Environ Monit Assess (2025) 197:310

https://doi.org/10.1007/s10661-025-13738-8

RESEARCH

Long‑term solar radiation forecasting in India using EMD,

EEMD, and advanced machine learning algorithms
T. RajasundrapandiyanLeebanon ·
N. S. Sakthivel Murugan · K. Kumaresan ·
Andrew Jeyabose

Received: 1 July 2024 / Accepted: 4 February 2025 / Published online: 18 February 2025
© The Author(s) 2025

Abstract Solar radiation plays a critical role in the Trivandrum) were used for training and testing. The
carbon sequestration processes of terrestrial ecosys- optimal model was identified based on performance
tems, making it a key factor in environmental sustain- metrics, including the highest linear correlation coef-
ability among various renewable energy sources. This ficient (R), and the lowest mean absolute error (MAE)
study integrates two advanced signal processing tech- and root mean square error (RMSE). The results
niques—empirical mode decomposition (EMD) and indicate that EEMD integrated with ML algorithms
ensemble empirical mode decomposition (EEMD)— consistently outperformed EMD-based approaches.
with machine learning (ML) algorithms, including Among the ML models evaluated, EEMD integrated
multilayer perceptron (MLP), random forest regres- with MLP achieved the best performance across all
sion (RFR), support vector regression (SVR), and locations, with RMSE = 0.332, MAE = 0.26, and
ridge regression, to forecast long-term solar radiation. R2 = 0.99. Furthermore, a comparative analysis with
Meteorological data spanning 13 years (2000–2012) previous studies demonstrated that the proposed
from seven locations across India (Bhubaneswar, approach provides superior accuracy, underscoring its
Chennai, Delhi, Hyderabad, Nagpur, Patna, and efficacy in solar radiation forecasting.

Keywords Long-term solar radiation · Machine

T. RajasundrapandiyanLeebanon (*) · N. S. S. Murugan
Department of Electrical and Electronics Engineering, learning · Empirical mode decomposition · Ensemble
TamilNadu College of Engineering, Coimbatore, empirical mode decomposition · Indian cities
Tamil Nadu, India
e-mail: [email protected]
N. S. S. Murugan Introduction
e-mail: [email protected]

K. Kumaresan Renewable energy sources, especially solar, geother-

Department of Mechanical Engineering, Park College mal, and wind energy are acknowledged for their
of Engineering & Technology, Coimbatore, TamilNadu, environmental benefits. Among these, solar energy
India
stands for its accessibility, easy availability and its
e-mail: [email protected]
non-polluting characteristics for its nonemission of
A. Jeyabose (*) harmful gases such as carbon dioxide (­ CO2) by Singh
Department of Computer Science and Engineering, et al. (2012). Solar energy plays a vital role in energy
Manipal Institute of Technology, Manipal Academy
security and sustainability (Alanazi et al., 2016). It
of Higher Education, Manipal, Karnataka 576104, India
e-mail: [email protected] is the primary source of energy in the atmosphere,

Vol.: (0123456789)
 310 Page 2 of 36 Environ Monit Assess (2025) 197:310

accounting for more than 99.97% of all untapped Machine learning for solar forecasting
energy (Blal et al., 2020). The sun’s energy received
as electromagnetic radiations is further converted into Machine learning models have been applied in this
power (Premalatha & Valan Arasu, 2016). The meas- field of solar radiation forecasting to capture complex
uring instruments such as pyranometers and pyrhe- relationships. Artificial neural network (ANN) is an
liometers are used to monitor the daily global solar efficient ML-based model for solar forecasting. For
radiation (GSR). However, this method is too expen- example Meenal and Selvakumar (2018) utilized seven
sive and time consuming. Solar forecasting is con- different Indian cities, integrates multiple empirical
sidered as an efficient alternative for effective energy models, in which, artificial neural networks (ANN)
usage. Approxiamately four million exajoules (1 achieved good results. Srivastava et al. (2019) com-
EJ = 1018 J) of solar energy reach the earth’s surface pared four ML techniques such as multivariate adaptive
in which 5 × 104 EJ is claimed to be harvested (Kabir regression splines (MARS), M5 (decision tree machine
et al., 2018). Energy imbalances can be tackled with learning methods), classification and regression tree
solar forecasting by reducing it to 19.65% (Kaur (CART), and random forest (RF) used for solar radia-
et al., 2016). In recent days, solar energy has attracted tion forecasting, concluding that RF model provided
many researchers which leads to the fast development best results for projecting solar radiation. Maldonado
of long-term solar forecasting methods. et al. (2019) developed a strategic framework for auto-
mated selection of lag in time series analysis with
support vector regression (SVR). This strategy was
Statistical methods for solar forecasting applied on four datasets for short-term predictions of
solar forecasting. However, the computational cost of
Short-term solar forecasting methods are divided the algorithms is observed to increase exponentially
into main categories: traditional statistical methods with the number of preselected variables. Vennila et al.
and artificial intelligence methods (Jebaraj & Iniyan, (2022) suggest that machine learning techniques, par-
2006; Yin et al., 2019). Traditional methods consists ticularly hybrid models and artificial neural networks,
of dynamic approach and empirical model. Based on are effective for improving the accuracy of solar energy
the sunshine duration, the dynamic approaches pre- production forecasting by accounting for weather
dicts short-term radiation while empirical approaches variability and optimizing model performance. Anu-
estimate GSR which leads to achieve high predictive radha et al. (2021) investigated random forest regres-
accuracy by Mohanty et al. (2016). Notable statisti- sor (RFR), SVM regressor, and linear regressor for
cal models includes ARIMA (autoregressive inte- the solar power–based generation forecast. The RFR
grated moving average) models, Bayesian estimation yielded the best results with accuracy of 94.01, RMSE
models, and Kalman filtering models, in which the 27.32, MAE 12.45, and MSE 746.48 respectively.
ARIMA algorithm captures temporal structures in Hedar et al. (2021) investigated the GPR technique for
time series data and are suitable for linear temporal the prediction of solar-based radiation. The proposed
patterns but not suitable for nonlinear temporal pat- model gave an RMSE of 421.15 W/m2 respectively.
terns in real world scenarios by Siami-Namini et al. Mohanad et al. (2018) adopted six daily climate vari-
(2018); the Angstrom–Prescott sunshine–based ables for the eleven major locations, and optimal data
empirical model is a widely used approach for esti- driven technique based on SVR model using PSO
mating global solar radiation with the performance algorithm. The PSO-SVR model outperformed MARS
metrics of mean bias error (MBE), mean percentage and SVR. Zhang and Hong (2019) support energy
error (MPE), and root mean square error (RMSE) policy and managers of power systems, a new fore-
(Mohanty et al., 2016; Ronno et al., 2017). Due to casting model, the CEEMDAN- SVRQDA model is
various climatic conditions, empirical models needs proposed to provide more accurate forecasts. Ghimire
long processing time for the prediction of GSR et al. (2023) developed an electricity forecast model by
(Perveen et al., 2018). Based on the recent studies, using multi-head self-attention transformer. This model
empirical models and hybrid support vector machine achieves high accuracy and low predictive errors.
are less efficient compared to machine learning tech- The main objective of this work is to improve power
niques (Fan et al., 2019, 2020). demand point prediction. Kundra and Sadawarti (2015)

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 3 of 36 310

introduce a novel approach of hybridization of cuckoo an exploration of existing research gaps. Wang et al.
search and particle swarm optimization for remote (2022) demonstrated the comparison between single,
sensing image classification. This model achieves an stand-alone models to hybrid models for solar radiation
accuracy of 96.33%. Deo et al. (2022) present an arti- prediction. According to their findings, single models
ficial intelligence (AI) approach that incorporates total offer lack of accuracy which indicates a clear need for
sky conditions, focusing on the impact of cloud cover advancement in modeling strategies and the develop-
variations to accurately model PPFD at 5-min time- ment of hybrid techniques. To increase the forecasting
scales. The novelty and contribution lie in developing accuracy of the ML and deep learning (DL) models,
the first deep learning AI method for real-time PPFD the feasible hybrid option is to utilize signal processing
forecasting, effectively capturing the influence of techniques to remove noise from data thereby enhanc-
cloud properties on measured photosynthetically active ing the signal clarity and extraction of relevant features
radiation. Dong et al. (2018) proposed seasonal SVR required to improve the prediction accuracy.
with CCS, called the support vector regression (SVR) For instance, AL‐Musaylh et al. (2021) introduce
with chaotic cuckoo search (SSVRCCS) model, which a novel empirical wavelet transform (EWT) tech-
is designed to enhance forecasting accuracy by effec- nique to analyze daily gas consumption demand pat-
tively capturing the non-linear and cyclical patterns terns in Melbourne, Australia, and forecast future
of electric load variations. This model achieves more demand using a hybrid decision tree (M5 model tree)
high forecasting accurate levels. Almarzooqi et al. model. The EWT algorithm decomposes the data
(2024) proposed hybrid framework which employs a into intrinsic mode functions (IMFs), capturing fre-
fast trainable statistical learning technique based on the quency patterns and stochastic behaviors, and signifi-
truncated-regularized kernel ridge regression model cantly improves forecasting accuracy, with the EWT-
for optimal prediction of grid-connected solar photo- M5 model tree outperforming traditional methods
voltaic (PV) power plants. Ahmed et al. (2024) develop by achieving a lower RRMSE of 29.19%. Aghmadi
a hybrid convolutional neural network long short-term et al. (2021) have combined EMD technique with
memory bidirectional gated recurrent unit forecast sys- back propagation in neural networks (BPNN) tech-
tem (CLSTM-BIGRU) trained to accurately predict nique for a hybrid-based solar forecasting. The model
significant wave height at multiple forecasting hori- has produced a root mean square error (RMSE) of
zons. Hong et al. (2019) present hybrid kernel–based 28.13 W/m2. Li et al. (2018) have proposed a hybrid
SVR to stimulate the motion of a floating platform EMD-ANN–based model. The model has yielded
with EEMD to forecast the motion data with reliable an optimal correlation coefficient of 0.93 in terms of
accuracy and effectiveness with a chaotic efficient bat monthly predicted values. To anticipate intra-hour
algorithm to receive an optimized parameter. solar photovoltaic energy for Vitoria-Gasteiz, Spain,
Rodríguez et al. (2022) have suggested a hybrid tech-
Literature‑based research gaps and motivation for the nique employing Daubechies wavelets and fast Fou-
proposed model rier neural network (FFNN). During validation, the
hybrid technique has attained an RMSE of 35.7 W/m2.
The application of machine learning in forecasting In order to perform solar forecasts for Odisha, India,
solar radiation holds significant promise; however, Majumder et al. (2018) presented variational mode
still uncertainties remain constant in developing mod- decomposition (VMD) with extreme learning machine
els and methodologies (Sivakumar et al., 2023). The technique. For a 15-min timeframe forecast, the pro-
primary cause of resulting randomness in predicted posed methodology has produced RMSE and mean
values caused by the inherent variability of solar radia- absolute percentage error (MAPE) values of 0.011 and
tion, which is amplified by environmental factors such 1.244, respectively. An innovative hybrid model for
as cloud cover, temperature, and other local weather electricity demand forecasting is proposed, integrat-
conditions and air quality such as particulate matter ing the VMD method, self-recurrent mechanism, tent
and dust, latitude, and season. Understanding these mapping function, out-bound-back mechanism, CS
complexities is crucial for enhancing predictive accu- algorithm, and SVR model, referred to as the VMD-
racy. To address these challenges, a comprehensive SR-SVRCBCS model. The performance of the model
literature analysis has been conducted, followed by receives the significance under 95 confident levels

