Engineering Applications of Artificial Intelligence 92 (2020) 103688

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

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

A particle swarm optimisation-trained feedforward neural network for

predicting the maximum power point of a photovoltaic array
Sadeq D. Al-Majidi a,b , Maysam F. Abbod a ,∗, Hamed S. Al-Raweshidy a
a Department of Electronic and Computer Engineering, College of Engineering, Design and Physical Sciences, Brunel University London, Uxbridge UB8 3PH, United
Kingdom
b
Department of Electrical Engineering, College of Engineering, University of Misan, Amarah 62001, Iraq

ARTICLE INFO ABSTRACT

Keywords: In this paper, a feedforward Artificial Neural Network (ANN) technique using experimental data is designed for
Artificial neural network (ANN) predicting the maximum power point of a photovoltaic array. An ANN model training strategy is challenging
Fuzzy logic control (FLC) due to the variations in the training and the operation conditions of a photovoltaic system. In order to improve
Maximum power point tracking (MPPT)
ANN model accuracy, the Particle Swarm Optimisation (PSO) algorithm is utilised to find the best topology
Photovoltaic (PV)
and to calculate the optimum initial weights of the ANN model. Hence, the dilemma between computational
Perturb and observe (P&O)
Efficiency of MPPT (𝜂 MPPT)
time and the best-fitting regression of the ANN model is addressed, as well as the mean squared error
being minimised. To evaluate the proposed method, a MATLAB/Simulink model for an installed photovoltaic
system is developed. Experimental data of a sunny and cloudy day are utilised to determine the average
efficiency of this proposed method under varying atmospheric conditions. The results show that the optimised
feedforward ANN technique based on the PSO algorithm using real data predicts the maximum power point
accurately, achieving hourly average efficiencies of more than 99.67% and 99.30% on the sunny and cloudy
day, respectively.

1. Introduction
Nowadays, electrical energy is a vital element for maintaining
lifestyles and as a result, the demand for it is sharply rising. To generate
this energy, classical sources, such as fossil fuel, are mostly used.
Unfortunately, these cause air pollution and global warming owing
to CO2 emissions from. To solve the problem of increasing demand
for electrical energy and minimising the side effects of burning fossil
fuels, using renewable energy resources has been widely proposed.
Photovoltaic (PV) arrays, wind turbines and hydropower stations are
the major renewable energy sources, coming from sunlight, wind and
fast running water, respectively. PV systems are considered the most
attractive energy source owing to their lower cost when compared to
other renewables (Bakos and Soursos, 2002). In addition, they can be
almost installed everywhere and with different capacities.
Several studies have pointed out that the surface of earth receives
power radiation from the sunlight at about 1.8×1011 MW and that
this is much more than the total global demand for electrical power
(Shah et al., 2015). However, the major problem of PV systems low
efficiency, because power generation from its module mostly depends
upon climatic conditions, which can result in lost power of up to 25%
(Román et al., 2006). To improve the efficiency of the PV system, a
maximum power point tracker (MPPT) technique has been designed,
as shown in Fig. 1, which can achieve maximum power generation for
various atmospheric conditions. The principle work of this technique
is that it feeds an appropriate duty cycle (D) to the power conversion
system, such as a DC–DC converter, based on the output and/or input
of the PV module. In addition, it improves the stability and reliability of
the PV generation when it is connected to a grid. In recent years, several
MPPT techniques have been proposed that address efficiency, cost,
and form of implementation, which are divided into two major types:
classical techniques, such as the perturbation and observation (P&O)
algorithm (Kamala et al., 2017); and artificial intelligence techniques
(AI), including the fuzzy logic controller (FLC) (Al-Majidi et al., 2018a)
and artificial neural network (ANN) (Bahgat et al., 2005).
The classic methods, such as P&O algorithm, are widely used for
PV-MPPT controllers because they have simple design and low imple-
mentation costs (Enany et al., 2016). However, such controllers have
a slow tracking speed, high fluctuation, and a drift problem (Al-Majidi
et al., 2018b). Hence, AI techniques based on MPPT controllers have
been proposed to address these issues. These do not need complex
mathematics and exact parameters when managing the application
system (Kwong et al., 2016; Mellit and Kalogirou, 2008). In particular,
the FLC method is a very attractive MPPT technique for PV systems,
because it has a faster tracking speed and less oscillation than the

∗ Corresponding author.
E-mail address: maysam.abbod@brunel.ac.uk (M.F. Abbod).

