Energy xxx (2014) 1e10

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Parameter identification of solar cells using artificial bee colony

optimization
Diego Oliva a, Erik Cuevas b, *, Gonzalo Pajares a
a
Dpt. Ingeniería del Software e Inteligencia Artificial, Facultad Informa

tica, Universidad Complutense, 28040 Madrid, Spain
b
Departamento de Electro
nica, Universidad de Guadalajara, CUCEI, Av. Revolucio
n 1500, Guadalajara, Jal, Mexico

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

Article history: In order to improve the performance of solar energy systems, accurate modeling of current vs. voltage (I
Received 30 October 2013 eV) characteristics of solar cells has attracted the attention of various researches. The main drawback in
Received in revised form accurate modeling is the lack of information about the precise parameter values which indeed charac-
2 May 2014
terize the solar cell. Since such parameters cannot be extracted from the datasheet specifications, an
Accepted 3 May 2014
Available online xxx
optimization technique is necessary to adjust experimental data to the solar cell model. Considering the I
eV characteristics of solar cells, the optimization task involves the solution of complex non-linear and
multi-modal objective functions. Several optimization approaches have been proposed to identify the
Keywords:
Solar cell modeling
parameters of solar cells. However, most of them obtain sub-optimal solutions due to their premature
Photo voltaic cells convergence and their difficulty to overcome local minima in multi-modal problems. This paper proposes
Artificial bee colony the use of the ABC (artificial bee colony) algorithm to accurately identify the solar cells' parameters. The
ABC algorithm is an evolutionary method inspired by the intelligent foraging behavior of honey bees. In
comparison with other evolutionary algorithms, ABC exhibits a better search capacity to face multi-
modal objective functions. In order to illustrate the proficiency of the proposed approach, it is
compared to other well-known optimization methods. Experimental results demonstrate the high per-
formance of the proposed method in terms of robustness and accuracy.
© 2014 Elsevier Ltd. All rights reserved.

1. Introduction of their parameter values. For the mathematical model, the Current
vs. Voltage (IeV) characteristics that rule the behavior of a solar cell
The increase in the cost of fossil fuels and their probable is considered. Several approaches have been proposed in order to
depletion, air pollution, global warming phenomenon, and severe model such a behavior from different point of views [7e12].
environmental laws have resulted in renewable energy sources In practical terms, there exist two equivalent electronic circuits
gaining the attention of many nations to produce electricity. Solar that model the behavior of a solar cell. Such circuits are known as
energy is one of the most promising renewable sources that is SD (single diode) and DD (double diode) models [13]. Irrespective
currently being used worldwide to contribute to meeting rising of the model selected, it is necessary to estimate or identify all its
demands for electric power. It has been reported that solar PV parameters such as photo-generated current, diode saturation
(photovoltaic) is the fastest growing power-generation technology current, series resistance, and diode ideality factor. Depending on
in the world, with an annual average increase of 50% between 2004 the model (SD or DD), two different sets of parameters must be
and 2011 [1]. PV is not only capable of directly converting solar identified: five for the SD and seven for the DD. The main problem
energy to electricity but also is an emission-free distributed gen- is to identify the optimal parameter values which, when applied to
eration unit that would supply power at the load site. the selected model, produce the best possible approximation to the
Solar cell accurate modeling has received significant attention in experimental data obtained by the true solar cell [13].
recent years [2e6]. The modeling of PV cells consists in two steps: The methods employed to solve the problem of PV parameter
the mathematical model formulation and the accurate estimation identification can be divided in two groups: deterministic and
heuristic. Some examples of deterministic methods involve
methods such as least squares [14], Lambert W-functions [15], and
* Corresponding author. Tel.: þ52 33 1378 5900x7715. the iterative curve fitting [16]. Deterministic techniques impose
E-mail addresses: [email protected] (D. Oliva), [email protected], cuevas@ several model restrictions such as convexity and differentiability in
inf.fu-berlin.de (E. Cuevas), [email protected] (G. Pajares).

http://dx.doi.org/10.1016/j.energy.2014.05.011
0360-5442/© 2014 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
2 D. Oliva et al. / Energy xxx (2014) 1e10