Vol.: (0123456789)
 310 Page 4 of 36 Environ Monit Assess (2025) 197:310

(Zhang & Hong, 2019). Similar to this, there are a few 4. The multi-layer perceptron (MLP) with EEMD has
research on ML and hybrid models in the literature for given better performance compared with EMD in
predicting solar radiation; the authors, Li et al. (2018), terms of MAE, RMSE, and correlation coefficient.
Sivakumar et al. (2023), Mohanty et al. (2016), Shar-
iff and Duzan (2018), and Mohammadi et al. (2015),
have conducted solar forecasting for Indian regions. Solar radiation forecasting frameworks
Majumder et al. (2018) and Maldonado et al. (2019) and evaluation metrics
have demonstrated the integration of machine learning
(ML) methods with various signal processing methods To achieve the objectives of this research work, the
like EMD, VMD, and wavelet transforms. Despite of Indian regions has been identified from the relevant
this advancement, the potentials of ensemble empiri- datasets. The selection of locations and datasets
cal mode decomposition (EEMD) algorithm remain are based on specific criteria, including geographic
unexplored in the field of solar forecasting. Based on diversity and the availability of data. Well suitable
this, the following research gaps are explored. framework is designed for the implementation of
proposed model. To evaluate the effectiveness of
• This paper presents a novel hybrid model that the proposed algorithm, established performance
effectively addresses non-stationarity issues in metrics, including mean absolute error (MAE), root
multiple predictor inputs through a self-adaptive mean squared error (RMSE), and R squared (R2)
approach, while generating accurate forecasts of have been identified.
long-term solar radiation. This model demon-
strates enhanced potential for practical applica- Data regions and data acquisition
tions.
• To achieve this goal, EEMD addresses the mode The research work focus on seven major cities
mixing problem and improves robustness. known for its significant solar potential. The location
• The algorithms such as MLP, SVM, KNN, and includes Bhubaneswar, Chennai, Delhi, Hyderabad,
Ridge are becoming more prominent due to Nagpur, Patna, and Trivandrum. The Indian Mete-
recent advancements in AI; however, their inte- orological Department (IMD) in Pune provides the
gration with EEMD and other signal-processing input datasets needed for the analysis (https://​imdpu​
approaches for solar forecasting applications yield ne.​gov.​in/). Table 1 lists the geographic information
significant improvement in its performance. for the seven cities, including longitude, latitude, and
the time period during which data was collected. Due
to the inability to collect the most recent data for all
cities within the same timeframe, outdated datasets
Key contributions of the research work with varying timeframes were utilized. To ensure bet-
ter comparison of the hybrid models, homogeneous
1. The proposed methods focus on solar radiation parameters were selected, despite the differences in
forecasting with relevance of signal processing time periods among the chosen cities. The parameters
techniques such as empirical mode decomposition considered include bright sunlight hours (BHSS), day
(EMD) and ensemble empirical mode decomposi- length, and minimum (Tmin) and maximum (Tmax)
tion (EEMD) with four unique ML algorithms. temperatures.
2. Effectively decompose non-linear and non-sta-
tionary solar data using signal processing tech- Solar forecasting–based framework
niques such as EMD and EEMD based on tem-
poral-spatial distribution and its computational Forecasting-related research typically involves a large
complexity in terms of series of intrinsic mode datasets, as noted in the “Data regions and data acqui-
functions (IMFs). sition” section. The datasets contains a wide range of
3. Four ML techniques are applied to the extracted attributes and values, which leads to the improvement
features to show their efficacy in performance in of data quality. Various traits and values in the data-
terms of regression. set lead to the effective analysis of the framework.

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 5 of 36 310

Table 1  Geographical Cities Latitude Longitude Time period Climatic classification

parameters for seven
locations Bhubaneswar 20.296 85.824 2003–2008 Savanna tropical
Chennai 13.082 80.271 2003–2011 Savanna tropical
Delhi 28.704 77.102 2003–2011 Monsoon-influenced
humid subtropical and
semi-arid
Hyderabad 17.385 78.486 2000–2008 Tropical wet and dry
Nagpur 21.145 79.088 2004–2010 Savanna tropical
Patna 25.594 85.137 2000–2008 Humid subtropical
Trivandrum 8.524 76.936 2005–2012 Savanna tropical

Figure 1 shows the systematic architectural diagram Evaluation metrics for validating model performance
for solar radiation forecasting. To improve the qual-
ity of the dataset, the noisy values were removed and The performance metrics namely, RMSE, MAE,
missing values were addresses using mean imputa- and R, are utilized for this study to access the
tion. After the pre-processing phase, significant fea- result obtained using each model. RMSE repre-
tures such as BHSS, Tmin, and Tmax were identified. sents degree in dispersion which results in various
The signal processing methods such as EMD and patterns. For the best accuracy, values of RMSE
EEMD were applied to the pre-processed dataset to and MAE should be adjacent to zero while R value
generate decomposed values as IMF. Lastly, the ML should be nearer to 1. Equations (1), (2), and (3)
algorithms are loaded with these data as inputs to cal- describe the equations to compute RMSE, MAE,
culate the RMSE and MAE, respectively. and R2 values, respectively by Hedar et al. (2021).

Fig. 1  Signal processing-incorporated machine learning techniques for solar radiation forecasting framework

Vol.: (0123456789)
 310 Page 6 of 36 Environ Monit Assess (2025) 197:310

EMD process receives instant frequency data from

√ n
1∑
Root mean square error (RMSE) = (X (i) − Xpredicted (i))2 (1)
n k=1 actual
non-stationary and non-linear data which allows
splitting each complex dataset into several limited
1 ∑|
n
|2 and often small components. Empirical methods are
Mean absolute error (MAE) = | Xactual (i) − Xpredicted (i)|
n k=1 | | considered for processing decomposition signals to
(2) create inputs for MLP. Each signal is classified as
∑
(Xi− Xmean )(Yi− Ymean )
four or five IMFs to generate a series of signals and
Correlation coefficient (R) = � can reflect the specific physical meaning of the origi-
∑
(Xi− Xmean )2 (Yi− Ymean )2 nal signal. The IMFs should satisfy the subsequent
(3) conditions:

• The number of endpoints and zero crossings

must be equal or differ by a maximum of one
Methodologies during the whole duration of a single IMF.
• For each data location, the average envelope deter-
The dataset utilized in this research comprises data mined by the neighborhood maximum and enve-
from seven cities in India: Bhubaneswar, Chennai, lope determined by the local maximum is zero.
Delhi, Hyderabad, Nagpur, Patna, and Trivandrum.
The features considered include BHSS, minimum The original signal can be represented using
temperature (T_min), and maximum temperature (T_ Eq. (4).
max). To enhance the efficacy of the proposed analy-
∑
sis, constant features such as latitude, longitude, and y(x) = yx (a) + R(a) (4)
day length were excluded. The dataset was partitioned x
into training and test sets in a 75–25 ratio to ensure
The steps carried out in EMD could be described
robust evaluation and validation of the classification
in an algorithmic form is as follows.
models. This partitioning facilitates adequate data
Figures 2, 3, 4, 5, 6, 7, and 8 represent the IMFs
availability for model training while preserving a sig-
generated for the seven states we have considered in
nificant subset for testing.
this study by the EMD technique.
Empirical mode decomposition (EMD) is employed
to decompose the signal into intrinsic mode func-
tions (IMFs) based on local extrema. Each IMF must
Ensemble empirical mode decomposition
adhere to specific properties, including zero mean
and symmetry. Ensemble empirical mode decomposi-
The effect of modes merging is the fundamental
tion (EEMD) augments the original signal with white
downside of EMD. This happens when vibrations
noise to mitigate mode mixing and enhance robust-
with different time spans are stored in one IMF or
ness. Once the IMFs reach a normalized level, the fea-
vibrations with the same time scale are sifted into
tures extracted from these IMFs are utilized as input
different IMFs. Wu and Huang (2009) suggested
for four distinct regression models for solar radiation
a leading version of the EMD that solves this dif-
forecasting: multi-layer perceptron (MLP) regression,
ficulty. It marks a significant advancement in the
random forest regression (RFR), ridge regression, and
EMD approaches (Wu & Huang, 2009). Process of
support vector regression (SVR).
the EEMD is developed as follows:
Equation (5) represents the final outcome
Signal processing techniques acquired after the EEMD process. Here, an ensem-
ble of data sets yi(a) is created by adding white
Empirical mode decomposition noise xik(a) to the original time series y(a).

