Predictive Models For Photovoltaic Electricity Production in Hot Weather Conditions

energies
Article
Predictive Models for Photovoltaic Electricity
Production in Hot Weather Conditions
Jabar H. Yousif 1 , Hussein A. Kazem 2 and John Boland 3, *
1 Computing & Information Technology, Sohar University, P.O. Box 44, Sohar 311, Oman;
jyousif@soharuni.edu.om
2 Faculty of Engineering, Sohar University, P.O. Box 44, Sohar 311, Oman; h.kazem@soharuni.edu.om
3 Centre for Industrial and Applied Mathematics, University of South Australia, Adelaide 5095, Australia
* Correspondence: john.boland@unisa.edu.au; Tel.: +61-8-830-23449; Fax: +61-8-830-25785
Academic Editor: Francesco Calise

Received: 23 May 2017; Accepted: 6 July 2017; Published: 11 July 2017
Abstract: The process of finding a correct forecast equation for photovoltaic electricity production
from renewable sources is an important matter, since knowing the factors affecting the increase in
the proportion of renewable energy production and reducing the cost of the product has economic
and scientific benefits. This paper proposes a mathematical model for forecasting energy production
in photovoltaic (PV) panels based on a self-organizing feature map (SOFM) model. The proposed
model is compared with other models, including the multi-layer perceptron (MLP) and support
vector machine (SVM) models. Moreover, a mathematical model based on a polynomial function for
fitting the desired output is proposed. Different practical measurement methods are used to validate
the findings of the proposed neural and mathematical models such as mean square error (MSE),
mean absolute error (MAE), correlation (R), and coefficient of determination (R2 ). The proposed
SOFM model achieved a final MSE of 0.0007 in the training phase and 0.0005 in the cross-validation
phase. In contrast, the SVM model resulted in a small MSE value equal to 0.0058, while the MLP
model achieved a final MSE of 0.026 with a correlation coefficient of 0.9989, which indicates a strong
relationship between input and output variables. The proposed SOFM model closely fits the desired
results based on the R2 value, which is equal to 0.9555. Finally, the comparison results of MAE for the
three models show that the SOFM model achieved a best result of 0.36156, whereas the SVM and
MLP models yielded 4.53761 and 3.63927, respectively. A small MAE value indicates that the output
of the SOFM model closely fits the actual results and predicts the desired output.
Keywords: solar electricity prediction; artificial neural networks; photovoltaic; machine learning;
self-organizing feature map (SOFM)
1. Introduction
The primary source of energy used to generate electricity is fossil fuel, which is a non-renewable
energy source (RES) since it will run out in the future. Therefore, it is important to explore renewable
energy sources such as solar, geothermal, wind, etc., which are geographically site-specific. Solar energy
is one of the best options for Gulf Cooperation Council (GCC) countries. Solar energy presents as
both thermal and light components. The latter can be converted into electricity using solar cells or
photovoltaic (PV) technology, which is proven to work, and has been used for a long time. An important
reason for utilizing PV technology is to reduce its price, which has indeed occurred, especially in the
last decade, as depicted in Figure 1. Global production has increased by a factor of 370 since 1992 [1]
as shown in Figure 2.
Energies 2017, 10, 971; doi:10.3390/en10070971 www.mdpi.com/journal/energies

