Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Renewable Energy 206 (2023) 135–147

Contents lists available at ScienceDirect

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

Deep learning based long-term global solar irradiance and temperature


forecasting using time series with multi-step multivariate output
Narjes Azizi , Maryam Yaghoubirad , Meisam Farajollahi *, Abolfzl Ahmadi
Department of Energy Systems, School of Advanced Technologies, Iran University of Science and Technology, Narmak, Tehran, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: Solar radiation’s intermittent and fluctuating nature poses severe limitations for most applications. Accurate
Solar irradiance forecasting prediction of solar radiation is an essential factor in predicting the output power of a photovoltaic power system.
Deep learning For this purpose, the potential of the 20 MW solar photovoltaic power plant in Zahedan city has been evaluated
Time series
in this article. With the help of monthly data (1984–2021) and MLP, LSTM, GRU, CNN, and CNN-LSTM models,
Long-term prediction
Multi-step multivariate output
solar radiation and temperature are predicted for the next ten years. CNN exhibits the best performance
compared to other models with four input parameters: global horizontal irradiance, temperature, surface pres­
sure, relative humidity (RH), and two outputs of temperature and radiation. The root mean square error values
for global horizontal irradiance and temperature were 12.68 W/m2 and 1.75 ◦ C, respectively. Relative humidity
exhibited more significant effect on the model in comparison with surface pressure. Finally, the average annual
power output for ten years from 2022 to 2031 is calculated and predicted to be 50.37 GWh.

forecasting since the 1970s [10]. It was shown that ARIMA in short-time
horizons and Numerical Weather Prediction (NWP) models with
1. Introduction increasing horizons are more accurate [11,12]. The acceptance of the
ARIMA models is due to their statistical properties and the renowned
Solar energy, a common source of renewable energy, has random, Box-Jenkins method in the model-building process. However, their main
intermittent, and uncertain characteristics. This variability is due to the limitation is the pre-assumed linear form of the model [13]. Recently,
variety of cloud cover, aerosol, and participating gases in the atmo­ Artificial Neural Networks (ANN) have been widely investigated and
sphere [1,2], which may lead to defects in the reliability and stability of used in time series forecasting due to learning from the features pre­
power grid systems [3]. Proper forecasting and accurate solar radiation sented in the data, inferring the unobservable part of noisy data, and
analysis could be the assistance in reducing risk and enabling asset non-linearity [14]. One study compared the performance of different
exploitation in the most cost-effective way possible. By using the precise Neural Networks (NNs) and statistical models such as ARMA, ARIMA,
prediction of solar radiation, it is possible to investigate the economic and SARIMA in forecasting large-scale PV plants’ output power. The
justification of the solar power plant at a given location and significantly results showed that NNs are more accurate than proposed statistical
reduce costs by the placement optimization [4,5]. Solar power predic­ models and require less computing time [15].
tion can be investigated by two main approaches: 1) based on fore­ ANN model performs better in Mean Absolute Percentage Error
casting time horizon 2) based on input data type. The forecasting (MAPE) and Correlation Coefficient (R) than empirical regression
methods of Solar Radiation (SR) based on historical data, which is one of models [16]. Guermoui et al. used the Multi-Layer Perceptron (MLP)
the methods based on input data, is classified into four classes, including model to predict Daily Global Solar Radiation (DGSR) in Ghradia [17]. It
(a) statistical, (b) persistence, (c) machine-learning, and (d) hybrid was shown that the ANN models in MAPE and R exhibited better per­
method. The classifications of the approaches and methods are shown in formance compared to empirical regression models [18]. Radial Basis
Fig (1) [6–8]. Function (RBF) and MLP models also were utilized for long-term data
Many efforts have been made to predict solar radiation, and solar from eight different stations in Oman, and RBF models were preferred
forecasting can be considered a time series problem. Autoregressive due to fewer computational requirements [19]. MLP and Multi Linear
Moving Average (ARMA) and Autoregressive Integrated Moving Regression (MLR) networks using nine input features were proposed,
Average (ARIMA) [9] have been utilized for applications of solar

* Corresponding author.
E-mail address: farajollahi@iust.ac.ir (M. Farajollahi).

https://doi.org/10.1016/j.renene.2023.01.102
Received 6 September 2022; Received in revised form 26 January 2023; Accepted 27 January 2023
Available online 7 February 2023
0960-1481/© 2023 Elsevier Ltd. All rights reserved.
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Nomenclature MAE Mean Absolute Error


MAPE Mean Absolute Percentage Error
Symbol MLP Multilayer Perceptron
Evap Evaporation MLR Multi Linear Regression
ISC Short circuit current MSE Mean Square Error
PO Output power NN Neural Network
RHMAX Relative Humidity at maximum temperature NOCT Nominal Operating Cell Temperature
RHMIN Relative Humidity at minimum temperature NRMSE Normalized Root Mean Square Error
TMAX maximum temperature NWP Numerical Weather Prediction
TMIN minimum temperature PV Photovoltaic
VOC Open circuit voltage PW Precipitable Water
R Corrolation Coeffident
Greek symbols R2 coefficient of determination
α Temperature Coefficients of ISC RBF Radial Basis Function
β Temperature Coefficients of VOC ReLU Rectified Linear Unit
acronyms RH Relative Humidity
ANN Artificial Neural Network rMABE Relative Mean Absolute Bias Error
ARMA Autoregressive Moving Average RMSE Root Mean Square Error
ARIMA Autoregressive Integrated Moving Average RNN Recurrent Neural Network
CLSTM Convolutional Long Short-Term Memory rRMSE Relative Root Mean Square Error
CNN Convolutional Neural Netwerk SARIMA Seasonal Autoregressive Integrated Moving Average
DGSR Daily Global Solar Radiation SH Sunshine Hour
DNN Deep Neural Network SP Surface Pressure
DPT Dew Point Temperature SR Solar Radiation
FF Fill Factor SVM Support Vector Machine
G Solar radiation SZA Solar Zenith Angle
GHI Global Horizontal Irradiance T Temperature
GRU Gated Recurrent Unit WD Wind Direction
LSTM Long Short-Term Memory WS Wind Speed

