A GAN-based approach to solar radiation prediction: data augmentation and model optimization for Saudi Arabia

Traditional approaches

Machine learning approaches

Machine learning techniques, including random forest and gradient boosting, have shown notable improvements in solar radiation prediction. Chaibi et al. (2018) combined satellite and ground-based observations using these models in Riyadh, enhancing accuracy compared to models trained on standalone datasets. However, ML models often require region-specific retraining and are computationally intensive, which limits their scalability.

Deep learning approaches

Deep learning models, such as artificial neural networks (ANNs) and convolutional neural networks (CNNs), further advance prediction capabilities by learning complex nonlinear relationships. Hanif et al. (2024), for instance, incorporated meteorological variables like wind speed and humidity into ANN models in Dammam, significantly improving daily predictions. CNNs are proficient at extracting spatial features, while long short-term memory (LSTM) networks excel at modeling temporal dependencies (Song et al., 2020). Despite their strong performance, DL models require large volumes of high-quality data and careful preprocessing to prevent overfitting.

Hybrid models

Hybrid models, which integrate physical, statistical, and ML approaches, have also gained prominence. Hedar et al. (2021) combined weather simulation outputs with ML models to improve solar radiation forecasts in eastern Saudi Arabia, demonstrating enhanced performance under extreme conditions. Similarly, Khan & Khan (2024) implemented a CNN-LSTM hybrid model in the UAE to capture spatiotemporal dependencies, with promising results applicable to the Saudi context. Nevertheless, these hybrid approaches are computationally expensive and often rely on extensive data infrastructure, posing challenges for real-time implementation in data-scarce regions.

Emerging hybrid frameworks (Delarami, Mohammadbeigi & Gharib, 2024) and edge-computing optimizations (Minh et al., 2022) have further improved the feasibility of generative data augmentation for renewable energy systems, supporting the scalable deployment goals outlined in ‘Future Directions’.

GAN-based approaches

GANs, introduced by Goodfellow et al. (2020), comprise two neural networks—a generator and a discriminator—that compete in an adversarial process. The generator synthesizes data, while the discriminator assesses its authenticity. This dynamic training process enables GANs to produce realistic synthetic data, effectively augmenting training datasets and improving model performance (Fayaz et al., 2024).

Recent advancements in GAN architectures, such as StyleGAN and BigGAN, have expanded their capabilities. StyleGAN enables fine-grained control over generated outputs, allowing synthetic solar radiation data to be customized for specific regional climates (Kousounadis-Knousen et al., 2023). BigGAN, on the other hand, enhances scalability and fidelity through the use of larger batch sizes and advanced normalization strategies (Quaicoo et al., 2024). These innovations are especially promising for improving prediction accuracy and adaptability in data-limited settings like Saudi Arabia.

Summary

The evolution of solar radiation prediction methodologies reflects a clear transition from traditional statistical and physical models to more advanced ML, DL, and GAN-based techniques. While early models provide valuable foundational understanding, their limitations in handling nonlinearities and adapting to real-time environments necessitate the adoption of modern approaches. ML and DL models offer superior spatial-temporal modeling capabilities but require extensive data and computational resources. GANs emerge as a compelling alternative, addressing data scarcity while improving predictive accuracy through realistic data augmentation. Future research should prioritize the adaptation of GAN architectures to Saudi Arabia’s diverse climatic regions and the integration of these models with real-time data pipelines to maximize their practical impact on solar radiation forecasting.

Solar radiation modeling

Solar radiation modeling is fundamental to understanding the atmospheric and environmental factors that influence solar energy availability. A robust modeling framework ensures that the synthetic data generated by the proposed system aligns with real-world patterns and captures the inherent variability of solar radiation (Izzi, Martella & Longo, 2024).

The modeling process begins with the calculation of global horizontal irradiance (GHI), which is derived from three components: direct normal irradiance (DNI), diffuse horizontal irradiance (DHI), and the cosine of the solar zenith angle. This relationship is expressed as:

(1) $$GHI=DNI\cdot \cos (\theta )+DHI.$$

This equation serves as the foundation for simulating the direct, diffuse, and angular components of solar radiation—an essential element for photovoltaic (PV) system optimization.

To enhance this baseline approach, radiative transfer models are employed. These models simulate the interaction between solar radiation and atmospheric elements such as aerosols, clouds, and gases (Zhang et al., 2024). By incorporating these physical processes, radiative transfer models ensure that the synthetic data accurately reflects region-specific solar radiation characteristics—particularly vital in the context of Saudi Arabia’s varied climatic zones.

