You are on page 1of 8

Renewable Energy 60 (2013) 71e78

Contents lists available at SciVerse ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Articial neural network-based model for estimating the produced


power of a photovoltaic module
lam c, S.A. Kalogirou d
A. Mellit a, b, *, S. Sag
a

Faculty of Sciences and Technology, Renewable Energy Laboratory, Jijel University, Ouled-Aissa, P.O. Box .98, Jijel 18000, Algeria
Unit de dveloppement des quipements solaires (UDES), Bousmail, Tipaza 42000, Algeria
c
Technical Education Faculty, Marmara University, Istanbul 34722, Turkey
d
Department of Mechanical Engineering and Materials Science and Engineering, Cyprus University of Technology, P.O. Box 50329, Limassol 3603, Cyprus
b

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 18 November 2012
Accepted 23 April 2013
Available online 18 May 2013

In this paper, a methodology to estimate the prole of the produced power of a 50 Wp Si-polycrystalline
photovoltaic (PV) module is described. For this purpose, two articial neural networks (ANNs) have been
developed for use in cloudy and sunny days respectively. More than one year of measured data (solar
irradiance, air temperature, PV module voltage and PV module current) have been recorded at the
Marmara University, Istanbul, Turkey (from 1-1-2011 to 24-2-2012) and used for the training and validation of the models. Results conrm the ability of the developed ANN-models for estimating the power
produced with reasonable accuracy. A comparative study shows that the ANN-models perform better
than polynomial regression, multiple linear regression, analytical and one-diode models. The advantage
of the ANN-models is that they do not need more parameters or complicate calculations unlike implicit
models. The developed models could be used to forecast the prole of the produced power. Although, the
methodology has been applied for one polycrystalline PV module, it could also be generalized for largescale photovoltaic plants as well as for other PV technologies.
2013 Elsevier Ltd. All rights reserved.

Keywords:
Photovoltaic module
Modelling
Produced power
ANN
Forecasting

1. Introduction
As reported by the IEA [1], global photovoltaic capacity has been
increasing at an average annual growth rate of more than 40% since
2000 and it has signicant potential for long-term growth over the
next decades. By 2050, PV will provide 11% of global electricity
production (4500 TWh per year), corresponding to 3000 GW of
cumulative installed PV capacity.
In countries like Turkey, photovoltaic research and development
activities are still mainly undertaken across a range of universities,
government and industry facilities and the projects are mainly
nanced by the research programme of State Planning Organization (DPT) and The Scientic & Research Council (TUBITAK) [2,3].
As the performance of photovoltaic systems is inuenced by the
magnitude of the insolation and atmospheric conditions, more accurate models of photovoltaic cell/module are required to estimate
the produced power and generally to analyse the photovoltaic
systems performance. As the modelling of photovoltaic cells/

* Corresponding author. Abdus Salam International Centre for Theoretical Physics, Strada Costiera, 11, 34151 Trieste, Italy. Tel.: 213 (0)551 998 982.
E-mail addresses: a.mellit@yahoo.co.uk, amellit@ictp.it (A. Mellit).
0960-1481/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.renene.2013.04.011

modules is one of the most essential areas in photovoltaics research,


numerous methods have been developed for modelling the IeV
characteristic and estimating the maximal power. These can be
generally classied into two types: explicit I f(V) and implicit
I f(I, V) models. Explicit models employ a simple analytical
expression based on assumptions and they need less computational
effort. However, implicit models are relatively more accurate than
the explicit ones, and they have the disadvantage of introducing a
series of parameters which are difcult or even impossible to obtain
from solar cells manufacturers (i.e., the series resistance, RS; the
shunt resistance, RSh; the dark saturation current, I0; photogenerated current Iph and the diode ideality factor, n). Even if
these parameters can be obtained empirically, designers of photovoltaic systems often nd difculties in applying such models [4].
Articial neural networks (ANNs), genetic algorithm (GA),
neuro-fuzzy inference system (ANFIS) and particle swarm optimization (PSO) techniques have been investigated in order to model
and extract the PV cell/module parameters, as well as to estimate
the maximum power.
In Ref. [5] the authors used a neural network to estimate the
maximum power generation from a PV module using environmental information. The proposed network can be utilized for the
prediction of the next days generation from the PV systems by

