A novel physics-inspired method for efficient and robust ultra-short-term photovoltaic power forecasting

Accurate and efficient ultra-short-term photovoltaic power forecasting is crucial for ensuring grid stability under high renewable energy penetration. However, existing multimodal learning methods based on sky images suffer from two key limitations: computational inefficiency due to the dense encoding of redundant image sequences, and a “black-box” paradigm lacking clear physical guidance and robustness. This paper proposes a novel, physics-inspired end-to-end prediction framework to address these challenges. The framework's core is a “dynamic-guided static” decoupled spatio-temporal encoder. It introduces a physics-inspired Multi-scale Cloud Morphology Decomposer (MCMD) with a parameter-free decomposition stage to explicitly quantify the physical dynamics of cloud evolution from historical sequences. This dynamic information then guides a visual attention module to focus on key predictive regions in the latest image, significantly enhancing feature representation while reducing computational complexity. Furthermore, an innovative Prototypical Contrastive Alignment (ProCA) loss is introduced to enhance robustness by aligning multimodal features at a conceptual level, effectively resisting single-modal data noise. Extensive experiments on the public dataset demonstrate the model's state-of-the-art (SOTA) performance. Qualitatively, it exhibits exceptional generalization ability and stability, particularly under challenging overcast conditions. Quantitatively, this superiority is confirmed as it achieves a minimum root mean square error (RMSE) of 2.4874 kW with only 0.30 M model parameters, striking an optimal balance between accuracy and efficiency.