Probabilistic wind power forecasting with a conditional spatio-temporal mixture-of-experts diffusion model

The rapid expansion of wind energy necessitates highly accurate ultra-short-term power prediction techniques to address integration and dispatch challenges caused by wind volatility. While point forecasting provides a single estimate, probabilistic forecasting offers a more comprehensive view by capturing the distribution of future outputs, enabling risk-aware decision-making. However, direct modeling of such distributions remains challenging due to complex spatio-temporal dependencies and high network optimization burdens. To address these issues, this study proposes the spatio-temporal mixture-of-experts diffusion (MOE-STD) framework for wind power probabilistic forecasting. The framework introduces a Channel Attention Enhancement Module (CAEM) to identify key dependencies between turbine output and influencing factors. A Mixture of Spatial Experts (MOSE) dynamically learns both simple and complex spatial correlations among wind farms, while a Mixture of Temporal Experts (MOTE) captures continuous trends and sudden fluctuations in time-series data. These modules are embedded within a diffusion framework, which learns to reconstruct future output distributions via multi-step prediction, rather than direct forecasting. Validated on real-world wind farm datasets, the proposed model consistently outperforms both point and probabilistic forecasting methods in terms of accuracy and reliability, demonstrating a superior ability to capture uncertainty in ultra-short-term wind power forecasting.