72

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

using the predicted information from the weather ofces. According to the authors, the proposed method gives more accurate prediction compared to the prediction obtained by using the
conventional multiple regression models.
An application of radial basis function (RBF) for solar array
modelling and maximum-power point (MPP) prediction was presented by Al-Amoudi and Zhang [6]. The proposed RBF model can
lead to energy saving and it can calculate MPPs accurately without
searching around the optimal power point.
Abdulhadi et al. [7] have developed a neuro-fuzzy model to
predict solar cell short-circuit current and open-circuit voltage.
According to the authors, the model can be extended beyond the
bounds of measured data by incorporating a priori knowledge
derived from theory and manufacturers data.
In Ref. [8] the authors used a neural network-based approach
for improving the accuracy of the electrical equivalent circuit of a
PV module. The equivalent circuit parameters of a PV module
mainly depend on solar irradiation and temperature. The
dependence on environmental factors on the circuit parameters
is investigated by using a set of currentevoltage curves. It is
shown that the relationship between these two parameters is
nonlinear and cannot be easily expressed by any analytical
equation.
Almonacid et al. [9] developed a method to obtain the characteristic curve of a PV module at any given condition using an MLP
taking into account that the voltage could be given as a function of
the current, irradiance and the module temperature. It has been
reported that, the proposed ANN introduces an accurate prediction
for Si-crystalline PV modules performance when compared with
the measured values.
A particle swarm optimization (PSO) algorithm and a cluster
analysis are performed to t the calculated currentevoltage characteristic of a PV module by Sandrolini et al. [10]. This approach
allows one to obtain a set of parameters, which is reasonable and
representative of the physical system.
In Ref. [11] the authors applied a genetic algorithm to identify
the electrical parameters of PV cells/modules, which were used to
determine the maximum power. According to the authors, the GA is
a very efcient technique compared to other methods.
Implementation of an intelligent photovoltaic module on
recongurable Field Programmable Gate Array (FPGA) was developed by Mekki et al. [12]. The authors designed an MLPphotovoltaic module, which permits the performance evaluation
of the PV module using only environmental parameters and involves less computational effort. According to the authors, the device can also be used for predicting the output electrical energy
from the PV module and for a real time simulation.
Almonacid et al. [13] developed an ANN model to generate the
IeV curves of thin-lm PV Copper Indium Selenide (CIS) modules
for any solar irradiance and module cell temperature. According to
the authors, the results obtained were very promising and the
developed ANN performs better than other conventional
techniques.
Generalised regression neural network (GRNN) used to predict
the operating current of the photovoltaic module was developed in
Ref. [14]. The proposed GRNN model accepts as an input the PV cell
temperature, irradiance and PV voltage, while the PV current was
the output of the network. Results demonstrated that the GRNN
provides a better prediction of the current than a ve-parameter
analytical model.
In Ref. [15] the authors used an ANFIS for modelling and simulation of photovoltaic power supply system. The ANFIS was developed to model the delivered and consumed power generation by
the PV power supply (PVPS) system. It has been demonstrated
that, the developed model can predict and simulate the different

