You are on page 1of 23

Soft Computing (2020) 24:15249–15271

https://doi.org/10.1007/s00500-020-04861-4

METHODOLOGIES AND APPLICATION

An improved particle swarm optimization with clone selection


principle for dynamic economic emission dispatch
Shuqu Qian1 · Huihong Wu1 · Guofeng Xu2

Published online: 17 March 2020


© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
In this paper, an improved particle swarm optimization algorithm (PSOCS) that integrates with a clone selection (CS) principle
of artificial immune system is proposed to solve dynamic economic emission dispatch (DEED) problem. Classical particle
swarm optimization method is easy to fall into stagnation when no particle discovers a position that is better than its previous
best position. To overcome the disadvantage, the CS mechanism is used to evolve the personal best swarm (i.e., P best ) at every
generation. The fittest particles in P best will be cloned independently and proportionally to their fitness. In order to force
PSOCS jump out of stagnation, a hybrid mutation scheme (called R/1orCB/1) is developed to mutate the clones generated.
A constrain-handling approach is utilized to repair infeasible solutions for enhancing the ability of adapting to the DEED
problem with various strong constraints. In numerical experiments, the proposed PSOCS is applied to solve three test cases
(5-unit, 10-unit, and 15-unit systems) with nonsmooth fuel cost and emission functions. Simulation results indicate that the
PSOCS can find the high-quality solutions for the DEED problem, when compared with the most recent methods reported in
the literature.

Keywords Dynamic economic emission dispatch · Particle swarm optimization · Clonal selection principle ·
Strong constraints · Hybrid mutation

1 Introduction is an important topic related to applied mathematics. In order


to solve the ED problem, the selection of optimization meth-
Economic dispatch (ED) is a key problem of power system ods is particularly important issue. A number of scholars
optimal operation. It is necessary to build models for the have tried to mix different optimization algorithms to form a
electricity dispatching principles and plans in order to ensure hybrid intelligent optimization algorithm. The hybrid intelli-
fair competition and real-time ED by various power genera- gent optimization algorithm could fully utilize the advantages
tion enterprises. In recent years, the research on ED mainly of various algorithms, complement each other’s advantages,
focuses on the optimal start–stop problem of the units and the and overcome the flaws in a single intelligent optimization
spot market daily power generation plan. However, the study algorithm with satisfying results. In Neto et al. (2017), in
on the power system dispatching model and related algorithm order to improve the global searching capability and pre-
vent the convergence to local minima, the greedy randomized
adaptive search procedure (C-GRASP) and differential evo-
Communicated by V. Loia.
lution (DE) were combined to form a hybrid algorithm (called
B Shuqu Qian C-GRASP-DE) for ED problem. Elsakaan et al. (2018) pro-
shuquqian@163.com posed an enhanced Moth-Flame algorithm (EMFO) for ED
Huihong Wu with value-point effects and emissions. In Silva Chavez et al.
asuwhh@163.com (2019), the EMFO combines the merits of the traditional
Guofeng Xu MFO and lev́y flight to increase the diversity of population.
Nationalwind@sina.com A robust adaptive hybrid optimization approach was pro-
1 posed to solve the ED problem in large-scale power grids. In
School of Science, Anshun University, Anshun 561000,
Guizhou, China Gholamghasemi et al. (2019), an improved particle swarm
2 optimization (PSO) with a phasor angle was used to solve
University of Engineering of Nanjing, Nanjing 211167, China

123
15250 S. Qian et al.

different types of convex and nonconvex/nonsmooth eco- conventional methods that make use of derivatives and gra-
nomic load dispatch problems (LDE) in different benchmark dients, in general, are not easy to locate the global optimal
power systems. In Singh and Dhillon (2019), ameliorated solution. In recent years, metaheuristic intelligence algo-
gray wolf optimization algorithm was proposed to solve eco- rithms, such as genetic algorithms (GAs) (Basu 2008), PSO
nomic power LDE that synergizes between exploration and (Mason et al. 2017), DE (Basu 2014; Zhang et al. 2015),
exploitation aspects. In Yuan and Yang (2019), a new hybrid harmony search algorithm (HSA) (Niu et al. 2014), teach-
intelligent algorithm based on the PSO and artificial fish ing learning-based optimization (TLBO) (Yang et al. 2014),
swarm algorithm (AFSA) was developed to optimize the biogeography-based optimization (BBO) (Ma et al. 2017),
ED of the electric power system. In Aragón et al. (2015), have been developed to solve the DEED problem efficiently.
an improved algorithm inspired on the T cell model of the Basu (2008) applied NSGAII to obtain the Pareto-optimal
immune system was proposed to solve ED problems. solutions of the DEED problem. A new constraint-handling
Nowadays, environmental problems have seriously method for the power balance constraint was proposed when
restricted the sustainable development of the world’s econ- NSGAII was employed to solve the DEED problem. The
omy and society. The economic emission dispatch (EED), output of the first (N − 1) generators for each time t is speci-
which optimizes the generation cost and emission level fied (generated) firstly, while the output of the N th generator
simultaneously based on the traditional ED, has gradu- is then found by solving a quadratic functions using a stan-
ally become an important approach to reduce emissions dard algebraic method. The comparison with the classical
of the power industry. Cardoso Bora et al. (2019) pro- search techniques demonstrates the superiority of the pro-
posed an improved nondominated sorting genetic algorithm posed algorithm. However, in the approach above, the power
II (NSGAII) approach incorporating a parameter-free self- level of the N th generator found is till over the lower or upper
tuning by reinforcement learning technique called learner boundary. The same author proposed a multiobjective differ-
nondominated sorting genetic algorithm (NSGA-RL) for the ential evolution (MODE) algorithm in Basu (2014), where
multiobjective EED problem. As an extension of the EED, the same constraint-handling technique is used. The results
the dynamic economic emission dispatch (DEED), which is obtained from the MODE are compared with those obtained
more in line with the actual short-term dispatching require- by NSGAII. It is seen from the comparison that the MODE
ments, has gained more and more academic attention. DEED provides a competitive performance in terms of the Pareto-
serves to schedule the committed generator outputs with optimal front (PF) obtained. In Bahmanifirouzi et al. (2012),
the predicted load demands over a certain period of times an interactive fuzzy-based approach called fuzzy adaptive
so as to minimize both cost and emission simultaneously modified theta particle swarm optimization (FAMTPSO)
while satisfying ramp rate constraints, load demands con- was introduced to investigate the multiobjective day-ahead
straints, and power balance limit etc. (Jebaraj et al. 2017). DEED problem considering the effect of wind power gener-
DEED is normally solved by decomposition of the entire ators. In Niknam et al. (2012), an improved θ multiobjective
dispatch period into a number of small time intervals, over TLBO algorithm was proposed to solve the DEED problem.
which the load is assumed to be constant (Basu 2006). The fuzzy clustering technique is utilized to handle the size
Because of the large numbers of the EED involved in a of the repository and obtain profitable solutions from the
DEED calculation and the narrow and uneven search space decision maker’s point of view. In simulation experiments,
being made by the additional dynamic constraints, a fast the proposed algorithm can achieve a uniformly distributed
and accurate algorithm is desired by the industry for both PF. More recently, an improved multi-objective hybrid DE
online implementation and offline studies. Currently, the based on simulated annealing technique (MOHDE-SAT) was
available methods for solving the DEED problem can be developed to solve DEED problem (Zhang et al. 2015).
classified into two categories: (1) conventional mathemat- The MOHDE-SAT integrates the orthogonal initialization
ical methods and (2) heuristic approaches. Starting from method into the DE, and a modified mutation operator and
mathematical methods based on Lagrange multipliers, there archive retention mechanisms are utilized to control conver-
are gradient methods by Wood and Wollenberg (1996) and gence rate. In Mason et al. (2018), a novel multiobjective
quadratic programming by dos Santos Coelho and Mari- neural network trained with DE (MONNDE) was presented
ani (2006), and dynamic programming by Ross and Kim for the DEED problem. MONNDE utilizes neural network
(2007). These methods use gradient information to search the function approximators to address the DEED model. DE is
solution space near an initial starting point. However, most used to evolve neural networks capable of generating Pareto-
are not able to provide an optimal solution and usually get optimal fronts. In Shen et al. (2019), an efficient fitness-based
stuck a local minimum or require enormous computational DE (EFDE) algorithm and a constraint-handling technique
effort due to the nonlinear and nonconvex characteristics of were proposed for the DEED problem. In EFDE, an archive
generating units. Furthermore, a good starting point is vital containing the current and previous population is estab-
for those approaches to be executed successfully. Therefore, lished to provide more candidate solutions. DE/rand/1 and

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15251

DE/best/1 are used to generate mutant individuals. The infea- overcome the premature convergence of the classical PSO.
sible solutions are effectively avoided by a proposed repair In PSOCS, the multiobjective DEED problem is converted
technique. into a single-objective one using the weighting method. In
However, some issues happen when the above multiob- order to compensate the convergence speed and supply more
jective algorithms are adopted to solve the DEED problem. valuable information to adjust the trajectories of the parti-
Such as, (1) in order to maintain a good spread of solu- cles, the superior particles in P best population are selected to
tions, the global convergence is ignored when multiobjective clone, and then the clones will be mutated by using a novel
techniques are adopted to solve the DEED problem, and DE mutation strategy. A feasibility-based rule proposed by
(2) to provide a good compromise solution, an appropriate Deb (2000) is used to update the P best population at every
decision-making method is needed to find the optimal deci- generation. In additional, a repair scheme is adopted to han-
sion solution from the obtained Pareto-optimal solutions. As dle the real power balance constraints, which can improve
mentioned in Elaiw et al. (2013), in order to overcome the the feasibility ratio of individuals in the evolution popu-
drawbacks above, the multiobjective DEED problem can be lation. We examine the performance of the PSOCS on a
converted to a single-objective problem by a linear combina- series of DEED test systems. Simulation results are com-
tion of both objectives, which will yield meaningful results pared with that of several peer algorithms, which indicate
to a decision maker. In Pandit et al. (2012), the multiobjective that the PSOCS can obtain more superior performance than
DEED problem is converted to a single-objective optimiza- other algorithms. Moreover, in the experiment, we also make
tion problem by a fuzzy satisfying method and penalty factor a series of analysis regarding several important strategies in
of the price method, respectively. In Elaiw et al. (2013), the PSOCS.
Elaiw proposed two hybrid optimization methods (DE-SQP The major contributions of the paper are pointed out as
and PSO-SQP) to solve the DEED problem by aggregating follows:
the objectives and constraints using the weighting method.
The first algorithm combines DE and sequential quadratic 1. An improved PSO algorithm based on the clonal selection
programming (SQP). The second one is hybrid PSO and (CS) principle of the artificial immune system (PSOCS)
SQP. DE or PSO is used as a global optimizer, and SQP is proposed to solve DEED problem with large-scale con-
is used as fine-tuning to determine the final optimal solu- straints.
tion. For all we know, comparing the existing algorithms 2. The CS mechanism is incorporated to update the previous
in the literature, DE-SQP and PSO-SQP provide the best best particles to force the PSOCS jump out of stagnation.
results in terms of fuel cost and emission. However, the 3. A hybrid mutation scheme (called R/1orCB/1) is con-
computational cost is large since the two algorithms request structed to mutate the clones generated and enhance the
a number of iterations. In Mason et al. (2017), the cost exploration and exploitation capabilities of the PSOCS.
and emission functions (along with the penalty function) 4. Three test systems (5-unit, 10-unit, and 15-unit system)
were combined to form a single-objective function to be with nonsmooth fuel cost and emission functions are
minimized. A total of eight different PSO variations were solved by the PSOCS algorithm.
tested to determine which PSO variant performs best on the
DEED problem. Simulation results indicate that the PSO This paper is organized as follows: The problem formu-
with avoidance of worst locations (PSOAWL) performs sta- lation of DEED is shown in Sect. 2. The basic concepts of
tistically better the PSO under four different topologies. PSO and clone selection principle are introduced in Sect. 3.
Although the methods based on PSO can obtain a set of Section 4 presents the proposed PSOCS. The obtained results
desirable results, they also fall into premature convergence of three cases study are given in Sect. 5. Finally, conclusions
problem that all evolutionary algorithms suffer (Zhang et al. are drawn in the last part.
2015).
In conclusion, various optimization methods have been
developed to solve the DEED problem using multiobjective 2 Problem formulation of DEED
or single-objective formulation from the reviews above, but
the complexity of the DEED problem motivates the need In this section, we present the formulation of the DEED
for more efficient algorithms to precisely locate the opti- problem which attempts to minimize both cost and emis-
mal dispatch solution. Moreover, any improvement in the sion simultaneously, while satisfying equality and inequality
solution will have a significant impact on the economic and constraints including real power balance and ramp rate lim-
environmentally friendly operation of power systems. In this its. The following objectives and constraints are taken into
paper, an improved PSO that combines the CS mechanism account in the formulation of the DEED problem.
of the biological immune system (PSOCS) is proposed to
solve the DEED problem with value-point effects in order to

123
15252 S. Qian et al.