order to be correctly applied [24]. Therefore, they are very sensitive objective function evaluates the matching quality between a candi-
to the initial solution, and most often lead to local optima. As an date solution and the experimental data. Guided by the values of this
alternative to deterministic-based techniques, the problem of PV objective function, the set of encoded candidate solutions is evolved
parameter identification has also been handled through heuristic by using the operators defined by ABC so that the parameters that
methods. In general, they have demonstrated that they deliver produce the best possible approximation to the IeV measurements
better results than those based on deterministic approaches obtained by the true solar cell can be found. In order to illustrate the
considering accuracy and robustness [12,13,17e27]. In the litera- proficiency of the proposed approach, it is compared to other well-
ture, several heuristic approaches have been proposed in order to known optimization methods. Experimental evidence shows that
solve the problem of solar cell parameter identification. Such ABC exhibits no sensitivity to noisy conditions and high performance
methods include GA (genetic algorithms) [17,18,24], PSO (particle in terms of robustness and accuracy.
swarm optimization) [12,19], SA (simulated annealing) [20], HS The remainder of the paper is organized as follows. In Section 2,
(harmony search) [13], BFA (Bacterial Foraging Algorithm) [21], the problem of solar cell identification is defined. Section 3 de-
TBLO (teaching-learning based optimization) [23] and BMO (bird scribes the ABC algorithm. In Section 4, the problem of solar cell
matting optimization) [27]. Although heuristic methods present a identification is translated to an optimization task. Section 5 pre-
higher probability of obtaining a global solution in comparison with sents the experimental results and comparisons. In Section 6, the
deterministic ones, they have important limits [18]. In case of GA conclusions are stated, finally an appendix with the ABC algorithm
and PSO, they maintain a trend that concentrates toward local is presented.
optima, since their elitist mechanism forces premature conver-
gence [28,29]. Such a behavior becomes worse when the optimi- 2. Solar cell modeling
zation algorithm faces multi-modal functions [30,31]. On the other
hand, due to the fact that SA and HS are single-searcher algorithms, The modeling of PV cells consists in two steps: the mathematical
their performance is sensitive to the starting point of the search, model formulation and the accurate estimation of their parameter
having a lower probability to localize the global minimum in multi- values. In general, there exist two models: SD (single diode) and DD
modal problems than population algorithms such as GA and PSO (double diode) [13]. In this section these models are described and
[32,33]. Therefore, GA, PSO, SA, and HS present a bad performance their objective functions are formulated.
when they are applied to multi-modal and noisy objective
functions.
2.1. Double diode model (DD)
In order to identify the PV parameters as an optimization
problem, it is necessary to define an objective function. Such an
Solar cells are ideally modeled considering a photo-generated
objective function is built by using experimental data extracted
(Iph) current source which is shunted with a rectifying diode.
from IeV measurements of the solar cell. Since experimental data
However, in practical terms, the current source Iph is shunted by
contain noise as a consequence of an imperfect data collection
another diode which models the space charge recombination cur-
process, the objective function obtained presents high multi-modal
rent and other non-idealities. The model of solar cells also includes
and noisy characteristics [34,35]. Under these circumstances, most
a resistor connected in series with the cell shunt elements [13].
of the heuristic approaches present a bad performance [36].
Fig. 1 shows the equivalent circuit for the DD model.
In this paper, an alternative approach using the ABC (artificial
According to Fig. 1, the cell terminal current is computed as
bee colony) [37] method for determining the parameters of a solar
follows:
cell is presented. The ABC is an evolutionary algorithm inspired by
the intelligent behavior of honey bees. The performance of the ABC It ¼ Iph  Id1  Id2  Ish ; (1)
has been compared to other evolutionary methods such as GA and
PSO [38,39]. The results have shown that ABC produces optimal where It is the terminal current, Iph the photo-generated curren-
solutions when it faces multi-modal and noisy optimization prob- t,Id1,Id2 is the first and second diode currents whereas Ish is the
lems. Such characteristics have motivated the use of ABC to solve shunt resistor current. In order to appropriately model the solar
different types of engineering problems within several fields cell, there is used the Shockley diode equation; hence, Eq. (1) is
[40e45]. One relevant advantage of the ABC method is that it does rewritten as it is shown in Eq. (2).
not follow a local strategy for computing new solutions. Instead, the
ABC method uses a set of operators to build solutions from random
operations avoiding falling into local optimal.
ABC consists of three essential components: food source positions,
nectar amount, and several honey-bee classes. Each food source po-
sition represents a feasible solution for the problem under consider-
ation. The nectar amount for a food source represents the quality of
such a solution (represented by a fitness value). Each bee class sym-
bolizes one particular operation for generating new candidate food
source positions (i.e., candidate solutions). The ABC algorithm starts
by producing a randomly distributed initial population (food source
locations). After initialization, an objective function evaluates
whether such candidates represent an acceptable solution (nectar
amount) or not. Guided by the values of such an objective function,
candidate solutions are evolved through different ABC operations
(honey-bee types) until a termination criterion is met.
This paper presents the use of ABC to accurately estimate the
parameter of solar cells. In the approach, the estimation process is
considered as an optimization problem. The proposed approach
encodes the parameters of the solar cell as a candidate solution. An Fig. 1. Double diode model of solar cells.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
 D. Oliva et al. / Energy xxx (2014) 1e10 3

   Table 1
qðVt þ Rs $It Þ
It ¼ Iph  Isd1 exp 1 Upper and lower range of the solar cell parameters.
n1 $k$T

   Parameter Lower value Upper value
qðVt þ Rs $It Þ Vt þ Rs $It
 Isd2 exp 1  ; (2) Rs(U) 0 0.5
n2 $k$T Rsh
Rsh(U) 0 100
Iph(A) 0 1
where Isd1 and Isd2 are the diffusion and saturation current, Isd(mA) 0 1
respectively. Vt is the terminal voltage whereas the series and shunt n 1 2
resistances are represented by Rs and Rsh respectively. According to
the Shockley diode equation, q ¼ 1.602 1019 (coulombs) is the approach, the problem of parameter identification is considered as
magnitude of charge on an electron, k ¼ 1.380 1023 (J/K) is the an optimization problem where it is sought the parameter set that
Boltzmann constant, n1 and n2 are the diffusion and recombination produces the best approximation to the IeV measurements ob-
diode ideality factors, respectively. Finally, T is the cell temperature tained by the true solar cell. Therefore, it is necessary to define an
(K). Therefore, Eq. (2) has seven unknown parameters (Rs, Rsh, Iph, objective function that evaluates the matching quality between a
Isd1, Isd2, n1, and n2). An accurate identification of such parameters candidate parameter set and the experimental data. In this paper,
allows projecting the optimal performance of a solar cell, for that the problem of solar cell modeling is approached considering the
reason the estimation process is an important task. SD (Eq. (3)) as well as the DD (Eq. (2)) model. Thus, Eqs (2) and (3)
must be rewritten in order to reflex the difference with regard to
2.2. Single diode model (SD) experimental data. Thereby, for the DD model, the error function is
defined as follows:
In a solar cell, the diffusion (Isd1) and saturation (Isd2) currents
  