https://doi.org/10.1016/j.engappai.2020.103688
Received 13 September 2019; Received in revised form 16 March 2020; Accepted 28 April 2020
Available online 7 May 2020
0952-1976/© 2020 Elsevier Ltd. All rights reserved.
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
classical MPPT techniques. However, its main disadvantage is that, as
with the classic techniques, it still suffers from the drift phenomenon
associated with rapid change in the irradiance level (G) and operation
temperature (T). This is due to requiring a good knowledge in PV
systems to get an accurate defined membership functions of the FLC
model (Guenounou et al., 2014; Li et al., 2019; Nabipour et al., 2016).
Recently, MPPT with an ANN technique has been used to solve the issue
of conventional FLC-MPPT, which provides a heuristic output function
using numerical quantifying data. As a consequence, good knowledge
of the PV parameters is not required to design the optimised MPPT
controller (Salah and Ouali, 2011). However, the training strategy of
the ANN is a major weakness when using it as a prediction model
(Al-Majidi et al., 2019; Livieris, 2019; Plumb et al., 2005).
In this paper, a feedforward ANN technique is employed to predict
the maximum power point (MPP) of a PV array using a large real
training dataset. Those data are collected from experimental tests on
a PV array installed at Brunel University London, Uxbridge, United
Kingdom. To optimise the training strategy of the ANN model, the PSO
algorithm is utilised. This strategy is divided into two parts: determin-
ing the right topology and then, optimising the initial weights of the
feedforward ANN model. That is, the issue between the computational
time and the best-fitting regression of the distribution of the ANN
nodes is solved in the first part, while the global minimum training
error of the ANN model is addressed in the second. Consequently, the
predicting function of the proposed ANN method is improved under
various weather conditions. The G and T of the atmospheric conditions
are used as the inputs of the proposed ANN model, whilst the predicting
power (Pref.) is the output. This regulates the D of a DC–DC boost
converter after being compared with the PV’s actual power (Pact.)
using a proportional–integral (PI) controller. The D is converted into
the signal of the DC–DC converter using a pulse-width modulation
generator (PMW) to regulate the operating MPP of the PV array, as
shown in Fig. 1.
The results prove that the converging speed of the proposed method
for a transient state is faster than for the conventional ANN, FLC
and P&O techniques. Moreover, the power prediction is accurate and
the drift phenomenon for different atmospheric conditions is avoided.
Hence, the optimised ANN-MPPT method based on the PSO algorithm
achieves the highest efficiency on both sunny and cloudy days. Finally,
those optimisations are demonstrated to be simple to design by com-
parison with the hybrid method. The rest of this paper is organised
as follows: Section 2 reports on related works on the use of optimised
ANNs in PV applications. Sections 3 and 4 cover the principles the
feedforward ANN technique and PSO algorithm, respectively. Section 5
introduces the modelling of the PV system. Section 6 explains the
methodology of the collected training data, whilst Section 7 presents
the proposed method. In Section 8, the results are provided. Experimen-
tal measurement test results for sunny and cloudy days are provided
and discussed in Section 9, whilst Section 10 contains the conclusion.
2. Related works
The ANN approach is the leading technique used in PV-MPPT ap-
plications, because it is able to predict the MPP of PV arrays accurately
under various weather conditions (Messalti et al., 2017). Unfortunately,
as aforementioned, the ANN model training strategy is challenging in
relation to designing the optimised ANN-MPPT controller for PV sys-
tems due to the variations in the training data and operation conditions.
In order to improve model accuracy, several researchers have proposed
various optimisations, among them Hiyama et al. (1997) presented a
novel methodology based on a regression analysis method to select
effective data for training an ANN-MPPT model. The data set was
collected from experimental tests of a PV array installed at Kumamoto
University, Kumamoto, Japan, for one year and then, the measured data
for two days with a widely varying weather conditions were selected.
This proposal gives accurate predictions when compared with an ANN-
MPPT based on the total data. Afsin and Akkaya (2012a) used a genetic
2
Engineering Applications of Artificial Intelligence 92 (2020) 103688
algorithm (GA) to select automatically the effective data among all
those collected, resulting in a smaller mean squared error of the ANN
training network. Simulation and experimental results confirmed that
this proposed conception is effective for working with PV systems.
Chaouachi et al. (2010), classified the real data of a PV system
installed in Tokyo, Japan, into three multi-layers based on a fuzzy rules-
based before they fed them into an ANN for offline training. The results
of this proposed method reveal that it achieves the highest efficiency
when compared to the conventional ANN and P&O algorithms under
different climatic conditions. With the same idea, Fathabadi (2013)
used the Lambert W function with a feed foreword ANN technique to
calculate the characteristics of silicon and plastic PV cells. The major
contribution of this method is that it enhances the performance of the
ANN model in predicting the PV curve.
Akkaya et al. (2007) used a GA to optimise the size of the hid-
den layer of an ANN model using an evolutionary hybrid algorithm,
which resulted in minimising the mean squared error of the ANN
training. The experimental results prove that the proposed method is
efficient for controlling PV systems under varying atmospheric condi-
tions. Similarly, Afsin and Akkaya (2012b) used GA to find the optimal
numbers of neuron nodes for a multi-layer neural network. The results
demonstrated that the proposed technique is valid in that comparison
between the practical and simulation results showed good agreement.
In the same vein, Zhang and Bai (2005) adapted a GA to find the
optimal number of a radial basis function ANN for modelling PV arrays.
The results proved that the proposed method can accurately predict
optimised PV power under various conditions. Hamdi et al. (2019) used
a particle swarm optimisation (PSO) algorithm instead of a GA one
to find the main parameters of the radial basic function of an ANN
network using a novel adaptive strategy. The results provided evidence
that the proposed method enhances the efficiency of MPPT tracking.
However, the selecting the effective training data and determining
the topology of ANN model whose report in previous proposals are
considered basic criteria to enhance their performance. Hence, Duman
et al. (2018), used a hybrid PSO gravitational search algorithm (GSA)
to calculate a suitable activation function of ANN layers, resulting
in the achievement of an accurate power prediction. Unfortunately,
this modification is classified as a complex method. There are novel
optimisations algorithms based on Grey Wolf and Bee Colony algorithm
that have been proposed in recent years, which have been used to
enhance the performance of AI techniques such as in D’Urso et al.
(2016), Precup et al. (2017) and Hu et al. (2018). In this work, PSO
is utilised to improve the accuracy of ANN model due to its sampling
technique and fast optimisation delivery.
3. ANN algorithm
An ANN technique is a distributed processing technique, which
is able to save experimental knowledge of application systems (Al-
dunainawi et al., 2017). It does not require a good knowledge when
modelling an application system, but it does need accurate data to
predict output functions as close to reality. This algorithm converts
the training data to a non-linear mapping between inputs and output
nodes. ANN topology is classified in to two types; feedforward and
feedback networks. The first type is the most commonly deployed due
to its usage of less memory in the implementation stage (Baghban et al.,
2015). Furthermore, it has proven to be highly powerful when working
with non-linear systems, such as a PV array. In general, a multilayer
feedforward ANN has three layers, input, hidden and output, as shown
in Fig. 2. Moreover, the neurons of each layer are connected through
the weights of the other neurons and bias terms in the antecedent
layers. This distributed processing system is defined mathematically by
Eq. (1):
∑
𝑛
𝑦= 𝑤𝑖𝑗 𝑥𝑗 + 𝑏𝑗 (1)
𝑖=1
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 1. General diagram of a stand-alone PV system using an ANN-MPPT method.
