International Journal of Sustainable Engineering

ISSN: 1939-7038 (Print) 1939-7046 (Online) Journal homepage: www.tandfonline.com/journals/tsue20

PV power forecasting based on data-driven

models: a review

Priya Gupta & Rhythm Singh

To cite this article: Priya Gupta & Rhythm Singh (2021) PV power forecasting based on data-
driven models: a review, International Journal of Sustainable Engineering, 14:6, 1733-1755,
DOI: 10.1080/19397038.2021.1986590

To link to this article: https://doi.org/10.1080/19397038.2021.1986590

Published online: 06 Oct 2021.

Submit your article to this journal

Article views: 9914

View related articles

View Crossmark data

Citing articles: 97 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsue20
INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING
2021, VOL. 14, NO. 6, 1733–1755
https://doi.org/10.1080/19397038.2021.1986590

PV power forecasting based on data-driven models: a review

Priya Gupta and Rhythm Singh
Department of Hydro and Renewable Energy, Indian Institute of Technology Roorkee, Roorkee, India

ABSTRACT ARTICLE HISTORY

Accurate PV power forecasting techniques are a prerequisite for the optimal management of the grid and Received 18 January 2021
its stability. This paper presents a review of the recent developments in the field of PV power forecasting, Accepted 21 September 2021
mainly focusing on the literature which uses ML techniques. The ML techniques (sub-branch of artificial KEYWORDS
intelligence) are extensively used due to their ability to solve nonlinear and complex data structures. PV PV power forecasting;
power forecasting can either be direct, or indirect, which involves solar irradiance forecast model, plane forecast horizon; data-driven
of array irradiance estimation model, and PV performance model. This paper presents a review of both of models; machine learning;
these pathways of PV power forecasting based on the proposed methodology, forecast horizons and the deep learning; ensemble
considered input parameters. In case of unavailability of historical PV power for a new PV plant and in methods; solar radiation
case of failure of real-time data acquisition, indirect PV power forecasting can be a viable alternative. forecasting; POA irradiance;
Although the performance ranking of various ML models is complicated and no model is universal, recent PV performance models
studies suggest that methodologies like deep neural networks and ensemble or hybrid models outper­
form conventional methods for short-term PV forecasting. Recent articles also present the various
intelligent optimisation and data-preparation techniques to improve performance accuracy.

1. Introduction
A balance between generation and consumption has become a
Conventional energy sources are losing importance because of great challenge for electrical operators. Voltage fluctuations,
global warming and the rapid depletion of fossil fuels. Non- low power quality, and low stability issues are some other
conventional energy sources such as solar, wind, bio-mass, problems that arise due to the uncontrollable nature of solar
etc., are free from the above problems, and thus, are poised energy. Variability, uncertainty and asynchronous operations
to play a significant role in the future. Peters et al. (Peters and are the main technical challenges electrical operators have to
Buonassisi 2000) have analysed the effect of global warming on deal in integrating renewable energy sources with the power
the performance of PV installations in the future. Solar energy grid (Kroposki 2017). So, for optimal management of the
is becoming one of the most promising sources to supply electrical grid, accurate forecasting of solar PV power is
power for residential, commercial and industrial loads (Ren, required (Paulescu et al. 2013). Solar power forecasting is
Suganthan, and Srikanth 2015). Nevertheless, one of the major also required for scheduling, approximating the reserves, deal­
problems with renewable energy sources is their unpredictable ing generated electrical power, better operation of the power
nature. PV plants generate different amounts of electricity grid, reducing the cost of produced electrical energy and con­
based on the variation of solar radiation and other meteorolo­ gestion management. As the penetration of solar PV in the grid
gical variables (Ismail et al. 2015). There can be significant increases, the prediction of solar power also becomes more
variations in the power output not only over different days, but critical due to the above-mentioned problems in the power
also from one hour to another, or even over few minutes (Kasu system. Researchers also suggest using storage systems with
and Hara 2017). This has the potential to lead to load-genera­ renewable energy prediction to control electricity variation.
tion mismatch in the grid, thus rendering PV power forecast­ Storage systems absorb excess power, dampen the fluctuations
ing very important, particularly in grids with high penetration and maintain a continuous flow of electricity. This paper aims
of PV (Espinar et al. 2013). to discuss and compare different forecasting techniques to
estimate the PV power output in two different ways, i.e. (i)
direct forecasting that predicts the power directly by using
1.1. Challenges and the need to forecast solar PV power
historical data of PV power and (ii) indirect forecasting,
The integration of renewable energy sources with the electrical which uses solar irradiation forecasting, along with other
grid has been gaining more importance and is also creating meteorological variables that directly affect solar PV power
more challenges for electrical engineers and researchers. The production (Rana, Koprinska, and Agelidis 2016).
intermittent and uncontrollable nature of solar energy A number of techniques for PV power forecasting have
increases the complexity of grid management and adds to the been proposed based on physical assumptions of the atmo­
difficulty in balancing the generation and consumption of sphere. NWP is the most commonly used physical model of
electrical energy (Lara-Fanego et al. 2012; Raza et al. 2016). forecasting. The variability in the atmosphere and its

CONTACT Rhythm Singh [email protected] Department of Hydro and Renewable Energy, Indian Institute of Technology Roorkee, Roorkee,
Uttarakhand 247667, India
© 2021 Informa UK Limited, trading as Taylor & Francis Group
1734 P. GUPTA AND R. SINGH

unpredictable nature make the numerical weather prediction to predict radiation for N-step ahead forecast horizon. Data-
model computationally severe and complex (Sheng et al. 2018). driven models are another category of forecasting. These mod­
Due to the growth in the computer science field and its ability els are based on the extraction of useful information from the
to deal with non-linearity, the popularity of machine learning input training data, and based on this information, they pre­
techniques is increasing. Apart from direct and indirect PV dict the output. The performance of these methods is suscep­
power forecasting using ML techniques, this paper also tible to the quality of the training data. After a lot of research in
reviews, classifies and compares various plane of array irradi­ this field, it has been concluded that machine learning-based
ance estimation models and PV performance models in the models show better performance for short-term forecast hor­
field of PV power forecasting. izons (readers can refer to the references mentioned in Section
This paper will help researchers working in the field of solar 3 of this article). This section presents a detailed overview of
energy by providing useful information on PV power forecast­ data-driven models, plane of array irradiance estimation mod­
ing. The sections of this paper are organised as follows- Section els, and PV performance models in the field of PV power
2 presents an overview of PV power forecasting methodolo­ forecasting. (Figure 1) shows the block diagram of PV power
gies, Section 3 presents a detailed literature review in the field forecasting.
of solar power forecasting, Section 4 presents the development
of literature on PV power forecasting, and finally, Section 5
2.1. Data-driven models
concludes the complete work.
As the name suggests, data-driven models use historical data.
Models are trained, or extract some useful information from
2. Available PV power forecasting methodologies
the dataset, to predict the output. Length and quality of histor­
Solar power forecasting can be divided into two groups- (1) ical data are two main factors that affect the accuracy of
Direct forecasting, that directly forecasts PV power. (2) forecasting. Time series and ML models are the two main
Indirect forecasting, that is based on solar radiation forecast­ categories of data-driven models. (Figure 2) represents the
ing. Several researchers have been working in this domain over classification diagram.
the last few years. However, still, the accuracy of short-term
forecasting is a challenging task for the researchers. Solar 2.1.1. Time-series models
radiation/PV power forecasting is a non-linear problem, Time series is a part of statistics. Four researchers Slutsky,
which depends upon several weather parameters. For a non­ Walker, Yaglom and Yule, first composed the idea of time
linear system, finding the proper parameter estimation method series models such as auto regressive (AR), integrated (I) and
is a difficult task. Several methods have been proposed for moving average (MA) (Brockwell and Davis 2006). The com­
forecasting in the literature; two major categories are (i) phy­ bination of these three concepts produces ARMA and ARIMA
sical models and (ii) data-driven models. Choice of forecasting models. These are the simplest models in the field of times
model depends upon the forecast horizon and the selected series forecasting. These models use historical data points of
location. Numerical weather prediction is the most commonly the dependent variables, known as endogenous forecasting.
used physical model. It uses a dynamic atmosphere model (to Exogenous forecasting utilises a historical dataset of indepen­
use cloud motion and direction) and mathematical equations dent variables other than the dependent variable. For

Figure 1. Block diagram of PV power forecasting: (a) direct forecasting and (b) indirect forecasting.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1735

Time series Classification and data-

models ex: preparation ex: PCA, SOM
ARMA, ARIMA
Supervised learning ex: SVM,
Data-driven ANN, nearest neighbor,
models Regression, etc

Unsupervised learning ex:

Clustering
Machine learning
models
Ensemble learning ex:
Boosting, Bagging

Deep learning ex: RNN,

convolutionl network etc

Figure 2. Classification of data-driven models.

exogenous time series forecasting, these models are termed as kind of ANN technique used to convert the input vectors in
ARMAX and ARIMAX. There are some limitations associated low dimensional representation and is therefore termed
with the traditional ARMA and ARIMA models: dimensionality reduction technique.
Wavelet transform (WT): It is a potent tool to decompose a
(1) Missing values in the data set highly affect the accuracy time series in both time and frequency domain. It is handy for
of these models. extracting some hidden local information from the time-series
(2) Solar radiation and power forecasting is a non-linear data. It is more powerful than the Fourier transform. It is not
problem, but these models cannot effectively deal non- an ML technique, but applying wavelet transform with various
linearity of the problems (Voyant et al. 2017b). ML models, including SVM, ANN (can deal with non-linearity
(3) These models are usually applied for univariate fore­ of series) in the classification phase, improves the forecasting
casting applications. But many problems in the real results effectively (Eseye, Zhang, and Zheng 2018; Wen et al.
world depend upon more than one input variable. 2019).
(4) These models usually don’t work well for the multi-step Supervised learning: Supervised learning-based techniques
forecast horizon. learn from the training data set, which consists of some inputs
and their desired outputs, to establish a general mapping rule
2.1.2. Machine learning models for finding a relation between the inputs and the outputs. Some
ML is a sub-branch of computer science and also a sub-field of supervised learning-based ML techniques are as follows.
artificial intelligence techniques. Machine learning utilises his­ Regression: Regression analysis determines the functional
torical data points to establish a relation between the input and relationship between variables and forecasting. When only one
output, even if the relationship between them is complex. So it parameter is required, then the regression will be univariate,
is recommended to use proper data or prepare them to solve and in case two or more variables are involved, regression will
the problem effectively. Machine learning comprises a gamut be termed as multivariate. The type of regression model
of techniques; some of these have been discussed in the depends on the complexity or relationship between dependent
remainder of this section as follows. and independent variables and the number of independent
Data-preparation and classification techniques: For the variables. Two common types of regression techniques are
pre-processing of input data, data preparation and classifica­ SLR and MLR (Sobri, Koohi-kamali, and Rahim 2020).
tion techniques are utilised with the forecasting model. Some Artificial neural network (ANN): Neural network is an AI
of them are described as follows- technique that resembles the brain in two aspects. It can deal
Principal component analysis (PCA): It is a linear trans­ with both linear and non-linear regression and classification
formation technique used for dimensionality reduction and problems. NN is usually a combination of three layers (i.e.
feature extraction. This method transforms correlated vari­ input, hidden and output layer), activation function and sum­
ables into mutually uncorrelated variables using some ortho­ mation node.
gonal transformation. The principal components obtained Xm � Xd
from the Eigen vector of the covariance matrix can be lower y ¼ yðx; wÞ ¼ a0 þ j¼1
w j f w x
i¼1 ji i
(1)
or equal to the original variables. Most of the variance is
achieved by the first principal components that indicate a Equation (1) defines a standard mapping function for input x
high correlation between input variables (Alskaif et al. 2019). and output y with d inputs and m hidden layers. MLP is the
Self-Organising map (SOM): Self-organising map is a fea­ most commonly used structure of NN that is formed by an
ture extraction technique based on unsupervised learning. The input layer, one or more hidden layers, and one output layer.
technique is sometimes termed Kohonen map because it was Care must be taken while selecting the NN structure, as a too
first developed by professor Kohonen (Yang et al. 2014). It is a complex structure of NN can easily overfit the training data
1736 P. GUPTA AND R. SINGH

(Voyant et al. 2017b). The complexity of the neural network K-means clustering: Clustering is one of the most com­
depends upon the number of hidden layers and the dimension monly used data analysis techniques to find the subgroup in
of weights (w). Network topology depends on the connection the data points to form clusters. The data points in each cluster
of neurons, and for better functioning, it is an important are of similar nature. K-means clustering is an unsupervised
factor. Researchers are using hybrid methods for selecting ML technique. The main aim of this technique is to extract
NN topology that allows the neural network to achieve more useful information and relations between inputs in the form of
gain and perform at its maximum capacity. clusters. These will then be useful to predict the output in the
Support vector machine (SVM): SVM is a kernel-based ML forecasting problems.
technique used in both classification and regression tasks. The Hierarchical clustering: This is also a type of unsupervised
idea of a support vector machine was first introduced by Vapnik technique that builds the hierarchy of clusters in a tree struc­
(Cortes and Vapnik 1995). These models require some kernels, ture called ‘dendrogram’. A dendrogram is a tree structure
can make fast computations, avoid overfitting of training data, consisting of roots and leaves (Voyant et al. 2017b). Roots
and have good convergence. SVM models are simpler to imple­ contain all the information of the clusters, and leaves corre­
ment and faster in computation compared to neural networks. spond to individual information.
These are the reasons that SVM models are preferred for time Ensemble learning: Ensemble learning is a hybrid ML
series forecasting problems. Based on the type of kernel, SVM technique that uses weak learners to predict the output and
models are classified into L-SVM and K-SVM. L-SVM models then combines these outputs to obtain a final prediction.
use linear kernel functions, whereas K-SVM models use a non- Nowadays, researchers prefer ensemble models because they
linear kernel function and can transform the data into higher do not generate output by the estimation of only one model;
dimensional space. Prediction calculated by SVM can be given by: instead, it combines the output of many weak learners. The
Xn � reason that researchers are focusing on the use of ensemble
y ¼ i¼1
ai krbf ðxi ; x�Þ þ b (2) models is that these models improve the stability and accuracy
of single-stage machine learning algorithms. Depending on
Equation (2) represents the mathematical representation of how the weak learners are connected, two types of ensemble
predicted output value for the test case input x � , b is the bias learning models are formulated: Boosting and Bagging.
parameter, ai is the coefficient and xi is the training set input Boosting: Boosting is also an ML technique used in both
vector at ith instance with total n training vectors. classification and regression tasks. Boosting takes the advan­
And the commonly used radial basis function kernel is tage of some weak learners known as decision trees to predict
given by: the output. Boosting predicts the output in a stage-wise man­
� � ner, or in a serial form by optimising some loss function
� ðxp xq Þ
krbf xp ; xq ¼ exp (3) (Cheng 2016). Boosting combines many weak learners and
2σ 2
gives more preference to a bad classifier to form a strong
The radial basis function is the most commonly used kernel learner. Gradient boosting is one of the examples of boosting
in the support vector machine regression and classification technique.
problems (Fan et al. 2018). Equation (3) represents a general Bagging: Bagging is a machine learning algorithm that is
mathematical form of this type of kernel function. The ðxp also gaining importance in the field of forecasting. The basic
xq Þ represents the relation between two feature vectors in idea behind bagging is to generate multiple versions of deci­
terms of the squared Euclidean distance, and σ is the free sion trees or weak learners and get an aggregated strong pre­
parameter. The Gaussian function, polynomial kernel and dictor by combining the output of each decision tree. Weak
sigmoid kernel are some other kernel functions. learners can be formed either by taking one ML model with
Nearest neighbour: The k-NN neural network is one of the different sets of training data or by taking different ML fore­
simplest types of ML algorithm that is also used for both casting models. Bagging is a unique form of averaging fore­
classification and regression purposes. The k-NN algorithm casting technique (Bbeiman 1996). Random forest is one of the
assigns weights to all the neighbours. The weight of each examples of bagging techniques.
neighbour is inversely proportional to the distance of the Deep learning (DL): Deep learning is a sub-class of
neighbour to its centroid (Pedro and Coimbra 2015). It machine learning and can be supervised or unsupervised.
means the weight assigned to the minimum distance neigh­ Due to the ability of artificial neural networks to deal with
bour will be the highest. the non-linearity of the problem, researchers are focusing on
Unsupervised learning: Unsupervised learning does not the further development of ANN methods. Deep learning is a
need expert involvement to extract useful information from more advanced form of neural network with a memory unit
the input data, unlike supervised learning. This is why unsu­ and can apply in many fields, including forecasting. Some
pervised learning algorithms cannot be used directly for models based on supervised deep learning like recurrent
regression and classification problems. Unsupervised learn­ neural networks (Guan, Zhang, and Yan 2019), deep belief
ing-based models can find the hidden structure or useful networks, convolutional neural networks have started to pro­
information from the input data without knowing the output. duce excellent results in the field of solar energy forecasting.
Clustering is one of the crucial techniques of unsupervised Compared to other forecasting methods like NWP physical
learning to preprocess the data in the form of clusters of model, persistence and other statistical methods, DL-based
similar information. Some clustering techniques are as models can effectively extract deep features and information
follows- from the time series.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1737