qðVt þ Rs $It Þ
are different and independent. In the SD model, both currents are fDD ðVt ; It ; xÞ ¼ It  Iph þ Isd1 exp 1
n1 $k$T
combined by using a non-physical ideality factor n [13,17,20]. This
  
model, shown in Fig. 2, is widely used for modeling solar cells due qðVt þ Rs $It Þ Vt þ Rs $It
þ Isd2 exp 1 þ ;
to its simplicity. Different to the DD, the SD model has only five n2 $k$T Rsh
parameters to be identified. (4)
Under the SD model, Eq. (2) is reduced to the following
whereas for the SD model such function is formulated as Eq. (5).
equation:

  
  
qðVt þ Rs $It Þ
qðVt þ Rs $It Þ Vt þ Rs $It fSD ðVt ; It ; xÞ ¼ It  Iph þ Isd exp 1
It ¼ Iph  Isd exp 1  (3) n$k$T
n$k$T Rsh
Vt þ Rs $It
Consequently, the parameters to be identified are Rs, Rsh, Iph, Isd, þ (5)
Rsh
and n. In Table 1 are presented the range values for each parameter,
notices that such ranges are used for both SD and DD. In both functions (fDD andfSD ), the values of Vt and It are exper-
imentally collected from the solar cell. x is a vector that contains the
model parameters, where x ¼ ½Rs ; Rsh ; Iph ; Isd1 ; Isd2 ; n1 ; n2  is the
2.3. Parameter identification of a solar cell as an optimization
model parameters for DD and x ¼ ½Rs ; Rsh ; Iph ; Isd ; n for SD. Eqs (4)
problem
and (5) allow to evaluate the model quality of the candidate
parameter set x, assessing the difference between the real value It
The problem of modeling solar cells consists in accurately
and the computed by the identified model. Therefore, the parameter
identifying the parameters of Eqs (2) and (3). In the proposed
estimation is a process that minimizes the difference between the
measured data and the calculated current by adjusting the model
parametersx. Considering that the number of experimental data is N,
the objective function can be formulated by the RMSE (root mean
square error) as:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u1 X N   2
RMSEðxÞ ¼ t f c V c ; Ic ; x ; (6)
N c¼1 M t t

where M is the model type DD or SD.

As it is formulated in Eq. (6), the objective function is built by
using experimental data, extracted from IeV measurements of the
solar cell. The experimental data could be also obtained from the
solar cell datasheet. Since experimental data contain noise as a
consequence of an imperfect data collection process, the objective
function obtained presents high multi-modal and noisy character-
istics [34,35]. Under these circumstances, most of the heuristic
approaches present a bad performance [36].

3. Artificial bee colony algorithm

In this paper, an alternative approach that uses the ABC (artifi-

Fig. 2. Single diode model of solar cells. cial bee colony) [37] method for determining the parameters of a

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
4 D. Oliva et al. / Energy xxx (2014) 1e10

