Abstract

With global attention shifting toward clean and renewable energy, accurate solar radiation prediction is critical for optimizing solar energy systems. This project introduces a hybrid deep learning model that combines MobileNet for spatial image feature extraction with BiLSTM for sequential analysis, enhanced by a Multi-Head Attention mechanism for capturing temporal dependencies. The model achieved strong performance with MAE of 0.2355, MSE of 0.1159, and RMSE of 0.3405. Explainable AI tools such as SHAP, Grad-CAM, and LIME were integrated to provide interpretability and build user trust. The outcome offers an innovative and reliable approach for intelligent solar energy forecasting.

Key Contributions

• Proposed a lightweight yet robust MobileNet-BiLSTM hybrid architecture for time series solar radiation prediction.
• Integrated Multi-Head Attention to improve the model's ability to focus on important sequence features.
• Incorporated XAI tools for interpretability, making predictions more transparent and trustworthy.
• Evaluated with comprehensive error metrics—MAE, MSE, RMSE—demonstrating state-of-the-art results.

Conclusion

The study showcases how integrating spatial and temporal deep learning models with attention mechanisms can significantly improve the accuracy of solar radiation prediction. This approach provides a practical tool for optimizing solar panel operations and energy forecasting in real-world deployments.

Limitation

The model's accuracy is influenced by weather variability and camera sensitivity. Additionally, real-time prediction at scale would require further optimization and adaptation to diverse geographic datasets.

Future Work

Future developments include expanding the dataset to cover multi-regional solar fields, integrating real-time weather data, and deploying the model on edge devices for instant solar output prediction. Exploring transformer-based architectures for even better time series performance is also recommended.