Operating-State-Constrained Weather-Power Fusion Network for Regional Wind Farm Cluster Forecasting

Abstract

Intraday regional wind farm cluster power forecasting is an important basis for power dispatch, reserve allocation, and renewable energy integration. Weather-driven forecasting models usually rely on future weather forecasts to describe wind-field evolution during the forecast day. However, under the same weather situation, actual power output is also affected by the station-level generation level, short-term inertia, and synchronized multi-site variations before the forecast origin; if a model only focuses on future weather, it can produce deviations inconsistent with the true operating state. This paper proposes an operating-state-constrained weather-power fusion network. The method extracts weather-driven representations from meteorological forecasts, encodes station-level measured power available before the forecast origin as an operating-state representation, and uses this representation to constrain weather-information reading during decoding. Experiments on three years of data from 58 wind farms in Jilin Province show that the proposed method achieves a normalized MAE of 0.0779, a normalized RMSE of 0.1065, and an R2 of 0.8036 for intraday aggregate power forecasting. Compared with Weather-only, the MAE and RMSE are reduced by 6.84% and 6.57%, respectively; both errors are reduced by approximately 5% relative to the fusion baselines. Horizon-wise analysis shows that the gain is most pronounced at the beginning of the forecast day, with first-horizon MAE reduced by 47.94%. These results indicate that true power states before the forecast origin can effectively constrain the weather-to-power mapping and improve state adaptability in regional wind farm cluster forecasting.