2.1 Cost objective 


N 
N 
N
P Lt = pt,i Bi j pt, j + B0i pt,i + B00 (4)
i=1 j=1 i=1
The cost objective function considering the valve-point effect
is expressed as the sum of a quadratic and a sinusoidal func-
where Bi j , B0i , B00 are the transmission loss coeffi-
tion:
cients.
– Inequality constraint: Real power generation limit.
min F(p)

T 
N
= {(ai + bi pt,i + ci pt,i
2
) + |di sin[ei ( pimin − pt,i )]|}
t=1 i=1 (1) Pimin ≤ pt,i ≤ Pimax , i = 1, 2, . . . , N . (5)

where T is the number of hours during the scheduling period where Pimin and Pimax represent the minimum and max-
and ai , bi , ci , di , and ei are the cost coefficients for the ith imum generation limit of ith unit in M W .
thermal unit. N is the number of thermal power generators, – Inequality constraint: Generating unit ramp rate limit.
while F(p) presents the total cost of the system in $. p is the
output of the ith thermal unit at time t in M W . to avoid undue thermal stresses on the boiler and the com-
bustion equipment, the rate of change of the output power
2.2 Emission objective of each thermal unit must not exceed a certain rate during
increasing or decreasing the power output of each unit.
The atmospheric pollutants such as sulfur oxides (SX x ), This is mathematically represented as follows:
nitrogen oxides (NOx ) and carbon dioxide (CO2 ) cased by
fossil fuel-fired generator can be modeled separately. How- pt,i − pt−1,i
ever, for comparison purposes, the total emission of these ≤ U Ri , i = 1, 2, . . . , N , t = 1, 2, . . . , T
pollutants which is the sum of a quadratic and an exponen- (6)
pt−1,i − pt,i
tial function can be expressed as:
≤ D Ri , i = 1, 2, . . . , N , t = 1, 2, . . . , T .
min E(p)

T 
N where U Ri and D Ri are the up and down ramp rate limits
= {(αi + βi pt,i + γi pt,i
2
) + ηi exp(δi pt,i )} of ith unit in M W .
t=1 i=1
(2)

where αi , βi , γi , ηi , and δi are emission coefficient of ith 3 PSO algorithm and CS principle
unit. E(p) is the total emission of the system in lb.
In this section, we concentrate on different variants of the
2.3 Constraints PSO and the clonal selection principle, which are relevant
to the proposed algorithm. The first is several variants of the
The DEED problem is subject to the following constraints. PSO which are presented in Sect. 3.1. The second is the clonal
selection principle which is introduced in Sect. 3.2.
– Equality constraint: Real power balance.
The real power generation must balance the predicted 3.1 PSO algorithm
power demand plus the real power losses in the trans-
mission lines, at each time interval over the scheduling PSO algorithm was first proposed by Kennedy and Eberhart
horizon. (1995) which has become a popular field of study in recent
years due to its applicability to a broad range of research

N areas. The main ideal of PSO is to evaluate candidate indi-
pt,i = P Dt + P L t , t = 1, 2, . . . , T . (3) viduals (particles) and eventually move toward the global
i=1 best particle (i.e., gbest). Initially each particle is assigned
a random position in decision space and a velocity in the
where P Dt is the total demand in M W at time t and P L t range of [Vmin , Vmax ]. At each iteration (i.e., g), each parti-
is the total transmission loss in M W at time t. The most cle evaluates its position within the problem space defined
popular approach for finding an approximate value of the by a fitness function. If a new position has a better fitness
loss is by Kron’s loss formula as given in Eq. (4) which than the previous best position for that particle, the particle
can be expressed in the quadratic form as follows: will remember this new position as its personal best position

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15253

( pbest). Each particle updates its velocity, and as a result its where R1 and R2 are generated using abs(N (0, 1)). The
new position. mean of abs(N (0, 1)) is 0.798, and the variance is 0.36
Supposing that at the gth iteration, the position and veloc- according to the statistical knowledge.
g g g
ity of the ith particle in the population are xi = (xi,1 , xi,2 , More recently, Mason et al. (2017) extended the previous
g g g g g
. . . , xi,n ) and vi = (vi,1 , vi,2 , . . . , vi,n ), respectively. Its approach to develop a PSOAWL algorithm. The PSOAWL is
new velocity and position at the (g + 1)th iteration will be inspired by incorporating a negative stimulus on the motion
obtained by using the following formulas: of the particles in the form of the worst locations in the prob-
lem space. Their proposed velocity update method (denoted
g+1 g g g
vi, j = ωvi, j + c1 r1 ( pbesti, j − xi, j ) as Mason-updating method) is as follows:
g g
+ c2 r2 (gbest j − xi, j ), (7) g+1 g
g+1 g g+1
vi, j = χ (vi, j + l1 + l2 + l3 + l4 ),
xi, j = xi, j + vi, j , (8) g g
l1 = r1 c1 ( pbesti, j − xi, j ),
g g
where c1 and c2 are acceleration coefficients (nonnegative l2 = r2 c2 (gbest j − xi, j ),
constants), which control the influence of pbest and gbest  
l1 (12)
during the search process, and ω is the inertia weight to l3 = r3 c3 ,
g g
control the exploration capability of particles in the search 1 + |xi, j − pwor sti, j |
 
space; r1 and r2 are two random numbers within [0,1]. The l2
personal best position of the ith particle at the gth gener- l4 = r4 c4 g g .
g g 1 + |xi, j − gwor st j |
ation is denoted as pbesti , which can represent pbesti =
g g g
( pbesti,1 , pbesti,2 , . . . , pbesti,n ), while the global best par-
ticle in the whole population is denoted as gbest g , which can The new l3 and l4 terms take into account the worst loca-
g g g tion by using the difference between the particles position
represent gbest g = (gbest1 , gbest2 , . . . , gbestn ).
In order to improve the performance of the classical and the worst position. The absolute value of this difference
PSO, various velocity update schemes have been devel- is taken to ensure the extra velocity given to the particle
oped in recent years. In Clerc and Kennedy (2002), Clerc from l3 and l4 is in the same direction as l1 and l2 , respec-
g g
and Kennedy proposed a new constriction value χ which tively. In addition, one is added to |xi, j − pwor sti, j | and
g g
provided the PSO with a more stable convergence. The veloc- |xi, j − gwor st j | in order to avoid dividing by 0 when the
ity update method (denoted as Clerc-updating method) is particle is at the worst location and to limit the extra velocity
g
implemented as follows: given to the particle from t3 and t4 . The pwor sti, j repre-
g
sents the element of the personal worst position pwor sti of
g+1 g g g g
vi, j = χ × [ωvi, j + c1 r1 ( pbesti, j − xi, j ) the ith particle, while the gwor st j represents the element
g g (9) of the global worst position gwor st g of the whole popula-
+ c2 r2 (gbest j − xi, j )]
tion. In general, the parameters c1 and c2 = 1.845 and c3
where ω and χ are the inertial and constriction factors that and c4 = 0.205 are selected as the acceleration coefficients
have been defined, respectively, as: which gave the best overall performance according to Mason
et al. (2017). All ri values are random numbers in the range
g of [0,1].
ω = ωmax − (ωmax − ωmin ),
G  (10) In addition to improve the velocity update method, a
χ = 2/|2 − ϕ − ϕ 2 − 4ϕ|, combination of classical PSO with other optimization tech-
niques, such as operations research techniques, artificial
here ωmin and ωmax are initial and final weights, g and G are intelligence techniques, or other research areas, can also
the iteration counter and the maximum number of iterations, lead to getting a much better algorithm for complex opti-
and ϕ = c1 + c2 > 4. The terms c1 and c2 = 2.25 are mization problems. For instance, Djemame et al. (2019)
acceleration coefficients. proposed a quantum-inspired PSO algorithm (QPSO) that
Krohling and dos Santos Coelho (2006) analyzed Eq.(9) is a hybridization of the PSO algorithm and quantum com-
and concluded that interval [0.72,0.86] is a possible good puting principles. The QPSO needs no velocity vectors for
choice for χ . Therefore, the absolute value of the Gaussian particles and also has fewer parameters to adjust during
probability distribution with zero mean and unit variance the iteration process, making it easier to implement. The
abs(N (0, 1)) is introduced into the velocity update equation experimental results indicate that QPSO outperforms most
(denoted as Krohling-updating method) in Eq.(9). of the classic PSO in convergence speed and achieves bet-
ter levels for the reverse emergence real-world optimization
g+1 g g g g
vi, j = R1 ( pbesti, j − xi, j ) + R2 (gbest j − xi, j ), (11) problem. In Kalita and Singh (2020), the authors proposed

123
15254 S. Qian et al.

an improved PSO algorithm (Q-DMPSO) to deal with SVM


model selection problem in dynamic environment. The Q-
DMPSO uses multiswarm-based optimization with exclusion
and anti-convergence theory to select the optimal values for
the SVM hyperparameters. To maintain the diversity among
the swarms, it uses quantum particles to create repulsion
among the particles. The experiments performed using the
Q-DMPSO shown better results in comparison with other
techniques like traditional gird search, PSO, and chained
PSO. In Guo et al. (2017), to overcome the issue that existing
PSO algorithm could not adaptively adjust the key parameters
and easily fell into premature, a novel multiobjective cul-
tural PSO algorithm (MOCPSO) was proposed for solving
uncertain optimization problems. In MOCPSO, particles’s
locations and weight inertia are updated by normative knowl-
edge. The grid’s coverage degree is defined in topographic
knowledge. The solutions with smaller coverage degrees are
chosen as the global best or local best to ensure the uni-
Fig. 1 CS principle
form distribution of the PF. The accelerating coefficients
are dynamically adjusted in terms of the grid’s coverage
degree so as to balance exploration and exploitation. Similar-
ity, an improved PSO algorithm based on cultural algorithm 3.2 CS principle
(IPSOCA) was developed for constrained optimization prob-
lems (Wang et al. 2012). Fixed proportion elites are selected Over the last few years, there has been an ever-increasing
to construct the swarm of belief space. The belief space interest in the area of artificial immune systems (AIS) and
updates its normative knowledge and situational knowl- their applications. The CS principle extracted from AIS can
edge according to the elite particles, and the elite swarm lead to the development of useful computational tools which
in the belief space performs PSO operation according to the can be used to solve complex engineering tasks. A new opti-
update knowledge and generates new particles. The simula- mization algorithm, named CS algorithm, was proposed to
tion results of the six constrained functions verify the validity solve multimodal and combinational optimization tasks by
of the IPSOCA. In Ghasemi et al. (2019), a simple and effi- De Castro and Von Zuben (2002). The CS principle, which is
cient variant of PSO, phasor particle swarm optimization demonstrated in Fig. 1, indicates that some subpopulations
(PPSO), was proposed for unimodal and multimodal stan- of B lymphocytes (B cells) respond by producing a number
dard test functions and traditional benchmark functions. The of antibodies (Abs), when an animal is exposed to an antigen
PPSO is based on modeling the particle control parameters (Ag). Each B cell secretes a single type of Ab, which is rela-
with a phase angle (θ ), inspired from phasor theory in the tively specific for the Ag. By binding to these Abs and with
mathematics. This phase angle (θ ) converts PSO algorithm a second signal from accessory cells (e.g., T-helper cell), the
to a self-adaptive, trigonometric, balanced, and nonparamet- Ag stimulates the B cell to divide and mature into terminal
ric metaheuristic algorithm. In Zhai et al. (2019), an improved Ab-secreting cells, called plasma cells. The process of cell
simulated annealing particle swarm optimization (ISAPSO) division generates a clone population (i.e., a cell or set of cells
was proposed for unit commitment problem (UCP). The that are the progenies of a single cell). In addition, some B
ISAPSO consists of a two-layer structure, which reduces cells can differentiate long-lived B memory cells. Memory
the dimension of UCP. In the upper layer, the elitist strat- cells circulate through the blood, lymph, and tissues. When
egy PSO and simulated annealing technique are combined exposed to a second antigenic stimulus, commence to differ-
to jump out of the local optimum when solving UCP. In the entiate into plasma cells capable of producing high-affinity
lower layer, the interior point method is used to improve the Abs preselected for the specific Ag that had stimulated the
search efficiency of ISAPSO. The experimental results show primary response.
that ISAPSO can obtain better solutions than other algorithms The main features of the CS principle are included by two
in solving UCP problems. More researches about improved main mechanisms: (1) division and (2) maturation. In the
PSO can refer to Sengupta et al. (2019). process of division, the highest affinity antibodies compose a
new subpopulation, and then the selected antibodies will be
cloned independently and proportionally to their affinities,
generating a repertoire of clones: The higher the antibody’s

