132 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 26, NO.

1, MARCH 2011
Chebyshev Neural Network-Based Model for
Dual-Junction Solar Cells
Jagdish Chandra Patra, Member, IEEE
AbstractDesign and development process of solar cells can be
greatly enhanced by using accurate models that can predict their
behavior accurately. Recently, there has been a surge in research
efforts in multijunction (MJ) solar cells to improve the conver-
sion efciency. Modeling of MJ solar cells poses greater challenges
because their characteristics depend on the complex photovoltaic
phenomena and properties of the materials used. Currently, sev-
eral commercial complex device modeling software packages, e.g.,
ATLAS, are available. But these software packages have limita-
tions in predicting the behavior of MJ solar cells because of several
assumptions made on the physical properties and complex interac-
tions. Articial neural networks have the ability to effectively model
any nonlinear systemwith complex mapping between its input and
output spaces. In this paper, we proposed a novel Chebyshev neu-
ral network (ChNN) to model a dual-junction (DJ) GaInP/GaAs
solar cell. Using the ChNN, we have modeled the tunnel junction
characteristics and developed models to predict the external quan-
tumefciency, and IV characteristics both at one sun and at dark
levels. We have shown that the ChNN-based models performbetter
than the commercial software, ATLAS, in predicting the DJ solar
cell characteristics.
Index TermsChebyshev neural networks (ChNN), dual-
junction (DJ) solar cell, modeling, tunnel junction (TJ).
I. INTRODUCTION
A
VAILABILITY of effective modeling techniques to accu-
rately estimate the device behavior is quite important to
enhance the design and development process of solar cells. In the
case of analytical modeling, complex mathematical equations
are derived based on materials properties, e.g., optical absorp-
tion, hole and electron mobilities, various energy parameters,
and physical and chemical phenomena and interactions. Several
analytical modeling techniques based on complex mathematical
expressions to estimate the behavior of the solar cells have been
reported. Some of the recently published reports are modeling of
departure of IV characteristics under extreme conditions [1],
modeling for intermediate band solar cells [2], modeling of
monolithic InGaN/Si two-junction solar cells [3], degradation
modeling of InGaP/GaAs/Ge solar cells [4], modeling of solar
cells under solar concentration [5], and numerical simulation of
tunnel diodes [6]. The aim of electrical solar cell modeling is to
Manuscript received February 4, 2010; revised May 12, 2010; accepted
September 13, 2010. Date of publication November 1, 2010; date of current
version February 18, 2011. Paper no. TEC-00055-2010.
The author is with the School of Computer Engineering, Nanyang Techno-
logical University, Singapore, Singapore (e-mail: [email protected]).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TEC.2010.2079935
develop a link between analytical modeling based on materials
properties and the electrical device characteristics. The solar cell
is represented by an electrical circuit with series and shunt resis-
tances and a diode. The IV characteristics of the solar cell can
be simulated using the electrical circuit parameters. The main
solar cell parameters, e.g., the open-circuit voltage (V
OC
), short-
circuit current density (J
SC
), ll factor (FF), external quantum
efciency (EQE), and conversion efciency (), can be obtained
from the simulated IV characteristics. Using a network model,
Burgelman et al. have reported an effective modeling technique
for nanostructured solar cells [7]. Simulation and optimization
of metalinsulatorsemiconductor inversion-layer silicon solar
cells using 2-D numerical modeling and circuit simulation have
been reported in [8]. In addition, several commercial complex
device modeling software packages based on different analytical
and electrical modeling techniques are available [9], [10].
These modeling techniques are based on the mathematical re-
lationships and interactions that depend on physical, chemical,
and electronic properties of the materials used in the solar cells.
Although these techniques provide adequate solutions in most
of the devices, they may not yield accurate results in case of
complex multijunction (MJ) solar devices. One of the main rea-
sons for the inadequacy of analytical or electrical techniques is
due the fact that some of the complex phenomena and material
properties are not well known, e.g., the band-to-band recom-
bination, ShockleyReadHall scattering time, absorption co-
efcients for different materials, reective and resistive losses,
etc. [3]. Therefore, several assumptions and simplications are
made in the modeling techniques that lead to inaccuracies in the
predicted device behavior.
Articial neural networks (ANNs), due to their several unique