electrical data of the PVPS-system from only the ambient temperature, solar irradiation and clearness index.
RBF networks are utilized to predict the output characteristic of
a commercial PV module, using as an input the solar irradiance and
temperature [16]. Results show that the numerical values of the
computed IeV and PeV characteristics match closely those obtained from the experimental data. The RBF network can also be
used for other modelling purposes of solar cells such as the ve or
seven circuit parameters estimation.
Recently, in Ref. [17] a novel methodology based on articial
neural networks is proposed to determine the IeV curve of a PV
module operating under different conditions. The main contribution consists in incorporating the measurement of the spectrum as
an input of the model. According to the authors, the performance of
the network trained with spectral information improves over the
one without spectral information.
As can be concluded from the above brief review, modelling of
PV cell/array based on articial intelligence techniques such as
ANNs, GA, PSO, Neuro-Fuzzy, etc., was applied in different circumstances. These techniques have been proved more benecial
than classical models specically from the point of view of
simplicity and accuracy.
The main objective of this paper is to develop a simple and accurate ANN-model taking into account the kind of day (cloudy or
sunny) and then to examine its capability in order to estimate the
prole of the power produced for a 50 Wp Si-polycrystalline PV
module. For this purpose, two ANN-based models have been
investigated, the rst one (ANN-model 1) is used to estimate the
power produced in the case of cloudy days, and the second one
(ANN-model 2) is used for sunny days. To assess the performance of
the designed models, a comparison between polynomial regression, one-diode, analytical and multiple linear regression models is
carried out.
The rest of this paper is organized as follows. Section 2 provides
information on the database used and the system description. The
methodology of developing ANN-based models is presented in
Section 3. Evaluation of ANN-models and a comparative study are
given in Section 4.
2. System monitoring and database
The system consists of one PV module (ASE-50-DG/16) and an
MPP unit connected to a load resistance of 100 U, 5 W as indicated
in Fig. 1. In this experiment, system current and voltage values are
measured with a power analyzer. The power analyzer is connected
to a PC via an RS232 serial port. To avoid problems of possible
power outages, the PC and the power analyzer are fed by a UPS.
Ohmmeter measurements show that set resistance value is 5.5 U
including contact resistance. A power analyser (Lutron DW-6090)
data-logger has been used for recording data every 2 min (PV
current and PV voltage).
A weather station (Davis Vantage Pro2 Plus) has been used for
recording the meteorological data with 5 min time intervals (air
temperature and solar irradiance). These have been recorded at the
Marmara University, Istanbul, Turkey from 1-1-2011 to 24-2-2012.
As an example of recorded data, Fig. 2 shows the evolution of solar
irradiance, air temperature, PV power, PV current and PV voltage
(20 cloudy days and 16 sunny days).
3. Models development
An ANN-based schematic block diagram used to estimate the
prole of the power produced of the PV module is depicted in Fig. 3.
The employed ANN has 3 layers, an input layer, a single hidden
layer and an output layer. The input layer has 2 inputs: solar

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

73

Fig. 1. Photo of the monitoring photovoltaic system with weather station.

20 cloudy days

irradiance and air temperature, and its output layer has a single
output node which is the power produced by the PV module.
Therefore, the problem is to nd a relationship between the inputs
and the output based on the experimental data included in the
database described before. In other words, the problem can be
formulated as, whether it is possible to nd a simple formula,
which can be useful to estimate the power produced based on solar
irradiance and air temperature.
The investigated relationship can be given as:

16 sunny days

Irradiance (W/m)

1000

500

Air temperature (C)

PV Power (W)
PV Current (A)

100

200

300

400
500
Time, (h)

600

700

800

900

~ ~f G; T
P

(1)

where, ~f is an approximate function.


Pre-processing of the data included in the dataset is carried out
by using the following expression:

40
30
20

y ymin x  xmin xmax  xmin 1 ymax  ymin

(2)

10
0

PV Voltage (V)

40

100

200

300

400
500
Time, (h)

600

700

800

900

100

200

300

400
500
Time (h)

600

700

800

900

20
0

4
2
0

100

200

300

400
500
Time (h)

600

700

800

900

100

200

300

400
500
Time (h)

600

700

800

900

20
10
0

Fig. 2. An example of monitored data (solar irradiance, air temperature, PV power, PV


current and PV voltage).

where xmin  x  xmax and ymin  y  ymax, x is the original data


value and y is the corresponding normalized variable. The parameters ymin and ymax have been assumed to 1 and 1 respectively.
This pre-processing step on the data (input/output) allows the
network to perform more efciently.
From the dataset described above, two sub-databases have been
extracted:
- The rst one consists of data (T, G and P), which has met the
condition that the average daily solar irradiation is lower or
equal 400 W/m2/day (considered as cloudy days). In this subdatabase we have 5760 samples, 70% of samples have been
used for training the network (ANN-model 1), while the rest
30% were used to validate the network. The majority of these
days are from the period from 16th September to 14th May.
- The second set includes data, which has a condition that the
average daily solar irradiation is higher than 400 W/m2/day
(considered as sunny days). In this sub-database we have 2280
samples, 70% of samples have been used for training the
network (ANN-model 2), while the rest 30% were used to
validate the network. The majority of these days are from the
period from 15th May to 15th September.

