Multi‐Model Ensemble Knowledge Distillation With Physics Constraints for Numerical Weather Prediction Bias Correction

Abstract Accurate Numerical Weather Prediction (NWP) is of paramount importance for global climate change response and sustainable development. Although numerical models such as the Weather Research and Forecasting (WRF) model are widely applied in operational forecasting, they exhibit significant systematic biases under complex atmospheric conditions, which limits the reliability of refined meteorological services. Existing bias correction methods suffer from insufficient model diversity, low computational efficiency, and lack of physical consistency, making them inadequate to meet current demands. This study develops a multi‐model ensemble knowledge distillation framework with physics constraints to systematically improve WRF bias correction. Based on long‐term observational data from Chinese ground meteorological stations and WRF simulation data, a rigorous validation system with spatial, temporal, and feature isolation was employed for training and evaluation. Results demonstrate that the teacher ensemble model significantly enhances prediction accuracy across temperature, humidity, and wind speed variables. Through knowledge distillation, the lightweight student model achieves substantial speed improvements and dramatic memory reduction while further improving prediction accuracy beyond the teacher model performance. The framework exhibits excellent generalization capability across different terrains and temporal conditions, maintains physical consistency through conservation constraint mechanisms, and achieves effective extreme value control. This research provides a computationally efficient and physically consistent solution for WRF bias correction applications, achieving the unification of model diversity, computational efficiency, and physical rationality in intelligent NWP system, which holds significant value for refined meteorological services.