基于扩散模型卫星图像预测的多模态集成光伏发电功率预测

Accurate ultra-short-term photovoltaic (PV) power forecasting is crucial for grid stability and energy management, but it is often challenged by the intermittent and unpredictable nature of cloud cover. Since satellite images provide large-scale cloud information, most studies incorporate them by predicting cloud evolution using deterministic models. However, these models often produce blurry forecasts, failing to capture rapid cloud changes and leading to unreliable power predictions. To address this limitation, CloudDiff, a novel diffusion-based satellite image prediction model, is proposed to generate ensemble satellite image forecasts, which could better account for PV power forecasting uncertainty caused by short-term cloud variations. Spatiotemporal cloud uncertainty is modeled recurrently by conditioning the diffusion process on encoded cloud motion features, while a neural stochastic differential equation captures temporal variability in cloud dynamics. These image forecasts, combined with historical PV power and numerical weather prediction data, are fed into a multimodal forecasting model to produce calibrated ensemble PV power forecasts. Extensive experiments across multiple sites and regions demonstrate that the proposed method outperforms existing approaches in probabilistic forecasting of both satellite imagery and PV power at a 180-minute horizon, exhibiting strong robustness under diverse cloud conditions.