However, physical models alone often fall short in capturing local variability. To address this, empirical corrections are introduced, adjusting model outputs to account for unique climatic conditions in specific regions—such as the desert and coastal areas of the Kingdom (Daxini, 2024). These corrections bring the synthetic data into closer alignment with observed solar radiation patterns, enabling more effective generalization across regions with distinct meteorological profiles.

Machine learning and deep learning fundamentals

The complex and nonlinear nature of solar radiation data necessitates the application of advanced ML and DL techniques. These methods are capable of uncovering intricate patterns, enabling accurate forecasting even in climatically diverse regions (Bamisile et al., 2022).

At the core of ML approaches lies ensemble learning, which combines multiple base learners to enhance predictive performance (Alam et al., 2023). This process is mathematically represented as:

(2) $$\hat{y}=\sum _{i=1}^n{w}_i\cdot h_i(x)$$where $\hat{y}$ is the final predicted output, $w_i$ is the weight assigned to the $i$-th model, and $h_i(x)$ represents the prediction from the $i$-th model. Augmented datasets significantly improve ensemble learning by providing the diverse and high-quality data needed for more accurate and reliable predictions (Bamisile et al., 2022).

DL models like CNNs and LSTM networks offer additional capabilities. Training of these networks is typically achieved via backpropagation, which iteratively updates model weights to minimize prediction error (Yahiaoui & Assas, 2023):

(3) $$w_{ij}=w_{ij}-\eta \cdot \frac{\mathrm{\partial }L}{\mathrm{\partial }{w}_{ij}}$$where $w_{ij}$ is the weight between neurons $i$ and $j$, $\eta$ is the learning rate, and $\frac{\mathrm{\partial }L}{\mathrm{\partial }{w}_{ij}}$ is the gradient of the loss function with respect to the weight.

Generative adversarial networks

GANs have revolutionized data generation by addressing scarcity and improving the robustness of predictive models. Through adversarial training, GANs generate synthetic data that closely resembles real-world observations (Wen et al., 2023).

At the heart of a GAN are two neural networks—the generator (G) and the discriminator (D)—trained simultaneously in a competitive setting (Nematchoua, Orosa & Afaifia, 2022). This adversarial setup is captured in the GAN’s minimax objective function:

(4) $$\underset{G}{min}\underset{D}{max}V(D,G)={\mathbb{E}}_{x\sim p_{data}(x)}[\log D(x)]+{\mathbb{E}}_{z\sim p_z(z)}[\log (1-D(G(z)))].$$

Conditional GANs (cGANs) introduce additional flexibility by incorporating conditional inputs $y$ (Saxena & Cao, 2021):

(5) $$\underset{G}{min}\underset{D}{max}V(D,G)={\mathbb{E}}_{x,y\sim p_{data}(x|y)}[\log D(x|y)]+{\mathbb{E}}_{z\sim p_z(z),y\sim p_{data}(y)}[\log (1-D(G(z|y)|y))].$$

To overcome challenges like mode collapse, Wasserstein GANs (WGANs) are employed, using the Wasserstein distance metric (Khare, Wadhvani & Shukla, 2022):

(6) $$W(p_r,p_g)=\underset{||f|{|}_L\le 1}{sup}{\mathbb{E}}_{x\sim p_r}[f(x)]-{\mathbb{E}}_{x\sim p_g}[f(x)].$$

Evaluation metrics

Robust evaluation metrics are essential for validating both the quality of synthetic data and the performance of predictive models. The root mean square error (RMSE) is defined as Mustafa et al. (2022):

(7) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}\sum _{i=1}^n{({\hat{y}}_i-y_i)}^2}.$$

The mean absolute error (MAE) offers an intuitive measure of prediction errors (Boubaker et al., 2021):

(8) $$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{n}\sum _{i=1}^n|y_i-{\hat{y}}_i|.$$

The structural similarity index (SSIM) evaluates the similarity between real and synthetic data (Boubaker et al., 2021):

(9) $$\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}(x,y)=\frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}.$$

The Kullback-Leibler (KL) Divergence measures distribution alignment (Dairi, Harrou & Sun, 2021):

(10) $$D_{\mathrm{K}\mathrm{L}}(P||Q)=\sum _{i}P(i)\log \frac{P(i)}{Q(i)}.$$

By integrating these theoretical components—ranging from physical and statistical modeling to GAN-based data augmentation and advanced DL architectures—this framework addresses the core challenges of solar radiation forecasting in data-scarce regions. The proposed approach not only ensures high model performance but also aligns with the sustainable energy goals of Saudi Arabia’s Vision 2030.

