RESEARCH ARTICLE

www.advtheorysimul.com

Enhancing Solar Forecasting Accuracy with Sequential Deep

Artificial Neural Network and Hybrid Random Forest and
Gradient Boosting Models across Varied Terrains
Muhammad Farhan Hanif,* Muhammad Umar Siddique, Jicang Si,
Muhammad Sabir Naveed, Xiangtao Liu, and Jianchun Mi

necessitates a significant shift toward re-

Effective solar energy utilization demands improvements in forecasting due to newable sources due to the detrimental
the unpredictable nature of solar irradiance (SI). This study introduces and environmental impact of fossil fuels.[1,2]
rigorously tests two innovative forecasting models across different locations: Solar energy, particularly effective in re-
the Sequential Deep Artificial Neural Network (SDANN) and the Deep Hybrid gions with high solar exposure such as
Africa, the Southwestern US, China, Mo-
Random Forest Gradient Boosting (RFGB). SDANN, leveraging deep learning,
rocco, and Australia, is emerging as a
aims to identify complex patterns in weather data, while RFGB, combining crucial renewable contender.[3] The impor-
Random Forest and Gradient Boosting, proves more effective by offering a tance of advanced solar forecasting is rec-
superior balance of efficiency and accuracy. The research highlights the ognized by major energy stakeholders like
SDANN model’s deep learning capabilities along with the RFGB model’s the Australian Renewable Energy Agency
(ARENA), the National Grid, and the Cal-
unique blend and their comparative success over existing models such as
ifornia Energy Commission, offering sub-
eXtreme Gradient Boosting (XGBOOST), Categorical Boosting (CatBOOST), stantial benefits.[4] The solar Photovoltaic
Gated Recurrent Unit (GRU), and a K-Nearest Neighbors (KNN) and (PV) sector’s growth, with a 22% increase in
XGBOOST hybrid. With the lowest Mean Squared Error (147.22), Mean 2021, underscores its strong investment po-
Absolute Error (8.77), and a high R2 value (0.80) in a studied region, RFGB tential, dominated by China, the US, and In-
stands out. Additionally, detailed ablation studies on meteorological feature dia, with Asia expected to lead the market by
2050.[5] Embracing this trend, Pakistan tar-
impacts on model performance further enhance accuracy and adaptability. By
gets 30% renewable electricity by 2030, cap-
integrating cutting-edge AI in SI forecasting, this research not only advances italizing on its ample solar resources. Ad-
the field but also sets the stage for future renewable energy strategies and vancements in solar energy utilization rely
global policy-making. on accurate forecasting of Global Horizon-
tal Irradiance (GHI), a complex task due
to various meteorological factors.[6,7] In re-
sponse, diverse forecasting methods have
developed, including image-based models, integrations of ma-
1. Introduction
chine learning with Numerical Weather Prediction (NWP)
1.1. Background models,[8,9] and deep learning (DL) techniques, especially effec-
tive for bridging short to medium-term forecasting gaps.[10]
The anticipated rise in global energy demand, from 13 ter-
awatt (TW) currently to 46 TW by the end of the 21st century,
1.2. Related Work

M. F. Hanif, J. Si, M. S. Naveed, X. Liu, J. Mi In the realm of forecasting, the approach involves an extensive
Department of Energy & Resource Engineering
College of Engineering examination of various input variables, both internal and exter-
Peking University nal, aiming to generate precise forecasts for a defined future
Beijing 100871, China period. The accuracy of these forecasts is closely linked to the
E-mail: [email protected] precision of the input parameters and the forecast’s duration,
M. F. Hanif, M. U. Siddique categorized into short-term (𝑡 ≤ 1 day), medium-term (𝑡 ≤ 14
Department of Mechanical Engineering
FE&T Bahauddin Zakariya University Multan
days), and long-term (𝑡 > 14 days) intervals.[11,12] A variety of
Multan 60000, Pakistan methods are applied in this domain, encompassing physical, sta-
tistical, and machine learning models,[13] each finding its utility
The ORCID identification number(s) for the author(s) of this article in scenarios ranging from immediate operational planning[14] to
can be found under https://doi.org/10.1002/adts.202301289 extensive long-term strategic planning.[15] In the realm of long-
DOI: 10.1002/adts.202301289 term predictions, the introduction of DeepSI, a Bayesian deep

Adv. Theory Simul. 2024, 2301289 2301289 (1 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

learning framework utilizing Global Circulation Models (GCMs) Convolutional Networks (CNN)[45] are especially used for spa-
for SI forecasting, underscores the evolving sophistication in tial analysis of power production data, enabling precise localized
tackling climate change-related challenges.[16] predictions.[46]
Physical models in this area are distinguished by their de- Hybrid models that blend spatial and temporal analysis ap-
pendence on fundamental attributes of the photovoltaic system, proaches have shown promise in solar energy forecasting. For
foregoing the need for historical data. Inputs such as numer- instance, the ConvLSTM model combines convolutional layers
ical weather predictions (NWP), site-specific monitoring data, within LSTM units, offering both translation invariance and
and specific characteristics of the power plant are utilized,[17,18] an expansive analysis capability, which is crucial for accurate
leading to outputs like global horizontal irradiance (GHI) and SI predictions.[47] WaveNet and Temporal Convolutional Net-
direct normal irradiance (DNI), directly influencing the plant’s works (TCN) utilize causal dilated convolutions to enhance tem-
production capacity.[19] The innovative downscaling methodol- poral data interpretation.[38,48] The Transformer model employs
ogy Global precipitation mission (GPM)- based multitemporal an attention mechanism to focus predictions on specific time
weighted precipitation analysis (GMWPA), which utilizes the steps, representing a significant advancement in forecasting
GPM_IMERGDF Product and Advanced Spaceborne Thermal methodologies.[49] Notably, the hybrid CNN-LSTM model pro-
Emission and Reflection Radiometer (ASTER) DEM within an posed by Ghimire et al.[50] and Zang et al.[51] integrates pat-
empirical distribution-based frameworl (EDBF) Algorithm, un- tern recognition with time-series analysis, reducing data depen-
derscores the vital role of high-resolution spatial data in pre- dencies and improving forecasting accuracy. The integration of
cipitation analysis, particularly in areas like Mainland China.[20] attention mechanisms further bolsters model performance by
The precision of meteorological data critically determines the addressing both short-term and long-term temporal patterns
accuracy of these models, with their predictive efficacy de- effectively.[52] Additionally, research optimizing the eXtreme Gra-
clining amidst imprecise weather information.[21] This cate- dient Boosting (XGB) model for solar radiation forecasting has
gory encompasses methodologies like the ASHRAE and Hottel demonstrated remarkable accuracy enhancements, highlighting
equations.[22,23] the efficacy of hyperparametric tuning in this field.[53] However,
Conversely, statistical methods leverage historical data to dis- it is important to acknowledge the inherent challenges associated
cern patterns and relationships between input variables and with each model type. RNNs, for example, may face training diffi-
power production. In the solar energy realm, these method- culties due to vanishing gradients[54] and numerical instability,[55]
ologies are widespread, incorporating diverse techniques like while CNNs often require extensive training data and complex ar-
Markov Chains,[24] fuzzy logic,[25] and auto-regressive[26] mod- chitectures to achieve comprehensive receptive fields.[56]
els such as Nonlinear Autoregressive model with eXogenous in-
puts (NARX)[27] and Nonlinear Autoregressive Moving average
with eXogenous inputs (NARMAX).[28] The study on daylight uti- 1.3. Contribution
lization for energy saving in Karachi’s buildings, which identi-
fied optimal dome areas for energy efficiency, serves as a prac- Integrating sophisticated AI and machine learning for SI fore-
tical example of how forecasting can aid in enhancing build- casting faces challenges, including balancing computational ef-
ing energy conservation.[29] Despite generally being less com- ficiency with accuracy due to solar data’s complexity.[57] There is
plex than physical models, the reliance on statistical methods a research gap in effectively combining advanced AI with a deep
on historical data facilitates more detailed modeling of specific understanding of meteorological impacts on SI, often leading to
plant characteristics.[30,31] Nonetheless, the necessity for compre- issues like overfitting and data inconsistency.[58–62] Limited re-
hensive plant-specific data may introduce constraints in quickly gional testing also restricts models’ global applicability, crucial
changing or growing energy landscapes. for renewable energy adoption.[63]
Machine learning models in this domain deploy algorithms Our study explores solar energy potential using innovative
trained on extensive datasets of input-output pairs to tailor pre- deep learning techniques, focusing on SDANN and Hybrid
dictions to specific scenarios. These models encompass De- RFGB model. SDANN, with its complex multi-layered setup and
cision Trees,[32] Random Forests (RF),[33] and their various ReLU activation, excels in processing sequential data. On the
ensembles,[34] all adept at navigating the complex conditions of other hand, RFGB uniquely combines RF“s[64] robust decision-
different power plant sites. Support Vector Machines (SVM)[35] making and GB”s[65] precision. By leveraging the strengths of
and ANN[36] excel at identifying critical input factors and offer these models, we aim to significantly enhance SI prediction accu-
interpretability. Similar advancements are echoed in studies like racy, drawing on the proven effectiveness of ensemble methods
the one employing Optimized Data-Driven Machine Learning for superior performance. This approach positions our research
Models, where Support Vector Regressor (SVR),[37] Multilayer at the forefront of advancements in SI forecasting.
Perceptron (MLP),[38] and RF Regressor[39] optimized with al- Our research not only focuses on SI forecasting but also con-
gorithms such as Moth Flame Optimization (MFO), Grey Wolf ducts detailed ablation studies to assess how different meteoro-
Optimization (GWO), and Evolve Class Topper Optimization logical features—like temperature, humidity, wind speed, and at-
(ECTO) significantly improved SI predictions.[40] Deep Neural mospheric pressure—affect model performance. Through sys-
Networks (DNN)[41] are applied more selectively for their com- tematic examination, we aim to refine the accuracy and adapt-
prehensive approach to time-series data. Recurrent Networks ability of our models, gaining deeper insights into the dynamics
(RNN),[42] including specialized forms like Long Short-Term of SI forecasting. This approach enriches the field of renewable
Memory (LSTM)[43] and Gated Recurrent Unit (GRU),[44] excel energy forecasting with a nuanced understanding of the relation-
in temporal data modeling, thus improving prediction stability. ship between environmental factors and SI predictions.

