Interpretable deep learning framework for hourly solar radiation forecasting based on decomposing multi-scale variations

High-precision solar radiation forecasting is crucial for the economic and reliable operation of building energy systems and contributes to achieving sustainability goals. However, owing to multi-scale variations and instability of solar radiation, most methods face a zero-sum game of trade-offs between simplicity, reliability, and interpretability. In this study, we propose a reliable and interpretable deep learning framework by deconstructing the multi-scale variations of solar radiation. It integrates feature engineering using transformation matrices, Convolutional Neural Networks classification for local feature recognition, and Long Short-Term Memory prediction for global uncertainty learning. First, the temporal variability of solar radiation is transformed into a simple mathematical description using transformation matrices. Then, Convolutional Neural Network associate local features with embedded skewed labels of solar radiation, resulting in a robust output in the first stage. Finally, Long Short-Term Memory's global feature mining adapts to the instability to obtain the final output. The proposed model was trained and tested using meteorological data collected in Tokyo from January 1, 2000, to December 31, 2021. The results indicate that the proposed model can achieve high-precision predictions by leveraging historical data correlation and inference demonstrated by its high performance on untrained datasets with an R2 of 0.97. Additionally, it is addressing a part of solar radiation time series forecasting problems that previously required black-box models, by using reliable mathematical descriptions and expert knowledge. This enhances the interpretability of the deep learning framework through the ante-hoc design of the model. This study provides valuable insights into the precise geographical expansion of solar radiation datasets and practical adjustment strategies for building energy systems. © 2024 Elsevier Ltd