appl Math Comput [72] Niknam T. A new fuzzy adaptive hybrid particle swarm optimization
2013;219(20):10253–62. algorithm for non-linear, non-smooth and non-convex economic dispatch
[40] Banharnsakun A, Achalakul T, Sirinaovakul B. The best-so-far selection in problem. Appl Energy 2009;87(1):327–39.
Artificial Bee Colony algorithm. Appl Soft Comput 2011;11(2):2888–901. [73] Cai J, Li Q, Li L, Peng H, Yang Y. A hybrid FCASO-SQP method for solving the
[41] Zhu G, Kwong S. Gbest-guided artificial bee colony algorithm for numerical economic dispatch problems with valve-point effects. Energy
function optimization. Appl Math Comput 2010;217(7):3166–73. 2012;38(1):346–53.
[42] Gao WF, Liu SY, Huang LL. A global best artificial bee colony algorithm for [74] Khamsawang S, Jiriwibhakorn S. DSPSO-TSA for economic dispatch problem
global optimization. J Comput Appl Math 2012;236(11):2741–53. with nonsmooth and noncontinuous cost functions. Energy Convers Manage
[43] Gao WF, Liu SY. A modified artificial bee colony algorithm. Comput Oper Res 2010;51(2):365–75.
2012;9(3):687–97. [75] Tsai MT, Yen CW. The influence of carbon dioxide trading scheme on economic
[44] Sonmez M. Artificial Bee Colony algorithm for optimization of truss structures. dispatch of generators. Appl Energy 2011;88(12):4811–6.
Appl Soft Comput 2011;11(2):2406–18. [76] Xiong G, Shi D, Duan X. Multi-strategy ensemble biogeography-based
[45] Akay B, Dervis Karaboga D. Artificial bee colony algorithm for large-scale optimization for economic dispatch problems. Appl Energy
problems and engineering design optimization. J Intell Manuf 2013;111(11):801–11.
2012;23(4):1001–14. [77] Reddy AS, Vaisakh K. Shuffled differential evolution for economic dispatch
[46] Haluk G, Cengiz MT. Comparative performance analysis of artificial bee colony with valve point loading effects. Int J Electr Power Energy Syst
algorithm for automatic voltage regulator (AVR) system. J Franklin Inst 2013;46(3):342–52.
2011;348(8):1927–46. [78] Lohokare MR, Panigrahi BK, Pattnaik SS, Devi S, Mohapatra A. Neighborhood
[47] Sßahin AS, Kılıç B, Kılıç U. Design and economic optimization of shell and tube search-driven accelerated biogeography-based optimization for optimal load
heat exchangers using Artificial Bee Colony (ABC) algorithm. Energy Convers dispatch. IEEE Trans Syst Man Cybern 2012;42(5):641–52.
Manage 2011;52(11):3356–62. [79] Coelho LS, Mariani VC. Correction to ‘‘combining of chaotic differential
[48] Huang SJ, Liu XZ. Application of artificial bee colony-based optimization for evolution and quadratic programming for economic dispatch optimization
fault section estimation in power systems. Int J Electr Power Energy Syst with valve-point effect’’. IEEE Trans Power Syst 2006;21(3):1465.
2013;44(1):210–8. [80] Wang Y, Li B, Yuan B. Hybrid of comprehensive learning particle swarm
[49] Ayan K, Kiliç U. Artificial bee colony algorithm solution for optimal reactive optimization and SQP algorithm for large scale economic load dispatch
power flow. Appl Soft Comput 2012;12(5):1477–82. optimization of power system. Sci China-Inform Sci 2010;53(8):1566–73.
[50] Aydın D, Ozyon S. Solution to non-convex economic dispatch problem with [81] Wang Y, Li B, Weise T. Estimation of distribution and differential evolution
valve point effects by incremental artificial bee colony with local search. Appl cooperation for large scale economic load dispatch optimization of power
Soft Comput 2013;13(5):2456–66. systems. Inform Sci 2010;180(12):2405–20.
[51] Akay B, Karaboga D. A modified Artificial Bee Colony algorithm for real- [82] Abdullah MN, Bakar AHA, Rahim NA, Jamian JJ, Aman MM. Economic dispatch
parameter optimization. Inform Sci 2012;192(1):120–42. with valve point effect using iteration particle swarm optimization. In: 47th
62 D.C. Secui / Energy Conversion and Management 89 (2015) 43–62

international universities power engineering conference (UPEC), London, UK, [87] Sharma R, Samantaray P, Mohanty DP, Rout PK. Environmental economic load
Sept. 2012. p. 1–6. dispatch using multiobjective differential evolution algorithm. In: Proceeding
[83] Bhattacharjee K, Bhattacharya A, Dev SHn. Oppositional real coded chemical of international conference on energy, automation, and signal (ICEAS); 2011. p.
reaction optimization for different economic dispatch problems. Int J Electr 1–7. 28–30 Dec. 2011, DOI: 10.1109/ICEAS.2011.6147132.
Power Energy Syst 2014;55(2):378–91. [88] Basu M. Economic environmental dispatch using multi-objective differential
[84] Mann–Whitney U test, Wikipedia Web Site, <http://en.wikipedia.org/wiki/ evolution. Appl Soft Comput 2011;11(2):2845–53.
Mann%E2%80%93Whitney_U> [retrieved 25.02.14]. [89] Selvakumar AI, Thanushkodi K. Optimization using civilized swarm: solution
[85] Ghasemi A. A fuzzified multi objective interactive honey bee mating to economic dispatch with multiple minima. Electr Power Syst Res
optimization for environmental/economic power dispatch with valve point 2009;79(1):8–16.
effect. Int J Electr Power Energy Syst 2013;49(7):308–21. [90] Bhattacharya A, Chattopadhyay PK. Solving complex economic load dispatch
[86] Shayeghi H, Ghasemi A. A modified artificial bee colony based on chaos theory problems using biogeography-based optimization. Expert Syst Appl
for solving non-convex emission/economic dispatch. Energy Convers Manage 2010;37(5):3605–15.
2014;79(3):344–54.