attributes, e.g., ability to learn, adapt and generalize, and toler-
ance to noise and faults, have been successfully used to solve
several complex and highly nonlinear problems [11]. A mul-
tilayer perceptron (MLP) trained with back-propagation (BP)
algorithm is one of the most popular ANNs that has been used
in several applications.
Wide application of ANNs and articial intelligence tech-
niques for photovoltaic (PV) modeling is evident from the ex-
cellent review paper [12]. Recently, Abdulhadi et al. [13] have
used a neurofuzzy-based model to predict the I
SC
and V
OC
of solar cells. A radial basis function (RBF)-based controller
to increase PV plant efciency [14], ANN-based characteriza-
tion of Si-crystalline PV modules [15], and ANN techniques for
accurate estimation of electrical circuit parameters [16] have
been reported with impressive results. Some other applications
of ANNs include estimation of maximum power generation in
PV modules [17], prediction of maximum power point of PV
0885-8969/$26.00 2010 IEEE
PATRA: CHEBYSHEV NEURAL NETWORK-BASED MODEL FOR DUAL-JUNCTION SOLAR CELLS 133
systems using improved RBF networks [18] and PV system
design [19].
Difculties associated with measurement and characteriza-
tion of tunnel diodes and MJ solar cells have been reported
by Guter and Bett [20]. A 3-D-distributed model for high-
concentrator solar cells based on elementary units made up of
electrical circuits is presented in [21]. Baudrit and Algora [22]
modeled a GaInP/GaAs DJ solar cell and tunneling phenomenon
using the device simulator ATLAS from Silvaco [9]. They have
shown that the ATLAS is able to provide satisfactory perfor-
mance in predicting tunneling behavior, determination of EQE,
and IVcharacteristics of the dual-junction solar cell under solar
irradiation and darkness. However, the ATLAS results are not
accurate enough to match to that of measured ones. Substantial
deviations are found between the experimental and the ATLAS-
simulated results for the FF and conversion efciency computed
from the IV characteristics of the DJ cell.
A functional link ANN (FLANN) is a highly effective and
computationally efcient, and the fast learning single layer ANN
in which the hidden layer is replaced by a functional expansion
block for enhancement of the input pattern space [23]. The
Chebyshev neural network (ChNN) is one class of the FLANN
in which functional expansion is carried out using orthogonal
Chebyshev polynomials that is computationally more efcient
than the popularly used MLP networks. The ChNN has been
successfully applied in intelligent sensors [24], digital commu-
nication [25], and dynamic system identication [26]. In this
paper, we proposed ChNN-based models to characterize a DJ
GaInP/GaAs solar cell. We have shown that in comparison to
the ATLAS model, the ChNN-based model is able to predict
the DJ solar cell parameters more closely to that of the experi-
mental ones. In this paper, we have taken the data (experimental
and simulated through the ATLAS) from the article reported by
Baudrit and Algora [22].
The rest of this paper is organized as follows. The basic
theory of the ANN-based modeling scheme from a system iden-
tication perspective is given in Section II. The results of the
ChNN-based model for estimation of tunnel junction (TJ) char-
acteristics and its performance comparison with the experimen-
tal and ATLAS simulated results are provided in Section III. In
the next section, we provide results of the ChNN-based model
for estimation of EQE and IV characteristics of the DJ solar
cell, both under one sun and in darkness, and computed the FF
and conversion efciency. Performance comparison with exper-
imental and the ATLAS-simulated results, and the sensitivity
analysis of ChNN-based models are also provided. We provide
conclusion and few comments for further study in Section V.
II. SYSTEM IDENTIFICATION AND CHEBYSHEV NEURAL
NETWORK
Identication of an unknown system is an important task
in control of industrial plants and prediction and estimation
of parameters in any process or system. The pioneer work of
Narendra and Parthsarathy [27] has shown the effectiveness of
the ANN-based identication of nonlinear dynamic systems.
In an earlier paper [26], we have reported the ability of the
Fig. 1. Schematic of system identication.
ChNN for identication of nonlinear dynamic systems. Here,
we briey describe the system identication perspective of a
solar cell modeling, explain the theory of the ChNN and the
learning algorithm, based on our earlier paper [26].
A. Characterization and Identication of Systems
Let the model of a system is represented by an operator P that
transforms the input space U into an output space Y. The oper-
ator P is dened implicitly by the specied inputoutput pairs.
The objective is to categorize the class P to which P belongs.
For a given class P, P P, the identication problem is to
determine a class