solar cell is introduced. The ABC is an evolutionary algorithm The parameter fj;i is a random value selected between [1, 1], i
inspired by the intelligent behavior of honey-bees. The ABC algo- is an index that corresponds to the i-th food source and j is the
rithm has demonstrated to produce optimal solutions when it faces dimension problem index, then to generate the new source food
multi-modal and noisy optimization problems. using the employed bee operator, in a randomly way is selected a k
The ABC algorithm assumes the existence of a set of operations food source in the j dimension. If a parameter of an employed bee
that may resemble some features of the honeybee behavior. For food sourceBi;j exceeds the boundaries, it should be adjusted in
instance, each solution within the search space includes a param- order to fit the appropriate range. After this process, it is calcu-
eter set representing food source locations. The “fitness value” re- lated the fitness value associated with each solution. The fitness
fers to the food source quality that is linked to the food's location. value is used to evaluate the quality of a food source. For a mini-
The process mimics the bee's search for valuable food sources mization problem it can be obtained using the following
yielding an analogous process for finding the optimal solution. expression:
8
3.1. Biological profile >
< 1
if Ji  0
fiti ¼ 1 þ Ji (9)
>
:
The minimal model for a honeybee colony consists of three 1 þ absðJi Þ if Ji < 0
classes: employed bees, onlooker bees and scout bees. The
employed bees will be responsible for investigating the food where Ji is the objective function value with regard to the candidate
sources and sharing the information with recruit onlooker bees. solution xi . In our context, Ji represents the RMSE (Eq. (6)) value
They, in turn, will make a decision on choosing food sources by associated to a candidate model xi . The next process consist in
considering such information. The food source having a higher apply a greedy selection between the values of the employed bee
quality will have a larger chance to be selected by onlooker bees food sources contained in Bi and the initial food sources vector xi,
than those showing a lower quality. An employed bee, whose food that means: if the nectaramount (fitness value) of Bi is better, then
source is rejected as low quality by employed and onlooker bees, the solution xi is replaced by Bi otherwise, xi is preserved.
will change to a scout bee to randomly search for new food sources.
Therefore, the exploitation is driven by employed and onlooker
bees while the exploration is maintained by scout bees. 3.2.3. Select the food sources by the onlooker bees
In order to describe the onlooker phase, first it is necessary to
3.2. Description of the ABC algorithm explain that the number of onlooker bees corresponds to the food
source number. In this way the food sources are modified several
Similar to other swarm-based approaches, the ABC algorithm is times depending on the fitness value (Eq. (9)). For a food source
an iterative process. It starts with a population of randomly could be selected, it is necessary to obtain a probability factor that is
generated solutions or food sources. The following three operations computed based on the fitness.
are applied until a termination criterion is met [39]:
fit
Probi ¼ PN i (10)
p
1. Send the employed bees. i¼1
fiti
2. Select the food sources using the onlooker bees.
3. Determine the scout bees. Here, fiti corresponds to the fitness value of the i-th food source
and is related to the objective function of the food sourcei. If the
fitness of a food source increases, then the probability of be selected
3.2.1. Initializing the population by an onlooker is bigger. When a food source is selected a new value
The first step of the algorithm is to initialize the population of Np is obtained using Eq. (2), its fitness is computed and the greedy
food sources. Every food source is a d-dimensional vector con- process is applied to modify (or not) its position.
taining the parameters values to be optimized. Such values are
randomly and uniformly distributed between a bounded space. 3.2.4. Determine scout bees
  The final step is the scout bee process. Here the bees are applied
xi;j ¼ lj þ randð0; 1Þ$ uj  lj ;
(7) if a food source i cannot be improved through a predetermined trial
j ¼ 1; 2; …; d; i ¼ 1; 2; …; Np ;
“limit” number, then the food source is considered to be abandoned
and instead to be modified by and onlooker bee, is modified by
where xi;j is a food source, the index i corresponds to i-th food source
a scout bee using Eq. (7). The predefined “limit” is a counter
and j is the j-th dimension of the search space. lj and uj are the lower
assigned to each food source and is incremented when the fitness is
and the upper bound in each dimension. The indexes i and j will be
not improved.
used under the same definition in the remainder section.

3.2.2. Send employed bees 4. Parameter identification of solar cells using ABC
The employed bees are used to generate new solutions; the
number of this kind of bees is equal to the number of food sources. 4.1. Problem statement
According with the literature [37,38] the entire population is
divided in two ðNp =2Þ, one part corresponds to the employed bees The proposed approach encodes the parameters of the solar
and the rest to the onlooker bees. This division operates as part of cell as a candidate solution. The representation of such a
the search strategy, for that reason, it needs to be applied for all candidate solution (food source) depends on the model type: DD
problems where ABC is employed as optimization tool. or SD. Therefore, each food source uses seven elements for the
DD formulation and five for the SD model, as decision variables
Bi;j ¼ xi;j þ fi;j xi;j  xk;j ; c isk within the optimization algorithm. Thus, the estimation task is
(8)
k2rand 1; Np ; j2f1; 2; :::; dg faced as an optimization problem which can be stated as fol-
lows:

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
 D. Oliva et al. / Energy xxx (2014) 1e10 5

h i 
minimize : RMSEðXÞ; X ¼ x1 ; x2 ; …; xN ; xi ¼ xi;1 ; xi;2 ; …; xi;d d2½5; 7;
p
d ¼ 5 ðSDÞ
d ¼ 5 ðSDÞ 0  xi;1 ðRs Þ  0:5
0  xi;1 ðRs Þ  0:5 0  xi;2 ðRsh Þ  100
0  xi;2 ðRsh Þ  100 0  xi;3 Iph  1 (11)
subject to :
0  xi;3 Iph  1 0  xi;4 ðIsd1 Þ  1
0  xi;4 ðIsd Þ  1 0  xi;5 ðIsd2 Þ  1
1  xi;5 ðnÞ  2 1  xi;6 ðn1 Þ  2
1  xi;7 ðn2 Þ  2

where Np and d are the population size and the number of di- 5. Experimental results
mensions, respectively.
In order to prove the performance of the proposed approach,
4.2. Computational approach the algorithm has been tested using a commercial silicon solar cell
(from the R.T.C. Company of France) under the STC (standard test
The proposed algorithm has been implemented considering the conditions), with a diameter of 57 mm. During the data collection
two different solar cell models (SD and DD) whereas its efficiency is process, it is considered that the solar cell operates under the
evaluated using de RMSE criterion. As optimization technique, the following operating conditions: 1 sun ð1000W=m2 Þ at T ¼ 33 C;
ABC method is used to solve the problem of parameter identifica- however, in order to test the performance of the ABC method four
tion defined by Eq. (11). The computational procedure of the pro- more temperatures have been included T ¼ 25 C, T ¼ 50 C,
posed approach can be summarized into the Algorithm 2. T ¼ 75 C and T ¼ 100 C. In this section, two different results are
presented. In the first part, the proposed approach is employed to
Algorithm 2. Computational approach.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
6 D. Oliva et al. / Energy xxx (2014) 1e10

Table 2 5.1. ABC experimental results

Parameter setup for the ABC algorithm.

