Evaluating the incremental value of forward-looking ERA5 data for short-term wind power forecasting: a multi-model, multi-horizon study across two proximate wind farms

Abstract Short-term wind power forecasting relies on two types of information: the recent behavior of the turbines and the expected evolution of the atmosphere. Because the changing balance between these sources across forecast horizons is rarely quantified, studies reporting an overall benefit from meteorological data do not establish when this benefit emerges or what it replaces. This study measures this transition for two 48 MW wind farms in central Vietnam, using one year of SCADA and ERA5 data at 15-minute intervals, four forecasting models, and one-, two-, and four-hour horizons. Using grouped permutation importance, the share attributable to recent production history decreases from 71.2 % to 1.7 % for farm A and from 75.8 % to 8.9 % for farm B between one and 4 h, while the share associated with forward-looking meteorological information rises to 65.3 % and 42.5 %, respectively. For XGBoost, the recommended model, this shift occurs on a finer 30-minute grid at 1.86 h for farm A and 2.76 h for farm B; accuracy gains follow the same trend but are more modest, becoming significant only at 4 h, where forward-looking information reduces RMSE by 0.86 and 0.80 % points of rated power, respectively, using a paired moving-block bootstrap. Random Forest also reaches significance at shorter horizons but consistently has higher RMSE than XGBoost. This gain is offset by wind-speed forecast errors above approximately 1.2 m s −1 ; when sustained, it corresponds to a reserve saving of approximately 1.2 MW per 48 MW farm.