Advanced transformer network for solar irradiance prediction in data-scarce desert environments

Solar radiation forecasting remains a major challenge in desert regions, where dust, high temperatures, and rapidly changing atmospheric conditions a significant impact on energy availability. In this study, we investigate the use of a Transformer-based model incorporating Multi-Head Self-Attention and Layer Normalization to better capture the complex temporal dynamics of solar radiation. Compared with traditional recurrent models such as LSTMs, the Transformer can represent long-term dependencies in time series more effectively, making it particularly suitable for climatic prediction tasks. The model was trained and evaluated using solar radiation data from eight cities across the Algerian desert, each representing a different arid subregion. Its performance was benchmarked against several baseline models using rRMSE, MAPE, and R² metrics. The Transformer consistently outperformed the baselines, achieving the best accuracy in Illizi (rRMSE = ۹.۵۴%, MAPE = ۸.۰۹%, R² = ۰.۸۷۴۷). However, the findings indicated significant variability among locations, demonstrating that, although Transformer models have great promise for solar forecasting in arid zones, their reliability remains closely linked to local atmospheric characteristics. This highlights the importance of site-specific calibration to ensure robust and operationally useful predictions.

کلیدواژه ها:

Transformer Model ، Solar radiation ، prediction ، Machine-Learning ، Desert Climate