P P and

P

P such that

P approximates
P in some desired sense. In a static system, the spaces U and Y
are subsets of R
n
and R
m
, respectively [?]. Pattern recognition
is an example of the static identication problem. Using a deci-
sion function P, input sets U
i
R
n
are mapped into elements
y
i
R
m
for i = 1, 2, . . ., in the output space. The elements of
U
i
denote the pattern vectors corresponding to class y
i
. The
main objective is to determine

P such that
y y = P(u)

P(u) < (1)
where u U, is some desired small value > 0, and is a
dened norm on the output space. In (1),

P and P denote the
output of the identied model and the plant, respectively. The
error, e = y y, is the difference between the observed plant
output and the output generated by the model.
In the schematic diagram of the system identication shown
in Fig. 1, the input and output of the plant are represented by u
and p(u), respectively, where u is assumed to be a uniformly
bounded function of time. The objective is to construct a suit-
able model generating an output

P(u) that approximates the
plant output P(u). In the context of solar cell modeling, we
consider the plant to be the solar cell in which the inputoutput
relationship is represented by a nonlinear function given by
y
p
(k) = f(u(k)) (2)
where u(k) and y
p
(k) represent the input and the output of
a solar cell at the kth time instant, respectively. In this paper,
we used a ChNN to construct the nonlinear function f(.) by
using several measurement data as training set. For modeling
TJ characteristics, the model input and output are given by the
input voltage and the tunnel diode current, respectively, whereas,
for estimation of EQE of the DJ solar cell, the models input
and output are given by the radiation wavelength and EQE,
respectively.
134 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 26, NO. 1, MARCH 2011
Fig. 2. Schematic of a Chebyshev neural network.
B. ChNN
The structure of a ChNN is depicted in Fig. 2. It consists
of a functional expansion block and a single-layer perceptron
network. The main purpose of the functional expansion block is
to increase the dimension of the input pattern so as to enhance
its representation in a higher dimensional space. This enhanced
pattern is then used for modeling of the solar cell. Let us denote
an n-dimensional input pattern at kth instant by
X
k
= [x
1
(k), x
2
(k), . . . , x
n
(k)]
T
. (3)
Each element of the input vector is expanded to several terms
using Chebyshev polynomials. The Chebyshev polynomials
are a set of orthogonal polynomials obtained as a solution of
the Chebyshev differential equation. The rth-order Chebyshev
polynomial is denoted by T
r
(x), where 1 < x < 1. The rst
few Chebyshev polynomials are given by
T
0
(x) = 1
T
1
(x) = x
T
2
(x) = 2x
2
1
T
3
(x) = 4x
3
3x
T
4
(x) = 8x
4
8x
2
+ 1. (4)
The higher order Chebyshev polynomials may be generated
using the recursive formula
T
r+1
(x) = 2xT
r
(x) T
r1
(x). (5)
Using the Chebyshev polynomials, an n-dimensional input pat-
tern is expanded to an l-dimensional (l > n) enhanced pattern,
which is then applied to a single-layer perceptron. In the ChNN
schematic shown in Fig. 2, the values of n and l are 2 and 9,
respectively. The advantage of the ChNN over MLP is that the
Chebyshev polynomials are computationally more efcient and
takes much less time to train compared to an MLP network.
C. Learning of ChNN
The learning process involves updating of the weights (W) of
the ChNNin order to minimize a given cost function. Agradient
descent algorithm is used for learning where the gradient of a
cost function with respect to the weight is determined and the
weights are incremented by a fraction of the negative gradient
at each iteration. The well-known BP algorithm [11] is used to
update the weights of the ChNN. Considering the single-output
ChNN as shown in Fig. 2, its output is given by
y = tanh(s) (6)
where the linear sum s is given by
s =
l

i=1
w
i
T
i
(X) (7)
where X R
n
, X = [x
1
, x
2
, . . . , x
n
]
T
is the input vector,
T
i
(X), {i = 1, 2, . . . , l}, is an element of the expanded input
vector and W = [w
1
, w
2
, . . . , w
l
]
T
is the weight vector of the
ChNN.
The learning algorithm minimizes E
k
, the cost function at
kth instant
E
k
=
1
2
[d
k
y
k
]
2
=
1
2
e
2
k
(8)
where d
k
is the desired output at the kth instant, and e
k
denotes
the error term. In each iteration, an input pattern is applied, the
output of the ChNN is computed, and the error e
k
is obtained.
The error value is then used in the BP algorithm to minimize
the cost function until it is less than a predened small value.
The weights are updated as follows:
W
k+1
= W
k
+W
k
= W
k
+
_