2.2. Plane of array irradiance estimation models The only difference in various transposition models lies in
the sky-diffuse component. The sky-diffuse component of
The plane of array irradiance estimation model comprises two
POA irradiance depends on the amount of sky seen from the
empirical models, i.e. decomposition and transposition mod­
surface (Cui et al. 2019).
els. These two models are termed as combination models.
The reflected ground diffuse POA irradiance component is
(Figure 3) shows the structure diagram to the model plane of
a function of GHI, ground albedo (ρ) and surface tilt from
array irradiance using modelled GHI which can be obtained by
horizontal (β), as given in Equation (6).
utilising various data-driven and physical models.
1 cos β
Etg ¼ GHI � ρ � (6)
2.2.1. Decomposition models 2
Decomposition models are used to decompose the GHI data
into direct normal and diffuse horizontal irradiance. Two
2.3. PV performance models
substantial physical quantities in this context are clearness
index(Kt) and a diffuse fraction(DF) (Lave et al. 2015). The 2.3.1. PV cell model
accuracy of decomposition models depends upon the relation­ To establish PV power forecasting, PV cell model is the core
ship of the clearness index and the diffuse fraction. These component. The voltage-current equation of the single diode
models are empirical because their equations are not formally PV cell model (Huang et al. 2010) can be represented by
acquired from physical rules but involve some coefficients. Equation (7).
These coefficients are derived by collecting solar irradiance � � � �
data from various locations. qðV þ IRs Þ ðV þ IRs Þ
I ¼ Ip I0 exp 1 (7)
nkT Rsh

2.2.2. Transposition models Where Ip is the current generated from light, Io is the reverse
Transposition models transpose direct normal and diffuse saturation current, k is the Boltzmann constant, n is the ide­
horizontal irradiance into global solar irradiance on a tilted ality factor, q is the charge, Rs is the series resistance, Rsh is the
plane. Plane of array (POA) irradiance(Et) is divided into the shunt resistance, and T is the absolute temperature. The single
following components: (1) direct normal component of POA diode model is easy to implement. For a more accurate fore­
(Etb), (2) sky diffuse component of POA (Etd) and (3) cast, researchers also suggest the use of the two diode model.
reflected ground-diffuse component of POA (Etg). The equa­ The two diode model is comparatively complex but can pro­
tion of POA can be represented as in Equation (4). vide more precise results. Equation (8) represents the current-
voltage equation of the two diode models (Bouralou et al.
Et ¼ Etb þ Etd þ Etg (4) 2015).
� � � �
qðV þ IRs Þ
The direct irradiance incident on POA can be calculated I ¼ Ip I01 exp 1
� � kT � �
directly by direct normal irradiance, as given in Equation (5). qðV þ IRs Þ ðV þ IRs Þ
I02 exp 1 (8)
2kT Rsh
Etb ¼ DNI � cos ðAOI Þ (5)
The equivalent circuit of the two diode model is the same as
Where AOI is the angle of incidence of solar radiation on the one diode model with an additional parallel-connected
POA and DNI is direct normal irradiance on a horizontal diode. I01 and I02 are the reverse saturation current in the
plane. first and the second diodes, respectively.

Figure 3. Structure to model plane of array irradiance.

1738 P. GUPTA AND R. SINGH

2.3.2. Inverter model Sark 2019), (Dewangan, Singh, and Chakrabarti 2020), Deep
PV array and inverters are the main components of the grid- learning – (Chang and Lu 2020), (P. Li et al. 2020), (Zheng et
connected power system. Therefore, accurate simulation of an al. 2020), (Ali Jallal, Chabaa, and Zeroual 2020), (Yin et al.
inverter is the critical deciding factor for accurate forecasting. 2020), (Dewangan, Singh, and Chakrabarti 2020) A summary
Few inverters can achieve the efficiency specified on the inver­ of the literature on direct PV power forecasting is given in
ter datasheet because of the loss associated with the inverter (Table 1). Readers can refer to these references for more
components. Sandia National Laboratory proposed an inverter detail.
which considers the loss associated with the inverter compo­ It can be seen from (Table 1) that artificial neural network
nents and also shows the high accuracy (King, Gonzalez, and (ANN), and support vector machine (SVM) are the two
Galbraith 2007). Authors in (Cui et al. 2019; Liaoliao et al. most often used machine learning algorithms in the field of
2017) consider this inverter model for the application of PV PV forecasting. Several studies in the area of ANN validate
power forecasting. the better non-linear fitting ability of neural networks
compared to time series models. Broad research in neural
networks, from an early simple architecture to a late deep
3. Literature review
configuration, results from the better performance of these
In this section, we have analysed the state-of-the-art of litera­ networks. Sharadga, Hajimirza, and Balog (2020) recently
ture on solar PV power forecasting. The literature review has presented a new deep learning model Bi-LSTM (Bi-direc­
been organised into two groups, i.e. direct forecasting and tional LSTM), for PV power forecasting. After comparing
indirect forecasting. the results of various structures of neural networks (NNs),
and time-series models (ARMA, ARIMA, and SARIMA),
the prediction accuracy of NNs was reported higher with
3.1. Direct forecasting
less computational time. Unlike recurrent neural network
As discussed previously that ML techniques need some his­ (RNN), LSTM contains a memory unit that helps to keep
torical data for the training of the model, so, for direct PV the long spans data and can also solve the gradient descent
power forecasting, the availability of such historical PV problem. That’s why LSTM is helpful to extract the tem­
power data is an essential requirement. This can be possible poral information from the time-series data. Similarly, a
if the PV plant is old and equipped with data loggers and a deep belief network (DBN) has been presented by Chang
SCADA system. Papers based on data-driven models classi­ and Lu (2020) to learn the non-linear features from the
fication in Section 2 for the direct PV power forecasting, are previous PV power time-series data.
categorised as follows. The core of machine learning algorithms is the selection
Time series models – (Pedro and Coimbra 2012), (Liu, of suitable hyperparameters values that impact forecasting
Zhan, and Bai 2019), (Yin et al. 2020) Classification and accuracy. For example, C and gamma are the two impor­
data-preparation – (Yang et al. 2014), (Eseye, Zhang, and tant parameters of SVM, whose inappropriate values are
Zheng 2018), (Wen et al. 2019), (Chen et al. 2011), (Haque, responsible for the overfitting and underfitting issues.
Nehrir, and Mandal 2013), (Zhu et al. 2016), (Chang and Lu Presently, scholars are adopting intelligent optimisers for
2020), (P.Li et al. 2020), Supervised learning (1) NN- (Wen et hyperparameter tuning of machine learning algorithms.
al. 2019), (Chen et al. 2011), (Cococcioni, D’Andrea, and Pan et al. (2020) reported the improvement in R2 score
Lazzerini 2011), (Pedro and Coimbra 2012), (J. Liu et al. (coefficient of determination) from 0.991 to 0.997 by utilis­
2015), (Ramsami and Oree 2015), (Zhu et al. 2016), (Hossain ing an improved ACO to optimise SVM parameters.
et al. 2017), (J.Wang, Ran, and Zhou 2017), (Gigoni et al. Similarly, Zhou et al. (2020) adopted a genetic algorithm,
2018), (Raza, Mithulananthan, and Summerfield 2018), and Liu et al. (2020) adopted an improved chicken swarm
(Rosiek, Alonso-Montesinos, and Batlles 2018), (Yadav et al. optimisation to tune the hyperparameters of extreme learn­
2018), (Visser, Alskaif, and Van Sark 2019), (Sharadga, ing machine (ELM), and both reported a better forecast
Hajimirza, and Balog 2020), (Yin et al. 2020), (Dewangan, accuracy with the incorporation of optimisation algorithms.
Singh, and Chakrabarti 2020), (Zhou et al. 2020), (Z.F. Liu et Further, Zheng et al. (2020) and Ali Jallal, Chabaa, and
al. 2020) (2) SVM- (Rana, Koprinska, and Agelidis 2016), Zeroual (2020) presented deep learning-based models for
(Yang et al. 2014), (Eseye, Zhang, and Zheng 2018), (Shi et PV power forecasting by optimising the parameters with
al. 2012), (Gigoni et al. 2018), (L.Liu, Zhan, and Bai 2019), PSO and RODDPSO techniques, respectively.
(Visser, Alskaif, and Van Sark 2019), (VanDeventer et al. Besides this, researchers also combine different classifica­
2019), (Dewangan, Singh, and Chakrabarti 2020), (Pan et al. tion and data-preparation techniques before building any
2020) (3) Regression – (Sheng et al. 2018), (Almeida, machine learning-based forecasting model to ensure the
Perpiñán, and Narvarte 2015), (Ramsami and Oree 2015), extraction of reliable information from the time-series data.
(Wang, Su, and Shu 2016), (Gigoni et al. 2018), (L. Liu, As shown in (Table 1), Chang and Lu (2020) applied a grey
Zhan, and Bai 2019), (Visser, Alskaif, and Van Sark 2019) (4) theory on the raw data of PV power before building deep
Nearest neighbour – (Pedro and Coimbra 2012), (Gigoni et al. belief network to convert the irregular data into new data
2018), Ensemble learning (1) Boosting – (Visser, Alskaif, and with strong regularity. Similarly, Li et al. (2020) involves
Van Sark 2019) (2) Bagging – (Rana, Koprinska, and Agelidis applying wavelet packet decomposition to model the non-
2016), (El-Baz, Tzscheutschler, and Wagner 2018), (Gigoni et linearity of the original time-series data by decomposing the
al. 2018), (L.Liu, Zhan, and Bai 2019), (Visser, Alskaif, and Van signal into several sub-series. While comparing the results of
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1739

Table 1. Summary of literature review on direct PV power forecasting.

Forecast Data
Ref., location horizon duration Input variables Methodology Findings
(Chen et al. 2011) 1-d 1 year PV power, weather forecasted data RBFNN, SOM, K-fold ● The weather classification factor improves
China (solar irradiance, air temp., wind cross-validation the performance of the neural network.
speed and direction, cloud, sunshine
hours, air pressure, and humidity
(Cococcioni, 1-d 1 year Solar irradiance and time as FFNN ● The proposed model obtains errors less
D’Andrea, and exogenous, PV power as than 5% for 1-d ahead horizon and can
Lazzerini 2011) endogenous be extended for a broader forecast horizon.
Italy
(Pedro and Coimbra 1-h, 2-h 1 year PV power Persistent, ARIMA, K- ● GA-ANN improves the performance by
2012) USA 8 months NN, ANN, GA-ANN optimising the input variables.
(Shi et al. 2012) 1-d 10 months PV power, predicted daily temperature SVM model with ● The next day weather data report helps the
China from the weather report weather SVM model to obtain good results.
classification
(Haque, Nehrir, and 12-h NA Historical PV power, solar irradiance WT+fuzzy ARTMAP (FA) ● FF algorithm optimises the parameters of
Mandal 2013) +FF the FA model.
USA
(Yang et al. 2014) 1-d 1 year Historical PV power, temperature, SOM+LVQ SVR+Fuzzy ● SOM and LVQ classifiy the PV output power
China precipitation, solar irradiance inference method patterns into different weather types by
extracting some features.
(Almeida, Perpiñán, 1-d 1 year AC PV power and meteorological Non-parametric model, ● Forecasted meteorological variables from
and Narvarte variables quantile regression NWP are utilised in a non-parametric PV
2015) Spain forest model.
(J. Liu et al. 2015) 24-h NA PV power, meteorological data BP neural network ● The aerosol index reduces the prediction
China (temperature, humidity, wind error under cloudy days.
speed), Aerosol index
(Ramsami and Oree 24-h 11 months PV Power, air temperature, pressure, Single-stage GRNN, ● ANN-based models gave a better perfor­
2015) Mauritius humidity, rainfall, radiation, wind MLR, FFNN and with mance than MLR schemes.
direction, and speed SR
(Rana, Koprinska, 5– NA PV power, irradiance, wind data, Ensemble neural ● A univariate model with correlation-based
and Agelidis 60 min temperature, humidity network, SVR, variable selection, ensemble neural net­
2016) Australia correlation-basedwork, and with non- iterative approach
gives the best results for short-term fore­
variable selection
casting horizon.
(Wang, Su, and Shu 1-d 1 year PV power, solar insolation, pressure RPFLRM ● Regularised PFLRM is simpler and better
2016) China 4 months than PFLRM.
(Zhu et al. 2016) 1-d NA PV power Wavelet decomposition ● A hybrid model of WD and ANN predicts
China and ANN hybrid power with more accuracy and less com­
model putation time than the ANN model.
(Hossain et al. 2017) 1-h, 1-d 1 year PV power, Solar irradiance, wind speed, Extreme learning ● ELM model is an efficient machine learning
Malaysia ambient temperature, module machine (ELM) approach with high computational speed.
temperature
(Lave et al. 2017) NA 1 year Measured and forecasted data of PV Experimental ● The authors identified spatial and temporal
USA power, load, net load patterns in forecast errors to improve the
accuracy of the various ML forecasting
tools.
(J. Wang, Ran, and 1-h NA PV power, Solar irradiance ELM, OS-ELM, FOS-ELM ● FOS-ELM algorithm shows higher accuracy
Zhou 2017) China (ELM with forgetting followed by ELM and OS-ELM algorithm.
mechanism
(El-Baz, 1-d 1 year PV power, forecasted weather data Clear sky model, ● The deviation of estimated PV power and
Tzscheutschler, (temp., wind speed and direction, bagging regression the actual PV power measurements neces­
and Wagner 2018) humidity, and cloudiness factor) tree, probabilistic sitates the need for probabilistic analysis.
Germany analysis
(Eseye, Zhang, and 1-d 1 year SCADA PV power, NWP meteorological Wavelet decomposition ● The proposed method is suitable for short-
Zheng 2018) + SVM+PSO hybrid term forecasting. It is found that as the
China method forecast horizon increases, the accuracy of
the model decreases.
(Gigoni et al. 2018) 24-h 2 years Global, diffuse, and direct radiation GB (grey box model), ● Ensemble models provide good results
Italy 7 months data, temperature data QRF, K-NN, ANN, SVR, under all types of weather conditions.
ensemble of models
(Raza, 1-d 1 year PV power, meteorological data (air FNN, ELM, NewCF ● NewCF ensembles performed well as com­
Mithulananthan, temperature, humidity, wind speed, models LM, SCG, BP pared to FNN and ELM ensemble models.
and Summerfield solar irradiance) algorithms
2018) Australia
(Rosiek, Alonso- 3-h 1 year Global, direct, diffuse irradiance, wind Artificial neural ● The authors proposed a new methodology
Montesinos, and speed and relative humidity, power networks (ANN) with to forecast PV power using satellite-based
Batlles 2018) output, air temperature, cloud satellite images forecasted global irradiance.
Spain images
(Sheng et al. 2018) short term 6 days PV power, f photosynthetic active Weighted gaussian ● WLOF outperforms the other comparison
Spain forecast radiation, ambient temp., humidity, process regression models by extracting features from a high-
solar irradiance, precipitation and (WLOF) dimensional dataset.
wind speed, and direction
(Yadav et al. 2018) 24-h 1 year Solar radiation (SR), Back surface RBFNN, Relief attribute ● SR found more relevant for predicting daily
India module temperature evaluator algorithm array yield with rank value 0.0077.
(Continued)
1740 P. GUPTA AND R. SINGH

