A deep hybrid CNN–BiLSTM–attention framework with metaheuristic optimization and uncertainty quantification for solar power forecasting in dust-affected arid climates

This research presents an advanced hybrid deep learning framework that combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTMs), and an attention mechanism, enhanced by Particle Swarm Optimization (PSO) and Gaussian Processes (GP) for uncertainty evaluation. The model tackles key forecasting challenges caused by dust buildup on solar panels, especially in dry regions, which reduces solar irradiance transmission and increases prediction uncertainties. The framework was tested using a large dataset of 49 million observations from 2018 to 2024 across six Gulf Cooperation Council (GCC) countries. This data included various meteorological variables, air quality measures (such as PM 10 and PM 2 . 5 levels), and operational information from photovoltaic (PV) systems. The model showed excellent predictive performance, with an R 2 of 98.7%, a root mean square error (RMSE) of 73.9 kW, and a mean absolute percentage error (MAPE) of 4.4.1% on 5 MW PV plants. Compared to traditional methods, the proposed model outperformed others by 52 points: it surpassed the persistence model. 8%, ARIMA by 42.5%, and the basic LSTM by 42.4%. Its robustness was especially evident during extreme dust events, where PM 1 0 levels exceeded 500 μg/m 3 , yet it still achieved a 56.0% improvement over conventional approaches. The attention mechanism adapts by focusing on a small set of features early during dust events but spreads attention during peak periods, assigning up to 85.3% importance to key meteorological changes. The probabilistic forecast capability showed a 95.3% prediction interval coverage probability (PICP) and a narrow interval width (PINAW of 0.232), validated through GP- based uncertainty quantification. Including PM metrics enhanced performance by 97.0%, with attention mechanisms providing an additional 18.1% gain. Economic analysis indicates a 22%–25% reduction in PV operational costs across GCC countries. Moreover, dust cleaning operations based on accurate forecasts could recover 97.2.2–98.1% of energy lost due to dust while cutting water use by 20%–25%. Overall, the hybrid model effectively addresses the complex forecasting problems faced by dust-affected photovoltaic power systems.