UniWind Improves Day-Ahead Wind Power Forecasting with Physics-Informed AI

Accurate day-ahead wind power forecasting is crucial for efficient power system operations, yet it's complicated by the interplay of meteorological conditions, temporal patterns, and unobserved operational states like shutdowns or curtailment. Traditional physical models struggle with adaptability across different wind farms, while purely data-driven models often conflate meteorological effects with state-induced deviations. This research introduces UniWind, a new model designed to address these limitations. UniWind employs a physics-informed state routing mechanism to achieve unified and robust forecasting. It begins with a Physical Prior Estimator that creates a site-calibrated physical prior, combining site-conditioned monotonic warping with a shared physical power curve. This prior is then constrained by a physical upper-bound to act as a soft envelope for available wind power. The model further incorporates a Latent State Encoder to model operational state embeddings. These embeddings are used by a State-aware Power Corrector, which leverages knowledge-guided supervised state routing and bounded, state-specific expert correction to transform the physical prior into the final power forecasts. Extensive experiments on over 20 real-world datasets, including full-shot and cross-farm zero-shot scenarios, demonstrate UniWind's superior accuracy and robustness compared to existing forecasting methods.