74

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

4. Results and discussion


4.1. Evaluation of the ANN-models
Different architectures have been evaluated and the best one
was chosen by trial and error. The optimized one consists of 2 units
in the input layer, 3 units in the hidden layer and one unit in the
output layer as shown in Fig. 4.
This structure is applicable for both ANN-models with different
weights and bias. Thus, the relationship between the inputs and
output can be formulated as:

~
P

Fig. 3. ANN-based schematic block diagram.

3
X
i1

w00i ai b00

(6)

where, ai G; T 21 exp2w0i;1 G w0i;2 T b0i 1  1, i 1, 2, 3


Therefore, we try to develop two ANN-models:
-

ANN-model 1 used to estimate the prole of power produced


in the case of cloudy days.
ANN-model 2 used to estimate the prole of the power produced in the case of sunny days.

A set of pre-processed input and output data is used to train the


network during the training step. After application of each input,
the network computes its output (P) which is then compared with
the target output to produce an error (e). The performance function
used for training feed-forward neural networks is the mean
~ given
squared error with regularization performance function E
as [18]:

0
1
n
n
X
1@ X
2
2A
~
g
E
ei 1  g
wj
n
i1
j1

(3)

where g is the performance ratio.


This modied error is used in the LevenbergeMarquardt optimization to update the weights and biases of the network. After
sufcient number of iterations, the mean square error between the
target and network outputs settles down to a minimum value.
Linear, logarithmic sigmoid and hyperbolic tangent sigmoid are
the three most common transfer functions. In this study, a linear
function is used for the output layer while a hyperbolic tangent
sigmoid transfer function is used for the input and the hidden
layers given by:

2

1e

2
PN
i1

1

(4)

wi xi b

w00i ; w0i;1 ; w0i;2 are the weights and b00 ; b0i are the bias values of the
network.
The weight and bias values for both ANN models are reported in
the Appendix.
In order to examine the capability of the designed ANN models
to estimate the prole of the produced power accurately, four days
for each class (sunny and cloudy) have been considered. These have
not been used for the training of the ANN. Fig. 5 shows the evolution of the performance error for both ANN-models. As can be
observed the mean squared error (measured versus estimated energy) during the training process is about 104 for the rst model,
while for the second model is about 107. These results indicate
that the network weights and bias of the networks are well
adjusted and the models could reproduce the output data with
good accuracy especially for the second model used in the case of
sunny days.
Fig. 6 depicts the superposition curves between the monitored
and estimated proles of the power produced by the PV module. As
can be seen, the monitored power values are relatively close to the
estimated ones for both ANN-models. However, to assess the performance of the designed ANN-models, the correlation coefcient
(R), the root mean square error (RMSE), and the mean bias error
(MBE) between monitored (actual) and predicted energy produced
(this value is the cumulative sum of the power values along the day)
are estimated.
These results are reported in Table 1.With reference to the rst
ANN-model 1 (cloudy days), it should be noted that the correlation
coefcient is between 93% and 97%, which means that, both
measured and estimated energy are relatively close. The MBE varies
in the range of 0.7 and 1.1%, and the RMSE is less than 0.2%.
Concerning the second model ANN-model 2 (sunny days), the
correlation coefcient is between 96% and 97%, the MBE varies in

The network weights and biases are updated based on the


following expression:


1
Xk1 Xk  J T J mI
E

(5)

where J is the Jacobian matrix that contains rst derivatives of the


network errors with respect to the weights and biases, E is a vector
of network error and m is the Marquardt adjustment parameter.
Thus, the training of the network is completed and now the
network is ready for evaluating the produced power. A soft
computing program for hourly estimation of the produced power of
the considered PV module is developed under MatLab (Ver. 7.8,
2009) [19].

Fig. 4. The multi-layer perceptron architecture used for both ANN models.

