A Hybrid Dual Stream ProbSparse Self-Attention Network for spatial–temporal photovoltaic power forecasting

Abstract

Growing energy demand and increasing environmental challenges underscore the importance of precise forecasts for photovoltaic (PV) operations in renewable energy generation systems. At this stage, it is mainstream to combine both temporal and spatial factors to forecast PV power generation. However, there are fewer studies that consider factors at very large spatial scales. This paper proposes Hybrid Dual Stream ProbSparse Self-Attention Network (HDSPAN), a novel spatial–temporal photovoltaic power forecasting network architecture that can solve the above limitations. The model implements an encoder–decoder approach that extracts the required spatial–temporal information through a dual stream distilling mechanism. In addition, the ProbSparse self-attention mechanism is employed to improve model efficiency and reduce repetitive and redundant information processing. The hyperparameters are optimized using Tree-structured Patzen estimator to improve forecasting outcomes. This paper demonstrates the effectiveness of spatial–temporal PV forecasting by using ERA5 reanalyzed PV data as a case study. Our results show that the HDSPAN model achieves a 10% higher forecasting accuracy compared to the baseline models, significantly advancing PV power forecasting.