Table 1. (Continued).
Forecast Data
Ref., location horizon duration Input variables Methodology Findings
(Seyedmahmoudian 1 h, 2 h, 2100 hourly PV power, solar irradiance, air DEPSO (hybrid of ● The proposed model outperformed the sin­
et al. 2018) and 4 h samples temperature, relative humidity differential evolution gle-stage models (DE and PSO) for all con­
Australia for and PSO) sidered forecast horizons.
training
(J. Huang et al. 2019) 10 sec – 1 year Time series of solar PV power, sky Statistical and sky cam ● A combination of statistical and sky cam-
Australia 10 min images based model based approaches improves the forecast­
ing accuracy compared to an individual
approach.
(L. Liu, Zhan, and Bai 1-d 5 years Historical PV power, Wind speed, and Ensemble of SVM, MLP, ● The ensemble model performed better in
2019) Australia direction, weather temperature, GHI, MARS both the training and forecasting stage
humidity, DHI compared to single-stage models.
(Visser, Alskaif, and 1-d 3 years Past PV power generation, weather MLR, LASSO, L-SVM, K- ● GB and RF outperform both on individually
Van Sark 2019) variables (pressure, temperature, SVM, GB, RF, FFNN and aggregation of 152 PV sites.
Netherland cloud cover, wind vectors, clear sky ● Aggregation of PV sites improves the result
radiation, precipitation) by reducing the variability effect.
(Wen et al. 2019) 1-h, 3-h, 1 year PV power, solar irradiance, temperature WT+RBFNN+PSO (point ● Probabilistic analysis is helpful to deter­
Oregon 6-h forecast) with mine the uncertainty in the results of
bootstrap and QR as point forecasts.
the probabilistic ● The Bootstrap probabilistic method outper­
forecast forms the QR probabilistic method.
(VanDeventer et al. 1-h 278 days for PV power, solar irradiance, air GASVM (genetic ● The involvement of GA with SVM has
2019) Australia training temperature algorithm-based reduced the RMSE and MAPE by
SVM) 669.624 W and 98.76%, respectively.
(Akhter et al. 2019) NA NA PV power and weather conditions Metaheuristic and ● However, the study suggests that meta­
Malaysia machine learning heuristic models combine different techni­
models ques and show improved forecast
(Review article) accuracy, but a tradeoff is also necessary
between accuracy and other constraints
(cost and computational complexity).
(Chang and Lu 2020) 1-d 1 year PV power GT-DBN (grey theory- ● The proposed model mitigates over-fitting
Taiwan based deep belief problems by smoothing input data series.
neural network)
(P. Li et al. 2020) 1-h 2 years PV power, GHI, DHI, ambient WPD-LSTM ● The proposed model outperforms LSTM,
China temperature, wind speed, relative GRU, RNN, and MLP models
humidity
(Zheng et al. 2020) 30-min 1 year PV power, sunshine, temperature, wind PSO-LSTM ● PSO has been used to optimise the para­
China speed, humidity meters of LSTM.
The proposed model outper­
forms LSTM, ANN, and extreme
gradient boosting.
(Ali Jallal, Chabaa, 1-h 2 years PV energy, relative humidity, wind DNN-RODDPSO ● The proposed model shows high perfor­
and Zeroual 2020) speed and direction, GHI, DHI, mance with the integration of a selective
Morocco rainfall, temperature layer and an automatic input selection
hidden layer.
(Sharadga, 1,2,3 h 13/01/2010 PV power BI-LSTM, LSTM, FFNN, ● Neural network shows more accuracy and
Hajimirza, and ahead to 29/10/ MLP, LRNN, ARMA, less computation time compared to statis­
Balog 2020) USA 2010 ARIMA, SARIMA tical models for time-series prediction.
(Yin et al. 2020) 1-d 1 year PV power, solar irradiance, wind speed, ELM with the non- ● The proposed non-iterative correction
China temperature, humidity iterative correction algorithm helps to make the real-time cor­
method rection and to reduce the forecast error.
(Dewangan, Singh, 1-d 2 years NWP-derived variables like solar Persistence, linear ● The authors proposed a combined forecast
and Chakrabarti irradiance, cloud cover, regression, FFNN, methodology based on mean, median, lin­
2020) India precipitation, relative humidity, LSTM, SVR, GPR, ear, and non-linear regression to determine
pressure ensemble trees the best model from the various similar
level (performing) models.
(Zhou et al. 2020) 1-d January Max., min and average ambient SDA-GA-ELM (similar ● The proposed model outperformed the
China 2017- temperature, average GHI, and DHI day analysis-Genetic SVM, SDA-SVM, BPNN, persistence, SDA-
October algorithm-Extreme BPNN, ELM and SDA-ELM models for the
2018 learning machine) same dataset.
(Z.F. Liu et al. 2020) 5 min _ PV power, wind speed, radiation ICSO-ELM (Improved ● The proposed ICSO-ELM model is tested for
China intensity, temperature chicken swarm three weather conditions, and the results
optimiser-Extreme were found better than other ML models,
learning machine) including CSO-ELM.
(Pan et al. 2020) 5-min 2 years PV power, temperature, relative SVM-IACO (Improved ● The authors present some improvements
China humidity, GHI, DHI, sampling time, ACO) in the traditional ACO algorithm to opti­
and wind direction mise the parameters of the SVM model.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1741

classification and data-preparation techniques-based models al. 2018), (Meenal and Selvakumar 2018), (S. Sun et al. 2018),
with the single-stage models, it was reported that hybrid (Yao et al. 2018), (Cornejo-Bueno et al. 2019), (Yagli, Yang,
models perform better and show better performance accuracy and Srinivasan 2019), (Fan et al. 2020), (Prasad et al. 2020),
compared to single-stage models. However, hybrid models (YagYagli et al. 2020), (Huertas-Tato et al. 2020), (Gupta and
may increase the cost and computational complexity. Singh 2021) (3) Regression – (Lou et al. 2016), (4) Nearest
Therefore, a tradeoff between performance accuracy and neighbour –(Alfadda, Rahman, and Pipattanasomporn
other constraints is required while considering the hybrid 2018), (Ben Ammar, Ben Ammar, and Oualha 2020),
models (Akhter et al. 2019). Clustering (Unsupervised learning) – (Azimi, Ghayekhloo,
and Ghofrani 2016), (Jiménez-Pérez and Mora-López 2016),
(S. Sun et al. 2018), (Ben Ammar, Ben Ammar, and Oualha
3.2. Indirect forecasting
2020), Ensemble learning (1) Boosting – (Lou et al. 2016),
In the case of a newly developed PV plant, historical data of PV (Hassan et al. 2017), (Fouilloy et al. 2018), (Pedro et al. 2018),
power of such plant may not be available for the training of ML (Yagli, Yang, and Srinivasan 2019), (Lee et al. 2020), (Kumari
models. In that case, direct forecasting from such models is not and Toshniwal 2021) (2) Bagging – (Jiménez-Pérez and
possible. However, due to the high correlation between solar Mora-López 2016), (H. Sun et al. 2016), (Hassan et al.
irradiance and PV power, indirect forecasting can be imple­ 2017), (Ibrahim and Khatib 2017), (Voyant et al. 2017a),
mented in such cases. The procedure to forecast PV power by (Fouilloy et al. 2018), (Yagli, Yang, and Srinivasan 2019),
this pathway is given as follows. (YagYagli et al. 2020), (Babar et al. 2020), (Narvaez et al.
Step 1: Collect solar irradiance and meteorological data, PV 2020), (Lee et al. 2020), (Gupta and Singh 2021) Deep learn­
plant location information, PV panel and inverter ing – (Mishra and Palanisamy 2018), (Feng and Zhang 2020),
characteristics. (Narvaez et al. 2020), (Pang, Niu, and O’Neill 2020), (Zang et
Step 2: Utilise data-driven models to forecast solar irradi­ al. 2020), (Gao et al. 2020), (Kumari and Toshniwal 2021)
ance on a horizontal plane. (Researchers also use NWP mod­ Others – (Sharma and Kakkar 2018), (Kisi, Heddam, and
els, but this paper aims to focus on ML techniques) Yaseen 2019), (Munkhammar, Van Der Meer, and Widen
Step 3: Use combination models to calculate the plane of 2019), (Wu et al. 2019a), (Wu et al. 2019b).
array solar irradiance.
Step 4: POA irradiance is applied as an input in PV perfor­ 3.2.2. Plane of array irradiance estimation
mance models to forecast solar power. Decomposition models are used to find the three components
The details are given in the following three subsections. (global horizontal irradiation, diffuse and direct irradiation) of
solar irradiation on a horizontal plane by utilising any compo­
3.2.1. Solar radiation forecasting nent. Transposition models convert these horizontal compo­
Papers using data-driven models classified in Section 2, for nents to their equivalent in the plane of solar panels. Papers
the solar radiation forecasting on horizontal plane are cate­ using combination models for the plane of array irradiance
gorised as follows: Time series models – (Fouilloy et al. estimation are categorised as follows- (Lave et al. 2015), (Cui et
2018), (Marchesoni-Acland et al. 2019), Classification and al. 2019), (Tina and Ventura 2011), (Khahro et al. 2015),
data-preparation – (Deo, Wen, and Qi 2016), (Passow, Lee, and Panchula 2017), (Riley et al. 2017), (Oka et
(Shamshirband et al. 2016), (Bouzgou and Gueymard 2017), al. 2018), (Takilalte, Harrouni, and Yaiche 2020), (Meng,
(S. Sun et al. 2018), (Lan et al. 2019), (Prasad et al. 2019), Loonen, and Hensen 2020).
(Prasad et al. 2020), (Zambrano and Giraldo 2020), (Gupta
and Singh 2021), Supervised learning (1) NN –(Zeng and 3.2.3. PV performance models
Qiao 2013), (Shamshirband et al. 2015), (Azimi, Ghayekhloo, PV performance models consist of PV cell and inverter mod­
and Ghofrani 2016), (Çelik, Teke, and Yildirim 2016), els. PV cell is the basic unit of the photovoltaic system and
(Jiménez-Pérez and Mora-López 2016), (Aybar-Ruiz et al. plays a significant role in PV power forecasting. Inverter model
2016), (Gutierrez-Corea et al. 2016), (Wang et al. 2016), converts instantaneous DC power from PV cell into AC power.
(Bouzgou and Gueymard 2017), (Mousavi, Mostafavi, and Papers using PV performance models for PV power forecast­
Jiao 2017), (Voyant et al. 2017a), (Alfadda, Rahman, and ing are categorised as follows- (Cui et al. 2019), (Liaoliao et al.
Pipattanasomporn 2018), (Fouilloy et al. 2018), (Khosravi et 2017), (Tan and Deng 2017), (Durrani et al. 2018), (Ben
al. 2018), (Meenal and Selvakumar 2018), (Rao, Premalatha, Ammar, Ben Ammar, and Oualha 2020).
and Naveen 2018), (Benali et al. 2019), (Cornejo-Bueno et al. The summary of literature using the above three categorised
2019), (Lan et al. 2019), (Marzouq et al. 2019), (Yagli, Yang, models (solar radiation forecasting, plane of array irradiance
and Srinivasan 2019), (Y. Feng et al. 2020), (Marzouq et al. estimation, PV performance models) for PV power forecasting
2020), (Pang, Niu, and O’Neill 2020), (Ben Ammar, Ben is given in (Table 2). Readers can refer to these references for
Ammar, and Oualha 2020), (Zambrano and Giraldo 2020), more detail.
(H. Wang et al. 2020) (2) SVM – (Fan et al. 2018), (Zeng and As discussed in section 1.1, the physical or empirical
Qiao 2013), (Belaid and Mellit 2016), (Deo, Wen, and Qi model is one of the categories of PV forecasting. The empiri­
2016), (Jang et al. 2016), (J. Li et al. 2016), (Jiménez-Pérez cal models depend upon the physical assumptions of atmo­
and Mora-López 2016), (Shamshirband et al. 2016), (Baser sphere, mathematical equations, solar elevation, hour angle,
and Demirhan 2017), (Alfadda, Rahman, and solar declination angle, etc.; this increases the complexity of
Pipattanasomporn 2018), (Fouilloy et al. 2018), (Khosravi et the model and makes it more suitable for medium and long-
1742 P. GUPTA AND R. SINGH

