Explainable hourly global solar radiation forecasting using a CNN-BiLSTM model with temperature scaled softmax attention.

Abstract

Accurate global solar radiation (GSR) forecasting is vital for smart grids and resilient energy systems. However, the nonlinear and non-stationary nature of meteorological drivers challenges conventional approaches. This study proposes a lightweight, explainable hybrid deep learning architecture, CNN-BiLSTM-STAM, which integrates convolutional layers for inter-feature pattern extraction, bidirectional long short-term memory (BiLSTM) networks for temporal dependency learning, and a softmax temperature attention mechanism (STAM) with a learnable temperature to adaptively sharpen or smooth attention over time under varying weather conditions. The model is trained and evaluated using an hourly multivariate meteorological dataset collected at Alparslan Türkeş Science and Technology University in Sarıçam, Adana, Türkiye. Results show that CNN-BiLSTM-STAM outperforms classical machine-learning models, baseline deep networks, and recent attention-based variants for short and mid-term GSR forecasting, achieving RMSE = 66.62 W/m[Formula: see text], MAE = 43.29 W/m[Formula: see text], [Formula: see text] = 0.9381, and Pearson r = 0.9715. Additional validation on two independent public meteorological datasets, with the same architecture trained and evaluated separately on each dataset, yielded [Formula: see text] values of 0.9188 and 0.9537 and Pearson correlation coefficients of 0.9597 and 0.9768, supporting the external reproducibility of the proposed framework. To support operational trust, Shapley additive explanations (SHAP) are used to quantify feature importance and provide interpretable insights into seasonal and short-term atmospheric influences. With a compact footprint and strong predictive accuracy, the proposed framework supports deployable GSR forecasting for PV operation and scheduling, real-time energy management, and IoT-enabled monitoring systems.