EMD is a flexible method of generating an infinite set 1 ∑t

yi (a) = x (a)
k−1 ik
(5)
of IMF components to represent the source data. The t

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 7 of 36 310

Algorithm 1  EMD algorithm

Input: Metrological Data

Output: IMFs
Step 1: Initialize x as 1 and R0(a) as y(a)
Step 2: Extract the IMF generated at nth iteration
Step 3: Equate H0(a) = Rx-1(a)
Step 4: Determine the local minima and local maxima for H x-1(a)
Step 5: By projecting the cubed surfaces, build the enclosure N x-1(a) represented by the maxima and the enclosure
Mx-1(a) represented by the minima for Hx-1(a).
Step 6: Compute mean of local maxima and minima, mean= [N x-1(a) - Mx-1(a)] /2
Step 7: Determine the xth element. H(x)= Hx-1(a) – mean
Step 8: Equate yx(a)=Hx(a) and Rx(a)= Rx-1(a) – yx(a)
Step 9: Repeat Step 2 -8 until the required count of IMFs are obtained.

Fig. 2  IMFs generated for Bhubaneshwar using EMD

Fig. 3  IMFs generated for Chennai using EMD

Vol.: (0123456789)
 310 Page 8 of 36 Environ Monit Assess (2025) 197:310

Fig. 4  IMFs generated for Hyderabad using EMD

Fig. 5  IMFs generated for Delhi using EMD

Figures 9, 10, 11, 12, 13, 14, and 15 represent the methods. This is applied for the prediction of the
IMFs generated for the seven states we have consid- outcome of an occurrence based on the continuous
ered in this study by the EEMD technique. value of the relationship between variables extracted
from datasets. The projected capacities of various ML
ML techniques regression techniques in this problem of solar radia-
tion is analyzed and compared. The corresponding
The role of predicting solar radiation is known as sections elaborate on the functional description of the
a supervised regression problem in the ML based four ML algorithms utilized in this work.

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 9 of 36 310

Fig. 6  IMFs generated for Nagpur using EMD

Fig. 7  IMFs generated for Patna using EMD

Ridge regression magnitude of coefficients in order to calculate data

over fitting. The correlation coefficients demonstrate
The very first process in ridge regression is to nor- which independent variables that are highly asso-
malize the variables (both dependent and inde- ciated with the dependent variable and with each
pendent) by dividing by their standard deviations other; if the measure of fit of the model is a small
and removing their mean values. All ridge regres- value, it means that the model is well matched to
sion computations are dependent on standardized the dataset. A control term forces the learning algo-
variables in terms of standardization. The resulting rithm to adjust the data and keep the weights as low
regression coefficients are rescaled to their original as possible. The regulated terms have an “alpha”
when they are presented. It is used to measure the parameter that controls the regularization of the

Vol.: (0123456789)
 310 Page 10 of 36 Environ Monit Assess (2025) 197:310

Fig. 8  IMFs generated for Trivandrum using EMD

Algorithm 1  EEMD algorithm

Input: Metrological Data

Output: IMFs
Step 1: Include a noise (preferably white) in the target data
Step 2: Break down the input to IMFs with its addition of noise.
Step 3: Repeat Steps 1 and 2 with the inclusion of a varied noise for every iteration.
Step 4: As an end result, compute the ensemble averages of the decompositions respective IMFs.

Fig. 9  IMFs generated for Bhubaneshwar using EEMD

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 11 of 36 310

Fig. 10  IMFs generated for Chennai using EEMD

Fig. 11  IMFs generated for Hyderabad using EEMD

model, i.e., helps in reducing ranking variants. It is strong connection between the explanatory variables.
represented using Eq. (6). This is frequently the case when a high number of
explanatory variables are used in the analysis by Shar-
A = Bc + E (6)
iff and Duzan (2018). If the multi-collinearity is high,
In above equation, A represents the dependent even when the regressors are practically completely
variable(s), B denotes the independent variables, while correlated, one or more regressors are effectively elim-
C is the regression-based coefficients to be evalu- inated while utilizing the ridge method.
ated, and E depicts the error term. It allows a user to
choose a range of regressor values that a user would Multilayer perceptron
not be able to use if you used least—square is used.
Positively correlated variables can be combined, and Multilayer perceptron (MLP) is a fundamental tech-
ridge regression can be used to reduce multi-collinear- nique in machine learning (Kalogirou, 2001). It
ity. In regression analysis, multi-collinearity refers to a is structured as feedforward neural network composed

Vol.: (0123456789)
 310 Page 12 of 36 Environ Monit Assess (2025) 197:310

Fig. 12  IMFs generated for Delhi using EEMD

Fig. 13  IMFs generated for Nagpur using EEMD

Fig. 14  IMFs generated for Patna using EEMD

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 13 of 36 310

Fig. 15  IMFs generated for Trivandrum using EEMD

of a series of inputs, hidden layers, and an output element, while 𝝋 (y) depicts the map function of
layer. Guermoui et al. (2020) analyzed this model projecting into the high l-D feature set. SVR has
based on time series data in which the model identi- been proven to be an effective tool in real-value
fies the suitable input features. Their findings indicate function estimation.
that the MLP model can perform statistically well
than the other models. Rabehi et al. (2018) analyzed
f (y) = F𝝋(y) + a (7)
the work of MLP in conjuction with boosted decision
trees and linear regression models to select appropri- Random forest regression
ate inputs within each regression model. Although
there are several radiation prediction models, the cur- The RFR model is an additive based model that cre-
rent research indicates MLP regressor predicts multi- ates forecasts through combining the results with
ple target variables and produces best results for con- several different models. As a forecast value, it
tinuous values. uses the mean or median of all outputs. This type
of models can be written more explicitly as Eq. (8).
Support vector regression
F(x) = g0 (x) + g1 (x)g2 (x) + ...gn (x) (8)
SVM was introduced in the year 1995 to complete
where F(x) is the addition of simpler core models
classification problem. This is extended to the area
“gi” where “i” iterates from 1 to n, n being the high-
of regression and estimation problem, therefore, is
est count of models. In this method, all baselines
referred to as SVR. One of the important advan-
are created independently with different subsample
tages of SVR is that the computational intricacy of
of the data. Although, a large number of trees can
a problem has no bearing on measure of input space.
produce high computational costs and takes up a
In addition, it has superior generalization capabili-
lot of memory, and even slower predictions, which
ties with high predictive accuracy. In recent dec-
can lead to challenges, but they are parallelizable,
ades, solar radiation has been accurately predicted
which means that we can do this around the process
under various climatic conditions, for example, in
of dividing several machines which will run. This
humid and arid regions of China, by Almaghrabi
leads to faster computation time. Because of lower
et al. (2022), Singh et al. (2012), and Thombare
parameter settings and faster applications, this also
et al. (2022). SVR creates an optimal hyper plane
performs better than other ML methods (Belgiu &
by projecting data for training into a 1-D feature
Drăguţ, 2016). So, in our research, we considered
set that also depicts the non-linear connection of
RFR which also helps overcome missing values and
inputs and outputs. SVR function is as mentioned in
maintains the accuracy as the solar radiation for a
Eq. (7). The forecast results are denoted by f(y), F
long-term data is estimated.
is the l-dimensional weight factor, a is an adjusting

Vol.: (0123456789)
 310 Page 14 of 36 Environ Monit Assess (2025) 197:310

Results ML models at each station. Among the MLP-EMD

and MLP-EEMD models, in terms of the metrics
In the proposed model, EEMD is often preferred MAE and R, the MLP-EEMD has acquired the best
for its robust decomposition capabilities, effectively results and when evaluated in terms of RMSE the
addressing the “mode mixing” issue prevalent in MLP-EMD has yielded the best results.
EMD. It strikes a favorable balance between com- The RFR-EMD model ranks as the next opti-
putational complexity and accuracy, particularly mum model which is succeeded by RIDGE-EEMD.
in noisy signal scenarios. A comparative analysis However, the SVR-EEMD model gives the low-
is done between empirical mode decomposition est accuracy. The eight models trained and evalu-
(EMD) and ensemble empirical mode decomposi- ated for RMSE in Table 2, and the diagrammatic
tion (EEMD) to which the input data features such representation is shown in Fig. 16. Similarily MAE
as Tmax, Tmin, and bright sunshine hours are taken evaluation and diagrammatic representation is shown
from top metropolitan cities in India. The output in Fig. 17. As in the preceding situation, the MLP-
value from both EMD and EEMD will undergo four EEMD model (on average RMSE of 0.306) provides
ML models for calculating statistical errors such the best prediction accuracy with regard to the eight
as RMSE, MAE, and R measured individually for models. It has its ranking in the order 0.204 for Chen-
every metropolitan stations. Among the four ML nai, 0.23 for Bhubaneswar, 0.266 for Hyderabad, 0.3
models, it is observed that when evaluated individu- for Trivandrum, 0.3028 for Patna, 0.42 for Nagpur,
ally for each metric, the RFR and RIDGE models and 0.421 for Delhi. But as stated previously, RFR
had given a good accuracy as well but when com- and RIDGE given better results as well with the
puted in terms of the performance metrics taken RFR-EEMD acquiring an average RMSE: 0.221 and
as a whole, the MLP models had a best accuracy. RIDGE: 0.263. The SVR model has given the low-
In order to implement the MLP, based on trial and est accurate values when compared with the other
error process, the hidden layers count was deter- three models. Same as previous metric evaluation,
mined. The count for hidden layer was not fixed for the MLP has yielded best accuracy. But as previously
any of the locations. For instance, for Hyderabad, stated, the MLP-EEMD has outperformed the MLP-
six counts were used; whereas for Patna and Trivan- EMD with an MAE: 0.26. It has its ranking in the
drum, five counts were used. Here, activation func- order: Patna (0.19), Trivandrum (0.202), Hyderabad
tions used were tangent (tanh) and logistic and (0.212), Delhi (0.241), Bhubaneshwar (0.31), Chen-
rectified linear unit (relu) respectively. For optimi- nai (0.324), and Nagpur (0.3439). RFR and RIDGE
zation limited-memory Broyden–Fletcher–Gold- have given better results as well. But for this metric
farb–Shanno (lbfgs), stochastic GD (SGD), and RFR-EMD has given the better results with an aver-
Adam methods were used. As a result, the MLP age RMSE: 0.1811 and RIDGE-EEMD: 0.218. The
models (on average RMSE of EMD: 0.306 and SVR-EMD model has given the lowest accurate val-
EEMD: 0.332, MAE of EMD: 0.27 and EEMD: ues when compared with the other three models.
0.26, and R of EMD: 0.9919 and EEMD: 0.9941) Table 3 and Fig. 17 depicts the performance met-
exhibit the best prediction accuracy among the eight rics of various machine learning algorithms integrated