term forecasting applications. In contrast, artificial intelli­ et al. (2020) applied the crossover and mutation-based evolu­
gence applies a learning process to solve the non-linear and tionary algorithm, and Y. Feng et al. (2020) involved PSO to
complex relationship between the dependent and indepen­ optimise the parameters of neural networks.
dent variables (input and output data). Many articles reflect It is well known that the stochastic nature of PV data is
the suitability of machine learning and artificial intelligence the result of variability in solar irradiance, ambient tem­
techniques over physical models for short-term forecasting. perature, and other meteorological variables. This neces­
Recently, Ben Ammar, Ben Ammar, and Oualha (2020) pro­ sitates some data-preprocessing tools to deal with the
posed three methodologies, including empirical models, arti­ non-linear and non-stationary data by suitably extracting
ficial intelligence technique and the hybrid model based on a the embedded information. It can be concluded from the
neuro-fuzzy inference system for day-ahead solar irradiance previous literature that wavelet transformation is the most
and ambient temperature forecasting. The lowest normalised commonly used signal decomposition method. However,
mean bias error (NMBE) and normalised RMSE values were it is reported that the wavelet transformation-based ML
obtained using the hybrid model combining fuzzy logic and model achieved better performance than the single-stage
neural networks. Recent studies also show the potential of models, some drawbacks have been associated with them.
combining physical models with machine learning and intel­ Selection of the mother wavelet basis function is the pre­
ligence techniques. An analysis based on this idea has been requisite of the said data-preprocessing tool. A new multi-
proposed by Huertas-Tato et al. (2020). This study presents a resolution time-series decomposition tool (empirical mode
vision to integrate the results from four forecasting models decomposition (EMD)) proposed by Huang et al. (1998)
(satellite-based model, Smart Persistence, WRF-solar and may overcome the limitations of wavelet transformation.
CIADcast) with the popular ML algorithm (SVM) by three Recently, Gao et al. (2020) applied a hybrid deep learning
blending strategies (general, by horizons and regional). model which combines two popular deep neural networks
Further, a similar analysis has been performed in which to extract the temporal-spatial information from the
authors compared the performance of several ML models CEEMDAN-decomposed solar irradiance subseries.
(ELM, SVM, multi-layer perceptron (MLP), and Gaussian CEEMDAN, the empirical mode decomposition variant,
process (GP)) with and without considering the various phy­ can divide the noisy and volatile time series into several
sical models (CAMS, Heliostat-2 and SolarGIS) as input subseries termed intrinsic mode functions. Similarly,
variables (Cornejo-Bueno et al. 2019). The results demon­ Gupta and Singh (2021) proposed a univariate model for
strated that the ELM exhibits the lowest forecast error com­ 1–11 h ahead GHI forecasting by utilising the EMD and
pared to other regressors while including the physical models EEMD time-series decomposition techniques. The authors
as input. reported a tradeoff between these two techniques in terms
The researchers are adopting the various architectures of of time complexity and performance accuracy. Further,
neural networks in solar irradiance forecasting due to their Prasad et al. (2020) proposed a multivariate empirical
ability to model the nonlinearity and complexity of the data, as mode decomposition (MEMD) and singular vector
discussed in section 3.1. Recently, H. Wang et al. (2020) pro­ decomposition-based machine learning model for the
posed a four layers-based (one input layer, two linear and one week ahead solar irradiance forecasting. Unlike EMD
nonlinear layer) direct explainable neural network for 1-min and its variants, Multivariate empirical mode decomposi­
ahead solar irradiance forecasting. The authors concluded that tion can be applied in the multivariate setting.
the proposed deep learning model provides better perfor­ After reviewing the vast literature on machine learning
mance accuracy than the traditional neural network, SVM based-models for PV forecasting, it can be concluded that no
and XGBF, along with higher training efficiency. Further, model is universal. Different models experience different lim­
Feng and Zhang (2020) applied another deep learning archi­ itations such as model complexity in terms of time and archi­
tecture, i.e. deep convolutional neural network (CNN), for tecture, computational cost, instability issues, low performance
intra-hour solar irradiance forecasting. The authors developed for highly noisy and complex data, etc. Now-a-days, research­
this model by considering only sky images without any extra ers are interested in designing ensemble or hybrid models to
feature and reported a good range of normalised RMSE combine the rough prediction of several base learners. Several
(8.85%) and forecasting skill score (25.14%). recent articles on ensemble and hybrid models and their super­
Similar to the literature review on intelligent optimisation iority over single-stage models reflect this area’s popularity and
algorithms for direct PV power forecasting (section 3.1), some ongoing research. For example, recently, Zang et al. (2020)
recent articles based on these algorithms for the hyperpara­ proposed a deep learning hybrid model based on CNN-LSTM
meters tuning of ML models have also been reviewed in solar to model the spatiotemporal correlations between the neigh­
irradiance forecasting. For example, Fan et al. (2020) proposed bouring and the target site to predict 1-h ahead global hor­
a model to predict daily diffuse solar irradiance by optimising izontal irradiance. The authors combined the spatial and
SVM parameters with three intelligent meta-heuristic temporal features extraction ability of CNN and LSTM,
approaches (PSO, BAT, and WOA). Along with the recom­ respectively, and reported improved performance compared
mendation of BAT algorithm, the study also reported an to single-stage models. Similarly, Gao et al. (2020) used the
interesting finding of improvement in the prediction results same idea by adding an extra step of the data decomposition
by involving the air pollution parameters. Similarly, Marzouq technique for the week ahead solar irradiance forecasting.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1743

Further, Lee et al. (2020) presented a comparative analysis of 4.1. Growth of literature with time
various tree-based ensemble models (bagged trees, boosted
(Figure 4) shows the chronological development of the litera­
trees, random forest and generalised random forest) with the
ture reviewed in this paper. The first category is PV power
individual regressors (SVM and Gaussian process regressor).
direct forecasting, and the remaining three categories, i.e. solar
The results showed in the article reflect the superiority of
radiation forecasting, the plane of array irradiance estimation
ensemble models over single-stage models for short-term
and PV performance models, are the sub-categories of PV
solar irradiance forecasting. Later, Kumari and Toshniwal
power indirect forecasting.
(2021) proposed a novel hybrid model based on a tree-based
ensemble and deep learning approach for an hour ahead solar
irradiance forecasting. By considering reliable information
4.2. Literature classification based on methods deployed
extraction from the noisy raw data and ML model parameters
extraction, Prasad et al. (2019) proposed a multi-stage model. Data-preparation and dimensionality reduction methods and
The authors combined the MEMD and ant colony optimisa­ unsupervised learning-based algorithms cannot be directly
tion (ACO) in the first stage, and then the monthly solar applied for the forecasting applications, unlike supervised
irradiance forecasting was performed using the random forest learning. However, the application of these techniques with
ensemble model. forecasting models may increase the accuracy. (Figure 5)
As discussed earlier, in the indirect way of PV power shows the literature reviewed in this paper utilising data-
forecasting, once we have the solar irradiance data on a driven techniques in the field of PV power forecasting. In
horizontal plane, it needs to be converted into a corre­ (Figure 5), there may be some references that are common
sponding value on a tilted surface. Many studies in this in more than one data-driven method.
field have introduced various decomposition and transpo­
sition models based on mathematical equations to deter­
mine the solar irradiance in the plane of PV array from 4.3. Evolution of techniques with time
the horizontal one. Recently, Takilalte, Harrouni, and
Increasing the accuracy of PV power forecasting is a chal­
Yaiche (2020) adopted the two popular conventional
lenging task for researchers. A lot of research is going on
models Perrin Brichambaut and Liu and Jordan, to con­
in this field. (Figure 6) shows the evolution of some ML
vert the 5-min step global irradiance in the horizontal
techniques in this domain with time. It can be seen that
plane into the plane of PV array. The authors reported
deep learning, which is a newly developed sub-branch of
improved model accuracy by considering the cloudiness
machine learning, is being used in the field of PV power
factors (direct and diffuse) in the conventional models.
forecasting, and its popularity is increasing with time.
Similarly, Meng, Loonen, and Hensen (2020) applied a
Ensemble of predictors combines the features of two or
Perez transposition model to convert the hourly global
more models. As seen in most of the cases in Section 3,
horizontal solar irradiance into their corresponding values
its better forecasting accuracy and low error is attracting
in the tilted plane. The authors also presented an inter­
the researcher’s attention. Single-stage models like ANN
esting finding in this field by determining the set of tilt
and SVM are also some of the most commonly used
and azimuth angles to check the effect of these angles on
machine learning models because of the quality to deal
the profile of plane of array solar irradiance. Recently, Cui
with the non-linearity of the forecasting problem. In
et al. (2019) presented an article on indirect PV power
(Figure 6), there may be some references that are common
forecasting by considering the effect of the air mass, soil­
in more than one technique.
ing factor, and solar radiation incidence angle. It was
concluded from the results that the modified PV power
forecasting model shows improved performance accuracy
4.4. Forecast horizon
for different weather conditions than the traditional
model. The authors also reported the suitability of this Forecast horizon defines the span of time in the future for
model for not only newly developed PV plant but also for which the forecasts are to be made. For example 1-day ahead
the old plant in case of failure of real-time data forecast horizon for a specific site and testing period means the
acquisition. forecasts will be made up to one day in the future for the
daylight hours (Chaturvedi 2016).
Based on the span of time, forecasting is classified into four
4. Analysis of the development of literature on PV different types, and each forecast horizon has its own
power forecasting application.
Very short-term forecast: The forecast for few seconds to
This section presents the analysis of the development of
several minutes ahead comes in the very short-term forecast
literature on PV power forecasting based on the following
category.
factors: (1) chronological growth of literature, (2) methodol­
Short-term forecast: The forecast for 1-h or several hours
ogy used for forecasting, (3) evolution of techniques with
to 24 h or 1-week ahead comes in the short-term forecast
time, (4) forecast horizon and (5) the input parameters for
category.
data-driven models. The details are given in the following
Medium-term forecast: The forecast from 1 month to
subsections.
1 year ahead falls in the medium-term forecast category.
1744 P. GUPTA AND R. SINGH

Table 2. Summary of literature review on indirect PV power forecasting.

Forecast
Ref., location horizon Data duration Input variables Methodology Findings
Solar radiation forecasting
(Zeng and Qiao 2013) 1 h, 2 h, 3 h 1991–2005 Atmospheric transmissivity in 2D, Least-square support vector ● A linear model is worse than a non-
USA humidity, wind speed and sky machine, RBFNN linear in terms of capturing the
cover. non-linear characteristics.
● Sky-cover is found important site-
independent input variable.
(Shamshirband et al. 1-d 1995–2004 GHI, sunshine hour, maximum, Extreme learning machine ● ELM is a highly accurate and reli­
2015) Iran minimum, and average air (ELM) able model as compared to other
temperatures and relative ML models.
humidity
(Azimi, Ghayekhloo, 1-h, 2-h . . . NA Solar irradiance time-series data Transformation based (TB) ● The main feature of TB K-means
and Ghofrani 2016) .48 h K-means clustering, MLP clustering is to provide consistent
USA neural network results for different runs.
(Belaid and Mellit daily and 2 years Maximum, minimum, mean, SVM ● The proposed model requires few
2016) Algeria monthly maximum-minimum temperatures, parameters and low computational
horizontal daily global solar time to obtain good accuracy
radiation compared to other comparison
neural networks.
(Çelik, Teke, and 1-month 10 years Longitude, latitude, month, temp., ANN with LM algorithm ● ANN model accurately predict the
Yildirim 2016) sunshine hours, and altitude GSR with the use of limited
Turkey meteorological variables
(Deo, Wen, and Qi Daily and 1 year Sunshine hours, maximum SVM-WT ● SVM-WT model found accurate for
2016) Australia monthly 4 months temperature minimum both daily and monthly forecast
temperature, wind speed, horizons.
evaporation and precipitation.
(Jang et al. 2016) 15 to 4 years Satellite images for atmospheric support vector machine ● The proposed model shows better
South Korea 300 min satellite motion vector (AMV), irradiance (SVM) prediction accuracy compared to
data images, and cloud analysis images conventional time series and ANN
models.
(J. Li et al. 2016) 5 to 30 min 2 months Weather data (temperature, humidity, Hidden Markov Model ● It is observed that as the forecast
Australia atmospheric pressure) and solar (HMM) and SVM horizon increases forecast accuracy
data. regression of HMM and SVM regression mod­
els decreases.
(Lou et al. 2016) 1-month 5 years Global, diffuse and direct normal Boosted regression tree, ● Cloud cover and clearness index
China irradiance, clearness index, air logistic regression for are identified as the most impor­
temp, cloud cover, visibility, and multiple variables tant input variables to predict dif­
solar altitude fuse solar irradiance.
(Jiménez-Pérez and 1-d 3 years Humidity, atmospheric pressure, GHI, k-means clustering, ● SVM- C and SVM-R combination
Mora-López 2016) and air temperature decision trees, ANN, SVM outperforms the other four classi­
Spain regression (SVM-R), and fication and regression models
SVM classification (SVM- combination.
C)
(Aybar-Ruiz et al. NA NA Predictive variables from WRF, GGA–ELM ● The outputs from a numerical
2016) Spain previous solar radiation weather model (WRF) are applied
as input features for the ELM.
(Shamshirband et al. 1-d 2006–2012 Diffuse irradiance and GHI SVM-WT ● The authors considered the corre­
2016) Malaysia lation of cloudiness index with
clearness index only as an input
variable to predict diffuse solar
irradiance.
(H. Sun et al. 2016) 1-d 2001–2012 Sunshine hours, max and min air Random forest ● RF models with air pollution index
China temperature, air temperature (API) performed better than non-
difference, solar radiation, API API RF and empirical models.
(Gutierrez-Corea et al. 1–6 h 1 year GHI, air temp, humidity, MLP ● For short-term forecast horizon up
2016) Spain extraterrestrial solar radiation, to 3 h, nearest neighbour data
wind direction, clearness index, estimates play an important role.
zenith, wind speed
(Wang et al. 2016) 1-d 1961–2014 Sunshine durations, humidity, air MLP, GRNN, RBNN, and ● The MLP and RBNN models outper­
China temperature, GHI, max and min air improved Bristow- form the other comparable models.
temperature, water vapour Campbell (IBC)
pressure
(Baser and Demirhan 24-h 2000 and Solar radiation, mean air temperature, Fuzzy regression with SVM ● Linear, polynomial, Gaussian, and
2017) Turkey 2013 relative humidity, longitude, (FR-SVM) sigmoid kernel functions are used
latitude, altitude ● Gaussian kernel-based FR-SVM
approach found better.
(Bouzgou and 15-min, 1-h, NA GHI components Minimum redundancy- ● The proposed model combines two
Gueymard 2017) 24-h maximum relevance blocks: the first is dimensionality
Algeria (MRMR), ELM reduction methodology, and the
other is ELM to forecast solar
irradiance.
(Hassan et al. 2017) 1 h-24 h NA Horizontal global and diffuse Gradient boosting, bagging, ● Bagging and RF ensemble techni­
Egypt irradiance, normal irradiance and RF ques outperformed the gradient
boosting methodology.
(Continued)
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1745