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15255

affinity, the higher the number of clones generated for each uating each particle in the PSOCS as follows:
of the selected antibodies. In the process of maturation, the
repertoire is submitted to an affinity maturation inversely pro- min f (p) = ω × F(p) + (1 − ω) × E(p)
portional to the affinity, generating a population of maturated 
N
s.t. pt,i = P Dt + P L t t = 1, 2, . . . , T
clones: The higher the clone’s affinity, the smaller the muta- i=1
tion rate. After the division and maturation, a reselection pt,i − pt−1,i ≤ U Ri i = 1, 2, . . . , N ; t = 2, . . . , T
mechanism is adapted to select the clones with highest affin- pt−1,i − pt,i ≤ D Ri i = 1, 2, . . . , N ; t = 2, . . . , T
ity to be a candidate to enter the next population. (13)
The CS principle has shown considerably successful in
solving a variety of optimization problems. In Panigrahi et al. where p = ( p1,1 , p1,2 , . . . , pT ,N ), pt,i is the output of the
(2007), Panigrahi et al. applied CS algorithm to ED prob- ith thermal unit at time t. ω ∈ [0, 1] is the weighting fac-
lems, and the simulation results indicate that CS algorithm tor. F(p) and E(p) identify the total fuel cost and pollution
can achieve better results than PSO. However, to the best of emission of the system, respectively. T is the number of hours
our knowledge, there has been no published article describing during the study period. N is the number of thermal units of
the application to the DEED problem. In the PSOCS, the CS the system. The problem space which each individual in the
principle will be adopted to update each personal best loca- algorithm will explore will have T ×N dimensions. The num-
tion ( pbest) in P best population, which will be introduced in ber of equal constraints is T , while the number of unequal
Sect. 4.5. constraints is 2(T − 1)N .
The DEED problem will be solved as high-dimensional
optimization problem. Each dimension in the problem space
represents a power generator at a given time. The problem
4 The proposed PSOCS
space will have T × N dimensions. For example, the opti-
mization problem consists of 240 variables for 10-unit system
PSO is one of the modern heuristic algorithms suitable to
considering T = 24 time intervals and N = 10 generators.
solve complex optimization problems. It possesses the advan-
The number of equal and unequal constraints is 24 and 460,
tage of fast convergence and easy implementation. However,
respectively.
when tackling the DEED problem with value-point effects
According to the objective function above in Eq. (13), the
which comprises many local optima and constraints, it lacks
pseudo-code of PSOCS is displayed in Algorithm 1, where g
the ability of finding globally optimal solutions due to the
is denoted as iteration counter. At the begin of the PSOCS, an
drawbacks such as local optimal trapping, insufficient capa-
initial population (denoted as A), which contains N p parti-
bility to find nearby extreme points (Zhang et al. 2014). The
cles with an initial velocity of zero for each one, is produced
working mechanism of PSO only involves two fundamen-
according to Sect. 4.2 and an empty population P best , which
tal updating rules, so it is simple and easy to implement,
is utilized to store each particle’s personal best position pbest ,
and it can be flexibly combined with other search mecha-
is also created. All infeasible particles in A are repaired by
nisms to build an ensemble optimization algorithm. In this
using the repair strategy of the power balance equation intro-
paper, therefore, the classical PSO algorithm that combined
duced in Sect. 4.3. The objective f (p) and the degree of
CS principle is presented to solve the DEED problem. In the
constraint violation V (p) (introduced in Sect. 4.4) are calcu-
following, the implementation of the proposed PSOCS and
lated for each particle in A, and each particle of the repaired
the brief procedure are firstly described in Sect. 4.1, and then
A is copy into P best population to form the initial P best pop-
some main operations will be introduced in Sects. 4.2–4.5.
ulation. Then, at each generation, A is sorted according to
V (·) value in a increasing order. In order to keep one-to-
4.1 Implementation of PSOCS one mapping between each particle and its pbest , the order
of P best changes when A is sorted. The first half part of A
The goal of the DEED problem is to minimize both the is selected to form population Q. The velocity and position
cost and emissions involved in the power generation pro- of each particle in Q are updated by using Eqs. (8) and (12)
cess. This is a multiobjective optimization problem where and calculate the objective value and the degree of constraint
the two objectives are in conflict with each other. In order violation of each particle in Q. The first half part of A is
to solve the multiobjective optimization problem using the replaced with Q to create the new evolution population. For
PSOCS, aggregating the cost objective [Eq. (1)] and emission each particle (i.e., p) in Q, updating P best according to the
objective [Eq. (2)], the multiobjective DEED problem can be feasible selection criterion in Sect. 4.4, and obtain the updated
mathematically formulated as a nonlinear single-objective P best . Some highest fitness particles in P best are selected to
optimization problem using the weighting method as Eq. evolve by using the CS principle in Sect. 4.5. After evolving,
(13). Therefore, we define the objective function for eval- all infeasible particles in P best are repaired, calculating the

123
15256 S. Qian et al.

Algorithm 1 pseudo-code of PSOCS 4.3 Constraint-handling method


1: Let iteration g=1, initialize population A, and set initial P best =
and the initial velocity equal to zero for each particle. The DEED problem has a set of equality constraints [i.e.,
2: repeat real power balance equation Eq. (3)] and two sets of inequal-
3: Evaluate A, let P best = A.
4: Sort the particles in A in increasing order according to V (·). ity constraints [i.e., real power generation limit Eq. (5) and
5: Evolve population A ramp rate limit Eq. (6)], among which inequality constraints
5.1: Select the first half part of A to form Q, and update Q by can be properly handled by repair approach, while equal-
using the Clerc-updating method. ity constraints have been a headache problem. Moreover,
5.2: Evaluate Q, and replace the first half part of A using Q to
create new population. constraint-handling technique for equality constraints has
6: Evolve population P best great influence on the efficiency of solving this problem.
6.1: Update P best by using Q according to Sect. 4.4. Therefore, an efficient repair method for inequality con-
6.2: Evolve P best by using the CS principle in Sect. 4.5. straints is used to guarantee all solutions feasible.
6.3: Repair P best , and evaluate P best .
7: Determine the best particle (pbest ) from P best , and let g + +.
8: until the termination condition is met. 4.3.1 The repair method for the power generation limits
9: Output the pbest in the final population P best . and ramp rate limits

First, in the gth generation, the new bounds of the jth variable
objective and the degree of constraint violation of each par- are updated by:
ticle in P best . The best particle (pbest ) in P best is determined, 
and generation g is added one. This process is repeated gen-  min P jmin if t = 1,
eration after generation until the termination criterion is met. Pj = (16)
max{P jmin , pt−1, j − D R j } Otherwise.

4.2 Structure of particles encoding  max P jmax if t = 1,
Pj = (17)
min{P jmax , pt−1, j + U R j } Otherwise.
For the DEED problem, the main goal is to obtain the power
 
generation scheduling of those thermal units. The power out- where P jmin and P jmax represent the new lower and upper
put of each thermal unit is taken as a decision variable. Each bounds of the jth for the tth time interval ( j = 1, 2, . . . , N ×
particle consists of N × T decision variables. Thus, a particle T ; t = 1, 2, . . . , T ).
p can be represented as follows: Then, any variable pt, j beyond their new bound will be lim-
⎛ ited to the bound as follows:
⎧  
⎜ ⎪
⎨ Pj
min if pt, j ≤ P jmin ,
p = ⎝ p1,1 , p1,2 , . . . , p1,N , p2,1 , p2,2 , . . . , p2,N , . . . ,



pt, j = P jmax if pt, j ≥ P jmax ,

(18)
t=1 t=2 ⎪
⎩p
⎞ t, j Otherwise.

pT ,1 , pT ,2 , . . . , pT ,N ⎠ where Eq. (18) makes the power generation limits and ramp


t=T (14) rate limits satisfied.

4.3.2 The repair method for the power balance constraints


where T and N represent the number of hours and the num-
ber of power generators considered in the DEED problem,
For the equality constraints handling, the heuristic constraint-
respectively. All decision variables are distributed uniformly
handling method in Zhang et al. (2015) is used to properly
between their upper and lower generation limits which are
handle these constraints with coarse adjustment and fine-
produced using Eq. (15)
tuning technique, which can be presented in detail as follows:
pt,i = Pimin + rand × (Pimax − Pimin ),
Step 1 For infeasible solution
t = 1, 2, . . . , T ; i = 1, 2, . . . , N , (15) p = ( p1,1 , p1,2 , . . . , p1,N , p2,1 , p2,2 , . . . , p2,N , . . . ,



t=1 t=2
where rand is a uniform random number from [0,1]. Pimax pT ,1 , pT ,2 , . . . , pT ,N ), set the scheduling time t =


and Pimin are the maximum and minimum real power outputs t=T
of the ith generator, respectively. According to Eqs. (14) and 1;
(15), N p particles are generated to form the initial population Step 2 Calculate the violation value ΔPt between the total
A. output of thermal units and system load by formula

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15257

(19), if the violation |ΔPt | < ε (a positive tolerance 2.30


Krohling−updating method
value), then go to Step 5; Otherwise, go to Step 3; 2.28 Clerc−updating method

Average Fitness Value ( ×106)


2.26 Mason−updating method

N
2.24
ΔPt = pt, j − P Dt − P L t , (19)
j=1 2.22

2.20
Step 3 Modify the output of each thermal unit equally with
2.18
the above violation on coarse adjustment by formula
2.16
(20) for current time interval t, if the modified output
exceeds the boundary limit (i.e., the r th thermal unit 2.14

pt,r ), let it equal to the nearest bound of thermal out- 2.12


put in formulation (18), and go to Step 4; otherwise, 2.10
12 24 36 48 60
go to Step 5;
3
Fitness Function Evaluations (×10 )
ΔPt
pt, j = pt, j − , (20) Fig. 2 Curves of average fitness from three different velocity update
N
methods over 25 independent runs under 60,000 FFEs for each run
Step 4 Modify the output of the r th thermal unit with formu-
lation (20) on fine-tuning adjustment, and calculate If q wins, it replaces its corresponding pbest to survive to
the violation again, if |ΔPt | < ε is satisfied, go to P best ; if not, the corresponding pbest remains.
Step 5; otherwise, go to Step 2;
Step 5 If t < T , the scheduling time t = t + 1, go to Step 4.5 Evolve Pbest using CS principle
2; otherwise, exit all the repair procedures.
In order to improve the local exploration ability of the con-
4.4 Update Pbest using population Q ventional PSO, CS principle is applied to evolve the personal
best population (P best ) which ensures highly preferable posi-
Although the above repair strategy in constraint-handling tions in P best and increases the probability of finding a better
methods is adopted to handle each infeasible particle, some solution. At each generation, each pbest ∈ P best is evaluated
particles may be also infeasible since the ramp rate con- according to the following fitness function fit(·),
straints and power balance constraints cannot be considered
simultaneously in the repair process. Therefore, a feasible fit(p) = f (p) + V (p), (22)
selection criterion is designed to update P best as follows:
First, the violation degree V (p), for a given particle p, is here f (p) represents objective value of particle p and V (p)
defined as: is the violation degree of p. Then all particles in P best are
  sorted according to the fit(·) in an increasing order. Only the
 T N
V (p) = max |P Dt + P L t − pt,i | − ε, 0 first |P best |/Nc particles are selected to evolve by using the
t=1 i=1 CS principle, and here | · | is the cardinality of set. Nc is

N 
T referred to the clone size of each selected particle. In order
+ max{| pt,i − pt−1,i | − U Ri , 0}, to maintain the size of the evolution population (the total
i=1 t=2 (21) number of clones generated remains |P best |), only the first
|P best |/Nc particles are selected to clone. For instance, sup-
where ε is a permitted accuracy. Feasible particle satisfies pose that |P best | = 30 and Nc = 3, and then the first 10
V (p) = 0, while V (p) > 0 means that particle p is infeasible. particles after sorting are selected to produce 30 clones.
Then, for each particle q ∈ Q will be compared with In our implementation, we find that increasing the clone
its corresponding pbest in P best by the following selection size of each particle will improve the solution quality, while
criterion. yielding computationally expensive runs of PSOCS. After
exhaustive experimentation the clone size of each selected
– If q and pbest are both feasible, the particle that has lower particle is finally set to Nc = 3. For example, for each
objective value will win; best particle (pbest ) to be cloned, three identical clones are
– If one particle is feasible, while the other is infeasible, firstly created, and then each cloned pbest will be mutated
the feasible one will win; through the following hybrid mutation scheme (denoted as
– If q and pbest are both infeasible, the particle that has R/1orCB/1), generating three offsprings (i.e., pbest best
1 , p2 , and
lower violation value will win. best
p3 ).