cmax Np limit This experiment presents the results of the proposed approach
10,000 150 Np $d
when it is employed to extract the cell parameters considering the
single and double diode models. To this end, 26 measurements
from the physical solar cell are collected. Such samples, shown in
Table 3 Table 4 represent the experimental data set; this set is extensively
Extracted parameter after applying ABC for SD and DD.
used in the related literature. Here, it has been selected to maintain
Parameter Double diode Single diode compatibility to similar works reported in the literature. The
Rs ðUÞ 0.0364 0.0364 extracted parameters for the SD and DD model are shown in
Rsh ðUÞ 53.7804 53.6433 Table 3.
Iph ðAÞ 0.7608 0.7608 Since the SD model has five parameters and DD seven param-
Isd ðmAÞ e 0.3251 eters, there are parameters not available for one or other model in
Isd1 ðmAÞ 0.0407 e
Isd2 ðmAÞ 0.2874 e
Table 4. The inexistence of such parameters is represented by the
n1 1.4495 e symbol (e). In order to evaluate the accuracy of the identified
n2 1.4885 e model, four different performance indexes have been employed:
n e 1.4817 The relative errorRerror, the MAE (median absolute error) and its
RMSE 9.861 E-04 9.862 E-04
respective NMAE (normalized MAE), the NRMSE (normalized
RMSE), the MBE (median bias error) and the NMBE (normalized
median bias error).
The relative error Rerror evaluates the difference between the
extract the cell parameters using the single and double diode measured current Itmeasured and the calculated by the respective
models. Finally, in the second part, the results of the ABC-based modelItcalculated . Rerror , is calculated by:
approach are compared with other well-known similar
approaches.
Itmeasured  Itcalculated
In order to conduct such experiments, the ABC is configured Rerror ¼  100 (13)
considering the parameter values shown in Table 2. Once they have Itmeasured
been determined after intensive tests, they are kept for all The MAE (median absolute error) and its respective NMAE
experiments. (normalized MAE) are computed using Eqs. (14) and (15), the value
The parameter limit is computed asNp $d, where Np and d are the of N corresponds to the number of experimental data (for this work
population size and the number of dimensions, respectively. In the N ¼ 26).
experiments, the stop criterion is the maximum iteration
numbercmax. However, if the fitness value for the best candidate
solution remains unspoiled in 10% of the total number of cmax , then X
N
jI  Itcalculated j
tmeasured
MAE ¼ (14)
the algorithm is stopped. N
i¼1

Table 4
Terminal (Vt  It ) measurements and relative error values for: double and single diode models.

Data Vt ðV Þ It ðAÞ Itcalculated ðAÞ Rerror Normalized NRerror ð%Þ Itcalculated ðAÞ Rerror Normalized NRerror ð%Þ
Measured Measured ABC double diode ABC double diode ABC double diode ABC single ABC single ABC single
model model model diode model diode model diode model

1 0.2057 0.7640 0.7640 9.2908 E-05 36.8310 0.7641 0.0001 36.6608

2 0.1291 0.7620 0.7626 0.0006 22.8277 0.7626 0.0006 22.6153
3 0.0588 0.7605 0.7613 0.0008 18.1264 0.7613 0.0008 17.95165
4 0.0057 0.7605 0.7601 0.0003 47.2470 0.7601 0.0003 47.4202
5 0.0646 0.7600 0.7590 0.0009 61.7485 0.7590 0.0009 62.1258
6 0.1185 0.7590 0.7580 0.0009 62.0349 0.7580 0.0009 62.4765
7 0.1678 0.7570 0.7571 0.0001 36.5502 0.7571 0.0001 36.8073
8 0.2132 0.7570 0.7561 0.0008 59.6029 0.7561 0.0008 60.1225
9 0.2545 0.7555 0.7550 0.0004 48.8006 0.7550 0.0004 49.2543
10 0.2924 0.7540 0.7536 0.0003 46.9719 0.7536 0.0003 47.4304
11 0.3269 0.7505 0.7513 0.0008 17.2776 0.7513 0.0008 17.4642
12 0.3585 0.7465 0.7473 0.0008 18.3219 0.7473 0.0008 18.5101
13 0.3873 0.7385 0.7401 0.0016 0 0.7401 0.0016 0
14 0.4137 0.7280 0.7273 0.0006 54.4874 0.7273 0.0006 54.9849
15 0.4373 0.7065 0.7069 0.0004 28.2650 0.7069 0.0004 28.5214
16 0.4590 0.6755 0.6752 0.0002 45.2437 0.6752 0.0002 45.6933
17 0.4784 0.6320 0.6307 0.0012 70.0782 0.6307 0.0012 70.8228
18 0.4960 0.5730 0.5718 0.0011 65.8199 0.5718 0.0011 66.6002
19 0.5119 0.4990 0.4995 0.0005 24.8089 0.4995 0.0005 25.2670
20 0.5265 0.4130 0.4136 0.0006 23.4270 0.4136 0.0006 23.8922
21 0.5398 0.3165 0.3175 0.0010 14.2999 0.3175 0.0010 14.6298
22 0.5521 0.2120 0.2121 0.0001 34.7889 0.2121 0.0001 35.1718
23 0.5633 0.1035 0.1022 0.0012 68.7767 0.1022 0.0012 69.2431
24 0.5736 0.0100 0.0087 0.0012 7.57006 0.0086 0.0013 7.13579
25 0.5833 0.1230 0.1255 0.0025 100 0.1254 0.0024 100
26 0.5900 0.2100 0.2085 0.0014 2.6948 0.2084 0.0015 1.41935