Adv. Theory Simul. 2024, 2301289 2301289 (2 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 1. Annual Temperature Trends in Selected Data Sites a) Hyderabad b) Sukkur c) Turbat Reproduced with permission.[66] weatherspark.com.

To encapsulate, the primary objective of this research is three- mid summers with temperatures ranging from 23 °C to 44 °C.
fold, i.e., Specifically, June is the hottest month, recording average high
temperatures of 39 °C in Hyderabad, 43 °C in Sukkur and Tur-
1) To design and articulate both SDANN and RFGB modeling bat. In contrast, the winter season is relatively mild, with temper-
frameworks; atures fluctuating between 8 °C and 29 °C, with January being
2) To empirically validate these models using robust statistical the coolest month, showing average low temperatures of 24 °C
metrics, comparing them with existing models to ascertain in Hyderabad, 23 °C in Sukkur, and 25 °C in Turbat.[66]
their predictive prowess in the field of SI forecasting; The clear skies prevalent throughout the year offer ample op-
3) To evaluate the predictive accuracy of SI forecasting models portunities for outdoor activities and contribute to the region’s
across diverse geographies, understand how regional varia- solar power potential. The variation in daylight hours is notewor-
tions influence model performance and reliability; and thy; the shortest day of 2023 is expected to be December 22, with
4) To investigate how various meteorological features impact the daylight hours of 10′20″ in Turbat, 10′23″ in Sukkur, and 10′33″
performance of forecasting models. This involved the incre- in Hyderabad. Conversely, the longest day, June 21, 2023, will pro-
mental addition and analysis of various features, enhancing vide 13′57″ of daylight in Turbat, 13′54″ in Sukkur, and 13′43″ in
the understanding of their individual and collective effects on Hyderabad.
model accuracy. Complementing this graphical analysis, Table 1 furnishes
detailed SI data for each city. It presents the Direct Normal
The research methodology includes data gathering, model Irradiations (DNI), which is the measure of solar radiation
formulation, and assessment criteria, which are extensively de- received per unit area by a surface directly facing the sun, the
scribed in Section 2. Section 3 provides obtained results and their Diffuse Horizontal Irradiations (DIF), which is solar radiation
interpretation; Section 4 provides ablation studies and Section 5 received from the sky excluding the solar disk, and the GHI, the
presents detailed analysis and discussion of the obtained results. total amount of solar radiation received per unit area from above
Lastly, in Section 6, a summary of the research findings is pre-
sented, emphasizing the contributions made to the field of solar
energy forecasting.
Table 1. Key climatic metrics of selected data sites.

2. Methodology Attributes Hyderabad Sukkur Turbat

2.1. Data Sites Latitude 25.3924 N 27.39242 N 26.008 N

Longitude 68.3737 E 68.836899 E 63.0383 E
The cities of Hyderabad, Sukkur, and Turbat, situated in the Elevation (m) 13 67 129
provinces of Sindh and Baluchistan, Pakistan, have been metic- Average Temperature (°C) 28 27 27
ulously chosen as the testing grounds for our DL models. These Maximum Temperature (°C) 48.5 50.5 53.7
locales are distinguished by their advantageous meteorological Minimum Temperature (°C) 3 −1.5 −0.4
conditions, notably the substantial SI and the extensive duration Sunshine Hours in June 14 14 14
of diurnal sunlight across all seasons, including the cooler win-
Sunshine Hours in Dec 11 11 10
ter period. The presence of such climatic features renders these
Average Sunshine Hours 9.1 12 9
sites particularly suitable for the empirical evaluation and calibra-
Direct Normal Irradiations (DNI) 4.644 4.342 5.03
tion of algorithms intended for solar energy potential forecasting
(kW m−2 day−1 )
within our DL frameworks. As depicted in Figure 1, the tempera-
Diffuse Horizontal Irradiations 2.54 2.577 2.44
ture profiles of these cities throughout the year are instrumental
(DIF) (kW m−2 day−1 )
in understanding the climatic dynamics relevant to solar energy
Global Horizontal Irradiations 5.692 5.515 5.781
potential Figure 1 depicts that these urban centers are noted for
(GHI) (kW m−2 day−1 )
their unique meteorological patterns, which include hot and hu-

Adv. Theory Simul. 2024, 2301289 2301289 (3 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Table 2. Characterization of satellite-derived input features utilized in proposed models.

Attribute Name Unit Acronym Source Model Spatial Resolution

Solar Irradiance kWh m−2 day−1 SI CERES SYN 1-deg 1°×1°

Earth’s Skin Temperature oC Ts MERRA-2 M2T1NXRAD v5.12.4 0.5°×0.625°
Minimum Temperature oC Tmin
Maximum Temperature oC Tmax
Average Temperature oC T
Surface Pressure kPa PS
Relative Humidity % RH
Wind Speed m−1 s WS

by a horizontal surface. Notably, the GHI values for all three with the Ts, suggesting that these two parameters have an inverse
cities surpass 5 kWh m−2 per day, indicating a high potential for relationship. Other notable relationships include the moderate
solar energy harnessing.[67] negative correlation between T and RH, and the weak positive
Given this combination of high solar radiation, prolonged correlation between PS and T.
sunshine hours, and favorable temperature ranges, Hyderabad, It is important to note that the pair plot only reveals corre-
Sukkur, and Turbat emerge as ideal test sites for the development lations between the different parameters, not causation. It is
of solar power infrastructure. This potential aligns with the in- possible that there is a third factor that is driving the observed re-
creasing global demand for renewable energy sources and posi- lationships. For example, the positive correlation between T and
tions Pakistan to make a substantial contribution to the growth Tmax may be due to the fact that both parameters are influenced
of solar energy within the South Asian region. by solar radiation. Overall, the pair plot provides a valuable tool
for exploring and visualizing the relationships between multiple
parameters. By understanding these relationships, we can better
2.2. Data Collection understand the available datasets to incorporate them into mod-
els. Pair plots are essential for deep learning models because
For the purposes of this research, the dataset was meticulously they reveal correlations between different variables. This infor-
sourced from the National Aeronautics and Space Adminis- mation can be used to select the best features, fine-tune model
tration (NASA) Langley Research Center’s (LaRC) Prediction parameters, and identify outliers. SI being the target parameter
of Worldwide Energy Resource (POWER) Project.[68] This ini- in this study is the main focus of this plot. The hue determines
tiative, which receives funding through the NASA Earth Sci- the value of SI in a particular correlation, i.e., if the correlation is
ence/Applied Science Program, is renowned for its extensive and negative as shown at the bottom of plot, SI value decreases with
reliable long-term solar energy data. The scope of the data ac- an increase in the values of the input parameters. It is obvious
quired for this study encompasses a period from 1994 to 2021, from the top left of the plot that in case of a positive correlation,
comprising a total of 10002 daily entries each, which provides a the SI values increase with the corresponding increase of input
comprehensive foundation for our analysis. A variety of climatic parameters.
and environmental factors can be utilized to develop models for
predicting solar radiation. Based on our literature review and data
analysis,[69–73] we have selected eight input parameters for our SI 2.4. Data Pre-Processing
prediction models, detailed description of these parameters can
be seen in Table 2. In the process of tailoring robust and effective predictive mod-
els, we have employed several data pre-processing techniques.
The significance and choice of these techniques stem from their
2.3. Correlation Analysis well-established efficacy in enhancing model performance. Be-
low, we elaborate on these pre-processing steps and rationalize
The correlation for Sukkur dataset in the form of a pair plot, is their adoption:
depicted in Figure 2, The data statistics and correlation plots for
the other datasets are provided as supplementary data. The pair
plot provides a comprehensive visualization of the relationships 2.4.1. Data Cleaning and Handling Missing Values
between eight environmental parameters: T, Tmax , Tmin , PS, RH,
WS, Ts, and SI. Each panel in the pair plot depicts a scatterplot of A pivotal step in our pre-processing is data cleaning, which pri-
two parameters, with the color of the points indicating the value marily deals with the challenge of missing values. Incomplete
of a third parameter.[74,75] Upon inspection of the pair plot, several datasets, if not addressed, can substantially undermine the pre-
patterns and relationships emerge. For example, the temperature dictive accuracy of models, leading to outcomes that may be bi-
parameters (T, Tmax , and Tmin ) all exhibit strong positive correla- ased or even nonsensical. Various techniques can be utilized for
tions with each other, indicating that they tend to vary in unison. the purification of data, depending on the type of data, its format,
Conversely, the PS parameter exhibits a weak negative correlation and the desired responses.[76–78] To address this, we adopted a

Adv. Theory Simul. 2024, 2301289 2301289 (4 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 2. Pair Plot of data distribution for Sukkur.

strategy of eliminating rows with missing values. This approach, Building on the insights from the correlation analysis, where
often referred to as listwise deletion, ensures data integrity and we explored the relationships between environmental parame-
maintains consistency across the dataset. It is especially benefi- ters through pair plots, we adopted a data-driven approach to fea-
cial when the proportion of missing data is relatively minimal, ture selection. Our objective was to distill the dataset to variables
thereby minimizing the potential loss of valuable information. that exhibit significant correlations with SI, the dependent vari-
able of our study. This process was informed by both statistical
2.4.2. Feature Selection analyses and domain expertise, ensuring that retained features
are relevant for SI prediction.
Feature selection emerges as a pivotal pre-processing step in re- Initial Correlation Analysis: Leveraging the preliminary in-
fining our forecasting model. This stage is critical for mitigating sights from the pair plots, we conducted a correlation analysis
the risks of overfitting—particularly prevalent in complex mod- to quantitatively assess the relationship between each potential
els like SDANN—and for managing computational complexity in feature and SI. This helped us identify variables with statistically
ensemble methods such as RF and GB. Extraneous features can significant correlations as primary candidates for inclusion.
compromise model efficiency and accuracy, making judicious Exclusion of Non-informative Features: In alignment with our
feature selection essential. focus on relevance, features that lacked substantial insights into

Adv. Theory Simul. 2024, 2301289 2301289 (5 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