Energies 2017, 10, 971 2 of 19
Energies 2017, 10, 971 2 of 19
Energies 2017, 10, 971 2 of 19
100
100
(USD/W)
(USD/W)
10
Price
Price
Module
Module 10
1
1 0 0.01 0.1 1 10 100 1000 10000
PVPV
0 0.01 0.1 1 10 100 1000 10000
0.1
0.1 Cumulative Manufactured Capacity (GW)
Cumulative Manufactured Capacity (GW)
Figure 1. Projected
Projected economies of scale
scaleofofphotovoltaic (PV).
Figure 1. 1.
Figure Projectedeconomies
economies of
of scale of photovoltaic
photovoltaic (PV).
(PV).
100000
100000 PV power (MW) productions globally 1992-2015
PV power (MW) productions globally 1992-2015
10000
10000
(MW)
(MW)
1000
1000
power
power
100
100
PVPV
10
10
Year
Year
1 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
1 199260.1
199369.6
199477.6
199588.6
1996 126
1997 155
1998 201
1999 277
2000 386
2001 547
2002 748
20031140
20041782
20052474
20063733
20077056
20081066
20092140
20103000
20113000
20124000
20134100
20145400
2015
PV power 57.9
PV power 57.9 60.1 69.6 77.6 88.6 126 155 201 277 386 547 748 1140 1782 2474 3733 7056 1066 2140 3000 3000 4000 4100 5400
Figure 2. Total PV power (MW) productions globally 1992–2015.
Electricity demand in Oman is expected to increase due to thriving industrial areas such as
Electricity demand in Oman is expected to increase due to thriving industrial areas such as
Sohar, Alrusail, and Alduqum. Furthermore, the population of the Sultanate of Oman has grown
Electricity
Sohar, Alrusail,demand and in Oman isFurthermore,
Alduqum. expected to increase due toofthriving
the population industrial
the Sultanate of Omanareashas such as Sohar,
grown
dramatically. This will increase the demand for electricity, as illustrated in Figure 3. In 2016, the
dramatically.
Alrusail, and Alduqum.This will increase thethe
Furthermore, demand for electricity,
population of the as illustrated
Sultanate of in Figure
Oman has 3. In 2016,
grown the
dramatically.
maximum power demand reached 6000 MW, whereas the forecasted maximum power demand for
maximum
This 2018
will increase power the demand
demand reached
for 6000 MW,
electricity, as whereas
illustratedthe forecasted
in Figure maximum
3. In 2016, power
the demand for
maximum power
is expected to reach 6800 MW [2]. Thus, more electrical power plants are needed. Energy sources
demand 2018 is expected to reach 6800 MW [2]. Thus, more electrical power plants are needed. Energy sources
mustreached 6000 since
be diversified MW,approximately
whereas the forecasted
98% of the fuel maximum power electricity
used to generate demand for 2018 is
in Oman is natural
expected to
must be diversified since approximately 98% of the fuel used to generate electricity in Oman is natural
reachgas,
6800 MW2%
while [2].is Thus, moreMany
diesel fuel. electrical
studies power
have plants areevaluated,
explored, needed. Energy sources must
and investigated be diversified
the feasibility
gas, while 2% is diesel fuel. Many studies have explored, evaluated, and investigated the feasibility
sinceofapproximately
renewable energy 98% systems
of theand fuelconcluded that the priority
used to generate for renewable
electricity in Oman energy
is naturalgeneration
gas, whilein 2%
of renewable energy systems and concluded that the priority for renewable energy generation in
Oman
is diesel is
fuel. solar
Many energy. Oman
studies have has 342 long
explored, sunny days
evaluated, (between 10
and investigatedand 12 h) and large areas of free
Oman is solar energy. Oman has 342 long sunny days (between 10 and 12 the feasibility
h) and large areas of of
renewable
free
energylandsystems
in whichand PV concluded
systems canthat be installed
the [3,4]. for
priority Since most of the
renewable industrial
energy areas andinpopulation
generation Oman is solar
land in which PV systems can be installed [3,4]. Since most of the industrial areas and population
(55%) are in the northern part of Oman and solar radiation has a high concentration, approaching 955
energy. Oman
(55%) are inhasthe342 long part
northern sunny days and
of Oman (between 10 and has
solar radiation 12 h) andconcentration,
a high large areas ofapproaching
free land in955 which
W/m2 in summer, this region has become a target for PV investment.
PV systems
W/m2 in can summer, be installed
this region[3,4]. Since most
has become a target of for
thePV industrial
investment. areas and population (55%) are in
PV output is affected by many factors such as solar radiation, temperature, humidity, dust, rain,
the northern PV outputofisOman affected by many
solar factors suchhas as solar radiation, temperature, humidity, dust, rain, 2 in
and so on.part Solar radiation and and temperatureradiation
are the a high
main concentration,
factors, approaching
so the power output of the PV 955 W/m
system
summer,and so on. Solar
this region radiation
has becomeand temperature are the main factors, so the power output of the PV system
is a stochastic random process.aMoreover,
target forthe PVPV investment.
power fluctuation due to the solar radiation and
is a output
PV stochastic random process.
isfluctuations
affected by many Moreover,
factors the PV
such powerradiation,
as solar fluctuationtemperature,
due to the solar radiationdust,
humidity, and rain,
temperature also affects the PV system capital and operation costs. The uncertainty of
temperature fluctuations also affects the PV system capital and operation costs. The uncertainty of
and sotheon.
PVSolar radiation
system’s powerand temperature
output is the major aredrawback
the main factors,
of these so the power
systems. outputthis
Therefore, of the
issuePVhas system
the PV system’s power output is the major drawback of these systems. Therefore, this issue has
encouragedrandom
is a stochastic researchers to propose
process. modelsthe
Moreover, to forecast
PV power the PV output. These
fluctuation due models will be
to the solar used to and
radiation
encouraged researchers to propose models to forecast the PV output. These models will be used to
forecast the
temperature long-term production
fluctuations affectsof thePVPV systems,capital
which is useful in termscosts.
of theThetechnical and
forecast the long-term alsoproduction ofthe
the PVsystem
systems, which and operation
is useful in terms of the uncertainty
technical and of
economic points of view.
the PV system’s
economic power
points output is the major drawback of these systems. Therefore, this issue has
of view.
encouraged researchers to propose models to forecast the PV output. These models will be used to
forecast the long-term production of the PV systems, which is useful in terms of the technical and
economic points of view.
Energies 2017, 10, 971 3 of 19
Energies 2017, 10, 971 3 of 19
Figure
Figure 3. Oman
3. Oman peak
peak power
power demandfor
demand for 2011–2015
2011–2015 and
andprojection
projectionuntil 2018.
until 2018.
Artificial neural networks (ANN) have become increasingly useful for system modelling and the
Artificial
optimizationneural networks
of results. ANN (ANN) have become
computation increasingly
has increased usefultofor
our capacity system
analyze modelling
and and the
process data.
optimization
ANN is aofpowerful
results.modelling
ANN computation has increased
tool, which maps a complexour capacity
input to analyze
space into and process
simple output space. Indata.
ANN general, ANN applications
is a powerful tend towards
modelling tool, which mapsthe recognition,
a complexclassification, prediction,
input space into simplegeneralization,
output space. In
andANN
general, association of data.
applications tendThetowards
networkthearchitecture
recognition, determines the topology
classification, prediction, of the connections and
generalization,
between neurons (input, hidden, and output) [5]. There are various
association of data. The network architecture determines the topology of the connections topologies of ANN, such as between
feed
forward neural networks and recurrent neural networks. In addition, ANN can be classified, based
neurons (input, hidden, and output) [5]. There are various topologies of ANN, such as feed forward
on learning techniques, into two categories: namely, supervised and unsupervised [6]. In this paper,
neural networks and recurrent neural networks. In addition, ANN can be classified, based on learning
ANN is used to predict the productivity of a photovoltaic system installed in Oman. Data were
techniques,
measured intoand
two categories:
recorded for onenamely,
completesupervised
year in thisand unsupervised
study. The proposed [6]. In this paper,
self-organizing ANN is
feature
used to predict the productivity of a photovoltaic system installed in Oman.
map (SOFM) model was tested and verified in terms of speed and accuracy. Accuracy is defined as Data were measured
and recorded
the degree for
of one complete
a measured year
value thatinconforms
this study.
to a The
knownproposed
value or self-organizing feature
standard. In addition, map (SOFM)
the proposed
modelmodel
was tested and verified
is discussed, analyzed,inandterms of speed
compared withandotheraccuracy.
models likeAccuracy is defined
the multi-layer as the(MLP)
perceptron degree of
and support vector machine (SVM) models. Finally, the proposed model
a measured value that conforms to a known value or standard. In addition, the proposed model is compared with models in is
the literature for further verification.
discussed, analyzed, and compared with other models like the multi-layer perceptron (MLP) and
support vector machine (SVM) models. Finally, the proposed model is compared with models in the
2. Related Artificial Neural Networks-Based Work
literature for further verification.
Ogliari et al. [7] presented a hybrid intelligence computational evolutionary method for
2. Related Artificial
forecasting and Neural Networks-Based
analyzing Work
the predictions of photovoltaic systems based on comparing different
forecasting error techniques. The absolute hourly error, and daily absolute error, are the factors used
Ogliari et al. [7] presented a hybrid intelligence computational evolutionary method for forecasting
to validate and analyze the errors of the forecasting models. They implement a case study for
and analyzing
calculatingthe predictions
error values duringof photovoltaic systems
a clear sky day based(theoretical,
of radiation on comparing different
forecast, forecasting error
and actual).
techniques.Hernández et al. [8], developed an unsupervised 4 × 4 neurons self-organizing mapto
The absolute hourly error, and daily absolute error, are the factors used validate
(SOM) and and
analyze
the clustering k-means data processing system for analyzing the energy consumption patterns values
the errors of the forecasting models. They implement a case study for calculating error in
duringindustrial parks
a clear sky in of
day Spain. The proposed
radiation systemforecast,
(theoretical, is validated
andbased on real data and finding different
actual).
behavior patterns
Hernández thatdeveloped
et al. [8], are meaningful and could work4without
an unsupervised any prior
× 4 neurons knowledge about
self-organizing mapthe data. and
(SOM)
Bracale et al. [9] proposed a short term probabilistic forecast model for
the clustering k-means data processing system for analyzing the energy consumption patterns in predicting the hourly
power of a photovoltaic system based on the probability density function. The model implements a
industrial parks in Spain. The proposed system is validated based on real data and finding different
Bayesian autoregressive time series for solar radiation based on critical variables like cloud cover and
behavior patterns that are meaningful and could work without any prior knowledge about the data.
humidity. The proposed model is validated using the Monte Carlo simulation procedure based on a
Bracale
random et al. [9] proposed
sampling a short
of the clearness indexterm probabilistic forecast model for predicting the hourly
distribution.
power of aInphotovoltaic system based on the probability
wind energy systems, Sheela and Deepa [10] had used density function.
a hybrid neuralThe model
network implements a
model-based
Bayesian
SOMautoregressive
and radial basistime series(RBF)
functions for solar radiation
to predict wind based
speed. on critical
Their variables
experiments andlike cloud model
proposed cover and
humidity. The
results areproposed
close to model is validated
the actual results using the Monte
with small Carlo error.
percentage simulation procedure
Several researchersbasedare on a
randomimplementing
sampling of supervised feed forward
the clearness ANN techniques to forecast solar energy production [11–13].
index distribution.
In wind energy systems, Sheela and Deepa [10] had used a hybrid neural network model-based
SOM and radial basis functions (RBF) to predict wind speed. Their experiments and proposed model
results are close to the actual results with small percentage error. Several researchers are implementing
supervised feed forward ANN techniques to forecast solar energy production [11–13].
Energies 2017, 10, 971 4 of 19
Argiriou et al. [11] implemented a neural controller model based on feed forward back propagation
for hydronic heating plants in buildings. The controller was used to forecast variables such as
meteorological modules, ambient temperature, and solar irradiance. The proposed models are used on
real-scale office buildings during real operating conditions. A comparison of operational results and
the conventional controller is performed to evaluate the performance of the proposed numerical-based
simulation models. The results of the experiments and numerical models showed that the percentage
of energy saving is about 15% in North European weather conditions.
Khatib et al. [12] proposed a solar irradiation, system-based ANN, which uses data from 28 cities
in Malaysia. The proposed neural network model is used to forecast the clearness index, which is used
to establish a predictive global solar irradiation system for Malaysia. The predicted solar irradiation
system has a mean absolute percentage error (MAPE) of 5.92%, and yielded 7.96% for the root mean
squared error (RMSE) and 1.46% for the mean absolute error (MAE).
Dorvlo et al. [13] implemented an ANN technique to estimate the solar radiation-based clearness
index. They used long-term data to design a hybrid neural, net system-based RBF and MLP for Oman.
They claim that both the RBF and MLP models perform well. However, they suggested that RBF
models are more efficient since they require less computing power. The RBF is trained with data
obtained from different meteorological stations located in Sohar, Seeb, Masirah, Sur, Salalah, Sohar,
and Fahud.
A number of works have involved implementing the feedback neural network, such as the
recurrent neural network, and unsupervised learning or competition learning to predict renewable
energy systems. Thus, ANNs have been widely implemented in renewable energy power systems.
Moreover, several neural network techniques have been used to design and implement different phases
of renewable energy power systems based on problem requirements and their characteristics [14–37].
Kalogirou and Bojic [14] implemented an artificial neural network model for predicting the energy
consumption of a passive solar building of one room and a roof. A multilayer with a recurrent,
architecture-based, back-propagation learning algorithm was implemented. The dynamic thermal
building model was performed and tested for two seasons: winter and summer. The thickness of
walls in simulated buildings varied from 15 cm to 60 cm. The proposed model was much faster than
the dynamic simulation programs and achieved a good fit for actual data based on the value of the
coefficient of determination (R2 value), equal to 0.9985. Zhang and Chen [15] presented a predictive
trend analysis model for energy consumption based on RBF network learning. The compression
study with other feed-forward neural networks proves that the RBF model is faster and has a strong
approximation function. The particle swarm optimization (PSO) algorithm is used for optimizing the
performance of predicting energy consumption. Zhou et al. [16] proposed an artificial neural network
model for estimating solar irradiation in Chinet, Beijing. The results of the proposed model showed
that it can be used to evaluate the solar potential. Mohandes et al. [17] introduced the multilayer
perceptron neural network for estimating global solar radiation in Saudi Arabia. The data was collected
using 41 stations spread all over the country since 1971. These data were divided into training data sets
(from 31 stations) and the other data (from 10 stations) were used as unseen data test set. The result
of the proposed model proved the viability of the neural approach for modeling the solar radiation.
The neural model obtained a MAPE equal to 4.49. Also, Adnan et al. [18] implemented a multilayer
perceptron neural network for Turkey and it accurately predicted the actual data based on the value of
coefficient of determination (R2 ), which is equal to 0.958, with a MAPE equal to 6.7. Sozen et al. [19]
proposed the use of the ANN for evaluating the solar energy potential in Turkey. They used different
supervised learning algorithms like a scaled conjugate gradient (SCG), Levenberg-Marquardt (LM),
and Polak-Ribiere conjugate gradient (CGP). The data for this paper is collected from 17 stations
(namely cities) spread over Turkey for a period of 3 years (2000–2002). This model achieved a MAPE
equal to 2.41, and the coefficient of determination (R2 ) value was 99.99. Elminir [20] developed an
ANN model based on the Levenberg optimization function for predicting the insolation data for
different spectral bands of the Helwan monitoring station in Egypt. The proposed model obtained
Energies 2017, 10, 971 5 of 19
accuracies of 95%, 93%, and 96% for infrared, ultraviolet, and global insolation, respectively. Also, the
model was tested based on infrared, ultraviolet, and global insolation using data for about one year for
Aswan station, and it obtained accuracies of 95%, 91%, and 92%, respectively. Also, Elminir et al. [21]
implemented a multilayer Perceptron Neural Network Neural Network model to predict diffuse
fraction (KD) on an hourly and daily scale in Egypt. Hence, an effort is underway to implement the
ANN model regarding a first order polynomial relating KD with the clearness index (KT) and sunshine
fraction (S/S0). The ANN model showed that the model is suitable for predicting the diffuse fraction
in hourly and daily scales, in comparison with the regression models. Rehman and Mohandes [22]
utilized three feedforward neural network models for estimating global solar radiation (GSR) in the
city of Abha in Saudi Arabia. The data were collected for 1462 days during 1998–2002, including
factors like air temperature and relative humidity. The proposed model was tested using different
combinations of input/output. The first model achieved a mean squared error (MSE) equal to 0.00028
and a MAPE equal to 10.3. The second model obtained MSE equal to 0.0052 and a MAPE equal to
11.8%. The last model obtained MSE equal to 0.00003 and a MAPE equal to 4.9. The testing of the
proposed model proved that it was capable of estimating global solar radiation from temperature
and relative humidity accurately. The study in [23] deploys a multilayer feed forward (MLFF) neural
network model for evaluating the monthly average daily global solar irradiation on a horizontal
surface in Uganda. The input factors of the neural model are sunshine interval, temperature, cloud
cover, and location. The proposed model fits the actual data accurately, the correlation coefficient (R) is
0.974, the value of the mean bias error (MBE) is 0.059 MJ/m2 , and the value of RMSE is 0.385 MJ/m2 .
The predictive results are compared and tested with other works in the literature. Mehmet et al. [24]
applied a multilayered neural network model based on a resilient backpropagation learning algorithm
for predicting the mean monthly wind speed of any target station in Turkey. The data was collected
for a period between 1992 and 2001 by the Turkish State Meteorological Service (TSMS). The results
demonstrated the proposed neural model could forecast the actual data accurately. It achieved a
maximum value of mean absolute percentage error equal to 14.13% for the Antakya meteorological
station and it got a best value in the Mersin meteorological station equal to 4.49%. Also, the model
achieved a maximum value of the correlation coefficient (R) equal to 0.97 and a minimum value equal
to 0.67. Bosch et al. [25] developed an ANN model to determine the daily global irradiation at stations
located in a complex terrain. The collected data sets consisted of 3 years’ data of daily global radiation
from 12 different stations located in the north face of the Sierra Nevada National Park in the South
East of Spain. The output of the neural model was tested and compared with the actual data, which
proved that the neural model was very accurate. It showed that the proposed model was an efficient
and easy methodology for calculating solar radiation levels, with great results like the RMSE of 6.0%
and a MBE of 0.2%.
Fadare [26] developed a supervised MLFF neural network model based on the back-propagation
learning method for forecasting the production of solar energy potential in Nigeria. The data sets
used in this work were collected from 195 cities in Nigeria from the NASA geo-satellite database
over a period of 10 years (1983–1993). The input consisted of 7 variables and their locations (latitude,
longitude, and altitude), month, average of sunshine duration, temperature, and relative humidity.
The solar radiation intensity is the output variable of the model. The monthly mean of solar radiation
potential in northern and southern regions ranges from 7.01 to 5.62 and 5.43 to 3.54 kWh/m2 day.
The proposed model implemented different learning algorithms like the SCG and LM. Also, three
different numbers of hidden layers (5, 10, and 15) were used, and the models achieved correlation
coefficients (R) equal to 0.978, 0.971, and 0.956, correspondingly. Jiang [27] presented an artificial neural
network model for estimating the solar irradiation in China. The collected data includes locations over
all China for the period 1995–2004. The proposed model is tested and compared with measured values
using mean percentage error (MPE), MBE, and RMSE. It achieved a value of 1.55, −0.04, and 0.746,
respectively, with an accuracy of 94.81%.
Energies 2017, 10, 971 6 of 19
Energies 2017, 10, 971 6 of 19

