Improving the fine structure of intense rainfall forecast by a designed generative adversarial network

Abstract

Accurate short-term quantitative precipitation forecasting (QPF) is critical for disaster prevention, mitigation, and socio-economic activities. However, due to the inherent limitations of numerical weather prediction (NWP) models, precipitation forecasts still exhibit substantial inaccuracies. In recent years, deep learning (DL) techniques have been increasingly applied to improve precipitation forecasts, yet these approaches often produce overly smoothed outputs that fail to meet operational requirements for detail and accuracy. In this study, we propose a Generative Fusion Residual Network (GFRNet), a generative adversarial network (GAN)-based framework that integrates multi-source NWP forecasts to generate 3-hourly quantitative precipitation forecasts for North China up to 24 h in advance. GFRNet employs an adversarial learning mechanism to enhance spatial structure reconstruction, combined with a weighted loss function and carefully designed sampling strategies to address the long-tailed distribution of precipitation and improve model training efficiency. Using independent rainy-season datasets from 2022–2024, we comprehensively evaluate the performance of GFRNet against three NWP models, a linear ensemble method (MSEM), and a deep learning baseline model (FRNet). Results show that GFRNet consistently outperforms the NWPs and baseline models across light, moderate, and heavy rainfall thresholds. Compared to the China Meteorological Administration’s highest-resolution regional model (CMA-3KM), GFRNet improves Threat Scores (TS) by 4 %, 28 %, 35 %, and 19 % at the 0.1, 10, 20, and 40 mm thresholds, respectively, and improves Fractions Skill Scores (FSS) by 13 %, 18 %, and 15 % at the 10, 20, and 40 mm thresholds. Moreover, GFRNet consistently achieves the highest Multi-Scale Structural Similarity (MS-SSIM) scores and significantly reduces RMSE, demonstrating robust spatial structure recovery, stable intensity control, and strong generalization ability. These advantages are particularly pronounced in systemic high-impact heavy rainfall events, underscoring the model’s operational value. FRNet shows advantages in forecasting heavy precipitation but suffers from high BIAS and weaker generalization, limiting its practical applicability. MSEM exhibits robust performance in light and moderate precipitation scenarios but degrades significantly under extreme precipitation conditions. Overall, GFRNet dynamically fuses multi-source NWP information, balances precipitation intensity and spatial structure, achieves higher forecasting skill, and improves forecast quality across diverse precipitation regimes.