PrecipFusionNet: A Unified Deep Learning Model for Improving Numerical Precipitation Prediction

Abstract Numerical weather prediction (NWP) often suffers from substantial biases when forecasting extreme rainfall. Traditional corrections tend to underuse spatial information, while deep learning approaches typically struggle with data imbalance for rare events. In this study, we propose PrecipFusionNet, a novel deep learning framework designed to better capture localized heavy rainfall, which combines an Att‐UKAN architecture, a U‐Net‐derived design integrating convolutional, Kolmogorov‐Arnold Network, and attention modules, with a fusion loss function. PrecipFusionNet improves the prediction skill for extreme events (≥50 mm/day) by 50.42%. For short‐ to medium‐range forecasts (1–4 days), it increases the average Equitable Threat Score (ETS) over China by 35.03% compared to the Global Forecast System (GFS). At longer forecast lead times (72 and 96 hr), it matches the ETS performance of GFS forecasts at 24 and 48 hr. In addition, our data augmentation strategy improves robustness across different lead times and regions, maintaining high accuracy in cross‐domain evaluations (e.g., over North America) and demonstrating strong spatiotemporal generalization.