State-Adaptive Forecast Error Recalibration for probabilistic wind power forecasting

Reliable short-term wind-power forecasting requires uncertainty estimates that remain useful as operating conditions change. This study develops State-Adaptive Forecast Error Recalibration, a state-conditioned conformal post-processing framework for renewable-energy forecasting. Hourly Eskom wind generation from April 2019 to February 2024 is combined with regional meteorological predictors from the fifth-generation European Centre for Medium-Range Weather Forecasts reanalysis. A leakage-controlled chronological design assigns 55% of the model-ready sample to training, 10% to validation, 20% to conformal calibration and 15% to one untouched test evaluation. Validation selects linear regression at 1, 3, 6 and 24 h and CatBoost at 12 h. Untouched-test root mean squared errors are 107.23, 231.05, 342.20, 554.12 and 553.64 megawatts across the five horizons. At 90% nominal coverage, state conditioning raises empirical coverage from 0.8869 to 0.8984 at 3 h, from 0.8770 to 0.8869 at 6 h and from 0.8014 to 0.8140 at 12 h. Dependence-aware moving-block bootstrap inference confirms significant coverage differences at 1, 3, 6 and 12 h; however, the 1-hour increase does not reduce absolute coverage error. A significant Winkler-score improvement is observed only at 12 h. All 48 calibration states satisfy the minimum sample-size rule and no test forecast requires fallback. The framework can complement statistical and artificial-intelligence energy forecasters, while sequential adaptive conformal inference remains closer to nominal coverage at longer horizons.