Change-point-driven adaptive deep learning framework for extreme wind event prediction under non-stationary atmospheric conditions

Abstract

Extreme wind prediction remains challenging because wind dynamics are nonlinear, non-stationary, and frequently affected by abrupt regime transitions. This study proposes a change-point-driven adaptive deep learning framework that detects structural changes in multivariate wind data, segments the series into locally coherent regimes, and updates a lightweight predictor for post-change forecasting. The framework jointly performs multi-step wind-speed prediction and extreme-event detection. Experiments on the NREL WIND Toolkit and a controlled synthetic dataset show that the proposed method achieves the lowest RMSE of 1.86, the highest TSS of 0.81, the lowest FAR of 0.15, and the best Brier Score of 0.131. It also provides the lowest post-change RMSE, indicating faster recovery following regime transitions. Although the Transformer performs slightly better in nRMSE, MAE, CSI, and ETS, the proposed framework demonstrates stronger robustness to large forecasting errors, false alarms, and abrupt structural changes under non-stationary wind conditions.