Table 2  RMSE values City MLP RFR SVR RIDGE

obtained for the seven states
using EMD and EEMD EMD EEMD EMD EEMD EMD EEMD EMD EEMD

Chennai 0.204 0.381 0.3 0.15 1.23 0.93 0.317 0.24

Delhi 0.421 0.321 0.211 0.22 1.636 1.03 0.245 0.228
Hyderabad 0.266 0.227 0.301 0.234 0.56 1.05 0.31 0.3
Nagpur 0.42 0.416 0.252 0.239 2 2.03 0.247 0.262
Patna 0.303 0.58 0.159 0.17 0.349 0.924 0.225 0.287
Trivandrum 0.3 0.22 0.261 0.289 2.78 2.53 0.288 0.267
Bhubaneshwar 0.23 0.18 0.37 0.243 1.68 1.76 0.285 0.257

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 15 of 36 310

Fig. 16  Histograms for

RMSE values obtained for
the seven states using EMD
and EEMD

with EMD and EEMD for solar radiation forecasting for EEMD. In Nagpur, RFR performed similarly well,
are summarized across several cities. In Chennai, the yielding 0.1962 for EMD and 0.1957 for EEMD.
MLP model using EMD achieved the lowest RMSE Patna exhibited the best performance overall with RFR
of 0.158, while the EEMD integration resulted in a under EMD achieving the lowest RMSE of 0.126,
higher RMSE of 0.324. For random forest regres- while EEMD reached 0.14. Trivandrum and Bhubane-
sion (RFR), the EMD model produced an RMSE of swar showed varied results, with Trivandrum’s MLP
0.064, significantly lower than the EEMD’s 0.1055. yielding an RMSE of 0.27 (EMD) and 0.202 (EEMD),
In Delhi, the EMD-based MLP showed an RMSE of and Bhubaneswar’s RFR producing 0.285 for EMD
0.346, compared to 0.241 with EEMD, whereas RFR and 0.2 for EEMD. Overall, these results indicate sig-
delivered 0.163 for EMD and 0.1729 for EEMD. nificant variations in model performance depending
Hyderabad’s models displayed closely matched on both the machine learning approach and the data
results, with MLP achieving 0.225 for EMD and 0.212 processing technique used.

Vol.: (0123456789)
 310 Page 16 of 36 Environ Monit Assess (2025) 197:310

Fig. 17  Histograms for

MAE values obtained for
the seven states using EMD
and EEMD

Table 3  MAE values City MLP RFR SVR RIDGE

obtained for the seven states
using EMD and EEMD EMD EEMD EMD EEMD EMD EEMD EMD EEMD

Chennai 0.158 0.324 0.064 0.1055 1.08 0.792 0.25 0.19

Delhi 0.346 0.241 0.163 0.1729 1.47 0.779 0.2002 0.1816
Hyderabad 0.225 0.212 0.218 0.213 0.448 0.921 0.28 0.26
Nagpur 0.359 0.3439 0.1962 0.1957 1.847 1.76 0.2052 0.21723
Patna 0.2551 0.19 0.126 0.14 0.3158 0.862 0.1927 0.2426
Trivandrum 0.27 0.202 0.216 0.245 2.59 2.28 0.255 0.227
Bhubaneshwar 0.284 0.31 0.285 0.2 1.51 1.56 0.2231 0.211

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 17 of 36 310

Table 4  Correlation City MLP RFR SVR RIDGE

coefficient (R) values
obtained for the seven states EMD EEMD EMD EEMD EMD EEMD EMD EEMD
using EMD and EEMD
Chennai 0.998 0.992 0.964 0.975 0.14 0.076 0.93 0.955
Delhi 0.9975 0.9969 0.96 0.94 0.029 0.24 0.88 0.91
Hyderabad 0.994 0.998 0.98 0.97 0.31 0.4 0.94 0.9
Nagpur 0.993 0.998 0.9 0.97 0.22 0.42 0.84 0.95
Patna 0.984 0.98 0.97 0.89 0.63 0.39 0.875 0.94
Trivandrum 0.98 0.996 0.96 0.97 −0.051 0.276 0.92 0.98
Bhubaneshwar 0.997 0.9979 0.96 0.88 0.886 0.58 0.988 0.97

For our work, we have mainly considered the For our work, we mainly considered Perveen
paper (Meenal & Selvakumar, 2018) as our baseline et al. (2018) as our baseline model. To estimate
model. Meenal and Selvakumar 2018 have applied the global solar radiation (GSR), few models have
the SVM and ANN for the GSR forecast. The eight built for representing its empirical relationship. For
models trained and evaluated for R in Table 4, and the our work, we have used sunshine model, tempera-
diagrammatic representation is shown in Figs. 18, 19, ture model and hybrid model as a baseline model.
20, 21, 22, 23, 24 and 25. Tables 5, 6, 7, and 8 depict Table 5 depicts the training error of RMSE and R
the performance of our study with the baseline model values for Bhuvaneshwar. Comparing EMD and
for Bhubaneshwar, Chennai, Hyderabad, and Patna. EEMD with ANN, it is observed that the model
Figures 26, 27, 28, and 29 depict the R and RMSE EEMD yield best result. Perveen et al. (2018) have
values comparison obtained using our study and the applied the SVM and ANN for the GSR forecast.
baseline model. Similarly Tables 6, 7, and 8 depict the performance
Table 4 depicts the correlation coefficient values of our study with the baseline model for Chennai,
generated by the eight models. The maximum cor- Hyderabad, and Patna. Figures 26, 27, 28, and 29
relation coefficient was yielded by the MLP-EEMD depict the R and RMSE values comparison obtained
model with an average value of 0.9941 followed by using our study and the baseline model.
MLP-EMD (0.9919), RFR-EMD (0.956), RIDGE- The solar radiation is forecasted for the seven
EEMD (0.943), RFR-EEMD (0.942), RIDGE-EMD states by applying the SVM technique by the authors
(0.91), SVR-EEMD (0.34), and SVR-EMD (0.31) of our baseline model and generating multiple SVM
respectively in the order. The MLP-EEMD model has models. Based on the performance of the SVM mod-
correlation co-efficient ranked among the selected els, they have been allocated a rank. When compared
cities as follows: Hyderabad and Nagpur (0.998), with the SVM model ranked as best in the baseline
Bhubaneswar (0.9979), Delhi (0.9969) Trivandrum paper, our study has acquired better results using
(0.996), Chennai (0.992), and Patna (0.98). Figure 18 MLP. Table 9 illustrates the comparison of the results
shows the visual representation of R for all the sta- obtained in our study with the baseline model.
tions. Hence from the comparisons, it is found that Perveen et al. (2018) have applied the ANN tech-
MLP models have the best and SVR models have the nique for the solar forecast of the seven states which
least forecasting capability with least error using the is illustrated in Table 10. The authors have applied
selected three features—Tmin , Tmax, and BHSS. This the ANN technique for the sunshine model, temper-
could additionally be visually in terms of scatter plots ature model, and the hybrid models. Our study has
that represent the closeness of values predicted by indicated better results when applied with MLP for all
MLP vs. other algorithms. the outcomes yielded by the baseline model.
Figures 19, 20, 21, 22, 23, 24, and 25 represent Figure 30 illustrate the comparison between the
the variation between the values obtained for R by values acquired by all the machine learning mod-
the machine learning models applied in our study for els used in our study along with the values obtained
the seven states. It is observed that MLP has the best by the baseline model paper using ANN technique.
performance. Results indicate that MLP, RFR, and RIDGE have

Vol.: (0123456789)
 310 Page 18 of 36 Environ Monit Assess (2025) 197:310

Fig. 18  Histograms for

correlation coefficient (R)
values obtained for the
seven states using EMD and
EEMD

outperformed the ANN technique while the SVR tech- MLP-EEMD model, the existing prediction models
nique has not performed well in comparison with the have limited performance, as shown in Table 11.
ANN technique.
Table 10 compares the MLP-EEMD model
employed in this study for monthly solar radiation Discussion
forecasting with certain other GSR forecasting models
found in the literature, with the assessment conducted in In proposed work, the signal processing techniques
terms of RMSE. This is done to ensure that our research such as EMD and EEMD with novel ML algorithms
is effective. When compared to our newly constructed for solar radiation forecasting determines significant

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 19 of 36 310

Fig. 19  Scatter plots for

Bhubaneshwar based on the
eight models

efficacy in terms of errors RMSE, MAE, and cor- are applied with comparison with baseline models.
relation coefficient (R). The EEMD method pro- SVR deals with regression problems, to optimize
vides a significant improvement in the decomposi- the continuous valued function, while minimiz-
tion effect of the EMD method by reducing modal ing the prediction error. RFR is a meta-estimator
aliasing. The ML models like support vector regres- for regression problems, immune to overfitting and
sion (SVR), random forest regression (RFR), and learns spurious correlations of constructed model.
RIDGE regression, multi-layer perceptron (MLP) Ridge regression eliminates the bias coefficients and

Vol.: (0123456789)
 310 Page 20 of 36 Environ Monit Assess (2025) 197:310

Fig. 20  Scatter plots for

Chennai based on the eight
models

reduces the mean square error by shrinking the coef- algorithms, including long short-term memory
ficients of the model in order to reduce the problems (LSTM) and bidirectional LSTM (BiLSTM), were
of multi-collinearity, overfitting associated with applied in conjunction with the proposed decompo-
ordinary least squares regression. MLP is a univari- sition techniques. However, these algorithms exhib-
ate model used for forecasting problems by learning ited high error rates, with root mean square error
a series of past observations to predict the next value (RMSE) values of 0.82 and 0.90 and mean abso-
in the sequence. Additionally, several deep learning lute error (MAE) values of 0.60 and 0.62 scores.

