The Lightweight Hybrid Deep Learning Approach for Capturing Long-Term and Short-Term Constraints for an Accurate Solar Radiation Forecast

Abstract

Accurate solar radiation forecasting is essential for photovoltaic energy generation, smart grid stability, and renewable energy management. This study proposes a lightweight hybrid deep learning framework that combines a transformer encoder and Gated Rrecurrent Uunit (GRU) network for short-term solar radiation forecasting in Makkah and Madinah, Saudi Arabia. Hourly meteorological data from the NASA POWER dataset (2020–2025) were utilized, including solar radiation intensity, temperature, humidity, wind speed, cloud amount, rainfall, surface pressure, and dew point temperature. A preprocessing pipeline consisting of missing value treatment, outlier removal, normalization, timestamp alignment, and data cleaning was applied to improve data quality. Feature engineering techniques were incorporated to capture temporal dependency, meteorological interactions, weather dynamics, and solar variability patterns. The transformer encoder was used to learn long-range temporal dependencies through multi-head self-attention, while the GRU layer modeled sequential temporal dynamics efficiently. Hyperparameter optimization was performed using Bayesian optimization with Optuna. The experimental results demonstrate that the proposed transformer GRU framework achieved a Mean Absolute Error (MAE) of 0.014, Root Mean Square Error (RMSE) of 0.0219, and a coefficient of determination (R2) of 0.98. The proposed model outperformed ARIMA, LSTM, GRU, and XGBoost models while maintaining stable performance across varying weather conditions and forecasting horizons.