Received: 30 August 2018 

|   Revised: 15 February 2019 

|  Accepted: 18 February 2019

DOI: 10.1002/ese3.312

RESEARCH ARTICLE

Optimization methods using artificial intelligence algorithms to

estimate thermal efficiency of PV/T system

Mohammad Zamen1  | Alireza Baghban2  | S. Mohsen Pourkiaei3  |

1
Mohammad Hossein Ahmadi

1
Faculty of Mechanical and Mechatronics
engineering, Shahrood University of
Abstract
Technology, Shahrood, Iran Renewable energies, specifically solar energy has been employed in numerous ap-
2
Department of Chemical Engineering, plications while being CO2 emission free energy in comparison with fossil fuel re-
Amirkabir University of Technology
sources. The main purpose of this study is to predict thermal efficiency of
(Tehran Polytechnic), Mahshahr, Iran
3 photovoltaic-­thermal (PV/T) setups in regard with input temperature, recirculation
Faculty of New Sciences and
Technologies, Department of Renewable flow rate, and solar irradiation by modifying multilayer perceptron artificial neural
Energies, University of Tehran, Tehran, Iran network (MLP-­ANN), adaptive neuro-­fuzzy inference system (ANFIS), and least
Correspondence
squares support vector machine (LSSVM) approaches. For this goal, more than 100
Mohammad Hossein Ahmadi, Faculty of empirical measurements were performed on a fabricated water-­cooled PV/T setup.
Mechanical and Mechatronics engineering, Several numerical analyses are also carried out to assess the validity of the presented
Shahrood University of Technology,
Shahrood, Iran. models. It is confirmed that there is a great agreement between predictive models and
Emails: [email protected], actual data. The proposed ANN model provided the best performance due to the
[email protected]
mean squared error (MSE) and determination coefficient (R2) values of 0.009 and
1.00, respectively. Also, numerical comparisons with other recently developed
­models were performed.

KEYWORDS
adaptive neuro-fuzzy inference system, Intelligent models, least squares support vector machine,
optimization, MLP-ANN, photovoltaic/Thermal, thermal efficiency

1  |   IN T RO D U C T ION for thermal applications by solar thermal collectors.5 However,

the state of combination of photovoltaic cells and thermal solar
Solar energy is the unique type of energy with regard to quantity together leads to a variety of different types of PVT systems
and environmental points of view.1 In the past decades, the nu- such as: Air or water-­cooled PV/T, glazed and unglazed panels,
merous solar equipment have been designed with novelties and PVT with natural or forced circulation flow, etc. As a result,
different characteristics.1,2 Solar thermal and photovoltaics are various solar thermal collector types have been investigated and
the main approaches for solar energy utilization.3 Photovoltaic-­ examined in the past decade.6-8 PV/T efficiencies are being im-
thermal (PV/T) is one of the most leading solar thermal technolo- proved steadily, whereas the overall performance relies on sys-
gies, which provides thermal and electric power simultaneously. tem configuration and photovoltaic internal-­external elements.9
Solar irradiation has been employed by thermal and photovoltaic Several reviews presented latest developments and technologies
arrangements.4 A common photovoltaic cell has an efficiency of of flat plate PV/Ts.10-13 Also, the prospect and future challenges
4%-­17%, while efficiencies higher than 50% can be obtained of PV/T systems are reviewed as well.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original
work is properly cited.
© 2019 The Authors. Energy Science & Engineering published by the Society of Chemical Industry and John Wiley & Sons Ltd.

Energy Sci Eng. 2019;7:821–834.  |

wileyonlinelibrary.com/journal/ese3     821
 |
822      ZAMEN et al.