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 21 of 36 310

Fig. 21  Scatter plots for

Delhi based on the eight
models

Furthermore, the computational time required to direction of a linear relationship between two or more
execute these algorithms was significantly high. variables. It ranges from − 1 to + 1. For our work, we
The optimal results are achieved in terms of evalu- have mainly considered the article of Meenal and Sel-
ation criteria such as RMSE, R2, and MAE. RMSE is vakumar (2018) as our baseline model. Meenal and
a statistical value which assess the largest expected Selvakumar (2018) have applied the SVM and ANN
error in the forecasted data. MAE provides the dif- for the GSR forecast. The four models trained and
ference between two set of data. A correlation coef- evaluated for R in Table 4, and the diagrammatic rep-
ficient is a number that measures the strength and resentation is shown in Fig. 18. Tables 5, 6, 7, and 8

Vol.: (0123456789)
 310 Page 22 of 36 Environ Monit Assess (2025) 197:310

Fig. 22  Scatter plots for

Nagpur based on the eight
models

depict the performance of our study with the baseline models found in the literature, with the assessment
model for Bhubaneshwar, Chennai, Hyderabad, and conducted in terms of RMSE. This is done to ensure
Patna. that our research is effective. When compared to our
Figures 26, 27, 28, and 29 depict the R and RMSE newly constructed EEMD-MLP model, the exist-
values comparison obtained using our study and the ing prediction models have limited performance,
baseline model. Table 10 compares the MLP-EEMD as shown in Table 11. The comparative analysis of
model employed in this study for monthly solar radia- various machine learning models for temperature pre-
tion forecasting with certain other GSR forecasting diction across different locations in India and other

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 23 of 36 310

Fig. 23  Scatter plots for

Patna based on the eight
models

countries reveals significant performance variations. a support vector machine (SVM) model across sev-
For instance, Premalatha and Valan Arasu (2016) eral Indian cities, demonstrating lower RMSE values,
employed an artificial neural network (ANN) model with Bhubaneswar recording an RMSE of 0.4223 and
using multiple input parameters, including latitude, an R2 of 0.99. Belaid and Mellit (2016) utilized sup-
longitude, altitude, and various meteorological fac- port vector machine (SVM) and multilayer perceptron
tors, achieving a root mean square error (RMSE) of (MLP) models in Algeria, obtaining RMSE values of
3.6461 and an R2 value of 0.9272 in Mumbai, India. 1.524 and 1.596 respectively, with R2 values of 0.986
Conversely, Meenal and Selvakumar (2018) utilized and 0.97, indicating strong predictive capabilities.

Vol.: (0123456789)
 310 Page 24 of 36 Environ Monit Assess (2025) 197:310

Fig. 24  Scatter plots for

Hyderabad based on the
eight models

Anuradha et al. (2021) employed random forest (LSTM) networks for photovoltaic (PV) module
regression (RFR) using mean of daily meteorologi- temperature forecasting in China, recording RMSE
cal data, resulting in a significantly higher RMSE of values of 1.30, 1.40, and 2.04 at different time inter-
27.32, highlighting challenges in accuracy when uti- vals. Furthermore, Hedar et al. (2021) applied Gauss-
lizing broader datasets. Additionally, Siami-Namini ian process regression (GPR) to predict temperature
et al. (2018) explored the application of convolutional based on various wind and irradiance factors, though
neural networks (CNN) and long short-term memory no specific performance metrics were provided. In

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 25 of 36 310

Fig. 25  Scatter plots for

Trivandrum based on the
eight models

Turkey, Citakoglu (2015) evaluated multiple empiri- approaches based on specific environmental variables
cal equations, including the Abdalla, Angstrom, and data characteristics.
Bahel, and Hargreaves–Samani equations, yielding Recent advancements in temperature predic-
RMSE values ranging from 2.062 to 43.95, with cor- tion methodologies highlight the effectiveness of
responding R2 values indicating varying degrees of various machine learning models across diverse
fit. These findings illustrate the diversity in predictive geographical regions. Dong and Jiang (2019)
modeling techniques and the necessity for tailored employed the CS-hard-ridge-RBF model utilizing

Vol.: (0123456789)
 310 Page 26 of 36 Environ Monit Assess (2025) 197:310

Table 5  Comparison of our R RMSE

study vs. the baseline model
for R and RMSE values for Baseline model (sunshine models) SVM 0.9756 0.6886
Bhubaneshwar
ANN (training) 0.9711 0.7466
ANN (testing) 0.9559 0.9377
Baseline model (temperature models) SVM 0.9697 0.8354
ANN (training) 0.9723 0.7408
ANN (testing) 0.9334 1.2112
Baseline model (hybrid models) SVM 0.9919 0.4223
ANN (training) 0.977 0.68
ANN (testing) 0.952 0.9553
Our study EMD 0.997 0.23
EEMD 0.9979 0.18

Table 6  Comparison of R RMSE

our study vs. the baseline
model R and RMSE values Baseline model (sunshine models) SVM 0.9465 0.8083
for Chennai
ANN (training) 0.9784 0.5437
ANN (testing) 0.9739 1.4468
Baseline model (temperature models) SVM 0.8881 1.1434
ANN (training) 0.9043 1.1199
ANN (testing) 0.7727 2.2197
Baseline model (hybrid models) SVM 0.9365 0.8673
ANN (training) 0.9963 0.2472
ANN (testing) 0.9835 1.1848
Our study EMD 0.998 0.204
EEMD 0.992 0.381

Table 7  Comparison of our R RMSE

study vs. the baseline model
R and RMSE values for Baseline model (sunshine models) SVM 0.9805 0.6278
Hyderabad
ANN (training) 0.9968 0.2722
ANN (testing) 0.9847 0.7451
Baseline model (temperature models) SVM 0.9704 0.8384
ANN (training) 0.9782 0.7212
ANN (testing) 0.9725 0.8929
Baseline model (hybrid models) SVM 0.9911 0.4205
ANN (training) 0.9968 0.229
ANN (testing) 0.9894 0.5814
Our study EMD 0.994 0.266
EEMD 0.998 0.227

12 meteorological factors in the USA, achiev- performance spectrum. In Nigeria, Olatomiwa

ing RMSE values ranging from 324.63 to 417.78, et al. (2015) investigated several models, includ-
while the DE-hard-ridge-RBF model reported ing a polynomial support vector machine (SVM)
values from 388.8 to 472.18, indicating a broad yielded an RMSE of 1.5102 and an R2 of 0.74,

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 27 of 36 310

Table 8  Comparison of Baseline model (sunshine models) SVM 0.9573 0.8735

our study vs. the baseline
model R and RMSE values ANN (training) 0.9601 0.9214
for Patna
ANN (testing) 0.9244 1.2003
Baseline model (temperature models) SVM 0.9277 1.0769
ANN (training) 0.969 0.7192
ANN (testing) 0.927 1.1807
Baseline model (hybrid models) SVM 0.9746 0.6515
ANN (training) 0.9765 0.6508
ANN (testing) 0.9465 0.9828
Our study EMD 0.984 0.3028
EEMD 0.98 0.58

Fig. 26  R and RMSE comparison of our study vs. the baseline models for Bhubaneshwar

Fig. 27  R and RMSE comparison of our study vs. the baseline models for Chennai

Vol.: (0123456789)
 310 Page 28 of 36 Environ Monit Assess (2025) 197:310

Fig. 28  R and RMSE comparison of our study vs. the baseline models for Hyderabad

Fig. 29  R and RMSE comparison of our study vs. the baseline models for Patna

while radial support vector regression (SVR) (2016) and Pan et al. (2020), showed that RMSE
showed a higher RMSE of 1.9059 with an R2 of values reached 0.997, leading to enhanced accu-
0.588. Additionally, different adaptive neuro-fuzzy racy during peak hours and nighttime. Akarslan
inference systems (ANFIS) produced RMSE values et al. (2018) further reinforced the robustness of
between 1.5485 and 2.0628, indicating the impact ANN models, achieving an R2 value of 0.9936,
of algorithm choice on predictive accuracy. Thom- while Mubiru and Banda (2008) reported an RMSE
bare et al. (2022) analyzed the performance of arti- of 0.385 in Uganda, showcasing the versatility and
ficial neural networks (ANN) in India, with RMSE efficacy of these approaches in different climatic
values improving from 1.8019 for five inputs to contexts. These findings underscore the critical role
1.5899 for seven inputs. of model selection and parameter optimization in
Jiang et al. (2020) presented an ANN model for enhancing predictive performance in meteorological
temperature prediction in China with an RMSE of applications. The proposed work, employs a multi-
0.746, integrating factors such as solar radiation and layer perceptron (MLP) model enhanced with ensem-
atmospheric pressure. Similarly, Jiang and Dong ble empirical mode decomposition (EEMD), yielded

