Energy Conversion and Management 49 (2008) 33493355

Contents lists available at ScienceDirect

Energy Conversion and Management

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

Optimization and modeling of a photovoltaic solar integrated system

by neural networks
Mohd Sami S. Ashhab *
Department of Mechanical Engineering, The Hashemite University, Zarqa 13115, Jordan

a r t i c l e

i n f o

Article history:
Available online 22 July 2008
Keywords:
PV solar system
Neural networks
Modeling
Constrained optimization

a b s t r a c t
A photovoltaic solar integrated system is modeled with articial neural networks (ANNs). Data relevant
to the system performance was collected on April, 4th 1993 and every 15 min during the day. This input
output data is used to train the ANN. The ANN approximates the data well and therefore can be relied on
in predicting the system performance, namely, system efciencies. The solar system consists of a solar
trainer which contains a photovoltaic panel, a DC centrifugal pump, at plate collectors, storage tank,
a owmeter for measuring the water mass ow rate, pipes, pyranometer for measuring the solar intensity, thermocouples for measuring various system temperatures and wind speed meter. The complex
method constrained optimization is applied to the solar system ANN model to nd the operating conditions of the system that will produce the maximum system efciencies. This information will be very
hard to obtain by just looking at the available historical inputoutput data.
2008 Elsevier Ltd. All rights reserved.

1. Introduction
Jordan imports most of its energy demands in the form of petroleum products. In order to meet the countrys future increasing energy demands some of the conventional thermal energy systems
should be replaced by renewable energy systems. Thus, the development of renewable energy systems is very attractive. The share
of renewable energy in the total energy consumption was around
1% in year 2004 and is expected to rise up to 15% in year 2010 [1].
Solar energy is very promising in Jordan for its use in food canning,
paper production, air-conditioning and sterilization [2]. Water heating accounts for about 20% of the total energy consumption of residential buildings, and about 7% of business buildings [3]. Solar
heated water by at plate collectors is considered one of the commonly used, reliable and maintenance free sources of domestic
hot water in most of the countries [4]. Research in the area of solar
energy has been very active in Jordan due to its extreme importance.
An experimental study of the performance of a thermal photovoltaic solar integrated system which was designed and assembled
in Jordan has been carried out in [4]. The solar system consists of a
solar trainer which contains a photovoltaic panel, a DC centrifugal
pump, at plate collectors, storage tank, a owmeter for measuring
the water mass ow rate, pipes, pyranometer for measuring the solar intensity, thermocouples for measuring various system temperatures and wind speed meter. Further analysis of the system with
the aid of articial neural networks (ANNs) has been performed in
* Tel.: +962 5 390 3333x4639; fax: +962 5 382 6348.
E-mail address: sami@hu.edu.jo
0196-8904/$ - see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enconman.2007.10.036

[5]. In this paper, we utilize the ANN model of the solar system to
nd the optimum operating condition that will produce maximum
system efciencies. Historical inputoutput system data that was
collected experimentally is used to train an ANN that predicts
the collector, PV module, pump and total efciencies. The model
predicts the efciencies well and thus can be utilized to nd the
operating conditions of the system that will produce the maximum
efciencies. This is done using the complex method constrained
optimization [6] applied to the solar system ANN model. The optimum operating conditions will be very hard to obtain by just looking at the available historical inputoutput data. An ANN based
inverse controller is designed to achieve the optimum efciencies
which is difcult to design by just looking at the historical data.
The study and procedure can be carried out for other solar systems
towards achieving their best performance.
Articial neural networks (ANNs) have received a lot of attention in recent years due to their attractive capabilities in forecasting, modelling of complex nonlinear systems and control.
Applications of neural networks are numerous and include many
various elds. ANNs have been used for forecasting photovoltaic
solar integrated system efciencies [5], solar energy [7], space
weather [8], outdoor sound transmission [9], stream ow [10]
and wind waves [11].
In this paper, we deal with the inverse control problem of a
multi-input multi-output process whose historical inputoutput
data is available. The PV solar integrated system under consideration belongs to this set of processes. The objective of the controller is to determine the set of inputs that will drive the outputs of
the process to their desired target values. The available process