Fig. 2. Block diagram of an ANN system.
where, 𝑥𝑗 is the input training node, 𝑤𝑖𝑗 are the connection weights
associated with the input, hidden and layer nodes, 𝑏𝑗 is the bias of
the hidden and the output layer nodes and n is the number of input
signals. To learn the process of a feedforward ANN system, the back
propagation (BP) algorithm is usually used. It is a complex gradient
algorithm deployed for enhancing the performance of the ANN by
changing the weights of each node and the bias terms until the output
value at the output layer predicts the actual outputs as closely as
possible, thus resulting in a reduction in the training error. The mean
squared error (MSE) is usually chosen as the cost function, which is
given by Eq. (2):
1 ∑∑[
𝑛 𝑚
]∧
𝑀𝑆𝐸 = 𝑌 (𝑖) − 𝑇𝑗 (𝑖) 2
𝑛 𝑖=1 𝑗=1 𝑗
where, n is the number of input data, m is the number of output
signals, 𝑌𝑗 (i) is the real output, and 𝑇𝑗 (i) is the target output. There
are two main issues when designing a feedforward ANN: finding the
best topology of its structure (the number of hidden layers and units in
these); and optimising the initial weights of the training nodes.
3.1. Hidden layer size
Calculating the optimum number of hidden layers and units in each
is an important task for the feedforward ANN design, as this addresses
the issue between the computational time and best fitting regression of
the distribution of the ANN nodes (Basheer and Hajmeer, 2000; Akkaya
et al., 2007; Afsin and Akkaya, 2012b; Zhang and Bai, 2005; Hamdi
et al., 2019). If the ANN model has too many units in the hidden layers,
the computational time becomes too long, thus resulting in an over
Engineering Applications of Artificial Intelligence 92 (2020) 103688
Fig. 3. Equivalent circuit of a PV cell.
fitting regression. In contrast, if the ANN model has too few units in
hidden layers, then it will have low computational time when there
is a linear fitting regression. The most common method for finding the
hidden layer size is a trial and error technique. However, this technique
is inadequate as it requires a very long time.
3.2. Initial training weights
As aforementioned, the BP algorithm is used to learn the process
of the feedforward ANN model based on searching an error surface.
This processing search is changing as gradient descent regarding to the
incremental change in the weight (𝛥𝑊 ), As explained in Eq. (3);
𝑤𝑙𝑗𝑖 (𝑡) = 𝑤𝑙𝑗𝑖 (𝑡 − 1) + 𝛥𝑤𝑙𝑗𝑖 (𝑡) (3)
where 𝑤𝑙𝑗𝑖 (𝑡) and 𝑤𝑙𝑗𝑖 (𝑡 − 1) are the next and previous assumed training
weights, respectively. Each iteration in the BP algorithm has two steps:
a forward step to produce an updating solution; and a backward step to
compute and modify the MSE to new weights based on Eqs. (2) and (3).
(2)
This procedure continues until the optimal training weights of the ANN
model are determined. Many studies have pointed out that this method
will be failed to find the optimised training weights, because it mostly
depends on the size of 𝛥𝑊 (Basheer and Hajmeer, 2000; Aljarah et al.,
2018; Li et al., 1993). If 𝛥𝑊 is large, this can lead to accelerate training
and large fluctuating research on the error surface, thus resulting in
a non-converged optimising solution. Conversely, if 𝛥𝑊 is small, this
can lead to slower training and smooth fluctuating research on the
error surface, which could mean that the training process is stopped
before the global minimum error is found. Regarding this concept, the
assumed initial weights play a crucial role in designing an accurate
ANN model.
4. PSO algorithm
The PSO algorithm is considered a high-quality search tool in
engineering applications. The principle idea of this algorithm is that
3
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 4. The outdoor PV system installed at the Engineering campus of Brunel
University, London, UK.
it tries to find an optimised area, where each space has a degree of
possibility for a candidate solution (Vasumathi and Moorthi, 2012).
The movement of the PSO algorithm is inspired from the behaviour
of birds flocking, which depends on the individual and neighbouring
experiences of the PSO optimiser during each particle step. The proce-
dure of the PSO algorithm is divided into four steps. In the first, the PSO
optimiser starts the search within a random particle value. This particle
value is selected based on the degree of possibility of solution spaces
regarding several varying optimisations. In the second step, it compares
the previous and next best fitness values (𝑃𝑏𝑖 ) and (𝑃𝑙𝑖 ), respectively, to
search for optimised solutions in the same space. In the third step, the
best and global best positions (𝐺𝑏𝑖 ) are compared to select the global
fitness value. During this step, these positions are adapted and recorded
for the next step mathematically, as defined by Eqs. (4) and (5):
( ) ( )
𝑉𝑖𝑘+1 = 𝑤 × 𝑉𝑖𝑘 + 𝑟1 × 𝑐1 × 𝑃𝑏𝑖 − 𝑋𝑖𝑘 + 𝑟2 × 𝑐2 × 𝐺𝑏𝑖 − 𝑋𝑖𝑘 (4)
𝑋𝑖𝑘+1 = 𝑋𝑖𝑘 + 𝑉𝑖𝑘
where, 𝑋𝑖 is the current position of each particle, 𝑉𝑖 is the speed of the
search space, i is the optimisation vector, k is the number of iterations.
w represents the inertia weight factor of the speed, 𝑐1 is the cognitive
coefficient of the single particles, 𝑐2 is the social coefficient of all the
particles and 𝑟1 and 𝑟2 are the random velocity values of the search
space in the range 0 to maximum velocity. In the fourth step, the best
particle in terms of the fitness evaluation is determined and saved to
enhance the particle movement steps in each iteration. Those steps
continue to work until a stopping condition is achieved or the number
of iterations has ended. The stopping condition and the number of
iterations are proposed based on the required accuracy of the system
and control processing time.
Fig. 5. Power-Voltage characteristic of a PV module under: (a) varying values of G at a constant T of 25 ◦ C; (b) varying values of T at a constant G of 1000 W/m2 .
Engineering Applications of Artificial Intelligence 92 (2020) 103688
5. Photovoltaic system
A solar PV cell is an essential element in PV systems, the principal
work of which is converting the sunlight into electricity based on a
photon-voltage effect phenomenon. In an ideal solar cell, the series and
shunt resistances are not included. However, in practical case, they are
put in action and connected with the PV diode as shown in Fig. 3. This
is due to factors of magnitude resistance of the PV semiconductor and
non-optimum PN junction diode, resulting in, implementing a series
and shunt resistance, respectively.
Kirchhoff’s law, as given in Eq. (6), can provide the output current
from the solar PV cell:
𝐼𝑃 𝑉 = 𝐼𝐿 − 𝐼𝑑 − 𝐼𝑠ℎ (6)
where 𝐼𝐿 is the current generator diode, which is given in Eq. (7):
[ ( )]
𝐼𝐿 = 𝐺{𝐼𝑆𝐶 1 + 𝑘𝑎 𝑇 − 𝑇𝑆𝑇 𝐶 } (7)
where, 𝐼𝑆𝐶 is the short current of the PV cell circuit, ka is the tempera-
ture coefficient, 𝑇𝑆𝑇 𝐶 is the temperature operation of the PV cell circuit
under standard test conditions, and 𝐼𝑑 is the reviser current of the PV
diode circuit given by Shockley’s equation. (8):
{ ( ) }
𝑞𝑉𝑑
𝐼𝑑 = 𝐼0 exp −1 (8)
𝑛𝑘𝑇
where 𝐼𝑜 is the saturation current of the PV diode circuit, and 𝑉𝑑 is the
voltage across the PV diode, q is the electrical charge (1.69 × 10−19 C),
k is Boltzmann’s constant (1.38 × 10−23 J/K), and n is the standard PV
diode factor. A universal equation that describes the current–voltage
characteristic of the PV cell circuit is given by Eq. (9):
[ ( ( )) ] [ ]
𝑞 𝑉𝑃 𝑉 + 𝐼𝑅𝑆 𝑉 + 𝐼𝑅𝑆
𝐼𝑃 𝑉 = 𝐼𝐿 − 𝐼0 exp − 1 − 𝑃𝑉 (9)
𝑛𝑘𝑇 𝑅𝑠ℎ
where, 𝐼𝑃 𝑉 and 𝑉𝑃 𝑉 are the output PV current, and the output PV
voltage, respectively. Several PV cells are connected in series and
parallel topologies to obtain the required PV current and PV voltage,
(5) which is called a PV module, and then, several PV modules are con-
nected in various configurations to get a PV array. In this work, the
installed PV array consists of five 925 W modules (Sharp NU-S5E3E
185) of PV generation implemented in the Engineering Building of
Brunel University, Uxbridge, London, UK, as shown in Fig. 4. The
general parameters of this PV module are given in Table 1. Based on
Eqs. (6)–(9), the current–voltage (I–V) characteristics of the installed
PV array are simulated. As shown in Fig. 5, there are distinguished
points on the P-V curve of the PV array that recognised as the MPP.
In addition, the location of these points moves according to the values
of G and T. Moreover, the maximum power of PV array increases
when G increases, whilst a PV generator has high efficiency at a low T
than an increased one (Kadri et al., 2012). To address those concepts,
a conversion power system with an MPPT controller is implemented
4
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy Engineering Applications of Artificial Intelligence 92 (2020) 103688