123
15258 S. Qian et al.

Table 1 Statistical results of fitness values obtained by the PSOCS under three velocity update methods over 25 independent runs with 60,000
FFEs for each run
Velocity update methods Best value Worst value Average value SD

Krohling-updating method (Krohling and dos Santos Coelho 2006) 2.1254e+6 2.1174e+6 2.1200e+6 3.4699e–4
Clerc-updating method (Clerc and Kennedy 2002) 2.1197e+6 2.1165e+6 2.1175e+6 6.6360e–5
Mason-updating method (Mason et al. 2017) 2.1204e+6 2.1167e+6 2.1179e+6 8.9775e–4
Bold values indicate best results

2.30 2.22
PSOCS/R/1 PSOCS/R/1
2.28 PSOCS/B/1 PSOCS/CB/1
Average Fitness Value ( ×106)

Average Fitness Value ( ×106)


PSOCS/CB/1 2.20 PSOCS/R/1orCB/1
2.26 PSOCS/B/2
PSOCS/R/2
2.24 2.18

2.22
2.16
2.20

2.18 2.14

2.16
2.12
2.14

2.12 2.10
6 12 18 24 30 12 24 36 48 60
3
Fitness Function Evaluations (×10 ) Fitness Function Evaluations ( ×10 )
3

Fig. 3 Curves of average fitness from five different mutation strategies Fig. 4 Curves of average fitness from three mutation strategies over 25
over 25 independent runs under 60,000 FFEs for each run runs under 60,000 FFEs for each run


R/1(pbest ) if rand < 0.5,
pibest = (23)
C B/1(pbest ) otherwise,
mutation strategy in Eq. (23) is designed in the PSOCS, we
where i = 1, 2, 3, rand is a random number in the range of will explain the reason in detail in Sect. 5.3.
[0,1]. In Eq. (23), R/1(pbest ) and CB/1(pbest ) are referred to After that, the best one of the three offsprings will be
perform DE/rand/1 mutation and DE/current-to-best/1 muta- compared with the pbest according to the feasible selection
tion on particle pbest , respectively, as follows: criterion introduced in Sect. 4.4. If it is better than the pbest , it
will replace the pbest ; if not, the corresponding pbest remains.
R/1(pbest ) ⇔ pibest = prbest + F(prbest − prbest ), The CS process is a powerful stochastic optimization tech-
1 2 3
(24) nique, which has some distinguishing features. The better
C B/1(pbest ) ⇔ pibest = pgbest + F(prbest
1
− prbest
2
), particles of the P best population can be simultaneously oper-
ated on in the search space, and probabilistic transition rules
here pgbest is the global best particle in the P best popu- are used instead of deterministic transition rules. Each pbest is
lation. prbest
1
, prbest
2
, and prbest
3
are three mutually exclusive in fact a local optimum searcher, and the two mutation oper-
particles randomly chosen from the P best . F(F ∈ [0, 2]) ations with 0.5 probability can be regarded as a process of
is the amplification factor. In order to understand why the blind exploration in the search space. After a certain number

Table 2 Statistical results of


Mutation strategies Best value Worst value Average value SD
fitness values obtained by the
PSOCS under five mutation PSOCS/R/1 2.1286e+6 2.1213e+6 2.1247e+6 2.5997e–4
strategies over 25 independent
runs with 30,000 FFEs for each PSOCS/B/1 2.1556e+6 2.1228e+6 2.1325e+6 5.2810e–4
run PSOCS/CB/1 2.1333e+6 2.1187e+6 2.1222e+6 8.6821e–5
PSOCS/B/2 2.2509e+6 2.1764e+6 2.2014e+6 2.9967e–4
PSOCS/R/2 2.2304e+6 2.1839e+6 2.2014e+6 1.0582e–4
Bold values indicate best results

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15259

Table 3 Statistical results of fitness values obtained by the PSOCS under three mutation strategies over 25 independent runs with 60,000 FFEs for
each run
Mutation strategies Best value Worst value Average value SD

PSOCS/R/1 2.1217e+6 2.1179e+6 2.1195e+6 7.7192e–4


PSOCS/CB/1 2.1222e+6 2.1163e+6 2.1179e+6 1.3452e–4
PSOCS/R/1orCB/1 2.1197e+6 2.1161e+6 2.1175e+6 6.6360e–5
Bold values indicate best results

Table 4 Comparison of results obtained by several algorithms for 5-unit, and 10-unit, and 15-unit system
Instance Algorithm Cost ($) (104 ) ± SD Emission (lb)(104 ) ± SD Type

5-unit system DE-SQP (Elaiw et al. 2013) 4.4450 ± 0.0256 1.9616 ± 0.0195 SOAs
PSO-SQP (Elaiw et al. 2013) 4.4542 ± 0.0125 1.9772 ± 0.0098
PSOCS 4.3329 ± 0.0023 1.9934 ± 0.0018
PSO (Basu 2006) 5.0893 ± 0.0268 2.0163 ± 0.0395 MOAs
PPSO (Ghasemi et al. 2019) 4.8369 ± 0.0655 2.3685 ± 0.1028
MONNDE (Mason et al. 2018) 4.9884 ± 0.0170 1.8647 ± 0.0125
MOHDE-SAT (Zhang et al. 2015) 4.8214 ± 0.0056 1.8011 ± 0.0090
10-unit system DE-SQP (Elaiw et al. 2013) 246.88 ± 0.0362 315.64 ± 0.0128 SOAs
PSO-SQP (Elaiw et al. 2013) 247.01 ± 0.0362 315.07 ± 0.0425
PSOAWL (Mason et al. 2017) 254.63 ± 0.0236 294.55 ± 0.0526
PSOCS 252.69 ± 0.0095 298.00 ± 0.0102
NSGAII (Basu 2008) 252.26 ± 0.0269 309.94 ± 0.0694 MOAs
MODE (Basu 2014) 252.24 ± 0.0659 309.97 ± 0.0918
PSO (Mason et al. 2017) 260.44 ± 0.0853 310.75 ± 0.0548
PPSO (Ghasemi et al. 2019) 253.58 ± 0.0526 303.75 ± 0.0436
PSO-CSC (Huang et al. 2019) 252.47 ± 0.3216 305.24 ± 0.2546
MONNDE (Mason et al. 2018) 256.00 ± 0.0125 297.82 ± 0.0180
MOHDE-SAT (Zhang et al. 2015) 252.79 ± 0.0070 297.76 ± 0.0180
15-unit system PSO-SQP (Elaiw et al. 2013) 71.3682 ± 0.1026 30.2365 ± 0.0928 SOAs
PSOAWL (Mason et al. 2017) 70.6128 ± 0.0935 30.7726 ± 0.0346
PSOCS 71.0736 ± 0.0062 26.3625 ± 0.0028
PPSO (Ghasemi et al. 2019) 72.1272 ± 0.2610 26.3004 ± 0.1865 MOAs
PSO-CSC (Huang et al. 2019) 71.9347 ± 0.3216 26.3079 ± 0.2546
MONNDE (Mason et al. 2018) 71.1635 ± 0.0570 27.3906 ± 0.0296
MOHDE-SAT (Zhang et al. 2015) 71.9632 ± 0.0236 29.8269 ± 0.0178
The symbol ’SOAs’ and ’MOAs’ in type column represent the single-objective and multiobjective optimization algorithms, respectively

of generations, each pbest will find different optima, probably 5.1 Experimental setup
including the global optimum.
In order to verify the performance of the PSOCS, we apply
PSOCS to solve 5-, 10-, and 15-unit test systems. A schedul-
ing period of 24 h with 1-h intervals is considered in the
5 Numerical simulation and discussion experiments. The original data in the experiments are shown
in Tables 13, 14, 15, 16, 17, 18, 19 in ‘Appendix A,’ which
In this section, we first construct the experimental settings are taken from Mason et al. (2018). Tables 13 and 14 are the
and then analyze the superiority of different velocity update 5-unit generator coefficients and the B matrix, respectively.
methods and mutation strategies proposed in the PSOCS. Tables 15 and 16 are the 10-unit generator coefficients and the
Finally, simulation experiments are performed to compare B matrix, respectively. Tables 17 and 18 are the 15-unit gen-
the performance of the PSOCS with several peer algorithms erator coefficients and the B matrix, respectively. Table 19
on three practical power systems.

123
15260 S. Qian et al.

6 5
x 10 x 10
3.2 3.1
NSGAII MONNDE
MODE 3.05 PSO−SQP
DE−SQP PSOAWL
3.15 PPSO
PSO 3
PSOAWL PSO−CSC
PSO−SQP 2.95 MOHD−SAT
Emission (lb)

Emission (lb)
3.1 PSOCS
MONNDE
PPSO 2.9
PSO−CSC
3.05 2.85
MOHD−SAT
PSOCS 2.8
3
2.75
2.7
2.95
2.65
2.9 2.6
2.45 2.5 2.55 2.6 2.65 7.05 7.1 7.15 7.2 7.25
6 5
Cost ($) x 10 Cost ($) x 10

(a) 10-unit system (b) 15-unit system

Fig. 5 The best compromise solutions obtained by the compared algorithm

Table 5 Hourly generation (MW) schedule obtained from DEED using DE-SQP for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.000 135.000 73.000 120.383 171.926 121.887 99.089 120.000 51.971 12.327 19.585
2 150.000 135.000 83.743 120.671 171.247 128.104 129.089 120.000 52.224 42.327 22.407
3 150.000 135.000 162.519 169.063 174.656 120.938 129.565 120.000 80.000 44.798 28.540
4 150.000 135.000 207.128 191.717 224.656 160.000 128.788 120.000 80.000 44.171 35.461
5 150.000 135.000 215.593 240.702 243.000 160.000 129.901 120.000 80.000 45.141 39.338
6 150.000 157.443 295.466 285.163 243.000 160.000 130.000 120.000 80.000 55.000 48.073
7 150.000 221.021 296.156 300.000 243.000 160.000 130.000 120.000 80.000 55.000 53.176
8 178.991 227.559 339.854 300.000 243.000 160.000 130.000 120.000 80.000 55.000 58.404
9 258.991 307.559 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.549
10 301.948 371.599 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 79.547
11 369.698 396.163 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 87.861
12 395.446 419.005 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 92.451
13 345.451 382.968 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.418
14 265.451 302.968 338.149 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.567
15 185.451 222.968 337.989 300.000 243.000 160.000 130.000 120.000 80.000 55.000 58.408
16 150.000 142.968 278.540 250.000 243.000 160.000 129.193 120.000 80.000 43.869 43.569
17 150.000 135.000 216.068 241.831 243.000 160.000 129.760 120.000 80.000 43.677 39.336
18 150.000 151.780 294.957 291.309 243.000 160.000 130.000 120.000 80.000 55.000 48.046
19 227.099 231.780 288.885 299.032 243.000 160.000 130.000 120.000 80.000 55.000 58.796
20 307.099 311.780 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 74.879
21 266.143 300.405 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.548
22 186.143 220.405 262.223 250.000 223.705 160.000 130.000 120.000 80.000 44.309 48.785
23 150.000 140.405 184.022 200.000 222.372 124.957 129.491 90.000 80.000 42.800 32.045
24 150.000 135.000 161.901 150.000 174.553 123.014 129.120 120.000 52.835 12.800 25.223
The fuel cost is 2.4688e+6 ($), and pollution emission is 3.1564e+5 (lb)

