Improved Prediction of Extreme Rainfall Using a Machine Learning Approach

Under global warming, extreme precipitation in the Yellow River Basin (YRB) is increasing,threatening regional development and the environment. This trend highlights the need to improve our ability to predict extreme precipitation. We used an empirical orthogonal function (EOF) analysis and a method based on year-to-year increments (DY) method to extract the spatial patterns and principal components (PCs) of the DY of total precipition exceeding 95th percentile threshold in summer (R95pTOT-DY) in the YRB from 1962 to 2010. Prediction models of PCs during 2011–2022 were autonomously established using two sets of machine learning (ML) models (Light Gradient Boosting Machine(LightGBM) and Categorical Boosting (CatBoost)). Without human intervention, these models independently identified significant climatic factors from 114 ocean and circulation indices advanced by 1–6 months. LightGBM predicted time correlation coefficients (TCCs) of 0.53, 0.63, and 0.60 for the PCs, while CatBoost yielded TCCs of 0.50, 0.69, and 0.57, respectively. Notably, the PC3 prediction significantly outperformed traditional linear regression (LR). Improved R95pTOT-DY prediction was observed in the YRB's source area and its middle reaches. The ensemble models of LightGBM and CatBoost showed the best R95pTOT-DY prediction (regional average TCC of 0.51), with a reasonable performance in 12 years of forecasting R95pTOT (a multiyear average Pattern Correlation Coefficient (PCC) of 0.36), surpassing the traditional LR method. The SHapley Additive exPlanations (SHAP) analysis attributed PC3 enhancement to the North American Polar Vortex Area Index (NAPVAI). Overall, this ML-based incremental model enhances extreme precipitation prediction, aiding risk assessment and warnings in the YRB.