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

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence
• Meteorological Forecasting

Background:

• Accurate global solar radiation (GSR) forecasting is essential for the integration of renewable energy into smart grids.
• The inherent nonlinear and non-stationary characteristics of meteorological data pose significant challenges to traditional forecasting methods.
• Existing deep learning models often lack interpretability, hindering operational trust and deployment.

Purpose of the Study:

• To develop a novel, lightweight, and explainable hybrid deep learning architecture for enhanced GSR forecasting.
• To improve the accuracy and reliability of short and mid-term GSR predictions under diverse weather conditions.
• To provide interpretable insights into the factors influencing GSR variability for better energy management.

Main Methods:

• A hybrid deep learning model, CNN-BiLSTM-STAM, integrating convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM) networks, and a softmax temperature attention mechanism (STAM) was proposed.
• The model was trained and validated on hourly multivariate meteorological data from Türkiye, with further evaluation on two independent public datasets.
• Shapley additive explanations (SHAP) were employed to analyze feature importance and provide model interpretability.

Main Results:

• The CNN-BiLSTM-STAM model demonstrated superior performance compared to classical machine learning, baseline deep networks, and attention-based variants.
• Achieved high accuracy with RMSE = 66.62 W/m², MAE = 43.29 W/m², R² = 0.9381, and Pearson r = 0.9715 on the primary dataset.
• External validation on independent datasets confirmed the model's robustness and reproducibility, yielding R² values of 0.9188 and 0.9537.

Conclusions:

• The proposed CNN-BiLSTM-STAM framework offers a lightweight, explainable, and highly accurate solution for GSR forecasting.
• The model's ability to adapt attention mechanisms enhances its performance across varying meteorological conditions.
• The findings support the deployment of this framework for practical applications in PV operation, energy management, and IoT monitoring systems.