Fig. 1. Classification of solar power prediction approaches and methods.

and the performance evaluation criteria showed that the MLP model is radiation in five cities in Bangladesh. Among the three models, the
better compared to the MLR model [20]. GRU-based model had the best performance in terms of MAPE [25].
Recently, deep learning-based models have attracted much attention Deep learning techniques, including RNN, LSTM, and GRU, were pro­
in many applications due to their superior ability to learn complex posed for the prediction of solar energy, and it was observed that LSTM
patterns from raw data. Long Short-Term Memory (LSTM) and Gated and RNN are slightly better than the GRU model due to their capability
Recurrent Unit (GRU) are the most popular algorithms used for time- to preserve long-term dependencies [26]. Also, it was shown that GRU
series analysis since they can process sequential data well and are able and LSTM significantly improve forecasting accuracy when using
to model long-term independencies. They also have been utilized in load multivariate data compared to univariate data [27].
forecasting problems [21–23]. In the Photovoltaic (PV) power predic­ CNNs are also efficient in time-series problems [28]. In another
tion LSTM, RNN, and GRU were proposed and implemented [24]. Three study, CNN was implemented for short-term load forecasting. The model
NN models, including RNN, LSTM, and GRU were used to predict solar was observed to be very flexible and capable of handling forecasting

136
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Table 1
Summary of several long-term forecasts for solar radiation.
Ref Year Best Input Variable Forecasting Time Period/Location Evaluation Parameter
Model Horizon

[33] 2005 MLP Longitude, Altitude, Month of the year 1 month 1991–2010/Algeria RMSE <8%
R = 90–92%
[34] 2012 ARIMA SH, T, WS, RH, GSR monthly 1995–2007/Al-Ain R2 = 99.98%
[35] 2015 SVM Tmax, Tmin, SH, RH, Extraterrestrial SR, T monthly 1995–2007/Bandar Abbas MAPE = 3.2601%
MABE = 0.5104 MJ/m2d
RMSE = 0.6618 MJ/m2d
rRMSE = 3.6935%
[36] 2017 MLP SH, T monthly 2000–2009/India R = 99.68%
[37] 2018 NAR- GSR 2 years February 1, 2004 to April R2 = 96%
ANN 2018/Nigeria R = 98%
[38] 2019 SARIMA SR 10 years 1981–2017/Seoul R2 = 79%
RMSE = 33.18 Wh/m2
[39] 2019 CLSTM GSR 1 month 1 January 2006 to 31 July MAE = 13.131 W/m2
2014/Alice Springs RMSE = 32.872 W/m2
[40] 2019 LSTM Date, Tmin, Tropospheric Ozone, Latitude, Longitude, 10 years 1 January 1979 to 23 July R2 = 97–99%
Precipitation, WS, Elevation, Tmax, RH, SR 2014/Mowo
[41] 2019 RF Tmin, Tmax, Precipitation, Evap, Vapor Pressure, RHmin, monthly 1905–2018/Australia rRMSE = 14.04%
RHmax, Potential Evapotranspiration rMABE = 3.79%
[42] 2020 ARIMA SR 31 days 1 October 2018 to 31 MAPE = 3.7567
December 2019/Tetouan

tasks [29].
The hybrid method can be an effective way to enhance forecasting
performance because they are not only exploited for time series but also
have the ability to extract spatial and temporal characteristics that make
them appropriate for various applications [13]. It was shown that the
CNN-LSTM model has a higher predictive ability in PV power fore­
casting than the single CNN and LSTM models [30]. It was reported that
the hybrid of CNN-LSTM can improve the forecasting accuracy over
single models, including RNN, LSTM, CNN, and GRU [31].
The long-term forecast (more than one month), in addition to energy
tenders and security operations, helps officials plan electricity produc­
tion, transmission, and distribution [32]. Table 1 summarizes the results
of several studies on long-term solar radiation prediction. To the best of
our knowledge, there are only a few papers in the long-term prediction
area, which motivated us to work on these new emerging methods for
Fig. 2. Schematic of an MLP structure.
long-term solar power forecasting.
In this work, several models, including MLP, LSTM, GRU, CNN, and a
hybrid model called CNN-LSTM, are applied and implemented to predict making [17,43]. Fig (2) shows the schematic of an MLP structure.
the temperature and also the generated power of a 20 MW PV power
plant located at Zahedan city in the southeast of Iran for next ten years.
2.2. Recurrent neural network (RNN)
Two types of simulation are conducted: a) 2 inputs- 2 outputs, b) 4 in­
puts- 2 outputs to predict solar irradiation for exploiting the results to
Recurrent Neural Networks are a type of ANN designed to identify
examine the feasibility of a solar power plant’s optimal placement in a
patterns in sequential data, such as text, natural language, time series
specific region. The evaluation parameters of the models are compared,
data, stock markets, etc. The idea behind these NNs is that they allow
and it is observed that the CNN model with four inputs exhibits the best
cells to learn from previously connected cells. It can be said, in a way,
performance compared to other models. Also, the effect of features and
that these cells have “memory”. Therefore, they create more complex
the number of features are investigated, and it is concluded that relative
knowledge from the input data [44].
humidity significantly affects temperature prediction. Eventually, this
RNN networks make up for the shortcomings of feedforward neural
study tries to fill this gap with a prediction approach using multivariable
networks. Because feedforward networks can only accept and produce
inputs and two outputs as long-term forecasting, which helps one to
fixed-sized inputs and outputs and consider previous inputs in the same
have a more accurate economic and environmental analysis.
order. By considering the past inputs in the sequences, the recurrent
neural networks can capture the temporal dependencies that the feed­
2. Methodology
forward neural network cannot. The primary goal of RNNs is to process
and forecast sequential data. Classical RNN calculation can be expressed
2.1. Multilayer perceptron (MLP)
as Eqs. (1) and (2) [45]:

ANNs simplest and most common type is a fully connected feedfor­ ht = f (Uxt + ωht− 1 + b) (1)
ward ANN or MLP. It consists of three layers: a) the input layer, b) the
output layer, and c) the hidden layers, which are intermediate layers yt = σ t (Vht + c) (2)
between the input and output layers. MLP has onward connections from
input to output through consecutive hidden layers. Also, their links are where xt is the input of the training sample, ht is the hidden state of the
based on the weight structure, and their values vary through model model, yt is the output of the model all at time t. U, ω, and V are weight
training and create a stable framework with the capacity for decision- parameters, b and c are biases, and f is the activation function. Theo­
retically, RNNs can solve the training of sequential data with good

137
N. Azizi et al. Renewable Energy 206 (2023) 135–147

( )
ft = σ ωfh .ht− 1 + ωfx .xt + bf (3)

According to Eq. (3), the forget gate controls how much information
can be retained from the previous state to the current state. xt is the
input vector and ωfx and ωfh are weight matrices. ht− 1 stores the infor­
mation related to the preceding time step and bf represents the bias.
These three pieces of information are combined and then calculated by
the activation function, generally the sigmoid function Eq. (4) [47].
it = σ(ωih .ht− 1 + ωix .xt + bi ) (4)
According to Eq. (5), the input gate controls how much input data
information can be stored at the current moment [48,49].
ot = σ(ωoh .ht− 1 + ωox .xt + bo ) (5)

Fig. 3. The architecture of the RNN model.


ct = ωt × ωt− 1 + it × tanh(ωo × [ht− 1 , xt] + bo ) (6)

performance, but they are disposed to vanishing and explosion gradient ht = ot × tanh(ct ) (7)
problems when the sequential data is too long. The architecture of the
RNN model is shown in Fig (3). ct in Eq. (6) calculates data information from the beginning to the cur­
rent state. In Eq. (7), ht controls how much stored information from the
2.2.1. Long Short-Term Memory (LSTM) beginning to the current state, could be transferred to the next state. Fig
Long short-term memory networks were developed to overcome the (4) shows the structure of an LSTM.
vanishing and explosion gradient problem. The LSTM network has a
state cell that, in addition to the hidden state, controls the flow of in­ 2.2.2. Gated Recurrent Unit (GRU)
formation between time steps. LSTM consists of three gates, including Gated Recurrent Unit Neural Network is a modified version of LSTM,
forget gate, input gate, and output gate, which control the storing and which reduces the gates number while keeping long-term memory
deletion of information [46]. connections to solve the fading problem. GRUs consist of two gates,

Fig. 4. LSTM structure.

Fig. 5. GRU structure.

138
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Fig. 6. CNN structure.

namely, update and reset gates. The former defines what information is 2.4. CNN-LSTM
forgotten about a new input and what new information is added, and the
latter describes how much long-term or previous data is forgotten. Eqs. The main structure is CNN-LSTM, including the input layer, CNN
(8) and (9) are related to reset and forget gates, respectively [27]. layer, pooling layer, LSTM hidden layer, and full connection layer. The
CNN layer is used for processing spatial or locally related data, the LSTM
zt = σ(ωzh .ht− 1 + ωzx .xt + bz ) (8)
layer for capturing the characteristics of data related to time series, and
the fully connected layer for forecasting the output. The CNN-LSTM
rt = σ(ωrh .ht− 1 + ωrx .xt + br ) (9)
model will accept subsequences of the main sequence as blocks,
extract features from each block, then the LSTM layer to interpret the
(10)

ht = tanh(ωxh .xt + rt ωhh .ht− 1 )
features extracted from each block [53].
In Eq. (10), the input information xt and the previous time step in­
formation, ht− 1 are linearly combined, and the different matrices ω are 2.5. Multistep ahead forecasting
multiplied respectively on the right side. Then the reset gate rt and ωhh .
ht− 1 are multiplied together. Lastly, the new information of the current Multistep forecasting can predict all times during one stage [44]. Fig
state is calculated through an activation function, generally tanh func­ (7) shows a schematic of a multistep ahead forecasting strategy with the
tion [50]. direct method.

(11)

ht = zt .ht− 1 + (1 − zt )ht
3. Cost function
According to Eq. (11), the product of zt and ht− 1 shows the final data
information stored in the prior time step. This product and the infor­ To evaluate the performance of all models some criteria are defined
mation stored from the current memory to the final memory are equal to including: Root Mean Square Error (RMSE), coefficient of determination
the content of the output ht obtained by the final gate unit. Fig (5) in­ (R2), Mean Absolute Error (MAE), and Normalized Root Mean Square
dicates the architecture of GRU. Error (NRMSE). They can be calculated as follow [8,46]:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
2.3. Convolutional neural network (CNN) 1∑ n
RMSE = (y − ̂ y )2 (13)
n i=1
In deep learning, CNN is a feedforward deep neural network. A CNN
consists of an input, several hidden layers, and one output layer. Hidden ∑
n
y )2
(y − ̂
layers are typically a set of convolution layers with a modified activation 2
R =1 − i=1
(14)
function called linear unit (ReLU), followed by a pooling layer and dense ∑
n
y )2

layers. All these layers are stacked successively to create a CNN [51,52]. i=1

Fig (6) shows the structure of a CNN network. The formula of feature
extraction by a 1D convolution network is described as follows [39]: 1∑ n
MAE = y)
(y − ̂ (15)
(( ) ) n i=1
hkij = f ωk *x ij + bk (12)
In Eq. (13)–(15), y is the expected value, ̂
y is the predicted value, and
where f represents the activation function, ωk is the weight of the n is the number of observations
kernel connected to kth feature map, and the star (*) is an operator of the
convolutional process. RMSE
NRMSE = (16)
mean(observation)

Fig. 7. Schematic of a multistep ahead forecasting strategy.

