A novel wind power interval prediction method based on neural ensemble search and dynamic conformalized quantile regression

Abstract

Wind power interval prediction plays a crucial role in delivering accurate estimations of the potential range of wind power generation, enhancing the stability and reliability of the power system. In this study, a novel method named Conformalized Quantile Regression with Neural Ensemble Search (NESCQR) for wind power interval prediction is proposed. The NESCQR algorithm combines Neural Ensemble Search (NES) and dynamic conformalized quantile regression. The NES employs a forward selection strategy to identify an optimal subset of models, aiming to minimize prediction errors and consequently produce tighter prediction intervals (PIs). Meanwhile, the dynamic conformalization process allows the model to effectively adapt to temporal variations in the data, significantly improving its robustness. Experiments on four real datasets show that the proposed NESCQR algorithm can obtain extremely narrow prediction intervals while ensuring valid coverage rate, and effectively alleviate the quantile crossing phenomenon, providing reliable and effective help for decision-makers.