3350

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

data is used to train an articial neural net (ANN) model with one
hidden layer which gives a good approximation for continuous
functions. Control techniques can be applied to an ANN model
since it is an inputoutput mathematical description of a system.
To achieve our objective of driving the process outputs to their desired values we build a positive cost function that measures the
differences between the ANN outputs and their corresponding target values. We then use a modied version of the complex method
constrained optimization of [6] to minimize the cost function. The
resultant set of inputs will drive the ANN outputs as close as possible to their target values. An accurate ANN approximation is
essential in assuring that the controller achieves its objective.
The presented minimization algorithm offers the feature that it
can be applied to a subset of the inputs which gives the freedom
for setting the remaining inputs at values that satisfy other system
requirements.
2. Thermal system description and experimental data
The PV solar integrated system is shown in Fig. 1 and it consists
of the following components as numbered in the gure:
(1) A solar trainer STR-811/EV (Electronic Veneta) which contains a photovoltaic panel. The panel area is 72.5  31 cm2
and it consists of 36 single crystal half cells each of area
7.5  7.5 cm2. The nominal voltage of the panel is 7.5 V
whereas the current depends on the solar intensity. The current ranges from 0 to 1 A. Note that one photovoltaic panel
was used.
(2) A DC centrifugal pump: 14 X, 24 V, 2 A, 14 L/min, 5 m.
(3) Three at plate collectors with the following specications
for each collector: 1.24 m2 area, 4 bar maximum pressure,
2690 W/m2 capacity, 7 pipes with 0.5 in. diameter.
(4) A 60  100  50.5 cm storage tank.
(5) A owmeter for measuring the water mass ow rate in the
range 20100 g/s.
(6) Flexible pipes at the inlet and outlet of the pump.
(7) The Kipp and Zonen solar integrator pyranometer for measuring the solar intensity.
(8) Thermocouples for measuring various system temperatures.
(9) The edra ve wind speed meter. It measures the wind speed
in the temperature range 5 to 50 C.

The photovoltaic panel converts the solar energy to electrical

power that drives the DC pump. The pump drives the water in
the pipes to the at plate collectors passing through the owmeter.
The collectors heat the water by the solar energy and the heated
water ows into the storage tank which is connected to the DC
pump.
Data relevant to the system performance was collected on April,
4th 1993 and every 15 min during the day from 8:45 in the morning until 4:45 in the afternoon. The following measurements were
recorded: water inlet temperature to the collectors, water outlet
temperature of the collectors, ambient temperature, temperature
of the photovoltaic cells, solar intensity, wind speed, water mass
ow rate, photovoltaic panel current and voltage. Using this data
four different system efciencies were calculated, namely, the at
plate collectors efciency, the photovoltaic panel (module) efciency, the DC pump efciency and the overall system efciency.
The collector efciency is the ratio of heat energy gained by
water to the incident solar radiation on the collectors and is written as

gc

MC p T o  T i
I s Ac

where M is the water mass ow rate, Cp is water heat specic value,

To is the outlet water temperature, Ti is the inlet water temperature,
Is is the solar intensity and Ac is the solar collector area. The PV panel efciency is the ratio of power produced by the panel to the incident solar radiation on the panel, that is

gm

IV
I s Ac

where I is the module current and V is the module voltage. The efciency of the pump is calculated as

gp

qgQH
IV

where q is the water mass density, g is the gravity acceleration, Q is

the water volume ow rate and H is the pump head. The total efciency is dened as the product of the pump efciency and the
module efciency and is written as

gt gp gm

Fig. 1. Photograph showing the PV solar system.