Table 2. (Continued).
Forecast
Ref., location horizon Data duration Input variables Methodology Findings
Solar radiation forecasting
(Ibrahim and Khatib 1-d 1 year Sunshine ratio, ambient temp, month RF-FA ● Firefly algorithm optimises the
2017) Malaysia and day no., hours per day, and number of trees and leaves per
humidity tree for the RF forecasting model.
(Mousavi, Mostafavi, 1-d 1995–2014 Average, max and min of air temp, ANN-SA (SA-simulated ● ANN-SA model outperforms the
and Jiao 2017) Iran wind speed, pressure, and relative annealing) traditional ANN and SVM models
humidity to predict GHI.
(Voyant et al. 2017a) NA 1 year Measured hourly horizontal global MLP ● The authors discussed the effect of
France radiation four kinds of uncertainties on fore­
casted output
(Alfadda, Rahman, 1-h 1 year Ground measured GHI, DHI, and DNI, MLP, SVR, kNN and decision ● It is found that DNI is 3–4 times
and 9 months the hour of the day, the month of tree more sensitive to aerosol than GHI.
Pipattanasomporn the year, zenith, aerosol index data,
2018) USA Angstrom Exponent α, Wind speed
and direction,
(Fan et al. 2018) 1-d 2 years Daily meteorological (sunshine Support Vector Machine ● PM2.5, PM10, O3, and AQI found
China duration and air temperature) and (SVM) more relevant air pollution
air pollution data (PM2.5, PM10, parameters.
SO2, NO2, CO, O3, AQI)
(Fouilloy et al. 2018) 1–6 h Odeillo GHI time series ARMA, MLP, GP, SVR, ● ARMA or MLP found efficient mod­
France 2 years Standard and pruned els for low variability location,
Ajaccio regression trees, Boosted ARMA or bagged regression for
2 years and bagged regression medium variability location, and
Tilos 1 year trees, Random forests bagged regression or random for­
est for high variability location
(Khosravi et al. 2018) 1-h 6 years Temp, pressure, wind speed, Adaptive neuro FIS, ● MLFNN and SVR outperformed for
Brazil humidity, past time-series solar MLFFNN, RBFNN, SVR, FIS exogenous forecasting.
radiation ● MLFNN, SVR, and adaptive neuro
FIS outperformed for endogenous
forecasting.
(Meenal and 1-month NA Month, longitude, sunshine duration, Empirical models, SVM and ● Sunshine duration, max temp,
Selvakumar 2018) humidity, max, and min ANN month, and latitude are the most
India temperature, latitude influencing parameters according
to WEKA.
(Mishra and 1 h, 2 h, 3 h, 2010–2011 Solar irradiance, air temp, humidity, Recurrent neural network ● The proposed deep learning fore­
Palanisamy 2018) 4h wind speed, direction and station casting model outperforms other
USA pressure ML models in terms of forecasting
accuracy.
(Pedro et al. 2018) 5–30 min 1 year GHI, DNI and sky images Point forecast (persistence, ● The authors performed determinis­
USA KNN, GB, genetic tic and probabilistic forecast of GHI
boosting) and and DNI using various ML
probabilistic forecast techniques.
(persistence ensemble,
KNN and GB)
(Sharma and Kakkar 1 h-48 h 5 years Solar radiation FoBa, Leap Forward, spike ● Leap Forward, FoBa, and Cubist
2018) India slab, Cubist, bag Earth models performed well for large
GCV datasets, spikslab with less dataset
and bag Earth GCV found inde­
pendent with a dataset.
(Rao, Premalatha, and 1-d 3 years Max-min temperature, DT, ANN ● ANN models outperformed the
Naveen 2018) approximate sunshine hours(So), empirical models.
India terrestrial radiation(Ho), theoretical
sunshine hours(Sa)
(S. Sun et al. 2018) 1, 3, and 6 8 years Solar irradiance Decomposition-clustering- ● The DCE learning approach
China step 5 months ensemble (DCE) learning employed EEMD for decomposi­
ahead (EEMD-LSSVR-K-LSSVR) tion, K-means for clustering analy­
sis, and LSSVR for ensemble
learning.
(Yao et al. 2018) 1-d 3 years GHI, sunshine hours, temp, humidity SVM (SVM-1, SVM-2) ● Air quality index parameter
China 3 months and AQI improved the results of GHI pre­
diction accuracy.
(Marchesoni-Acland 10–240 min 2 years GHI, satellite albedo, local short-term ARMAX ● PGM-S (past ground measure­
et al. 2019) France variability ments with satellite data) model
outperformed.
(Benali et al. 2019) 1–6 h 3 years Global, beam and diffuse radiation Smart persistence, ANN and ● Normal and diffuse horizontal irra­
Algeria data RF diance showed more sensitivity to
meteorological variables than glo­
bal irradiance.
(Cornejo-Bueno et al. NA 1 year Cloud index, clear sky solar radiation SVR, extreme learning ● ELM was found as the best per­
2019) Spain model, several reflectivity machine(ELM), GP, MLP forming regressor among all the
machine learning regressors and
physical models.
(Continued)
1746 P. GUPTA AND R. SINGH

Table 2. (Continued).
Forecast
Ref., location horizon Data duration Input variables Methodology Findings
Solar radiation forecasting
(Kisi, Heddam, and 1-month 1983–2010 Solar irradiance, air temperature Dynamic evolving neural- ● DENFIS model provided accurate
Yaseen 2019) and 1968– FIS model (DENFIS) SR prediction results as compare
Turkey 2015 for to MARS, M5Tree, and LSSVR
stations models.
(Lan et al. 2019) 1-d 1 year Time-series irradiance data Discrete Fourier transform ● The proposed model extracts the
China (DFT), PCA, Elman based frequency features, unlike tradi­
neural network tional feature extraction in the time
domain
(Marzouq et al. 2019) 1-d 6 years GHI, humidity, rainfall, air Evolutionary artificial neural ● The proposed method used
Morocco temperature, wind direction and networks genetic operators to predict the
speed solar radiation and automatically
find the most affecting input
parameters.
(Munkhammar, Van 1-d 1 year GHI, clear sky irradiance Markov-chain mixture ● MCM model outperforms the
Der Meer, and (MCM) distribution and copula model and shows nearly
Widen 2019) copula model the same results with benchmark
Sweden quantile regression probabilistic
model.
(Prasad et al. 2019) 1 month 1905–2018 Tmax, Tmin, Rain, vapour pressure, max MEMD-ACO-RF ● The proposed model outperformed
Australia and min relative humidity, the other comparable models in
Potential Evapotranspiration terms of good performance
(FAO5). metrices and low forecast error.
(Wu et al. 2019a) USA 24-hour May to Lifting condensation level, inversion Analog ensemble ● AnEn outperformed persistence
September strength tropospheric and Weather research and fore­
2014–2017 total water mixing ratio, water casting model in terms of low
potential centred RMSE and RMSE values.
temperature, surface dew point
temperature and precipitable
water
(Wu et al. 2019b) 1-d 2001–2015 GHI, sunshine duration, humidity, KNEA and CatBoost, ● The proposed two models accu­
China max and min temperature, and rately forecast the GHI in the
precipitation humid region.
(Yagli, Yang, and 1-h 4 years Solar irradiance 68 machine learning ● The authors performed an analysis
Srinivasan 2019) models to determine which model is more
USA successful under which climate.
(Fan et al. 2020) 1-d 3 years GHI, diffuse solar irradiance, sunshine SVM-PSO, SVM-BAT and ● PSO, BAT, and WOA are heuristic
China duration, max and min temp, SVM-WOA algorithms.
PM2.5, PM10 and O3 ● SVM-BAT enhanced performance
accuracy followed by SVM-PSO
and SVM-WOA
(Y. Feng et al. 2020) 1-d 1961–2016 Solar irradiance, sunshine hours, PSO-ELM ● PSO improves the forecasting accu­
China relative humidity, average, max racy by optimising parameters of
and min air temp. ELM.
(Prasad et al. 2020) 1-week 1905–2018 Potential evapotranspiration (FAO56), MEMD-SVD- RF ● The authors designed a MEMD-SVD
Australia vapour pressure, evaporation, based machine learning model for
relative humidity at max and min reducing the feature dimension.
temp., rainfall, max and min temp.
(Yagli et al. 2020) 1-h 3 years Ground and satellite-based solar Cubist, SVR, GLMNET, ● Satellite-based measurements can
Singapore irradiance Random Forest, PPR be taken as an alternate option in
solar resource forecasting due to
the scarcity of reliable Ground-
based data sources.
(Babar et al. 2020) 1-d _ GHI, clear sky indices, solar zenith Random forest and CLARA- ● The random forest has been found
Norway A2, ERA5 satellite-based as a good alternative to both ERA5
models and CLARA-A2 reanalyses and
satellite models.
(C. Feng and Zhang 10 min to 6 years GHI, DNI, DHI, temperature, relative Deep convolutional neural ● CNN has been developed for uni­
2020) USA 60 min humidity, wind speed, atmosphere network(CNN) variate data (GHI only).
ahead pressure ● Other variables are used in bench­
mark models.
(Narvaez et al. 2020) Daily (24 h, 2006–2017 GHI, DHI, DNI, solar zenith, Machine learning (Random ● The authors proposed a methodol­
Colombia 72 h, temperature, wind speed forest, quantile ogy on site adaption technique
168 h), mapping), deep learning (selection of best of both satellite
weekly (LSTM, GRU) and in-situ measurements) using
(168 h, ML algorithms.
336 h,
672 h)
(Marzouq et al. 2020) 1–6 h ahead 1 year Solar irradiance HAEANN (History and ● For all three scenarios, the pro­
Morocco architecture evolutionary posed model outperformed the
ANN) smart persistence, RF, and decision
tree models.
(Continued)
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1747

Table 2. (Continued).
Forecast
Ref., location horizon Data duration Input variables Methodology Findings
Solar radiation forecasting
(Pang, Niu, and 10 min, May 22nd to Solar irradiance, dry-bulb air ANN, ANN_Moving window, ● Moving window algorithm can
O’Neill 2020) USA 30 min, 29 May temperature, time of day RNN, RNN_Moving slightly improve the results.
60 min 2016 window ● RNN model outperformed the ANN,
but it requires large computational
time.
(Ben Ammar, Ben 1-d _ Temperature, solar irradiance Empirical models, FFNN, ● By utilising forecasted (ANFIS out­
Ammar, and ANFIS (Adaptive neuro- performs) solar irradiance and
Oualha 2020) FIS) temperature, PV power has been
Tunisia calculated by applying a mathe­
matical model.
(Zambrano and 1-h 2005–2017 GHI ANN, Algorithms for the ● The paper presents some algo­
Giraldo 2020) selection of sites rithms for selecting similar pattern
Colombia (Mahalanobis distance, sites for the prediction of solar
kernel machines, PCA, irradiance without on-site training
SFS) measurements.
(Zang et al. 2020) 1-h 2006–2012 GHI, wind speed, and direction, dew CNN-LSTM ● The proposed model combines the
China point temp, precipitable water, spatiotemporal correlation of data
solar zenith, relative humidity, to predict the GHI effectively.
temperature
(Gao et al. 2020) 1-h 6 years Solar irradiance CEEMDAN-CNN-LSTM ● Paper presents an advanced tech­
China nique to decompose the highly
noisy solar irradiance time series
into different IMFs.
(H. Wang et al. 2020) 1-min 2018 Solar radiation, relative humidity, sun Direct explainable neural ● Straightforward, and understand­
China altitude, temperature, wind speed, network (DXNN) able input-output mapping of the
global and zenith illuminance, etc. proposed prediction model is the
critical finding of this research
work.
(Lee et al. 2020) 1-h _ Solar radiation, month, day of month Boosted trees, bagged ● Ensemble learning models outper­
Saudi Arabia and hour of the day trees, Random forest, formed SVM and Gaussian process
generalised random regression for 16 evaluations out of
forest 18.
(Huertas-Tato et al. 15 min to 2 years GHI, DHI, DNI Satellite model, WRF-solar, ● Blended physical model using SVM
2020) Spain 6h smart persistence, outperforms physical models.
CIADcast, SVM
(Kumari and 1-h 10 years GHI, relative humidity, temperature, XGBF-DNN(Extreme ● The proposed model has reduced
Toshniwal 2021) wind speed and direction, gradient boosting- Deep the forecasting error by 40% com­
India pressure, clear sky index neural network) pared to other traditional ML
models.
(Gupta and Singh 1–11 h 5 years Time series of GHI EEMD-RF, EMD-RF, EEMD- ● A univariate model has been pre­
2021) India SVM, EMD-SVM sented for 1–11 h ahead GHI fore­
casting comparing the two
adaptive time-series decomposi­
tion techniques.
Plane of array irradiance estimation
(Tina and Ventura NA NA Global solar radiation for horizontal Perez, Klucher, ANN ● Perez and Klucher models are use­
2011) and 45° and 60° tilted surface ful to compute POA irradiance
Italy when horizontal solar radiations
are available.
(Khahro et al. 2015) NA NA Sunshine hours, temperature, relative Diffuse solar irradiance ● Tilt angle adjustment two times in
Pakistan humidity, pressure, global solar model a year was found advantageous.
radiation
(Lave et al. 2015) NA NA GHI, DHI, POA, clearness index, solar 12 decomposition and 4 ● Erbs and DIRINT decomposition
USA elevation angle transposition models models showed better
performance.
● Hay/Davis model outperformed in
transposition models.
(Passow, Lee, and NA NA Measured GHI Decomposition models ● DIRINT model gave better results,
Panchula 2017) (DIRINT, Erbs, Reindl-3) followed by Erbs and Reindl-3
USA model.
(Riley et al. 2017) NA NA GHI, DHI, POAI Transposition models(Hays/ ● By tuning satellite-derived GHI
California Davis, Perez) with ground-based GHI, a reduc­
tion in error of modelled POAI
(evaluated by transposition mod­
els) can be observed.
(Oka et al. 2018) NA NA Global horizontal irradiance Four decomposition and 3 ● Erbs and Perez found the most sui­
Japan transposition models table decomposition and transpo­
sition models, respectively.
(Continued)
1748 P. GUPTA AND R. SINGH

Table 2. (Continued).
Forecast
Ref., location horizon Data duration Input variables Methodology Findings
Solar radiation forecasting
(Takilalte, Harrouni, 5-min NA GHI Perrin Brichambaut, Liu- ● By taking advantage of the cloudi­
and Yaiche 2020) Jordan models in ness factor, the proposed model
Algeria conjunction with outperformed traditional models
cloudiness factor and recently developed techniques
like ANN.
(Meng, Loonen, and 1-h NA PV generation and off-site solar Perez transposition model ● The authors present a novel tech­
Hensen 2020) irradiance nique to find the tilt and azimuth
Netherland angles for any distributed photo­
voltaic systems by utilising the
Perez transposition model.
PV performance models
(Liaoliao et al. 2017) NA NA Solar irradiance, atmospheric Three decomposition, 3 ● Out of the 9 models, the model
China pressure, ambient temp., wind transposition models, with Liu-Jordan performed better
speed and PV performance as compared to other models.
model
(Tan 2017) ultra short 1 year Cloud cover, solar irradiance Adaboost, KNN clustering, ● The authors introduced the con­
China term and Markov Chain, hotel cept of the attenuation coefficient
model, PV engineering of solar radiation to improve the
model efficiency of forecasted power
(Durrani et al. 2018) 1-d ahead 2010–2015 Temperature, pressure, humidity, Combination model, Feed ● The GHI predicted from MFFNN
United States wind speed and direction, sunshine forward neural network, outperforms the persistence
hours and cloud cover photovoltaic model model.
● PV model determines PV power
through a mathematical equation.
(Cui et al. 2019) 15 min NA Solar irradiance, ambient Three decomposition and ● Air mass and soiling effect
China temperature, wind speed, four transposition improved the forecasting results
atmospheric pressure, models, PV electric ● Erbs+Liu-Jordan model was found
model better.

35
30
No. of papers

25
PV power direct
20 forecasting
15 PV performance models
10
5 Plane of array irradiance
0 estimation
Solar radiation
forecasting

Year
Figure 4. Growth of literature with time.

Long-term forecast: The forecast from 1-year to several direction and speed, sunshine duration, humidity and
years ahead comes in the long-term forecast category. clou: cover, are also used to estimate solar power genera­
(Figure 7) shows the classification of reviewed literature for tion. The correlation factor of several weather variables
various forecast horizons. For very short-term and short-term with PV power varies, depending upon the location of the
forecast horizon, the researchers recommend machine learn­ PV plant. Based on the nature of input parameters, PV
ing models. In (Figure 7 and 8), there may be some references power forecasting can be classified as follows.
that are common in more than one forecast horizon Historical data-based forecasting: This type of forecasting
classification. identifies a relationship between the input and the output data.
Historical solar radiation data/PV power is used as the input to
predict next-day solar radiation/PV power.
4.5. Input parameters Historical and meteorological data-based forecasting-
This type of forecasting combines historical solar irradiation/
In solar energy research, various parameters are essential.
PV power with meteorological input parameters like wind
Solar irradiance is the most important parameter in this
direction and speed, humidity, air temperature, cloud cover,
area of study. The variability of solar irradiance directly
aerosol index, air pollution, sky images, etc. to predict solar
affects the power generation of the PV plant. Several
irradiation/PV power.
meteorological variables, for example temperature, wind
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1749

100 92
90

No. of Papers
80
70
60
50
40 25
30 18 13
20 5 4
10
0
Time series Classification Supervised Unsupervised Ensemble Deep learning
models and data- learning learning learning models
preparation
techniques
Data-driven methods
Figure 5. Classification of data-driven methods-based literature.

12
10
No. of Papers

Time series models

8
6
Neural Network
4 SVM
2 Ensemble
Deep learning
0 Ensemble
SVM Deep learning
2011 2012 Neural Network
2013 2014
2015 2016 Time series models
2017 2018
2019 2020
Year
Figure 6. Evolution of techniques with time.

