Weather–power mapping inference for wind power forecasting across heterogeneous sites

Abstract Accurate wind power forecasting is critical for increasingly weather-sensitive power systems. Forecast accuracy depends on the meteorology–power mapping, which varies across farms and evolves with operating conditions. Existing approaches either fit site-specific mappings from limited local data, making them vulnerable to poorly represented operating conditions, or learn a shared mapping across farms and risk blurring local response differences. The challenge is to exploit cross-farm data while retaining mappings that adapt to local and evolving conditions. Here we develop UniWind, which infers target-specific mappings from a shared cross-farm prior rather than relearning each farm’s complete mapping. Using data from 576 wind farms spanning all nine wind-energy environments in China, UniWind learns a shared weather–power prior to reduce reliance on conditions represented at any single farm, and uses target-farm observations to infer local deviations, adapting this prior to the current site and operating regime. On the held-out test sets of these farms, UniWind achieves 82.98% mean accuracy, outperforming the strongest baseline by 1.28 percentage points. It further achieves 84.05% accuracy on 142 farms excluded from training. UniWind also improves forecasts of rare ramp and low-power events, reduces estimated upward-reserve requirements by 5.0 MW per 100 MW at 95% reliability, and cuts fleet-wide model-update computation by 96.7%.