A Distributed Photovoltaic Power Forecasting Method Assisted by Satellite Cloud Imagery

Abstract

Owing to their geographical dispersion and high deployment costs, distributed photovoltaic (DPV) stations often lack access to high-precision meteorological data, which degrades power forecasting accuracy and threatens grid operational stability. Accordingly, developing a low-cost ultra-short-term power forecasting model is critical for power system dispatching. This study proposes a site-level ultra-short-term power forecasting method incorporating satellite cloud imagery (SCI) for DPV systems. First, three core forecasting challenges are analyzed: cross-modal data correlation establishment, spatial alignment between heterogeneous data structures, and spatiotemporal correlation extraction among dispersed stations. Second, a forecasting model taking satellite imagery and multi-station power outputs as inputs is constructed. It aligns multimodal data via a data embedding layer, and integrates multi-source features through cuboid self-attention modules and an encoder–decoder architecture. Third, a Multi-Scale Correlation Mechanism (MSCM) is proposed to capture spatiotemporal associations across geographically dispersed stations. Experimental results based on a real-world dataset from Hebei Province show strong performance. The full model with satellite cloud imagery achieves a Root Mean Square Error (RMSE) of 0.112 and a Mean Absolute Error (MAE) of 0.059, outperforming baseline models including the backpropagation (BP) neural network, long short-term memory (LSTM) network, and Transformer model.