Böer as a leading researcher has built a “Solar One” build- Presently, researchers has focused a great attention on the
ing in 1973-­1974.14 The aim of air PV/T systems is to maxi- employment of new computational approaches for detecting
mize the received heat and keep the cell temperature low by the optimum state of the energy systems.35-37
the air cooling application. Later in 80s a Hendrie et al,15 in- Argiriou and colleagues presented a neural controller
vestigated PV/T air heating systems numerically. An unglazed model for residential hydronic heating plants.38 The results
PV/T system was fabricated and studied at the University of of experiments showed that the energy could be saved up
Patras.16 Subsequently, various elements affecting the perfor- to 15%. Khatib and colleagues designed a solar irradiation
mance of the system were investigated. A PV/T setup was system-­based ANN.39 The results indicated that the model
constructed at Politecnico di Milano,17 to analyze the opera- is capable to predict solar irradiation by a mean absolute
tional parameters such as the air flow, collector tilt, and the percentage error (MAPE) of 5.92%, and root mean squared
air gap. Elsafi et al investigated a compound parabolic con- error (RMSE) of 7.96%. Dorvloa and colleagues designed a
centrated (CPC) system experimentally.18 Additionally, CPC hybrid ANN system based on Radial Basis Function (RBF)
equipped by fins have been modeled and built.19 An analysis and MLP.40 The results indicated that RBF configurations are
for single and double fan arrangement were performed, as more efficient owing to their lower process power demand,
well. Tonui et al investigated the thin flat metallic sheet setup against the MLP. Sulaiman and colleagues designed a hy-
for the collector system.20 Many studies have performed with brid multilayer feed forward neural network (HMLFNN) to
the aim of optimum operational conditions analysis of collec- predict the power of an on-­grid PV system.41 The designed
tor and PV/T systems. Some studies have focused on PV/T model employed an artificial immune system to optimize the
designs with different radiation area's and different length to training procedure as well.
width ratios.21,22 A dual channel hybrid PV/T collector was Mellit et al employed an ANFIS utilizing a back propaga-
studied numerically.23 Various air PV/T designs have been tion learning pattern to estimate the daily PV power system
investigated mathematically by Amori et al in a comparative performance.42 Izgi and colleagues designed an ANN model
study.24 Kumar et al evaluated the air type PV/T systems for a 750 W PV system.43 Also, the designed model was able
equipped by air cooling channels.4 to predict hours of high productivity, as well.
The water based PV/T setups are the most well-­known Based on the experimental data for fabricated serpentine
arrangement among others,6 whereas the air based models PV/T system prediction of thermal performance of the sys-
are easier to utilize and have more benefits. Absorbed heat tem by the mean of multilayer perceptron ANN (MLP-­ANN),
by the collectors could be used for household hot water/air ANFIS, and LSSVM methods is the main objective of this
consumption.25 Applying internal implements in PV/T sys- study.
tems have been investigated in order to enhance the overall
efficiency.26-28
From the geometric point of view, there are various types
2  |  THEORY
of thermal absorber. Allan et al investigated flat plate PVT
2.1  |  Multilayer perceptron ANN
systems with 4 absorbing plates: Parallel, Serpentine, Header-­
(MLP-­ANN)
Riser, and Bionic.29 They found that the serpentine cooling
is more efficient than Header-­Riser. The overall efficiency Artificial neural network is a learning algorithm inspired
of Serpentine is about 62% and that of PVT system using by natural neural networks of the brain, as a group of inter-
Header-­ riser is about 59%. Charalambous et al compared linked nodes skilled to process exchanged messages within
PVT system using serpentine absorber and header-­riser in the the network. Generally, each node generates an output by
same conditions. It was reported that the thermal efficiency the mean of a nonlinear summation of inputs. Also, the pa-
of the PVT was 4% more in serpentine that in header-­riser rameter of “weight” represents the intensity of the signal
prototype.30 Recently, several PV/T collector designs have changes at an interface.44 Three major activation functions
been presented due to their significant advantages.2 Kramer are as follows:
et al have performed an important examination to evaluate
Linear: f (x) = x (1)
PV/T market progress and PV power plants cost analysis.31
On the other hand, the artificial intelligence approaches
could be expressed as proper cost reduction alternatives.32,33 1
The predictive models are capable to provide output for un- Sigmoid: f (x) = (2)
1 + e−x
trained input elements. An ANN model is applied by Caner
et al to provide a significant estimation of solar air collectors ex − e−x
Hyperbolic tangent: f (x) = (3)
efficiency.34 Also, a predictive model has been presented by ex + e−x
Varol and colleagues,28 to predict and evaluate PCM perfor- Furthermore, the bias parameter is also determined as a key
mance in a solar air collectors. element besides the weight parameter. A MLP-­ANN is a
ZAMEN et al.
|
     823

O2i = wi = 𝛽Ai (X)𝛽Bi (Y) (9)

The 3rd layer normalizes the obtained firing strength

elements:
w
O3i = w̄ i = ∑ i (10)
i wi

The 4th layer processes the linguistic expressions of out-

puts as follows:
F I G U R E   1   The photovoltaic-­thermal (PVT) plate after
O4i = w̄ i fi = w̄ i (mi X1 + n1 X2 + ri ) (11)
connecting half-­pipe
In the final step, all the output rules are applied together
feedforward type network with nonlinear activation function. as follows:
It is noteworthy that the training approach in MLP-­ANN is ∑
backpropagation algorithm.45,46 � wi fi
5
Oi = w̄ i fi = ∑i (12)
i i wi

2.2  |  Adaptive Neuro-­Fuzzy Inference 2.3  |  Least squares support vector machine
System (ANFIS) (LSSVM)
The idea of fuzzy logic was presented by Zade for the first Support vector machine are controlled learning approaches
time.47 Applying fuzzy logic basics leads to more simple with correlated learning methods that evaluate data sets for
advances and exact result. Employing fuzzy logic and ANN categorizing or regression study. In SVM method, the func-
approaches simultaneously provides online learning capabil- tion is explained as53,54:
ity of the ANN and flexibilities of fuzzy logic together.48,49
Employing the fuzzy logic and ANN at the same time, forms f (x) = wT (x)𝜑(x) + b (13)
the foundation of ANFIS. Fuzzy inference systems have 2
main arrangements: (a) Mamdani, and (b) Takagi-­Sugeno.50-52
Mamdani type, applies logical description for the progress of
fuzzy if-­then limits, while; Takagi-­Sugeno type produces the
if-­then constrains. The presented if-then laws are used in a
general ANFIS arrangement containing 2 input elements:
IfX1 is A1 and X2 is B1 then f1 = m1 X1 + n1 X2 + r1 (4)

IfX1 is A2 and X2 is B2 then f2 = m2 X1 + n2 X2 + r2 (5)

If X1 is A1 and X2 is B2 then f3 = m3 X1 + n3 X2 + r3 (6)

If X1 is A2 and X2 is B1 thenf4 = m4 X1 + n4 X2 + r4 (7)

which Ai and Bi (i = 1,2) are fuzzy collections for X1 and X2,

and f represents the output as well.
Normally, ANFIS configuration contains 5 layers.
Equation (8) represent the corresponding Gaussian member-
ship function:
[ ]
1 −1 (X − Z)
Oi = 𝛽(X) = exp (8)
2 𝜎2
The Gaussian membership function elements will be set
at their optimum values to provide the most exact answers.
The 2nd layer defines the statements’ reliability by the
mean of firing strength elements:
F I G U R E   2   Photovoltaic-­thermal (PVT) system during test
824     
| ZAMEN et al.