139
N. Azizi et al. Renewable Energy 206 (2023) 135–147

balance relation in the module using the Nominal Operating Cell Tem­
perature (NOCT). In the NOCT condition, the ambient temperature is 20

C, solar radiation is 800 W/m2, and average wind speed is 1 m/s, and
the electric charge is assumed to be zero, which is determined as follows
[56]:
TC = Ta + (NOCT − 20) × G / 800 (20)
Ta is the ambient temperature. The Filling Factor (FF) is a parameter
that determines the maximum power extracted from the cell in the VOC
and ISC interaction. FF, is defined as presented in Eq. (21):
VOC − ln(VOC + 0.72)
FF = (21)
VOC + 1

5. Solar energy potential and data set

Iran has vast desert regions and cloudless skies, which are favorable
for solar energy, and this energy has a very different potential in each
area [57]. Calculations show that the number of hours of solar radiation
in Iran exceeds 2800 h per year. Iran’s solar radiation is 2.5 times larger
compared to European countries. It has high solar energy potential with
about 300 sunny days per year and average solar radiation of 2200 kWh/
m2 [58]. Fig (8) represents the Global Horizontal Irradiation (GHI) atlas
in Iran, which shows the higher value of solar irradiation for the
southeast section, which makes it a good candidate to demonstrate a
solar power plant.
Fig. 8. Global horizontal irradiation of Iran [59]. Zahedan is a city located in the southeast of Iran (latitude of 29.47,
longitude of 60.90) which has one of the highest solar irradiance den­
sities with the highest solar irradiation level of 8 kWh/m2/day in June,
4. Energy analysis
and the lowest level of 3.79 kWh/m2/day in December. The meteoro­
logical information of Zahedan from 1984 to 2021 (456 months) has
In this section, mathematical relationships related to PV potential
been received monthly from the NASA website. Fig (9) shows Zahedan’s
examination are discussed. The output power generated by a PV panel is
monthly average daylight hours (2001–2020) per month; the average
obtained using Eq. (17) [54]:
daylight is 4394 h per year.
PO = VOC ISC FF (17)
6. Pre-processing and feature selection
that VOC stands for the open circuit voltage, ISC stands for the short
circuit current, which are respectively the highest voltage and current 6.1. Pre-processing
values that a cell can produce. VOC is given by Eq. (18) [55]:
VOC (T) = VOC,25◦ C [1 + β(TC − 25)] (18) Pre-processing and feature selection are essential steps in deep
learning approaches to do well data collection process. Pre-processing
where VOC 25◦ C is the open circuit voltage at a temperature of 25 ◦ C, β is causes the clean input data to be easily used in deep learning models
the temperature coefficient of VOC and TC is the working temperature of and affects the model’s accuracy and results. Therefore, the following
the cell. ISC is also calculated using Eq. (19): steps have been conducted to prepare and pre-process the data.
/
ISC (T) = VSC,25◦ C [1 + α(TC − 25)] × G 1000 (19) 1. The average of each month was used to fill out the missing data in
that month.
where ISC,25◦ C is the short circuit current at 25 ◦ C, α is the temperature
2. The features of a dataset are usually given at different scales. To
coefficient of ISC and G is the solar radiation.
ensure that all these values are on the same scale and to add uni­
Working temperature of the cell, TC , can be calculated by the energy
formity to the data set, a Min-Max scale converts all values between

Fig. 9. Monthly Averaged Daylight Hours per month in Zahedan [60].

140
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Fig. 10. The correlation matrix.

0 and 1 and removes noise from the data. It slows down and sim­ radiation in a region, as they have been validated in previous studies
plifies the learning process of the presented models. Also, it is [63]. Several features, including the month and meteorological data,
necessary to use normalized data in neural networks for better per­ such as Relative Humidity (RH), Wind Speed (WS), Wind Direction
formance of the error propagation process. Eq. (22) shows the data (WD), Temperature (T), GHI, specific humidity, and Surface Pressure
normalization [61]: (SP), were considered effective input parameters for different ANN
models. Finally, optimal features were selected by the Pearson correla­
x − min(x)
xscaled = (22) tion method. These features highly correlate with the target outputs,
max(x) − min(x)
GHI, and temperature. This step aims to help the development of the
Time series data can be expressed as supervised learning. Given a most accurate models that learn from the most relevant data to provide
sequence of numbers for a time series dataset, the data is reconstructed the most accurate predictions [64,65]. Fig (10) shows the correlation
to convert to a supervised learning problem. This is done using the matrix for Zahedan; according to this matrix, features with a correlation
previous time steps as the inputs and the next time step as the outputs above 0.7 with solar radiation have been selected as effective features
[53]. for forecasting. Temperature, SP, and RH have the highest correlation
with solar radiation. Temperature is positively correlated, while SP and
6.2. Feature selection RH negatively correlate with the target variable. This relationship be­
tween the three variables with temperature is also valid.
Feature selection affects the learning process and can lessen the The simulation process used in the current study is shown in Fig (11).
number of variables to increase the model’s accuracy. It can remove This method is constructed by four main stages: a) data gathering and
redundant or unimportant information and preserve the most critical data pre-processing, b) model training, c) hyperparameters tuning, and
features to reduce computational costs and improve overfitting and d) model testing and validation.
multicollinearity problems [62].
Some meteorological parameters significantly affect the global solar

Fig. 11. The simulation process implemented for forecasting output solar power.

141
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Table 2
Specifications of MLP, LSTM, GRU, CNN, and CNN-LSTM structure.
Model Layer 1 Layer 2 Layer 3 Layer 4

Type/neurons Activation Type/ Activation Type/ Activation Type/ Activation Batch Epoch
function neurons function neurons function neurons function size

MLP Dense/50 ReLU Dense/50 ReLU Dense/120 Linear – – 10 200


LSTM Lstm/40 ReLU Dense/90 tanh Dense/120 Linear – – 20 200
GRU GRU/50 ReLU Dense/50 tanh Dense/120 Linear – – 10 200
CNN Conv1d/50, ReLU Flatten – Dense/50 tanh Dense/120 Linear 30 300
Kernel/3
CNN- Conv1d/16, ReLU Lstm/60 ReLU Dense/200 ReLU Dense/120 Linear 60 100
LSTM Kernel/3

Table 3
GHI evaluation parameters of MLP, LSTM, GRU, CNN, CNN-LSTM models.
RMSE (W/m2) MAE (W/m2) R2 (%) NRMSE (%)