Fig. 6. A block diagram of the system for collecting the measurement data.

between the PV array and a load or grid to get maximum continual

power generation. Consequently, its efficiency will be enhancing under
various atmospheric conditions.
Moreover, the fluctuation of the PV output power due the vari-
ation of weather conditions is considered a major issue for the grid
distribution of PV system (Shah et al., 2015). That is, the stability
of the PV power generation can be severely affected when weather
conditions are high variant, especially for a large-scale PV plant. Hence,
the proposed system will also improve the stability and reliability of
the PV generation when it connects to a grid (Fathabadi, 2016). A
DC–DC boost converter is a popular type of power conversion of PV
system, because it easily adapts the MPPT controller to the system. The
principle work of this converter is that it provides the voltage level at
an output side higher than at the input side. The main equation of the
DC–DC boost converter is given in Eq. (10): Fig. 7. Schematic diagram of the training methodology.

𝑉 1
𝐺𝑛 = 𝑜 =
𝑉𝑖 (1 − 𝐷)
where, 𝐺𝑛 is the amplifier voltage, 𝑉𝑜 and 𝑉𝑖 are the voltage level at the
input and output sides of the DC–DC converter, respectively, and D is
the duty cycle of the designed MPPT controller, which is converted to a
signal using PWM, as mentioned earlier. The value of D plays a crucial
role in adjusting the operating work of the PV power conversion system
at the MPP under various weather conditions.
To generate D, the proposed MPPT controller based on an ANN
model has been designed for this research. The G and T of atmospheric
conditions are used as the inputs of the proposed ANN model, whereas
the predicting power (Pref.) of PV array at the MPP as the output.
Under the same atmospheric conditions, the actual PV power (Pact.)
is measured based on a PV operating Simulink. Then, these powers are
compared, and the error is given to the PI controller to generate the D
of the DC–DC converter. The signal control of the PI controller is given
by Eq. (11):
( ) 𝐾
𝐷 = 𝐾𝑃 𝑃𝑎𝑐𝑡. − 𝑃𝑟𝑒𝑓 . + 𝐼 (𝑃𝑎𝑐𝑡. − 𝑃𝑟𝑒𝑓 . ) (11)
𝑆
(10)
where, 𝐾𝑃 and 𝐾𝐼 represent the proportional and integral gain of the
PI controller, respectively. To optimise the predicting function of the
ANN model, PSO is utilised. This optimisation is divided into two parts;
finding the best topology and then, determining initial weights of the
ANN model.
6. PV data acquisition system
As shown in Fig. 6, the PV array, atmospheric station and micro-grid
are installed at Brunel University London, Uxbridge, United Kingdom
for collecting the real training dataset. The PV array consists of five PV
modules connected inline and the array is connected to the low-voltage
(LV) distribution network using a Sunny Boy inverter. This inverter can
be adapted to the small-scale PV system and inbuilt communication
systems. On the other side, the atmospheric station includes, Pyra-
nometer, Hygrometer, Anemometer and Wind vane sensors, utilised for