1150

1100

1050
Solar irradiance (W/m2)

1000

950

900

850

800
10:50
11:00
11:00
11:10
11:10
11:20
11:20
11:30
11:30
11:40
11:40
11:50
11:50
12:00
12:00
12:00
12:10
12:15
12:20
12:30
12:30
12:30
12:40
12:40
12:50
12:50
13:00
13:00
13:10
13:20
13:30
F I G U R E   3   Data set for solar
Time irradiance

100

90

80

70
Thermal efficiency (%)

60

50

40

30

20

10

0
0 200 400 600 800 1000 1200 1400
Solar radiation (Watt)
F I G U R E   4   Bubble curve of investigated parameters

By applying the following cost function, values w and b val- ⎧ y − wT 𝜑(x ) − b ≤ 𝜀 + 𝜉 ,k = 1,2, … ,N
ues are obtained: ⎪ kT k k
∗
⎨ W 𝜑(xk ) + b − yk ≤ 𝜀 + 𝜉 k
,k = 1,2, … ,N
(15)
∑N ⎪ 𝜉k ,𝜉 ≥ 0
∗
1
Cost function = W T + c ⎩ k
2
(𝜉k − 𝜉k∗ ) (14)
k=1
The lowest amounts of the function are related to the most Suykenes and Vandewalle employed the least squares mod-
exact answers. The aforementioned function is corresponding ification of the SVM approach in order to shorten the solution
to the following constraints as well: calculations.55,56 They presented a novel cost function as follows:
ZAMEN et al.
|
     825

F I G U R E   5   Parallel diagram of different variables used in the present experiment

( )
1 1 ∑ 2
N ( ) ||xk − x||2
Cost function = W T W + 𝛾 e (16) k x,xk = exp − (20)
2 2 k=1 k 𝜎2
We have:
The variance value should be tuned in radial basis func-
Yk = W T 𝜑(xk ) + b + ek (17) tion. Consequently, optimum values of γ and σ2 should be
The Lagrangian of this problem is determined as: determined in order to maximize the prediction accuracy:

1 ∑ ( exp. )2
N
1 ∑ 2 ∑
N N
1 Hi − Hical.
L(w,b.e.a) = wT w + 𝛾
2 2 k=1
ek − ak (wT 𝜑(xk ) + b + ek − yk ) (18) MSE =
N i=1 (21)
k=1

3  |   EXPERIM ENTAL PROCEDU R E

In order to specify the optimum point of the problem, the
AND DATA PREPARATION
saddle point of the Lagrangian should be applied:
∑ Generally, PVT systems include 2 main units: solar cells unit
⎧ 𝜕L = 0 ⇒ w = Nk=1 ak 𝜑(xk )
⎪ 𝜕L
𝜕w
∑N and thermal units. A 90W polycrystalline silicon solar panel
⎪ 𝜕b = 0 ⇒ k=1 ak = 0 with 0.73 m2 area is used in this study. An aluminum sheet
⎨ 𝜕L = 0 ⇒ a = 𝛾 ,k = 1,2, … ,N (19)
⎪ 𝜕ek k ek with 3 mm thickness was used as absorber plate that a ser-
⎪ 𝜕L = 0 ⇒ w 𝜑(x ) + b + e y = 0,k = 1,2, … ,N pentine tube is connected to this plate (Figure  1).
⎩ 𝜕ak T k k k
The panel then connected to a water tank. A tank with a
Solving the aforementioned equations will determine the capacity of 100 liters is utilized to handover the heat from
LSSVM characteristics. Along with γ, kernel function ele- PVT setup. Also, an electric water pump is employed for
ments play role as tuning elements as well. The radial basis water circulation. Figure 2 illustrates the installed setup
function is defined as follows: during the experiments of this study.
 |
826      ZAMEN et al.

F I G U R E   6   Scatter matrix plot for all measured variables

A flowmeter, Pyranometer, solar system power analyzer and Figure 3 shows the variation in the solar irradiance data that
3 thermometers were used as measurement devices for data col- collected during test days.
lection. The water temperatures were meseared at 3 points of One of the main parameter that evaluated was the water
input, output and midpoint of the tank. The circulation flow rate mass flow rate that changed from 0.5 up to 4 lit/min and
is regulated by the flowmeter. The solar irradiation is logged by other parameter such as the inlet and outlet water tem-
the Pyranometer and operational conditions of photovoltaics, peratures, solar irradiation, electrical specifications of the
including maximum power, short-­circuit current, open circuit module such as I-­V curve, maximum voltage, current, and
voltage, efficiency, etc. are recorded by solar power analyzer. power are recorded. Also the effect of inlet temperature of
The tests were done in summer days and At near noon, the water is tested. This temperature vared in an interval
when the radiation is almost constant and in maximum value. between 20 to 45°C.
ZAMEN et al.
|
     827

F I G U R E   7   Construction of suggested artificial neural network (ANN) model

