Getting Uncertainty Right: Local Calibration for Wind Power Forecasting

Abstract

Wind energy forecasting has increasingly shifted from point to probabilistic approaches to support risk-aware decision-making. However, most existing methods evaluate models using global calibration metrics, which fail to capture reliability in real-time operations. In this paper, we emphasize the importance of local calibration for trustworthy decision support in wind power generation. We propose a set of local calibration metrics to assess probabilistic forecasts at a finer temporal scale. Furthermore, we incorporate the Adaptive Conformal Inference (ACI) framework as a model-agnostic, post-hoc approach to improve calibration. Extensive experiments on real-world wind power datasets using deep probabilistic forecasting models show that ACI consistently enhances local calibration performance, with improvements of up to 30% in the proposed metrics. These results highlight the significance of local calibration and demonstrate the effectiveness of ACI in improving the reliability of probabilistic forecasts for real-time, risk-aware decision-making.