Lite-BiAttn: A Lightweight CNN-BiLSTM Attention Model for Accurate Solar Radiation Forecasting

Reliable short-term prediction of global horizontal irradiance (GHI) is essential for maximizing photovoltaic (PV) power generation, improving grid reliability, and enabling efficient integration of renewable energy sources. This study proposes Lite-BiAttn, a lightweight hybrid deep learning framework that combines one-dimensional convolutional neural networks (1D CNN), a single-layer Bidirectional Long Short-Term Memory (BiLSTM), and a lightweight dot-product attention mechanism for accurate GHI forecasting. Hourly meteorological observations were obtained from the NASA POWER database for Zurich, Switzerland. Zurich was selected due to its seasonal climatic variability, solar-resource potential, and relevance to PV applications, as well as its prominence as a Swiss tourism destination, with the Zurich region recording a record 7.56 million overnight stays in 2025. To enhance temporal representation, a 24 h sliding window together with cyclical hour and month encoding was adopted, enabling the model to capture both short-term sequential patterns and periodic seasonal behavior. Comprehensive experiments demonstrate that the proposed framework consistently outperforms conventional machine learning and deep learning approaches, achieving an RMSE of 0.052, MAE of 0.036, MAPE of 0.071, and an R2 of 0.978. The framework provides accurate and efficient short-term GHI forecasting for Zurich. Further multi-location validation is needed to assess its broader applicability and edge deployment potential.