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
 D. Oliva et al. / Energy xxx (2014) 1e10 7

Table 5
Performance indexes for: double and single diode models. X
N
jI  Itcalculated j=Itmeasured
tmeasured
NMAE ¼ (15)
Model RMSE NRMSE MAE NMAE MBE NMBE i¼1
N
(%) (%) (%)
The normalized RMSE is defined as follows:
Double 9.8619 E-04 62.53 8.2918 E-04 0.47 5.8807 E-07 1.19872
diode RMSE
Single 9.8629 E-04 62.70 8.3034 E-04 0.49 1.5448 E-06 1.21547 NRMSEðxÞ ¼ ; (13)
maxðItcalculated Þ  minðItcalculated Þ
diode

Fig. 3. For the DD model: (a) Measured voltage vs. ABC computed current for different temperatures, (b) Measured voltage vs. ABC-power at different temperatures, (c) RMSE
evolution. SD model: (d) Measured voltage vs. ABC computed current for different temperatures, (e) Measured voltage vs. ABC-power at different temperatures, (f) RMSE evolution.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
8 D. Oliva et al. / Energy xxx (2014) 1e10

where minðItcalculated Þ and maxðItcalculated Þ are the minimum and Table 7

maximum values of Itcalculated over the existent N samples. Comparison results for the SD model.

Finally, the MBE (median bias error) and the NMBE (normalize Parameter ABC HS PSO GA BFA
MBE) are computed as follows: Rs ðUÞ 0.0364 0.0366 0.0354 0.0299 0.0325
Rsh ðUÞ 53.6433 53.5946 59.0120 42.3729 50.8691
X
N
ðI  Itcalculated Þ Iph ðAÞ 0.7608 0.7607 0.7607 0.7619 0.7602
tmeasured
MBE ¼ (16) Isd ðmAÞ 0.3251 0.3049 0.4000 0.8087 0.8000
i¼1
N
n 1.4817 1.4753 1.5033 1.5751 1.6951
RMSE 9.862 E-04 9.510 E-04 0.0013 0.0190 0.029
Mean 0.0010 0.0039 0.2544 0.0551 0.0152
MBE
NMBE ¼ (17) STD 1.497 E-05 0.7268 0.0289 0.0735 0.0586
maxðItcalculated Þ  minðItcalculated Þ
Table 5 presents the results of the experiment. Such results
include the experimental data (Vt and It ), the obtained results
(Itcalculated ), the respective relative errors (Rerror ) and their algorithm present better performance than other approaches. The
normalized values for both models. Mean corresponds to the average values of the RMSE after 35 in-
In Table 5 are presented the values of RMSE, NRMSE (normalized dependent experiments. The STD value can be interpreted as a
RMSE), MAE, NMAE, MBE and NMBE (normalized MBE) for the stability index which reflects the algorithm capacity to produce the
proposed approach based on ABC. same result when it is executed several times. Likewise, Table 7
Considering the model parameters of Table 4, it is possible to shows the comparison analysis for the single diode model. The
obtain the power (P ¼ I  V ) characteristics of the solar cell. results show that the proposed algorithm performs better in
Fig. 3 shows the graphs or current vs. voltage at different tem- comparison with the HS, PSO GA and BFA algorithms in terms of the
peratures (T ¼ 25 C, T ¼ 50 C, T ¼ 75 C and T ¼ 100 C), the po- averaged RMSE and STD values.
wer, and the fitness values for the double and single diode
models. 6. Conclusions
Fig. 3 presents the results obtained by the proposed approach
considering the two diode models and five different temperatures. In this paper, the use of ABC (artificial bee colony) to accu-
Form Fig. 3, it is possible to analyze that the ABC-based approach rately estimate the parameter of solar cells has been presented.
obtains better models producing an accurate approximation to In the approach, the estimation process is considered as an
experimental data; however, the influence of the temperatures optimization problem. The proposed approach encodes the
evidently modifies the current values in the model. This fact af- parameters of the solar cell as a candidate solution. An objec-
fects directly the power values. Besides, the evolution of the tive function evaluates the matching quality between a candi-
optimization process shows that the proposed method allow to date solution and the experimental data. Guided by the values
find appropriate solar cell models in a reduced number of of this objective function, the set of encoded candidate solu-
generations. tions is evolved by using the operators defined by ABC so that
the parameters that produce the best possible approximation to
the IeV measurements obtained by the true solar cell can be
5.2. Comparisons with other approaches found.
The proposed approach has been compared with other
In order to demonstrate the performance of the proposed similar techniques proposed in the literature such as HS, PSO GA
approach, its results have been compared to those produced by and BFA. The efficiency of the algorithm has been evaluated in
other similar implementations reported in the literature, for solar terms of accuracy and robustness. Experimental results provide
cell modeling. The methods used in the comparison are: HS (har- evidence on the outstanding performance, accuracy and
mony search) () [13], PSO (particle swarm optimization) [12], GA convergence of the proposed algorithm in comparison to such
(genetic algorithms) [18] and BFA (Bacterial Foraging Algorithm) methods.
[21]. In the comparison, all the algorithms have been executed 35 Although the results offer evidence to demonstrate that the
times so that it can be computed their averaged RMSE values and standard ABC method can yield good results on both diode models,
their respective mean values and STD (standard deviation). Tables 6 the aim of our paper is not to devise an SC algorithm that could beat
and 7 present the results obtained from this analysis, for the DD and all currently available methods, but to show that harmony search
SD model, respectively. algorithms can be effectively considered as an attractive alternative
Table 6 presents the comparison analysis for the double diode for this purpose.
model. From the results, it is possible to see that the ABC-based