3. Methodology
3. Methodology
In this2017,
work, 24 PV modules were installed at Sohar University in Oman. The latitude
Energies 10, 971 6 of 19 and
longitudeInofthis
Sohar-Oman,
work, 24 PVthe secondwere
modules largeinstalled
city, is 24 20 N, 56
at Sohar 40 E. The
University in Sohar
Oman.zone has a very
The latitude andgood
3. Methodology
longitude of Sohar-Oman, the second large city, is 24 20 N, 56 40 E. The Sohar zone
solar energy potential and, therefore, any PV system investment in this zone is expected to be very has a very good
solarElectrical
feasible. energy potential and, therefore,
parameters such as any PV system
voltage, investment
current, and in thiswere
power zone recorded
is expectedandto bemonitored
very
In this work, 24 PV modules were installed at Sohar University in Oman. The latitude and
feasible.
for a longitude Electrical
period ofofabout parameters
one year. such as voltage, current, and power were recorded and monitored for
Sohar-Oman, the The PVlarge
second module rating
city, is 24 20 is
N,140 W,E.13.9%
56 40 efficiency,
The Sohar zone has7.91 A maximum
a very good
a period of about one year. The PV module rating is 140 W, 13.9% efficiency, 7.91 A maximum current,
solar
current, energy
17.7 potential voltage,
V maximum and, therefore,
8.68 Aany PVcircuit
short systemcurrent,
investmentandin22.1
thisVzone
openis circuit
expected to be very
voltage. Figure 4
17.7 V maximum voltage, 8.68 A short circuit current, and 22.1 V open circuit voltage. Figure 4
illustrates the configurations of the PV systems. Electrical parameters of the grid connected PVfor
feasible. Electrical parameters such as voltage, current, and power were recorded and monitored system
illustrates the configurations of the PV systems. Electrical parameters of the grid connected PV
a period
(power, of about
current, and one year. The
voltage) werePVrecorded
module rating
and is 140 W, 13.9% efficiency, 7.91 A maximum current,
monitored.
system (power, current, and voltage) were recorded and monitored.
17.7 V maximum voltage, 8.68 A short circuit current, and 22.1 V open circuit voltage. Figure 4
illustrates the configurations of the PV systems. Electrical parameters of the grid connected PV
system (power, current, and voltage) were recorded and monitored.
Figure 4. The PV system.

Global solar radiation was measured between July 2013 and August 2014 using a BF5 Sunshine
Global (Version
Sensor solar radiation was measured
1, Cambridge, England). between
The sensors July 2013 and
generate the August
data under2014 using aconditions,
different BF5 Sunshine
and some
Sensor (Version of these data are out of range or cannot be used directly. Therefore, proper pre-processing
Global solar radiation was measured between July 2013 and August 2014 using a BF5 Sunshine and
1, Cambridge, England). The sensors generate the data under different conditions,
actions
someSensor
of theseand
datadata
are cleaning mustorbecannot
done be
beforeusedusing theseTherefore,
data. Theproper
calibration accuracy and
(Version 1, out of range
Cambridge, England). The directly.
sensors generate pre-processing
the data under different actions
conditions,
sensitivity were ±0.12% and 1 mV/0.5 W m , respectively, and the range of irradiance was 0–1250 W
−2
and data cleaning
and some of thesemust
databe
aredone
out ofbefore
range using
or cannot these data.directly.
be used The calibration accuracy
Therefore, proper and sensitivity
pre-processing
m−2. The operating temperature−2was −20–+70 °C and an RHT2 temperature sensor (Version 1, −2
were actions
±0.12%and and data cleaning W
1 mV/0.5 must
m be, respectively,
done before usingand thethese data.ofThe
range calibration
irradiance accuracy
was 0–1250andWm .
Cambridge, England) was used. The data was downloaded using a DL2e Data Logger and specific
sensitivity temperature
The operating were ±0.12% andwas 1−mV/0.5
20–+70 W◦m C−2 , respectively,
and an RHT2 and the range of
temperature irradiance
sensor was 0–1250
(Version 1, W
Cambridge,
software
−2. The (Version 5, Cambridge, England). The average temperature in most Gulf countries is
m
England) operating
was used. temperature
The data was 5.was −20–+70
downloaded °C and an RHT2 temperature sensor (Version 1,
symmetric, as illustrated in Figure Therefore, weusingcan saya the
DL2e Data
results Logger
of this workand specific
can be suitablesoftware
for
Cambridge, England) was used. The data was downloaded using a DL2e Data Logger and specific
(Version 5, Cambridge,
implementation England).
in the whole Gulf The average
region, with temperature in most Gulf countries is symmetric, as
adjustable conditions.
software (Version 5, Cambridge, England). The average temperature in most Gulf countries is
illustrated in Figure 5. Therefore, we can say the results of this work can be suitable for implementation
symmetric, as 40
illustrated in Figure 5. Therefore, we can say the results of this work can be suitable for
in theimplementation
whole Gulf region, in the with
wholeadjustable
Gulf region,conditions.
with adjustable conditions.
37
34
40
31
Temperature
37
28
34
25
31
Temperature
22
28
19
25
16
22
13
19
10
16
Jan Fab Mar Apr May Jun Jul Aug Sep Oct Nov Dec
13 Month
10
OMAN UAE KSA QATAR BAHRAIN KUWAIT
Jan Fab Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Month
Figure 5. Average monthly temperature for the Gulf Cooperation Council (GCC) from 1901 to 2015
OMAN UAE KSA QATAR BAHRAIN KUWAIT
(source: worldbank.org).
Energies 2017, 10, 971 7 of 19
4. Self-Organizing Feature Map Architecture and Design

SOFM was used to transfer multidimensional data to lower dimensional spaces. The differences
between the input vector and the vectors of other neurons (Dij ) are calculated as in Equation (1):
q
Dij = X l − Wij = ( x1 − wij1 )2 + ....... + ( xn − wijn )2 (1)

The best matching unit (BMU) is the winning node whose weight vector is the most similar to the
input vector.
D(k1, k2) = mini,j (Di,j ), where k1 and k2 are the indexes of the winner-neuron. The weights of
the winner node and its neighbor neurons are then adjusted. The neighborhood function is used to
determine the neighborhood of a neuron as follows:
h(ρ, t) = exp(ρ2 /2σ2 (t))
where ρ is the distance to the winner-neuron, which is computed as in Equation (2):

q
ρ= ( k 1 − i )2 + ( k 2 − j )2 (2)
hat h(ρ,t) or the French hat h(ρ) as in Equation (3):