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

75

Fig. 5. Evolution of the performance error.

between 350 W/m2/day and 450 W/m2/day both models could


be used.
After the evaluation step, ANN models can be used with an
approximate error of 5% for estimating the prole of the produced
power. Furthermore, the developed ANN models can also be used to
forecast the prole of the produced power based on forecasted
solar irradiance and air temperature. In this case, numerical
weather prediction (NWP) models could be used to forecast the
solar irradiance and air temperature. The total error is the sum of
errors generated by the ANN model and forecaster model, so this
error should be considered in the estimation the power produced.
If the ANN-models are employed to forecast the power produced (based on the forecasted solar irradiation and air temperature by the NWP), the average daily solar irradiation can be
calculated in order to choose which ANN-model (cloudy or sunny
days) is suitable for each day.
Additionally, the designed ANN-models could be employed to
detect faults in the PV module based on the losses in power
(comparison between measured and predicted), and then a procedure could be integrated in order to detect which kind of fault is
occurred, (shadow, PV module degradation, dust accumulation,

the range of 0.94 and 0.98%, and the RMSE is also less than 0.2%.
Both are well within acceptable performance values.
With reference to the aforementioned results, it is clearly shown
that the second ANN-model 2 (sunny days) is relatively more accurate than the rst ANN-model 1 (cloudy days), which is
conrmed by the performance evolution of the MSE plotted in
Fig. 5.
The mean error of the ANN models is less than 2% and therefore
taking into account the accuracy in the measurement of solar
irradiance, air temperature, PV module current, voltage, and losses
in connection wires, which is approximately 3%, hence, the total
error is approximately 5%.
It should be noted that, the classication of the days into sunny
and cloudy, improves signicantly the ANN-based models, unlike a
single database containing all days, reported in almost all reviewed
papers that used ANNs.
Therefore, the ANN-model 1 could be applied in the period from
16th September to 14th May, whereas the ANN-model 2 is suitable
for the period from 15th May to 15th September. However, in some
days (frequently occurred in spring and autumn period) where the
mean average daily solar irradiation is approximately in the range

Cloudy days: 15 -18 November 2011


14

Sunny days: 7- 10 August 2011

40

Estimated

Estimated
Monitored

12

30

25

ANN model 2

PV module power (W)

ANN model 1

10
PV module power (W)

Monitored

35

20
15

10
2

0
160

180

200

220
Time(h)

240

260

280

0
160

180

200

220
Time(h)

240

260

280

Fig. 6. Comparison between measured (monitored data) and estimated (ANN) prole of the produced power of the employed PV module for four days (cloudy and sunny).

76

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

Table 1
Comparison between measured (actual) and ANN predicted energy in both cloudy
and sunny day models.
Type of days

Mean daily
measured
energy
(kWh/day)

Cloudy days
1st day 15/11/2011
2nd day 16/11/2011
3rd day 17/11/2011
4th day 18/11/2011
Sunny days
1st day 7/8/2011
2nd day 8/8/2011
3rd day 9/8/2011
4th day 10/8/2011

Mean daily
estimated
energy
(kWh/day)

ANN-model 1
2.51
3.26
0.81
1.93
1.46
2.57
1.55
2.51
ANN-model 2
11.66
12.21
11.89
12.73
11.51
12.09
11.42
12.39

RMSE
(%)

MBE
(%)

R
(%)

(10). The designed models for sunny and cloudy days are given
as:


2
~ PR 0:16550:0016TCell G0:0371G0:0001G sunny days
P
0:18620:0018TCell G0:0328G0:0001G2 cloudy days
(12)

0.11
0.12
0.12
0.11

0.74
1.11
1.10
0.96

97.1
93.5
96.5
96.1

0.13
0.10
0.10
0.11

0.94
0.83
0.98
0.97

96.3
97.0
96.5
96.7

4.2.3. One diode model


The well-known equivalent circuit of a single junction solar cell
is shown in Fig. 7. The IeV characteristic of a photovoltaic module is
given by Markvart [21]:

 VIRs

V IRs
I Iph I0 enVt Ns  1 
Rsh

(13)

etc.). This application is more benecial for large-scale photovoltaic


plants.