Model input features train test train test train test train test

MLP_I T, GHI 11.01 13.60 8.52 11.02 96.92 95.28 4.51 5.40
MLP_II T,GHI,SP,RH 10.41 13.33 8.05 10.80 97.24 95.47 4.26 5.29
LSTM_I T,GHI 11.76 13.32 9.13 10.82 96.48 95.48 4.81 5.29
LSTM_II T,GHI,SP,RH 11.10 13.64 8.64 11.07 96.85 95.25 4.54 5.42
GRU_I T,GHI 11.24 14.22 9.54 12.85 96.79 94.84 4.60 5.65
GRU_II T,GHI,SP,RH 10.66 13.08 8.22 10.66 97.11 95.63 4.36 5.20
CNN_I T,GHI 10.80 13.62 8.44 10.91 97.04 95.25 4.42 5.42
CNN_II T,GHI,SP,RH 10.46 12.87 8.15 10.42 97.22 95.77 4.28 5.12
CNN-LSTM_I T,GHI 11.45 13.62 8.85 11.16 96.67 95.27 4.69 5.41
CNN-LSTM_II T,GHI,SP,RH 11.13 13.87 8.66 11.29 96.84 95.09 4.55 5.52

7. Results and discussion predict GHI as the first output and then to predict temperature as the
second one. It is worth mentioning that hyperparameters need to be
Table 2 indicates the specifications of MLP, LSTM, GRU, CNN, and readjusted if the priority in the prediction changes.
CNN-LSTM models with a hidden layer. The input features are organized Increasing the number of features from 2 to 4, causes the following
in two ways: a) two features, including T, GHI, and b) four features changes in the models.
selected from Pearson’s coefficient method, including T, GHI, RH, and
SP. The expected result is multistep ahead forecasting with two outputs • In the MLP model, the GHI evaluation parameters improved, but
of T and GHI. there was no change in the temperature evaluation parameters.
In all models, 50% of the data is allocated to the train data and the • GRU and CNN models are improved to evaluate GHI and temperature
rest to the test data. The optimization algorithm is Adam’s algorithm, parameters.
and the cost function is Mean Square Error (MSE). Developing an ar­ • The LSTM model has led to the deterioration of the GHI evaluation
chitecture with two hidden layers almost did not improve the prediction parameters, and overfitting has also occurred and, to some extent,
accuracy. Considering that the number of input data was low, using all has led to the deterioration of the temperature evaluation parame­
data as the batch size worsened the results of the evaluation parameters. ters. Suppose that the ReLU activation function is used in the hidden
In CNN-related models, pooling and dropout layers resulted in the layer of the LSTM model; then, by increasing the input parameters,
worsening of the evaluation parameters. Except for the CNN model, there will be no change in the temperature and GHI evaluation pa­
which had 24 months of input data, 12 months were used as input data rameters. However, this function was chosen because evaluation
in the rest of the models. parameters of the temperature for the tanh activation function
Tables 3 and 4 show the evaluation parameters of MLP, LSTM, GRU, exhibit better condition.
CNN, and CNN-LSTM models for the two features of temperature and • The CNN-LSTM model led to the deterioration of temperature and
GHI. As previously mentioned, two different cases were simulated with GHI evaluation parameters, and it also happened in the case of both
two different inputs. The first case with two features (T, GHI), and the overfitting parameters.
second case with four features (T, GHI, RH, SP). The model is expected to

Table 4
Temperature evaluation parameters of MLP, LSTM, GRU, CNN, CNN-LSTM models.
RMSE (◦ C) MAE (◦ C) R2(%) NRMSE (%)

Model input features train test train test train test train test

MLP_I T,GHI 1.27 1.80 0.99 1.45 97.23 94.72 7.15 9.87
MLP_II T,GHI,SP,RH 1.24 1.80 0.97 1.46 97.36 94.72 6.98 9.87
LSTM_I T,GHI 1.33 1.76 1.03 1.44 96.92 94.94 7.54 9.66
LSTM_II T,GHI,SP,RH 1.29 1.74 1 1.42 97.15 95.56 7..25 9.55
GRU_I T,GHI 1.28 1.80 1.00 1.44 97.20 94.71 7.19 9.89
GRU_II T,GHI,SP,RH 1.23 1.71 0.96 1.39 97.39 95.22 6.94 9.39
CNN_I T,GHI 1.25 1.85 0.98 1.50 97.31 94.42 7.05 10.14
CNN_II T,GHI,SP,RH 1.22 1.75 0.95 1.41 97.45 95.00 6.86 9.60
CNN-LSTM_I T,GHI 1.32 1.72 1.04 1.39 97.00 95.20 7.44 9.41
CNN-LSTM_II T,GHI,SP,RH 1.30 1.80 1.02 1.46 97.10 94.82 7.32 9.88

142
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Table 5
GHI evaluation parameters of CNN model.
NRMSE (%) R2 (%) MAE (W/m2) RMSE (W/m2)

Model Input parameters train test train test train test train test

CNN_III SP,GHI,T 11.05 13.14 8.72 10.51 96.89 95.58 4.53 5.23
CNN_IV RH,GHI,T 10.34 13.15 8.10 10.64 97.29 95.58 4.23 5.23

Table 6
Temperature evaluation parameters of CNN model.
NRMSE (%) R2 (%) MAE (◦ C) RMSE (◦ C)

Model Input parameters train test train test train test train test

CNN_III SP,GHI,T 1.28 1.82 1.00 1.48 97.21 94.59 7.17 9.98
CNN_IV RH,GHI,T 1.21 1.77 0.95 1.43 97.49 94.51 6.80 9.68

Fig. 12. Loss function’s value as a function of epoch for the CNN_II model.

Fig. 13. Forecasted average values of GHI on test data for CNN-II model.

Fig. 14. Forecasted average values of temperature on test data for CNN-II model.

143
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Fig. 15. Histogram diagram of predicted and actual values of (a) GHI and (b) temperature data of CNN-II model on test data.

Fig. 16. Regression plot of the predicted and actual values of (a) GHI and (b) temperature data of CNN-II model on test data.