1.8

1.6

1.4

1.2

1
RMSE

0.8

0.6

0.4

0.2

0
0 200 400 600 800 1000
Iteration

F I G U R E   8   Artificial neural network (ANN) performance during F I G U R E   1 0   Root mean squared error (RMSE) values for each
different iterations iteration

Also, the diagram shown in Figure 5 represents the

investigated parameters variations at different thermal
efficiency values. As can be seen, the flowrate ranges
between 0.5 and 4 lit/min, the inlet temperature ranges
between 21.5 and 46.8°C, solar radiation ranges between
1162 and 1152 watt, and thermal efficiency ranges be-
tween 18.56 and 84.4.
Also, the diagram shown in Figure 6 represents the Scatter
matrix plot for all measured variables. The scatter diagram
depicts the relation of 2 objectives, which is named correla-
tion. It should be noted that the closer the data points become,
F I G U R E   9   Construction of typical adaptive neuro-­fuzzy
inference system (ANFIS)
the greater coherence between variables exists. If the straight
line of data points goes from the high values in the y axis to
Bubble curve of the thermal efficiency results in different the high values in the x-­axis, then the objects are in a nega-
solar radiation and inlet water temperatures is shown in Figure  4. tive correlation. As it is shown in Figure 6, due to scattering
As shown in this figure, the maximum thermal efficiency of behavior of variables, the correlations between variables is
80% has been achieved in lowest inlet water temperature. feeble.
 |
828      ZAMEN et al.

models, whereas the testing dataset validates the results. In

order to homogenize data sets, all operational data points
were normalized in range of [−1,+1] by the Eq. (22):
D − Dmin
Dn = 2 −1 (22)
Dmax − Dmin

Thermal efficiency is the output parameter while the other

parameters such as inlet temperature, flow rate, and solar ra-
diation) are input parameters of the presented models.

4.2  |  Model development

4.2.1  |  MLP-­ANN
The following equation is used in this investigation for output
data set of the designed MLP-­ANN:
n ( )
∑ 1
Z= W3i + b3 (23)
i=1 1 + e−(xi Wi )
ANN arrangement is also trained to define optimized out-
put factors. This goal is attained by regulating weight and bias
­elements. Error function is presented as follows:
∑∑( j j,l
)
E= r i −o i
j i (24)

Optimization procedure is carried out through the

Levenberg-­Marquardt technique, as well.
The configuration of MLP-­ANN is illustrated in Figure 7.
Performance of the designed network is presented in
Figure  8 on the subject of the MSE calculated for predicted
values of the predictive MLP-­ANN model.

4.2.2  | ANFIS
Schematic structure of a typical ANFIS with double inputs is
presented in Figure  9.
Particle swarm optimization (PSO) technique is carried
F I G U R E   1 1   Trained membership function for input elements out to train the designed ANFIS model. Total ANFIS param-
eters number is determined as follows:
4  |  M O D E L S IMP L E ME N TAT I ON
NT = Nc .Nv .NMF (25)
4.1  |  Preprocessing procedure
The Gaussian membership function which is employed in
Three artificial intelligence techniques including MLP-­ this study contains 2 membership function elements of Z and
ANN, ANFIS, and LSSVM are used in this investigation σ2. Thermal efficiency, inlet temperature, flow rate, and solar
in a platform of MATLAB 2016 software, to specify the radiation, are implemented variables, and primary number of
thermal efficiency as a function of inlet temperature, water clusters is set to 5. Consequently, amount of 40 ANFIS ele-
flow rate, and solar irradiation. The data set of 100 opera- ments is provided. RMSE among experimental and predicted
tional points were used in order to develop the predictive values is taken into account as the cost function in PSO pro-
models. cess used in determination of the optimal ANFIS elements.
The total dataset consist of 2 subgroups: Train and Test Figure  10 depicts the RMSE values for each iteration.
(25% test, 75% train). The training dataset is used for de- Figure 11 shows the trained membership function for
termination of the corresponding elements in the proposed input elements as well.
ZAMEN et al.
|
     829

F I G U R E   1 2   Least squares support vector machine (LSSVM)-­GA chart

T A B L E   1   Comprehensive information of the 3 designed models

LSSVM Value/comment ANFIS Value/comment ANN Value/comment

Kernel function RBF Membership Function Gaussian Input layer 3
γ 161098.776177 MF parameters 40 Hidden layer 5
σ2 582.561403 Clusters 5 Output layer 1
Training data 75 Training data 75 Hidden layer activation function Logsig
Testing data 25 Testing data 25 Output layer activation function Purelin
Population 100 Population 50 Method Levenberg-­
Marquardt
Iteration 1000 Iteration 1000 Training data 75
C1 1 C1 1 Testing data 25
C2 2 C2 2 Max iterations 1000

4.2.3  | LSSVM
1 ∑ ( exp. )2
N
MSE = Hi − Hical. (26)
Least squares support vector machine also uses 2 N i=1
adaptable elements in its arrangement. These param-
­
eters are the r­egulation parameter (γ), and the kernel pa- N | exp. cal. |
rameter (σ2 from the radial basis function formulation). 100 ∑ ||Hi − Hi || (27)
ARD (%) = exp.
LSSVM method also uses the genetic algorithm (GA) to N i=1 Hi
provide the optimized elements as it is depicted in Figure 12.
( )0.5
1 ∑ ( exp. )2
N

4.3  |  Models’ evaluation STD =

N − 1 i=1
Hi − Hical. (28)

Different statistical indexes (ie MSE, percentage of aver-

age relative deviation (ARD%), standard deviation (STD),
RMSE, and coefficient of determination (R2)) are used in the ( )0.5
1 ∑ ( exp. )2
N
evaluation of designed models. Following determinations are RMSE = Hi − Hical. (29)
N i=1
presented for the aforementioned indexes:
830     
| ZAMEN et al.