where: Iph is the photocurrent of the cell; I0 is the dark saturation


current of the diode, n is the diode ideality factor, Vt is the thermal
voltage and Ns is the number of cells connected in series.
The thermal voltage is given by:

4.2. Comparative study

Vt

In this subsection, we try to compare the performance of the


designed ANN-models with different models (e.g. polynomial
regression, multiple linear regression, analytical and one-diode). A
brief review of each model is given below.

where: K is Boltzmanns constant, q is the charge of the electron,


TSTC  K.
As Eq. (13) is an implicit {I f(I, V)} formula, an iterative method
(NewtoneRaphson) is used to identify the ve parameters (I0, Iph, Rs,
Rsh and n) of the one-diode model, and then we calculate the
maximum power for different irradiance and PV cell temperature.
Fig. 8 shows the IeV and PeV characteristics of the photovoltaic
panel employed at STC conditions. The estimated maximum power
is 51.02 W, and the error between the nominal power provided by
the manufacturer on the datasheet and the one estimated by the
one-diode model is approximately 0.05%.
The values of the identied ve parameters of the PV
panel employed are: Rsh 675.9409 (U), Rs 0.1991 (U),
I0 3.3963  107 (A), Iph 3.3010 (A) and n 1.3925.

4.2.1. Multiple linear regression model


In multiple linear regression (MLR), the relationship between
the inputs and outputs is given as:

Yi a b1 Xi;1 b2 Xi;2 ::: bp Xi;p i

(7)

where a and bj are coefcients, Xi and Yi are the measured variables


and i are the errors. A least squares approach is used for estimating
the coefcients. In our case, we have two input variables and one
output variable, so that Eq. (7) can be written as:

PG; T a b1 G b2 T

(8)

The same database has been used for estimating the coefcients
of the multiple linear regression model given by Eq. (8). The
designed MLR models for sunny and cloudy days are given as:

~ MLR
P

0:2716 0:0415G 0:295T sunny days


0:7644 0:0157G 0:121T cloudy days

(9)

4.2.2. Polynomial regression model


A generic polynomial regression (PR) model to simulate the
performance of a selected PV system is given by International Energy Agency [20]:

P A B,Tcell ,G C,G D,G2

(10)

A K TSTC
q

(14)

4.2.4. Analytical model


A simplied algebraic equation was proposed in Ref. [22] to give
the maximum power:




G
Pref 1  g T  Tref
Gref

Subscript ref refers to standard testing conditions (Gref 1000


W/m2, Tref 25  C) and g is the maximum power correction factor
for temperature; it ranges from 0.005 to 0.003  C1 in crystalline silicon, whereas good results are achieved assuming
g 0.0035  C1.
In order to make a comparison between ANN-models and the
aforementioned models (multiple linear regression, polynomial
regression, analytical and one-diode model), two days (cloudy and
sunny) have been chosen that were not included in the subdatabases used in the training step.

where A, B, C and D are polynomial coefcients.


For variations in air temperature and irradiance, the cell temperature (in  C) can be estimated quite accurately with linear
approximation, given by:

Tcell T



NOCT  20
G
0:8

(11)

The same database has been also used for estimating the
coefcients of the polynomial regression model given by Eq.

(15)

Fig. 7. The one-diode equivalent circuit.

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

77

STC conditions

STC conditions

50
45

3.5

40
3
35
PV power (W)

PV current (A)

2.5

1.5

30
25
20
15

1
10
0.5

5
0

0
0

10

15
PV Voltage

20

25

10

15
PV Voltage

20

25

Fig. 8. The IeV and PeV characteristics of the photovoltaic module employed at STC conditions.

Simulation results are depicted in Fig. 9. As can be seen, the


prole of the power estimated by the ANN-models is very close to
the measured one for both days. Table 2 reports the mean relative
error and the correlation coefcient between the measured (actual)
and the estimated energy.