Acknowledgements
Table 6
Comparison results for the DD model. The first author acknowledges The National Council of Science
Parameter ABC HS PSO GA BFA and Technology of Mexico (CONACyT) for the doctoral Grant
number 215517 and The Youth Institute of Jalisco (IJJ) for partially
Rs ðUÞ 0.0364 0.0354 0.0325 0.0364 0.0351
Rsh ðUÞ 53.7804 46.8269 43.1034 53.7185 60.0000 support this research.
Iph ðAÞ 0.7608 0.7617 0.7623 0.7608 0.7609
Isd1 ðmAÞ 0.0407 0.1245 0.4767 0.0001 0.0094
Appendix
Isd2 ðmAÞ 0.2874 0.2547 0.0100 0.0001 0.0453
n1 1.4495 1.4943 1.5172 1.3355 1.3809
n2 1.4885 1.4998 2.0000 1.4810 1.5255 ABC computational procedure
RMSE 9.861 E-04 0.0013 0.0166 0.3604 0.0012
Mean 0.0010 0.0683 0.0715 0.0229 0.0245 The complete ABC Algorithm can be summarized by the in-
STD 3.285 E-05 0.5111 0.3109 0.0199 0.3697
structions listed in Algorithm 1.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
 D. Oliva et al. / Energy xxx (2014) 1e10 9

Algorithm 1. Artificial bee colony method.

References [15] Ortiz-Conde A, Garcia Sanchez FJ, Muci J. New method to extract the model
parameters of solar cells from the explicit analytic solutions of their illumi-
nated IeV characteristics. Sol Energy Mater Sol Cells 2006;90(3):352e61.
[1] Renewables. Global status report http://www.ren21.net/globalstatusreport/;
[16] Chan DSH, Phillips JR, Phang JCH. A comparative study of extraction methods
2010.
for solar cell model parameters. Solid-State Electron 1986;29(3):329e37.
[2] Ishaque K, Salam Z, Mekhilef S, Shamsudin A. Parameter extraction of solar
[17] AlRashidi MR, AlHajri MF, El-Naggar KM, Al-Othman AK. A new estimation
photovoltaic modules using penalty-based differential evolution. Appl Energy
approach for determining the IeV characteristics of solar cells. Sol Energy
2012;99:297e308.
2011;85(7):1543e50.
[3] Orioli A, Gangi AD. A procedure to calculate the five-parameter model of
[18] Jervase JA, Bourdoucen H, Al-Lawati A. Solar cell parameter extraction using
crystalline silicon photovoltaic modules on the basis of the tabular perfor-
genetic algorithms. Meas Sci Technol 2001;12(11):1922e5.
mance data. Appl Energy 2013;102:1160e77.
[19] Wei H, Cong J, Lingyun X, Deyun S. Extracting solar cell model parameters
[4] Sandrolini L, Artioli M, Reggiani U. Numerical method for the extraction of
based on chaos particle swarm algorithm. In: International conference on
photovoltaic module double-diode model parameters through cluster anal-
electric information and control engineering (ICEICE); 2011. pp. 398e402.
ysis. Appl Energy 2010;87:442e51.
[20] El-Naggar KM, AlRashidi MR, AlHajri MF, Al-Othman AK. Simulated annealing
[5] Amrouche B, Guessoum A, Belhamel M. A simple behavioural model for solar
algorithm for photovoltaic parameters identification. Sol Energy 2012;86(1):
module electric characteristics based on the first order system step response
266e74.
for MPPT study and comparison. Appl Energy 2012;91:395e404.
[21] Rajasekar N, Krishna Kumar Neeraja, Venugopalan Rini. Bacterial Foraging
[6] Bonanno F, Capizzi G, Graditi G, Napoli C, Tina GM. A radial basis function
Algorithm based solar PV parameter estimation. Sol Energy 2013;97:255e65.
neural network based approach for the electrical characteristics estimation of
[22] Jiang LL, Maskell DL, Patra JC. Parameter estimation of solar cells and modules
a photovoltaic module. Appl Energy 2012;97:956e61.
using an improved adaptive differential evolution algorithm. Appl Energy
[7] Han L, Koide N, Chiba Y, Mitate T. Modeling of an equivalent circuit for dye-
2013;112:185e93.
sensitized solar cells. Appl Phys Lett 2004;13:2433e5.
[23] Niu Q, Zhang H, Li K. An improved TLBO with elite strategy for parameters
[8] Villalva MG, Gazoli JR, Filho ER. Comprehensive approach to modeling and
identification of PEM fuel cell and solar cell models. Int J Hydrogen Energy
simulation of photovoltaic arrays. IEEE Trans Power Electron 2009;24(5):
2014;39(8):3837e54.
1198e208.
[24] Appelbaum J, Peled A. Parameters extraction of solar cells e a comparative
[9] Huld T, Gottschalg R, Beyer HG, Topic M. Mapping the performance of a PV
examination of three methods. Sol Energy Mater Sol Cells 2014;122:164e73
modules, effects of module type and data averaging. Sol Energy 2010;84:
[*****clasicos y GA].
324e8.
[25] Gong W, Cai Z. Parameter extraction of solar cell models using repaired
[10] Xiao W, Lind MGJ, Dunford WG, Capel A. Real-time identification of optimal
adaptive differential evolution. Sol Energy 2013;94:209e20.
operating points in photovoltaic power systems. IEEE Trans Ind Electron
[26] Siddiqui MU, Abido M. Parameter estimation for five- and seven-parameter
2006;53(4):1017e26.
photovoltaic electrical models using evolutionary algorithms. Appl Soft
[11] Chegaar M, Ouennough Z, Guechi F, Langueur H. Determination of solar cells
Comput 2013;13:4608e21.
parameters under illuminated conditions. J Electron Devices 2003;2:17e21.
[27] Askarzadeh A, Rezazadeh A. Extraction of maximum power point in solar cells
[12] Ye M, Wang X, Xu Y. Parameter extraction of solar cells using particle swarm
using bird mating optimizer-based parameters identification approach. Sol
optimization. J Appl Phys 2009;105(9):094502e8.
Energy 2013;90:123e33.
[13] Askarzadeh Alireza, Rezazadeh Alireza. Parameter identification for solar cell
[28] Hrstka Ondrej, Ku^cerov
a Anna. Improvements of real coded genetic algo-
models using harmony search-based algorithms. Sol Energy 2012;86(11):
rithms based on differential operators preventing premature convergence.
3241e9.
Adv Eng Softw 2004;35:237e46.
[14] Easwarakhanthan T, Bottin J, Bouhouch I, Boutrit C. Nonlinear minimization
[29] OstadmohammadiArani Behrooz, Mirzabeygi Pooya, ShariatPanahi Masoud.
algorithm for determining the solar cell parameters with microcomputers. Sol
An improved PSO algorithm with a territorial diversity-preserving scheme
Energy 1986;(4):1e12.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011
10 D. Oliva et al. / Energy xxx (2014) 1e10

