Abstract Meso- and micro-scale weather processes—such as cold air pools (CAPs) and nocturnal warming (NW)—exert significant influence on near-surface air temperatures in mountainous regions but are often overlooked in conventional statistical downscaling methods. This study develops an interpretable XGBoost-based framework that explicitly incorporates these processes to improve hourly temperature downscaling of ERA5 reanalysis data. Using observational data from 16 meteorological stations across ~20 km2 in the Zhangjiakou Competition Zone of the 2022 Beijing Winter Olympic Games during the 2017–2022 snow seasons, we analyze the spatial heterogeneity of 2-meter temperatures and ERA5 biases, along with their potential drivers. Multiple downscaling schemes are evaluated, comparing direct temperature prediction with bias-based modeling and assessing the impact of predictors associated with CAP and NW formation (e.g., radiative cooling, wind shear, and temperature advection). Results indicate that CAPs and NW occur on approximately 70% of nights, co-occurring in 57% of cases. Bias-based models consistently outperform direct downscaling approaches, particularly when meso- and micro-scale predictors are included. The best-performing model (Scheme D) reduces RMSE by up to 70% on mountain peaks, 40% in valleys, and 49% on slopes compared to raw ERA5 data, and achieves 4–9% improvements over conventional bias-based XGBoost models during NW-CAP nights. SHAP analysis enhances interpretability by linking model predictors to physical processes. This study underscores the importance of incorporating meso- and micro-scale dynamics in machine learning-based downscaling and offers practical insights for improving high-resolution temperature forecasting in complex terrain.
Yue et al. (Tue,) studied this question.