2
h(ρ, t) = exp(−ρ /σ (t)) 1 − 2 ρ2
2 2
(3)
σ (t)
Thus weights of all the neurons are updated as in Equation (4):
Wij (t + 1) = Wij (t) + α(t)h((ρ, t)( x l (t) − Wij (t)) (4)
where α(t) is the learning rate. Finally, the weight vector of the winner-neuron or its adjacent is updated.
The back-propagation learning algorithm is usually implemented in the MLP model to propagate the
errors through the network. Error correction learning ei (n) is defined as in Equation (5):
ei ( n ) = d i ( n ) − y i ( n ) (5)
Gradient descent learning is performed to adapt each weight in the network, as in Equation (6):
wij (n + 1) = wij (n) + ηδi (n) + x j (n) (6)
where the local error δi (n) is computed from ei (n) at the output processing element (PE). The constant
step size is η.
The SOFM is designed and implemented using the Neuro-Solutions software package. Figure 6
illustrates the design of SOFM, which has two PEs as input layers (solar radiation and ambient
temperature), one output layer (the PV current), and one hidden layer. The data for this paper are
generated from the twenty-four PV modules, which are shown in Figure 4 for a period of about one
year (July 2013 until August 2014). The panel generates the data every hour. Therefore, we will use
the average of each day as one data set. This work uses about 245 data sets (days), which splits into
three categorized (40% as training data sets, 40% for the cross-validation data set, and 20% for testing
data sets). The size and dimension of the unsupervised output space is 5 × 5. The neighborhood
shape is significant in defining the organizational rules within the neural field. The final shape of
the feature map greatly depends on the size of the initial and final radii of the neighborhoods. The
final radius should have a small value, normally one or two PEs wide. Therefore, the initial radius
is determined to be equal to 2 and the final radius is determined to be equal to 0. The learning rule
and non-linearity selected for the hidden layer is momentum function with a step size of 1 and a
Energies 2017, 10, 971 8 of 19
momentum rate of 0.7. The TanhAxon is used as a transfer function in the four nearest neighbor PEs
and output layers. The unsupervised learning rate is started with 0.01 and decays to 0.001. In order
Energies 2017, 10, 971 8 of 19
to enhance the behavior of the proposed SOFM model, several epochs (1000, 2000, 5000, and 10,000)
are implemented.
of the proposedThe training
SOFM process
model, severalisepochs
terminated when5000,
(1000, 2000, all the
andweights
10,000) are less than or equal
are implemented. The to a
specific value.
training process is terminated when all the weights are less than or equal to a specific value.
Figure 6. The self-organizing feature map (SOFM) architecture.

Figure 6. The self-organizing feature map (SOFM) architecture.

The NeuroSolutions package was applied for various methods to measure the performance of
The NeuroSolutions
neural package
models and estimate the was
errorsapplied for various
in the model outputmethods to measureerror
data. A well-known the performance
measuring of
neural models
method and
is the estimate
MSE, the
which is theerrors
result in the model
of dividing the output data.
summation A well-known
of the squared error error measuring
(SSE) by the
population
method n, as defined
is the MSE, which in
is Equations
the result(7)
of and (8): the summation of the squared error (SSE) by the
dividing
population n, as defined in Equations (7) and (8): 1 (7)
=
n
1
MSE = SSE. (7)
SSE =  ( xni − x ) 2 (8)
i =1
n
Here n is the number of observations, SSE i =is ∑
x 2
xi − xof) the ith observation, and x is the mean (8)
the(value
i =1
value of all the observations.
Here n is Moreover,
the number theof observations,
correlation xi is (r)
coefficient theisvalue
used toofdetermine
the ith observation, and x is the
the linear dependence meantwo
between value of
all the
or observations.
more data sets. The values of r in the range of (−1, 1). The correlation is computed as in Equation
(9):
Moreover, the correlation coefficient (r) is used to determine the linear dependence between two
∑
or more data sets. The values of r in the range .of (−−1, 1).∑ The correlation
∑ is computed as in Equation (9):
= (9)
∑ − ∑ ∑ − ∑
n (∑in=1 x.y) − (∑in=1 x )(∑in=1 y)
r =
where x is the first dataseth {x1,...,xn} containing n values
r i hand y is the other dataset {y . n} containing
i 1,...,y (9)
n 2 n 2 n 2 n 2
n value. 2
n ( ∑ i =1 x ) − ( ∑ i =1 x ) n ( ∑ i =1 y ) − ( ∑ i =1 y )
5.1. Epoch Results

where x is the first dataset {x1 , ..., xn } containing n values and y is the other dataset {y1 , ..., yn } containing
n value. The experiments undertaken are implemented for different numbers of epochs (1000, 2000, 5000,
and 10,000) in the training and testing of the SOFM model for the purpose of enhancing the
performance
5.1. Epoch Resultsof the proposed models. Figure 7 demonstrates the results of MSE for the SOFM-1000
model using 1000 epochs achieved 0.0005 in the training phase and 0.0006 in the cross-validation
The experiments
phase. The comparisonundertaken are implemented
of the proposed SOFM model and for the
different
desirednumbers of epochsin(1000,
output is illustrated Figure 2000,
5000,8.and
It clearly indicates that the SOFM model predicts the actual results well. For the sake of testing the the
10,000) in the training and testing of the SOFM model for the purpose of enhancing
performance
stability ofofthe
thenetwork,
proposed models.
several Figure
different 7 demonstrates
epochs the results
were implemented. of MSE
The graph forfinal
for the the MSE
SOFM-1000
of
modeltheusing 1000
proposed epochs
SOFM modelachieved 0.0005
using 2000 in achieved
epochs the training phase
0.0001 in theand 0.0006
training in and
phase the 0.0001
cross-validation
in the
cross-validation
phase. The comparison phase, as illustrated
of the proposedinSOFM
Figuremodel
9. On the
andother hand, the
the desired desired
output is output and actual
illustrated in Figure 8.
Energies 2017, 10, 971 9 of 19
It clearly indicates that the SOFM model predicts the actual results well. For the sake of testing the
stability of the network, several different epochs were implemented. The graph for the final MSE of
the proposed SOFM model using 2000 epochs achieved 0.0001 in the training phase and 0.0001 in the
cross-validation phase, as illustrated in Figure 9. On the other hand, the desired output and actual
network output using 2000 epochs are presented in Figures 10–12 illustrated the SOFM-5000, and
Energies 2017, 10, 971 9 of 19
SOFM-10000 models, which achieved MSEs of 0.00006, and 0.00003, respectively, in the training phase.
Energies 2017, 10, 971 9 of 19
Tablenetwork
1 showsoutput
the results of minimum and final MSE for the SOFM network. Moreover,
using 2000 epochs are presented in Figures 10, 11 and 12 illustrated the SOFM-5000, these models
achieved
and MSEs
SOFM-10000 of
network output using0.0001,
models,in which
2000 the cross
epochs validation
achieved
are MSEs of
presented phase, as 10,
in 0.00006,
Figures depicted
and11 and in
0.00003, the Table 2.
12respectively,
illustrated The
in
the the experiments
training
SOFM-5000,
proveandthat
phase. increasing
Table 1 the
shows number
the results of
of epochs
minimum in the
and training
final MSE and
for learning
the SOFM phases
network.
SOFM-10000 models, which achieved MSEs of 0.00006, and 0.00003, respectively, in the training leads to
Moreover, a reduction
these
in thephase.
valueTable
models of MSE.
achieved In addition,
MSEs of the
0.0001, experiments
in the cross prove
validationthat there
phase, is
as a strong
depicted
1 shows the results of minimum and final MSE for the SOFM network. Moreover, theserelation
in the between
Table 2. Theinput
experiments
and output
models variables prove
achieved basedthat
MSEs ofincreasing
on0.0001, the number
the correlation of
factor
in the cross epochs in the training
of the SOFM-1000,
validation and learning
SOFM-2000,
phase, as depicted phases leads
SOFM-5000,
in the Table 2. Theto and
a reduction
experiments
SOFM-10000 in the which
prove
models, value of MSE.
that increasing
are In
theaddition,
0.9915, theepochs
number 0.9996,
0.9996, of experiments
andin 0.9997prove
the training that
andthere is a phases
learning
respectively. strong leads
relationto
abetween input
reduction andvalue
in the output
of variables based on the correlation
MSE. In addition, experimentsfactor
proveofthat
the SOFM-1000, SOFM-2000,
there is a strong relation
SOFM-5000,
between
Table and
input
1. Results ofSOFM-10000
and minimum models,
output variables
and final which
based
mean are
the0.9915,
onsquare error0.9996,
correlation 0.9996,
factor
(MSE) for of and
thethe 0.9997 respectively.
SOFM-1000,
self-organizing SOFM-2000,
feature map
SOFM-5000, and
(SOFM-1000) SOFM-10000 models, which are 0.9915, 0.9996, 0.9996, and 0.9997 respectively.
network.
Table 1. Results of minimum and final mean square error (MSE) for the self-organizing feature map
(SOFM-1000)
Table network.
1. Results of minimum and final mean square error (MSE) for the self-organizing feature map
Best Networks Training Cross Validation
(SOFM-1000) network.
Best Networks Training Cross Validation
Epoch # 999 999
BestEpoch
Minimum MSE #
Networks 999
Training
0.0007 999
Cross Validation
0.0005
Minimum
Final MSE #MSE
Epoch 0.0007
999
0.0005 0.0005
9990.0006
Final MSE
Minimum MSE 0.0005
0.0007 0.0006
0.0005
Final MSE 0.0005 0.0006
0.0001
0.0001
0.001
0.001
0.01
MSEMSE
0.01 Training MSE