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

3351

We will model later in the paper the PV solar system with articial
neural networks (ANNs) based on the above experimental data and
efciencies.
3. Neural network description
The articial neural net mathematical model that represents the
structure shown in Fig. 2 is dened as [12]

Y f U W o  tanhW i  U Bi Bo

where Y is a column vector which contains the q outputs of the process, U is a column vector that contains the p inputs of the process,
Wo is a matrix of size q  n that contains the weights of the neural
net model from the hidden layer to the outputs with n being the
number of neurons in the hidden layer, Wi is a matrix of size
n  p that contains the weights of the neural net model from the inputs to the hidden layer, Bi (not shown in Fig. 2) is a column vector
of size n that contains the biases from the inputs to the hidden layer
and Bo (not shown in Fig. 2) is a column vector of size q which contains the biases from the hidden layer to the outputs.
Each input uj, j = 1, 2, . . ., p has lower and upper bounds, Lbj and
Ubj, respectively. These bounds are calculated from the available
inputoutput data. Lbj is the minimum value of the jth input over
the given input data whereas Ubj is the maximum value of the jth
input over the same data. If all the inputs lie within their lower and
upper bounds then the estimated output by the ANN should lie
within the given output data range. The output bounds are dictated
by the ANN and the input bounds.
The weights and biases of the ANN are determined by training
with the historical inputoutput data. Backpropagation is an example of a training algorithm. The available data is divided into two
parts: one part is used for training the net whereas the other usually smaller part is used to test the performance of the ANN. The
number of hidden neurons n affects the performance of the neural
net over the training and test sets of data. More neurons make the
tting of data more accurate over the training region. It is more
important to check the generalization performance of the model
over the test set of data since it was not used to calculate the
parameters of the model. The number of nodes is usually chosen
by trying different values and selecting the one that gives best results over both the training and test regions.
4. Optimization of neural networks
The inverse control of a process is to nd the inputs to that process that will drive its outputs as close as possible to given desired
values. In order to solve this problem a mathematical model of the
process is needed. In this paper, we deal with processes whose historical inputoutput data are available. We model the process with

Fig. 3. Optimization algorithm demonstration.

articial neural networks. Optimization techniques are then used

to determine the inputs to the process that will drive its outputs
as close as possible to the given desired goal. Fig. 3 illustrates these
ideas. A desired signal D (which contains the same number of elements as in the output signal Y) is passed to the inverse controller.
The controller determines the input signal U to the process such
that the output Y of the neural net model is as close as possible
to the desired signal D. The controllers ability in driving the real
output vector Yr to its target value will depend on the neural net
model accuracy.
The controller operation is characterized by minimizing a positive cost function that measures the differences between the model
outputs and their desired values. If there is a set of inputs at which
the cost function is equal to zero (minimum) then this is an optimum input. Generally speaking, the outputs are not equal to their
target values at the initial inputs. Thus, the inverse controller coordinates the inputs or part of them to achieve the desired outputs if
possible. Optimization over a subset of the inputs is a feature of the
developed algorithm that leaves the freedom to the user to determine the remaining inputs such that other essential requirements
of the process are met.
Given a neural net model of a process in the form given in Eq.
(5), we dene the cost function J as the weighted sum

q
X

ki di  yi 2

where yi is the ith output of the net, di is the desired value of the ith
output and ki is the ith output weight that has a positive value
which depends on the scale and importance of that output. If all
the outputs are on the same scale and have the same importance
then all the weights are set at 1. The cost function is expressed
explicitly in terms of the model outputs. However, since the outputs
are functions of the process inputs through the neural net model
representation given in Eq. (5), the cost function is expressed
implicitly in terms of the inputs. Note that the cost function J is
non-negative and is equal to zero if and only if all of the q outputs
are at their desired values. In order to drive the outputs as close as
possible to their target values we need to minimize J. Optimization
techniques are best in tackling such a problem.
We are now ready to dene our inverse control problem: Given
a process with p inputs and q outputs and a neural net model for
this process. Assume that pc of the p inputs (pc is less than or equal
to p) are allowed to change by the controller. These are called the
control inputs. We denote the vector that contains the control inputs by Uc. It is required to minimize the cost function J subject
to the following constraints