SI prediction, such as the “Date” column, were excluded. This de- This formula was applied to each feature in our dataset to en-
cision was supported by the lack of a direct or meaningful correla- sure consistency and uniformity across the data used for training
tion with SI, underscoring our commitment to model simplicity and testing our models.
and performance. Implementation: Our code implementation of feature scaling
Iterative Refinement: Our selection process is iterative, em- utilizes the StandardScaler from Scikit-learn, which automates
ploying techniques such as dropping features with limited pre- the scaling process according to the formula above. In this pro-
dictive value, exemplified by the exclusion of the “Date” column. cess, the function fit_transform[81] calculates the mean (𝜇) and
This decision-making is facilitated by our code snippet, which standard deviation (𝜎) of each feature in the training set and
systematically refines the dataset to focus on variables directly scales the features accordingly. The same scaling parameters (𝜇)
influencing SI. and (𝜎) are then applied to the test set using transform, ensuring
Empirical Evaluation: The ablation studies section later elab- that the model sees the test data in the same scale as the training
orates on this process, detailing how features were evaluated and data.
selected based on their predictive power for SI. This includes These pre-processing steps, rooted in best practices, are
temperature metrics (Ts , Tmax , Tmin ), PS, RH, and WS, each sub- foundational in our machine learning endeavors, enabling our
stantiated by their observed influence on SI variability. models—whether the SDANN or the hybrid approach—to effec-
Integration of Findings: By the end of our analysis, the chosen tively learn from the data and produce reliable, robust predic-
features are those demonstrated through both visual explorations tions.
and statistical validation to have significant predictive power for
SI, ensuring our model is both efficient and effective.
This approach to feature selection ensures our models are built 2.5. Model Architecture
on a foundation of relevant, impactful variables. The insights
gained from initial data exploration, combined with the detailed Throughout the model formulation of SDANN and the hy-
evaluations in our ablation studies, guide our final feature set, bridized RFGB and validation stages, we leverage a suite of
aligning with our goal of enhancing SI forecasting accuracy. Python’s preeminent libraries. These encompass Matplotlib,
Scikit-learn, KERAS, Seaborn, Pipeline, and Pandas.[82,83] The
system setup is anchored by the robust NVIDIA Tesla T4 GPU,
2.4.3. Data Splitting equipped with NVIDIA’s Turing architecture, and is configured
without display activation, indicating its dedicated use in high-
To ensure a robust evaluation of the model’s performance, it is performance computer operations. The system is running with a
imperative to divide the dataset into distinct training and testing driver version of 535.104.05 and a CUDA version of 12.2, ensur-
subsets. While the training set plays a crucial role in building the ing compatibility with contemporary machine-learning frame-
model, the test set offers an objective and unbiased assessment of works and libraries. The Tesla T4’s memory and processing ca-
the model’s generalization capability. For this purpose, we parti- pabilities are leveraged to support intensive computational tasks,
tioned the data into an 80-20 distribution, allocating the majority underlined by its integration into the Google Compute Engine
for training and the remainder for validation. backend,[84] highlighting the system’s readiness for scalable and
efficient data processing tasks.
2.4.4. Feature Scaling
2.5.1. Sequential Deep Artificial Neural Network (SDANN)
The significance of feature scaling in predictive modeling, par-
ticularly for models like SDANN, is paramount. Features mea-
The development of SDANN is a comprehensive process that in-
sured on different scales can significantly impede the net-
volves several crucial steps. The six steps involved in the SDANN
work’s convergence during the training phase, as gradient de-
development process are data splitting, model architecture se-
scent algorithms—commonly used in training these networks—
lection, input selection, model structure determination, model
achieve optimal performance when features are standardized.
calibration/training (parameter optimization), and model valida-
While ensemble models such as RF and GB do not inherently re-
tion (Figure 3b).[85,86] The selection of inputs is a vital aspect of
quire feature scaling,[79] standardizing features can contribute to
SDANN development, as the inclusion of too many correlated
model uniformity, simplifying the interpretation of hybrid mod-
or unnecessary inputs can lead to an increase in data training
els and potentially enhancing training efficiency.
time and overfitting. On the other hand, omitting crucial inputs
To standardize our features, we employed the following stan-
can result in a model that fails to fully explain the input-output
dardization technique,[80] which normalizes the features to have
relationship.[69–71,73]
a mean of zero (𝜇 = 0) and a standard deviation of one (𝜎 = 1).
The SDANN model structure determined the complete flow
The mathematical formula for standardizing a feature X is given
and structure of the model as shown in Figure 3a. The inclu-
by:
sion of eight pivotal parameters, such as T, Tmax , Tmin , Ts, RH,
X −𝜇 SH, WS, and PS reinforces the model’s capability to integrate
Xscaled = (1) a broad spectrum of variables essential for accurate forecasting.
𝜎
The number of input nodes in the model is equal to the number
here: X is the original feature vector, 𝜇 is the mean of the feature of parameters used. One of the distinguishing features of our
vector, 𝜎 is the standard deviation of the feature vector. SDANN model, setting it apart from conventional models, is its

Adv. Theory Simul. 2024, 2301289 2301289 (6 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 3. a) Structure of SDANN Model b) Overall SDANN Model Process.

multi-layered architecture comprising six hidden layers which is where (hl ) is the output from layer (l), (f ) denotes the ReLU acti-
a deliberate enhancement from the traditional single-layer net- vation function, (Wl ) is the weight matrix, (bl ) is the bias vector,
works. This multi-layered approach is not arbitrary; it is grounded and (hl−1 ) is the input from the previous layer (with (h0 ) being the
in the need to capture the inherent complexity of the variables initial input features).
that influence SI, such as temperature, solar heating, wind speed, Output Layer: The final prediction of SI is obtained from the
and pressure[87,88] ensuring the model’s focus on variables essen- output layer, formulated as:
tial for accurate forecasting. ( )
In developing the SDANN model’s architecture, not only did ŷ = f Woutput . hL + boutput (3)
we focus on incorporating a broad spectrum of variables, but
we also carefully selected the Adam optimizer for its robust- Model Training: Training involves optimizing the weight ma-
ness and efficiency in handling deep neural networks. The choice trices and bias vectors across all layers to minimize the loss func-
of Adam as the optimizer is supported by its demonstrated su- tion, specifically the Mean Squared Error (MSE) between the pre-
periority in various deep learning applications, as highlighted dicted and actual values of SI:
in related studies, including one that assessed its performance 1 ∑n ( )2
against other renowned optimizers like SGDNesterov, AdaGrad, MSE = ŷ i − yi (4)
n i=1
and RMSProp, among others.[89] This optimizer, selected for its
proven effectiveness and compatibility with complex architec- The division of the dataset into training and testing sets, with
tures like ours, further refines our model’s capability to integrate 80% allocated for training and the remainder for validation, cou-
and analyze the essential forecasting variables within its multi- pled with the employment of regularization techniques, ensures
layered structure. the model’s robustness and generalizability.
Model Architecture and Mathematical Representation: The This detailed approach, from the structured architecture to
model initiates with an input layer corresponding to the num- the strategic layering and meticulous training, underscores the
ber of selected features, followed by multiple hidden layers, and SDANN’s design sophistication. It is this architecture, grounded
concludes with an output layer, structured as follows: in both theoretical underpinnings and empirical validation, that
Input Layer: The input layer receives the features, setting the equips the SDANN to adeptly model SI forecasting, setting it
stage for the model’s deep learning process. apart from conventional models.
Hidden Layers: The architecture progresses through six hid- Our journey involved navigating multiple libraries, with the
den layers, with the transformation at each layer (l ) represented Standard Scalar library being pivotal in normalizing our diverse
by: dataset.[82,83,90,91] The model’s structure, facilitated by the Sequen-
tial feature, reflects a perfect blend of complexity and clarity.
( ) Notably, the uniform distribution employed for neuron weight
hl = f Wl . hl−1 + bl (2) initialization ensures balanced weight assignment, crucial for

Adv. Theory Simul. 2024, 2301289 2301289 (7 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Table 3. Optimal Architecture of the Purposed Models. tion (̂yRF ) for a given input (X) can be expressed as the average
output of (N) decision trees:
Hyperparameters Range of Values/Description

SDANN Hidden Layers 6 1 ∑N

ŷ RF = T . (X ) (5)
Hidden Neuron [500, 400, 300, 200, 99, 25] N i=1 i
Learning rate 0.001
Epochs 100
where (Ti ) represents the (ith ) decision tree in the ensemble.
Batch Size 10 Gradient Boosting: GB, on the other hand, is a three-
Activation function [ReLU] component algorithm that consists of an optimized loss function,
Solver [Adam] predictions made by a weak learner, and an additive model. The
Loss Function Mean Squared Error loss function is determined based on the problem being solved,
RF Random State 42 and common loss functions such as the square error for regres-
N Estimators [70, 140, 210, 280] sion or logarithmic loss for classification are used. The weak
Max Depth [1–10] learner used in GB is typically a decision tree, which is trained to
Min Samples Leaf [0.1, 0.2, 0.3, 0.4, 0.5]
make predictions and minimize the loss function. The decision
tree is built in an additive manner, meaning that as new trees
GB Random State 42
are added to the model, existing trees are not altered. A gradi-
N Estimators [70, 90, 110, 130]
ent descent approach is used to reduce the loss by parameteriz-
Max Depth [1 to 10]
ing the trees and moving them in a direction that minimizes the
Min Samples Leaf [0.1, 0.2, 0.3, 0.4, 0.5] error.[94–96]
Hybrid RFGB Initial Estimator RF with max. depth = 2, min. Its prediction (̂yGB ) is derived from iteratively adding weighted
samples leaf = 0.1, N predictions (hi ) of (M) weak learners:
estimators = 280
N Estimators 110
∑M
Max Depth 4 ŷ GB = ∝i . hi . (X ) (6)
i=1
Min Samples Leaf 0.1

with (hi ) being the (ith ) weak learner’s prediction, and (∝i ) its cor-
model stability. With the optimizer algorithm at its helm, the responding weight.
model’s weight values were meticulously fine-tuned to optimize Meta Model (RFGB): Our innovative RFGB model integrates
performance. The complete details of the SDANN model struc- the RF and GB models’ strengths by initially applying RF for a ro-
ture can be found in Table 3. A uniform distribution was used to bust baseline prediction, which is then refined through GB’s se-
represent the weights in the neurons for every hidden layer, and quential improvement technique. The hybrid model’s final pre-
an optimizer algorithm was employed to adjust the values of the diction (̂yRFGB ) combines the RF base model’s output with GB’s
weights. enhancements:

∑M
ŷ RFGB = ∝j . hj . (X ) (7)
j=1
2.5.2. Hybrid Random Forest Gradient Boosting (RFGB)

SI forecasting, a domain with its unique set of challenges, man- Parameter Optimization: Hyperparameter tuning is critical
dates the employment of advanced techniques to optimize pre- for optimizing the RFGB model’s performance. For RF, param-
dictive accuracy. In this milieu, we present the hybrid RFGB eters such as the number of trees (N), maximum depth of trees
model – an innovative approach aiming to synergize the robust- (d), and minimum samples per leaf node (msl ) are adjusted. Sim-
ness of RF with the refined precision of GB. ilarly, for GB, we fine-tune the number of stages (M), depth of
Random Forest (RF): The working principle of the RF algo- trees, and minimum samples per leaf, among others, to find the
rithm is based on the concept of bagging (bootstrap aggregating). optimal model configuration. This process is facilitated by grid
In this technique, a large number of decision trees are trained on search optimization, systematically exploring a range of parame-
random subsets of the training data, and the outputs of all the ter combinations to identify those yielding the highest predictive
trees are combined to produce a final output.[92,93] The algorithm accuracy.
works by dividing a node into a subset of features rather than By leveraging the RF and GB models’ complementary
using the most significant feature, as is the case in traditional de- strengths, our RFGB model offers a sophisticated approach to
cision trees (Figure 4). Such diversity ensures that the model isn’t SI forecasting. The outlined mathematical framework underpins
overly skewed by any particular data trend, making it especially the model’s structure and operational logic, ensuring a deep un-
apt for handling the complex and varied data associated with SI derstanding of its functionality without delving into specific cod-
forecasting. ing implementations. The hybrid RFGB model detailed hyperpa-
The RF model operates on the principle of ensemble learn- rameters can be visualized in Table 3, Also the variation of best
ing, aggregating the outputs of multiple decision trees trained hyperparameters for different datasets is also presented in Ap-
on random data subsets. Mathematically, the RF model’s predic- pendix 1 in Supporting Information.

Adv. Theory Simul. 2024, 2301289 2301289 (8 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 4. Visualization of Hybrid RFGB Model.

2.6. Statistical Indicators for Evaluation SDANN, a model engineered to identify complex patterns
within extensive datasets, through rigorous accuracy testing.
The evaluation of the accuracy of a predictive model is a cru- Subsequently, the effectiveness of the Hybrid RFGB model is
cial aspect in the field of statistics and machine learning. In re- scrutinized.
gression analysis, the comparison between the original and pre- Furthermore, the comparative analysis is enriched by juxtapos-
dicted values is used to determine the effectiveness of the model ing the SDANN and RFGB models against four contemporary
and to identify sources of error. Several metrics are employed to models within the domain. This comparative study is designed
measure the error rate of prediction and the performance of the to situate the contributions of the SDANN and RFGB models
model in regression analysis. These include Mean Absolute Error within the current landscape of SI forecasting methodologies.
(MAE), MSE, Root Mean Squared Error (RMSE), and Coefficient Moreover, the adaptability and efficacy of the SDANN and
of Determination (R2 ).[97] RFGB models are validated across diverse geographical terrains,
encompassing Ahvaz (Iran),[98] Bikaner (India),[99] Alice Springs
3. Results (Australia),[100] Las Vegas (USA),[101] Marrakech (Morocco),[102]
and Luxor (Egypt).[103] This phase aims to ascertain the models’
In this section, the outcomes of deploying advanced deep performance consistency across varied climatic and geographic
learning models for SI forecasting are meticulously de- conditions, thereby broadening the understanding of their global
tailed. Initially, the research evaluates the performance of applicability in SI forecasting.

Adv. Theory Simul. 2024, 2301289 2301289 (9 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 5. KDE of SDANN Predictions a) Hyderabad b) Sukkur c) Turbat.

By leveraging a comprehensive suite of statistical indicators, Overall, the SDANN model demonstrates a high degree of ac-
this research delineates a quantitative evaluation of each model’s curacy in SI forecasting,[105] with potential for further optimiza-
accuracy and overall performance, underpinning the significant tion, particularly in refining its performance in capturing the full
advancements introduced through this study in the realm of solar spectrum of SI, especially in regions with more complex meteo-
energy forecasting. rological conditions.

3.1. SDANN Predictions 3.2. RFGB Predictions

The SDANN model, when applied to the SI forecasting in the The results depicted in Figure 6 convey the performance of the
cities of Hyderabad, Sukkur, and Turbat, exhibits commendable hybrid RFGB model in estimating SI for three cities: Hyderabad,
predictive performance, albeit with variations reflective of re- Sukkur, and Turbat. The kernel density estimations (KDE) for ac-
gional solar profiles. tual and predicted SI values are presented across three subfig-
Figure 5 presents the kernel density estimations (KDE) of ures.
actual versus predicted SI values for three distinct cities— Hyderabad (Figure 6a) demonstrates a model with high pre-
Hyderabad, Sukkur, and Turbat—using the SDANN model. The dictive accuracy, as evidenced by the substantial overlap of the
KDE is a non-parametric way to estimate the probability density actual and predicted density curves, particularly around the pri-
function of a continuous variable, providing a smooth curve rep- mary peak at 60 kWh m−2 day−1 and a secondary peak at 125 kWh
resentative of the data’s distribution.[104] m−2 day−1 . This suggests that the RFGB model has a robust ca-
In Hyderabad, the model’s density plot (Figure 5a) reveals pacity for capturing the SI distribution in Hyderabad.
a pronounced peak at ≈125 kWh m−2 day−1 , suggesting a In Figure 6b, representing Sukkur, the model also displays
high probability of occurrence around this SI value. The pre- commendable performance, with the predicted SI values closely
dicted values closely shadow the actual measurements, delin- tracking the actual data. The most prominent peak is observed
eating a narrow divergence and a high kernel density over- around 100 kWh m−2 day−1 . The minor underestimation by
lap. Moving to Sukkur, the density plot (Figure 5b) displays a the model at the apex of the peak is indicative of a slight di-
peak ≈120 kWh m−2 day−1 , indicative of the model’s precision vergence but does not detract significantly from the overall
in capturing the median irradiance values. Occasional devia- accuracy.
tions are noted, yet they remain within an acceptable range, un- Contrastingly, Figure 6c for Turbat shows a more pronounced
derscoring the model’s adeptness at handling the region’s SI discrepancy between the model’s predictions and actual SI val-
variability. ues, with the model predicting a broader and bimodal distribu-
The closeness of the two curves in each subfigure is an indica- tion of SI values. This discrepancy suggests that the model may
tor of the model’s predictive accuracy. In the cases of Hyderabad require further tuning to accommodate the unique environmen-
and Sukkur, the actual and predicted curves are almost superim- tal factors or SI patterns specific to Turbat.
posed, suggesting that the SDANN model has a high level of accu- The RFGB model, fine-tuned through an exhaustive search
racy in predicting SI for these cities. The peaks of the KDE curves of parameter space using GridSearchCV, indicates that the op-
indicate the most probable SI values, with both cities showing timal parameters derived from the cross-validation process have
prominent peaks at similar SI levels, which demonstrates the led to a high degree of predictive accuracy for Hyderabad and
model’s effectiveness in these instances. Sukkur.[64,65] The divergence observed in Turbat’s SI predictions
Turbat’s subfigure (c), however, shows a divergence between highlights the necessity for a more nuanced approach to model
the actual and predicted curves. The predicted curve is broader tuning, perhaps incorporating a more detailed analysis of local
with a shifted peak, indicating variance from the actual SI values. climatic variables or employing a region-specific model training
This suggests that for Turbat, the model’s predictions are less ac- regimen.
curate compared to the other two cities. The broader spread of In summary, the RFGB model demonstrates a commendable
the predicted curve could imply overfitting, a model’s sensitiv- level of accuracy in predicting SI, particularly in Hyderabad
ity to certain variables, or variability in the city’s SI not captured and Sukkur. The slight prediction discrepancies observed in
during the model’s training. Turbat provide a basis for future enhancements to the model,

Adv. Theory Simul. 2024, 2301289 2301289 (10 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 6. KDE of RFGB Predictions a) Hyderabad b) Sukkur c) Turbat.

which could involve more complex or localized approaches 3.3. Model Assessment
to data analysis and model training. These findings affirm
the potential of the RFGB model as a sophisticated tool in Figure 7 presents a detailed comparative analysis of the SDANN
the field of SI forecasting, with a proven capability to adapt and RFGB models, employing a segment of test data for in-depth
to and capture the unique characteristics of varied regional examination and the entire test dataset for frequency analysis.
climates. This approach ensures a thorough evaluation of the models’

Figure 7. Bar Graph Comparison of Actual and Predicted SI: a) Hyderabad, b) Sukkur, and c) Turbat. Frequency Plot of Actual and Predicted SI: d)
Hyderabad, e) Sukkur, and f) Turbat.