and enhanced explorationeexploitation balance. Swarm Evol Comput [37] Karaboga D. An idea based on honey bee swarm for numerical optimization.
2013;11:1e15. Technical report-TR06. Erciyes University, Engineering Faculty, Computer
[30] Qing Ling, Gang Wu, Zaiyue Yang, Qiuping Wang. Crowding clustering genetic Engineering Department; 2005.
algorithm for multimodal function optimization. Appl Soft Comput 2008;8: [38] Karaboga D, Basturk B. On the performance of artificial bee colony (ABC) al-
88e95. gorithm. Appl Soft Comput 2008;8(1):687e97.
[31] Li Minqiang, Lin Dan, Kou Jisong. A hybrid niching PSO enhanced with [39] Karaboga D, Akay B. A comparative study of artificial bee colony algorithm.
recombination-replacement crowding strategy for multimodal function Appl Math Comput 2009;214:108e32.
optimization. Appl Soft Comput 2012;12:975e87. [40] Karaboga N. A new design method based on artificial bee colony algorithm for
[32] Niksirat Malihe, Ghatee Mehdi, Mehdi Hashemi S. Multimodal K-shortest digital IIR filters. J Franklin Inst 2009;346:328e48.
viable path problem in Tehran public transportation network and its solution [41] Pan Q-K, Fatih Tasgetiren M, Suganthan PN, Chua TJ. A discrete artificial bee
applying ant colony and simulated annealing algorithms. Appl Math Model colony algorithm for the lot-streaming flow shop scheduling problem. Inf Sci;
2012;36:5709e26. 2011. http://dx.doi.org/10.1016/j.ins.2009.12.025.
[33] Wang Chia-Ming, Huang Yin-Fu. Self-adaptive harmony search algorithm for [42] Kang F, Li J, Xu Q. Structural inverse analysis by hybrid simplex artificial bee
optimization. Expert Syst Appl 2010;37:2826e37. colony algorithms. Comput Struct 2009;87:861e70.
[34] Li Jun-hua, Li Ming. An analysis on convergence and convergence rate esti- [43] Zhang C, Ouyang D, Ning J. An artificial bee colony approach for clustering.
mate of elitist genetic algorithms in noisy environments. Optik 2013;124: Expert Syst Appl 2010;37:4761e7.
6780e5. [44] Karaboga D, Ozturk C. A novel clustering approach: Artificial Bee Colony (ABC)
[35] Pan Hui, Wang Ling, Liu Bo. Particle swarm optimization for function opti- algorithm. Appl Soft Comput 2011;11:652e7.
mization in noisy environment. Appl Math Comput 2006;181:908e19. [45] Ho SL, Yang S. An artificial bee colony algorithm for inverse problems. Int J
[36] Beyer Hans-Georg. Evolutionary algorithms in noisy environments: theoret- Appl Electromagn Mech 2009;31:181e92.
ical issues and guidelines for practice. Compute Meth Appl Mech Eng
2000;186:239e67.

Please cite this article in press as: Oliva D, et al., Parameter identification of solar cells using artificial bee colony optimization, Energy (2014),
http://dx.doi.org/10.1016/j.energy.2014.05.011