Training MSE
Cross Validation MSE
0.1
0.1
100 100
199 199
298 298
397 397
496 496
595 595
694 694
793 793
892 892
991 991
1
1
1
1 Epoch number
Epoch number
Figure 7. The mean square error (MSE) for the proposed SOFM model using 1000 epochs.
8000
8000
7000
7000
6000
6000
5000
/Am/Am
5000
4000
Ipv Ipv
4000
3000
3000
2000
2000
1000
10000
0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
1 Ipv
3 /Am
5 Output
7 9 11 Ipv13/Am15 17 Day
19 number
21 23 25 27 29 31 33 35
Ipv /Am Output Ipv /Am Day number
Figure 8. The testing results of the desired output and the proposed SOFM model output using 1000
epochs.8. The testing results of the desired output and the proposed SOFM model output using 1000
Figure
Figure 8. The testing results of the desired output and the proposed SOFM model output using
epochs.
1000 epochs.
In order for the proposed SOFM model to be accurate and fit the actual results well, the
coefficient of determination
In order R2 is SOFM
for the proposed used, which
modelis to
defined as in Equation
be accurate and fit (10):
the actual results well, the
coefficient of determination R2 is used, which is defined
∑ −as in Equation (10):
=1− (10)
∑∑ −−
=1− (10)
∑ −
Energies 2017, 10, 971 10 of 19
In order for the proposed SOFM model to be accurate and fit the actual results well, the coefficient
of determination R2 is used, which is defined as in Equation (10):
( ∑ i ( y i − f i )2
R2 = 1 − . (10)
( ∑ i ( y i − f y i )2
Energies 2017, 10, 971 10 of 19
yi . is2017,
whereEnergies the 10,
observed
971 value of the actual output, fi is the predicted value, and y. i is the 10 arithmetic
of 19
where is the observed value of the actual output, f is the predicted value, and i
mean value of the observed targets. The preferred model is that which gets a coefficient determination
i is the arithmetic
where value
mean is the observed value of the actual
Theoutput, fi is the predicted value, and gets
i is the arithmetic
R2 value closer to of the prediction
1. The observed targets.
trend-line preferred
model model
for testing is SOFM-2000
the that which model a is
coefficient
calculated as
mean value of
determination the observed
R2 value targets.
closer to 1. The preferred
The prediction trend-line model
modelisfor
that which
testing gets a coefficient
the SOFM-2000 model
in Equation (11): 2
determination
is calculated asRinvalue closer
Equation to 1.
(11): 4
The prediction
3
trend-line2
model for testing the SOFM-2000 model
is calculated as in Equation (11): − 6.3087x + 150.59x − 1429.2x + 9474.5
y = 0.0884x (11)
y = 0.0884x4 − 6.3087x3 + 150.59x2 − 1429.2x + 9474.5 (11)
y = 0.0884x4 − 6.3087x3 + 150.59x2 − 1429.2x + 9474.5 (11)
0.0001
0.0001
0.001
0.001
0.01
MSE
0.01
MSE
0.1
Training MSE
0.1
Training MSE
1 200 399 598 797 996 1195 1394 1593 1792 1991
1 1 200 399 598 797 996 1195 1394 1593 1792 1991
1 Epoch number
Epoch number
Figure 9. The MSE for the proposed SOFM model using 2000 epochs.
9000
9000
8000
y = 0.0884x4 - 6.3087x3 + 150.59x2 - 1429.2x + 9474.5
8000
y = 0.0884x4 - 6.3087x3 + 150.59x2 - 1429.2x + 9474.5
7000
7000
6000
6000
Ipv/Am
5000
Ipv/Am
5000
4000
4000
3000
3000 Ipv /Am
2000
Ipv /Am
2000
1000
1000
0
0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Day number
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Day number
Figure 10. The testing results of the desired output and the forecasting SOFM output using 2000
Figure 10. The testing results of the desired output and the forecasting SOFM output using 2000
epochs.
Figure 10. The testing results of the desired output and the forecasting SOFM output using 2000 epochs.
epochs.
Figures 13 and 14 show the predicted trend-line models using a biquadratic polynomial function
Figures 134 for
Figures
of degree and 14 show
13SOFM-5000
and 14 showthe predicted
the
and predictedtrend-line
SOFM-10,000,trend-line modelsusing
models
respectively. usinga biquadratic
a biquadratic polynomial
polynomial function
function
of degree
of degree In4 for4 for SOFM-5000
SOFM-5000
general, the termand and SOFM-10,000,
SOFM-10,000,
accuracy respectively.
is defined respectively.
to mean the goodness of fit of measured results to the
In In general,
desired
general, the the
output. Theterm
term SOFMaccuracy isisdefined
models
accuracy defined
achieved toan
to mean the
thegoodness
accuracy
mean rate of of
goodness fit fit
99%.
of of measured
MLP
of results
and SVM
measured to theto the
obtained
results
desired
accuracy
desired output.
output.rates The
TheofSOFM SOFM
92% and models achieved
62%, respectively.
models an accuracy rate of 99%. MLP and SVM
achieved an accuracy rate of 99%. MLP and SVM obtained accuracy obtained
accuracy rates of 92% and 62%, respectively.
rates of 92% and 62%, respectively.
Energies 2017, 10, 971 11 of 19
Energies 2017, 10, 971 11 of 19
Energies 2017, 10, 971 11 of 19
Energies 2017,5E-05
10, 971 11 of 19
5E-05
5E-05
0.0005
0.0005
0.0005
MSE
0.005 Training MSE

MSE
0.005 Cross Validation

Training MSE MSE
MSE
0.005 Cross
Training MSEValidation MSE
0.05
0.05
0.05
1 500 999 1498 1997 2496 2995 3494 3993 4492 4991
1 500 999 1498 1997 2496 2995 3494 3993 4492 4991
1 500 999 1498 Epoch2496
1997 numbers2995 3494 3993 4492 4991
Epoch numbers
Figure 11. The testing of MSE for the proposed SOFM model using 5000 epochs.
Epoch
Figure 11. The testing of MSE for the numbers
proposed SOFM model using 5000 epochs.
0.00001
0.00001
0.00001
0.0001
0.0001
0.0001
0.001
0.001
MSE
0.001
MSE
0.01
MSE
0.01 Training MSE