5. Conclusion evolution of techniques, forecast horizon and input para­

meters has also been presented. Some key points from the
The paper presents a review of the recently published
literature are as follows.
papers on PV power forecasting along two pathways, i.e.
direct forecasting and indirect forecasting. Direct forecast­
(1) Researchers suggest an indirect way of PV power fore­
ing techniques directly forecast the PV power output. The
casting when direct PV power forecasting is not possi­
literature for indirect PV power forecasting is classified
ble using ML techniques. For example, for a newly
into three categories, i.e. solar radiation forecasting, plane
developed PV plant.
of array irradiance estimation and PV performance mod­
(2) As shown in the paper, many data-driven techniques
els. These three categories are the three steps of the
are available for forecasting. Some are beginning to be
methodology for forecasting PV power in an indirect
used more frequently (bagging, boosting, decision
way. Analysis of the development of literature on
trees, etc.), some are often used (ANN, SVM, ARIMA,
PV power forecasting based on some factors like chron­
MLR), and some have not been used so frequently
ological growth of literature, methods deployed, the
(deep learning).

60
49
50
No. of Ppapers

40 33
30
Solar radiation forecasting
20
9 7 PV power direct forecasting
10 5
0 0 0
0
Very short Short term Medium term Long term
term
Forecast Horizon
Figure 7. Classification with respect to forecast horizon.
1750 P. GUPTA AND R. SINGH

Solar radiation
50 Solar radiation
40 forecasting
30 PV performance models
20
10
PV
0 S.R.+M.V. PV power direct
power+M.V.
forecasting

M.V. Meteorological variables

PV power S.R. Solar Radiation

Figure 8. Classification with respect to input parameters.

(3) Ensemble learning-based forecasting models out­

AnEn Analog Ensemble
perform the single-stage forecasting models (ANN, ANN Artificial neural network
SVM, k-NN, etc.) in most of the cases. ACO Ant colony optimisation
(4) Deep learning is becoming a more powerful tool in ARIMA Auto regressive integrated moving average
AQI Air quality index
time series forecasting problems due to its ability to ARMAX ARMA with exogenous inputs
discover the complex non-linearity of the time series BP Back propagation
data. The ability of deep learning models to overcome BAT Bat algorithm
BI-LSTM Bi directional LSTM
the limitations of traditional time series models BPNN Back propagation neural network
(ARMA, ARIMA) makes them the best option for CatBoost Gradient boosting with categorical features support
time series forecasting. CF Cascade forward
CLARA-A2 Cloud, Albedo, Radiation Dataset Edition 2
(5) From the literature on the plane of array irradiance CEEMDAN Complete ensemble empirical mode decomposition adaptive
estimation models, it can be said that, decomposi­ noise
tion models’ results (modelled and measured) vary DNI Direct normal irradiance
DNN- Deep neural network- randomly occurring distributed delayed
with location, whereas transposition models’ results RODDPSO particle swarm optimisation
vary more with the type of model. ECMWF European centre for medium range weather forecasts
EEMD Ensemble Empirical Mode Decomposition
ERA5 ECMWF Reanalysis 5
To conclude, we cannot label any model as a universal FA Firefly algorithm
model. One model can be superior to others in some FFNN Feed forward neural network
cases, whereas some other models might perform better FIS Fuzzy inference system
FoBa Adaptive forward backward greedy algorithm
in other conditions. GA-ANN Genetic algorithm-artificial neural network
GGA Grouping genetic algorithm
GHI Global horizontal irradiance
Highlights GRNN General regression NN
GLMNET Generalised linear model with elastic net regulariser
1. A review of PV power forecasting using data-driven models has been GRU Gated recurrent
done. GPR Gaussian process regression
2.Two pathways for forecasting have been reviewed: direct and indirect. IBL Inversion based height
3. Direct forecasting directly forecasts the PV power output. IMF Intrinsic mode function
4. Indirect forecasting consists of solar irradiance forecast, plane of array K-NN K-nearest neighbour
irradiance estimation and PV performance model.
5. Analysis of the development of literature in the area has also been
presented.

Acknowledgements
Nomenclature
This work was made possible with the fellowship support provided
by the Ministry of Education, formerly the Ministry of Human
AnEn Analog Ensemble Resource Development, Government of India.
ANN Artificial neural network
ACO Ant colony optimisation
ARIMA Auto regressive integrated moving average
AQI Air quality index Disclosure statement
ARMAX ARMA with exogenous inputs
BP Back propagation No potential conflict of interest was reported by the author(s).
(Continued)
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1751

Funding Babar, B., L. T. Luppino, T. Boström, and S. N. Anfinsen. 2020.

“Random Forest Regression for Improved Mapping of Solar
This work was supported by the Ministry of Education (formerly Ministry Irradiance at High Latitudes.” Solar Energy 198: 81–92.
of Human Resource Development), Government of India. doi:10.1016/j.solener.2020.01.034.
Baser, F., and H. Demirhan. 2017. “A Fuzzy Regression with Support
Vector Machine Approach to the Estimation of Horizontal Global
Notes on contributors Solar Radiation.” Energy 123: 229–240. doi:10.1016/j.
energy.2017.02.008.
Priya Gupta is currently pursuing a Ph.D. in Solar Energy in the Bbeiman, L. E. O. 1996. “Bagging.” Predictors 140: 123–140.
Department of Hydro and Renewable Energy at IIT Roorkee, India. She Belaid, S., and A. Mellit. 2016. “Prediction of Daily and Mean Monthly
received the Master of Technology degree (Gold Medal) in Power System Global Solar Radiation Using Support Vector Machine in an Arid
from NIT Kurukshetra (2019), and a Bachelor of Technology degree in Climate.” Energy Conversion and Management 118: 105–118.
Electrical Engineering from G. B. Pant Engineering College, Pauri doi:10.1016/j.enconman.2016.03.082.
Garhwal (2016). Her main research areas are machine learning, deep Ben Ammar, R., M. Ben Ammar, and A. Oualha. 2020. “Photovoltaic
learning, time-series data, solar resource forecasting, MPPT control of Power Forecast Using Empirical Models and Artificial Intelligence
PV modules. She is a recipient of the Ministry of Education, Government Approaches for Water Pumping Systems.” Renewable Energy 153:
of India fellowship for Postgraduate research as well as Doctoral 1016–1028. doi:10.1016/j.renene.2020.02.065.
Research. Benali, L., G. Notton, A. Fouilloy, C. Voyant, and R. Dizene. 2019. “Solar
Rhythm Singh is currently working as an Assistant Professor in the Radiation Forecasting Using Artificial Neural Network and Random
Department of Hydro and Renewable Energy at IIT Roorkee, India. He Forest Methods: Application to Normal Beam, Horizontal Diffuse and
holds a Ph.D. in Energy Science and Engineering from IIT Bombay Global Components.” Renewable Energy 132: 871–884. doi:10.1016/j.
(2018), a Master of Technology degree in Energy Studies from IIT Delhi renene.2018.08.044.
(2009), and a Bachelor of Technology (Gold Medal) degree in Electrical Bouralou, A., C. A. Messaoud Hamouda, M. Mostefaoui, M. Sadok,
Engineering from MNNIT Allahabad (2007). His research interests lie at M. Mostefaoui, and S. Lachtar. 2015. “Modeling and Simulation
the intersection of technological analysis and policy support for enabling a of Photovoltaic Module and Array Based on One and Modeling
transition to renewable energy technologies. He was awarded the Best and Simulation of Photovoltaic Module and Array Based on One
Poster Award at the 2017 IEEE 44th Photovoltaic Specialists Conference and Two Diode Model Using Matlab/Simulink.” Energy Procedia
held in Washington D.C., USA, for his paper on solar resource variability 74: 864–877. doi:10.1016/j.egypro.2015.07.822.
assessment. He is an active reviewer for around 13 reputed journals Bouzgou, H., and C. A. Gueymard. 2017. “Minimum Redundancy –
published by Elsevier, Taylor and Francis, IEEE, IET, and AIP publishing. Maximum Relevance with Extreme Learning Machines for Global
Solar Radiation Forecasting: Toward an Optimized Dimensionality
Reduction for Solar Time Series.” Solar Energy 158: 595–609.
ORCID doi:10.1016/j.solener.2017.10.035.
Brockwell, P. J., and R. A. Davis. 2006. “Springer Series in Statistics
Rhythm Singh http://orcid.org/0000-0003-4510-5845 Advisors.”
Çelik, Ö., A. Teke, and H. B. Yildirim. 2016. “The Optimized Artificial
Neural Network Model with Levenberg-Marquardt Algorithm for
References Global Solar Radiation Estimation in Eastern Mediterranean Region
of Turkey.” Journal of Cleaner Production 116: 1–12. doi:10.1016/j.
Akhter, M. N., S. Mekhilef, H. Mokhlis, and N. M. Shah. 2019. “Review on jclepro.2015.12.082.
Forecasting of Photovoltaic Power Generation Based on Machine Chang, G. W., and H. J. Lu. 2020. “Integrating Gray Data
Learning and Metaheuristic Techniques.” IET Renewable Power Preprocessor and Deep Belief Network for Day-Ahead PV
Generation 13 (7): 1009–1023. doi:10.1049/iet-rpg.2018.5649. Power Output Forecast.” IEEE Transactions on Sustainable
Alfadda, A., S. Rahman, and M. Pipattanasomporn. 2018. “Solar Irradiance Energy 11 (1): 185–194. doi:10.1109/TSTE.2018.2888548.
Forecast Using Aerosols Measurements: A Data Driven Approach.” Solar
Chaturvedi, D. K. 2016. “Solar Power Forecasting : A Review.”
Energy 170: 924–939. doi:10.1016/j.solener.2018.05.089.
International Journal of Computers and Applications 145: 28–50.
Ali Jallal, M., S. Chabaa, and A. Zeroual. 2020. “A Novel Deep Neural Network
Chen, C., S. Duan, T. Cai, and B. Liu. 2011. “Online 24-h Solar Power
Based on Randomly Occurring Distributed Delayed PSO Algorithm for
Forecasting Based on Weather Type Classification Using Artificial
Monitoring the Energy Produced by Four Dual-axis Solar Trackers.”
Neural Network.” Solar Energy 85 (11): 2856–2870. doi:10.1016/j.
Renewable Energy 149: 1182–1196. doi:10.1016/j.renene.2019.10.117.
solener.2011.08.027.
Almeida, M. P., O. Perpiñán, and L. Narvarte. 2015. “PV Power Forecast Using a
Nonparametric PV Model.” Solar Energy 115: 354–368. doi:10.1016/j. Cheng, H. 2016. “Hybrid Solar Irradiance Now-casting by Fusing Kalman
solener.2015.03.006. Fi Lter and Regressor.” Renewable Energy 91: 434–441. doi:10.1016/j.
Alskaif, T., S. Dev, L. Visser, M. Hossari, and W. Van Sark. 2019. “On the renene.2016.01.077.
Interdependence and Importance of Meteorological Variables for Cococcioni, M., E. D’Andrea, and B. Lazzerini. 2011. “24-Hour-ahead
Photovoltaic Output Power Estimation.” Conference Record of the Forecasting of Energy Production in Solar PV Systems.” International
IEEE Photovoltaic Specialists Conference 2117–2120. doi:10.1109/ Conference on Intelligent System Design Applicatio ISDA 1276–1281.
PVSC40753.2019.8981308. doi:10.1109/ISDA.2011.6121835.
Aybar-Ruiz, A., S. Jiménez-Fernández, L. Cornejo-Bueno, C. Casanova- Cornejo-Bueno, L., C. Casanova-Mateo, J. Sanz-Justo, and S. Salcedo-
Mateo, J. Sanz-Justo, P. Salvador-González, and S. Salcedo-Sanz. 2016. Sanz. 2019. “Machine Learning Regressors for Solar Radiation
“A Novel Grouping Genetic Algorithm-Extreme Learning Machine Estimation from Satellite Data.” Solar Energy 183: 768–775.
Approach for Global Solar Radiation Prediction from Numerical doi:10.1016/j.solener.2019.03.079.
Weather Models Inputs.” Solar Energy 132: 129–142. doi:10.1016/j. Cortes, C., and V. Vapnik. 1995. „Support-Vector Networks.„ Machine
solener.2016.03.015. Learning 20: 273–297. doi: 10.1007/BF00994018
Azimi, R., M. Ghayekhloo, and M. Ghofrani. 2016. “A Hybrid Method Cui, C., Y. Zou, L. Wei, and Y. Wang. 2019. “Evaluating Combination
Based on A New Clustering Technique and Multilayer Perceptron Models of Solar Irradiance on Inclined Surfaces and Forecasting
Neural Networks for Hourly Solar Radiation Forecasting.” Energy Photovoltaic Power Generation.” IET Smart Grid 2 (1): 123–130.
Conversion and Management 118: 331–344. doi:10.1016/j. doi:10.1049/iet-stg.2018.0110.
enconman.2016.04.009.
1752 P. GUPTA AND R. SINGH