Overview

The methodology for this study, illustrated in Fig. 1, outlines a comprehensive framework designed to address data scarcity in solar radiation prediction across Saudi Arabia. As shown in the figure, our approach integrates advanced data generation techniques with state-of-the-art predictive modeling to ensure accurate and scalable forecasts tailored to the Kingdom’s diverse climatic regions, including desert, coastal, and mountainous areas.

Synthetic data generation

Climate-conditioned cGAN framework

Our climate-aware synthesis system uses a conditional GAN (cGAN) architecture (Fig. 2) consisting of:

• • Generator network:

  • –

    Five fully-connected layers (512 neurons each) with ReLU activation

  • –

    Tanh output layer for bounded solar irradiance values

  • –

    Concatenates random noise ( $z\sim \mathcal{N}(0,1)$) with meteorological conditions

• • Discriminator network:

  • –

    Four 1D convolutional layers (kernel size = 3, stride = 2) with LeakyReLU ( $\alpha =0.2$)

  • –

    Final dense layer with sigmoid activation

Predictive model integration

Evaluation metrics

The framework was evaluated using:

• • RMSE: (14) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}\sum _{i=1}^n{(y_i-{\hat{y}}_i)}^2}.$$

• • MAE: (15) $$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{n}\sum _{i=1}^n|y_i-{\hat{y}}_i|.$$

• • SSIM (Boubaker et al., 2021): (16) $$\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}(x,y)=\frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}.$$

Computing infrastructure

The implementation used:

• Hardware: NVIDIA RTX 3080 Ti GPU, 32 GB RAM

• Software: Python 3.10, TensorFlow 2.12

• Frameworks: Keras, scikit-learn

Algorithms and code

The implementation of the GAN framework in this study combines two essential components to generate high-quality synthetic solar radiation data tailored to specific environmental conditions:

• •

  Generator: The generator G creates synthetic solar radiation data that closely mirrors real-world patterns. By incorporating conditional inputs $\mathbf{c}$ (e.g., temperature, humidity), the generator customizes the outputs to reflect specific environmental contexts. The transformation can be represented as: (17) $${\mathbf{x}}_{synth}=G(\mathbf{z}|\mathbf{c}),\mathbf{z}\sim \mathcal{N}(0,1)$$

• where $\mathbf{z}$ is random noise and $\mathbf{c}$ represents the conditional meteorological parameters.

• •

  Discriminator: The discriminator D evaluates the authenticity of generated data through the adversarial loss: (18) $${\mathcal{L}}_{adv}=\mathbb{E}[\log D({\mathbf{x}}_{real}|\mathbf{c})]+\mathbb{E}[\log (1-D(G(\mathbf{z}|\mathbf{c})|\mathbf{c}))].$$

To overcome challenges such as mode collapse, we employed Wasserstein GANs with gradient penalty (WGAN-GP) (Gulrajani et al., 2017):

(19) $$L_{\mathrm{W}\mathrm{G}\mathrm{A}\mathrm{N}-\mathrm{G}\mathrm{P}}={\mathbb{E}}_{{\mathbf{x}}_{\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{l}}}[D({\mathbf{x}}_{\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{l}})]-{\mathbb{E}}_{{\mathbf{x}}_{\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}}}[D({\mathbf{x}}_{\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}})]+\lambda {\mathbb{E}}_{\hat{\mathbf{x}}}[{(||{\mathrm{\nabla }}_{\hat{\mathbf{x}}}D(\hat{\mathbf{x}}){||}_2-1)}^2]$$where $\hat{\mathbf{x}}$ is sampled along straight lines between real and generated data pairs.

Data availability

This study utilizes the following meteorological datasets:

• EUMETSAT SEVIRI satellite data (European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), 2023), which provides high-resolution radiance imagery covering the Arabian Peninsula from 2020 to 2023.

• Saudi National Center for Meteorology (NCM) (Saudi National Center for Meteorology, 2024), offering ground-based observations including solar irradiance, temperature, and humidity from multiple Saudi cities between 2018 and 2023.

The processed data, including merged satellite and ground observations, feature-engineered variables, and synthetic data samples, are provided as Supplemental Material to support the full reproducibility of the study.

