Machine learning subseasonal forecast model of summer precipitation: a case study for the upper Yangtze River Basin

Subseasonal forecast of regional precipitation remains highly challenging due to the chaotic nature of the atmosphere and complex variability.It is particularly true for the summer precipitation over Upper Yangtze River Basin (SP_UYZ), where intricate terrain and the synergistic effects of multiple predictors further complicate forecasting.This study develops a three-layer Gated Recurrent Unit model for the subseasonal prediction (GRU_S2S) of SP_UYZ, integrating the preprocessing of K-Means temporal clustering, predictor screening and the reconstruction.During the 2014-2022 hindcast period, the GRU_S2S model achieves averaged anomaly correlation coefficient skills of 0.34 and 0.36 at lead 30 and 60 days, respectively, outperforming the European Centre for Medium-Range Weather Forecasts (ECMWF) with that of 0.00 at lead 30-day.Beyond the well-documented roles of tropical Pacific intraseasonal oscillation and the Eurasian atmospheric circulation, GRU_S2S model reveals that the atmospheric circulation anomalies originating in South Pacific serves an additional key source of predictability.