E
k
W
k
_
(9)
where the learning parameter is set between 0 and 1. The
gradient of the cost function (8) is given by
E
k
W
= e
k
y
k
W
. (10)
The update rule for the weight w
i
is given by
w
i,k+1
= w
i,k
+e
k
y
k
w
i
. (11)
Since tanh(.) function is used at the output node, the update
rule becomes
w
i,k+1
= w
i,k
+e
k
(1 y
k
)
2
T
i
(X). (12)
To improve convergence, a momentum term is added to the
update rule as follows:
w
i,k+1
= w
i,k
+w
i,k
+w
i,k1
(13)
where the momentum factor is set between 0 and 1.
The mean square error (MSE) during training in dB at the kth
iteration is dened as
MSE
train
(k) = 10log10

1
N
p
N
L
N
p

p=1
N
L

i=1
_
e
(p)
i
(k)

(14)
where N
p
denotes the number of training patterns used. In this
paper, the number of output nodes N
L
= 1.
PATRA: CHEBYSHEV NEURAL NETWORK-BASED MODEL FOR DUAL-JUNCTION SOLAR CELLS 135
Fig. 3. ANN-based schematic diagram for estimation of external quantum
efciency of a DJ solar cell.
D. Estimation of Solar Cell Characteristics
An ANN-based schematic diagram to estimate the EQE of
a DJ solar cell is shown in Fig. 3. The input to the ANN is
the wavelength () of irradiation and its target output is the
measured EQE (q). A set of input and target output (taken from
the experimental measurements) is given to the ANN during
the training phase. After application of each input, the ANN
computes its output ( q) that is then compared with the target
output to produce an error. This error is used in the BP algorithm
to update the weights of the ANN. After a sufcient number of
iterations, the MSE (14) between the target and ANN outputs
settles down to a minimum value. Thus, the training of the ANN
is completed and now the ANN is ready for estimation of solar
cells EQE. In a similar way, for estimation of the TJ and IV
characteristics, different sets of input-target data sets are used
during training of the ANN.
III. MODELING OF TUNNEL JUNCTION CHARACTERISTICS
One of the reasons of the lack of accurate modeling tech-
niques for MJ solar cells is due to the difculties in modeling of
highly complex phenomena in tunnel diodes. Here we provide
modeling of a tunnel diode using a the ChNN. For this modeling
purpose, we have taken the experimental and ATLAS-simulated
data form [22].
A. Principles of Tunneling Phenomenon
In a highly doped pn junction, tunneling of charge carriers
through the bandgap is an important part of transport mech-
anism. The two fundamental tunneling mechanisms in a pn
junction are: 1) the direct transition from band to band and
2) the transition via traps, i.e., trap-assisted tunneling (TAP). A
schematic of the two tunneling mechanisms is shown in Fig. 4.
The principles of band-to-band tunneling phenomenon can be
explained by using the Kane model [28] in which the tunneling
current through a potential barrier is described analytically. The
tunneling probability for charge carriers in a parabolic potential
barrier is given by
T
t
exp
_
2
_
r
2
r
1
_
2m
h
2
(V (r) E)
_
1/2
dr
_
(15)
where r
1
and r
2
are the turning points of the wave functions,
m is the tunneling mass, and V (r) is the potential barrier. The
tunneling current can be computed using the FermiDirac distri-
Fig. 4. Schematic of tunneling mechanisms.
bution f
C
(E) and f
V
(E), and the density states of the electrons
D
n
(E) and holes D
p
(E), as follows:
J
t