10
Measured
Polynomial model
ANN-model 1
One diode model
MLR model
Analytical model

9
8

PV Power (W)

Cloudy day :
22/01/2012

6
5
4
3
2
1
0

10

15

20

Time (h)

50
Measured
Polynomial model
ANN-model 2
One diode model
MLR model
Analytical model

45
40

PV Power (W)

35

Sunny day :
15/08/2011

With reference to Table 2, the following key statements can be


made:
- ANN-based models estimate the prole of the produced power
with reasonable accuracy.
- The ANN-model 2 (sunny days) is more accurate than the ANNmodel 1 (cloudy days)
- Polynomial regression, multiple linear regression and analytical models provide good results for sunny days since the correlation coefcient (R) is between 94% and 96% and the mean
relative error (MRE) does not exceed 5% but is not so accurate
for the cloudy days. The multiple linear regression model is the
least accurate.
- One-diode model provides nearly the same MRE (4.4%) for both
days, and it provides also good results for sunny days. At STC
conditions, the one-diode model provides good accurate results (the error was 0.05%), however, in outdoor tests the results are not as accurate.
- According to the MRE the models accuracy can be classied as:
B Sunny days: ANN-model, polynomial regression model,
analytical model, one-diode model and multiple linear
regression model.
B Cloudy days: ANN-model, multiple linear regression model,
one-diode model and analytical model polynomial regression model.
The effectiveness of the split of the available one-year dataset
into two different ones representative of two typical days (sunny

30

Table 2
Comparison between measured and estimated energy by ANN-models and other
models (polynomial regression, analytical, multiple linear regression and one-diode
model) for both days considered.

25
20
15

Models
10
5
0

10

15

20

Time (h)

Fig. 9. Measured versus estimated power by different models (polynomial, ANN, onediode, multiple linear regression and analytical models) in a day.

ANN-model
Polynomial regression
Multiple linear regression
One-diode model
Analytical model

Cloudy day
22/1/2012

Sunny day
15/8/2011

R (%)

MRE (%)

R (%)

MRE (%)

96.40
91.30
92.10
93.68
91.20

2.50
5.30
3.21
4.44
5.27

97.22
96.81
94.45
94.87
95.10

2.30
2.42
5.67
4.38
4.22

78

A. Mellit et al. / Renewable Energy 60 (2013) 71e78

and cloudy) has been shown. As it is proved, this enhanced the


prediction accuracy of the developed models.

ANN-model 2 (sunny day)

5. Conclusion

w0ij

A simple ANN architecture has been developed for modelling


and estimating the prole of the power produced by a 50 Wp
polycrystalline photovoltaic module. The capability of the ANN
for estimating the PV power produced has been veried with
reasonable accuracy (error is about 5%). It has been demonstrated that the ANN-models perform better than polynomial
regression, multiple linear regression, analytical and one-diode
models.
The developed ANN models can be applied to forecast the PV
power produced based on the forecasted solar irradiance and air
temperature (e.g. from forecaster or numerical weather prediction
models). Furthermore, if we consider the developed ANN models as
a basic particular model of the employed PV panel, they could be
used also to estimate the produced power output of PV arrays in
stand-alone or grid-connected photovoltaic systems in this
location.
The advantage of the ANN-models is that, they do not need
additional parameters, which are not always available, unlike
implicit models that need more complicated calculations and
parameters which are not readily available. For example, the implicit model needs calculations that are more complicated and
some parameters (like, Rs and Rsh) which are not given by the
manufacturer (in the datasheet). These parameters could be
estimated experimentally, which is difcult, or by a numerical
approach.
Simplicity, accuracy and practicability are the main advantages
of the developed ANN models when a sufcient quantity of
experimental data is available to estimate the PV power produced.
Although the method has been applied for polycrystalline silicon, it
could be also used for other type of photovoltaic technologies such
as a-Si, CIGS, CIS and CdTe.
Acknowledgements
The rst author would like to thank the ICTP, Trieste (Italy) for
providing materials and the computer facilities for achieving the
present work.
Appendix
ANN-model 1 (cloudy day)