Deo, R. C., X. Wen, and F. Qi. 2016. “A Wavelet-coupled Support Vector Haque, A. U., M. H. Nehrir, and P. Mandal. 2013. “Solar PV Power
Machine Model for Forecasting Global Incident Solar Radiation Using Generation Forecast Using a Hybrid Intelligent Approach.” The IEEE
Limited Meteorological Dataset.” Applied Energy 168: 568–593. Power & Energy Society General Meeting 1–5. doi:10.1109/
doi:10.1016/j.apenergy.2016.01.130. PESMG.2013.6672634.
Dewangan, C. L., S. N. Singh, and S. Chakrabarti. 2020. “Combining Hassan, M. A., A. Khalil, S. Kaseb, and M. A. Kassem. 2017. “Exploring
Forecasts of Day-ahead Solar Power.” Energy 202: 117743. the Potential of Tree-based Ensemble Methods in Solar Radiation
doi:10.1016/j.energy.2020.117743. Modeling.” Applied Energy 203: 897–916. doi:10.1016/j.
Durrani, S. P., S. Balluff, L. Wurzer, and S. Krauter. 2018. apenergy.2017.06.104.
“Photovoltaic Yield Prediction Using an Irradiance Forecast Hossain, M., S. Mekhilef, M. Danesh, L. Olatomiwa, and S. Shamshirband.
Model Based on Multiple Neural Networks.” Journal of Modern 2017. “Application of Extreme Learning Machine for Short Term
Power Systems and Clean Energy 6 (2): 255–267. doi:10.1007/ Output Power Forecasting of Three Grid-connected PV Systems.”
s40565-018-0393-5. Journal of Cleaner Production 167: 395–405. doi:10.1016/j.
El-Baz, W., P. Tzscheutschler, and U. Wagner. 2018. “Day-ahead jclepro.2017.08.081.
Probabilistic PV Generation Forecast for Buildings Energy Huang, J., M. M. Khan, Y. Qin, and S. West. 2019. “Hybrid Intra-
Management Systems.” Solar Energy 171: 478–490. doi:10.1016/j. hour Solar PV Power Forecasting Using Statistical and
solener.2018.06.100. Skycam-based Methods.” Conference Record of the IEEE
Eseye, A. T., J. Zhang, and D. Zheng. 2018. “Short-term Photovoltaic Solar Photovoltaic Specialists Conference 2434–2439. doi:10.1109/
Power Forecasting Using a Hybrid Wavelet-PSO-SVM Model Based PVSC40753.2019.8980732.
on SCADA and Meteorological Information.” Renewable Energy 118: Huang, N. E., Z. Shen, S. R. Long, M. C. Wu, H. H. Snin, Q. Zheng,
357–367. doi:10.1016/j.renene.2017.11.011. N. C. Yen, C. C. Tung, and H. H. Liu. 1998. “The Empirical Mode
Espinar, B., J. Aznarte, R. Girard, A. M. Moussa, B. Espinar, J. Aznarte, R. Decomposition and the Hubert Spectrum for Nonlinear and Non-
Girard, A. M. Moussa, and G. Kariniotakis. 2013. „Photovoltaic stationary Time Series Analysis.” Proceedings of the Royal Society of
Forecasting : A State of the Art to Cite This Version.„ 5th European
London. Series A: Mathematical, Physical and Engineering Sciences
PV-Hybrid and Mini-Grid Conference. Tarragona, Spain, pp. 250–255.
454 (1971): 903–995. doi:10.1098/rspa.1998.0193.
hal–00771465.
Huang, Y., J. Lu, C. Liu, X. Xu, W. Wang, and X. Zhou. 2010.
Fan, J., L. Wu, F. Zhang, H. Cai, X. Wang, X. Lu, and Y. Xiang. 2018.
“Comparative Study of Power Forecasting Methods for PV Stations.”
“Evaluating the Effect of Air Pollution on Global and Diffuse Solar
2010 International Conference on Power System Technology :
Radiation Prediction Using Support Vector Machine Modeling Based
Technological Innovations Making Power Grid Smarter,
on Sunshine Duration and Air Temperature.” Renewable and
POWERCON2010 1–6. doi:10.1109/POWERCON.2010.5666688.
Sustainable Energy Reviews 94: 732–747. doi:10.1016/j.
Huertas-Tato, J., R. Aler, I. M. Galván, F. J. Rodríguez-Benítez, C. Arbizu-
rser.2018.06.029.
Barrena, and D. Pozo-Vázquez. 2020. “A Short-term Solar Radiation
Fan, J., L. Wu, X. Ma, H. Zhou, and F. Zhang. 2020. “Hybrid Support
Forecasting System for the Iberian Peninsula. Part 2: Model Blending
Vector Machines with Heuristic Algorithms for Prediction of Daily
Approaches Based on Machine Learning.” Solar Energy 195: 685–696.
Diffuse Solar Radiation in Air-polluted Regions.” Renewable Energy
doi:10.1016/j.solener.2019.11.091.
145: 2034–2045. doi:10.1016/j.renene.2019.07.104.
Ibrahim, I. A., and T. Khatib. 2017. “A Novel Hybrid Model for Hourly
Feng, C., and J. Zhang. 2020. “SolarNet: A Sky Image-based Deep
Global Solar Radiation Prediction Using Random Forests Technique
Convolutional Neural Network for Intra-hour Solar Forecasting.”
and Firefly Algorithm.” Energy Conversion and Management 138: 413–
Solar Energy 204: 71–78. doi:10.1016/j.solener.2020.03.083.
425. doi:10.1016/j.enconman.2017.02.006.
Feng, Y., W. Hao, H. Li, N. Cui, D. Gong, and L. Gao. 2020. “Machine
Ismail, A. M., R. Ramirez-Iniguez, M. Asif, A. B. Munir, and F.
Learning Models to Quantify and Map Daily Global Solar Radiation
Muhammad-Sukki. 2015. “Progress of Solar Photovoltaic in ASEAN
and Photovoltaic Power.” Renewable and Sustainable Energy Reviews
Countries: A Review.” Renewable and Sustainable Energy Reviews 48:
118: 109393. doi:10.1016/j.rser.2019.109393.
399–412. doi:10.1016/j.rser.2015.04.010.
Fouilloy, A., C. Voyant, G. Notton, F. Motte, C. Paoli, M. L. Nivet, E.
Jang, H. S., K. Y. Bae, H. S. Park, and D. K. Sung. 2016. “Solar Power
Guillot, and J. L. Duchaud. 2018. “Solar Irradiation Prediction with
Prediction Based on Satellite Images and Support Vector Machine.”
Machine Learning: Forecasting Models Selection Method Depending
IEEE Transactions on Sustainable Energy 7 (3): 1255–1263.
on Weather Variability.” Energy 165: 620–629. doi:10.1016/j.
doi:10.1109/TSTE.2016.2535466.
energy.2018.09.116.
Jiménez-Pérez, P. F., and L. Mora-López. 2016. “Modeling and
Gao, B., X. Huang, J. Shi, Y. Tai, and J. Zhang. 2020. “Hourly Forecasting
Forecasting Hourly Global Solar Radiation Using Clustering and
of Solar Irradiance Based on CEEMDAN and Multi-strategy CNN-
Classification Techniques.” Solar Energy 135: 682–691. doi:10.1016/j.
LSTM Neural Networks.” Renewable Energy 162: 1665–1683.
solener.2016.06.039.
doi:10.1016/j.renene.2020.09.141.
Kasu, M., and S. U. T. Hara. 2017. “High Speed Measurements of
Gigoni, L., A. Betti, E. Crisostomi, A. Franco, M. Tucci, F. Bizzarri, and D.
Generated Power and Its Relationship to Weather Observations at
Mucci. 2018. “Day-Ahead Hourly Forecasting of Power Generation
Yoshinogari Mega Solar Power Plant.” 2017 IEEE 44th
from Photovoltaic Plants.” IEEE Transactions on Sustainable Energy
Photovoltaic Specialists Conference PVSC 1–4. doi:10.1109/
9 (2): 831–842. doi:10.1109/TSTE.2017.2762435.
PVSC.2017.8366441.
Guan, Y., B. Zhang, and X. Yan. 2019. “Accurate Short-term
Khahro, S. F., K. Tabbassum, S. Talpur, M. B. Alvi, X. Liao, and L.
Forecasting for Photovoltaic Power Method Using RBM
Dong. 2015. “Evaluation of Solar Energy Resources by Establishing
Combined LSTM-RNN Structure with Weather Factors
Empirical Models for Diffuse Solar Radiation on Tilted Surface and
Quantification. iSPEC 2019-2019 IEEE Sustain.” Power and
Analysis for Optimum Tilt Angle for a Prospective Location in
Energy Conference: Grid Modernization for Energy Revolution
Southern Region of Sindh, Pakistan.” International Journal of
797–802. doi:10.1109/iSPEC48194.2019.8975309.
Electrical Power & Energy Systems 64: 1073–1080. doi:10.1016/j.
Gupta, P., and R. Singh. 2021. “Univariate Model for Hour Ahead
ijepes.2014.09.001.
Multi-step Solar Irradiance Forecasting.” 2021 IEEE 48th
Photovoltaic Specialists Conference (PVSC). Fort Lauderdale, FL, Khosravi, A., R. N. N. Koury, L. Machado, and J. J. G. Pabon. 2018.
USA, 0494–0501. doi:10.1109/PVSC43889.2021.9519002. “Prediction of Hourly Solar Radiation in Abu Musa Island Using
Gutierrez-Corea, F. V., M. A. Manso-Callejo, M. P. Moreno-Regidor, and Machine Learning Algorithms.” Journal of Cleaner Production 176:
M. T. Manrique-Sancho. 2016. “Forecasting Short-term Solar 63–75. doi:10.1016/j.jclepro.2017.12.065.
Irradiance Based on Artificial Neural Networks and Data from King, D. L., S. Gonzalez, and G. M. Galbraith, Boyson W.E. 2007.
Neighboring Meteorological Stations.” Solar Energy 134: 119–131. “Performance Model for Grid-Connected Photovoltaic Inverters.”
doi:10.1016/j.solener.2016.04.020. Sandia National Laboratory, United States, 38. doi: 10.2172/920449.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1753

Kisi, O., S. Heddam, and Z. M. Yaseen. 2019. “The Implementation of Marzouq, M., Z. Bounoua, H. El Fadili, A. Mechaqrane, K. Zenkouar,
Univariable Scheme-based Air Temperature for Solar Radiation and Z. Lakhliai. 2019. “New Daily Global Solar Irradiation
Prediction: New Development of Dynamic Evolving Neural-fuzzy Estimation Model Based on Automatic Selection of Input
Inference System Model.” Applied Energy 241: 184–195. doi:10.1016/ Parameters Using Evolutionary Artificial Neural Networks.”
j.apenergy.2019.03.089. Journal of Cleaner Production 209: 1105–1118. doi:10.1016/j.
Kroposki, B. 2017. „Integrating High Levels of Variable Renewable jclepro.2018.10.254.
Energy into Electric Power Systems.„ Journal of Modern Power Meenal, R., and A. I. Selvakumar. 2018. “Assessment of SVM, Empirical
System and Clean Energy Vol. 5: 831–837. doi:10.1007/s40565-017- and ANN Based Solar Radiation Prediction Models with Most
0339-3. Influencing Input Parameters.” Renewable Energy 121: 324–343.
Kumari, P., and D. Toshniwal. 2021. “Extreme Gradient Boosting and doi:10.1016/j.renene.2017.12.005.
Deep Neural Network Based Ensemble Learning Approach to Forecast Meng, B., R. C. G. M. Loonen, and J. L. M. Hensen. 2020. “Data-driven
Hourly Solar Irradiance.” Journal of Cleaner Production 279: 123285. Inference of Unknown Tilt and Azimuth of Distributed PV Systems.”
doi:10.1016/j.jclepro.2020.123285. Solar Energy 211: 418–432. doi:10.1016/j.solener.2020.09.077.
Lan, H., C. Zhang, Y. Y. Hong, Y. He, and S. Wen. 2019. “Day-ahead Mishra, S., and P. Palanisamy. 2018. “Multi-time-horizon Solar
Spatiotemporal Solar Irradiation Forecasting Using Frequency-based Forecasting Using Recurrent Neural Network.” 2018 Energy
Hybrid Principal Component Analysis and Neural Network.” Applied Conversion Congress and Exposition ECCE 2018: 18–24. doi:10.1109/
Energy 247: 389–402. doi:10.1016/j.apenergy.2019.04.056. ECCE.2018.8558187.
Lara-Fanego, V., J. A. Ruiz-Arias, D. Pozo-Vazquez, and F. J. T.-P. J. Mousavi, S. M., E. S. Mostafavi, and P. Jiao. 2017. “Next Generation
Santos-Alamillos. 2012. “Evaluation of the WRF Model Solar Prediction Model for Daily Solar Radiation on Horizontal Surface
Irradiance Forecasts in Andalusia.” Solar Energy 86 (8): 2200–2217. Using a Hybrid Neural Network and Simulated Annealing Method.”
doi:10.1016/j.solener.2011.02.014. Energy Conversion and Management 153: 671–682. doi:10.1016/j.
Lave, M., C. Hansen, Y. Ueda, and K. Hakuta. 2017. “Targeted Evaluation enconman.2017.09.040.
of Utility-Scale and Distributed Solar Forecasting.” 2017 IEEE 44th Munkhammar, J., D. Van Der Meer, and J. Widen. 2019. “Probabilistic
Photovoltaic Specialists Conference PVSC 1–6. doi:10.1109/ Forecasting of the Clear-sky Index Using Markov-chain Mixture
PVSC.2017.8366281. Distribution and Copula Models.” Conference Record of the IEEE
Lave, M., W. Hayes, A. Pohl, and C. W. Hansen. 2015. “Evaluation of Photovoltaic Specialists Conference 2428–2433. doi:10.1109/
Global Horizontal Irradiance to Plane-of-array Irradiance Models at PVSC40753.2019.8980952.
Locations across the United States.” IEEE Journal of Photovoltaics 5 (2): Narvaez, G., L. Felipe, M. Bressan, and A. Pantoja. 2020. “Machine
597–606. doi:10.1109/JPHOTOV.2015.2392938. Learning for Site-adaptation and Solar Radiation Forecasting.”
Lee, J., W. Wang, F. Harrou, and Y. Sun. 2020. “Reliable Solar Irradiance Renewable Energy. doi:10.1016/j.renene.2020.11.089.
Prediction Using Ensemble Learning-based Models: A Comparative Oka, N., Y. Takahashi, K. Fujiwara, K. Hidaka, and H. Morita. 2018.
Study.” Energy Conversion and Management 208: 112582. doi:10.1016/ “Power Generation Evaluation of Large-Scale Photovoltaic Systems
j.enconman.2020.112582. Located on Inclined Plane. „2017 IEEE 44th Photovoltaic Specialists
Li, J., J. K. Ward, J. Tong, L. Collins, and G. Platt. 2016. “Machine Conference (PVSC). Washington, DC, 1973–1978. doi: 10.1109/
Learning for Solar Irradiance Forecasting of Photovoltaic System.” pvsc.2017.8366287. [Accessed: 15 Sept 2020]
Renewable Energy 90: 542–553. doi:10.1016/j.renene.2015.12.069. Pan, M., C. Li, R. Gao, Y. Huang, H. You, T. Gu, and F. Qin. 2020.
Li, P., K. Zhou, X. Lu, and S. Yang. 2020. “A Hybrid Deep Learning Model “Photovoltaic Power Forecasting Based on a Support Vector Machine
for Short-term PV Power Forecasting.” Applied Energy 259: 114216. with Improved Ant Colony Optimization.” Journal of Cleaner
doi:10.1016/j.apenergy.2019.114216. Production 277: 123948. doi:10.1016/j.jclepro.2020.123948.
Liaoliao, W., C. Chenggang, Y. Ning, and C. Hui. 2017. “Evaluation of Pang, Z., F. Niu, and Z. O’Neill. 2020. “Solar Radiation Prediction Using
Solar Irradiance on Inclined Surfaces Models in the Short-term Recurrent Neural Network and Artificial Neural Network: A Case
Photovoltaic Power Forecasting.” The Journal of Engineering Study with Comparisons.” Renewable Energy 156: 279–289.
2017 (13): 2226–2230. doi:10.1049/joe.2017.0726. doi:10.1016/j.renene.2020.04.042.
Liu, J., W. Fang, X. Zhang, and C. Yang. 2015. “An Improved Photovoltaic Passow, K., M. Lee, and A. F. Panchula. 2017. “Evaluation of Diffuse
Power Forecasting Model with the Assistance of Aerosol Index Data.” Decomposition Models.” 2017 IEEE 44th Photovoltaic Specialists
IEEE Transactions on Sustainable Energy 6 (2): 434–442. doi:10.1109/ Conference PVSC 2017: 1–5. doi:10.1109/PVSC.2017.8366172.
TSTE.2014.2381224. Paulescu, M., E. Paulescu, P. Gravilla, and V. Badescu. 2013 “Weather
Liu, L., M. Zhan, and Y. Bai. 2019. “A Recursive Ensemble Model for Modeling and Forecasting of PV Systems Operation.” In Green Energy
Forecasting the Power Output of Photovoltaic Systems.” Solar Energy and Technology. Springer, London. doi:10.1007/978-1-4471-4649-0.
189: 291–298. doi:10.1016/j.solener.2019.07.061. Pedro, H. T. C., and C. F. M. Coimbra. 2012. “Assessment of
Liu, Z. F., L. L. Li, M. L. Tseng, and M. K. Lim. 2020. “Prediction Short- Forecasting Techniques for Solar Power Production with No
term Photovoltaic Power Using Improved Chicken Swarm Optimizer - Exogenous Inputs.” Solar Energy 86 (7): 2017–2028. doi:10.1016/j.
Extreme Learning Machine Model.” Journal of Cleaner Production 248: solener.2012.04.004.
119272. doi:10.1016/j.jclepro.2019.119272. Pedro, H. T. C., and C. F. M. Coimbra. 2015. “Nearest-neighbor
Lou, S., D. H. W. Li, J. C. Lam, and W. W. H. Chan. 2016. “Prediction of Methodology for Prediction of Intra-hour Global Horizontal and
Diffuse Solar Irradiance Using Machine Learning and Multivariable Direct Normal Irradiances.” Renewable Energy 80: 770–782.
Regression.” Applied Energy 181: 367–374. doi:10.1016/j. doi:10.1016/j.renene.2015.02.061.
apenergy.2016.08.093. Pedro, H. T. C., C. F. M. Coimbra, M. David, and P. Lauret. 2018.
Marchesoni-Acland, F., P. Lauret, A. Gomez, and R. Alonso-Suarez. 2019. “Assessment of Machine Learning Techniques for Deterministic and
“Analysis of ARMA Solar Forecasting Models Using Ground Probabilistic Intra-hour Solar Forecasts.” Renewable Energy 123: 191–
Measurements and Satellite Images.” Conference Record of the IEEE 203. doi:10.1016/j.renene.2018.02.006.
Photovoltaic Specialists Conference 2445–2451. doi:10.1109/ Peters, I. M., and T. Buonassisi. 2000. “The Impact of Global Warming on
PVSC40753.2019.8980821. Silicon PV Energy Yield in 2100.” 5–7.
Marzouq, M., H. El Fadili, K. Zenkouar, Z. Lakhliai, and M. Amouzg. Prasad, R., M. Ali, P. Kwan, and H. Khan. 2019. “Designing a Multi-stage
2020. “Short Term Solar Irradiance Forecasting via a Novel Multivariate Empirical Mode Decomposition Coupled with Ant
Evolutionary Multi-model Framework and Performance Assessment Colony Optimization and Random Forest Model to Forecast
for Sites with No Solar Irradiance Data.” Renewable Energy 157: 214– Monthly Solar Radiation.” Applied Energy 236: 778–792. doi:10.1016/
231. doi:10.1016/j.renene.2020.04.133. j.apenergy.2018.12.034.
1754 P. GUPTA AND R. SINGH