Lbi 6 ui 6 Ubi ; i 1; 2; . . . ; p

Fig. 2. Neural net structure.

i1

We will use a modied version of the complex method described in [6] to solve this constrained optimization problem. The
algorithm will be applied to the control inputs only. The remaining

3352

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

inputs will stay constant at their given initial values. These inputs
are usually set at values (within their given constraints) that satisfy
some other desired conditions for the process. The complex method is based on the generation and maintenance of a pattern of
search points and the use of projections of undesirable points
through the centroid of the remaining points as the means of nding new trial points. In the presence of inequality constraints, it is
evident that if the new point is infeasible, then it is a simple matter
to retract it toward the centroid until it becomes feasible. The complex method was proposed as a better alternative to the simplex
direct search-method [6]. The considerations that led to the simplex modication are
1. The destruction of the regularity of the simplex.
2. Many of the points of the regular initial simplex will be
infeasible.
3. The points must be generated sequentially rather than being
essentially simultaneously dened using the formula for the
regular simplex.
It was proposed that the set of M trial points be generated randomly and sequentially. p samples are required to dene a point
in p dimensions. Each newly generated point is tested for feasibility, and if infeasible it is retracted toward the centroid of the previously generated points until it becomes feasible. The total
number to be used, M, should be no less than pc + 1 but can be
larger. The optimization algorithm can be summarized as follows
[6]:
Given an initial strictly feasible point Uo (set of inputs), reection parameter a, and termination parameters e and d.
(1) Generate the initial set of M feasible points. For each point
m = 1, . . ., M  1,
(a) Sample pc times to determine the point Um. Note that the
inputs that are not allowed to change are left constant at
their initial values.
(b) If Um is infeasible, calculate the centroid Ua of the current
set of points and reset

1
U m U m U a  U m
2
Repeat until Um becomes feasible.
(c) If Um is feasible, continue with (a) until M points are
available.
(d) Evaluate J(Um), for m = 0, 1, . . ., M  1.
(2) Carry out the reection step.
(a) Select the point UR such that

JU R max JU m  J max
(b) Calculate the centroid Ua and the new point

U n U a aU a  U R
(c) if Un is feasible and J(Un) is greater than or equal to Jmax,
retract half the distance to the centroid Ua. Continue until
J(Un) becomes less than Jmax.
(d) If Un is feasible and J(Un) < Jmax, go to step 4.
(e) If Un is infeasible, go to step 3.
(3) Adjust for feasibility.
(a) Reset violated variable bounds:If uni < Lbi , set uni Lbi . If
uni > Ubi , set uni Ubi .Note that the subscript i denotes
the ith input. This step will be applied only to the control
inputs since the other xed inputs are set within their constraints initially.

(b) If the resulting Un is infeasible, retract half the distance to

the centroid. Continue until Un is feasible, then go to step
2(c).
(4) Check for termination.
(a) Calculate

1 X m
U
M

and J

1 X
JU m
M

(b) If

X
JU m  J2 6 e or
m
X

kU m  Uk 6 d

Terminate. Otherwise, go to step 2(a).

The search is terminated when the pattern of points has shrunk
so that the points are sufciently close together and/or when the
differences between the cost function values at the points become
small enough. Numerical experiments has been performed with
this algorithm [6], and on this empirical basis it is recommended
to use a = 1.3 and M = 2pc. On the other hand, good results are reported with M = pc + 2. The a > 1 compensates for shrinking of the
complex caused by halving the distances, while the large number
of vertices is intended to prevent the complex from collapsing
and attening out along the constraints. The setting at the value
of the bound is intended to avoid unnecessary contraction of the
complex. The complex method requires that the feasible region
be a convex set. This requirement arises in two places: in the calculation of the centroid and in the retraction of a feasible but
unsatisfactory point. For nonconvex regions the method could fail
to converge satisfactorily. In practice, the method is widely used
and has successfully solved numerous nonconvex problems. Thus,
the above situation must arise with low probability.