5
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 8. Search history of the PSO–ANN algorithm.
recording the atmospheric parameters of the PV generation. The Sunny
Boy controller and the inverter are connected to the atmospheric station
using an RS485 cable. To read the electrical power measurements data
and the weather conditions, a PC was connected to the Sunny Boy
controller via an RS232 cable.
A supervisory control and data acquisition system (SCADA) was
used to monitor and control the system, linked to the university local
area network (LAN). The Sunny Boy controller read data every 5 min
periodically in daytime and switched off at night. It then turns on every
15 min to exam the weather conditions otherwise it returns to power-
save mode. The data were recorded on PC and stored as a Microsoft
Excel Sheet. To avoid failure in the commination system, an external
modem was installed with the system to send an alarm signal to the
system’s operator if any such problems occurred. For one year, about
48,500 data readings were collected and recorded comprising the G
level and T operation of the atmospheric station, with the output being
the measured power of the PV array at the MPP.
7. Proposed method
Using a MATLAB-Simulink model, an intelligent ANN-MPPT tech-
nique based on a large and real data set of a PV system is proposed.
The inputs of the ANN method consist of the G level and T operation
of climatic conditions, while the maximum power measurement of the
PV installed array at the MPP is the output. As mentioned earlier,
the accurate prediction of PV power using the ANN technique mostly
depends on the training strategy of ANN network. This starts with
Fig. 9. The best validation performance of (a) the conventional ANN and (b) the optimised ANN.
6
Engineering Applications of Artificial Intelligence 92 (2020) 103688
Fig. 10. A 3D surface of the output verse inputs of the ANN model.
selecting the topology of the ANN model and then, optimising its initial
weight values. To address this strategy, two algorithms are developed
in the form of a hybrid PSO–ANN technique. The parameters of the
PSO optimiser are set as having the following values: c1 = 1.49618,
c2 = 1.49618, w = 0.7298 and k = 50. In addition, the mean square
error is considered as the object function. A schematic diagram of the
training methodology of the PSO–ANN algorithm is presented in Fig. 7.
7.1. PSO–ANN algorithm — selected as the best topology of an ANN
network
In the first part of this modification, the PSO algorithm is employed
with the ANN model to find the best topology of the feedforward
ANN network. Hence, a hybrid algorithm is used to test progressively
increase the number of neurons in hidden layer without requiring the
user to preselect the number, which may be inaccurate. The main steps
of this are provided in Algorithm 1. The lower and upper boundaries
of the neuron numbers in this algorithm are 10 to 20. In this work, one
hidden layer with two inputs and one output of the neural network is
achieved with minimum training error, while the optimised number of
neurons in it is 18 neurons. This topology will be used in the second
modification to find the optimal initial weights of the ANN model.
7.2. PSO–ANN algorithm — determining the initial weights of the ANN
model
Once the topology of ANN network has been selected, a hybrid
algorithm based on the PSO and ANN method is designed to find the
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 11. PV array system for the proposed method versus the P&O, FLC and ANN methods under theoretical climate conditions: (a) power, (b) voltage, and (c) duty cycle.
optimised initial weights of the ANN model. These are determined to
improve the output prediction of the model when the assumed initial
weight values are correcting. To this end, the PSO algorithm is utilised
with the ANN technique. The main steps of the hybrid algorithm are
described in Algorithm 2. The lower and upper boundaries of the
weight values in this algorithm are - 0.9 to 0.9. As a result of running
this hybrid algorithm the optimised initial weights are obtained. The
search history of the algorithm is presented in Fig. 8. Then, the optimal
initial weights are used to train the ANN model using the ‘‘nntool’’
command of MATLAB. Next, the optimised initial weights are replaced
with the standard training weights in a field of the initial weights of the
7
Engineering Applications of Artificial Intelligence 92 (2020) 103688
‘‘nntool’’ box. Consequently, the performance of the ANN model based
on the optimised training strategy using the real data achieves better
prediction than using conventional ANN. This is because of lower MSE
and lesser number of epochs, about 0.00068 and 17 respectively, while
those of the non-optimal ANN are about 0.0079 and 68, respectively,
as shown in Fig. 9. The 3D surface of the optimised ANN model depicts
that the output power increases smoothly, with an increase in the
radiance level and with a decrease in the ambiance temperature, as
shown in Fig. 10. Table 2 presents the rudimentary statistical analysis
of the proposed algorithm. Notably, this proposal is simpler to design,
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy Engineering Applications of Artificial Intelligence 92 (2020) 103688

Fig. 12. Experimental measurement tests for a sunny and a cloudy day of: (a) irradiation; and (b) temperature.

because it does not need an extra unit in the implementation stage to Step Command
improve its accuracy.
1 Load data;
Algorithm 1. The main steps of the first proposed algorithm to find the 2 Set the testing and training samples,
optimised topology of the ANN 3 Select the number of neurons in the hidden layer
regarding to the first modification;
Step Command 4 Set the number of initial weights of the ANN;
1 Load data; 5 Set the main parameters of the PSO algorithm;
2 Set the number of the hidden layer; 6 Generate (for loop) to calculate the random weights of
3 Set the main parameters of the PSO algorithm the ANN using ‘‘netff’’
4 Initialise the random number for the neurons of the hidden 7 Generate (for loop) to determine the optimised initial
layer; weights after training the ANN using the
5 Generate (for loop) to calculate the MSE for each particle ‘‘neural_model’’ ‘‘findfitness’’, ‘‘train’’ and ‘‘net’’
using ‘‘netff’’, ‘‘train’’ and ‘‘net’’ commands and Eq. (2); commends.
6 Update the MSE value for each particle; 8 Calculate the MSE value based on Eq. (2) for each
7 Compare the 𝑃𝑙𝑖 with the 𝑃𝑏𝑖 of the PSO algorithm for each particle;
particle; 9 Update the MSE value for each particle;
8 Compare the 𝑃𝑏𝑖 with the 𝐺𝑝𝑖 of the PSO algorithm for each 10 Compare the 𝑃𝑙𝑖 with the 𝑃𝑏𝑖 of the PSO algorithm for
particle; each particle;
9 Update the velocity and position values of the PSO algorithm 11 Compare the 𝑃𝑏𝑖 with the 𝐺𝑝𝑖 of the PSO algorithm for
by Eqs. (4) and (5); each particle;
10 If the maximum iterations are reached or the stopping 12 Update the velocity and position values of the PSO
condition is achieved, print the result (the number of algorithm by Eqs. (4) and (5);
neurons of the hidden layer), otherwise return to step 5. 13 If the maximum iterations are reached or the stopping
condition is achieved, print the result (the optimal
initial weights of the ANN), otherwise return to step 7.
Algorithm 2. The major steps of the second proposed algorithm to find
the initial weights of the ANN.