0.01 Cross Validation…
Training MSE
Cross
Training MSEValidation…
0.1
Cross Validation…
0.1
0.1
1 1000 1999 2998 3997 4996 5995 6994 7993 8992 9991
1 1 1000 1999 2998 3997 4996 5995 6994 7993 8992 9991
Epoch numbers
1 1 1000 1999 2998 3997 4996 5995 6994 7993 8992 9991
1
Epoch numbers
Figure 12. The testing of the MSE for the proposed
Epoch SOFM model using 10,000 epochs.
numbers
Figure 12. The testing of the MSE for the proposed SOFM model using 10,000 epochs.
12. The testing of the MSE for the proposed SOFM model using 10,000 epochs.
Figure 9000
Figure 12. The testing of the MSE for the proposed SOFM model using 10,000 epochs.
9000
8000
9000 y = 0.089x4 - 6.3486x3 + 151.45x2 - 1435.3x + 9486.5
8000 y = 0.089x4 - 6.3486x3 + 151.45x2 - 1435.3x + 9486.5
7000
8000 y = 0.089x4 - 6.3486x3 + 151.45x2 - 1435.3x + 9486.5
7000
6000
7000
6000
5000
6000
5000
Day
4000
5000
Day
4000
Day
3000
4000 Ipv /Am
3000
2000 Ipv /Am
3000
2000 Ipv Ipv
/Am/Am Output
1000 Ipv /Am Output
2000
1000 Ipv /Am Output
1000 0
01 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
0 1 3 5 7 9 11 13Day
15 number
17 19 21 23 25 27 29 31 33 35
1 3 5 7 9 11 13 15 17Day19number
21 23 25 27 29 31 33 35
Figure 13. Testing the results of the desired data and the forecasting SOFM model using 5000 epochs.
Day number
Energies 2017, 10, 971 12 of 19
Energies 2017, 10, 971 12 of 19
9000
8000
Energies 2017, 10, 971 12 of 19
7000
9000
6000
Ipv/Am
8000
5000
7000
4000
6000
3000
Ipv/Am
Ipv /Am
5000
2000 Ipv /Am Output
4000 Poly. (Ipv /Am Output)
1000
3000
0 Ipv /Am
2000 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 Ipv /Am Output
Days number Poly. (Ipv /Am Output)
1000
Figure 14. Testing the results of the desired data and the forecasting SOFM model using 10,000 epochs.
Figure 14. Testing
0 the results of the desired data and the forecasting SOFM model using 10,000 epochs.
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
It is noted that the network results arenumber
Days non-linear functions, as shown in Figures 10, 13, and 14.
It is notedthe
Therefore, that the network
equation results
of the trend lineare
doesnon-linear
not perfectly functions, as shown
fit the results. For thein Figures
purpose of 10, 13 and 14.
ensuring
Therefore, Figure
that the 14. Testing
theproposed
equation thethe
models
of results of the
achieve
trend an
line desired
does data
excellent and
not fittingthe curve
perfectlyforecasting
fit for
thetheSOFM model
desired
results. using
Forresults, 10,000 epochs.
the regression
the purpose of ensuring
models
that the were used
proposed to create
models a newan
achieve fitting curve. The
excellent 2D-CurveTable
fitting curve for the software
desired wasresults,
used tothecreate the
regression
new It is noted
fitting model that
andtheitnetwork
was then results
comparedare non-linear
with the functions,
output of as shown
neural in Figures
models. First, 10,
the 13,
dataandwas14.
models were used
Therefore, the
to createofa the
equation
new fitting
trend line
curve.
does
The
not
2D-CurveTable
perfectly fit the
software
results. For the
was used
purpose of
to create the
ensuring
cleaned and then used to draw the predicted data. The spline and Fourier smoothing estimation was
new fitting
that the model
proposedand it was then compared with the output ofthe
neural models. First, the data was
performed in order models
to comeachieve
up withanexcellent
excellent fitting
smoothing curve for
of data anddesired results,
good fitting. Thetheexperiment
regression
cleaned and then
models were used toto create
draw athe predicted data. The spline and Fourier smoothing estimation was
determines theused
confidence limitnew fitting
to 95% andcurve.
then The 2D-CurveTable
computes the fittingsoftware was used
model, which is theto create
best fit the
to
performed
new in order
fitting model to come
and it up
was with
then excellent
compared smoothing
with
the original data. The best model is computed as in Equation (12): the of
output data
of and
neural good
models. fitting.
First, The
the experiment
data was
determines
cleaned the confidence
and then used limit
to draw to 95% and thendata.
the predicted computes the and
The spline fitting model,
Fourier which isestimation
smoothing the best fitwas to the
performed in order toycome
= a + bxup +with
cx2 +excellent
dx3 + ex4 +smoothing
fx5 + gx6 + ofhx7data
+ ix8and
+ jx9good
+ kx10fitting. The experiment(12)
original data. The best model is computed as in Equation (12):
determines the confidence
This polynomial modellimit to 95%
obtained and then computes
a coefficient the fitting
of determination model,
R2 value which and
of 0.9554 is theanbest fit to
F-value
the original data.
of 96.5827. The ycurve The best model 2 is 3
computed 4 as 5
in 6
Equation 7
(12): 8 9 10
= a + of bxthe+ cxproposed
+ dx +model ex + fx + gx + hx
in Equation (12)+ is
ix illustrated
+ jx + kxin Figure 15. The (12)
confidence limits are plotted
y = a +inbx
green
+ cx2and
+ dxthe forecasted
3 + ex interval
4 + fx5 + gx 6 + hx7 +values
ix8 + jxare plotted
9 + kx10 in red. The result
(12)
of thepolynomial
This regression model
modelisobtained
plotted in blue using of
a coefficient a dotted-line.
determination 2 value of
The Rproposed model
0.9554 forecasts
and an the
F-value
This
electrical polynomial
current model obtained a coefficient of determination R 2 value of 0.9554 and an F-value
of 96.5827. The curveproduction on a daily
of the proposed basis.
model inItEquation
is clear that theispolynomial
(12) illustratedmodel forecasts
in Figure the future
15. The confidence
of 96.5827. Thewith
curve of the proposed model indata
Equation (12)soisforillustrated in Figure 15. The
limitsdata
arefor 60 days
plotted in greena good
andfit.
the The original
forecasted input
interval only does
values are plotted 35 days
in and
red.the forecasting
The result of the
confidence limits are plotted in green
model is producing the current for 60 days. and the forecasted interval values are plotted in red. The result
regression
of themodel is plotted
regression model inis blue
plotted using a dotted-line.
in blue The proposed
using a dotted-line. model model
The proposed forecasts the electrical
forecasts the
current production
electrical currentonproduction
a daily basis. It is clear
on a daily basis.that the polynomial
It is clear model model
that the polynomial forecasts the future
forecasts data for
the future
60 days with
data for a
60good
days fit.
withThe original
a good input
fit. The datainput
original onlydatadoesonly
so for
does35sodays and
for 35 theand
days forecasting model is
the forecasting
producing
modelthe current for
is producing the60current
days. for 60 days.
Figure 15. The result of the forecasting proposed model illustrated in the Equation (13).
Figure 16 shows the Comparison of the Desired Output and the Forecasted production of SOFM
for 56 days. The actual amount of current produced (Ipv/Am) that is generated from the installed
Figure
Figure 15. The
15. The result
result of the
of the forecastingproposed
forecasting proposed model
model illustrated
illustratedininthe Equation
the (13).(13).
Equation
Figure 16 shows the Comparison of the Desired Output and the Forecasted production of SOFM
Figure 16 shows
for 56 days. the Comparison
The actual of the Desired
amount of current Output
produced andthat
(Ipv/Am) theisForecasted production
generated from of SOFM
the installed
for 56 days. The actual amount of current produced (Ipv/Am) that is generated from the installed
Energies 2017, 10, 971 13 of 19
Energies 2017, 10, 971 13 of 19
renewable power
renewable powerstation PVPV
station appears
appearsininblue.
blue. And theforecasted
And the forecastedproduction
production of current
of current fordays
for 56 56 days
is is
Energies 2017, 10, 971 13 of 19
shown in red.
shown in red.
renewable power station PV appears in blue. And the forecasted production of current for 56 days is
shown in red. Comparison of the Desired Output and the Forecasting production of SOFM
8000 8000
Comparison of the Desired Output and the Forecasting production of SOFM
7000 7000
8000 8000
6000 6000
7000 7000
5000 5000
Ipv/AmIpv/Am
6000 6000
4000 4000
5000 5000
3000 3000
4000 4000
2000 2000
3000 3000
1000 1000
2000 2000
0 0
1000 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 1000
Day Number
0 Ipv /Am Ipv /Am Output0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55
Day Number
Figure 16. Comparison of the desired output and the forecasting
Ipvproduction
/Am of SOFM.
Figure 16. Comparison of the desired output and the forecasting productionIpv
of/Am Output
SOFM.
5.2. Self-Organizing Feature Map, Support Vector Machine and Multi-Layer Perceptron Comparison
Figure 16. Comparison of the desired output and the forecasting production of SOFM.
5.2. Self-Organizing Feature Map, Support Vector Machine and Multi-Layer Perceptron Comparison
The comparison study covers all data sets (training, cross validation, and testing the results of
5.2. proposed
The
the Self-Organizing
comparisonmodels) Feature
study Map, Support
covers
to ensure that theVector
all data setsMachine
measurement and
(training, Multi-Layer
cross
factors Perceptron
validation,
will apply andComparison
for input testing the results
and output data of
sets of
the proposedTheneural models.
models)
comparison The comparison
tostudy
ensure that all
covers thedatawassets
applied
measurement using
(training, several
factors
cross willmeasuring
applyand
validation, forcriteria to verify
input the
testing and the
output
results of data
results
sets of
theneuralof the
proposed proposed
models.
models) tomodels,
The comparison
ensure such aswas
that the MSE, the MAE
applied
measurement and coefficient
using
factorsseveral offor
determination.
measuring
will apply inputcriteria Figure
and output 17 the
to verify
data
demonstrates
sets of neural the comparison
models. The of the
comparison final MSE
was of
appliedthe MLP,
using SVM,
several and SOFM
measuring
results of the proposed models, such as MSE, the MAE and coefficient of determination. Figure 17 models
criteria based
to on
verify the
the
training
results ofand
thecross-validation phases. This figurethe
shows thatandthe proposed SOFM model has Figure
achieved
demonstrates theproposed
comparison models,
of the such as MSE,
final MSE of MAE the MLP, coefficient
SVM, andofSOFM determination.
models based 17 on the
lower MSE values
demonstrates the in comparison
comparison of thewith
finalthe
MSEMLP of andMLP,
the SVMSVM, models.
and The
SOFM experiments
models prove
based that
onachieved
the
training and cross-validation phases. This figure shows that the proposed SOFM model has
training
training the
andnetwork multi-times
cross-validation phases.withThis
a varied
figurenumber of epochs
shows that is necessary
the proposed SOFM to model
enhance hasthe results
achieved
lowerofMSE
neuralvalues
models,in comparison
as depicted in with
Table the2. MLP and SVM models. The experiments prove that training
lower MSE values in comparison with the MLP and SVM models. The experiments prove that
the network multi-times with a varied number of epochs is necessary to enhance the results of neural
training the network multi-times with a varied number of epochs is necessary to enhance the results
The MSE
2. forintraining
models, as depicted
of neural models,inasTable
depicted Table and2. cross validation using different types of neural models
1
The MSE for training and cross validation using different types of neural models
0.11
0.01
0.1
MSE value
0.001
0.01
MSE value
0.0001
0.001
0.00001
0.0001
MLP- SVM- SOFM- SOFM- SOFM- SOFM-
1000 1000 1000 2000 5000 10000
0.00001 0.0104068 0.12746142 0.00049031 0.00015558 0.0001613 0.000125
MSE for Cross Validation
MLP- SVM- SOFM- SOFM- SOFM- SOFM-
MSE for Training 0.0058885
1000 0.10187602
1000 0.0006664
1000 0.00010121
2000 0.00006662
5000 0.00003749
10000
MSE for Cross Validation 0.0104068 0.12746142 0.00049031 0.00015558 0.0001613 0.000125
Figure 17. The comparison results of the final MSE for the multi-layer perceptron (MLP), support
MSE for Training 0.0058885 0.10187602 0.0006664 0.00010121 0.00006662 0.00003749
vector machine (SVM) and SOFM models.
Energies 2017, 10, 971 14 of 19
Table 2. The comparison results of MSE in training and cross-validation, mean absolute error (MAE),
R, and accuracy for multi-layer perceptron (MLP), support vector machine (SVM) and SOFM models.
Energies 2017, 10, 971 Neural 14 of 19