In Table 3, the best evaluation parameters regarding radiation Obviously, increasing epochs leads to a decrease in MAE value, and the
belong to the CNN_II model. In Table 4, the best evaluation parameters learning process is completed. Large number of epochs (larger than 300)
are observed for GRU_II model, which is very close to the results of causes an increase in the probability of overfitting, and the model loses
CNN_II model; eventually, the CNN_II model was selected for forecasting its efficiency; however, in lower epoch values, the learning process does
as the best model. not exhibit acceptable performance.
Moreover, the effect of RH and SP were investigated separately on The performance of the proposed model to follow the test data for the
the CNN_II model. Table 5 summarizes the evaluation parameters for GHI and temperature as output are shown in Fig (13) and Fig (14),
CNN_III and CNN_IV models, which consider SP and RH as features with respectively.
GHI and temperature. GHI parameters are improved compared to the In addition to monthly forecasting, another essential target is to
CNN_I model. However, according to Table 6, it is approved that the RH forecast the distribution of observations. In other words, a histogram
feature enhances the evaluation parameters of the temperature in enables one to understand the reliability of the model. Fig (15) shows
comparison with the CNN_I model, but SP has no constructive effect. the distribution of forecast observations for the two features of tem­
Fig (12) shows the loss function’s value as a function of epoch. perature and GHI on the test data. Also, the graph of kernel density

Fig. 17. GHI diagram by month from 1984 to 2031 with CNN-II model.

144
N. Azizi et al. Renewable Energy 206 (2023) 135–147

Fig. 18. Temperature diagram by month from 1984 to 2031 with CNN-II model.

years with acceptable accuracy and efficiency, which was the aim of this
Table 7
study. This type of prediction is very useful and advantageous for those
Specifications of the selected PV module.
who want to design and analyze the feasibility of the solar power plant
Parameter Value in a specific region from an economic and environmental point of view.
Model number YLM72CELL 40 mm Series Now, one is able to investigate the potential of constructing a solar
Maximum power 340 W power plant with a rated power of 20 MW. The rated power of each
Short Circuit Current (A) 9.35
given panel is 340 W. Therefore, 58,824 PV panels are required to
Open Circuit Voltage (V) 47.3
Temperature Coefficients of VOC − 0.32/◦
provide the desired capacity of this power plant. The specifications of
(%)
∁ the selected panel used in the proposed power plant are summarized in
Temperature Coefficients of ISC +0.05/◦

Table 7.
(%) According to Eq. (16)–(20), two parameters of temperature and GHI
Efficiency 17.5%
are needed to predict the potential of PV panels to generate power.
Lifetime 20 years
NOCT 46 ± 20 C Therefore, considering the GHI and the temperature obtained from the
Standard Test Condition Irradiation = 800 W/m2, temperature = 20 ◦ ∁, CNN_II model, the output power of this power plant is calculated. Fig
AM = 1.5 (19) shows the annual power output of the PV plant for the years from
1984 to 2031.
It should be noted that in all the calculations, some parameters, such
as humidity, dust, freezing, and cloudy weather, have not been consid­
ered in the numerical calculations of the production power.

8. Conclusion

The aim of this study was to predict the potential power generation
of the 20 MW solar power plant located in a specific region in Iran called
Zahedan city for the next ten years. Several deep learning models,
including MLP, LSTM, GRU, CNN, and CNN-LSTM were implemented to
predict the temperature and GHI to calculate solar power produced by
the power plant. Two main types of models were proposed and simu­
lated. In the first type of model, two features (T and GHI) were
considered as inputs, and in the second type, four features (T, GHI, RH,
SP) were selected using the Pearson coefficient. The simulation results
confirmed the better performance of the CNN_II model compared to
Fig. 19. The annual power output of the 20 MW solar power plant located in other models to accurately follow the actual test data with four selected
Zahedan for years from 1984 to 2031. inputs. The results demonstrated that the value of RMSE, R2, and MAE
for GHI prediction was 12.87 W/m2, 95.77%, and 10.42 W/m2,
estimation has been plotted for a better understanding and more respectively. These values were 1.75 ◦ C, 95%, and 1.41 ◦ C for temper­
straightforward comparison of the predicted and actual values. Ac­ ature prediction. It was observed that adding the SP or RH feature to the
cording to the kernel density diagram, it is confirmed that the model has CNN_I model improves the evaluation parameters; however, the RH had
the capability to predict GHI and temperature due to reasonable con­ a greater effect compared to the SP. The range of monthly GHI variation
sistency between the trends of real and predicted values in the kernel for the next ten years is 151.5–341.5 W/m2 and, for temperature, be­
density diagram. tween 7 and 30.5 ◦ C. Finally, the predicted power generation of the solar
Another way to examine the efficiency of the model is to fit a power plant was calculated to be 50.37 GWh for years between 2022 and
regression line between the actual and predicted values. The accumu­ 2031.
lation of the points closed to y = x line, approves the good agreement
between the predicted values and real ones (Fig (16)). It is also observed CRediT authorship contribution statement
that there is a significant correlation between the temperature and GHI
actual and predicted data. Narjes Azizi: Simulation and developing the code and, Software.
The final simulation results (Fig (17) and (18)), also confirm the Maryam Yaghoubirad: Simulation and developing the code and,
capability of the proposed CNN_II model to predict GHI for the next ten Software. Meisam Farajollahi: Mentoring and, Supervision, and