90 90
Train Exp. Test Exp. Train LSSVM Test LSSVM (A)
80 80 Train: y = 1.0003x –0.0146, R² = 0.9996

Actual heat transfer coeficient

Test: y = 0.9895x + 0.5823, R² = 0.9998
70 70
Heat transfer coeficient

60 60

50 50

40 40

30 Train
30
Test
20 20 Best line (train)
Best line (test)
10 10

0 0
0 20 40 60 80 100 120 0 10 20 30 40 50 60 70 80 90
Data index Predicted heat transfer coefficient
90 90
Train Exp. Test Exp. Train ANFIS Test ANFIS
80 (B)
80 Train: y = 0.9986x + 0.0718, R² = 0.9992

Actual heat transfer coeficient

70 Test: y = 1.0122x –0.5863, R² = 0.9995
Heat transfer coeficient

70
60 60
50 50
40 40
Train
30 30
Test
20 20 Best line (train)
Best line (test)
10 10
0 0
0 20 40 60 80 100 120 0 10 20 30 40 50 60 70 80 90
Data index Predicted heat transfer coefficient
90 90
Train Exp. Test Exp. Train ANN Test ANN
80 80 (C) Train: y = 1x -3E-05, R² = 1
Actual heat transfer coeficient

70 Test: y = 0.9971x + 0.2213, R² = 0.9999

Heat transfer coeficient

70
60 60
50 50
40 40
30 Train
30
Test
20 20 Best line (train)
10 Best line (test)
10
0 0
0 20 40 60 80 100 120 0 10 20 30 40 50 60 70 80 90
Data index
Predicted heat transfer coefficient

F I G U R E   1 3   Experimental vs. predicted thermal efficiency F I G U R E   1 4   Regression plot between experimental versus
value predicted thermal efficiency value by: (A) Least squares support
∑N exp. vector machine (LSSVM), (B) Adaptive neuro-­fuzzy inference system
i=1
(Hi − Hical. )2 (ANFIS), and (C) Artificial neural network (ANN)
R2 = 1− ∑N exp.
(30)
i=1
(Hi − H̄ exp. )2

Figure  13 depicts concurrent scheme of experimental for

5  |   R ES U LTS A N D D IS C U S S ION proposed models. As it is shown, among the proposed
models; LSSVM model seems to be strongly accurate due
Prediction of the thermal performance of PV/T as a function to less deviated estimations in comparison with actual
of input water temperature, flow rate, and solar radiation is values.
investigated using 3 approaches (ie MLP-­ANN, ANFIS, and Regression results of the are presented in Figure 14.
LSSVM) using an experimental dataset of 100 data points. As it is shown in Figure  15, the deviation assessment of all
Comprehensive information of the 3 designed models (ie proposed models is presented. LSSVM method seems to have less
MLP-­ANN, ANFIS, and LSSVM) are brought in Table 1. deviation due to its higher integrated data points near the zero line.
Different numerical methods were used to evaluate Average relative deviations of 0.22, 0.38, and 0.45 are achieved
the effectiveness and consistency of the designed models. by MLP-­ANN, ANFIS, and LSSVM methods, respectively.
ZAMEN et al.
|
     831

2.5 T A B L E   2   Statistical error analyses for all models

(A)
2 Model MSE RMSE MRE R2 STD
1.5
Relative deviation (%)

LSSVM
1 Test 0.067 0.258 0.446 1.000 0.136
0.5 Train 0.072 0.268 0.373 1.000 0.184
0 Total 0.071 0.266 0.392 1.000 0.173
–0.5 Train Test ANFIS
–1 Test 0.118 0.344 0.379 0.999 0.279
–1.5 Train 0.154 0.393 0.379 0.999 0.340
0 10 20 30 40 50 60 70 80 90
Total 0.145 0.381 0.379 0.999 0.324
Heat transfer coefficient
ANN
3
Test 0.033 0.181 0.216 1.000 0.165
2 (B)
Train 0.000 0.022 0.031 1.000 0.016
1
Relative deviation (%)

Total 0.009 0.093 0.076 1.000 0.087

0
–1
–2 versus hat values (H) the possible outlying will be determined.
–3 Diagonal parameters of the hat matrix which are calculated by
Train Test
–4 Equation (31), are the hat values applied in definition of the
–5 probability zone:
–6 ( )−1
0 10 20 30 40 50 60 70 80 90 H = X XT X XT (31)
Heat transfer coefficient
Feasible region is a squared area restricted to cut-­off and
2.5
(C) warning leverage values. Warning leverage value is deter-
2 mined as follows:
Relative deviation (%)

1.5 k+1
H∗ = 3 (32)
n
1 A Cut-­off value of 3 is suggested for the standardized
residual (R). Feasible region is limited by R = ± 3 lines on
0.5
Train Test vertical axis and H = 0 and H = H* = 0.09 on horizontal axis.
0 Outlying data points are those located outside the feasible re-
gion. As it is shown in the William's plot of Figure 16, almost
–0.5
0 10 20 30 40 50 60 70 80 90 all data are located in the feasible region.
Heat transfer coefficient