Adv. Theory Simul. 2024, 2301289 2301289 (11 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

predictive performance for SI across Hyderabad, Sukkur, and 3.4. Comparative Analysis
Turbat.
In Hyderabad (Figure 7a), the SDANN model’s predictions In this research, we have undertaken an exhaustive comparative
ranged from 15.1 to 145.0 kWh m−2 day−1 , with an average pre- analysis of various machine learning models for the prediction
diction of 101.7, closely aligning with the actual SI values that of SI. Our study benchmarks the performance of our proposed
spanned 14.9 to 136.0 kWh m−2 day−1 and an average of 101. models against a selection of four different algorithms previ-
The slight average difference of 0.002 suggests a high degree ously suggested by other researchers. This comparison encom-
of accuracy. The RFGB model, meanwhile, predicted SI values passes a broad spectrum of models, including XGBOOST,[106]
within a narrower range of 53.2 to 132.2 kWh m−2 day−1 , av- Cat Boost,[107] GRU,[32] and a hybrid model KNN-XGBOOST.[108]
eraging at 100.8, resulting in a marginal average difference of Each model has been rigorously evaluated for its efficiency and
−0.00043. The RFGB model’s predictions exhibited a superior precision in predicting SI. It is pertinent to highlight that dur-
accuracy for Hyderabad, with a maximum and minimum dif- ing this comparative analysis, adjustments were made to the hy-
ference of 97.93 and −41, respectively. The bar graphs reflect a perparameters of these models to better align them with the spe-
close match between predicted and actual values, while the fre- cific characteristics of our dataset, as detailed in Appendix 2 (sup-
quency plot (Figure 7d) depicts the consistency of both SDANN plementary data). These customizations were essential to ensure
and RFGB models over actual data. that the comparison is not only equitable but also pertinent to
Sukkur’s predictions (Figure 7b) indicated that both models the unique aspects of our dataset and the specific objectives of
performed similarly well, with SDANN’s predicted SI values our study. The aim of these modifications was to provide a trans-
ranging from 15.8 to 134.1 kWh m−2 day−1 , averaging 95.0, and parent and unbiased evaluation of how our newly proposed mod-
RFGB’s ranging from 46.7 to 129.7 kWh m−2 day−1 , averaging els, SDANN and RFGB, compare with established algorithms in
96.5. The actual SI values for Sukkur recorded a maximum of the field. This comparison using statistical indicators (Appendix
133.8, a minimum of 12.4, and an average of 96.4 kWh m−2 day−1 . 3 (supplementary data)) as illustrated in Figure 8 serves to vali-
The SDANN model had an average difference of −0.0021, while date the effectiveness and appropriateness of our models for SI
RFGB showed a slight variance of 0.0003, with maximum and prediction.
minimum differences of 69.4 and −48.7, respectively. The scatter Examining the Hyderabad dataset, we observe that while the
plots demonstrate RFGB’s minor predictive deviations, especially XGBOOST model incurred the highest error with an MSE of
at the tails of the distribution, while the frequency plot (Figure 7e) 246.04, the CAT Boost model demonstrated a commendable pre-
showcases both models’ close prediction frequencies around the cision with the lowest MSE of 165.33. The RFGB and SDANN
actual SI values. models showcased their predictive robustness with very compet-
In the case of Turbat (Figure 7c), the SDANN model exhib- itive MSE values of 176.21 and 177.98, respectively. This trend is
ited a prediction range for SI values from 60.2 to 149.5 kWh mirrored in the RMSE and MAE values, where both models are
m−2 day−1 , with an average prediction at 101.7, whereas the in close contest, but the R2 scores stand out, jointly reflecting a
RFGB model demonstrated predictions within a range from 68.8 high correlation with actual SI values at 0.74 for both RFGB and
to 124.8 kWh m−2 day−1 , averaging at 103.5. The recorded ac- SDANN, affirming their reliability.
tual SI values were 13.8, 103.9, and 137.7 kWh m2 day−1 , re- The Sukkur dataset presents a similar narrative. Here, the
spectively. Here, the SDANN model showed an average differ- RFGB model excels, not only improving upon its Hyderabad per-
ence of 0.005 from the actual values, closely followed by the formance with an MSE of 147.22 but also surpassing CAT Boost’s
RFGB model with a marginally different average difference of MSE of 158.45. The SDANN follows closely, ensuring tight com-
−0.0006. Despite the slight variance in their average differences, petition with an MSE of 152.91. Both models maintain their con-
both models displayed commendable performance in predict- sistency in accuracy as reflected in the RMSE and MAE scores.
ing SI for Turbat, with the RFGB model illustrating a narrower The RFGB model edges ahead slightly with an R2 of 0.80, indicat-
gap in the maximum and minimum differences of 83.4 and ing a marginally stronger predictive accuracy than the SDANN’s
−57.6, respectively. The frequency plot (Figure 7f) for Turbat in- also impressive 0.79.
dicates a significant overlap in predictions by both models with When shifting focus to the Turbat dataset, the metrics interest-
the actual SI values, highlighting their comparable accuracy in ingly shift. Here, although the XGBOOST and CAT Boost mod-
forecasting SI. els report lower MSE and RMSE values, their R2 scores of 0.36
Collectively, the bar and frequency plots across all regions and 0.71, respectively, do not necessarily translate to the most re-
showcase the comparable prediction capabilities of both the liable predictions. The RFGB and SDANN models, with closer
SDANN and RFGB models. While the SDANN model provides R2 scores of 0.43 and 0.44, underscore their comparable perfor-
valuable insights and exhibits commendable accuracy, the RFGB mance in this challenging environment. Despite a higher MSE
model also demonstrates high-quality predictions, particularly in and RMSE, which could be attributed to Turbat’s unique climatic
the alignment of its forecasts with the actual SI values. These factors, the consistency of RFGB and SDANN in the R2 scores
findings highlight the potential of both models as reliable and suggests a resilience in their predictive ability.
effective tools for SI forecasting, each contributing significantly As the scatter graphs of Figure 9 provide a visual testament
to the domain of solar energy management and planning. The to the performance of the forecasting models within the Sukkur
results emphasize the utility of integrating diverse modeling ap- dataset, the prowess of the RFGB and SDANN models becomes
proaches to enhance the accuracy and reliability of solar forecast- readily apparent. These graphs show that for the RFGB and
ing systems. SDANN models, the data points on the scatter plots closely

Adv. Theory Simul. 2024, 2301289 2301289 (12 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 8. Statistical validation of compared Models a) Hyderabad dataset, b) Sukkur dataset, and c) Turbat dataset.

Figure 9. Scatter plot of predicted over actual SI for Sukkur dataset a) XGBOOST, b) Cat Boost, c) GRU, d) KNN-XGBOOST, e) SDANN, and f) RFGB.

Adv. Theory Simul. 2024, 2301289 2301289 (13 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 10. Comparison Plot of Actual and Predicted SI across various geographies: a) Ahvaz, b) Bikaner c) Alice Springs, d) Las Vegas, e) Marrakech,
and f) Luxor.

adhere to the line of ideal prediction, underscoring their accu- not only in Sukkur but also in Hyderabad, where they both ex-
racy in capturing the true SI values in the Sukkur region. The hibit high accuracy levels.
proximity of the scatter points to this line in the RFGB plot is While the RFGB and SDANN models stand out as the more
particularly indicative of its precision. reliable models for solar forecasting, the varying degrees of per-
In contrast, the scatter plots for the XGBOOST, CAT Boost, formance across different models and regions, especially Turbat,
GRU, and KNN-XGBOOST models reveal a more scattered dis- highlights the ongoing need for model enhancements and the
tribution of points away from the ideal prediction line. This visual possible integration of localized factors to refine their forecast-
spread is more pronounced in the Turbat dataset plots, where ing capabilities.
prediction variance is most noticeable, reflecting these models’
struggle to match the accuracy seen with the RFGB and SDANN
models. 3.5. Evaluating the Predictive Accuracy of SI Forecasting Models
The scatter plots depicted in Figure 9 offer a clear visual com- Across Diverse Geographies
parison of the predictive success across various models within the
Sukkur dataset. The RFGB and SDANN models demonstrate su- Upon close examination of the SI forecasting capabilities of the
perior performance, as shown by the data points’ close alignment SDANN and RFGB models across varied geographical datasets,
to the line of perfect prediction, particularly for the RFGB model discernible patterns and differences in model performance
which shows slightly better precision as evidenced by lower er- emerge. The 3D bar plots, illustrated in Figure 10a–f, serve as
ror rates and higher R2 values. These scatter plots provide com- a visual representation of these comparisons and facilitate a
pelling evidence of the efficacy of the RFGB and SDANN models quantitative analysis through various performance metrics in
in the field of SI forecasting. For a more comprehensive under- Table 4.
standing, the scatter plots for the Hyderabad and Turbat datasets Starting with the Ahvaz, Iran dataset (Figure 10a), the SDANN
are made available in the supplementary data. These plots rein- model registers a robust R2 score of 0.83, signifying a strong
force the strong performance of the RFGB and SDANN models, correlation with the actual SI measurements. In contrast, the

Adv. Theory Simul. 2024, 2301289 2301289 (14 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Table 4. Statistical validation of proposed models across various geogra- fering a reliable alternative in the advanced domain of solar
phies. forecasting.