8
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 13. MPPT predicting power on a sunny day using the proposed method versus: (a) the conventional P & O method; (b) conventional FLC method; and (c) the conventional
ANN method.
8. Results and discussion
To assess the performance, a MATLAB/Simulink model for the
studied PV system is designed for three popular methods, namely, con-
ventional P&O, FLC and ANN-MPPT as well as the proposed ANN-MPPT
method. This system consists of a PV array, DC/DC boost converter
and a load, whilst the array consists of five PV modules connected in
series. The main parameters of the PV array are given in Table 1. The
load resistance is used in this simulation with value 25 Ω. The solar G,
which used in this simulation, is rapidly decreased from 1000 to 200
W/m2 at 1 to 2 s and then, it is increased from 200 to 1000 W/m2
at 3 to 4 s, whilst the input T is kept constant at 25 ◦ C. As shown in
9
Engineering Applications of Artificial Intelligence 92 (2020) 103688
Fig. 11(a), the predictive power of the proposed ANN method is perfect
during varying atmospheric conditions owing to its optimised training
strategy. In addition, it is more robust in delivering the optimal MPP
under increasing and decreasing radiation. Whereas the conventional
methods for the ANN, FLC and P&O techniques are lost the optimal
right direction when the input G rapidly changed, as presented in
Fig. 11(b). Notably, although this phenomenon is very effective on the
conventional P&O-MPPT method by comparison with the conventional
ANN and FLC-MPPTs, it almost avoids the problem when the input G
decrease rapidly.
Another advantage of this proposal is that the convergence time
to reach the tracking power from the transient state into steady state
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Fig. 14. MPPT power prediction on a cloudy day using the proposed method versus: (a) the conventional P & O method, (b) conventional FLC method; and (c) the conventional
ANN method.
conditions is the fastest when compared with the conventional ANN,
FLC and P&O-MPPT methods, being about 0.06 s, 0.08 s, 0.11 s and
0.13 s, respectively, as shown in the zoomed in part of Fig. 11(a).
Moreover, the fluctuation around the MPP for the steady-state condi-
tions of the proposed ANN method is the lowest by comparison with
the other methods, thus resulting in less consumption time, as shown
in the zoomed in part of Fig. 11(a) and (c). Furthermore, it is more
precise at predicting the MPP under steady-state conditions, because
the optimal MPP duty cycle of the proposed method is more centred
than those of the conventional methods. Whilst the MPP duty cycle of
the conventional ANN and FLC-MPPT methods are to the left and right
10
Engineering Applications of Artificial Intelligence 92 (2020) 103688
of the optimal MPP duty cycle of the conventional P&O-MPPT method,
respectively. Consequently, the dissipated power in the proposed ANN-
MPPT method is the lowest when compared with the conventional
ANN, FLC and P&O-MPPT methods. Hence, the predicting power of
the proposed ANN-MPPT method along with those of the conventional
ANN, FLC and P&O-MPPT methods, after they have reached the MPP,
is about 924.60, 924.00 W, 923.25 W and 922.50 W, respectively, as
shown in the zoomed in part of Fig. 11(a).
Table 3 reports a comparative study covering the main properties
of the conventional ANN, FLC and P&O-MPPT methods as well as the
optimised ANN-MPPT. Regarding Table 3, the converging time and
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy Engineering Applications of Artificial Intelligence 92 (2020) 103688

Fig. 15. The average efficiency of the predicting power for the conventional P&O, FLC and ANN as well as the proposed method on a (a) sunny day and (b) cloudy day.

Table 1 Table 2
PV module properties. The rudimentary statistical analysis of the proposed algorithm.
Parameters Value Training ANN The number of MSE (Average ± STD)
no. architecture weights and
Cell number 48
biases
Dimensions 1.318 × 994× 46 mm
Nominal power 185 W 1 2 × 10 × 1 41 0.00795 ± 1.2 × 10−3
Open circuit voltage 30.2 V 2 2 × 12 × 1 49 0.00781 ± 1.8 × 10−4
Maximum power voltage 24 V 3 2 × 20 × 1 81 0.00678 ± 1.4 × 10−4
short circuit current 8.54 A 4 2 × 17 × 1 67 0.000769 ± 3.22 × 10−5
Maximum power current 7.71 A 5 2 × 18 × 1 73 0.000689 ± 1.03 × 10−5
Temperature Coefficient (Pmax) −0.485%/0 C
Temperature Coefficient (Isc) +0.053%/0 C
Temperature Coefficient (Voc) −104 mV/0 C Table 3
A comparative study covering the main properties of the conventional ANN, FLC and
P&O-MPPT method as well as the optimised ANN-MPPT.
MPPT Tracking time (s) Oscillation Power (W) Drift Problem
tracking power of the proposed method are the fastest and highest, Optimised ANN 0.06 Low 924.60 Avoidable
when compared with the conventional ANN, FLC and P&O methods. In Conventional ANN 0.08 Low 924.00 Suffering
addition, the fluctuation around the MPP is the least of all the methods. Conventional FLC 0.11 Medium 923.25 Suffering
Conventional P&O 0.13 High 922.50 Suffering
Moreover, the predicting power of the optimised ANN avoids the drift
problem under instability conditions.

9. Experimental measurement tests and a cloudy day on 5th July 2018 (06.00 am - 19.00 pm) and 30th
July 2018 (06.00 am −19.00 pm), respectively, have been used. The G
To assess the prediction of the proposed ANN method under dif- and T of those days are presented in Fig. 12(a) and (b), respectively.
ferent climatic conditions, experimental measurement tests of a sunny Then, these are applied to the studied PV system with the conventional