145
N. Azizi et al. Renewable Energy 206 (2023) 135–147

writing the paper. Abolfzl Ahmadi: Writing – review & editing. [30] K. Wang, X. Qi, H. Liu, A comparison of day-ahead photovoltaic power forecasting
models based on deep learning neural network, Appl. Energy 251 (2019), 113315.
[31] P. Kumari, D. Toshniwal, Deep learning models for solar irradiance forecasting: a
Declaration of competing interest comprehensive review, J. Clean. Prod. 318 (2021), 128566.
[32] A. Sharma, A. Kakkar, Forecasting daily global solar irradiance generation using
machine learning, Renew. Sustain. Energy Rev. 82 (2018) 2254–2269.
The authors declare that they have no known competing financial [33] A. Mellit, M. Benghanem, A.H. Arab, A. Guessoum, A simplified model for
interests or personal relationships that could have appeared to influence generating sequences of global solar radiation data for isolated sites: using artificial
neural network and a library of Markov transition matrices approach, Sol. Energy
the work reported in this paper.
79 (5) (2005) 469–482.
[34] H.A. Hejase, A.H. Assi, Time-series regression model for prediction of mean daily
References global solar radiation in Al-Ain, UAE, Int. Sch. Res. Notices 2012 (2012).
[35] K. Mohammadi, S. Shamshirband, C.W. Tong, M. Arif, D. Petković, S. Ch, A new
hybrid support vector machine–wavelet transform approach for estimation of
[1] M.S. Alfailakawi, S. Michailos, D.B. Ingham, K.J. Hughes, L. Ma, M. Pourkashanian,
horizontal global solar radiation, Energy Convers. Manag. 92 (2015) 162–171.
Multi-temporal resolution aerosols impacted techno-economic assessment of
[36] R. Meenal, A.I. Selvakumar, Assessment of SVM, empirical and ANN based solar
concentrated solar power in arid regions: case study of solar power tower in
radiation prediction models with most influencing input parameters, Renew.
Kuwait, Sustain. Energy Technol. Assessments 52 (2022), 102324.
Energy 121 (2018) 324–343.
[2] D. Arvizu, et al., Direct Solar Energy, 2011.
[37] C.G. Ozoegwu, Artificial neural network forecast of monthly mean daily global
[3] Y. Dong, H. Jiang, Global solar radiation forecasting using square root
solar radiation of selected locations based on time series and month number,
regularization-based ensemble, Math. Probl Eng. 2019 (2019).
J. Clean. Prod. 216 (2019) 1–13.
[4] Y. Wang, D. Millstein, A.D. Mills, S. Jeong, A. Ancell, The cost of day-ahead solar
[38] M.H. Alsharif, M.K. Younes, J. Kim, Time series ARIMA model for prediction of
forecasting errors in the United States, Sol. Energy 231 (2022) 846–856.
daily and monthly average global solar radiation: the case study of Seoul, South
[5] E. Park, K. Yoo, J.Y. Ohm, S.J. Kwon, Case study: renewable electricity generation
Korea, Symmetry 11 (2) (2019) 240.
systems on Geoje Island in South Korea, J. Renew. Sustain. Energy 8 (1) (2016),
[39] S. Ghimire, R.C. Deo, N. Raj, J. Mi, Deep solar radiation forecasting with
015904.
convolutional neural network and long short-term memory network algorithms,
[6] M.N. Akhter, S. Mekhilef, H. Mokhlis, N. Mohamed Shah, Review on forecasting of
Appl. Energy 253 (2019), 113541.
photovoltaic power generation based on machine learning and metaheuristic
[40] A. Abayomi-Alli, M.O. Odusami, O. Abayomi-Alli, S. Misra, G.F. Ibeh, Long short-
techniques, IET Renew. Power Gener. 13 (7) (2019) 1009–1023.
term memory model for time series prediction and forecast of solar radiation and
[7] V.P. Lonij, A.E. Brooks, A.D. Cronin, M. Leuthold, K. Koch, Intra-hour forecasts of
other weather parameters, in: 2019 19th International Conference on
solar power production using measurements from a network of irradiance sensors,
Computational Science and its Applications (ICCSA), IEEE, 2019, pp. 82–92.
Sol. Energy 97 (2013) 58–66.
[41] R. Prasad, M. Ali, P. Kwan, H. Khan, Designing a multi-stage multivariate empirical
[8] U.K. Das, et al., Forecasting of photovoltaic power generation and model
mode decomposition coupled with ant colony optimization and random forest
optimization: a review, Renew. Sustain. Energy Rev. 81 (2018) 912–928.
model to forecast monthly solar radiation, Appl. Energy 236 (2019) 778–792.
[9] G. Reikard, Predicting solar radiation at high resolutions: a comparison of time
[42] B. Belmahdi, M. Louzazni, A. El Bouardi, One month-ahead forecasting of mean
series forecasts, Sol. Energy 83 (3) (2009) 342–349.
daily global solar radiation using time series models, Optik 219 (2020), 165207.
[10] E. Boileau, Discussion d’un modèle statistique en météorologie solaire, Rev. Phys.
[43] F. Rodríguez, M. Genn, L. Fontán, A. Galarza, Very short-term temperature
Appl. 14 (1) (1979) 145–152.
forecaster using MLP and N-nearest stations for calculating key control parameters
[11] G. Reikard, S.E. Haupt, T. Jensen, Forecasting ground-level irradiance over short
in solar photovoltaic generation, Sustain. Energy Technol. Assessments 45 (2021),
horizons: time series, meteorological, and time-varying parameter models, Renew.
101085.
Energy 112 (2017) 474–485.
[44] S. Suradhaniwar, S. Kar, S.S. Durbha, A. Jagarlapudi, Time series forecasting of
[12] M. Diagne, M. David, P. Lauret, J. Boland, N. Schmutz, Review of solar irradiance
univariate agrometeorological data: a comparative performance evaluation via
forecasting methods and a proposition for small-scale insular grids, Renew.
one-step and multi-step ahead forecasting strategies, Sensors 21 (7) (2021) 2430.
Sustain. Energy Rev. 27 (2013) 65–76.
[45] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT press, 2016.
[13] G.P. Zhang, Time series forecasting using a hybrid ARIMA and neural network
[46] N. Elizabeth Michael, M. Mishra, S. Hasan, A. Al-Durra, Short-term solar power
model, Neurocomputing 50 (2003) 159–175.
predicting model based on multi-step CNN stacked LSTM technique, Energies 15