_
E
V
E
C
[f
C
(E) f
V
(E)] T
t
D
n
(E)D
p
(E)dE. (16)
From (16), one can see that the tunneling current depends on
the tunneling probability and the supplied functions for the elec-
trons and holes. These are implemented in device simulators for
numerical simulation of the tunneling current. The device sim-
ulator computes the carrier transport in the device by solving
the electronhole continuity equations and the Poisson equa-
tion. The tunneling mechanism is incorporated by an additional
recombination term in the continuity equation that is done by
adopting either a local or a nonlocal tunneling model [6].
B. Performance Comparison
In order to compare the performance of the ATLAS-modeled
tunneling behavior, an isolated p
++
-GaAS/n
++
GaAS tun-
nel diode grown by the metal organic vapour phase epitaxy
(MOVPE) technique has been investigated. The simulated TJ
structure used in the ATLAS is shown in Fig. 5 [22]. The tunnel
diode consists of a TJ made by two degenerately doped p
++
-
and n
++
-GaASlayers along with two enclosing barrier layers, in
order to minimize the doping effect. The CAP and the substrate
layers were also included in the simulation. Local and nonlocal
TAT tunneling mechanisms have been included in the simu-
lation using the ATLAS. The results of the ATLAS-simulated
TJ characteristics and the experimentally measured values are
plotted in Fig. 6. It can be seen that the ATLAS response and
experimental values match quite closely.
In our ChNN-based model, the 1-D input pattern was ex-
panded to a 12-D pattern (including bias unit) by using
Chebyschev polynomials. Thus, the ChNN architecture is rep-
resented as {121}. We selected 60% of the experimentally
measured data points as training set to train the ChNN. The
input voltage (V) and the tunnel diode current (A/cm
2
), respec-
tively, were used as input and target output to the ChNN. Both
the input and target output to the ChNN were normalized be-
tween 0.9 and 0.9. The learning parameter and momentum
factor used in the BP algorithm were selected as 0.16 and 0.5,
136 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 26, NO. 1, MARCH 2011
Fig. 5. TJ structure used in ATLAS simulation.
respectively, after several experiments to give the best results.
The training continued for 50 000 iterations that resulted in an
MSE (14) value of 27.8 dB. To verify the effectiveness of the
ChNN-based model, the whole measured data set was applied
to the trained ChNN. The TJ characteristics estimated by the
ChNN-based model are shown in Fig. 6.
To determine effectiveness of the models quantitatively, we
computed correlation coefcient (CC) between the experimen-
tal and the model-predicted outputs. The CC between the ex-
perimental and ChNN outputs is denoted as CC
ChNN
, and the
CC between the experimental and ATLAS outputs is denoted
as CC
ATLAS
. Another quantitative measure we used for per-
formance comparison is the MSE between the experimentally
measured data and the estimated output from the ATLAS sim-
ulator or the ChNN-based model. The MSE in dB between the
experimental data and the ATLAS data is given by
MSE
ATLAS
= 10 log10
_
1
N
N