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 29 of 36 310

Table 9  Comparison table of baseline models vs. our study

Testing

0.9553
0.9828
0.5814
1.1848
using R and RMSE

ANN (baseline-
City R RMSE

hybrid model)
Bhubanesh- Baseline SVM 0.9919 0.4223

Training

0.6508
0.2472
war model

0.229

0.68
Our study MLP-EMD 0.997 0.3
MLP-EEMD 0.9979 0.22

ANN (baseline-tem-

Testing

1.2112
1.1807
0.8929
2.2197
Chennai Baseline SVM 0.9465 0.8083
model

perature model)
Our study MLP-EMD 0.998 0.204
MLP-EEMD 0.992 0.381

Training

0.7408
0.7192
0.7212
1.1199
Hyderabad Baseline SVM 0.9911 0.4205
model
Our study MLP-EMD 0.998 0.266

Testing
MLP-EEMD 0.994 0.227

ANN (baseline-sun-

0.9377
1.2003
0.7451
1.4468
Patna Baseline SVM 0.9746 0.6515
model

shine model)
Our study MLP-EMD 0.984 0.3028

Training
Table 10  State wise comparison of RMSE values with all algorithms used in our study vs. baseline models (ANN)

0.7466
0.9214
0.2722
0.5437
MLP-EEMD 0.98 0.58
Delhi Our study MLP-EMD 0.9975 0.421
MLP-EEMD 0.9969 0.321

EEMD

0.2868
Nagpur Our study MLP-EMD 0.993 0.42

0.257
0.24
0.3
MLP-EEMD 0.998 0.4156
Trivandrum Our study MLP-EMD 0.98 0.3
RIDGE
MLP-EEMD 0.996 0.22

0.285
0.225
0.317
EMD

0.31
EEMD

impressive results with RMSE values ranging from

0.924
1.76
1.05
0.93
0.18 to 0.58 across various Indian cities, indicating
a high predictive accuracy with R2 values approach- 0.3488
ing 0.999. These findings underscore the efficacy of
EMD
SVR

1.68
0.56
1.23

machine learning approaches in meteorological pre-

dictions, highlighting the importance of model selec-
EEMD

0.1502

tion and input parameter optimization for enhancing

0.243
0.234
0.17

forecast reliability.
However, a significant limitation of the empirical
0.159
0.301

mode decomposition (EMD) algorithm is the phe-

EMD
RFR

0.37
0.3

nomenon of mode mixing, where a single intrinsic

mode function (IMF) can contain signals of vary-
EEMD

ing scales. Although the ensemble empirical mode

0.227
0.381

0.18
0.58

decomposition (EEMD) method addresses the mode

mixing issue, it results in residual white noise that
0.3028

remains non-negligible after reconstruction, com-

0.266
0.204
EMD
MLP

0.23

plicating the prediction process. To overcome this

limitation, the future work focus on implement-
Bhubaneshwar

ing a few signal decomposition techniques such

Hyderabad

as CEEMD and ICEEMD using deep learning

Chennai

algorithms.
Patna
City

Vol.: (0123456789)
 310 Page 30 of 36 Environ Monit Assess (2025) 197:310

Fig. 30  State wise comparison of RMSE values with all algorithms used in our study vs. baseline model for ANN

Conclusion 0.212 for Hyderabad, 0.4156 for Nagpur, 0.58

for Patna, 0.22 for Trivandrum, and 0.18 for Bhu-
This paper presents a novel method for forecasting baneshwar. When evaluated against MAE values,
solar radiation in various cities throughout India. our model yielded 0.324 for Chennai, 0.321 for
The cities included in the study are Delhi, Chen- Delhi, 0.227 for Hyderabad, 0.3439 for Nagpur,
nai, Hyderabad, Nagpur, Patna, Trivandrum, and 0.19 for Patna, 0.202 for Trivandrum and 0.31 for
Bhubaneswar. The approach utilized four machine Bhubaneshwar. Finally, when testing for R val-
learning algorithms: MLP, SVR, ridge regression, ues, the model gave 0.992 for Chennai, 0.9969 for
and RFR, along with two signal processing tech- Delhi, 0.998 for Hyderabad, 0.998 for Nagpur, 0.98
niques, EMD and EEMD. Among all the models for Patna, 0.996 for Trivandrum, and 0.9979 for
developed, the MLP-EEMD technique yielded the Bhubaneshwar. When assessed in terms of aver-
most optimum results in terms of the evaluation ages, the model has achieved an RMSE of 0.33 and
metrics. Among all the models built, the MLP- R of 0.994 respectively. In order to further evalu-
EEMD technique yielded the most optimum results ate the efficiency of our proposed work, a compre-
in terms of the evaluation metrics such as RMSE, hensive comparison with 18 prior studies, includ-
has achieved 0.381 for Chennai, 0.241 for Delhi, ing a baseline model, revealed that our model had

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 31 of 36 310

Table 11  Comparison of proposed model with other works

Author/reference Model Input parameters Location RMSE testing R R2

Premalatha and Valan ANN latitude, longitude, Mumbai, India 3.6461 0.9272 -
Arasu (2016) altitude, year,
month, mean ambi-
ent air temperature,
mean station level
pressure, mean wind
speed, and mean
relative humidity
Meenal and Selvaku- SVM Tmax(monthly mean Bhubaneswar, India 0.4223 0.9919 -
mar (2018) maximum tem- Chennai, India 0.8083 0.9465 -
peratures), Tmin
Hyderabad, India 0.4205 0.9911 -
(monthly mean
minimum tempera- Patna, India 0.6515 0.9746 -
tures), S (monthly
mean daily bright
sunshine hours), S0
(maximum possible
monthly mean daily
sunshine hours),
H0 (monthly mean
daily extraterrestrial
radiation on hori-
zontal surface)
ANN Tmax (monthly mean Bhubaneswar, India 0.9377 0.9559 -
maximum tem- 1.2112 0.9334 -
peratures), Tmin
0.9553 0.952 -
(monthly mean
minimum tempera- Chennai, India 1.4468 0.9739 -
tures), S (monthly 2.2197 0.7727 -
mean daily bright 1.1848 0.9835 -
sunshine hours), S0
Hyderabad, India 0.7451 0.9847 -
(maximum possible
monthly mean daily 0.8929 0.9725 -
sunshine hours) 0.5814 0.9894 -
Patna, India 1.2003 0.9244 -
1.1807 0.927 -
0.9828 0.9465 -
Belaid and Mellit SVM Measured temperature Algeria 1.524 0.986 -
(2016) MLP (tmin), extraterres- 1.596 0.97 -
trial solar radiation
(H0)
Anuradha et al. (2021) RFR Mean daily values India 27.32 - -
for air temperature,
humidity, wind
speed and direction,
visibility, average
values of pressure,
wind speed, and
electricity generated
Siami-Namini et al. CNN and LSTM PV module tem- China 7.5 min interval:1.30 - -
(2018) perature, the current, 15 min interval:1.40 - -
voltage, frequency,
30 min interval:2.04 - -
phases, and PV
power

Vol.: (0123456789)
 310 Page 32 of 36 Environ Monit Assess (2025) 197:310

Table 11  (continued)
Author/reference Model Input parameters Location RMSE testing R R2

Hedar et al. (2021) GPR Air temperature, aver- China 421.15r - -

age wind direct at
3 m, average wind
speed at 3 m, diffuse
horizontal irradiance
Citakoglu (2015) Abdalla equation Average sunshine Turkey 43.95 - 0.906
duration hours (n),
monthly average
daylight hours (N),
average temperature
(Tmean), average
humidity(Rhmean)
Angstrom equation Extra-terrestrial radia- 2.234 - 0.927
tion (Ra), sunshine
hours (n) and the
maximum monthly
mean sunshine
hours (N)
Bahel equation Daylight hours (N), 2.062 - 0.925
monthly mean sun-
shine duration (n)
Hargreaves–Samani Extra-terrestrial radia- 18.192 - 0.809
equation tion (Ra), monthly
mean maximum air
temperature (Tmax),
monthly mean mini-
mum air tempera-
ture (Tmin)
ANN Month number (M), 1.65 - 0.93
ANFIS extraterrestrial 1.691 0.926
radiation (Ra) aver-
MLR 4.745 - 0.81
age temperature
(Tmean) and average
relative humidity
(RHmean)
Dong and Jiang (2019) CS-hard-ridge-RBF 12 meteorological USA 324.63–417.78 - -
DE-hard-ridge-RBF factors 388.8—472.18 - -
Olatomiwa (2015) SVM-polynomial Monthly mean Nigeria 1.510218 - 0.74
SVR-radial maximum temperature 1.905994 - 0.588
(Tmax), monthly
ANFIS 2.0628 0.547
mean minimum
ANFIS-ACO temperature (Tmin) 1.5485 - 0.731
ANFIS-DE and monthly mean 1.601 - 0.693
ANFIS-GA sunshine duration 1.7696 - 0.635
(n)
ANFIS-PSO 1.8015 - 0.567
Thombare et al. ANN Five inputs India 1.80193 - -
(2022) Six inputs 1.68021 - -
Seven inputs 1.58994 - -
Jiang et al. (2020) ANN Solar radiation, sun- China 0.746 - -
shine hours, rainfall,
humidity, temp,
ATM (atmospheric
pressure)

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 33 of 36 310

Table 11  (continued)
Author/reference Model Input parameters Location RMSE testing R R2