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15261

Table 6 Hourly generation (MW) schedule obtained from DEED using PSO-SQP for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.000 135.000 73.000 64.961 171.682 123.916 128.888 85.157 80.000 43.246 19.849
2 150.000 135.000 153.000 114.961 173.023 123.107 100.000 90.000 80.000 13.246 22.336
3 150.000 135.000 183.107 148.174 173.302 124.412 130.000 120.000 80.000 42.565 28.560
4 150.000 135.000 202.242 198.174 223.302 160.000 129.737 120.000 80.000 42.989 35.445
5 150.000 135.000 220.982 248.174 231.266 160.000 129.728 120.000 80.000 44.183 39.333
6 150.000 138.847 300.982 298.174 243.000 160.000 130.000 120.000 80.000 55.000 48.003
7 150.000 218.847 298.313 300.000 243.000 160.000 130.000 120.000 80.000 55.000 53.160
8 177.856 228.694 339.855 300.000 243.000 160.000 130.000 120.000 80.000 55.000 58.405
9 257.856 308.694 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.550
10 303.065 370.480 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 79.545
11 369.280 396.580 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 87.860
12 395.309 419.143 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 92.451
13 345.392 383.026 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.418
14 265.392 303.026 338.149 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.567
15 185.392 223.026 337.990 300.000 243.000 160.000 130.000 120.000 80.000 55.000 58.408
16 150.000 143.026 275.337 250.000 243.000 160.000 129.724 120.000 80.000 46.477 43.564
17 150.000 135.000 217.160 241.247 243.000 160.000 130.000 120.000 80.000 42.926 39.333
18 150.000 199.626 297.160 241.605 243.000 160.000 130.000 120.000 80.000 55.000 48.391
19 226.569 232.309 296.317 291.604 243.000 160.000 130.000 120.000 80.000 55.000 58.799
20 306.569 312.309 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 74.878
21 265.501 301.047 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 70.548
22 185.501 221.047 263.079 250.000 223.161 160.000 130.000 120.000 80.000 43.998 48.786
23 150.000 141.047 185.036 200.000 182.675 160.000 128.527 120.000 53.583 42.919 31.788
24 150.000 135.000 105.036 177.918 173.326 123.033 128.968 120.000 52.460 43.599 25.341
The fuel cost is 2.4701e+6 ($), and pollution emission is 3.1507e+5 (lb)

gives the 24-h power demands for the 5-, 10-, and 15-unit MODE (Basu 2014), DE-SQP (Elaiw et al. 2013), MOHD-
system. SAT (Zhang et al. 2015), MONNDE (Mason et al. 2018),
In the PSOCS, the population size N p = 30, the maximum PSO (Basu 2006), PSOAWL (Mason et al. 2017), PSO-SQP
number of the fitness function evaluations (FFEs) is set to a (Elaiw et al. 2013), PPSO (Ghasemi et al. 2019), PSO-CSC
fixed value (60,000) which is used to terminate runs of the (Huang et al. 2019). In the compared algorithms, NSGAII
PSOCS. We apply Eq. (9) to update the velocity of each par- and MODE belong to the evolutionary algorithms. DE-SQP,
ticle, and the parameters ωmax = 0.9, ωmin = 0.4 as shown MOHDE-SAT, and MONNDE are the hybrid methods com-
in (Elaiw et al. 2013). Considering the high dimension of the bining two or more approaches. PSO, PSOAWL, PSO-SQP,
DEED problem, the amplification factor F ∈ U (0.0,0.9), the PPSO, and PSO-CSC are the optimization methods based on
crossover rate C R ∈ U (0.8,1.0), here U (a, b) represents a the particle swarm optimization principle.
uniform distribution random number between a and b. The The settings of the main parameters for each compared
ε is defined as 0.1/g, and here g is the number of iterations. algorithm are chosen according to the suggestions in the cor-
For comparison purposes, we shall solve the DEED problem responding literature:
with a weighting factor ω = 0.5 in the objective function
f (p). The PSOCS is developed using C++ program, and the
system configuration is Pentium IV processor with 3.2 GHz (1) NSGAII (Basu 2008) The population size is selected as
speed and 1 GB RAM. To guarantee a fair comparison among 30, and the maximum number of FFEs is set as 60,000.
the algorithms, the experimental results are based on 25 The crossover probability of pc = 0.9 and a mutation
independent runs using random starting population for each probability of pm = 1/n (where n is the number of
run. decision variables) are used. Here, distribution indexes
We select several current reported algorithms to make a for crossover and mutation operators as ηc = 10 and
comparison with the PSOCS. They are NSGAII (Basu 2008), ηm = 10 are used, respectively.

123
15262 S. Qian et al.

Table 7 Hourly generation (MW) schedule obtained from DEED using PSOAWL for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.000 135.000 126.795 96.642 95.541 142.991 96.827 85.325 79.459 47.058 19.637
2 150.000 135.000 109.824 115.187 141.701 134.267 96.139 115.325 79.983 54.998 22.424
3 150.000 142.275 138.617 133.652 180.892 160.000 126.125 120.000 80.000 55.000 28.561
4 151.172 179.911 169.797 173.366 222.600 160.000 130.000 120.000 80.000 55.000 35.847
5 171.171 213.862 182.343 180.831 226.995 160.000 130.000 120.000 80.000 55.000 40.202
6 226.624 222.381 212.534 227.739 243.000 160.000 130.000 120.000 80.000 55.000 49.279
7 235.129 255.406 236.513 241.246 243.000 160.000 130.000 120.000 80.000 55.000 54.294
8 262.951 285.124 255.345 244.428 243.000 160.000 130.000 120.000 80.000 55.000 59.849
9 303.248 310.044 299.428 294.428 243.000 160.000 130.000 120.000 80.000 55.000 71.148
10 336.659 343.041 333.949 300.000 243.000 160.000 130.000 120.000 80.000 55.000 79.649
11 379.873 386.014 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 87.886
12 406.197 408.287 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 92.483
13 361.747 366.703 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.450
14 303.248 309.534 295.086 299.269 243.000 160.000 130.000 120.000 80.000 55.000 71.138
15 260.774 283.652 254.098 249.269 243.000 160.000 130.000 120.000 80.000 55.000 59.793
16 203.992 222.267 185.200 199.269 243.000 160.000 130.000 120.000 80.000 55.000 44.727
17 171.171 213.862 182.343 180.831 226.995 160.000 130.000 120.000 80.000 55.000 40.202
18 226.624 232.377 210.473 219.917 243.000 160.000 130.000 120.000 80.000 55.000 49.390
19 263.693 285.616 255.774 242.785 243.000 160.000 130.000 120.000 80.000 55.000 59.868
20 323.247 330.845 321.019 284.277 243.000 160.000 130.000 120.000 80.000 55.000 75.389
21 310.085 353.909 309.772 234.277 243.000 160.000 130.000 120.000 80.000 55.000 72.042
22 230.085 273.909 229.772 184.277 242.208 159.713 128.551 119.986 54.871 54.605 49.976
23 150.085 193.909 149.772 134.710 195.329 159.988 125.607 120.000 80.000 55.000 32.400
24 150.000 135.000 124.458 120.415 169.379 159.544 95.607 119.965 80.000 55.000 25.369
The fuel cost is 2.5490e+6 ($), and pollution emission is 2.9478e+5 (lb)

(2) MODE (Basu 2014) The population size, scaling factor, simulations, a network with 3 hidden neurons is imple-
crossover constant, and the maximum number of FFEs mented. Moreover, parameter tuning revealed that these
are selected as 30, 0.75, 1, and 60,000, respectively. parameters gave the best performance for MONNDE.
(3) DE-SQP (Elaiw et al. 2013) The control parameters of (6) PSO (Basu 2006) The population size is set as N = 30,
the DE are chosen as: N = 30, F = 0.423, and C R = and the maximum number of FFEs is set as 60,000.
0.885. The maximum number of FFEs is selected as The c1 and c2 = 0.205 are selected as the acceleration
600,000 according to the parameter setting in (Elaiw coefficients which gave the best overall performance.
et al. 2013). (7) PSOAWL (Mason et al. 2017) The population size is set
(4) MOHDE-SAT (Zhang et al. 2015) The main parameters as N = 30. The maximum number of FFEs is set as
of MOHDE-SAT are set as follows: The population size 300,000. The parameters c1 and c2 = 1.845 and c3 and
is set as 30, crossover rate is set as 0.3, and the archive c4 = 0.205 are selected as the acceleration coefficients.
set is also set as 30, the maximum number of FFEs is This parameter sweep showed the best value for the
set as 60,000. constriction to be χ = 0.72984 as is the case for the
(5) MONNDE (Mason et al. 2018) The DE differential standard PSO.
weight F = 0.5 and crossover probability C R = 0.9. (8) PSO-SQP (Elaiw et al. 2013) The population size N =
Chromosome mutation probability pm = 0.4, chro- 30, the maximum velocity V max = 0.5 × P max , the
mosomes C = 3. The maximum number of iterations V min = −0.5 × P min . The parameters ωmax = 0.9, and
I max = 5 × con (where con is the number of connec- ωmin = 0.4, c1 = c2 = 2.25. The maximum number
tion weights in the current network configuration), the of FFEs is set as 600,000 according to the parameter
number of DE agents N = con, (N = 4 if con < 4). setting in (Elaiw et al. 2013).
In all 5 and 10 generator unit simulations, a network (9) PPSO (Ghasemi et al. 2019) The swarm size is N = 30,
with 4 hidden neurons is implemented. For the 15-unit and the maximum number of FFEs is set as 60,000.

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15263

Table 8 Hourly generation (MW) schedule obtained from DEED using PSO-CSC for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.001 135.001 140.339 120.412 122.863 113.067 93.059 85.309 52.061 43.422 19.534
2 150.001 135.000 184.129 130.975 123.232 122.449 93.060 85.313 53.264 54.981 22.405
3 150.003 142.290 191.490 180.833 172.736 122.455 99.830 92.034 80.000 55.000 28.670
4 150.004 222.267 207.806 181.133 172.771 123.836 129.590 120.000 80.000 55.000 36.406
5 150.392 222.265 242.182 210.416 177.126 133.291 129.593 120.000 80.000 55.000 40.266
6 207.208 222.266 256.778 241.236 222.599 142.494 129.592 120.000 80.000 55.000 49.172
7 226.620 222.911 297.349 241.459 222.602 160.000 130.000 120.000 80.000 55.000 53.941
8 226.625 297.182 297.383 246.722 222.702 160.000 130.000 120.000 80.000 55.000 59.614
9 285.126 309.531 315.507 296.717 243.000 160.000 130.000 120.000 80.000 55.000 70.882
10 320.620 353.071 339.849 300.000 243.000 160.000 130.000 120.000 80.000 55.000 79.540
11 375.522 390.352 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 87.874
12 392.250 422.196 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 92.445
13 352.301 376.126 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.427
14 288.601 309.533 313.208 295.588 243.000 160.000 130.000 120.000 80.000 55.000 70.930
15 226.624 298.267 297.401 245.601 222.738 160.000 130.000 120.000 80.000 55.000 59.631
16 192.755 222.268 243.066 240.536 182.214 133.325 129.592 120.000 80.000 55.000 44.756
17 151.285 222.269 238.051 213.683 178.526 131.865 129.591 120.000 80.000 55.000 40.270
18 210.386 222.267 254.570 241.227 222.601 141.569 129.592 120.000 80.000 55.000 49.212
19 226.624 301.658 295.687 244.106 222.601 160.000 130.000 120.000 80.000 55.000 59.677
20 303.255 321.610 339.989 294.103 243.000 160.000 130.000 120.000 80.000 55.000 74.958
21 291.576 309.520 321.814 284.090 243.000 160.000 130.000 120.000 80.000 55.000 70.999
22 215.764 229.523 253.232 234.164 222.548 137.559 129.591 120.000 80.000 55.000 49.380
23 150.007 215.002 190.499 184.181 172.735 122.470 100.802 94.034 79.998 55.000 32.728
24 150.000 135.002 185.149 171.275 132.203 122.431 93.061 85.312 80.000 54.999 25.431
The fuel cost is 2.5247e+6 ($), and pollution emission is 3.0524e+5 (lb)

According to the PPSO algorithm, all other control individual runs is to check the statistical significance in the
parameters are obtained by choosing appropriate and differences in performance between different velocity update
proficient functions based on periodic sin and cos schemes.
trigonometric functions. The value of ω is set to zero Various velocity update approaches were suggested in
(Ghasemi et al. 2019). previous researches to improve the convergence character-
(10) PSO-CSC (Huang et al. 2019) The parameters of PSO- istics of the classical PSO (Mason et al. 2017). In this
CSC are set as follows: The swarm size is N = 30, section, in order to demonstrate the effects of the suggested
and the maximum number of FFEs is set as 60,000. velocity update methods for the DEED problem, three rep-
The acceleration coefficient is c1 = 1.5, c2 = 2.0, and resentative update approaches are also considered. They are:
the inertia weight χ = 0.4. The three parameters of Krohling-updating method (Krohling and dos Santos Coelho
the CSC are set to be that τ = 150, δ1 = 0.9, and 2006), Clerc-updating method (Clerc and Kennedy 2002),
δ2 = 1e–03. and Mason-updating method (Mason et al. 2017). The aver-
age fitness curves obtained by the PSOCS with corresponding
velocity update method over 25 independent runs under
5.2 Comparison of different velocity update 60,000 FFEs for each run are shown in Fig. 2. The best
methods (lowest), worst (highest), and average fitness values obtained
by the PSOCS in 25 independent runs are summarized in
In order to determine which velocity update methods in the Table 1. Observing Fig. 2, it is evident that the different
PSOCS performs best for the DEED problem, each variant of update methods affect significantly the convergence speed
PSOCS based on different velocity update equations is uti- of the PSOCS. The Clerc-updating method is the best one
lized to solve the 10-unit system. The fitness values obtained in the three approaches, while Krohling-updating method
by each variant are averaged over 25 independent runs, each presents the slowest convergence speed. Although the aver-
run consisting of 60,000 FFEs. The purpose of conducting 25

123
15264 S. Qian et al.