5. Application to the PV solar integrated system

The proposed optimization algorithm is applied to the PV solar
integrated system discussed in Section 2. Based on the PV solar
system data (as described above) obtained in April 4th, 1993 in
Amman, Jordan we train a neural net model. The data contains
27 patterns. Each pattern consists of 7 inputs and four outputs.
The inputs are the water inlet temperature to the collectors, water
outlet temperature of the collectors, ambient temperature, temperature of the photovoltaic cells, solar intensity, wind speed and
water mass ow rate. Note that the photovoltaic panel voltage is
not considered as an input since is has a constant value of 7.5 V.
The PV panel current information is implicitly contained in the
water mass ow rate data. The system outputs are the at plate
collectors efciency, the photovoltaic panel (module) efciency,
the DC pump efciency and the overall system efciency. 21 of
the 27 patterns are used to train the articial neural net model
for the system, whereas the remaining 6 data patterns are used
to test the performance of the model. The training was done with
the software package Matlab. We ran experiments for different
numbers of hidden neurons. It was observed that the quality of
the results depends on the number of hidden neurons. We choose
the neural net with four hidden neurons. This is the smallest number of hidden neurons with acceptable least square errors over the
training and test regions. The real and predicted (ANN) values of
the various system efciencies over the training and test regions
are plotted in Figs. 47 as functions of the index of the time of
the day which is from 8:45 in the morning until 4:45 in the afternoon. Note that the predicted values are close to the real ones.

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

3353

Fig. 4. ANN training and test results for the collector efciency.

Fig. 5. ANN training and test results for the PV module efciency.

Therefore, the developed neural net model has acceptable

accuracy.
Examining the four efciencies as plotted in Figs. 47 we note
that the maximum collector efciency is 61.67% in the early morning, the maximum module efciency is 4.16% in the late afternoon,
the maximum pump efciency is 13.38% around the solar noon,

and the maximum total efciency is 0.42% in the late morning.

The maximum outputs occur around the solar noon as can be seen
from Fig. 8 which shows the basic experimental readings (outlet
temperatureinlet temperature, mass ow rate, solar intensity
and module current) as functions of the index of the time of the
day. These readings are the main factors that determine the

3354

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

Fig. 6. ANN training and test results for the pump efciency.

Fig. 7. ANN training and test results for the total efciency.

outputs (see Section 2). Note that the pump efciency occur also
around the solar noon since the losses do not change a lot with
temperatures for the pump, whereas the other optimum efciencies might move away from this solar noon zone due to high losses.
The results obtained above are very useful since the ANN model
can be relied on in predicting the system performance. In addition,
the ANN can be utilized to nd the operating conditions of the sys-

tem that will produce the maximum system efciencies. This information will be very hard to obtain by just looking at the available
historical inputoutput data. Knowing the operating conditions of
the PV solar system that will produce the optimum efciencies will
aid in the design of better and more efcient solar systems. An ANN
based constrained optimization algorithm was developed in Section 4 and can be used to achieve the optimum efciencies which

M.S.S. Ashhab / Energy Conversion and Management 49 (2008) 33493355

3355

Fig. 8. Basic experimental readings: outlet temperatureinlet temperature, mass ow rate, solar intensity and module current.