F I G U R E   1 5   The deviation assessment of all proposed models 5.2  |  Sensitivity analysis

Reliance of the goal variable on input elements is strongly
Numerical error analyses are also carried out to evaluate specified by the mean of sensitivity analysis. This examina-
the designed models in training, test and overall sections. tion is performed concerning a relevancy parameter (r) vary-
The results are presented in Table 2 for all 3 developed ing from −1 to +1. Greater values of r show higher effects of
models. the corresponding element on goal variable. Relevancy factor
is determined as follows:
5.1  |  Outlier detection ∑N
(Xk,i − X̄ k )(yi − ȳ )
i=1
r= �
Mostly, reliability of a proposed model is under the influ- ∑N ∑N (33)
(Xk,i − X̄ k )2 i=1 (yi − ȳ )2
ence of data set which employed for model improvement.57 i=1

Outliers are separate or set of datum/data which behave in

contrast with majority of the data set. Hence, recognition and Relevancy factors of the study are presented in Figure 17.
omission of outliers in a data set is essential for achieving As it is shown, inlet temperature is the most effective pa-
higher validities. Leverage examination is carried out to de- rameter on thermal efficiency due to the relevancy factor
tect possible outliers. By plotting standardized residuals (R) equal to 0.6.
832     
| ZAMEN et al.

Valid Data
6 (A) Suspected Data
Leverage limit
Standardized Residual limit
4

2
Standard residual

0
0 0.02 0.04 0.06 0.08 0.1
–2

–4

–6

–8
Hat value

Valid Data
6 (B) Suspected Data F I G U R E   1 7   Relevancy factors for all input parameters
Leverage limit
Standardized Residual limit
4 T A B L E   3   Numerically comparison of presented model in this
study and other models
2
Standard residual

Model RMSE R2 References

0
0 0.02 0.04 0.06 0.08 0.1 ELM 0.1937 0.9523 32
–2 ANN 0.347 0.8495 32

–4 GP 0.347 0.8494 32
SVM-­Wavelet 0.3013 0.8871 33
–6
SVM-­RBF 0.3344 0.8593 33
–8 SVM-­FFA 0.3339 0.8598 33
Hat value
LSSVM 0.266 1.00 Current work
Valid Data ANFIS 0.381 1.00 Current work
6 (C) Suspected Data
Leverage limit ANN 0.093 1.00 Current work
Standardized Residual limit
4

2 presented model in this study and other aforementioned mod-

Standard residual

0
els. As it is shown, our designed models are more accurate than
0 0.02 0.04 0.06 0.08 0.1 2 other models, for estimating PV-­T thermal performance.
–2

–4
6  |  CONCLUSION
–6
Three intelligent algorithms of MLP-­ ANN, ANFIS, and
–8 LSSVM were implemented to constructing a relationship be-
Hat value
tween thermal efficiency of solar collector and inlet tempera-
F I G U R E   1 6   William's plot for: (A) Least squares support ture, flow rate, and solar radiation. Graphical and statistical
vector machine (LSSVM), (B) Adaptive neuro-­fuzzy inference system methods were employed to determine the credibility of the pro-
(ANFIS), (C) Artificial neural network (ANN) posed models in accurate prediction of the thermal efficiency.
The proposed ANN model provided the best performance com-
pared to ANFIS and LSSVM models due to the mean squared
5.3  |  Comparison with other models
error (MSE) and determination coefficient (R2) values of 0.009
Mojumder and colleagues have designed 3 models to estimate and 1.00, respectively. Implementation of the outlier detection
the efficiency of PV/T systems utilizing Extreme Learning analysis with the purpose of determining possible outlying data
Machine (ELM), Genetic Programming (GP), and ANN points was also carried out showing the feasibility of the pro-
models.32 Also, Mojumder and colleagues investigated the posed models for majority of data points. Almost all data are
PV-­T thermal efficiency by 3 types of support vector machine located in the feasible region. Sensitivity study was also ac-
method.33 Table 3 presents numerically comparisons of the complished to specify effects of different input parameters on
ZAMEN et al.
|
     833