Data Set Model R2 MAE MSE RMSE

4. Ablation Studies
Ahvaz, Iran SDANN 0.83 0.594 0.666 0.816
RFGB 0.79 0.702 0.834 0.913 4.1. Feature Impact Analysis
Bikaner, India SDANN 0.71 0.560 0.546 0.739
RFGB 0.70 0.568 0.563 0.750
An ablation study is conducted involving both the SDANN
Alice Springs, Australia SDANN 0.72 0.643 0.835 0.914
and the RFGB models, and a comprehensive investigation into
feature[109] set impacts on SI forecasting is conducted across
RFGB 0.66 0.768 1.029 1.014
three geographically distinct regions: Hyderabad, Sukkur, and
Las Vegas, USA SDANN 0.84 0.635 0.705 0.839
Turbat detailed in Appendices 4–9 (Supporting Information). The
RFGB 0.80 0.735 0.883 0.939 study delves into the efficacy of different feature subsets, culmi-
Marrakech, Morocco SDANN 0.75 0.676 0.804 0.897 nating in a comparative assessment that underscores the supe-
RFGB 0.75 0.710 0.831 0.912 rior predictive capability of the full feature set configuration.
Luxor, Egypt SDANN 0.85 0.407 0.306 0.553 The SDANN model’s feature analysis for the Hyderabad
RFGB 0.78 0.514 0.454 0.674 (Figure 11a) dataset began with temperature statistics as the base-
line input features, yielding an R2 of 0.71. The addition of PS ini-
tially showed a decrease in performance, with the R2 dipping to
0.69. Yet, as the feature set expanded, the model’s performance
RFGB model, with an R2 of 0.79, shows a comparable but slightly improved, particularly with the inclusion of RH and WS. The
less correlation. The SDANN’s precision is further evidenced by complete feature set achieved an R2 of 0.74, MAE of 8.87, and an
lower MAE, MSE, and RMSE values of 0.594, 0.666, and 0.816, MSE of 177.98, indicating a robust predictive framework when
respectively, as compared to the RFGB’s 0.702, 0.834, and 0.913. utilizing all available variables.
The pattern of model performance carries through to the Bikaner, In the Sukkur dataset (Figure 11b), the SDANN model pre-
India dataset (Figure 10b), where both models showcase a closer sented a similar trend, with the full feature set reaching an R2 of
R2 value (0.71 for SDANN and 0.70 for RFGB) and near-identical 0.79. However, a notable anomaly emerged when the subset [Ts ,
MAE, MSE, and RMSE scores, suggesting a nearly equivalent Tmax , Tmin , WS, RH, PS] was examined, where the R2 plummeted
predictive capability in this region. to −12.74, suggesting an aberration in the data or model overfit-
Moving to the Alice Springs, Australia dataset (Figure 10c), the ting. This highlights the need for careful feature engineering and
SDANN maintains its leading position with a higher R2 score model validation to ensure the reliability of predictive insights.
of 0.72 against the RFGB model’s 0.66, indicating a more con- The Turbat dataset (Figure 11c) revealed the challenges inher-
sistent fit with the actual data. The SDANN’s MAE, MSE, and ent in SI prediction for the region, with the highest R2 , attained
RMSE figures—0.643, 0.835, and 0.914—are lower than those of being 0.44 for the SDANN model using all features. The adverse
the RFGB model, which are 0.768, 1.029, and 1.014, respectively. R2 values observed with certain feature subsets, such as −16.14
However, as we shift our focus to the Las Vegas, USA dataset for the subset including PS, indicate extreme model sensitivity
(Figure 10d), while the SDANN model still leads with an R2 of to the feature set composition and a potential overestimation of
0.84 over RFGB’s 0.80, the difference in MAE, MSE, and RMSE is the model’s complexity, necessitating a more refined approach to
modest, underscoring the competitive nature of the RFGB model model tuning and validation.
in this American context. Comparatively, the RFGB model exhibited consistent improve-
In the Marrakech, Morocco dataset (Figure 10e), both mod- ment in performance with the inclusion of additional features
els record an R2 of 0.75, indicating their similar proficiency in across all datasets. For Hyderabad (Figure 11d), the incremental
capturing the variation in SI values. The SDANN model has feature inclusion led to a maximum R2 of 0.74, and for Sukkur
marginally better MAE and MSE values (0.676 and 0.804) com- (Figure 11e), the full feature set achieved an R2 of 0.80. However,
pared to the RFGB’s 0.710 and 0.831, while the RMSE scores are similar to the SDANN, the Turbat dataset (Figure 11f) posed a
closely matched at 0.897 and 0.912, respectively. This suggests greater challenge, with a maximum R2 of 0.43, indicating room
that either model could be deployed with nearly equal effective- for model enhancement or alternative predictive strategies.
ness in the Moroccan setting. The trend continues with the Luxor, Overall, Figure 11 of feature analysis for both SDANN and
Egypt dataset (Figure 10f), where the SDANN outshines with the RFGB models, clearly illustrates that the use of a comprehen-
highest R2 score observed at 0.85 and notably lower MAE, MSE, sive suite of input features generally provides the best results in
and RMSE values of 0.407, 0.306, and 0.553. The RFGB model, SI forecasting. This finding emphasizes the importance of utiliz-
while trailing with an R2 of 0.78 and error metrics of 0.514, 0.454, ing a full range of available environmental data to improve model
and 0.674, still represents a viable alternative for forecasting in accuracy. Nevertheless, the anomalies and challenges noted, es-
the Egyptian climate. pecially in the Turbat dataset, indicate that a one-size-fits-all
Through this analytical narrative, it becomes evident that approach may not be universally applicable. Thus, model cus-
while the SDANN model frequently exhibits superior accu- tomization and further exploration of feature interactions remain
racy across the datasets, the RFGB model remains a strong pivotal for advancing SI forecasting in varied geographical con-
contender, showcasing commendable predictive quality and of- texts.

Adv. Theory Simul. 2024, 2301289 2301289 (15 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

Figure 11. Line Plot for Feature Impact analysis; a) SDANN Hyderabad, b) SDANN Sukkur, c) SDANN Turbat, d) RFGB Hyderabad, e) RFGB Sukkur,
and e) RFGB Turbat.

4.2. SHAP Analysis predominantly positive SHAP values, indicating that higher
maximum temperatures are usually associated with higher
Building on the ablation studies, the SHapley Additive exPlana- predicted values of SI, which is consistent with the general
tions (SHAP)[110] analysis provides a deeper interpretative layer understanding of SI dynamics.
to the RFGB model’s functionality by measuring the impact of The PS feature predominantly displays positive SHAP values,
each feature on the model’s predictions for SI. SHAP analysis suggesting its critical role in increasing the model’s SI predic-
is a sophisticated method used to interpret machine learning tions. SH and RH display a complex interaction with the SI
models.[111] It explains the output of any model by computing the
contribution of each feature to a prediction. SHAP values provide
an understanding of how much each predictor variable impacts
the model’s output. In essence, SHAP values can explain the pre-
diction of an instance by quantifying the contribution of each fea-
ture to the prediction.
The SHAP values, as visualized in the plot for Sukkur dataset
in Figure 12, reflect the magnitude and direction of a feature’s
impact on the model output. Features that push the prediction
higher are represented with positive SHAP values, while those
that lower the prediction are indicated by negative values. The
color gradient from blue to pink denotes the low to high range of
the feature values.
In this analysis, we observe that temperature-related
features—Tmin , Tmax , and Ts —have a mixed influence on
the model’s predictions. Tmax particularly stands out with its Figure 12. SHAP Analysis Plot for Sukkur Dataset.

Adv. Theory Simul. 2024, 2301289 2301289 (16 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

predictions. Their SHAP values are distributed across both the and regional customization. Additionally, the complexity inher-
positive and negative spectrum, suggesting that the relationship ent in these sophisticated models, especially the RFGB, intro-
between humidity and SI is not straightforward and may depend duces computational demands that could challenge resource con-
on the interaction with other factors within the model. WS, straints.
with its cluster of points leaning toward the negative side of the Future directions could involve exploring more diverse
SHAP value spectrum, suggests an inverse relationship with datasets, particularly for regions presenting complex meteorolog-
SI predictions. This could imply that in the context of Sukkur, ical conditions, to further refine model accuracy. Advanced fea-
higher wind speeds might correlate with factors that reduce SI, ture engineering and the potential application of techniques such
such as increased cloud cover or atmospheric disturbances. as transfer learning to the SDANN model may offer new avenues
The distribution of SHAP values across the features not only for enhancing predictive versatility and effectiveness.
reveals the individual impact of each feature but also hints at the In essence, this research not only highlights the robust pre-
underlying complexity of their interactions. For example, tightly dictive performance of the RFGB and SDANN models but also
grouped SHAP values indicate a consistent impact of a feature on marks a significant contribution to the domain of renewable en-
the model output, while a wider dispersion suggests variability in ergy forecasting. The insights gleaned from this study serve as a
influence, which could be due to the interplay with other features foundation for future innovations in technology and sustainable
in the model. energy policy, driving forward the capabilities of SI forecasting in
This SHAP analysis is instrumental in understanding the con- support of renewable energy development and management.
tribution of each feature within the RFGB model and thus can
be leveraged to fine-tune the model for even greater accuracy in
SI prediction. The insights drawn from this analysis are espe-
6. Conclusion
cially useful for stakeholders in the solar energy sector, providing The comprehensive analysis presented in this research article de-
a clearer picture of the factors that significantly influence SI. lineates the efficacy of advanced deep learning models for SI fore-
casting, underscored by quantitative assessments derived from
5. Discussion and Analysis extensive data across three different datasets. Our investigation
leveraged the SDANN and RFGB models, revealing significant
Our investigation into SI forecasting through advanced machine insights into their predictive capacities and comparing them with
learning models, specifically the RFGB and SDANN models, un- established models like XGBOOST, CAT Boost, GRU, and KNN-
derscores a pivotal advancement in predictive capabilities across XGBOOST. Additionally, the ablation study conducted further il-
various geographical locations. These models, set against the luminated the impact of various features on the models’ perfor-
backdrop of a comprehensive comparison with contemporaries mance. The keynote conclusions are as follows:
such as XGBOOST, CatBOOST, GRU, and KNN-XGBOOST,
alongside thorough ablation studies, have elucidated a rich land- 1) RFGB model showcased exceptional predictive capabilities,
scape of insights. distinguishing itself with remarkable efficiency across differ-
The nuanced calibration of the RFGB model, balancing com- ent datasets. Notably, the Sukkur dataset, achieved an impres-
plexity with the imperative to avoid overfitting, has been instru- sive R2 value of 0.80, alongside the lowest Mean Squared Error
mental in its standout performance. This precision in model de- (MSE) of 147.22 and Mean Absolute Error (MAE) of 8.77.
sign is mirrored in its adeptness at capturing intricate environ- 2) SDANN model, renowned for its intricate architecture,
mental patterns, which is particularly notable in datasets where demonstrated high precision in SI forecasting, particularly ex-
these characteristics are well-represented. Similarly, the SDANN celling in the Hyderabad dataset where it mirrored actual SI
model, with its deep neural network architecture, has shown an values with significant accuracy. With a compelling R2 value
impressive aptitude for accuracy across diverse settings, evidenc- of 0.74, akin to RFGB’s, and closely matched MSE and MAE
ing the power of advanced deep learning in SI forecasting. The values in regions such as Hyderabad and Sukkur, SDANN
scatter plots in Figure 9 and supplementary data for the Hyder- proved its adeptness in capturing the nuanced dynamics of
abad and Turbat datasets visually reinforce the high degree of pre- SI.
dictive accuracy achieved by both models, particularly within the 3) Geographical validations across Ahvaz, Bikaner, Alice
Sukkur dataset where their data points closely align with the line Springs, Las Vegas, Marrakech, and Luxor showcased the
of ideal prediction. adaptability of both SDANN and RFGB models to diverse
Ablation studies further illuminate the significant impact of climates, underlining their predictive accuracy’s sensitivity
deliberate feature selection on the RFGB model’s performance, to regional factors. The SDANN model, in particular, demon-
showcasing systematic improvements in accuracy as the fea- strated its versatility, despite variations in performance
ture set expanded. This meticulous approach to model refine- indicators such as R2 and MAE, emphasizing the importance
ment reveals the importance of aligning model inputs with the of tailoring models to specific geographical conditions.
nuanced dynamics of climatic factors, yet also highlights areas 4) Comparative analysis revealed the SDANN and RFGB mod-
where enhancements are needed, particularly in environments els’ strengths against traditional forecasting methods like XG-
with unique complexities. BOOST, CAT Boost, GRU, and KNN-XGBOOST. Specifically,
Despite the demonstrated strengths, it is imperative to ac- RFGB demonstrated a slight edge in performance, with both
knowledge certain limitations. Tailoring these models to specific models surpassing others in metrics such as R2 values. For
regional climates might necessitate adjustments, a process that example, in certain datasets, RFGB achieved an R2 value of
is less a limitation and more an opportunity for optimization 0.80, closely followed by SDANN, showcasing their superior

