Seasonal forecast of winter precipitation over China using machine learning models

Abstract

This study employs two machine learning (ML) models, namely the light gradient boosting machine (LightGBM) and the extreme gradient boosting (XGBoost) model, to perform seasonal forecasts for winter precipitation over China. The seasonal forecast results derived from a dynamic model are used for the purpose of comparison. The predictors employed in the two ML models consist of climate variables from the preceding autumn. The ML models are trained using data during 1979 to 2010 without additional predictor selection process. Consequently, the ML models autonomously extract useful information and then apply it to perform seasonal forecasts of winter precipitation for 2011–2020. Results indicate that the two ML models exhibit reasonable forecast skill and demonstrate superior performance over various regions when compared to the dynamic model. Additionally, the ensemble forecasts of the two ML models and the dynamic model exhibit higher skill in China compared to the dynamic model forecast alone. The forecast skill of the ML models can be attributed to their skillful forecast of the first empirical orthogonal function mode (EOF1) of the winter precipitation. In addition, the split time count method and the Shapley additive explanation (SHAP) values analysis are performed to further understand and explain the ML models forecast results. Results show that the sea surface temperature (SST), especially the SST over the western Pacific plays an important role in the ML model seasonal forecasts of winter precipitation over China, consistent with previous studies. The outcomes of this investigation provide valuable insights that can enhance the practical implementation and understanding of ML models in the field of seasonal forecasting.