i=1
[x
i
y
i
]
2
_
(17)
where x
i
and y
i
denote the experimental and ATLAS-estimated
data, respectively. Similarly, MSE
ChNN
is dened.
The estimated characteristics and scatter plot along with the
CC and MSE performances of the ChNN-based model are also
shown in Fig. 6. It can be seen that the ChNN-based model
provides better performance than the ATLAS model in terms of
both CC and MSE.
IV. MODELING OF THE DJ SOLAR CELL
In this paper, the main objective of simulation of the tun-
nel diode is to use the TJ as an interconnection between the
subcells of a DJ solar cell. After successful modeling of TJ
characteristics, we study further to develop the ChNN-based
models for the DJ solar cell. For this purpose, we have investi-
gated a GaInP/GaAs DJ cell simulated by the ATLAS using the
layer thickness and doping levels as given in Table I. We used a
ChNN-based model to estimate the EQE and IV characteristics
both at one sun and dark levels.
After several simulations (as discussed in Section IV-E), we
found the optimum parameters of the ChNN as follows: the
number of input N
input
= 12, learning rate, = 0.16, momen-
Fig. 6. Estimated TJ characteristics obtained from (a) and (b) ATLAS and (c)
and (d) ChNN models.
TABLE I
STRUCTURE OF THE DJ SOLAR CELL
tum factor, = 0.5, the number of iterations, N
nitr
= 50 000
and percent of the measurement data used for training N
trgsample
= 60%. These optimum values were used in the following
sections.
A. Modeling of External Quantum Efciency
The EQE is an important parameter in design and devel-
opment of solar cells. It depends on the physical properties,
structure, and thickness of the materials used, and the incident
irradiation wavelength, in a complex manner. The EQE is de-
ned as
EQE = SR()
h
q
(18)
where SR() is the spectral response (in A/W) at the incident
wavelength , h is the energy of the incident photon, and q is
the electronic charge.
In a GaInP/GaAs DJ cell, due to the resonant cavity effect
occurring between the top cell back surface eld (BSF) and the
bottom cell window layer, the EQE suffers from oscillations in
the bottom GaAs cell. In order to accurately model the EQE
in a DJ cell, the ATLAS incorporates a characteristics matrix
PATRA: CHEBYSHEV NEURAL NETWORK-BASED MODEL FOR DUAL-JUNCTION SOLAR CELLS 137
Fig. 7. Estimated EQEs obtained through (a) and (b) ATLAS and (c) and (d)
ChNN models.
approach that relates total tangential components of the electric
and magnetic elds at the multilayer boundaries [22]. The ex-
perimental results and the ATLAS model output are plotted in
Fig. 7.
We modeled EQE of the DJ solar cell with a ChNN using
the optimum parameters. The radiation wavelength () and the
computed EQE from the experimental measurement are the in-
put and the target output to the ChNN, respectively. The training
continued for 50 000 iteration to achieve an MSE level (14) of
29.7 dB. After completion of training, all the data points were
used to estimate the EQE. The results of the ChNN-based model
are also shown in Fig. 7. It can be seen that the ChNN-based
model performs better than the ATLAS model in terms of both
MSE and CC.
B. Modeling of IV Characteristics at One Sun
The experimentally measured IV characteristics of the DJ
solar cell at one sun along with the ATLAS-simulated char-
acteristics are shown in Fig. 8. It can be seen that there is a
good amount of deviation between the two characteristics. We
used a two-layer ChNN {121} with the optimum parameters
to model the IV characteristics. The input and the target output
to the ChNN, respectively, were the voltage (in V) and current
density (in mA/cm
2
). The training data set was applied and up-
dating of ChNN weights were carried out for 50 000 iterations
to produce an MSE level (14) of 28.1 dB. The testing of the
ChNN was done with all the experimental data. The results of
the ChNN model are shown in Fig. 8. The response charac-
teristics of the ChNN model almost overlap the experimental
characteristics. This fact indicates the excellent performance of
the ChNN model.
Fig. 8. Estimated IV characteristics at one sun obtained through (a) and (b)
ATLAS and (c) and (d) ChNN models.
Fig. 9. Estimated IV characteristics at dark obtained through (a) and (b)
ATLAS and (c) and (d) ChNN models.
C. Modeling of IV Characteristics at Dark
The experimentally measured IV characteristics of the DJ
solar cell at dark along with the ATLAS-simulated characteris-
tics are shown in Fig. 9. The ChNN was trained in a similar way
as that at one sun. Once again, the response of the ChNN model
is found to be superior to that of the ATLAS model, in terms of
both CC and MSE.
Performance comparison of the ChNN-based models for the
DJ solar characteristics is shown in Table II. It can be seen that
for all the four DJ solar characteristics the ChNN-based models
perform better than the ATLAS model in terms of both CC and
MSE.
D. Computation of DJ Solar Cell Parameters
We computed the major parameters the DJ solar cell, e.g., J
SC
,
V
OC
, FF, and , from the IV characteristics and these values are
tabulated in Table III. It may be noted that the experimentally
138 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 26, NO. 1, MARCH 2011
TABLE II
PERFORMANCE COMPARISON BETWEEN THE CHNN- AND ATLAS-BASED
MODELS
TABLE III
COMPARISON OF COMPUTED DJ SOLAR CELL PARAMETERS
Fig. 10. Sensitivity analysis of ChNN-based models in terms of correlation
coefcient. (a) Alpha. (b) Iteration. (c) Input node. (d) Sample data.
measured values were taken under the standard test conditions:
the air mass (AM) 1.5 spectrum with an incident power density
of 1000 W/cm
2
and a temperature of 25

