Reserve-risk-oriented regime-calibrated interval forecasting for wind power integration under extreme fluctuations

Abstract

With increasing wind penetration, reliable prediction intervals are essential for reserve preparation and risk-aware operation. However, satisfactory aggregate coverage may conceal substantial under-coverage during ramp and high-volatility conditions. This study proposes a regime-calibrated interval forecasting framework that integrates Lightweight Temporal Feature Adaptation, extreme-regime-aware quantile learning, and horizon- and regime-dependent conformal recalibration. An author-pretrained iTransformer is used as the temporal backbone, whose parameters are frozen while lightweight adapters and a quantile forecasting head are optimized for the target task. The framework is evaluated on two source wind farms with installed capacities of 99 MW and 200 MW and independently tested on an unseen 240 MW high-altitude complex-terrain wind farm. On the source farms, the proposed method achieves prediction interval coverage probabilities of 0.903 and 0.901 at the 90 % nominal level, with normalized interval widths of 0.144 and 0.150. Compared with standard conformalized quantile regression, interval width is reduced by 23.8 % and 26.8 %, respectively. The reserve-relevant shortfall rate is also reduced to 0.074 and 0.081. The results demonstrate that the proposed framework improves interval reliability, sharpness, and cross-site robustness under operationally critical wind power fluctuations.