is difcult to design by just looking at the historical data. It is

worth to mention that this optimization algorithm has been utilized to nd the optimum operating conditions for a hot-wire thermal ow sensor [13] and a camless engine [14].
The inverse controller or optimization algorithm objective in this
application is to determine the water inlet temperature to the collectors, water outlet temperature of the collectors, ambient temperature, temperature of the photovoltaic cells, solar intensity, wind
speed and water mass ow rate that will achieve the best at plate
collectors, photovoltaic panel (module), DC pump and overall system efciencies. The desired values of the four efciencies are set
at 100% to make sure that we obtain their maximum values at the
end of the optimization algorithm. The accuracy of the articial neural network has been veried by Figs. 47 as we mentioned above.
This will guarantee the accuracy of the inverse controller since the
controller is based on optimizing the ANN model of the PV solar integrated system. An accurate ANN will make sure that the controller
produces the inputs which will achieve the best system performance. The inverse controller calculates the inputs as water inlet
temperature to the collectors = 21.80 C, water outlet temperature
of the collectors = 26.64 C, ambient temperature = 18.61 C, temperature of the photovoltaic cells = 26.22 C, solar intensity = 799.52 W/m2, wind speed = 2.59 m/s and water mass ow
rate = 78.92 g/s. The corresponding efciencies are at plate collectors = 60.72%, photovoltaic panel = 4.10%, DC pump = 12.86% and total = 0.47%. Note that the rst three efciencies are very close to their
maximum values and the total efciency exceeds its maximum value calculated from the data. We used e = d = 106 to terminate the
algorithm as described in the previous section. The optimization
method coordinates the inputs to compromise these efciencies
for best performance. In summary, the controller described in this
paper achieves its objective.
6. Conclusion
A PV solar integrated system is modeled with an articial neural
network. The net predicts well the collector, PV module, pump and

total efciencies as functions of various system measurements or

inputs. A modied version of the complex constrained optimization method was developed to nd the operating conditions that
will produce the optimum efciencies which is difcult to locate
by just looking at the historical data. The optimization algorithm
achieves its objective of producing the maximum efciencies.
The procedure can be applied to other solar systems in order to attain the best system performance.

References
[1] Agenda 21. Preliminary Report Prepared by Several Organizations, Amman;
2000.
[2] Odeh S, Nijmeh S, Ashhab MS, Zakaria Y, Amra A. Design of a single-axis
tracking parabolic collector for moderate temperature applications. In: ISES
conference, EuroSun, Freiburg-Germany; 2004. p. 52732.
[3] E, Crawly D, Centolella P. Energy efciency and the environment American
Council for an Energy-Efcient Economy (ACE); 1991.
[4] Hammad MA. Experimental study of the performance of a solar thermal
photovoltaic integrated system. Renew Energy 1994;4:897905.
[5] Ashhab MS. Thermal photovoltaic solar integrated system analysis using
neural networks. In: The
third international conference on
thermal
engineering: theory and applications, Amman-Jordan; 2007. p. 5936.
[6] Reklaitis G. Engineering optimization. New York: John Wiley & Sons; 2002.
[7] Reddy K, Ranjan M. Solar resource estimation using articial neural networks
and comparison with other correlation models. Energy Convers Manage
2003;44:251930.
[8] Vandegriff J, Wagstaff K, Ho G, Plauger J. Forecasting space weather: predicting
interplanetary shocks using neural networks. Adv Space Res 2005;36:23237.
[9] Mungiole M, Keith Wilson D. Prediction of outdoor sound transmission loss
with an articial neural network. Appl Acoust 2006;67:32445.
[10] Chiang Y, Chang L, Chang F. Comparison of static-feedforward and dynamicfeedback neural networks for rainfallrunoff modeling.
J Hydrol
2004;290:297311.
[11] Makarynskyy O. Improving wave predictions with articial neural networks.
Ocean Eng 2004;31:70924.
[12] Haykin S. Neural networks: a comprehensive foundation. New Jersey: PrenticeHall; 1998.
[13] Ashhab M, Al-Salaymeh A. Optimization of hot-wire thermal ow sensor based
on a neural net model. Appl Therm Eng 2006;26:94855.
[14] Ashhab M. A combination of neural net modeling and constrained
optimization towards inverse control. In: The 4th IASTED international
conference on modelling, simulation and optimization, Hawaii, USA; 2004. p.
6671.