Dynamical and Machine Learning Hybrid Seasonal Prediction of Summer Rainfall in China

Abstract

Seasonal prediction of summer rainfall is crucial to reduction of regional disasters, but currently it has a low prediction skill. We developed a dynamical and machine learning hybrid (MLD) seasonal prediction method for summer rainfall in China based on circulation fields from the Chinese Academy of Sciences (CAS) Flexible Global Ocean-Atmosphere-Land System Model finite volume version 2 (FGOALS-f2) operational dynamical prediction model. Through selecting optimum hyperparameters for three machine learning methods to obtain the best fit and least overfitting, an ensemble mean of the random forest and gradient boosting regression tree methods was shown to have the highest prediction skill measured by the anomalous correlation coefficient. The skill has an average value of 0.34 in the historical cross-validation period (1981–2010) and 0.20 in the 10-yr period (2011–2020) of independent prediction, which significantly improves the dynamical prediction skill by 400%. Both reducing overfitting and using the best dynamical prediction are important in applications of the MLD method and in-depth analysis of these warrants a further investigation.