Predicting Summer Precipitation in China: A Hybrid Downscaling Model Using the XGBoost Method

ABSTRACT The influencing factors of summer precipitation in China are complex, making prediction challenging, with the need to further improve its forecast accuracy. In the rapid development of machine learning algorithms, a key research topic is how to extract the physical factors that influence summer precipitation in China, and how to effectively integrate these factors with numerical models using machine learning methods to construct prediction models. This study develops a hybrid downscaling model based on the XGBoost method (referred to as the XGBoost model), and evaluates its performance by analysing the potential impacts of circulation and sea surface temperature (SST) anomalies on China's summer precipitation both in the observation and state‐of‐the‐art model. The results indicate that by introducing key climate factors such as sea level pressure (SLP), 500 hPa geopotential height (GH5) from global climate models (GCMs), and SST in the tropical central and eastern Pacific from observation, the XGBoost model can effectively capture the complex nonlinear relationships between large‐scale climate information and local precipitation. Compared with the original results from the BCCCSM1. 1m model, the XGBoost model shows significant improvements in the accuracy of predicting summer precipitation in China. In the cross‐validation for the period of 1994–2016, the anomaly correlation coefficient for summer precipitation increased from −0. 001 to 0. 34, the RMSE (Root Mean Square Error) reduction rate in most regions exceeded 40%, and the prediction score was significantly improved in most years. The XGBoost model shows obviously better performance than the original model results, validating the potential of the XGBoost method in handling climate data and providing new approaches and insights for regional precipitation prediction.