Table 9 Hourly generation (MW) schedule obtained from DEED using PPSO for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.000 135.000 149.279 120.414 122.866 101.197 93.059 85.311 52.288 46.170 19.585
2 150.000 135.000 173.070 130.836 122.870 114.423 93.060 85.312 72.875 54.999 22.447
3 150.006 148.528 186.760 180.830 172.733 122.450 99.600 90.801 80.000 55.000 28.708
4 153.830 222.267 204.986 180.847 172.733 123.186 129.590 120.000 80.000 55.000 36.439
5 186.295 222.267 236.579 188.862 172.783 129.263 129.591 120.000 80.000 55.000 40.639
6 226.623 222.266 253.162 238.724 216.351 135.736 129.592 120.000 80.000 55.000 49.453
7 226.625 240.143 283.855 241.251 222.600 157.036 129.593 120.000 80.000 55.000 54.102
8 226.630 302.688 293.662 245.093 222.611 160.000 130.000 120.000 80.000 55.000 59.685
9 294.330 309.533 308.382 294.765 243.000 160.000 130.000 120.000 80.000 55.000 71.011
10 308.380 365.158 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 79.538
11 373.548 392.321 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 87.869
12 413.075 401.437 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 92.512
13 353.500 374.930 340.000 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.429
14 297.481 309.534 305.540 294.502 243.000 160.000 130.000 120.000 80.000 55.000 71.056
15 226.937 302.438 294.197 244.510 222.605 160.000 130.000 120.000 80.000 55.000 59.687
16 223.973 222.439 236.593 223.123 177.361 131.054 129.590 120.000 80.000 55.000 45.133
17 190.074 222.267 233.443 190.032 172.737 127.538 129.591 120.000 80.000 55.000 40.680
18 226.625 222.266 250.566 236.189 221.569 135.642 129.591 120.000 80.000 55.000 49.449
19 226.626 301.627 295.245 244.576 222.600 160.000 130.000 120.000 80.000 55.000 59.674
20 303.249 322.996 338.152 294.576 243.000 160.000 130.000 120.000 80.000 55.000 74.972
21 300.576 309.534 305.858 291.139 243.000 160.000 130.000 120.000 80.000 55.000 71.107
22 226.619 229.535 252.570 241.156 202.979 140.053 129.590 120.000 80.000 55.000 49.502
23 150.001 216.169 187.310 191.157 172.734 122.451 99.622 90.303 80.000 55.000 32.746
24 150.000 136.171 184.041 171.191 132.536 122.130 93.061 85.313 79.995 54.999 25.437
The fuel cost is 2.5358e+6 ($), and pollution emission is 3.0375e+5 (lb)

age fitness value provided by the Mason-updating method considered to integrate the PSOCS for the 10-unit system.
is very close to that of the Clerc-updating method obtained The scenarios are as: DE/rand/1, DE/best/1, DE/current-
in the late stage, Mason-updating method presents a slower to-best/1, DE/rand/2, and DE/best/2. To make a conve-
convergence speed in the early stage of iteration. The con- nient description, PSOCS/R/1, PSOCS/B/1, PSOCS/CB/1,
clusions above are obvious in Table 1. Comparing the results PSOCS/R/2, and PSOCS/B/2 are, respectively, used to
obtained by the Clerc-updating method and those of the other denote PSOCS with DE/rand/1, PSOCS with DE/best/1,
two update methods shows that the Clerc-updating method PSOCS with DE/current-to-best/1, PSOCS with DE/rand/2,
provides a best average fitness value (i.e., 2.1175e+6) from and PSOCS with DE/best/2 in the figures and tables. The
Table 1. The standard deviations (SD) obtained by the three curves of average fitness from the five mutation schemes over
update methods through 25 independent runs are calculated 25 independent runs, 30,000 FFEs at each run, are displayed
as 3.4699e–4, 6.6360e–5, and 8.9775e–4, respectively. The in Fig. 3. Statistical results of the fitness values obtained by
smaller SD confirms the capability and reliability of the the corresponding PSOCS over 25 independent runs under
PSOCS with the Clerc-updating method to find the best solu- the five mutation strategies are presented in Table 2. From the
tion in 25 independent runs. So we can conclude that the convergence graphs in Fig. 3, it is evident that the PSOCS/R/1
Clerc-updating method can obtain the best results for solv- and PSOCS/CB/1 offer faster convergence speed than the
ing the DEED problem. other three variants of the PSOCS, which indicates that R/1
and CB/1 mutation strategies are better than other mutation
5.3 Comparison of different mutation strategies strategies for the DEED problem. The 2-th slowest are R/2
and B/2. Observing Table 2, it can be seen that PSOCS/CB/1
In this section, in order to demonstrate the effects of different performs best, and the second best is the PSOCS/R/1. We fur-
mutation strategies on the PSOCS based on the Clerc- ther investigate the performance of the PSOCS/R/1 and the
updating method for the DEED problem, five commonly PSOCS/CB/1, and it can be seen from Fig. 3 that PSOCS/R/1
used mutation schemes of DE to the cloned particles are has a faster convergence speed at the early stages of evo-

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15265

Table 10 Hourly generation (MW) schedule obtained from DEED using MONNDE for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.276 135.003 80.723 64.278 124.050 116.571 130.000 120.000 80.000 55.000 19.903
2 150.627 136.864 77.487 105.059 133.926 143.570 130.000 120.000 80.000 55.000 22.535
3 163.872 153.809 100.783 150.830 181.751 150.794 130.000 120.000 80.000 55.000 28.843
4 197.664 173.521 132.654 181.173 215.453 156.870 130.000 120.000 80.000 55.000 36.339
5 219.391 188.362 151.921 194.345 223.614 157.916 130.000 120.000 80.000 55.000 40.552
6 208.649 223.439 219.403 239.907 240.766 159.892 130.000 120.000 80.000 55.000 49.060
7 251.027 246.935 233.422 240.061 240.151 159.816 130.000 120.000 80.000 55.000 54.415
8 231.769 279.366 274.404 262.447 242.396 159.976 130.000 120.000 80.000 55.000 59.360
9 299.712 323.442 306.199 278.075 242.872 159.996 130.000 120.000 80.000 55.000 71.299
10 355.471 356.463 318.800 283.543 242.923 159.998 130.000 120.000 80.000 55.000 80.200
11 390.301 386.504 334.726 294.669 242.912 159.997 130.000 120.000 80.000 55.000 88.111
12 403.456 412.430 339.767 299.624 242.309 159.914 130.000 120.000 80.000 55.000 92.503
13 363.633 368.227 338.365 298.475 242.816 159.992 130.000 120.000 80.000 55.000 84.510
14 306.271 324.333 302.584 274.474 242.760 159.992 130.000 120.000 80.000 55.000 71.417
15 272.270 282.998 254.710 242.494 238.894 159.595 130.000 120.000 80.000 55.000 59.964
16 223.394 225.580 184.908 202.871 220.879 156.359 130.000 120.000 80.000 55.000 44.993
17 206.132 198.416 159.529 194.016 220.358 157.001 130.000 120.000 80.000 55.000 40.455
18 206.965 222.448 220.448 241.261 240.996 159.909 130.000 120.000 80.000 55.000 49.030
19 252.110 269.788 266.917 259.447 242.262 159.970 130.000 120.000 80.000 55.000 59.497
20 327.451 335.026 314.141 282.953 242.941 159.998 130.000 120.000 80.000 55.000 75.513
21 316.243 328.677 296.251 267.162 242.350 159.964 130.000 120.000 80.000 55.000 71.650
22 236.420 249.814 216.311 217.539 217.975 154.674 130.000 120.000 80.000 55.000 49.735
23 157.977 191.954 136.749 168.040 181.192 143.530 130.000 120.000 80.000 55.000 32.445
24 154.060 143.690 113.832 121.771 158.621 132.543 130.000 120.000 80.000 55.000 25.520
The fuel cost is 2.5600e+6 ($), and pollution emission is 2.9782e+5 (lb)

lutions, while it presents a slower speed at the late stages. that the PSOCS/R/1orCB/1 can obtain better (smaller) statis-
However, PSOCS/CB/1 presents a slower convergence speed tical performance in terms of both the best, worst, and average
at the early states, but it can maintain a continuous descent fitness values compared with the other mutation schemes.
during the whole process. Thus, we can conclude that the PSOCS with R/1orCB/1
Based on the above consideration, we further study the mutation strategy can obtain the best results for the DEED
performance of the PSOCS with R/1, CB/1, and the pro- problem.
posed R/1orCB/1 under 60,000 FFEs for the 10-unit system
in order to judge the convergence behaviors of the three 5.4 Comparison with previously published results
mutation schemes. Figure 4 displays the curves of aver-
age fitness obtained in 25 independent runs by PSOCS/R/1, In this section, we present the experimental results obtained
PSOCS/CB/1, and PSOCS/R/1orCB/1 on the 10-unit sys- by the PSOCS on the DEED problem under three kinds of
tem. Table 3 lists the statistical information attained by each test cases (5-, 10-, and 15-unit systems). For comparison pur-
algorithm. Comparing PSOCS/R/1orCB/1, PSOCS/R/1, and poses, we shall solve each test case with weight ω = 0.5 for
PSOCS/CB/1 in Fig. 4, it is clear that PSOCS/R/1orCB/1 DEED model Eq. (13) by using the single-objective optimiza-
converges smoothly to the optimum solution without any tion algorithms: DE-SQP, PSOAWL, PSO-SQP, and PSOCS.
sudden oscillations. However, PSOCS/R/1 entraps into local For the multiobjective optimization algorithms, for example,
optima and fails to reach the global optimum solution even NSGAII, MODE, MOHDE-SAT, MONNDE, PPSO, PSO,
if 60,000 FFEs. Moreover, PSOCS/R/1orCB/1 gets to the and PSO-CSC, the fuzzy-based approach (Wu et al. 2010)
global solution after fewer numbers of iterations in compari- is used to pick out the best compromise solution in the
son with the PSOCS/CB/1. This indicates that the R/1orCB/1 Pareto-optimal solutions obtained by these multiobjective
mutation strategy proposed in Sect. 4.5 is effective in improv- algorithms. All algorithms are performed 25 independent
ing the solution quality and the convergence rate because of runs with random population under 60,000 fitness function
its enhanced exploiting capability. In Table 3, it is clear to see evaluations (FFEs) in each run in order to obtain the statis-

123
15266 S. Qian et al.

