Enhanced Extended-Range Probabilistic Forecasting of Persistent Extreme Precipitation using a Deep Residual Network

Based on subseasonal-to-seasonal (S2S) datasets obtained from four models (ECMWF, CMA, UKMO, and NCEP) and ground-observed precipitation data for April–September 1997–2022, an improved deep residual network (ResNet) model was developed for the extended-range probabilistic forecasting of persistent extreme precipitation (PEP). The deterministic and probabilistic forecasts of the improved ResNet model were compared with those of five other methods (single-model ensemble, multi-model ensemble, principal component regression, and Bayesian model averaging). Results demonstrated that the improved ResNet model outperformed other methods in probabilistic forecasting. The enhanced performance of the ResNet model is attributed to the improvement in spatial correlation coefficients, which stems from its ability to deeply extract PEP features from the ECMWF S2S data and the optimization of the ResNet structure. As a result, the improved ResNet model achieves a positive Brier skill score for forecast lead times up to 30 d, highlighting its considerable potential for extended-range probabilistic forecasting of PEP in eastern China.