thermal efficiency where inlet temperature revealed the highest 18. Elsafi AM, Gandhidasan P. Comparative study of double-­pass flat
impact on thermal efficiency of solar collector regarding higher and compound parabolic concentrated photovoltaic–thermal sys-
relevancy factor equal to 0.6. Consequently, these models are tems with and without fins. Energy Convers Manag. 2015;98:59‐68.
19. Othman MYH, Yatim B, Sopian K, Bakar MNA. Performance
user-­friendly and can be reliable value in order to investigate
analysis of a double-­pass photovoltaic/thermal (PV/T) solar collec-
effective parameters in solar collectors.
tor with CPC and fins. Renew Energy. 2005;30:0000‐0000.
20. Tonui JK, Tripanagnostopoulos Y. Improved PV/T solar collectors
R E F E R E NC E S with heat extraction by forced or natural air circulation. Renew
Energy. 2007;32:623‐637.
1. Kannan N, Vakeesan D. Solar energy for future world:-­A review. 21. Farshchimonfared M, Bilbao JI, Sproul AB. Channel depth, air
Renew Sustain Energy Rev. 2016;62:1092‐1105. mass flow rate and air distribution duct diameter optimization of
2. Pandey AK, Tyagi VV, Jeyraj A, Selvaraj L, Rahim NA, Tyagi photovoltaic thermal (PV/T) air collectors linked to residential
SK. Recent advances in solar photovoltaic systems for emerging buildings. Renew Energy. 2015;76:27‐35.
trends and advanced applications. Renew Sustain Energy Rev. 22. Farshchimonfared M, Bilbao JI, Sproul AB. Full optimisation and
2016;53:859‐884. sensitivity analysis of a photovoltaic–thermal (PV/T) air system
3. Sarhaddi F, Farahat S, Ajam H, Behzadmehr A. Exergetic perfor- linked to a typical residential building. Sol Energy. 2016;136:15‐22.
mance assessment of a solar photovoltaic thermal (PV/T) air col- 23. Su D, Jia Y, Huang X, Alva G, Tang Y, Fang G. Dynamic performance
lector. Energy Build. 2010;42:2184‐2199. analysis of photovoltaic–thermal solar collector with dual channels
4. Kumar A, Baredar P, Qureshi U. Historical and recent development for different fluids. Energy Convers Manag. 2016;120:13‐24.
of photovoltaic thermal (PVT) technologies. Renew Sustain Energy 24. Amori KE, Abd-AlRaheem MA. Field study of various air based
Rev. 2015;42:1428‐1436. photovoltaic/thermal hybrid solar collectors. Renew Energy.
5. Chow TT. A review on photovoltaic/thermal hybrid solar technol- 2014;63:402‐414.
ogy. Appl Energy. 2010;87:365‐379. 25. Liang R, Zhang J, Ma L, Li Y. Performance evaluation of new type
6. Daghigh R, Ruslan MH, Sopian K. Advances in liquid based pho- hybrid photovoltaic/thermal solar collector by experimental study.
tovoltaic/thermal (PV/T) collectors. Renew Sustain Energy Rev. Appl Therm Eng. 2015;75:487‐492.
2011;15:4156‐4170. 26. Touafek K, Haddadi M, Malek A. Design and modeling of a pho-
7. Hamid SA, Othman MY, Sopian K, Zaidi SH. An overview of pho- tovoltaic thermal collector for domestic air heating and electricity
tovoltaic thermal combination (PV/T combi) technology. Renew production. Energy Build. 2013;59:21‐28.
Sustain Energy Rev. 2014;38:212‐222. 27. Tripanagnostopoulos Y. Aspects and improvements of hy-
8. Hasan MA, Sumathy K. Photovoltaic thermal module concepts and brid photovoltaic/thermal solar energy systems. Sol Energy.
their performance analysis: a review. Renew Sustain Energy Rev. 2007;81:1117‐1131.
2010;14:1845‐1859. 28. Varol Y, Koca A, Oztop HF, Avci E. Forecasting of thermal en-
9. Lee Y-S, Tong L-I. Predicting high or low transfer efficiency of ergy storage performance of Phase Change Material in a solar
photovoltaic systems using a novel hybrid methodology combining collector using soft computing techniques. Expert Syst Appl.
rough set theory, data envelopment analysis and genetic program- 2010;37:2724‐2732.
ming. Energies. 2012;5:545‐560. 29. Allan J, Dehouche Z, Stankovic S, Mauricette L. Performance test-
10. Dincer I, Rosen MA, Exergy: energy, environment and sustainable ing of thermal and photovoltaic thermal solar collectors. Energy
development, Oxford, UK: Elsevier Science; 2013. Sci Eng. 2015;3:310‐326.
11. Hussain F, Othman MYH, Sopian K, Yatim B, Ruslan H, Othman 30. Charalambous PG, Kalogirou SA, Maidment GG, Yiakoumetti
H. Design development and performance evaluation of photovol- K. Optimization of the photovoltaic thermal (PV/T) collector ab-
taic/thermal (PV/T) air base solar collector. Renew Sustain Energy sorber. Sol Energy. 2011;85:871‐880.
Rev. 2013;25:431‐441. 31. Kramer K, Helmers H. The interaction of standards and innova-
12. Chandrasekar M, Rajkumar S, Valavan D. A review on the thermal tion: hybrid photovoltaic–thermal collectors. Sol Energy. 2013;98:
regulation techniques for non integrated flat PV modules mounted 434‐439.
on building top. Energy Build. 2015;86:692‐697. 32. Mojumder JC, Ong HC, Chong WT, Izadyar N, Shamshirband
13. Reddy SR, Ebadian MA, Lin C-X. A review of PV–T systems: S. The intelligent forecasting of the performances in PV/T collec-
thermal management and efficiency with single phase cooling. Int tors based on soft computing method. Renew Sustain Energy Rev.
J Heat Mass Transf. 2015;91:861‐871. 2017;72:1366‐1378.
14. Böer KW, Tamm G. Solar conversion under consideration of en- 33. Mojumder JC, Ong HC, Chong WT, Shamshirband S. Application
ergy and entropy. Sol Energy. 2003;74:525‐528. of support vector machine for prediction of electrical and thermal
15. Raghuraman P. Analytical predictions of liquid and air photovol- performance in PV/T system. Energy Build. 2016;111:267‐277.
taic/thermal, flat-­plate collector performance. J Sol Energy Eng. 34. Caner M, Gedik E, Keçebaş A. Investigation on thermal perfor-
1981;103:291‐298. mance calculation of two type solar air collectors using artificial
16. Tripanagnostopoulos Y, Tzavellas D, Zoulia I, Chortatou M, Hybrid neural network. Expert Syst Appl. 2011;38:1668‐1674.
PV/T systems with dual heat extraction operation, In Proceedings of 35. Baghban A, Ahmadi MA, Shahraki BH. Prediction carbon dioxide
the 17th PV solar energy conference, Munich 2001 Oct 22 (pp. 22‐26). solubility in presence of various ionic liquids using computational
17. Aste N, Beccali M, Chiesa G. Experimental evaluation of the per- intelligence approaches. J Supercrit Fluids. 2015;98:50‐64.
formance of a prototype hybrid solar photovoltaic-­thermal (PV/T) 36. Ghazani SHHN, Baghban A, Mohammadi AH, Habibzadeh S.
air collector for the integration in sloped roof. Proceedings of Absorption of CO2-­rich gaseous mixtures in ionic liquids: a com-
EPIC. 2002 Oct. putational study. J Supercrit Fluids. 2018;133:455‐65.
 |
834      ZAMEN et al.