C. It can be seen that

the parameters obtained through the ChNN-based model match
closely to that obtained through experiment.
E. Sensitivity Analysis of ChNN-Based Model
In this section, we study the effect of variation of ChNN
parameters on the performance of the ChNN-based model in
predicting the DJ solar cell characteristics. We varied the learn-
ing parameter (), number of training iterations N
nitr
, number
of input nodes (N
input
), and percent of measurement data used
for training (N
trgsample
), one at a time, while keeping the other
parameters at their optimum values. The results of this analysis
with CC performance measure are shown in Fig. 10.
The four DJ solar cell model characteristics are denoted as
TJ (tunnel junction characteristics), EQE, IVSun (IV charac-
teristics at one sun), and IVDark (IV characteristics at dark).
For the four characteristics, as increases from 0.05 to 0.50
Fig. 11. Sensitivity analysis of ChNN-based models in terms of MSE. (a)
Alpha. (b) Iteration. (c) Input node. (d) Sample data.
the CC remains steady about 0.995 for TJ, EQE, and IVDark.
However, the IVSun characteristics seem to be quite sensitive to
variation of . The maximum value of CC (0.9983) for IVSun
characteristics was obtained at = 0.16. As the number of it-
erations increased from 5000 to 60 000, the CC increases from
0.9861 to 0.9981 for IVSun characteristics; however, for other
three characteristics, it remains steady at 0.995. The inuence of
the number of input nodes, i.e., the number of functional expan-
sion, is quite prominent for all the four characteristics. It can be
seen that the at N
input
= 12, the CC is optimum for all the four
characteristics. Finally, it can be seen that the percent of sample
data used for training the ChNN affects the CC performance
substantially. The optimum value of CC was obtained when
(N
trgsample
= 60%). The MSE performance with variation of the
ChNN parameters is shown in Fig. 11. At the selected optimum
ChNN parameters ( = 0.16, N
nitr
= 50 000, N
input
= 12, and
N
trgsample
= 60%), the MSE performance is found to be quite
satisfactory for all the four ChNN models.
V. CONCLUSION
In this paper, we proposed a novel ChNN-based technique
for modeling of complex DJ solar cell characteristics. Using
single-layer ChNNs, we have modeled the TJ behavior and es-
timated the EQE and IV characteristics at one sun and dark
levels with quite satisfactory results. With extensive computer
simulations, we have shown that the ChNN-based models could
achieve superior performance than the models generated by the
device simulator, the ATLAS by Silvaco [9]. Especially, for the
estimation IV characteristics, the ChNN-based models match
perfectly to that of the experimental characteristics. The electri-
cal parameters, e.g., J
SC
, V
OC
, J
max
, V
max
, FF and , computed
from the IV characteristics match closely to those obtained
from the experimentally measured characteristics. Due to their
low-computational complexity, the ChNNs are preferable in
modeling of complex and nonlinear phenomena. Furthermore,
because of their exibility and adaptability, the ANNs can be
used for other modeling aspects of MJ solar cells.
PATRA: CHEBYSHEV NEURAL NETWORK-BASED MODEL FOR DUAL-JUNCTION SOLAR CELLS 139
ACKNOWLEDGMENT
The author would like to thank the anonymous reviewers
whose positive and constructive comments and suggestions
helped to enhance the quality of this paper to a great extent.
REFERENCES
[1] A. Cuevas and J. Tan, Analytical and computer modelling of sunsVoc
silicon solar cell characteristics, Solar Energy Mater. Solar Cells, vol. 93,
pp. 958960, 2009.
[2] A. S. Lin, W. Wang, and J. D. Phillips, Model for intermediate band solar
cells incorporating carrier transport and recombination, J. Appl. Phys.,
vol. 105, pp. 064512-1064512-8, 2009.
[3] L. Hsu and W. Walukiewicz, Modeling of InGaN/Si tandem solar cells,
J. Appl. Phys., vol. 104, pp. 024507-1024507-7, 2008.
[4] S.-I. Sato, H. Miyamoto, M. Imaizumi, K. Shimazaki, C. Morioka,
K. Kawano, and T. Ohshima, Degradation modeling of InGaP/GaAs/Ge
triple-junction solar cells irradiated with various-energy protons, Solar
Energy Mater. Solar Cells, vol. 93, pp. 768773, 2009.
[5] A. Cheknane, H. S. Hilal, J. P. Charles, B. Benyoucef, and G. Campet,
Modelling and simulation of InGaP solar cells under solar concentration:
Series resistance measurement and prediction, Solid State Sci., vol. 8,
pp. 556559, 2006.
[6] M. Hermle, G. Letay, S. P. Philipps, and A. W. Bett, Numerical simulation
