A hybrid day-ahead wind speed prediction model based on refined WRF and attention-LSTM for coastal complex terrain

Abstract Accurate day-ahead wind speed forecasting remains challenging over complex coastal terrain due to strong land–sea thermal contrasts, terrain-induced flow variability, and rapidly changing atmospheric conditions. This study investigates the characteristics of Weather Research and Forecasting (WRF) model wind speed forecast errors and proposes a coordinated error-correction approach to improve prediction reliability. Forecast errors are decomposed into trend and fluctuation components. A key methodological contribution is the HC–DTW-guided trend similarity identification, which enables conditional reconstruction of the fluctuation component based on analogous historical patterns. These components are jointly modeled to capture their distinct temporal behaviors and interactions under varying meteorological conditions. Results show that the proposed approach effectively reduces forecast errors, with the mean absolute error (MAE) decreased by 29.0% and the root mean square error (RMSE) reduced from 1.61 m/s to 1.26 m/s. Further analysis reveals that the coordinated strategy improves the representation of nonstationary error structures linked to evolving atmospheric states. Cross-seasonal validation demonstrates robust performance under different meteorological regimes. The findings highlight the importance of multi-scale error characterization and coordinated modeling in capturing the dynamics of wind speed forecast errors over coastal terrain, providing both improved prediction performance and insights into the underlying atmospheric processes governing forecast uncertainty.