37. Baghban A, Mohammadi AH, Taleghani MS. Rigorous modeling 49. Zarei K, Atabati M, Moghaddary S. Predicting the heats of com-
of CO 2 equilibrium absorption in ionic liquids. Int J Greenh Gas bustion of polynitro arene, polynitro heteroarene, acyclic and cy-
Control. 2017;58:19‐41. clic nitramine, nitrate ester and nitroaliphatic compounds using
38. Argiriou AA, Bellas-Velidis I, Kummert M, André P. A neural net- bee algorithm and adaptive neuro-­fuzzy inference system. Chemom
work controller for hydronic heating systems of solar buildings. Intell Lab Syst. 2013;128:37‐48.
Neural Networks. 2004;17:427‐440. 50. Jang J-SR, Neuro-Fuzzy and soft computing. In: Jang J-SR, Sun C-
39. Khatib T. Optimization of a grid-­ connected renewable energy T, Mizutani E. A Comput. Approach to Learn. Mach. Intell. Saddle
system for a case study in Nablus, Palestine. Int J Low-­Carbon River, NJ: Prentice Hall, Inc., 1997: 614.
Technol. 2013;9:311‐318. 51. Nikravesh M, Zadeh LA, Aminzadeh F. Soft computing and intelli-
40. Dorvlo ASS, Jervase JA, Al-Lawati A. Solar radiation estimation gent data analysis in oil exploration. Amsterdam, the Netherlands:
using artificial neural networks. Appl Energy. 2002;71:307‐319. Elsevier; 2003.
41. Sulaiman SI, Rahman TKA, Musirin I, Shaari S. An artificial 52. Lee KH. First course on fuzzy theory and applications. Berlin,
immune-­ based hybrid multi-­ layer feedforward neural network Germany: Springer Science & Business Media; 2006.
for predicting grid-­connected photovoltaic system output. Energy 53. Suykens JAK, Van Gestel T, De Brabanter J. Least squares support
Procedia. 2012;14:260‐264. vector machines. Singapore: World Scientific; 2002.
42. Mellit A, Kalogirou SA. ANFIS-­ based modelling for pho- 54. Suykens JAK, Vandewalle J. Least squares support vector machine
tovoltaic power supply system: a case study. Renew Energy. classifiers. Neural Process Lett. 1999;9:293‐300.
2011;36:250‐258. 55. Hemmati-Sarapardeh A, Shokrollahi A, Tatar A, Gharagheizi F,
43. Izgi E, Öztopal A, Yerli B, Kaymak MK, Şahin AD. Short–mid-­ Mohammadi AH, Naseri A. Reservoir oil viscosity determination
term solar power prediction by using artificial neural networks. Sol using a rigorous approach. Fuel. 2014;116:39‐48.
Energy. 2012;86:725‐733. 56. Suykens JAK, Van Gestel T, De Moor DBJB, Vandewalle J. Least
44. Haykin S. Neural networks: a comprehensive foundation, New Squares Support Vector Mach. 2002.
York, NY: Prentice Hall PTR; 1994. 57. Rousseeuw PJ, Leroy AM. Robust regression and outlier detection.
45. Rosenblatt F. Principles of neurodynamics: perceptrons and the New York, NY: John wiley & Sons; 2005.
theory of brain mechanisms. Washingt. DC: Spartan; 1962 (n.d.).
46. Rumelhart DE, Hinton GE, Williams RJ. Learning internal rep-
resentations by error propagation, California Univ San Diego La How to cite this article: Zamen M, Baghban A,
Jolla Inst for. Cognitive Science. 1985. Pourkiaei SM, Ahmadi MH. Optimization methods
47. Zadeh LA. Information and control. Fuzzy Sets. 1965;8:338‐353. using artificial intelligence algorithms to estimate
48. Safari H, Nekoeian S, Shirdel MR, Ahmadi H, Bahadori A,
thermal efficiency of PV/T system. Energy Sci Eng.
Zendehboudi S. Assessing the dynamic viscosity of Na–K–Ca–
Cl–H2O aqueous solutions at high-­pressure and high-­temperature
2019;7:821–834. https://doi.org/10.1002/ese3.312
conditions. Ind Eng Chem Res. 2014;53:11488‐11500.
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.