Mohammadi et al. ANN Latitude, longitude, India 4.5 - -

(2015) altitude, time of day,
months, relative
humidity, rainfall,
air temperature, long
wave, length, and
wind speed
Jiang and Dong (2016) ANN Latitude, altitude, China 0.867 - 0.97
sunshine percentage,
clearness index
Pan et al. (2020) ANN Latitude, longitude, China 0.86 0.89 -
altitude
Akarslan et al. (2018) ANN Latitude, longitude, Turkey 0.9936 -
altitude, month,
mean land surface
temperature
Mubiru and Banda ANN Sunshine duration, Uganda 0.385 0.974 -
(2008) maximum tempera-
ture, cloud cover,
latitude, longitude,
altitude
Our study MLP-EEMD Bright sunshine Bhubaneswar, India 0.18 0.9979 -
hours(S), maximum Chennai, India 0.381 0.992 -
temperature (Tmax),
Delhi, India 0.321 0.9969 -
minimum tempera-
ture (Tmin) Hyderabad, India 0.227 0.998 -
Nagpur, India 0.4156 0.998 -
Patna, India 0.58 0.98 -
Trivandrum, India 0.22 0.996 -

outperformed its potential as reliable, cost effective, Data availability The data used in this research will be made
and time efficient solution to adapt for the long term available upon reasonable request to the authors.
solar forecasting. In our study, the machine learning Declarations
model outperformed the deep learning model due to
the relatively small size and nature of our dataset. Ethical approval All authors have read, understood, and have
For future work, we plan to update the dataset by complied as applicable with the statement on “Ethical respon-
extending it over more years and explore the adop- sibilities of Authors” as found in the Instructions for Authors.
tion of deep learning algorithms.
Declarations The authors have not used the AI tools to write
Author contribution T.R.L., and S.M., Conceptualization; or generate any content of this manuscript. However, they are
S.M., and K.K., Data curation; K.K., and A.J., Formal analy- used to correct spelling and grammar errors.
sis; T.R.L., and S.M., Investigation; T.R.L., and S.M., Meth-
odology; K.K., and A.J., Resources; S.M., and K.K., Supervi- Competing interests The authors declare no competing inter-
sion; T.R.L. S.M., K.K., and A.J., Validation; T.R.L., and S.M., ests.
Writing—original draft; K.K., and A.J., Writing—review &
editing. All authors read and approved the final version of the
manuscript. Open Access This article is licensed under a Creative Com-
mons Attribution 4.0 International License, which permits
Funding Open access funding provided by Manipal Academy use, sharing, adaptation, distribution and reproduction in any
of Higher Education, Manipal. medium or format, as long as you give appropriate credit to the

Vol.: (0123456789)
 310 Page 34 of 36 Environ Monit Assess (2025) 197:310

original author(s) and the source, provide a link to the Crea- Blal, M., Khelifi, S., Dabou, R., Sahouane, N., Slimani, A.,
tive Commons licence, and indicate if changes were made. The Rouabhia, A., Ziane, A., Neçaibia, A., Bouraiou, A.,
images or other third party material in this article are included & Tidjar, B. (2020). A prediction models for estimat-
in the article’s Creative Commons licence, unless indicated ing global solar radiation and evaluation meteorological
otherwise in a credit line to the material. If material is not effect on solar radiation potential under several weather
included in the article’s Creative Commons licence and your conditions at the surface of Adrar environment. Meas-
intended use is not permitted by statutory regulation or exceeds urement, 152, 107348. https://​doi.​org/​10.​1016/j.​measu​
the permitted use, you will need to obtain permission directly rement.​2019.​107348
from the copyright holder. To view a copy of this licence, visit Citakoglu, H. (2015). Comparison of artificial intelligence
http://creativecommons.org/licenses/by/4.0/. techniques via empirical equations for prediction of solar
radiation. Computers and Electronics in Agriculture, 118,
28–37. https://​doi.​org/​10.​1016/j.​compag.​2015.​08.​020
Deo, R. C., Grant, R. H., Webb, A., Ghimire, S., Igoe, D.
References P., Downs, N. J., Al-Musaylh, M. S., Parisi, A. V., &
Soar, J. (2022). Forecasting solar photosynthetic photon
Aghmadi, A., El Hani, S., Mediouni, H., Naseri, N., & El flux density under cloud cover effects: Novel predictive
Issaoui, F. (2021). Hybrid solar forecasting method based model using convolutional neural network integrated
on empirical mode decomposition and back propagation with long short-term memory network. Stochastic Envi-
neural network. E3S Web of Conferences, 231, 02001. ronmental Research and Risk Assessment, 36(10), 3183–
https://​doi.​org/​10.​1051/​e3sco​nf/​20212​31020​01 3220. https://​doi.​org/​10.​1007/​s00477-​022-​02188-0
Ahmed, A. A. M., Jui, S. J. J., AL-Musaylh, M. S., Raj, N., Dong, Y., Zhang, Z., & Hong, W. (2018). A hybrid seasonal
Saha, R., Deo, R. C., & Saha, S. K. (2024). Hybrid deep mechanism with a chaotic cuckoo search algorithm with
learning model for wave height prediction in Austral- a support vector regression model for electric load fore-
ia’s wave energy region. Applied Soft Computing, 150, casting. Energies, 11(4), 1009. https://​doi.​org/​10.​3390/​
111003. https://​doi.​org/​10.​1016/j.​asoc.​2023.​111003 en110​41009
Akarslan, E., Hocaoglu, F. O., & Edizkan, R. (2018). Novel short Dong, Y., & Jiang, H. (2019). Global solar radiation
term solar irradiance forecasting models. Renewable Energy, forecasting using square root regularization‐based
123, 58–66. https://​doi.​org/​10.​1016/j.​renene.​2018.​02.​048 ensemble. Mathematical Problems in Engineer-
Alanazi, Mohana, Abdulaziz Alanazi, and Amin Khodaei. ing, 2019(1). https://​doi.​org/​10.​1155/​2019/​96209​45
“Long-term solar generation forecasting.” In 2016 Fan, J., Wu, L., Zhang, F., Cai, H., Zeng, W., Wang, X., & Zou,
IEEE/PES transmission and distribution conference and H. (2019). Empirical and machine learning models for
exposition (T&D), pp. 1–5. IEEE, 2016. https://​doi.​org/​ predicting daily global solar radiation from sunshine dura-
10.​1109/​TDC.​2016.​75198​83 tion: A review and case study in China. Renewable and
Almaghrabi, S., Rana, M., Hamilton, M., & Rahaman, M. Sustainable Energy Reviews, 100, 186–212. https://​doi.​
S. (2022). Solar power time series forecasting utilising org/​10.​1016/j.​rser.​2018.​10.​018
wavelet coefficients. Neurocomputing, 508, 182–207. Fan, J., Wu, L., Ma, X., Zhou, H., & Zhang, F. (2020). Hybrid
https://​doi.​org/​10.​1016/j.​neucom.​2022.​08.​016 support vector machines with heuristic algorithms for
Almarzooqi, A. M., Maalouf, M., El-Fouly, T. H. M., Kat- prediction of daily diffuse solar radiation in air-polluted
zourakis, V. E., El, M. S., & Chrysikopoulos, C. V. regions. Renewable Energy, 145, 2034–2045. https://​doi.​
(2024). A hybrid machine-learning model for solar org/​10.​1016/j.​renene.​2019.​07.​104
irradiance forecasting. Clean Energy, 8(1), 100–110. Ghimire, S., Nguyen-Huy, T., AL-Musaylh, M. S., Deo, R. C.,
https://​doi.​org/​10.​1093/​ce/​zkad0​75 Casillas-Pérez, D., & Salcedo-Sanz, S. (2023). Integrated
AL-Musaylh, M. S., Al-Daffaie, K., & Prasad, R. (2021). Gas multi-head self-attention transformer model for electricity
consumption demand forecasting with empirical wavelet demand prediction incorporating local climate variables.
transform based machine learning model: A case study. Energy and AI, 14, 100302. https://​doi.​org/​10.​1016/j.​
International Journal of Energy Research, 45(10), egyai.​2023.​100302
15124–15138. https://​doi.​org/​10.​1002/​er.​6788 Guermoui, M., Melgani, F., Gairaa, K., & Mekhalfi, M. L.
Anuradha, K., Erlapally, D., Karuna, G., Srilakshmi, V., & (2020). A comprehensive review of hybrid models for solar
Adilakshmi, K. (2021). Analysis of solar power gen- radiation forecasting. Journal of Cleaner Production, 258,
eration forecasting using machine learning techniques. 120357. https://​doi.​org/​10.​1016/j.​jclep​ro.​2020.​120357
E3S Web of Conferences, 309, 01163. https://​doi.​org/​10.​ Hedar, A., Almaraashi, M., Abdel-Hakim, A. E., & Abdul-
1051/​e3sco​nf/​20213​09011​63 rahim, M. (2021). Hybrid machine learning for solar
Belaid, S., & Mellit, A. (2016). Prediction of daily and radiation prediction in reduced feature spaces. Energies,
mean monthly global solar radiation using support vec- 14(23), 7970. https://​doi.​org/​10.​3390/​en142​37970
tor machine in an arid climate. Energy Conversion and Hong, W., Li, M., Geng, J., & Zhang, Y. (2019). Novel chaotic
Management, 118, 105–118. https://​doi.​org/​10.​1016/j.​ bat algorithm for forecasting complex motion of floating
encon​man.​2016.​03.​082 platforms. Applied Mathematical Modelling, 72, 425–443.
Belgiu, M., & Drăguţ, L. (2016). Random forest in remote sens- https://​doi.​org/​10.​1016/j.​apm.​2019.​03.​031
ing: A review of applications and future directions. ISPRS Jebaraj, S., & Iniyan, S. (2006). A review of energy models.
Journal of Photogrammetry and Remote Sensing, 114, Renewable and Sustainable Energy Reviews, 10(4), 281–
24–31. https://​doi.​org/​10.​1016/j.​isprs​jprs.​2016.​01.​011 311. https://​doi.​org/​10.​1016/j.​rser.​2004.​09.​004