MSE for MSE for Cross
Model No. Networks R MAE Accuracy
Training Validation
Type results of MSE in training and cross-validation, mean absolute error (MAE),
Table 2. The comparison
R, and
1 accuracy for multi-layer perceptron
MLP-1000 0.0058(MLP), support vector machine
0.010 (SVM) and
0.9849 SOFM models.
3.639 92%
2 SVM-1000 0.1018 0.1274 0.9820 4.537 61%
Model Neural Networks MSE for MSE for Cross
3 SOFM-1000 0.0005 0.0006 0.9915
R MAE0.456 Accuracy
93%
No. Type Training Validation
4 SOFM-2000 0.0001 0.0001 0.9996 0.531 99%
5 1 MLP-1000
SOFM-5000 0.0058
0.00006 0.010
0.0001 0.9849
0.9996 3.639
0.435 92% 99%
6 2 SVM-1000
SOFM-10000 0.1018
0.00003 0.1274
0.0001 0.9820
0.9997 4.537
0.3615 61% 99%
3 SOFM-1000 0.0005 0.0006 0.9915 0.456 93%
4 SOFM-2000 0.0001 0.0001 0.9996 0.531 99%
MAE is 5
appraising theSOFM-5000
proper fit of 0.00006
predicting 0.0001
values 0.9996values.
with desired 0.435 It measures
99%
the forecast
6 SOFM-10000 0.00003 0.0001 0.9997 0.3615 99%
error in a time series analysis. In statistics, MAE is a factor used to measure the difference between the
predicted value, the true value, and the expected range of errors that can be expected in the forecasting
MAE is appraising the proper fit of predicting values with desired values. It measures the
model. It is computed as in Equation (13):
forecast error in a time series analysis. In statistics, MAE is a factor used to measure the difference
between the predicted value, the true value, and the expected range of errors that can be expected in
1 n (13):
the forecasting model. It is computedMAE
as in Equation
= k f −y k
n ∑ i i (13)
i
1
= ‖ − ‖ (13)
where yi is the observed value of the actual output and fi is the predicted value.
Thewhere
model isthat
the achieves a lower
observed value value
of the actualof MAEand
output is considered to be
fi is the predicted the best performing model.
value.
Figure 18 depicts the that
The model comparison
achieves aof MAE
lower values
value using
of MAE differenttotypes
is considered be theofbest
neural models.
performing Obviously,
model.
Figure 18 depicts
the SOFM-10000 modelthe comparison
achieved of MAE
a lower values
MAE. usingit different
Thus, types ofasneural
is considered models.
the best Obviously,
fitting model. Lastly,
the SOFM-10000 model achieved a lower MAE. Thus, it is considered as the best fitting model. Lastly,
the SOFM models achieved a high accuracy of 99% in comparison with the MLP and SVM models,
the SOFM models achieved a high accuracy of 99% in comparison with the MLP and SVM models,
which obtained accuracy values of 92% and 61% respectively.
which obtained accuracy values of 92% and 61% respectively.
5 4.53761
4.5
4 3.63927
3.5
3
MAE value
2.5
2
1.5
1 0.45613 0.53121 0.43577 0.36156
0.5
0
MLP-1000 SVM-1000 SOFM-1000 SOFM-2000 SOFM-5000 SOFM-10000
Neural Networks Type with the number of epoch
Figure 18. The comparison of MAE using different types of neural network models.
Figure 18. The comparison of MAE using different types of neural network models.
5.3. Comparison with the Relevant Models
5.3. Comparison with the Relevant Models
Table 3 illustrates several representative examples of ANN techniques applied to the modelling
or prediction
Table of PV
3 illustrates energyrepresentative
several and solar radiation in the period
examples of ANN1998–2016. The results
techniques from the
applied articles
to the modelling
presented in Table 3 show that, in general, the errors associated with predictions (monthly, daily,
or prediction of PV energy and solar radiation in the period 1998–2016. The results from the articles
hourly, and every minute) are between 3% and 14%. Moreover, MLP, SVM, and SOFM are the main
presented in Table
techniques 3 show
used to modelthat, in general,
or predict theand
PV energy errors
solar associated with predictions
radiation. However, it is noted that(monthly,
the MLP daily,
hourly, is
andthe most commonly used technique and can be used with exogenous parameters or coupledare
every minute) are between 3% and 14%. Moreover, MLP, SVM, and SOFM the main
with
techniques
otherused to model
predictors. The or predict
results PVinenergy
shown Table 3 and solarthe
illustrate radiation.
comparison However,
of forecastitaccuracy.
is notedFactors
that the MLP
suchcommonly
is the most as MAPE, MBE,used RMSE, MAE, R,and
technique R2 and
canaccuracy
be used arewith
usedexogenous
to verify the parameters
proposed models. In the with
or coupled
R or R2 andThe
other predictors. RMSE or MSE
results terms, in
shown theTable
ANN model best fits
3 illustrate thethecomparison
experimentalof data. The best
forecast results of Factors
accuracy.
the proposed neural model achieved an MSE value of 0.00003. Considering the same year (2016),
such as MAPE, MBE, RMSE, MAE, R, R2 and accuracy are used to verify the proposed models. In the
RMSE values of 10.09% and 5.60% were obtained in references [33,34], respectively. Furthermore, the
R or R2 proposed
and RMSE model MSE terms,
or (SOFM) has thethe ANN
highest modelofbest
accuracy 99%.fits the experimental data. The best results
of the proposed neural model achieved an MSE value of 0.00003. Considering the same year (2016),
Energies 2017, 10, 971 15 of 19
RMSE values of 10.09% and 5.60% were obtained in references [33,34], respectively. Furthermore, the
proposed model (SOFM) has the highest accuracy of 99%.
Table 3. Comparison of production accuracy results.
Network
Authors/Reference Year Location Model
Type
Awada et al. [12] 2012 Malaysia MLP MBE: −1.12, RMSE: 8.57, MAPE: 7.29
Mohandes et al. [17] 1998 Saudi Arabia MLP MAPE: 4.49
Adnan et al. [18] 2005 Turkey MLP MAPE: 6.70. R value: 0.958
Elminir et al. [21] 2007 Egypt MLP SD: 9.00
MLP-FF1 MAPE: 10.3, MSE: 0.00028
Rehman and Mohandes [22] 2008 Saudi Arabia MLP-FF2 MAPE: 11.8, MSE: 0.0052
MLP-FF3 MAPE: 4.49, MSE: 0.00004
Mubiru et al. [23] 2008 Uganda MLP-FF MBE: 0.059, RMSE: 0.385, R value: 0.971
MLP-Max. MAPE: 14.13%, R value: 0.97
Mehmet et al. [24] 2007 Turkey
MLP-Min. MAPE: 4.49%, R value: 0.67
Bosch et al. [25] 2008 Spain MLP-Max. RMSE: 6%, MBE:0.2
MLP-LM SD: 0.027, R value: 0.978
Fadare [26] 2009 Nigeria
MLP-SCG SD: 0.1104, R value: 0.961
MBE: −0.04, RMSE: 0.746, MPE: 1.55,
Yingni J. [27] 2008 China MLP
Accuracy 94.81%
Graditi et al. [33] 2016 Italy MLP RMSE: 10.09, R2 : 0.978
Huang et al. [34] 2016 USA SVM MAPE: 3.29, RMSE: 5.60, R value: 0.998
MSE: 0.0001, 0.00007, 0.00003;
SOF
MAE: 0.361, 3.369, 4.537;
Current study 2017 Oman MML
R value: 0.999, 0.984, 0.982;
PSVM
Accuracy: 99%, 92%, 61%
6. Sensitivity Analysis
Sensitivity analysis testing helps us to understand the relative importance among the inputs of
the neural model and demonstrates how the output of the model varies in response to the variation of
an input. Moreover, it helps to determine and eliminate an irrelevant input which reduces the data
collection cost and can sometimes improve a network’s performance. This test includes solar radiation
and temperature. The sensitivity analysis of the SOFM model supports the view that solar radiation
is the most influential variable on the output of the PV, as might be expected. Figure 19 depicts the
sensitivity analysis results. It is well known that temperature and solar radiation play a central role in
PV power generation and efficiency. Both power generation and efficiency have inverse and direct
relationships with temperature and solar radiation, respectively. This result is important because, as
the temperature in Oman is relatively high, which reduces the PV power productivity, the effect is
low due to high solar radiation. Therefore, the large quantity of solar radiation in Oman is reflected
positively in the PV performance, which reduces the importance of the reduction due to the increase
in temperature.
depicts the sensitivity analysis results. It is well known that temperature and solar radiation play a
central role in PV power generation and efficiency. Both power generation and efficiency have
inverse and direct relationships with temperature and solar radiation, respectively. This result is
important because, as the temperature in Oman is relatively high, which reduces the PV power
productivity, the effect is low due to high solar radiation. Therefore, the large quantity of solar
Energies 2017, 10, 971in Oman is reflected positively in the PV performance, which reduces the importance of the 16 of 19
radiation
reduction due to the increase in temperature.
Sensitivity About the Mean
0.05
0.04
Sensitivity value
0.03
0.02
0.01
0
Amb Tempr Solar Rad
Input variables
Figure 19. The sensitive analysis test for the SOFM network.
Figure 19. The sensitive analysis test for the SOFM network.
7. Conclusions
The main objective of this paper was to propose mathematical models to forecast the production of
electrical current from photovoltaic panels using neural network techniques. Though the use of neural
network methodologies to control and forecast current production is not new and many researchers
have used and implemented them previously (15–33), it is worth noting that most of these applications
are based on the use of supervised feed-forward neural networks. On the other hand, few studies
have focused on recurrent neural networks and unsupervised models. The conclusions of this study
contribute to the following:
(1) This paper is to propose a mathematical model to predict the production of electrical current from
photovoltaic panels using the SOFM neural model, which is based on an unsupervised feedback
neural network employed to improve the performance of the network. Moreover, this helps to
reduce the error rate in predicting results generated by the neural model, when compared with
the actual output. The data sets are divided into three sets (40% of the data sets for training
the network, 20% of the data sets for determining the errors in training data set by using the
cross-validation techniques, and 40% of the data sets for testing the output results of the network).
(2) The second contribution is performing the comparative study with the other two models (MLP
and SVM) in order to ensure reliable results. The comparison study relied on several methods to
measure the performance of the network, including the testing of the input and output data sets.
These methods included MSE, MAE, the correlation factor, and the coefficient of determination.
The proposed SOFM model achieves a final MSE of 0.0007 in the training phase and 0.0005 in
the cross validation phase. This MSE is very small in comparison with other neural networks,
such as that for the SVM model, which achieved an MSE of 0.0058. Moreover, the MLP model
achieved an MSE of 0.026. The experiments prove that there is a strong relationship between
input and output variables based on a correlation coefficient of 0.9989. In addition, the proposed
SOFM model best fits the desired values based on the coefficient of determination R2 value of
0.9555. For validating the results of the proposed model, MAE was used to confirm that the SOFM
model closely fits the desired output, resulting in an MAE value of 0.361. SVM and MLP obtained
MAE values of 4.537 and 3.639, respectively. Table 2 shows a comparison of MSE training, MSE
cross-validation, MAE, R, and accuracy for the MLP, SVM, and proposed SOFM model.
(3) Another contribution involves implementing the sensitivity analysis to determine the degree
of importance among the inputs of the neural model and demonstrating how the output of
the model varies in response to a variation of an input. The sensitivity analysis of the SOFM
model confirms that solar radiation is the most influential variable on the output of PV systems.
Moreover, several epochs were implemented (1000, 2000, 5000, and 10,000), which helps to
enhance the results of the proposed model. In addition, the accuracy of the proposed SOFM
Energies 2017, 10, 971 17 of 19
model is compared with the accuracy of the SVM and MLP models. It was found that the accuracy
of the SOFM model is 99%, while the accuracy of SVM and MLP are 62% and 91%, respectively.
(4) The last contribution of this paper involves implementing the regression techniques to generate a
close-fitting model for the results of the SOFM model and the desired data. The best regression
model was computed as in Equation (12), which is based on a polynomial equation. A coefficient
of determination R2 value of 0.9554 and an F-value of 96.582 were obtained. The curve of the
proposed model of Equation (12) is illustrated in Figure 15. Finally, the proposed SOFM model
predicts well and fits actual values based on a coefficient determination R2 value of 0.9555. Actual
values were gathered from solar irradiation and the production of the PV system of solar cells
and the Photovoltaic research laboratory, which is installed at Sohar University.
Acknowledgments: The research leading to these results has received Research Project Grant Funding from the
Research Council of the Sultanate of Oman, Research Grant Agreement No. ORG SU EI 11 010. The authors would
like to acknowledge support from the Research Council of Oman.
Author Contributions: All the researchers were worked equally. Each of them helped in the writing of this paper
and the design of the practical model and then proposed mathematical models.
Conflicts of Interest: The authors declare no conflict of interest.
Nomenclature
SOFM Self-Organizing Feature Map
MLP Multilayer Perceptron
MLFF Multilayered feed-forward
SVM Support Vector Machine
MBE Mean Bias Error
RMSE Root Mean Square Error
MSE Mean Square Error
MAE Mean Absolute Error
LM Levenberg Marquardt
CGP Polak-Ribiere Conjugate Gradient
SCG Scaled conjugate gradient
R Correlation
R2 Coefficient of determination
GCC Gulf Cooperation Council
PV Photovoltaic
GSR global solar radiation
ANN Artificial neural networks
SOM Self-Organizing Map
MAPE mean absolute percentage error
RBF Radial Basis Functions
BMU Best Matching Unit
k1, k2 Indexes of the winner-neuron
ρ Distance to the winner-neuron
α(t) Learning rate
ei (n) Error correction learning
δi (n) Local error
SSE Summation of the squared error
xi Value of the ith observation
x Mean value of all the observations
x First dataset {x1 , ..., xn }
y Other dataset {y1 , ..., yn } containing n value
Energies 2017, 10, 971 18 of 19
References
1. Kazem, H.A.; Khatib, T. Photovoltaic Power System Prospective in Oman, Technical and Economical Study, 1st ed.;
LAP LAMBERT Academic Publishing: Saarbrücken, Germany, 2013; ISBN 978-3659372957.
2. Authority for Electricity Regulation in Oman. Study on Renewable Resources; Final Report; Authority for
Electricity Regulation in Oman: Muscat, Oman, 2008.
3. Altunkaynak, A.; Ozger, M. Comments on Temporal significant wave height estimation from wind speed by
perceptron Kalman filtering. Ocean Eng. 2004, 31, 1245–1255.
4. Fakham, H.; Lu, D.; Francois, B. Power control design of a battery charger in a hybrid active PV generator
for load-following applications. IEEE Trans. Ind. Electron. 2011, 58, 85–94.
5. Yousif, J.H. Information Technology Development; Academic Publishing: Saarbrücken, Germany, 2011;
ISBN 9783844316704.
6. Kalogirou, S.A. Applications of artificial neural networks for energy systems. Appl. Energy 2000, 67, 17–35.
7. Ogliari, E.; Grimaccia, F.; Leva, S.; Mussetta, M. Hybrid predictive models for accurate forecasting in PV
systems. Energies 2013, 6, 1918–1929.
8. Hernández, L.; Baladrón, C.; Aguiar, J.M.; Carro, B.; Sánchez-Esguevillas, A. Classification and clustering of
electricity demand patterns in industrial parks. Energies 2012, 5, 5215–5228.
9. Bracale, A.; Caramia, P.; Carpinelli, G.; Di Fazio, A.R.; Ferruzzi, G. A Bayesian method for short-term
probabilistic forecasting of photovoltaic generation in smart grid operation and control. Energies 2013, 6,
733–747.
10. Sheela, K.S.G.; Deepa, S.N. An efficient hybrid neural network model in renewable energy systems.
In Proceedings of the 2012 IEEE International Conference on Advanced Communication Control and
Computing Technologies (ICACCCT), Ramanathapuram, India, 23–25 August 2012.
11. Argiriou, A.A.; Bellas-Velidis, I.; Kummert, M.; André, P. A neural network controller for hydronic heating
systems of solar buildings. Neural Netw. 2012, 17, 427–440.
12. Awada, A.; Pasupuleti, J.; Khatib, T.T.N.; Kazem, H.A. Modeling and characterization of a photovoltaic
array based on actual performance using cascade-forward back propagation artificial neural network. J. Sol.
Energy Eng. 2015, 137. [CrossRef]
13. Dorvlo, A.S.S.; Jervaseb, J.; Al-Lawatib, A. Solar radiation estimation using artificial neural networks.
Appl. Energy 2015, 71, 307–319.
14. Kalogirou, S.A.; Bojic, M. Artificial neural networks for the prediction of the energy consumption of a passive
solar building. Energy 2015, 25, 479–491.
15. Zhang, Y.; Chen, Q. Prediction of Building Energy Consumption Based on PSO-RBF Neural Network.
In Proceedings of the 2014 IEEE International Conference on System Science and Engineering (ICSSE),
Shanghai, China, 11–13 July 2014.
16. Zhou, J.; Wu, Y.Z.; Yan, G. Solar radiation estimation using artificial neural networks. J. Sol. Energy 2005, 26,
509–512.
17. Mohandes, M.; Rehman, S.; Halawani, T.O. Estimation of global solar radiation using artificial neural
networks. Renew. Energy 1998, 14, 179–184.
18. Adnan, S.; Arcaklioglu, E.; Ozalp, M.; Caglar, N. Forecasting based on neural network approach of solar
potential in Turkey. Renew. Energy 2005, 30, 1075–1090.
19. Sozen, A.; Arcaklioglu, E.; Ozalp, M.; Kanit, E.G. Use of artificial neural networks for mapping of solar
potential in Turkey. Appl. Energy 2004, 77, 273–286.
20. Elminir, H.K. Estimation of solar radiation components incident on Helwan site using neural networks.
Sol. Energy 2005, 79, 270–279.
21. Elminir, H.K.; Azzam, Y.A.; Younes, F.I. Prediction of hourly and daily diffuse fraction using neural network,
as compared to linear regression models. Energy 2007, 32, 1513–1523.
22. Rehman, S.; Mohandes, M. Artificial neural network estimation of global solar radiation using air temperature
and relative humidity. Energy Policy 2008, 36, 571–576.
23. Mubiru, J.; Banda, E.J.K.B. Estimation of monthly average daily global solar irradiation using artificial neural
networks. Sol. Energy 2008, 82, 181–187.
24. Mehmet, B.; Sahin, B.; Yasar, A. Application of artificial neural networks for the wind speed prediction of
target station using reference stations data. Renew. Energy 2007, 32, 2350–2360.
Energies 2017, 10, 971 19 of 19
25. Bosch, J.L.; Lopez, G.; Batlles, F.J. Daily solar irradiation estimation over a mountainous area using artificial
neural network. Renew. Energy 2008, 33, 1622–1628.
26. Fadare, D.A. Modeling of solar energy potential in Nigeria using an artificial neural network model.
Appl. Energy 2009, 86, 1410–1422.
27. Jiang, Y. Prediction of monthly mean daily diffuse solar radiation using artificial neural networks and
comparison with other empirical models. Energy Policy 2008, 36, 3833–3837.
28. Lam, J.C.; Wan, K.K.W.; Yang, L. Solar radiation modeling using ANNs for different climates in China.
Energy Convers. Manag. 2008, 49, 1080–1090.
29. Kazem, H.A.; Khatib, T.; Sopian, K.; Elmenreich, W. Performance and feasibility assessment of a 1.4kW roof
top grid-connected photovoltaic power system under desertic weather conditions. Energy Build. 2014, 82,
123–129.
30. Zhang, N.; Behera, K.P. Solar radiation prediction based on recurrent neural networks trained by
Levenberg-Marquardt backpropagation learning algorithm. In Proceedings of the 2012 IEEE PES Innovative
Smart Grid Technologies (ISGT), Washington, DC, USA, 16–20 January 2012.
31. Yona, A.; Senjyu, T.; Funabashi, T.; Mandal, P.; Kim, C. Decision Technique of Solar Radiation Prediction
Applying Recurrent Neural Network for Short-Term Ahead Power Output of Photovoltaic System. Smart Grid
Renew. Energy 2013, 4, 32–38.
32. Capizzi, G.; Napoli, C.; Bonanno, F. Innovative Second-Generation Wavelets Construction with Recurrent
Neural Networks for Solar Radiation Forecasting, Neural Networks and Learning Systems. IEEE Trans.
Neural Netw. Learn. Syst. 2012, 23, 1805–1815.
33. Graditi, G.; Ferlito, S.; Adinolfi, G. Comparison of Photovoltaic plant power production prediction methods
using a large measured dataset. Renew. Energy 2016, 90, 513–519.
34. Huang, C.; Bensoussan, A.; Edesess, M.; Tsui, K.L. Improvement in artificial neural network-based estimation
of grid connected photovoltaic power output. Renew. Energy 2016, 97, 838–848.
35. Kazem, H.A.; Yousif, J.H.; Chaichan, M.T. Modelling of Daily Solar Energy System Prediction using Support
Vector Machine for Oman. Int. J. Appl. Eng. Res. 2016, 11, 10166–10172.
36. Yousif, J.H.; Kazem, H.A. Modeling of Daily Solar Energy System Prediction using Soft Computing Methods
for Oman. Res. J. Appl. Sci. Eng. Technol. 2016, 13, 237–244.
37. Kazem, H.A.; Khatib, T.; Sopian, K. Sizing of a standalone photovoltaic/battery system at minimum cost for
remote housing electrification in Sohar, Oman. Energy Build. 2013, 6, 108–115.
© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Predictive Models For Photovoltaic Electricity Production in Hot Weather Conditions