11
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Table 4
Comparative study regarding the hourly average efficiency for the proposed method
and the conventional ANN, FLC and P&O-MPPT techniques.
Atmospheric Average efficiencies
Proposed Conventional Conventional Conventional
method ANN FLC P&O
Sunny day 99.68% 99.17% 98.90% 98.18%
Cloudy day 99.30% 97.43% 94.69% 88.21%
P&O, ANN and FLC-MPPTs as well as the optimised ANN method, as
shown in Figs. 13 and 14. In the first case, the predictive power of the
conventional P&O method is almost good on the sunny day; however, it
drifts away from the right prediction on the cloudy day, as presented in
the zoomed in parts of Figs. 13(a) and 14(a). The issue becomes much
worse regarding the power of the conventional P&O method, when the
solar G of the cloudy day is rapidly changed, as shown in Fig. 14(a). In
the second case, when comparing the proposed ANN method with the
conventional FLC one, the predicting power of the latter method suffers
from the deviation on both the sunny and cloudy days, as shown in the
zoomed in part of Figs. 13(b) and 14(b). However, the problem can be
seen as having a minimum effect when compared with the conventional
P&O method.
In the third case, whilst the drift problem seems to have very little
effect on the power prediction of the conventional ANN method on
both days, it is not very precise at estimating the MPP, as shown in
the zoomed in part of Figs. 13(c) and 14(c). Notably, the prediction
function of the conventional ANN on the cloudy day has a higher
deviation when compared with that on the sunny day. In contrast,
the power prediction of the proposed ANN method is as close to the
reality during both sunny and cloudy days due to its optimised training
strategy. The average MPPT efficiency formula based on time samples
is used to determine the power predicting efficiency of the MPPT
methods, as given in Eq. (12) (Boukenoui et al., 2016):
∑𝑛
𝑃𝑜𝑢𝑡 (𝑘)
𝜂.𝑀𝑃 𝑃 𝑇 (𝑎𝑣𝑎𝑟𝑔𝑒) % = ∑𝑛𝑘=1 (12)
𝑘=1 𝑚𝑎𝑥 (𝑘)
𝑃 J. Supercrit. Fluids 98, 50–64. http://dx.doi.org/10.1016/j.supflu.2015.01.002.
where, 𝑃𝑜𝑢t is the output power of the PV system with the MPPT con-
troller, 𝑃𝑚𝑎𝑥 is the theoretical maximum power of the PV generation, k
is the sample time and n is the number of time samples. The actual and
theoretical output power are divided into 12 samples for each hour to
calculate the prediction efficiency of the MPPT methods under varying
atmospheric conditions. As presented in Fig. 15(a), and (b), as well as
Table 4, the hourly average efficiency of the proposed method achieved
the highest efficiency on both sunny and cloudy days by comparison
with the conventional ANN, FLC and P&O-MPPT techniques.
10. Conclusion
An optimised feedforward artificial neural network technique based
on the Particle Swarm Optimisation algorithm using real data has been
utilised for modelling a current–voltage characteristic and predicting
the maximum power point of photovoltaic arrays. This optimisation
was divided into two parts: selecting the best topology and then,
optimising the initial weights of the feedforward ANN model. Accord-
ingly, the problem between the computational time and best-fitting
regression of the distribution of ANN nodes was solved in the first part,
whilst the mean squared error of the ANN model was reduced in the
second optimisation. Consequently, the predicting power of the ANN-
MPPT controller has been improved under various weather conditions
when compared with the conventional ANN, FLC, and P&O methods.
In addition, the converging speed of the proposed method has been
enhanced under a transient state. As a result, hourly efficiencies of more
than 99.67% and 99.30% on sunny and cloudy days, respectively, have
been achieved. Moreover, this controller also improves the stability and
reliability of the PV generation when it connects to a grid. Finally,
12
Engineering Applications of Artificial Intelligence 92 (2020) 103688
the optimisations have been demonstrated to be simple to design.
Future studies should to use the PSO optimisation algorithm or another
optimiser to improve the accuracy of the ANN model, specifically in the
context of rapid change in atmospheric conditions.
Acknowledgements
The first author would like to thank the Iraqi Ministry of Higher
Education and Scientific Research for its financial support for his PhD
study at Brunel University London.
References
Afsin, A., Akkaya, R., 2012a. Training data optimization for ANNs using genetic
algorithms to enhance MPPT efficiency of a stand-alone PV system. Turk. J. Elec.
Eng. Comp. Sci. 20 (2), 241–254. http://dx.doi.org/10.3906/elk-1101-1051.
Afsin, A., Akkaya, R., 2012b. A genetic algorithm optimized ANN-based MPPT algorithm
for a stand-alone PV system with induction motor drive. Sol. Energy 86, 2366–2375.
http://dx.doi.org/10.1016/j.solener.2012.05.006.
Akkaya, R., Kulaksiz, A., Aydogdu, O., 2007. DSP Implementation of a PV system with
GA-MLP-NN based MPPT controller supplying BLDC motor drive. Energy Convers.
Manage. 48, 210–218. http://dx.doi.org/10.1016/j.enconman.2006.04.022.
Al-dunainawi, Y., Abbod, M.F., Jizany, A., 2017. A new MIMO ANFIS-PSO based
NARMA-L2 controller for nonlinear dynamic systems. Eng. Appl. Artif. Intell. 62
(March), 265–275. http://dx.doi.org/10.1016/j.engappai.2017.04.016.
Al-Majidi, S.D., Abbod, M.F., Al-Raweshidy, H.S., 2018a. A novel maximum power point
tracking technique based on fuzzy logic for photovoltaic systems. Int. J. Hydrogen
Energy 43 (31), 14158–14171. http://dx.doi.org/10.1016/j.ijhydene.2018.06.002.
Al-Majidi, S.D., Abbod, M.F., Al-Raweshidy, H.S., 2018b. A Modified P&O-MPPT based
on Pythagorean Theorem and CV-MPPT for PV Systems. In: Proceedings of the 2018
53rd International Universities Power Engineering Conference UPEC, Glasgow, UK,
pp. 1–6. http://dx.doi.org/10.1109/UPEC.2018.8542049.
Al-Majidi, S.D., Abbod, M.F., Al-raweshidy, H.S., 2019. Design of an efficient maximum
power point tracker based on ANFIS using an experimental photovoltaic. Electronics
8 (8), 858. http://dx.doi.org/10.3390/electronics8080858.
Aljarah, I., Faris, H., Mirjalili, S., 2018. Optimizing connection weights in neural
networks using the whale optimization algorithm. Soft Comput. 22 (1), 1–15.
http://dx.doi.org/10.1007/s00500-016-2442-1, Springer Berlin Heidelberg.
Baghban, A., Ahmadi, M.A., Shahraki, B.H., 2015. Prediction carbon dioxide solubility
in presence of various ionic liquids using computational intelligence approaches.
Bahgat, A.B.G., Helwa, N.H., Ahmad, G.E., El-Shenawy, E.T., 2005. Maximum power
point traking controller for PV systems using neural networks. Renew. Energy 30,
1257–1268. http://dx.doi.org/10.1016/j.renene.2004.09.011.
Bakos, G.C., Soursos, M., 2002. Technical feasibility and economic viability of a grid-