Table 11 Hourly generation (MW) schedule obtained from DEED using MOHDE-SAT for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 150.000 135.000 89.653 120.062 125.615 127.555 94.166 120.000 80.000 13.506 19.557
2 150.000 135.000 94.699 112.685 166.204 134.260 124.166 120.000 80.000 15.347 22.361
3 150.000 136.792 138.117 132.452 207.810 160.000 129.599 120.000 80.000 31.718 28.488
4 150.308 170.868 177.259 180.255 224.746 160.000 130.000 120.000 79.994 48.300 35.73
5 155.885 209.911 185.270 181.474 242.713 160.000 130.000 120.000 80.000 54.792 40.045
6 224.021 222.777 211.851 230.597 243.000 160.000 130.000 119.998 80.000 55.000 49.244
7 232.619 248.431 245.336 241.904 243.000 159.973 130.000 120.000 79.918 55.000 54.181
8 242.716 284.643 261.977 258.219 243.000 160.000 130.000 120.000 80.000 55.000 59.555
9 303.062 309.609 294.650 299.813 243.000 160.000 130.000 120.000 80.000 55.000 71.134
10 334.807 338.759 340.000 300.000 243.000 160.000 130.000 120.000 80.000 54.998 79.564
11 379.135 387.198 339.597 300.000 242.988 159.977 130.000 120.000 80.000 55.000 87.895
12 408.234 406.281 340.000 300.000 242.989 159.999 130.000 120.000 80.000 55.000 92.503
13 360.783 367.707 339.954 300.000 243.000 160.000 130.000 120.000 80.000 55.000 84.444
14 302.140 313.007 293.222 299.103 243.000 160.000 130.000 119.999 80.000 54.689 71.16
15 250.058 282.665 265.960 249.103 243.000 160.000 130.000 119.985 80.000 54.863 59.634
16 218.084 221.602 188.653 203.501 243.000 160.000 130.000 120.000 80.000 33.987 44.827
17 159.017 216.308 186.888 181.011 243.000 160.000 129.989 120.000 80.000 43.882 40.095
18 226.667 223.550 208.819 231.011 242.379 160.000 129.973 119.898 80.000 54.999 49.296
19 250.659 270.935 273.056 252.866 243.000 160.000 130.000 119.998 80.000 54.990 59.504
20 313.900 321.211 325.188 299.140 243.000 160.000 130.000 120.000 80.000 54.646 75.085
21 303.480 305.818 304.972 292.915 243.000 160.000 130.000 120.000 80.000 54.914 71.099
22 223.892 225.818 226.079 242.915 234.511 160.000 130.000 120.000 80.000 33.993 49.208
23 150.000 146.453 146.079 192.915 210.343 160.000 130.000 120.000 79.972 28.105 31.867
24 150.000 135.000 106.583 142.915 172.480 157.966 129.479 120.000 80.000 14.812 25.235
The fuel cost is 2.5279e+6 ($), and pollution emission is 2.9776e+5 (lb)

tical results. For the DE-SQP, PSO-SQP, and PSOAWL, the mise solution obtained by PSOCS dominates that of most
number of FFEs is set the same as in (Elaiw et al. 2013; Ma other algorithms in Fig. 5b.
et al. 2017). Table 4 provides the comparison of the fuel cost As far as the computational cost (i.e., the number of FFEs)
and emission for the 5-, 10-, and 15-unit systems using the is concerned, PSOCS requires 60,000 FFEs to obtain the
PSOCS with respect to the most recent methods reported in reported results in Table 4, compared against 600,000 FFEs
the literature. According to the comparison results in Table 4, used by DE-SQP and PSO-SQP (Elaiw et al. 2013), 300,000
we can see that the PSOCS is efficient for solving the three FFEs by PSOAWL (Ma et al. 2017). From the statistical
test systems. For ω = 0.5, it is clear to see that the PSOCS information above, it can be seen that the PSOCS can obtain
can obtain better compromise solution in terms of both the the reported results in Table 4 in fewer numbers of FFEs. So
minimum fuel cost and the minimum emission compared we can conclude that the computational cost of the PSOCS
with other algorithms. Moreover, one can see from Table 4 is less than that of the aforementioned approaches.
that the small SD obtained by PSOCS indicates the opti- The best solutions of the DEED problem consisting of ten
mization performance of PSOCS is consistent over 25 runs. units with ω = 0.5 using the DE-SQP, PSO-SQP, PSOAWL,
For visualization, the best compromise solution obtained by and PSOCS methods are given in Tables 5, 6, 7, 8, 9, 10,
the compared algorithms is also plotted in Fig. 5a, b for 11, 12, respectively. The results of the 10-unit system for the
the 10-unit and 15-unit cases, respectively. It can be seen DEED problem show that our proposal PSOCS is efficient.
from Fig. 5a that the PSOCS can obtain the similar com-
promise solution with the MOHD-SAT and PSOAWL for
10-unit case. When compared to other approaches except 6 Conclusion and further work
MOHDE-SAT and POSAWL, PSOCS performs better than
other approaches. However, for 15-unit case, the compro- Dynamic economic emission dispatch (DEED) problem is
an extremely challenging research topic in the field of opti-

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15267

Table 12 Hourly generation (MW) schedule obtained from DEED using PSOCS for 10-unit system
t P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Loss

1 151.546 135.531 109.679 106.636 120.628 127.535 77.137 118.816 57.762 50.354 19.624
2 151.377 135.180 103.447 111.763 139.096 142.975 101.277 119.499 76.990 50.815 22.419
3 152.701 136.598 135.380 139.915 178.470 160.000 129.249 120.000 79.799 54.422 28.534
4 152.757 150.295 185.302 186.165 222.832 159.926 129.954 119.898 79.888 54.590 35.607
5 198.703 156.725 196.801 179.807 243.000 160.000 130.000 120.000 80.000 55.000 40.037
6 224.658 217.784 227.086 219.684 243.000 160.000 130.000 120.000 80.000 55.000 49.213
7 226.300 222.490 270.757 248.719 242.934 159.880 129.961 119.935 79.939 54.937 53.852
8 241.992 270.886 283.470 253.369 242.622 159.623 129.480 119.606 79.647 54.713 59.407
9 293.463 315.050 298.548 300.000 243.000 160.000 130.000 120.000 80.000 55.000 71.062
10 333.903 339.788 339.990 299.967 242.985 159.978 129.994 119.993 79.975 54.993 79.565
11 378.744 387.204 339.996 299.996 242.997 159.997 129.998 119.996 79.962 54.995 87.884
12 405.617 409.031 339.983 299.984 242.991 159.972 129.965 119.977 79.991 54.975 92.485
13 359.087 369.459 339.957 299.998 243.000 159.997 129.987 119.994 79.980 54.986 84.444
14 289.837 307.024 310.997 299.865 242.851 159.861 129.870 119.871 79.869 54.870 70.914
15 237.840 278.021 270.440 262.590 242.694 159.926 129.916 119.634 79.895 54.458 59.414
16 174.154 223.209 197.023 216.785 243.000 160.000 130.000 120.000 80.000 54.216 44.387
17 171.745 174.755 190.954 194.600 242.980 159.919 129.987 119.967 79.971 54.972 39.849
18 210.825 223.547 216.912 240.091 242.542 159.551 129.631 119.644 79.656 54.693 49.092
19 236.568 248.364 280.264 281.862 243.000 160.000 130.000 120.000 80.000 55.000 59.058
20 305.883 313.979 339.776 299.262 242.999 159.998 129.995 120.000 80.000 54.996 74.889
21 299.393 338.321 286.559 283.824 242.898 159.846 129.973 119.908 79.870 54.901 71.494
22 220.476 259.052 209.193 234.366 233.235 156.634 129.921 119.339 71.956 43.408 49.580
23 157.317 187.619 132.210 184.366 197.246 149.799 129.611 101.201 78.017 46.979 32.364
24 152.156 135.513 112.367 140.723 166.796 151.990 127.522 94.339 78.744 49.186 25.335
The fuel cost is 2.5269e+6 ($), and pollution emission is 2.9800e+5 (lb)

mization, because the DEED problem is a nonsmooth and Some further works relevant to this paper need to be done
nonconvex optimization problem with multiple local optimal in the future. It is future work that the developed framework
points, which makes finding the global optimal challenging. can be introduced into other searching mechanisms, such as
In this paper, an improved particle optimization algorithm simulated annealing, ant colony, fish swarm. We only select
based on CS principle (PSOCS) is proposed to solve the a fixed set of relevant parameters in the PSOCS and do not
DEED problem. In the algorithm design, a repair scheme make any effort in designing adaptivity for the proposed
is adopted to handle the power balance equality constraints, algorithm. The question that how to develop an adaptive
improving the feasibility rate of the evolution population. mechanism will also be addressed in the future. Finally, all
Based on the operator, the degree that an infeasible particle studies only focus on the DEED model in this paper. How-
violation constraints can be reduced fast, and that particle ever, relevant studies on large-scale problems of economic
even is converted into a feasible one at once. In order to dispatch should be addressed by using the proposed algo-
compensate the convergence speed and supply more valu- rithm in the future.
able information to adjust the velocity of the particles in the
standard PSO, the excellent particles in P best are chosen to Acknowledgements The authors acknowledge the support from the
National Natural Science Foundation of China under Grants 61762001,
clone and then mutated by a proposed mutation scheme. The Creative Research Groups of the Natural Science Foundation of
proposed algorithm is used as a searching method to handle Guizhou of China under Grants Qian Jiao he KY zi 2019069 and
DEED problem. According to the experimental results, with 2018034, Provincial Science and Technology Foundation of Guizhou of
three test systems (5-, 10-, and 15-unit system) taken into China under Grants Qian ke he LH zi 20177047, and Innovation Foun-
dation of Nanjing Institute of Technology under Grants CKJC201603.
consideration under 60,000 fitness function evaluations, we
can find that the proposed algorithm can obtain superior per-
formance with respect to the most recent methods reported
in the literature.

123
15268 S. Qian et al.

Compliance with ethical standards Informed consent Informed consent was obtained from all individual
participants included in the study.
Conflict of interest The authors declare that they have no conflict of
interest.

Ethical standard All procedures performed in studies involving human


participants were in accordance with the ethical standards of the insti- A Appendix
tutional and/or national research committee and with the 1964 Helsinki
Declaration and its later amendments or comparable ethical standards.
This article does not contain any studies with animals performed by any See Tables 13, 14, 15, 16, 17, 18, 19.
of the authors.

Table 13 5-unit generator coefficients


Unit p min p max UR DR a b c d e α β γ η δ
MW MW MW/h MW/h $/h $/MWh $/MW2 h $/h rad/MW lb/h lb/MWh lb/MW2 h lb/h 1/MW

1 10 75 30 30 25 2.0 0.0080 100 0.042 80 − 0.805 0.0180 0.6550 0.02846


2 20 125 30 30 60 1.8 0.0030 140 0.040 50 − 0.555 0.0150 0.5773 0.02446
3 30 175 40 40 100 2.1 0.0012 160 0.038 60 − 1.355 0.0105 0.4968 0.02270
4 40 250 50 50 120 2.0 0.0010 180 0.037 45 − 0.600 0.0080 0.486 0.01948
5 50 300 50 50 40 1.8 0.0015 200 0.035 30 − 0.555 0.0120 0.5035 0.02075

Table 14 5-unit B matrix


0.000049 0.000014 0.000015 0.000015 0.00002
0.000014 0.000045 0.000016 0.00002 0.000018
0.000015 0.000016 0.000039 0.00001 0.000012
0.000015 0.00002 0.00001 0.00004 0.000014
0.00002 0.000018 0.000012 0.000014 0.000035

Table 15 10-unit generator coefficients


Unit p min p max UR DR a b c d e α β γ η δ
MW MW MW/h MW/h $/h $/MWh $/MW2 h $/h rad/MW lb/h lb/MWh lb/MW2 h lb/h 1/MW

1 150 470 80 80 786.7988 38.5397 0.1524 450 0.041 103.3908 −2.4444 0.0312 0.5035 0.0207
2 135 470 80 80 451.3251 46.1591 0.1058 600 0.036 103.3908 −2.4444 0.0312 0.5035 0.0207
3 73 340 80 80 1049.9977 40.3965 0.0280 320 0.028 300.3910 −4.0695 0.0509 0.4968 0.0202
4 60 300 50 50 1243.5311 38.3055 0.0354 260 0.052 300.3910 −4.0695 0.0509 0.4968 0.0202
5 73 243 50 50 1658.5696 36.3278 0.0211 280 0.063 320.0006 −3.8132 0.0344 0.4972 0.0200
6 57 160 50 50 1356.6592 38.2704 0.0179 310 0.048 320.0006 −3.8132 0.0344 0.4972 0.0200
7 20 130 30 30 1450.7045 36.5104 0.0121 300 0.086 330.0056 −3.9023 0.0465 0.5163 0.0214
8 47 120 30 30 1450.7045 36.5104 0.0121 340 0.082 330.0056 −3.9023 0.0465 0.5163 0.0214
9 20 80 30 30 1455.6056 39.5804 0.1090 270 0.098 350.0056 −3.9524 0.0465 0.5475 0.0234
10 10 55 30 30 1469.4026 40.5407 0.1295 380 0.094 360.0012 −3.9864 0.0470 0.5475 0.0234

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15269

Table 16 10-unit B matrix


(×10−6 ) 49 14 15 15 16 17 17 18 19 20
14 45 16 16 17 15 15 16 18 18
15 16 39 10 12 12 14 14 16 16
15 16 10 40 14 10 11 12 14 15
16 17 12 14 35 11 13 13 15 16
17 15 12 10 11 36 12 12 14 15
17 15 14 11 13 12 38 16 16 18
18 16 14 12 13 12 16 40 15 16
19 18 16 14 15 14 16 15 42 19
20 18 16 15 16 15 18 16 19 44