Adv. Theory Simul. 2024, 2301289 2301289 (17 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

handling of complex data interplays compared to XGBOOST, [4] B. Zakeri, K. Paulavets, L. Barreto-Gomez, L. G. Echeverri, S.
CAT Boost, GRU, and KNN-XGBOOST. Pachauri, B. Boza-Kiss, C. Zimm, J. Rogelj, F. Creutzig, D. Ürge-
5) Ablation studies provided critical insights into the impact Vorsatz, D. G. Victor, M. D. Bazilian, S. Fritz, D. Gielen, D. L.
of feature selection on model performance. For the RFGB McCollum, L. Srivastava, J. D. Hunt, S. Pouya, Energies (Basel) 2022,
15, 61141.
model, the systematic addition of features such as tempera-
[5] F. Sher, O. Curnick, M. T. Azizan, Sustainability 2021, 13, 1.
ture statistics, atmospheric pressure (PS), relative humidity
[6] M. Mele, A. R. Gurrieri, G. Morelli, C. Magazzino, Environ. Sci. Pollut.
(RH), and wind speed (WS) incrementally improved its accu- Res. Int. 2021, 28, 41127.
racy. In particular, the inclusion of all selected features ele- [7] M. Shahbaz, B. A. Topcu, S. S. Sarıgül, X. V. Vo, Renew Energy 2021,
vated the model’s R2 to 0.762 in Hyderabad and to an optimal 178, 1370.
0.801 in Sukkur, affirming the pivotal role of comprehensive [8] M. J. B. Kabeyi, O. A. Olanrewaju, Front. Energy Res. 2022, 9.
feature analysis in bolstering predictive precision. [9] N. Alsagr, S. van Hemmen, Environ. Sci. Pollut. Res. Int. 2021, 28,
25906.
The development and evaluation of the RFGB and SDANN [10] Y. Demirel, Energy 2016, 441.
models significantly advance SI forecasting, essential for [11] J. Wang, J. Yao, H. Chen, X. Li, J. Lin, Compar. Analysis Power Load
improving solar energy predictions and optimizing power Forecast. ML Meth. 2023, 12604, 126042.
generation. Both models demonstrated robust performances, [12] A. S. Elberawi, M. Belal, Future Comput. Informat. J. 2021, 6, 45.
[13] T. Ahmad, D. Zhang, C. Huang, Energy 2021, 231, 120911.
with the RFGB model excelling in adaptability and accuracy
[14] F. Gerges, M. C. Boufadel, E. Bou-Zeid, H. Nassif, J. T. L. Wang, Pro-
across diverse conditions, and the SDANN model showcasing
ceed. 2023 7th Int. Conf. Mach. Learn. Soft Comput. 2023, 102.
deep learning strengths, albeit with room for improvement in [15] P. Kumari, D. Toshniwal, J. Clean Prod. 2021, 318, 128566.
complex environments. These findings highlight the need for [16] F. Gerges, M. C. Boufadel, E. Bou-Zeid, H. Nassif, J. T. L. Wang,
model customization to regional climates and suggest exploring Knowl. Inf. Syst. 2024, 66, 613.
broader datasets and advanced techniques to enhance future [17] D. Harada, P. Chinnavornrungsee, S. Kittisontirak, N. Chollacoop,
model efficiency and adaptability. This research contributes S. Songtrai, K. Sriprapha, J. Yoshino, T. Kobayashi, Jpn. J. Appl. Phys.
valuable insights into SI forecasting, with implications for global 2023, 62, SK1056.
renewable energy strategies. [18] R. A. A. Ramadhan, Y. R. J. Heatubun, S. F. Tan, H. J. Lee, Renew
Energy 2021, 178, 1006.
[19] P. Seshadri, B. P. Bagavat, A. Kumar B, K. H., K. G., S. S., Electric
Supporting Information Power Comp. Syst. 2021, 49, 1379.
[20] S. Ullah, Z. Zuo, F. Zhang, J. Zheng, S. Huang, Y. Lin, I. Iqbal, Y. Sun,
Supporting Information is available from the Wiley Online Library or from M. Yang, L. Yan, Remote Sens (Basel) 2020, 12, 1.
the author.
[21] S. Gbémou, J. Eynard, S. Thil, E. Guillot, S. Grieu, Energies 2021, 4,
3192.
[22] A. K. Mandal, R. Sen, S. Goswami, B. Chakraborty, Symmetry (Basel)
Conflict of Interest 2021, 13.
The authors declare no conflict of interest. [23] M. Golam, R. Akter, J. M. Lee, D. S. Kim, IEEE Geosci. Remote Sens.
Lett. 2022, 19, 1.
[24] A. T. Dahiru, C. W. Tan, A. L. Bukar, K. Y. Lau, C. L. Toh, S. Salisu,
Data Availability Statement 2021 IEEE Conf. Energy Convers. (CENCON) 2021, 214.
[25] M. Viswavandya, B. Sarangi, S. Mohanty, A. Mohanty, Adv. Intell.
The data that support the findings of this study are openly available in Syst. Comput. 2020, 990, 751.
POWER | Data Access Viewer at https://power.larc.nasa.gov/data-access- [26] M. A. Hassan, N. Bailek, K. Bouchouicha, S. C. Nwokolo, Renew En-
viewer/, reference number 0.
ergy 2021, 171, 191.
[27] I. Sansa, Z. Boussaada, N. M. Bellaaj, Energies (Basel) 2021, 14,
Keywords 6920.
[28] C. McHugh, S. Coleman, D. Kerr, Mach. Learn. Appl. 2022, 9,
artificial neural network (ANN), deep learning (DL), gradient Boosting 100383.
(GB), machine learning (ML), random Forest (RF), solar forecasting [29] I. Iqbal, I. Ahmed, ICESP 2015, 2014, 1.
[30] M. Massaoudi, I. Chihi, L. Sidhom, M. Trabelsi, S. S. Refaat, H. Abu-
Received: December 30, 2023 Rub, F. S. Oueslati, IEEE Access 2021, 9, 36571.
Revised: April 7, 2024 [31] H. Gao, S. Qiu, J. Fang, N. Ma, J. Wang, K. Cheng, H. Wang, Y. Zhu,
Published online: D. Hu, H. Liu, J. Wang, Sustainability 2023, 15.
[32] A. Mellit, A. M. Pavan, V. Lughi, Renew Energy 2021, 172, 276.
[33] R. Kim, K. M. Kim, J. H. Ahn, J. Korean Soc. Environm. Engin. 2021,
43, 347.
[1] J. L. Holechek, H. M. E. Geli, M. N. Sawalhah, R. Valdez, Sustainabil- [34] N. M. Sabri, M. El Hassouni, Energ. Sourc., Part A: Recov., Utilizat.,
ity 2022, 14. Environ. Effects 2022, 44, 6303.
[2] S. A. Qadir, H. Al-Motairi, F. Tahir, L. Al-Fagih, Energy Reports 2021, [35] Y. Ledmaoui, A. El Fahli, A. Elmaghraoui, M. El Aroussi, R. Saadane,
7, 3590. A. Chehri, 2023 Int. Conf. Digital Age Technol. Adv. Sustainable De-
[3] D. Bogdanov, M. Ram, A. Aghahosseini, A. Gulagi, A. S. Oyewo, M. velop. (ICDATA) 2023, 120.
Child, U. Caldera, K. Sadovskaia, J. Farfan, L. De Souza Noel Simas [36] M. F. Hanif, M. S. Naveed, M. Metwaly, J. Si, X. Liu, J. Mi, M. F. Hanif,
Barbosa, M. Fasihi, S. Khalili, T. Traber, C. Breyer, Energy 2021, 227, M. S. Naveed, M. Metwaly, J. Si, X. Liu, J. Mi, AIMS Energy 2024, 12,
120467. 350.

