Solar Radiation Forecasting with Hybrid Deep Learning Framework Integrating Feature Factorization

Deep learning models play a vital role in high-precision solar radiation prediction, leveraging their exceptional capabilities in capturing intricate relationships between input and output parameters. Due to the temporal dependence of solar radiation sequences and complexity of multidimensional variations, however, conventional deep learning models often face a trade-off between complex structures, gradient explosions, and accuracy. Furthermore, most deep learning-based solar radiation prediction models are considered black-box models, complicating efforts to adapt or modify the models based on theoretical considerations. In this work, a prior knowledge-based feature factorization layer that effectively addresses the temporal dependencies in solar radiation prediction problems is proposed. It is embedded into the CNN-LSTM-SA framework to extract spatiotemporal features from the non-stationary and nonlinear time series. The suggested model is learned and validated through weather data gathered in Tokyo and is contrasted with four traditional machine learning methods (linear regression, support vector machine, random forest, and back propagation neural network) and typical combined deep learning models. It demonstrates superior performance in both accuracy and interpretability with R2 of 0.93.