Vol:. (1234567890)
Environ Monit Assess (2025) 197:310 Page 35 of 36 310

Jiang, H., & Dong, Y. (2016). A nonlinear support vector networks. Solar Energy, 82(2), 181–187. https://​doi.​org/​
machine model with hard penalty function based on glow- 10.​1016/j.​solen​er.​2007.​06.​003
worm swarm optimization for forecasting daily global solar Olatomiwa, L., Mekhilef, S., Shamshirband, S., Mohammadi,
radiation. Energy Conversion and Management, 126, 991– K., Petković, D., & Sudheer, C. (2015). A support vector
1002. https://​doi.​org/​10.​1016/j.​encon​man.​2016.​08.​069 machine–Firefly algorithm-based model for global solar
Jiang, C., Mao, Y., Chai, Y., & Yu, M. (2020). Day-ahead radiation prediction. Solar Energy, 115, 632–644. https://​
renewable scenario forecasts based on generative adver- doi.​org/​10.​1016/j.​solen​er.​2015.​03.​015
sarial networks. International Journal of Energy Research, Pan, M., Li, C., Gao, R., Huang, Y., You, H., Gu, T., & Qin, F.
45(5), 7572–7587. https://​doi.​org/​10.​1002/​er.​6340 (2020). Photovoltaic power forecasting based on a support
Kabir, E., Kumar, P., Kumar, S., Adelodun, A. A., & Kim, vector machine with improved ant colony optimization.
K. (2018). Solar energy: Potential and future prospects. Journal of Cleaner Production, 277, 123948. https://​doi.​
Renewable and Sustainable Energy Reviews, 82, 894–900. org/​10.​1016/j.​jclep​ro.​2020.​123948
https://​doi.​org/​10.​1016/j.​rser.​2017.​09.​094 Perveen, G., Rizwan, M., & Goel, N. (2018). Develop-
Kalogirou, S. A. (2001). Artificial neural networks in renew- ment of empirical models for forecasting global solar
able energy systems applications: A review. Renewable energy. 2018 2nd IEEE International Conference on
and Sustainable Energy Reviews, 5(4), 373–401. https://​ Power Electronics, Intelligent Control and Energy Sys-
doi.​org/​10.​1016/​s1364-​0321(01)​00006-5 tems (ICPEICES), 317–324. https://​doi.​org/​10.​1109/​
Kaur, A., Nonnenmacher, L., Pedro, H. T., & Coimbra, C. F. icpei​ces.​2018.​88973​95
(2016). Benefits of solar forecasting for energy imbalance Premalatha, N., & Valan Arasu, A. (2016). Prediction of solar
markets. Renewable Energy, 86, 819–830. https://​doi.​org/​ radiation for solar systems by using ANN models with dif-
10.​1016/j.​renene.​2015.​09.​011 ferent back propagation algorithms. Journal of Applied
Kundra, H., & Sadawarti, H. (2015). Hybrid algorithm of Research and Technology, 14(3), 206–214. https://​doi.​org/​
cuckoo search and particle swarm optimization for natural 10.​1016/j.​jart.​2016.​05.​001
terrain feature extraction. Research Journal of Informa- Rabehi, A., Guermoui, M., & Lalmi, D. (2018). Hybrid mod-
tion Technology, 7(1), 58–69. https://​doi.​org/​10.​3923/​rjit.​ els for global solar radiation prediction: A case study.
2015.​58.​69 International Journal of Ambient Energy, 41(1), 31–40.
Li, F., Wang, S., & Wei, J. (2018). Long term rolling predic- https://​doi.​org/​10.​1080/​01430​750.​2018.​14434​98
tion model for solar radiation combining empirical mode Rodríguez, F., Azcárate, I., Vadillo, J., & Galarza, A. (2022).
decomposition (EMD) and artificial neural network Forecasting intra-hour solar photovoltaic energy by
(ANN) techniques. Journal of Renewable and Sustainable assembling wavelet based time-frequency analysis with
Energy, 10(1). https://​doi.​org/​10.​1063/1.​49992​40 deep learning neural networks. International Journal
Majumder, I., Dash, P., & Bisoi, R. (2018). Variational mode of Electrical Power & Energy Systems, 137, 107777.
decomposition based low rank robust kernel extreme https://​doi.​org/​10.​1016/j.​ijepes.​2021.​107777
learning machine for solar irradiation forecasting. Energy Ronno, C. K. (2017). Estimation of global radiation using
Conversion and Management, 171, 787–806. https://​doi.​ angstrom-type models at selected sites in Kenya. Afri-
org/​10.​1016/j.​encon​man.​2018.​06.​021 can Journal of Education, Science and Technology,
Maldonado, S., González, A., & Crone, S. (2019). Automatic 4(2), 63–70.
time series analysis for electric load forecasting via sup- Shariff, N. S., & Duzan, H. M. (2018). An application of pro-
port vector regression. Applied Soft Computing, 83, posed ridge regression methods to real data problem.
105616. https://​doi.​org/​10.​1016/j.​asoc.​2019.​105616 International Journal of Engineering & Technology,
Meenal, R., & Selvakumar, A. I. (2018). Assessment of SVM, 7(4.30), 106–108. https://​doi.​org/​10.​14419/​ijet.​v7i4.​30.​
empirical and ANN based solar radiation prediction models 22061
with most influencing input parameters. Renewable Energy, Siami-Namini, S., Tavakoli, N., & Siami Namin, A. (2018).
121, 324–343. https://​doi.​org/​10.​1016/j.​renene.​2017.​12.​005 A comparison of ARIMA and LSTM in forecasting time
Mohammadi, K., Shamshirband, S., Tong, C. W., Alam, K. A., series. 2018 17th IEEE International Conference on
& Petković, D. (2015). Potential of adaptive neuro-fuzzy Machine Learning and Applications (ICMLA), 1394–
system for prediction of daily global solar radiation by 1401. https://​doi.​org/​10.​1109/​icmla.​2018.​00227
day of the year. Energy Conversion and Management, 93, Singh, V. P., Vaibhav, K., & Chaturvedi, D. K. (2012). Solar
406–413. https://​doi.​org/​10.​1016/j.​encon​man.​2015.​01.​021 power forecasting modeling using soft computing
Mohanad, S. A., Ravinesh, C. D., & Yan, L. (2018). Particle approach. Nirma University International Conference on
swarm optimized–support vector regression hybrid model Engineering (NUiCONE), 2012, 1–5. https://​doi.​org/​10.​
for daily horizon electricity demand forecasting using cli- 1109/​nuico​ne.​2012.​64932​68
mate dataset. E3S Web of Conferences, 64, 08001. https://​ Sivakumar, M., George, S. T., Subathra, M. S., Leebanon, R.,
doi.​org/​10.​1051/​e3sco​nf/​20186​408001 & Kumar, N. M. (2023). Nine novel ensemble models for
Mohanty, S., Patra, P. K., & Sahoo, S. S. (2016). Prediction and solar radiation forecasting in Indian cities based on VMD
application of solar radiation with soft computing over and DWT integration with the machine and deep learn-
traditional and conventional approach – A comprehensive ing algorithms. Computers and Electrical Engineering,
review. Renewable and Sustainable Energy Reviews, 56, 108, 108691. https://​doi.​org/​10.​1016/j.​compe​leceng.​2023.​
778–796. https://​doi.​org/​10.​1016/j.​rser.​2015.​11.​078 108691
Mubiru, J., & Banda, E. (2008). Estimation of monthly aver- Srivastava, R., Tiwari, A., & Giri, V. (2019). Solar radia-
age daily global solar irradiation using artificial neural tion forecasting using Mars, cart, m5, and random forest

Vol.: (0123456789)
 310 Page 36 of 36 Environ Monit Assess (2025) 197:310

model: A case study for India. Heliyon, 5(10), e02692. Advances in Adaptive Data Analysis, 01(01), 1–41. https://​
https://​doi.​org/​10.​1016/j.​heliy​on.​2019.​e02692 doi.​org/​10.​1142/​s1793​53690​90000​47
Thombare, S., Pande, V. N., Kulkarni, R. S., & Yin, J., Rao, W., Yuan, M., Zeng, J., Zhao, K., Zhang, C.,
Kakade, S. M. (2022). Prediction of solar power using Li, J., & Zhao, Q. (2019). Experimental study of multi-
linear regression. Advances in Sustainability Science variate time series forecasting models. Proceedings of the
and Technology, 709–718. https://​doi.​org/​10.​1007/​978-​ 28th ACM International Conference on Information and
981-​16-​9033-4_​53 Knowledge Management, 2833–2839. https://​doi.​org/​10.​
Vennila, C., Titus, A., Sudha, T. S., Sreenivasulu, U., Reddy, 1145/​33573​84.​33578​26
N. P., Jamal, K., Lakshmaiah, D., Jagadeesh, P., & Belay, Zhang, Z., & Hong, W. (2019). Electric load forecasting by
A. (2022). Forecasting solar energy production using complete ensemble empirical mode decomposition adap-
machine learning. International Journal of Photoenergy, tive noise and support vector regression with quantum-
2022, 1–7. https://​doi.​org/​10.​1155/​2022/​77974​88 based dragonfly algorithm. Nonlinear Dynamics, 98(2),
Wang, W., Yang, D., Hong, T., & Kleissl, J. (2022). An archived 1107–1136. https://​doi.​org/​10.​1007/​s11071-​019-​05252-7
dataset from the ECMWF ensemble prediction system for
probabilistic solar power forecasting. Solar Energy, 248,
64–75. https://​doi.​org/​10.​1016/j.​solen​er.​2022.​10.​062 Publisher’s Note Springer Nature remains neutral with regard
Wu, Z., & Huang, N. E. (2009). Ensemble empirical mode to jurisdictional claims in published maps and institutional
decomposition: A noise-assisted data analysis method. affiliations.

Vol:. (1234567890)