Prasad, R., M. Ali, Y. Xiang, and H. Khan. 2020. “A Double Sobri, S., S. Koohi-kamali, and N. A. Rahim. 2020. “Solar Photovoltaic
Decomposition-based Modelling Approach to Forecast Weekly Solar Generation Forecasting Methods : A Review Number of Day.” Energy
Radiation.” Renewable Energy 152: 9–22. doi:10.1016/j. Conversion and Management 156: 459–497. doi:10.1016/j.
renene.2020.01.005. enconman.2017.11.019.
Ramsami, P., and V. Oree. 2015. “A Hybrid Method for Sun, H., D. Gui, B. Yan, Y. Liu, W. Liao, Y. Zhu, C. Lu, and N. Zhao. 2016.
Forecasting the Energy Output of Photovoltaic Systems.” Energy “Assessing the Potential of Random Forest Method for Estimating
Conversion and Management 95: 406–413. doi:10.1016/j. Solar Radiation Using Air Pollution Index.” Energy Conversion and
enconman.2015.02.052. Management 119: 121–129. doi:10.1016/j.enconman.2016.04.051.
Rana, M., I. Koprinska, and V. G. Agelidis. 2016. “Univariate and Sun, S., S. Wang, G. Zhang, and J. Zheng. 2018. “A Decomposition-cluster­
Multivariate Methods for Very Short-term Solar Photovoltaic Power ing-ensemble Learning Approach for Solar Radiation Forecasting.” Solar
Forecasting.” Energy Conversion and Management 121: 380–390. Energy 163: 189–199. doi:10.1016/j.solener.2018.02.006.
doi:10.1016/j.enconman.2016.05.025. Takilalte, A., S. Harrouni, and M. R. M.-L.-L. Yaiche. 2020. “New
Rao, D. V. S. K., M. Premalatha, and C. Naveen. 2018. “Analysis of Approach to Estimate 5-min Global Solar Irradiation Data on Tilted
Different Combinations of Meteorological Parameters in Predicting Planes from Horizontal Measurement.„ Renewable Energy 145: 2477–
the Horizontal Global Solar Radiation with ANN Approach: A Case 2488. doi: 10.1016/j.renene.2019.07.165. [Accessed: 20 Dec 2020].
Study.” Renewable and Sustainable Energy Reviews 91: 248–258. Tan, J, C. Deng. 2017. “Ultra-Short-term Photovoltaic Generation
doi:10.1016/j.rser.2018.03.096. Forecasting Model Based on Weather Clustering and Markov
Raza, M. Q., N. Mithulananthan, and A. Summerfield. 2018. “Solar Chain.” 2017 IEEE 44th Photovoltaic Specialists Conference (PVSC).
Output Power Forecast Using an Ensemble Framework with Neural Washington, DC, USA, 2017: 1–5. doi:10.1109/PVSC.2017.8366516.
Predictors and Bayesian Adaptive Combination.” Solar Energy 166: Tina, G. M., and C. Ventura. 2011. “Sub-hourly Irradiance Models on the
226–241. doi:10.1016/j.solener.2018.03.066. Plane of Array for Photovoltaic Energy Forecasting Applications 1321–
Raza M.Q., M. Nadarajah, and C. Ekanayake. 2016. “On recent advances 1326.”
in PV output power forecast.” Solar Energy 136: 125–144. doi: 10.1016/ VanDeventer, W., E. Jamei, G. S. Thirunavukkarasu, M. Seyedmahmoudian,
j.solener.2016.06.073. T. K. Soon, B. Horan, S. Mekhilef, and A. Stojcevski. 2019. “Short-term PV
Ren, Y., P. N. Suganthan, and N. Srikanth. 2015. “Ensemble Methods for Power Forecasting Using Hybrid GASVM Technique.” Renewable Energy
Wind and Solar Power Forecasting — A State-of-the-art Review.” 140: 367–379. doi:10.1016/j.renene.2019.02.087.
Renewable and Sustainable Energy Reviews 50: 82–91. doi:10.1016/j. Visser, L., T. Alskaif, and W. Van Sark. 2019. “Benchmark Analysis of
rser.2015.04.081. Day-ahead Solar Power Forecasting Techniques Using Weather
Riley, E., M. Lave, E. Wu, J. Dise, T. Tirumalai, J. L. Bosch, and C. Predictions.” 2019 IEEE 46th Photovoltaic Specialists Conference.
Tammineedi. 2017. “On the Ability of Ground Based Global Chicago, IL, USA, 2019: 2111–2116. doi:10.1109/
Horizontal Irradiance Measurements to Reduce Error in Satellite PVSC40753.2019.8980899.
Derived Plane of Array Irradiance Data for Fixed Tilt Photovoltaic Voyant, C., G. Notton, C. Darras, A. Fouilloy, and F. Motte. 2017a.
Power Plants.” 2017 IEEE 44th Photovoltaic Specialists Conference “Uncertainties in Global Radiation Time Series Forecasting Using
(PVSC). Portland, OR, USA, 1–4. doi:10.1109/PVSC.2017.8366680. Machine Learning: The Multilayer Perceptron Case.” Energy 125:
Rosiek, S., J. Alonso-Montesinos, and F. J. Batlles. 2018. “Online 3-h 248–257. doi:10.1016/j.energy.2017.02.098.
Forecasting of the Power Output from a BIPV System Using Voyant, C., G. Notton, S. Kalogirou, M. L. Nivet, C. Paoli, F. Motte, and A.
Satellite Observations and ANN.” International Journal of Fouilloy. 2017b. “Machine Learning Methods for Solar Radiation
Electrical Power & Energy Systems 99: 261–272. doi:10.1016/j. Forecasting: A Review.” Renewable Energy 105: 569–582. doi:10.1016/
ijepes.2018.01.025. j.renene.2016.12.095.
Seyedmahmoudian, M., E. Jamei, G. S. Thirunavukkarasu, T. K. Soon, M. Wang, G., Y. Su, and L. Shu. 2016. “One-day-ahead Daily Power
Mortimer, B. Horan, A. Stojcevski, and S. Mekhilef. 2018. “Short-term Forecasting of Photovoltaic Systems Based on Partial Functional
Forecasting of the Output Power of a Building-integrated Photovoltaic Linear Regression Models.” Renewable Energy 96: 469–478.
System Using a Metaheuristic Approach.” Energies 11 (5): 1260. doi:10.1016/j.renene.2016.04.089.
doi:10.3390/en11051260. Wang, H., R. Cai, B. Zhou, S. Aziz, B. Qin, N. Voropai, L. Gan, and E.
Shamshirband, S., K. Mohammadi, H. Khorasanizadeh, P. L. Yee, M. Lee, Barakhtenko. 2020. “Solar Irradiance Forecasting Based on Direct
D. Petković, and E. Zalnezhad. 2016. “Estimating the Diffuse Solar Explainable Neural Network.” Energy Conversion and Management
Radiation Using a Coupled Support Vector Machine-wavelet 226: 113487. doi:10.1016/j.enconman.2020.113487.
Transform Model.” Renewable & Sustainable Energy Reviews 56: Wang, J., R. Ran, and Y. Zhou. 2017. “A Short-term Photovoltaic Power
428–435. doi:10.1016/j.rser.2015.11.055. Prediction Model Based on an FOS-ELM Algorithm.” Applied Science
Shamshirband, S., K. Mohammadi, P. L. Yee, D. Petković, and A. 7. doi:10.3390/app7040423.
Mostafaeipour. 2015. “A Comparative Evaluation for Identifying the Wang, L., O. Kisi, M. Zounemat-Kermani, G. A. Salazar, Z. Zhu, and W.
Suitability of Extreme Learning Machine to Predict Horizontal Global Gong. 2016. “Solar Radiation Prediction Using Different Techniques:
Solar Radiation.” Renewable and Sustainable Energy Reviews 52: 1031– Model Evaluation and Comparison.” Renewable & Sustainable Energy
1042. doi:10.1016/j.rser.2015.07.173. Reviews 61: 384–397. doi:10.1016/j.rser.2016.04.024.
Wen, Y., D. AlHakeem, P. Mandal, S. Chakraborty, Y.-K. Wu, T. Senjyu,
Sharadga, H., S. Hajimirza, and R. S. Balog. 2020. “Time Series S. Paudyal, and T.-L. Tseng. 2019. “Performance Evaluation of
Forecasting of Solar Power Generation for Large-scale Photovoltaic Probabilistic Methods Based on Bootstrap and Quantile Regression
Plants.” Renewable Energy 150: 797–807. doi:10.1016/j. to Quantify PV Power Point Forecast Uncertainty.” IEEE Transactions
renene.2019.12.131. on Neural Networks and Learning Systems 31 (4): 1134–1144.
Sharma, A., and A. Kakkar. 2018. “Forecasting Daily Global Solar doi:10.1109/tnnls.2019.2918795.
Irradiance Generation Using Machine Learning.” Solar Energy 82: Wu, E., M. Z. Zapata, L. Delle Monache, and J. Kleissl. 2019a.
2254–2269. doi:10.1016/j.rser.2017.08.066. “Observation-Based Analog Ensemble Solar Forecast in Coastal
Sheng, H., J. Xiao, Y. Cheng, Q. Ni, and S. Wang. 2018. “Short-Term Solar California.” Conference Record of the IEEE Photovoltaic Specialists
Power Forecasting Based on Weighted Gaussian Process Regression.” Conference 2440–2444. doi:10.1109/PVSC40753.2019.8980546.
IEEE Transactions on Industrial Electronics 65 (1): 300–308. Wu, L., G. Huang, J. Fan, F. Zhang, X. Wang, and W. Zeng. 2019b.
doi:10.1109/TIE.2017.2714127. “Potential of Kernel-based Nonlinear Extension of Arps Decline
Shi, J., W. J. Lee, Y. Liu, Y. Yang, and P. Wang. 2012. “Forecasting Power Model and Gradient Boosting with Categorical Features Support for
Output of Photovoltaic Systems Based on Weather Classification and Predicting Daily Global Solar Radiation in Humid Regions.” Energy
Support Vector Machines.” IEEE Transactions on Industry Conversion and Management 183: 280–295. doi:10.1016/j.
Applications 48 (3): 1064–1069. doi:10.1109/TIA.2012.2190816. enconman.2018.12.103.
 INTERNATIONAL JOURNAL OF SUSTAINABLE ENGINEERING 1755

Yadav, A. K., V. Sharma, H. Malik, and S. S. Chandel. 2018. “Daily Array Zambrano, A. F., and L. F. Giraldo. 2020. “Solar Irradiance Forecasting
Yield Prediction of Grid-interactive Photovoltaic Plant Using Relief Models without On-site Training Measurements.” Renewable Energy
Attribute Evaluator Based Radial Basis Function Neural Network.” 152: 557–566. doi:10.1016/j.renene.2020.01.092.
Renewable and Sustainable Energy Reviews 81: 2115–2127. Zang, H., L. Liu, L. Sun, L. Cheng, Z. Wei, and G. Sun. 2020.
doi:10.1016/j.rser.2017.06.023. “Short-term Global Horizontal Irradiance Forecasting Based
Yagli, G. M., D. Yang, and D. Srinivasan. 2019. “Automatic Hourly Solar on a Hybrid CNN-LSTM Model with Spatiotemporal
Forecasting Using Machine Learning Models.” Renewable and Correlations.” Renewable Energy 160: 26–41. doi:10.1016/j.
Sustainable Energy Reviews 105: 487–498. doi:10.1016/j.rser.2019.02.006. renene.2020.05.150.
Yagli, G. M., D. Yang, O. Gandhi, and D. Srinivasan. 2020. “Can We Zeng, J., and W. Qiao. 2013. “Short-term Solar Power Prediction Using a
Justify Producing Univariate Machine-learning Forecasts with Support Vector Machine.” Renewable Energy 52: 118–127. doi:10.1016/
Satellite-derived Solar Irradiance? Appl.” Energy 259. doi:10.1016/j. j.renene.2012.10.009.
apenergy.2019.114122. Zheng, J., H. Zhang, Y. Dai, B. Wang, T. Zheng, Q. Liao, Y. Liang, F.
Yang, H. T., C. M. Huang, Y. C. Huang, and Y. S. Pai. 2014. “A Weather- Zhang, and X. Song. 2020. “Time Series Prediction for Output of
based Hybrid Method for 1-day Ahead Hourly Forecasting of PV Power Multi-region Solar Power Plants.” Applied Energy 257: 114001.
Output.” IEEE Transactions on Sustainable Energy 5 (3): 917–926. doi:10.1016/j.apenergy.2019.114001.
doi:10.1109/TSTE.2014.2313600. Zhou, Y., N. Zhou, L. Gong, and M. Jiang. 2020. “Prediction of
Yao, W., C. Zhang, H. Hao, X. Wang, and X. Li. 2018. “A Support Vector Photovoltaic Power Output Based on Similar Day Analysis, Genetic
Machine Approach to Estimate Global Solar Radiation with the Algorithm and Extreme Learning Machine.” Energy 204: 117894.
Influence of Fog and Haze.” Renewable Energy 128: 155–162. doi:10.1016/j.energy.2020.117894.
doi:10.1016/j.renene.2018.05.069.
Yin, W., Y. Han, H. Zhou, M. Ma, L. Li, and H. Zhu. 2020. “A Novel Non- Zhu, H., X. Li, Q. Sun, L. Nie, J. Yao, and G. Zhao. 2016. “A Power
iterative Correction Method for Short-term Photovoltaic Power Prediction Method for Photovoltaic Power Plant Based on Wavelet
Forecasting.” Renewable Energy 159: 23–32. doi:10.1016/j. Decomposition and Artificial Neural Networks.” Energies 9: 1–15.
renene.2020.05.134. doi:10.3390/en9010011.