SynQPF‐Net: Short‐Term Precipitation Forecasts by Integrating GraphCast Predictions and High‐Resolution Observational Analyses

Abstract While data‐driven global weather models (e.g., GraphCast) excel at large‐scale circulation forecasts, they exhibit systematic biases in heavy precipitation prediction due to European Centre for Medium‐Range Weather Forecasts Reanalysis (ERA5) uncertainties and neural networks' tendency to over‐smooth rainfall fields. To improve short‐term precipitation forecasts, we introduce SynQPF‐Net, a deep‐learning framework that fuses GraphCast outputs with high‐resolution observations. The network employs a dual‐stream spatiotemporal encoder and a U‐Net decoder to merge 0.25° GraphCast forecasts with 0.0625° precipitation analyses from the China Meteorological Administration (CMA) Land Data Assimilation System (CLDAS), optimized with a hybrid classification–regression loss that jointly learns exceedance probabilities and quantitative amounts for the different lead times (6–36 hr). Hindcasts over Southern China on the 2021 warm‐season test set show that for ≥10 mm rainfall, SynQPF‐Net raises the Critical Success Index from GraphCast's score of 0.151–0.314 at the 6 hr lead time, and from 0.146 to 0.211 at the 36 hr lead time. Interpretability analysis indicates that the model draws chiefly on CLDAS observations at short leads and increasingly on GraphCast circulation forecasts such as 700‐hPa specific humidity at longer leads, in line with physical reasoning. These results demonstrate that SynQPF‐Net generates rainfall forecasts with finer spatial structure and good physical realism, providing a new pathway for correcting data‐driven weather models through the integration of observation‐based analyses.