Skillful subseasonal prediction of rainfall and extreme events in East China based on deep learning

The subseasonal prediction with a lead time of 10–30 days is the gap between weather (<10 days) and climate (>30 days) predictions. Improving the forecast skill of extreme weather events at the subseasonal range is crucial for risk management of disastrous events. In this study, three deep-learning (DL) models based on the methods of convolutional neural network and gate recurrent unit are constructed to predict the rainfall anomalies and associated extreme events in East China at the lead times of 1–6 pentads. All DL models show skillful prediction of the temporal variation of rainfall anomalies (in terms of temporal correlation coefficient skill) over most regions in East China beyond 4 pentads, outperforming the dynamical models from the China Meteorological Administration (CMA) and the European Centre for Medium Range Weather Forecasts (ECMWF). The spatial distribution of the rainfall anomalies is also better predicted by the DL models than the dynamical models; and the DL models show higher pattern correlation coefficients than the dynamical models at lead times of 3 to 6 pentads. The higher skill of DL models in predicting the rainfall anomalies will help to improve the accuracy of extreme-event predictions. The Heidke skill scores of the extreme rainfall event forecast performed by the DL models are also superior to those of the dynamical models at a lead time beyond about 4 pentads. Heat map analysis for the DL models shows that the predictability sources are mainly the large-scale factors modulating the East Asian monsoon rainfall.