Code availability

The complete implementation, including preprocessing, training, and evaluation scripts, is provided as Supplemental Material. The repository includes:

• A README.md file with setup instructions, system requirements (Python 3.10+, CUDA 11.7), and usage examples.

• Modular source code organized under directories such as configs/, data/, src/, and models/.

• Pre-trained model weights and synthetic data samples for reproducibility.

Reproducibility steps

Experiments can be reproduced using the provided scripts and configuration files. The standard workflow involves:

• 1.

  Installing dependencies from requirements.txt.

• 2.

  Training models via: python src/trainer.py –config configs/desert_zone.yaml.

• 3.

  Evaluating performance using: python src/metrics.py.

Synthetic data generation

The synthetic data, generated using the cGAN architecture depicted in Fig. 2, successfully replicates statistical characteristics of solar radiation across climatic zones, with minor deviations in coastal regions (Wasserstein distance: 0.4997). The key findings are presented below.

Distribution analysis

Figures 3 and 4 demonstrate our architecture’s capacity to preserve statistical properties. The minor deviations in coastal regions (Wasserstein distance: 0.4997) stem from the conditional inputs’ sensitivity to humidity variations.

Table 2 shows that incorporating 30 m SRTM digital elevation model (DEM) data reduced the Wasserstein distance by 22% for mountainous zones, confirming terrain elevation as a critical predictor.

Table 2:

Impact of DEM data on synthetic data quality in mountainous regions.

Metric	Without DEM	With DEM	Improvement
Wasserstein distance	0.51	0.40	22%
MAE (Test set)	0.18	0.14	22%

DOI: 10.7717/peerj-cs.3189/table-2

Statistical metrics

The statistical fidelity of synthetic data is quantified in Table 3, demonstrating near-perfect alignment in Wasserstein distances (0.4993 for temperature, 0.4997 for solar radiation) between real and generated distributions. As evidenced in Table 3, the variance of synthetic data ( $9.28\times {10}^{-7}$) closely matches the original distributions. The alignment between real and synthetic data is further demonstrated in Fig. 5, which shows the overlaid temperature distributions. Similarly, Fig. 6 presents the overlaid distributions for solar radiation, confirming the framework’s ability to preserve key statistical properties.

Table 3:

Statistical metrics for evaluating synthetic data quality.

Metric	Temperature	Solar radiation
Wasserstein distance	0.4993	0.4997
Original mean	0.5	0.5
Synthetic mean	0.0011	0.9995
Original variance	0.5	0.5
Synthetic variance	$9.28\times {10}^{-7}$	$1.63\times {10}^{-7}$

DOI: 10.7717/peerj-cs.3189/table-3

Model evaluation and performance

Performance metrics

Table 4 demonstrates significant improvements across all metrics. The large Cohen’s d for RMSE (1.32) indicates strong practical significance. The RMSE improvements across different regions are visually presented in Fig. 7, highlighting the consistent benefits of data augmentation. The MAE reductions shown in Fig. 8 further validate the effectiveness of the GAN-augmented approach.

Table 4:

Model performance metrics for real and augmented datasets.

Metric	Real data	Augmented data	Improvement	p-value
RMSE	2.56	2.18	15.2%	<0.01
MAE	1.86	1.49	19.9%	0.003
SSIM	0.82	0.91	11.0%	0.012
KL-Divergence	0.67	0.51	23.9%	0.008

DOI: 10.7717/peerj-cs.3189/table-4

Regional evaluation

Regional performance gains are quantified in Table 5, with the most significant RMSE reduction (17%) occurring in mountainous regions. As demonstrated in Table 5, coastal areas showed a 12.7% improvement (2.68 $\to$ 2.34 RMSE), while desert regions achieved a 14.5% reduction. The regional performance variations are comprehensively illustrated in Fig. 9, demonstrating the framework’s adaptability to different climatic conditions.

Table 5:

Regional performance metrics.

Region	Real data RMSE	Augmented data RMSE
Desert	2.21	1.89
Coastal	2.68	2.34
Mountainous	3.02	2.67

DOI: 10.7717/peerj-cs.3189/table-5

Statistical analysis

All results were validated using rigorous statistical methods:

• • Bootstrapped confidence intervals:

  • –

    1,000 iterations of test-set resampling

  • –

    95% CIs computed via percentile method

• • Hypothesis testing:

  • –

    Paired t-tests ( $\alpha =0.05$ with Bonferroni correction)

  • –

    Effect sizes reported using Cohen’s d

The narrow confidence intervals (e.g., $\pm$1.6% for RMSE) and significant p-values (p < 0.01) confirm the robustness of our findings.

