SDF-Net: Solar-aware Decoupled Fusion Network for intra-hour forecasting of photovoltaic power

Accurate intra-hour PV power forecasting is critical for efficient operations of power grids. Many methods implement ground-based sky imagery to improve forecasting performance, but often underutilize image features and neglect sun-relevant regions, leading to degraded forecasting accuracy. To resolve such issues, we propose SDF-Net, a novel Solar-aware Decoupled Fusion Network. SDF-Net employs a decoupled fusion architecture to enhance feature utilization. Specifically it implements a Vision Transformer to extract global and local image features from historical sky image sequence and a Transformer to extract power features from historical PV power sequence. Then the global and local image features independently interact with power features in a decoupled manner, enabling efficient utilization of image features and enhancing all three features. Meanwhile, a solar-guided multi-head attention mechanism and a solar-aware cross-attention mechanism are designed to precisely guide SDF-Net’s attention to sun-relevant regions during feature extraction and cross-modal interaction, respectively. Finally, a triple-LSTM decoder with adaptive gating takes the three enhanced features to generate the final PV power prediction. Extensive experiments on the SKIPP’D dataset demonstrate SDF-Net’s superior performance across various forecasting horizons and weather conditions.