of tunnel diodes for multijunction solar cells, Progr. Photovoltaics: Res.
Appl., vol. 16, pp. 409418, 2008.
[7] M. Burgelman, M. Minnaert, and C. Grasso, Device modeling of nanos-
tructured solar cells, in Nanostructural Materials for Solar Energy Con-
version. T. Soga, Ed. Amsterdam, The Netherlands: Elsevier, 2006,
pp. 4580.
[8] B. Kuhlmann, A. G. Aberle, R. Hezel, and G. Heiser, Simulation and op-
timization of metalInsulatorSemiconductor inversion-layer silicon solar
cells, IEEE Trans. Electron. Devices, vol. 47, no. 11, pp. 21672178,
Nov. 2000.
[9] SILVACO. [Online]. Available: http://www.silvaco.com/
[10] Synopsys. [Online]. Available: http://www.synopsys.com/tools/tcad/
pages/ffdefault.aspx
[11] S. Haykin, Neural Networks, 2nd ed. Upper Saddle River, NJ: Prentice-
Hall, 1999.
[12] A. Mellita and S. A. Kalogirou, Articial intelligence techniques for
photovoltaic applications: A review, Prog. Energy Combustion Sci., vol.
34, pp. 574632, 2008.
[13] M. Abdulhadi, A. M. Al-Ibrahim, and G. S. Virk, Neurofuzzy-based solar
cell model, IEEE Trans. Energy Convers., vol. 19, no. 3, pp. 619624,
Sep. 2004.
[14] A. Varnham, A. M. Al-Ibrahim, G. S. Virk, and D. Azzi, Soft-computing
model-based controllers for increased photovoltaic plant efciencies,
IEEE Trans. Energy Convers., vol. 22, no. 4, pp. 873880, Dec. 2007.
[15] F. Almonacid, C. Rus, L. Hontoria, M. Fuentes, and G. Nofuentes, Char-
acterisation of Sicrystalline PV modules by articial neural networks,
Renew. Energy, vol. 34, pp. 941949, 2009.
[16] E. Karatepe, M. Boztepe, and M. Colak, Neural network based solar cell
model, Energy Convers. Manage., vol. 47, pp. 11591178, 2006.
[17] T. Hiyama and K. Kitabayashi, Neural network based estimation of max-
imum power generation from PV module using environmental informa-
tion, IEEE Trans. Energy Convers., vol. 12, no. 3, pp. 241247, Sep.
1997.
[18] C.-C. Liao, Genetic k-means algorithm based RBF network for photo-
voltaic MPP prediction, Energy, vol. 35, no. 2, pp. 529536, 2010.
[19] D. Paul, S. N. Mandal, D. Mukherjee, and S. R. B. Chaudhuri, Articial
neural network modeling for efcient photovoltaic system design, in
Proc. IEEE Intl. Conf. Adv. Comput. Theory Eng., Phuket, Thailand, Dec.
2008, pp. 5056.
[20] W. Guter and A. W. Bett, IV characterization of tunnel diodes and
multijunction solar cells, IEEE Trans. Electron Devices, vol. 53, no. 9,
pp. 22162222, Sep. 2006.
[21] B. Galiana, C. Algora, I. Rey-Stolle, and I. Vara, A 3-D model for
concentrator solar cells based on distributed circuit units, IEEE Trans.
Electron Devices, vol. 52, no. 12, pp. 25522558, Dec. 2005.
[22] M. Baudrit and C. Algora, Modeling of GaInP/GaAs dual-junction solar
cells including tunnel junction, presented at the Proc. 33rd IEEE Photo-
voltaic Spec. Conf., St. Diego, CA, May 2008.
[23] Y.-H. Pao, Adaptive Pattern Recognition and Neural Networks. Reading,
MA: Addison-Wesley, 1989.
[24] J. C. Patra, M. Juhola, and P. K. Meher, Intelligent sensors using compu-
tationally efcient Chebyshev neural networks, IET Sci. Meas. Technol.,
vol. 2, no. 2, pp. 6875, Mar. 2008.
[25] J. C. Patra, W. B. Poh, N. S. Chaudhari, and A. Das, Nonlinear chaneel
equalization with QAMsignal using Chebyshev articial neural network,
in Proc. IEEE Int. Joint Conf. Neural Netw. (IJCNN2005). Montreal,
Canada, Aug. 1995, pp. 32143219.
[26] J. C. Patra and A. C. Kot, Nonlinear dynamic system identication using
Chebyshev functional link articial neural network, IEEE Trans. Syst.,
Man, Cybern., Part B-Cybern., vol. 32, no. 4, pp. 505511, Aug. 2002.
[27] K. S. Narendra and K. Parthasarathy, Identication and control of dynam-
ical systems using neural networks, IEEE Trans. Neural Netw., vol. 1,
no. 1, pp. 427, Mar. 1990.
[28] E. O. Kane, Theory of tunneling, J. Appl. Phys., vol. 32, pp. 8391,
1961.
Jagdish Chandra Patra (M97) received the Ph.D.
degree from Indian Institute of Technology, Kharag-
pur, India, in 1996.
He was a Reserch Fellow at the School of Elec-
trical and Electronic Engineering, Nanyang Techno-
logical University (NTU), Singapore, in 1999. He
is currently an Assistant Professor at the School of
Computer Engineering, NTU. His research interests
include intelligent information processing using neu-
ral networks in the areas of information security, sen-
sor networks, image processing, digital communica-
tion, bioinformatics and multijunction solar cell modeling. He has more than
100 publications in these areas of research.
Dr. Patra is a member the Institution of Engineers, India.