Explainable uncertainty quantification in probabilistic wind power forecasting

Accurate wind power forecasting and reliable confidence intervals are essential for renewable energy integration and risk-aware grid operation. Forecast uncertainty arises from complex environmental conditions and turbine dynamics. However, conventional data-driven approaches often treat uncertainty as a black box, which limits operational transparency and fails to identify the physical sources of forecast variance. This study presents an explainable uncertainty quantification framework that combines physics-informed modeling, probabilistic deep learning, and sensitivity-based interpretability analysis. An operation-based turbine dynamics model derived from real-world offshore supervisory control and data acquisition (SCADA) measurements is embedded within a residual learning architecture to incorporate turbine operational characteristics into the forecasting process. Heteroscedastic Gaussian neural networks and quantile regression are employed to characterize both symmetric uncertainty and asymmetric operational limits. Unlike conventional probabilistic forecasting approaches, the hybrid framework explicitly associates uncertainty behavior with turbine operating states through sensitivity-driven analysis. The best performing hybrid framework achieves a relative root mean square error (rRMSE) of 3.3386% and a coefficient of determination of 0.9505 in deterministic forecasting. For probabilistic forecasting, the framework attains a normalized continuous ranked probability score (nCRPS) of 0.0057 ± 0.0098 . A perturbation-based sensitivity analysis further identifies how forecast uncertainty varies across different power production regimes and turbine control states. The results demonstrate that the hybrid framework improves forecast interpretability while maintaining high predictive accuracy, which supports real-time operational decision-making, dynamic turbine control, and intra-hour grid balancing in modern wind farms.