Adv. Theory Simul. 2024, 2301289 2301289 (18 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

[37] P. Pawar, N. Mithulananthan, M. Q. Raza, GPECOM 2020, 2020, [70] Z. Pang, F. Niu, Z. O’Neill, Renew. Energy 2020, 156, 279.
226. [71] Z. Wang, C. Tian, Q. Zhu, M. Huang, Energies (Basel) 2018, 11, 68.
[38] X. Huang, Q. Li, Y. Tai, Z. Chen, J. Zhang, J. Shi, B. Gao, W. Liu, [72] M. A. Munir, A. Khattak, K. Imran, A. Ulasyar, A. Khan, 1st Int. Conf.
Renew Energy 2021, 171, 1041. Electr., Commun. Comp. Engin., 2019, 24.
[39] K. Anuradha, D. Erlapally, G. Karuna, V. Srilakshmi, K. Adilakshmi, [73] R. Gallo, M. Castangia, A. Macii, E. Macii, E. Patti, A. Aliberti, Eng
E3S Web Conf. 2021, 309, 01163. Appl Artif Intell 2022, 116, 105493.
[40] M. Kumar, K. Namrata, N. Kumar, G. Saini, J. Grid Comput. 2023, [74] J. W. Emerson, W. A. Green, B. Schloerke, J. Crowley, D. Cook, H.
21, 28. Hofmann, H. Wickham, J. Computat. Graph. Statis. 2013, 22, 79.
[41] S. Boubaker, M. Benghanem, A. Mellit, A. Lefza, O. Kahouli, L. Kolsi, [75] N. J. Cox, Stata J 2009, 9, 621.
IEEE Access 2021, 9, 36719. [76] Z. Lv, W. Deng, Z. Zhang, N. Guo, G. Yan Proceed. - 2019 IEEE Intl.
[42] S. Nosouhian, F. Nosouhian, A. K. Khoshouei, A Review of Recurrent Conf Parallel Distrib.Process. Appl., Big Data Cloud Comput., Sus-
Neural Network Architecture for Sequence Learning: Comparison tain. Comput. Commun., Soc. Comput. Netw, 2019, 802.
between LSTM and GRU, https://www.preprints.org/manuscript/ [77] C. Zhang, S. Lin, D. Zhang, ACM Int. Conf. Proceed. Series 2022, 77.
202107.0252/v1. [78] P. Wang, Y. Liu, Q. Sun, Y. Bai, C. Li, Sustainability (Switzerland) 2022,
[43] C. H. Liu, J. C. Gu, M. T. Yang, IEEE Access 2021, 9, 17174. 14, 12286.
[44] A. Rai, A. Shrivastava, K. C. Jana, Optik (Stuttg) 2022, 252, 168515. [79] T. T. Le, W. Fu, J. H. Moore, Bioinformatics 2020, 36, 250.
[45] H. Raju, S. Das, Solar Physics 2021, 296, 1. [80] Z. Sultana, M. M. Rahman, L. Nahar, J. Appl. Computer Sci. Mathem.
[46] M. Jaihuni, J. K. Basak, F. Khan, F. G. Okyere, T. Sihalath, A. Bhujel, 2022, 16, 18.
J. Park, D. H. Lee, H. T. Kim, ISA Trans 2022, 121, 63. [81] P. P. Palmes, A. Kishimoto, R. Marinescu, P. Ram, E. Daly 2021 De-
[47] P. Kumari, D. Toshniwal, Appl. Energy 2021, 295, 117061. signing Machine Learning Pipeline Toolkit for AutoML Surrogate
[48] S. Ziyabari, Z. Zhao, L. Du, S. K. Biswas, IEEE Trans Ind Appl 2023, Modeling Optimization, arXiv preprint arXiv:2107.01253.
59, 5293. [82] S. Gholizadeh, J. Robot. Automat. Res. 2022, 3.
[49] E. M. Al-Ali, Y. Hajji, Y. Said, M. Hleili, A. M. Alanzi, A. H. Laatar, M. [83] I. Stančin, A. Jović, 2019 42nd International Convention on Infor-
Atri, Mathematics 2023, 11, 676. mation and Communication Technology, Electronics and Microelec-
[50] S. Ghimire, R. C. Deo, N. Raj, J. Mi, Appl. Energy 2019, 253, 113541. tronics (MIPRO), IEEE, pp. 977–982.
[51] H. Zang, L. Liu, L. Sun, L. Cheng, Z. Wei, G. Sun, Renew. Energy 2020, [84] Welcome to Colaboratory - Colaboratory, https://colab.research.
160, 26. google.com/.
[52] B. Gu, X. Li, F. Xu, X. Yang, F. Wang, P. Wang, Sustainability (Switzer- [85] W. Wu, G. C. Dandy, H. R. Maier, Environ. Model. Soft. 2014, 54,
land) 2023, 15. 108.
[53] M. Kumar, K. Namrata, N. Kumar, Adv. Electr. Electr. Engin. 2023, 20, [86] H. R. Maier, A. Jain, G. C. Dandy, K. P. Sudheer, Environ. Model.
549. Software 2010, 25, 891.
[54] O. Jogunola, A. S. Ajagun, B. Adebisi, A. M. Aibinu, J. A. Ojo, The [87] T. Vujičić, T. Matijević, J. Ljucović, A. Balota, Z. Ševarac, Central Eu-
5th Int. Conf. Fut. Netw. Distri. Sys. 2021, 214. ropean Conference on Information and Intelligent Systems, 2016.
[55] S. Sivarajan, S. D. S. S. Jebaseelan, Int. J. Electr. Electr. Engin. 2023, [88] P. Raut, A. Dani, Correlat. Bet. Num. Hidden Layers Accu. Artific. Neu-
10, 221. ral Net. 2020, 513.
[56] J. Yan, Y. Sha, Y. Zhang, T. Li, J. Zhang, 2023 3rd Int. Conf. Energ., [89] I. Iqbal, G. A. Odesanmi, J. Wang, L. Liu, Appl. Artif. Intell. 2021, 35,
Power Electr. Engin. (EPEE) 2023, 261. 697.
[57] P. Kumari, D. Toshniwal, J. Clean. Prod. 2021, 318, 128566. [90] A. Holzinger, Machine Learning for Health Informatics, Springer In-
[58] C. Halton, E. Rasure, Predictive Analytics: Definition, Model Types, and ternational Publishing, 2016, pp. 1–24.
Uses, Investopedia, 2023. [91] F. Psallidas, Y. Zhu, B. Karlas, M. Interlandi, S. Krishnan, B. Kroth,
[59] S. Manju, M. Sandeep, J. Clean Prod. 2019, 230, 116. V. Emani, W. Wu, C. Zhang, M. Weimer, A. Floratou, C. Curino, K.
[60] S. Ahmad, M. Parvez, T. A. Khan, O. Khan, Environm. Challeng. 2022, Karanasos, ACM SIGMOD Record 2022, 51, 30.
9, 100634. [92] A. Paul, D. P. Mukherjee, P. Das, A. Gangopadhyay, A. R. Chintha, S.
[61] Ü. Ağbulut, A. E. Gürel, Y. Biçen, Renewable Sustainable Energy Rev. Kundu, IEEE Transact. Image Process. 2018, 27, 4012.
2020, 135, 110114. [93] M. Schonlau, R. Y. Zou, Stata Journal 2020, 20, 3.
[62] S. Islam, N. K. Roy, Energy Reports 2023, 9, 6063. [94] C. Bentéjac, A. Csörgő, G. Martínez-Muñoz, Artif. Intell. Rev. 2021,
[63] B. K. Puah, L. W. Chong, Y. W. Wong, K. M. Begam, N. Khan, M. A. 54, 1937.
Juman, R. K. Rajkumar, Renew Energy 2021, 164, 908. [95] V. K. Ayyadevara in Pro Machine Learning Algorithms, Apress, New
[64] C. G. Villegas-Mier, J. Rodriguez-Resendiz, J. M. Álvarez-Alvarado, York 2018, 117.
H. Jiménez-Hernández, Á. Odry, Micromachines (Basel) 2022, 13, [96] G. Biau, B. Cadre in Advances in Contemporary Statistics and Econo-
1406. metrics: Festschrift in Honor of Christine Thomas-Agnan, Springer In-
[65] O. M. Mbah, C. I. Madueke, R. Umunakwe, M. N. Agba, J. Engin. Sci. ternational Publishing, Berlin 2021, 23.
2022, 9, E1. [97] Scikit-Learn Metrics and Scoring: Quantifying the Quality of Predic-
[66] Weatherspark.com Compare the Climate and Weather in Sukkur, tions, https://scikit-learn.org/stable/modules/model_evaluation.
Hyderabad, and Turbat, https://weatherspark.com/compare/y/ html.
106589∼106562∼106188/Comparison-of-the-Average-Weather-in- [98] M. Guermoui, F. Melgani, K. Gairaa, M. L. Mekhalfi, J Clean Prod
Sukkur-Hyderabad-and-Turbat. 2020, 258, 120357.
[67] Solargis Global Solar Atlas 2.0 (2018) GHI Map of Pakistan Utiliz- [99] N. Krishnan, K. R. Kumar, C. S. Inda, J Clean Prod 2023, 388, 135860.
ing Solargis Data, Energy Sector Management Assistance Program [100] S. Sheik Mohammed, F. Titus, S. B. Thanikanti, S. M. Sulaiman, S.
(ESMAP), World Bank Group, 2007, 1. Deb, N. M. Kumar, Sustainability 2022, 14, 3498.
[68] NASA, POWER Data Access Viewer v2.0.0 2023, https://power.larc. [101] Y. Wang, D. Millstein, A. D. Mills, S. Jeong, A. Ancell, Sol. Energy
nasa.gov/beta/data-access-viewer/. 2022, 231, 846.
[69] Ü. Ağbulut, A. E. Gürel, Y. Biçen, Renew. Sustainab. Energy Rev. 2021, [102] B. Belmahdi, M. Louzazni, M. Marzband, A. El Bouardi, Forecasting
135, 110114. 2023, 5, 172.

Adv. Theory Simul. 2024, 2301289 2301289 (19 of 20) © 2024 Wiley-VCH GmbH
 25130390, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adts.202301289 by Peking University Health, Wiley Online Library on [02/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advtheorysimul.com

[103] M. A. Ali, A. Elsayed, I. Elkabani, M. E. Youssef, G. E. Hassan, Environ [108] S. Husain, U. A. Khan, Environ. Prog. Sustain Energy 2022, 41,
Dev Sustain 2023, 1. e13782.
[104] Y. C. Chen, Biostat Epidemiol 2017, 1, 161. [109] J. Wang, H. Zhong, X. Lai, Q. Xia, Y. Wang, C. Kang, IEEE Trans. Smart
[105] L. Benali, G. Notton, A. Fouilloy, C. Voyant, R. Dizene, Renew Energy Grid 2019, 10, 1417.
2019, 132, 871. [110] Y. G. Lee, J. Y. Oh, D. Kim, G. Kim, J. Electr. Engin. Technol. 2023, 18,
[106] X. Li, L. Ma, P. Chen, H. Xu, Q. Xing, J. Yan, S. Lu, H. Fan, L. Yang, 579.
Y. Cheng, Energy Rep. 2022, 8, 1087. [111] C. Cakiroglu, S. Demir, M. Hakan Ozdemir, B. Latif Aylak,
[107] J. Fan, X. Wang, F. Zhang, X. Ma, L. Wu, J. Clean Prod. 2020, 248, G. Sariisik, L. Abualigah, Expert. Syst. Appl. 2024, 237,
119264. 121464.

Adv. Theory Simul. 2024, 2301289 2301289 (20 of 20) © 2024 Wiley-VCH GmbH