Extended evaluation and practical validation

To address the need for a comprehensive evaluation, we conducted an extended analysis covering both quantitative and practical aspects. Our framework was validated using additional metrics such as SSIM and Kullback-Leibler (KL) divergence, confirming statistical alignment between real and synthetic data distributions. Regional performance improvements were rigorously analyzed across desert, coastal, and mountainous zones, highlighting the framework’s adaptability to diverse climatic conditions.

Furthermore, we performed an empirical IoT-based validation using Raspberry Pi 4 devices equipped with BME680 sensors in Riyadh (desert), Jeddah (coastal), and Abha (mountainous) over a 30-day period. The resulting RMSE values of 0.12, 0.15, and 0.18 respectively demonstrate the framework’s practical generalization capability and real-world applicability. These field results complement the simulation outcomes and further establish the reliability of GAN-augmented predictions in operational settings.

Empirical IoT validation

As shown in Fig. 10, we deployed our proof-of-concept on Raspberry Pi 4 devices with BME680 sensors across Saudi Arabia. Field results over 30 days showed RMSE values of 0.12 in Riyadh (desert), 0.15 in Jeddah (coastal), and 0.18 in Abha (mountainous). While the system demonstrated edge-compatibility, sensor noise necessitated Kalman filtering (Priyanka et al., 2024), highlighting the importance of robust preprocessing for field deployments. Figure 10 details the hardware and data pipeline enabling these results.

Future directions

While this study has demonstrated the effectiveness of a GAN-based framework in addressing data scarcity and improving solar radiation prediction, there are several promising avenues for further research and development.

One key direction is the integration of real-time forecasting with IoT-enabled meteorological stations. By leveraging IoT devices, the framework can process real-time meteorological inputs, enabling immediate predictions and enhancing its applicability in dynamic environments like renewable energy operations and smart grids. This real-time capability is crucial for adaptive energy management systems, particularly in rapidly changing climatic zones.

Another potential advancement lies in cross-regional generalization. Expanding the framework to diverse regions beyond Saudi Arabia would validate its robustness and adaptability to various climatic conditions. Collaborations with international meteorological organizations could facilitate data sharing, ensuring that the framework remains applicable to global renewable energy challenges.

The incorporation of advanced GAN architectures, such as StyleGAN, represents another exciting opportunity. These advanced architectures can enhance the spatial and temporal resolution of synthetic data, capturing intricate patterns and variability in solar radiation. By improving the fidelity and realism of synthetic datasets, such advancements can further boost the accuracy of predictive models.

Lastly, exploring hybrid predictive models, such as Transformer-based architectures combined with GAN-generated data, offers the potential to enhance model scalability and accuracy. These hybrid approaches could capture both spatial and temporal dependencies more effectively, improving predictive performance across diverse climatic conditions.

By pursuing these focused advancements, the GAN-based framework can evolve into a comprehensive and adaptable solution for solar radiation prediction, addressing both regional and global challenges in sustainable energy development.

Limitations and future work

While the proposed framework shows promise, several challenges and opportunities for improvement remain:

• Region-specific variability: Capturing the nuanced solar radiation patterns of diverse climates requires further refinement.

• Data quality and availability: Limited high-resolution data in under-monitored regions restricts model generalization.

• Training stability: Issues like mode collapse necessitate more robust training strategies.

• Evaluation metrics: Developing specialized metrics tailored to solar radiation data is critical for accurate assessment.

• Architectural flexibility: Future work could explore dynamic layer scaling to adapt to input complexity.

While the framework demonstrates promising results, it has certain limitations. Generating high-quality synthetic data demands significant computational resources and careful hyperparameter tuning to mitigate issues such as mode collapse. Despite our regional adaptations, abrupt weather changes and extreme outliers may still challenge model robustness. Future research should explore ensemble GAN architectures and adaptive learning mechanisms to further enhance performance under highly dynamic conditions.

Addressing these limitations will advance the framework into a comprehensive solution for solar radiation forecasting, supporting global renewable energy goals.

Broader impact

Aligned with Saudi Arabia’s Vision 2030, this research contributes to sustainable energy development and innovation. By improving solar radiation forecasting, the framework lays the groundwork for smarter, greener energy systems, reinforcing the transition toward a sustainable future.