0:0509
w0ij 4 0:0241
0:3236
w00i 3:7164
b0i 5:7524

3
0:0163
0:0127 5
0:0007

1:4567

2:1704

0:0695 

3:3267 T and b00 5:0774

4:3399
4 0:0523
0:6371

w00i 1:5126
b0i 219:30

3:2848

3
0:1201
0:0033 5
0:0285
20:2166

0:4951 

18:617 T and b00 20:5570

References
[1] IEA. Technology roadmap, solar photovoltaic energy. Paris: OECD/IED; 2010.
[2] Dincer F. Overview of the photovoltaic technology status and perspective in
Turkey. Renewable and Sustainable Energy Reviews 2011;15:3768e79.
[3] IIcli S, Cubukcu M, Colak M. PV technology status and prospects. In: International Energy Agency (IEA) e photovoltaic power systems programme (PVPS),
annual report 2009. p. 114e6.
[4] Massi Pavan A, Mellit A, De Pieri D, Lughi V. A study on the mismatch effect
due to the use of different photovoltaic modules classes in large-scale solar
parks. Progress in Photovoltaics: Research and Applications 2012. http://
dx.doi.org/10.1002/pip.2266.
[5] Hiyama T, Kitabayashi K. Neural network based estimation of maximum power generation from PV module using environmental information. IEEE
Transactions on Energy Conversion 1997;12:241e7.
[6] Al-Amoudi A, Zhang L. Application of radial basis function networks for solararray modelling and maximum power-point prediction. IEE Proceedings of
Generation Transmission and Distribution 2000;147:310e6.
[7] Abdulhadi M, Al-Ibrahim AM, Virk GS. Neuro-fuzzy based solar cell model.
IEEE Transactions on Energy Conversion 2004;19:619e24.
[8] Karatepe E, Boztepe M, Colak M. Neural network based solar cell model. Energy Conversion and Management 2006;47:1159e78.
[9] Almonacid F, Rus C, Hontoria L, Fuentes M, Nofuentes G. Characterisation of
Si-crystalline PV modules by articial neural networks. Renewable Energy
2009;34:941e9.
[10] Sandrolini L, Artioli M, Reggiani U. Numerical method for the extraction of
photovoltaic module double-diode model parameters through cluster analysis. Applied Energy 2010;87:442e51.
[11] Zagrouba M, Sellami A, Bouacha M, Ksouri M. Identication of PV solar cells
and modules parameters using the genetic algorithms: application to
maximum power extraction. Solar Energy 2010;84:860e6.
[12] Mekki H, Mellit A, Kalogirou SA, Messai A, Furlan G. FPGA-based implementation of a real time photovoltaic module simulator. Progress in Photovoltaics: Research and Applications 2010;18:115e27.
[13] Almonacid F, Rus C, Hontoria L, Muoz FJ. Characterisation of PV CIS module
by articial neural networks. A comparative study with other methods.
Renewable Energy 2010;35:973e80.
[14] Naci Celik A. Articial neural network modelling and experimental verication of the operating current of mono-crystalline photovoltaic modules. Solar
Energy 2011;85:2507e17.
[15] Mellit A, Kalogirou SA. ANFIS-Based modelling for photovoltaic power supply
system: a case study. Renewable Energy 2011;36:250e9.
[16] Bonanno F, Capizzi G, Graditi G, Napoli C, Tinab GM. A radial basis function
neural network based approach for the electrical characteristics estimation of
a photovoltaic module. Applied Energy 2012;97:956e61.
[17] Piliougine M, Elizondo D, Mora-Lpez L, Sidrach-de-Cardona1 M. Photovoltaic
module simulation by neural networks using solar spectral distribution.
Progress in Photovoltaics: Research and Applications, http://dx.doi.org/10.
1002/pip.2209.
[18] MacKay DJC. A practical Bayesian framework for back propagation networks.
Neural Computation 1992;4:448e72.
[19] MATLAB. The MathWorks (Neural Networks toolboxes), Inc.; Natick: 2010.
http://www.matworks.com.
[20] International Energy Agency. PVPS task 2, performance reliability and analysis of
photovoltaic systems. Available from: http://www.iea-pvps-task2.org/; 2008.
[21] Markvart T. Solar electricity. John Wiley & Sons Ltd; 1994.
[22] Menicucci D, Fernandez JP. Users manual for PVFORM: a photovoltaic system
simulation program for stand-alone and grid interactive applications. Albuquerque, USA: Sandia National Laboratories, SAND85-0376; 1988.

You might also like