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Predictive Models For Photovoltaic Electricity Production in Hot Weather Conditions

Uploaded by

Copyright:

Available Formats

energies

Academic Editor: Francesco Calise

Energies 2017, 10, 971; doi:10.3390/en10070971 www.mdpi.com/journal/energies

0 0.01 0.1 1 10 100 1000 10000

Energies 2017, 10, 971 6 of 19

Figure 4. The PV system.

4. Self-Organizing Feature Map Architecture and Design

h(ρ, t) = exp(ρ2 /2σ2 (t))

where ρ is the distance to the winner-neuron, which is computed as in Equation (2):

hat h(ρ,t) or the French hat h(ρ) as in Equation (3):

Thus weights of all the neurons are updated as in Equation (4):

Wij (t + 1) = Wij (t) + α(t)h((ρ, t)( x l (t) − Wij (t)) (4)

wij (n + 1) = wij (n) + ηδi (n) + x j (n) (6)

Figure 6. The self-organizing feature map (SOFM) architecture.

5. Self-Organizing Feature Map Architecture and Design

5.1. Epoch Results

0.01 Training MSE

0.005 Training MSE

0.005 Cross Validation

0.01 Training MSE

Energies 2017, 10, 971 Neural 14 of 19

Table 3. Comparison of production accuracy results.

Sensitivity About the Mean

You might also like