Attention-Enhanced CNN–BiLSTM Framework with Temporal Feature Engineering for Global Horizontal Irradiance Forecasting

Abstract

Global Horizontal Irradiance (GHI) forecasting is crucial for optimizing PVs, stabilizing smart grids, and integrating renewable energy resources. This study introduces the following novel forecasting framework: temporal feature engineering, attention-enhanced deep sequence learning, explainable AI analysis, and robustness evaluation for short-term forecasting of GHI in arid regions. To enhance temporal representation learning, the sliding-window sequence construction method and cyclical temporal encoding were added to represent short-term temporal dependencies and the periodic variations in solar irradiance. The NASA POWER database was used to provide hourly meteorological and irradiance data to develop and evaluate the model at Jubail, Saudi Arabia. The proposed CNN–BiLSTM–Attention framework performed better than conventional benchmark models of machine learning and deep learning, achieving a Root Mean Square Error (RMSE) of 0.088, Mean Absolute Error (MAE) of 0.059, Mean Absolute Percentage Error (MAPE) of 13.6%, and a coefficient of determination (R2) of 0.912 in comparative experiments. These results were also validated by ablation analysis, indicating the role of convolutional feature extraction, bidirectional temporal learning, and attention-based temporal weighting in predictive performance. Seasonal and atmospheric-condition evaluations also showed consistent performance across different seasonal and atmospheric conditions.