[14] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artificial neural networks:: the
(6) (2022) 2150.
state of the art, Int. J. Forecast. 14 (1) (1998) 35–62.
[47] X. Huang, et al., Hybrid deep neural model for hourly solar irradiance forecasting,
[15] H. Sharadga, S. Hajimirza, R.S. Balog, Time series forecasting of solar power
Renew. Energy 171 (2021) 1041–1060.
generation for large-scale photovoltaic plants, Renew. Energy 150 (2020)
[48] F.A. Gers, D. Eck, J. Schmidhuber, Applying LSTM to time series predictable
797–807.
through time-window approaches, in: Neural Nets WIRN Vietri-01, Springer, 2002,
[16] R. Kumar, R. Aggarwal, J. Sharma, Comparison of regression and artificial neural
pp. 193–200.
network models for estimation of global solar radiations, Renew. Sustain. Energy
[49] F.A. Gers, J. Schmidhuber, F. Cummins, Learning to forget: continual prediction
Rev. 52 (2015) 1294–1299.
with LSTM, Neural Comput. 12 (10) (2000) 2451–2471.
[17] L. Hontoria, J. Aguilera, J. Riesco, P. Zufiria, Recurrent neural supervised models
[50] M.V. Assis, L.F. Carvalho, J. Lloret, M.L. Proença Jr., A GRU deep learning system
for generating solar radiation synthetic series, J. Intell. Rob. Syst. 31 (1) (2001)
against attacks in software defined networks, J. Netw. Comput. Appl. 177 (2021),
201–221.
102942.
[18] M. Guermoui, A. Rabehi, S. Benkaciali, D. Djafer, Daily global solar radiation
[51] P. Li, J. Zhang, P. Krebs, Prediction of flow based on a CNN-LSTM combined deep
modelling using multi-layer perceptron neural networks in semi-arid region,
learning approach, Water 14 (6) (2022) 993.
Leonardo Electron. J. Pract. Technol. 28 (2016) 35–46.
[52] D. Cannizzaro, A. Aliberti, L. Bottaccioli, E. Macii, A. Acquaviva, E. Patti, Solar
[19] A.S. Dorvlo, J.A. Jervase, A. Al-Lawati, Solar radiation estimation using artificial
radiation forecasting based on convolutional neural network and ensemble
neural networks, Appl. Energy 71 (4) (2002) 307–319.
learning, Expert Syst. Appl. 181 (2021), 115167.
[20] A.F. Mashaly, A. Alazba, MLP and MLR models for instantaneous thermal efficiency
[53] J. Brownlee, Deep Learning for Time Series Forecasting: Predict the Future with
prediction of solar still under hyper-arid environment, Comput. Electron. Agric.
MLPs, CNNs and LSTMs in Python, Machine Learning Mastery, 2018.
122 (2016) 146–155.
[54] A.K. Pandey, P.C. Pant, O.S. Sastry, A. Kumar, S.K. Tyagi, Energy and exergy
[21] D. Skrobek, et al., Implementation of deep learning methods in prediction of
performance evaluation of a typical solar photovoltaic module, Therm. Sci. 19
adsorption processes, Adv. Eng. Software 173 (2022), 103190.
(suppl. 2) (2015) 625–636.
[22] D. Skrobek, et al., Prediction of sorption processes using the deep learning methods
[55] G. Makrides, B. Zinsser, G.E. Georghiou, M. Schubert, J.H. Werner, Temperature
(long short-term memory), Energies 13 (24) (2020) 6601.
behaviour of different photovoltaic systems installed in Cyprus and Germany, Sol.
[23] Y. Chu, M. Li, C.F. Coimbra, D. Feng, H. Wang, Intra-hour irradiance forecasting
Energy Mater. Sol. Cell. 93 (6–7) (2009) 1095–1099.
techniques for solar power integration: a review, iScience 24 (10) (2021), 103136.
[56] M. Yaghoubirad, N. Azizi, A. Ahmadi, Z. Zarei, S.F. Moosavian, Performance
[24] H. Chen, X. Chang, Photovoltaic power prediction of LSTM model based on Pearson
assessment of a solar PV module for different climate classifications based on
feature selection, Energy Rep. 7 (2021) 1047–1054.
energy, exergy, economic and environmental parameters, Energy Rep. 8 (2022)
[25] A.F. Faisal, A. Rahman, M.T.M. Habib, A.H. Siddique, M. Hasan, M.M. Khan,
68–84.
Neural networks based multivariate time series forecasting of solar radiation using
[57] A. Shahsavari, F. Yazdi, H. Yazdi, Potential of solar energy in Iran for carbon
meteorological data of different cities of Bangladesh, Results Eng. 13 (2022),
dioxide mitigation, Int. J. Environ. Sci. Technol. 16 (1) (2019) 507–524.
100365.
[58] M. Dehghani, M. Feylizadeh, An overview of solar energy potential in Iran, Int. J.
[26] I. Jebli, F.-Z. Belouadha, M.I. Kabbaj, A. Tilioua, Deep learning based models for
Curr. Life Sci. 4 (9) (2014) 7173–7180.
solar energy prediction, Adv. Sci. 6 (2021) 349–355.
[59] Solargis. https://solargis.com/.
[27] J. Wojtkiewicz, M. Hosseini, R. Gottumukkala, T.L. Chambers, Hour-ahead solar
[60] NASA, Daylight hours in zahedan city. https://power.larc.nasa.gov/data-access
irradiance forecasting using multivariate gated recurrent units, Energies 12 (21)
-viewer/.
(2019) 4055.
[61] Y. Zhao, Y. Cen, Data Mining Applications with R, Academic Press, 2013.
[28] A.P. Wibawa, A.B.P. Utama, H. Elmunsyah, U. Pujianto, F.A. Dwiyanto,
[62] Y. Zhou, Y. Liu, D. Wang, X. Liu, Y. Wang, A review on global solar radiation
L. Hernandez, Time-series analysis with smoothed convolutional neural network,
prediction with machine learning models in a comprehensive perspective, Energy
J. big Data 9 (1) (2022) 1–18.
Convers. Manag. 235 (2021), 113960.
[29] W. He, Load forecasting via deep neural networks, Procedia Comput. Sci. 122
(2017) 308–314.

146
N. Azizi et al. Renewable Energy 206 (2023) 135–147

[63] A. Azadeh, A. Maghsoudi, S. Sohrabkhani, An integrated artificial neural networks ANFIS method: application in the west-Central Jordan, Future Internet 14 (3)
approach for predicting global radiation, Energy Convers. Manag. 50 (6) (2009) (2022) 79.
1497–1505. [65] H. Zhou, Z. Deng, Y. Xia, M. Fu, A new sampling method in particle filter based on
[64] H. Fraihat, A.A. Almbaideen, A. Al-Odienat, B. Al-Naami, R. De Fazio, P. Visconti, Pearson correlation coefficient, Neurocomputing 216 (2016) 208–215.
Solar radiation forecasting by Pearson correlation using LSTM neural network and

147

You might also like