connected PV installation for low cost electricity production. Energy Build. 34,
753–758. http://dx.doi.org/10.1016/S0378-7788(01)00142-6.
Basheer, I.A., Hajmeer, M., 2000. Artificial neural networks: fundamentals, computing,
design, and application. J. Microbiol. Methods 43, 3–31. http://dx.doi.org/10.
1016/S0167-7012(00)00201-3.
Boukenoui, R., Salhi, H., Bradai, R., Mellit, A., 2016. A new intelligent MPPT method
for stand-alone photovoltaic systems operating under fast transient variations of
shading patterns. Sol. Energy 124, 124–142. http://dx.doi.org/10.1016/j.solener.
2015.11.023.
Chaouachi, A., Kamel, R.M., Nagasaka, K., 2010. A novel multi-model neuro-fuzzy-
based MPPT for three-phase grid-connected photovoltaic system. Sol. Energy 84
(12), 2219–2229. http://dx.doi.org/10.1016/j.solener.2010.08.004.
Duman, S., Yorukeren, N., Altas, I.H., 2018. A novel MPPT algorithm based on
optimized artificial neural network by using FPSOGSA for standalone photovoltaic
energy systems. Neural Comput. Appl. 29 (1), 257–278. http://dx.doi.org/10.1007/
s00521-016-2447-9.
D’Urso, P., De Giovanni, L., Massari, R., 2016. GARCH-based robust clustering of time
series. Fuzzy Sets and Systems 305, 1–28. http://dx.doi.org/10.1016/j.fss.2016.01.
010.
Enany, M.A., Farahat, M.A., Nasr, A., 2016. Modelling and evaluation of main maximum
power point tracking algorithms for photovoltaics systems. Renew. Sustain. Energy
Rev. 58, 1578–1586. http://dx.doi.org/10.1016/j.rser.2015.12.356.
Fathabadi, H., 2013. Novel neural-analytical method for determining silicon/plastic
solar cells and modules characteristics. Energy Convers. Manag. 76, 253–259.
http://dx.doi.org/10.1016/j.solener.2015.11.047.
Fathabadi, H., 2016. Novel high efficiency DC/DC boost converter for using in pho-
tovoltaic systems. Sol. Energy 125, 22–31. http://dx.doi.org/10.1016/j.enconman.
2013.07.055.
Guenounou, O., Dahhou, B., Chabour, F., 2014. Adaptive fuzzy controller based MPPT
for photovoltaic systems. Energy Convers. Manag. 78, 843–850. http://dx.doi.org/
10.1016/j.enconman.2013.07.093.
S.D. Al-Majidi, M.F. Abbod and H.S. Al-Raweshidy
Hamdi, H., Regaya, C. Ben, Zaafouri, A., 2019. Real-time study of a photovoltaic system
with boost converter using the PSO-RBF neural network algorithms in a MyRio
controller. Sol. Energy 183 (February), 1–16. http://dx.doi.org/10.1016/j.solener.
2019.02.064.
Hiyama, T., Ieee, S.M., Kitabayashi, K., 1997. Neural network based estimation of max-
imum power generation from PV module using environmental information. IEEE
Trans. Energy Convers. 12 (3), 241–247. http://dx.doi.org/10.1109/60.629709.
Hu, W., Yu, Y., Gu, W., 2018. Parameter estimation of fractional-order arbitrary
dimensional hyperchaotic systems via a hybrid adaptive artificial bee colony
algorithm with simulated annealing algorithm. Eng. Appl. Artif. Intell. 68 (August
2016), 172–191. http://dx.doi.org/10.1016/j.engappai.2017.10.002.
Kadri, R., Andrei, H., Gaubert, J., Ivanovici, T., Champenois, G., Andrei, P., 2012.
Modelling of the photovoltaic cell circuit parameters for optimum connection model
and real-time emulator with partial shadow conditions. Energy 42 (1), 57–67.
http://dx.doi.org/10.1016/j.energy.2011.10.018.
Kamala, V., Premkumar, K., Bisharathu, A., Ramaiyer, S., 2017. A modified Perturb
& Observe MPPT technique to tackle steady state and rapidly varying atmo-
spheric conditions. Sol. Energy 157 (August), 419–426. http://dx.doi.org/10.1016/
j.solener.2017.08.059.
Kwong, C.K., Huimin, J., Luo, X.G., 2016. AI-based methodology of integrating affective
design, engineering, and marketing for defining design specifications of new
products. Eng. Appl. Artif. Intell. 47, 49–60. http://dx.doi.org/10.1016/j.engappai.
2015.04.001.
Li, G., Alnuweiri, H., Wu, Y., 1993. Acceleration of back propagations through initial
weight pre-training with delta rule training a BP network. In: IEEE International
Conference on Neural Networks. pp. 580–585. http://dx.doi.org/10.1109/ICNN.
1993.298622.
Li, X., Wen, H., Hu, Y., Jiang, L., 2019. A novel beta parameter based fuzzy-logic
controller for photovoltaic MPPT application. Renew. Energy 130, 416–427. http:
//dx.doi.org/10.1016/j.renene.2018.06.071.
Livieris, I.E., 2019. Improving the classification efficiency of an ANN utilizing a
new training methodology. Informatics 6 (1), 1–17. http://dx.doi.org/10.3390/
informatics6010001.
13
Engineering Applications of Artificial Intelligence 92 (2020) 103688
Mellit, A., Kalogirou, S.A., 2008. Artificial intelligence techniques for photovoltaic
applications: A review. Prog. Energy Combust. Sci. 34, 574–632. http://dx.doi.org/
10.1016/j.pecs.2008.01.001.
Messalti, S., Harrag, A., Loukriz, A., 2017. A new variable step size neural networks
MPPT controller: Review, simulation and hardware implementation. Renew. Sus-
tain. Energy Rev. 68 (October 2016), 221–233. http://dx.doi.org/10.1016/j.rser.
2016.09.131.
Nabipour, M., Razaz, M., Seifossadat, S.G.H., Mortazavi, S.S., 2016. A novel adaptive
fuzzy membership function tuning algorithm for robust control of a PV-based
Dynamic Voltage Restorer (DVR). Eng. Appl. Artif. Intell. 53, 155–175. http:
//dx.doi.org/10.1016/j.engappai.2016.04.007.
Plumb, A.P., Rowe, R.C., York, P., Brown, M., 2005. Optimisation of the predictive
ability of artificial neural network (ANN) models: A comparison of three ANN
programs and four classes of training algorithm. Eur. J. Pharm. Sci. 25, 395–405.
http://dx.doi.org/10.1016/j.ejps.2005.04.010.
Precup, R.E., David, R.C., Petriu, E.M., 2017. Grey Wolf optimizer algorithm-based
tuning of fuzzy control systems with reduced parametric sensitivity. IEEE Trans.
Ind. Electron. 64 (1), 527–534. http://dx.doi.org/10.1109/TIE.2016.2607698.
Román, E., Alonso, R., Ibañez, P., Elorduizapatarietxe, S., Goitia, D., 2006. Intelligent
PV module for grid-connected PV systems. IEEE Trans. Ind. Electron. 53 (4),
1066–1073. http://dx.doi.org/10.1109/TIE.2006.878327.
Salah, C. Ben, Ouali, M., 2011. Comparison of fuzzy logic and neural network in
maximum power point tracker for PV systems. Electr. Power Syst. Res. 81 (1),
43–50. http://dx.doi.org/10.1016/j.epsr.2010.07.005.
Shah, R., Mithulananthan, N., et al., 2015. A review of key power system stability chal-
lenges for large-scale PV integration. Renew. Sustain. Energy Rev. 41, 1423–1436.
http://dx.doi.org/10.1016/j.rser.2014.09.027.
Vasumathi, B., Moorthi, S., 2012. Implementation of hybrid ANN – PSO algorithm
on FPGA for harmonic estimation. Eng. Appl. Artif. Intell. 25 (3), 476–483.
http://dx.doi.org/10.1016/j.engappai.2011.12.005.
Zhang, L.Ã., Bai, Y.F., 2005. Genetic algorithm-trained radial basis function neural
networks for modelling photovoltaic panels. Eng. Appl. Artif. Intell. 833–844.
http://dx.doi.org/10.1016/j.engappai.2005.02.004.