Table 17 15-unit generator coefficients


Unit p min p max UR DR a b c d e α β γ η δ
MW MW MW/h MW/h $/h $/MWh $/MW2 h $/h rad/MW lb/h lb/MWh lb/MW2 h lb/h 1/MW

1 150 455 80 120 671 10.1 0.000299 100 0.042 80.0000 − 0.8050 0.0180 0.655 0.02846
2 150 455 80 120 574 10.2 0.000183 140 0.040 50.0000 − 0.5550 0.0150 0.5773 0.02446
3 20 130 130 130 374 8.8 0.001126 160 0.038 60.0000 − 1.3550 0.0105 0.4968 0.02270
4 20 130 130 130 374 8.8 0.001126 180 0.037 45.0000 − 0.6000 0.0080 0.4860 0.01948
5 150 470 80 120 461 10.4 0.000205 200 0.035 30.0000 − 0.5550 0.0120 0.5035 0.02075
6 135 460 80 120 630 10.1 0.000301 450 0.041 103.3908 − 2.4444 0.0312 0.5035 0.02070
7 135 465 80 120 548 9.8 0.000364 600 0.036 103.3908 − 2.4444 0.0312 0.5035 0.02070
8 60 300 65 100 227 11.2 0.000338 320 0.028 300.3910 − 4.0695 0.0509 0.4968 0.02020
9 25 162 60 100 173 11.2 0.000807 260 0.052 300.3910 − 4.0695 0.0509 0.4968 0.02020
10 25 160 60 100 175 10.7 0.001203 280 0.063 320.0006 − 3.8132 0.0344 0.4972 0.02000
11 20 80 80 80 186 10.2 0.003586 310 0.048 320.0006 − 3.8132 0.0344 0.4972 0.02000
12 20 80 80 80 230 9.9 0.005513 300 0.086 330.0056 − 3.9023 0.0465 0.5163 0.02140
13 25 85 80 80 225 13.1 0.000371 340 0.082 330.0056 − 3.9023 0.0465 0.5163 0.02140
14 15 55 55 55 309 12.1 0.001929 270 0.098 350.0056 − 3.9524 0.0465 0.5475 0.02340
15 15 55 55 55 323 12.4 0.004447 380 0.094 360.0012 − 3.9864 0.0470 0.5475 0.02340

Table 18 15-unit B matrix


(×10−5 ) 14 12 7 −1 −3 −1 −1 −1 −3 5 −3 −2 4 3 −1
12 15 13 0 −5 −2 0 1 −2 −4 −4 0 4 10 −2
7 13 76 − 1 − 13 −9 −1 0 −8 − 12 − 17 0 − 26 111 − 28
−1 0 −1 34 −7 −4 11 50 29 32 − 11 0 1 1 −6
−3 −5 − 13 −7 90 14 −3 − 12 − 10 − 13 7 −2 −2 − 24 −3
−1 −2 −9 −4 14 16 0 −6 −5 −8 11 −1 −2 − 17 3
−1 0 −1 11 −3 0 15 17 15 9 −5 7 0 −2 −8
−1 1 0 50 − 12 −6 17 168 82 79 − 23 − 36 1 5 − 78
−3 −2 −8 29 − 10 −5 15 82 129 116 − 21 − 25 7 − 12 − 72
−5 −4 − 12 32 − 13 −8 9 79 116 200 − 27 − 34 9 − 11 − 88
−3 −4 − 17 − 11 7 11 −5 − 23 − 21 − 27 140 1 4 − 38 168
−2 0 0 0 −2 −1 7 − 36 − 25 − 34 1 54 −1 −4 28
4 4 − 26 1 −2 −2 0 1 7 9 4 −1 103 − 101 28
3 10 111 1 − 24 − 17 −2 5 − 12 − 11 − 38 − 4 − 101 578 − 94
−1 −2 − 28 − 26 −3 3 −8 − 78 − 72 − 88 168 28 28 − 94 1283

123
15270 S. Qian et al.

Table 19 24-h power demands Deb K (2000) An efficient constraint handling method for genetic algo-
rithms. Comput Methods Appl Mech Eng 186(2):311–338
Hour 5 Units (MW) 10 Units (MW) 15 Units (MW) Djemame S, Batouche M, Oulhadj H et al (2019) Solving reverse emer-
gence with quantum PSO application to image processing. Soft
1 410 1036 1171
Comput 24(23):6921–6935
2 435 1110 1254 dos Santos Coelho L, Mariani VC (2006) Combining of chaotic differ-
3 475 1258 1415 ential evolution and quadratic programming for economic dispatch
4 530 1406 1581 optimization with valve-point effect. IEEE Trans Power Syst
21(2):989–996
5 558 1480 1664 Elaiw AM, Xia X, Shehata AM (2013) Hybrid DE-SQP and hybrid
6 608 1628 1829 PSO-SQP methods for solving dynamic economic emission dis-
7 626 1702 1909 patch problem with valve-point effects. Electr Power Syst Res
103(8):192–200
8 654 1776 1992 Elsakaan AA, El-Sehiemy RA, Kaddah S et al (2018) An enhanced
9 690 1924 2152 moth-flame optimizer for solving non-smooth economic dispatch
10 704 2022 2254 problems with emissions. Energy 157(8):1063–1078
Ghasemi M, Akbari E, Rahimnejad A et al (2019) Phasor particle swarm
11 720 2106 2344
optimization: a simple and efficient variant of PSO. Soft Comput
12 740 2150 2394 23(19):9701–9718
13 704 2072 2304 Gholamghasemi M, Akbari E, Asadpoor M et al (2019) A new solution
14 690 1924 2152 to the non-convex economic load dispatch problems using phasor
particle swarm optimization. Appl Soft Comput 79(6):111–124
15 654 1776 1992 Guo Y, Yang Z, Wang C et al (2017) Cultural particle swarm opti-
16 580 1554 1745 mization algorithms for uncertain multi-objective problems with
17 558 1480 1664 interval parameters. Nat Comput 16:527–528
Huang H, Lv L, Ye S et al (2019) Particle swarm optimization with con-
18 608 1628 1829 vergence speed controller for large-scale numerical optimization.
19 654 1776 1992 Soft Comput 23(12):4421–4437
20 704 1972 2204 Jebaraj L, Venkatesan C, Soubache I et al (2017) Application of dif-
ferential evolution algorithm in static and dynamic economic or
21 680 1924 2148
emission dispatch problem: a review. Renew Sustain Energy Rev
22 605 1628 1828 77(2017):1206–1220
23 527 1332 1506 Kalita DJ, Singh S (2020) SVM hyper-parameters optimization using
24 463 1184 1337 quantized multi-PSO in dynamic environment. Soft Comput
25(24):1225–1241
Kennedy J, Eberhart R (1995) Particle swarm optimization. In: Proceed-
ings of IEEE international conference on neural networks, 1995,
vol 4, pp 1942–1948
References Krohling RA, dos Santos Coelho L (2006) Coevolutionary particle
swarm optimization using Gaussian distribution for solving con-
Aragón VS, Esquivel SC, Coello Coello CA (2015) An immune algo- strained optimization problems. IEEE Trans Syst Man Cybern Part
rithm with power redistribution for solving economic dispatch B Cybern 36(6):1407–1423
problems. Information Sciences 24(262):609–632 Ma HP, Yang ZL, You PC et al (2017) Multi-objective biogeography-
Bahmanifirouzi B, Farjah E, Niknam T (2012) Multi-objective stochas- based optimization for dynamic economic emission load dispatch
tic dynamic economic emission dispatch enhancement by fuzzy considering plug-in electric vehicles charging. Energy 135:102–
adaptive modified theta particle swarm optimization. J Renew Sus- 111
tain Energy 4(2):23–45 Mason K, Duggan J, Howley E (2017) Multi-objective dynamic
Basu M (2006) Particle swarm optimization based goal-attainment economic emission dispatch using particle swarm optimisation
method for dynamic economic emission dispatch. Electr Power variants. Neurocomputing 270(12):188–197
Compon Syst 34(9):1015–1025 Mason K, Duggan J, Howley E (2018) A multi-objective neural network
Basu M (2008) Dynamic economic emission dispatch using nondomi- trained with differential evolution for dynamic economic emission
nated sorting genetic algorithm-II. Int J Electr Power Energy Syst dispatch. Electr Power Energy Syst 100(6):201–221
30(2):140–149 Neto J, Reynoso-Meza G, Ruppel T et al (2017) Solving non-smooth
Basu M (2014) Multi-objective differential evolution for dynamic economic dispatch by a new combination of continuous GRASP
economic emission dispatch. Int J Emerg Electr Power Syst algorithm and differential evolution. Int J Electr Power Energy
15(2):141–150 Syst 84:13–24
Cardoso Bora T, Cocco Mariani V, dos Santos Coelho L (2019) Niknam T, Golestaneh F, Sadeghi MS (2012) θ-multiobjective teach-
Multi-objective optimization of the environmental-economic dis- ing learning-based optimization for dynamic economic emission
patch with reinforcement learning based on non-dominated sorting dispatch. IEEE Syst J 6(2):341–352
genetic algorithm. Appl Therm Eng 146(1):688–700 Niu Q, Zhang HY, Li K et al (2014) An efficient harmony search
Clerc M, Kennedy J (2002) The particle swarm-explosion, stability, and with new pitch adjustment for dynamic economic dispatch. Energy
convergence in a multidimensional complex space. IEEE Trans 65(1):25–43
Evol Comput 6(1):58–73 Pandit N, Tripathi A, Tapaswi S et al (2012) An improved bacterial
De Castro LN, Von Zuben FJ (2002) Learning and optimization using foraging algorithm for combined static/dynamic environmental
the clonal selection principle. IEEE Trans Evol Comput 6(3):239– economic dispatch. Appl Soft Comput 12(11):3500–3513
251

123
An improved particle swarm optimization with clone selection principle for dynamic economic… 15271

Panigrahi BK, Yadav SR, Agrawal S et al (2007) A clonal algorithm to Wu LH, Wang YN, Yuan XF et al (2010) Environmental/economic
solve economic load dispatch. Electr Power Syst Res 77(10):1381– power dispatch problem using multi-objective differential evolu-
1389 tion algorithm. Electr Power Syst Res 80(9):1171–1181
Ross D, Kim S (2007) Dynamic economic dispatch of generation. IEEE Yang Z, Kang L, Niu Q et al (2014) A self-learning TLBO based
Trans Power Appar Syst PAS–99(6):2060–2068 dynamic economic/environmental dispatch considering multiple
Sengupta S, Basak S, Peters R (2019) Particle swarm optimization: a plug-in electric vehicle loads. J Mod Power Syst Clean Energy
survey of historical and recent developments with hybridization 2(4):298–307
perspectives. Mach Learn Knowl Extr 1(1):157–191 Yuan G, Yang W (2019) Study on optimization of economic dispatching
Shen X, Zou D, Duan N et al (2019) An efficient fitness-based differ- of electric power system based on hybrid intelligent algorithms
ential evolution algorithm and a constraint handling technique for (PSO and AFSA). Energy 183(9):926–935
dynamic economic emission dispatch. Energy 186(2):115–121 Zhai Y, Liao X, Mu N et al (2019) A two-layer algorithm based on PSO
Silva Chavez J, Zamora-Mendez A, Arrieta Paternina M (2019) A for solving unit commitment problem. Soft Comput 24(3):15–24
hybrid optimization framework for the non-convex economic dis- Zhang Y, Gong DW, Geng N et al (2014) Hybrid bare-bones PSO for
patch problem via meta-heuristic algorithms. Electr Power Syst dynamic economic dispatch with valve-point effects. Appl Soft
Res 177:456–467 Comput 18(C):248–260
Singh D, Dhillon J (2019) Ameliorated grey wolf optimization for eco- Zhang H, Yue D, Xie X et al (2015) Multi-elite guide hybrid differ-
nomic load dispatch problem. Energy 169(2):398–419 ential evolution with simulated annealing technique for dynamic
Wang L, Cao C, Xu Z et al (2012) An improved particle swarm algorithm economic emission dispatch. Appl Soft Comput 34(C):312–323
based on cultural algorithm for constrained optimization. In: Hepzi
T (ed) Knowledge discovery and data mining, vol 135. Advances
in intelligent and soft computing. Springer, Berlin, pp 453–460 Publisher’s Note Springer Nature remains neutral with regard to juris-
Wood A, Wollenberg B (1996) Power generation operation and control. dictional claims in published maps and institutional affiliations.
Fuel Energy Abstr 37(3):90–93

123

You might also like