本文已被：浏览 21次 下载 26次
投稿时间：2025-02-18 修订日期：2026-03-22
投稿时间：2025-02-18 修订日期：2026-03-22
中文摘要: 通过优化生成对抗网络（generative adversarial network，GAN）提出两种深度学习短时强降水临近预报方法PhySGAN（PhyDNet结合SGAN）和PhyMGAN（PhyDNet结合MGAN），提供江淮地区未来3 h的降水量预报。通过预报技巧评分、“复杂场景”性能评估和典型应用实例分析了两种方法在2024年江苏汛期短时强降水（降水强度≥20 mm·h-1）的预报表现，得到以下结论： PhySGAN和PhyMGAN在不同检验时段的3 h内短时强降水TS较基础试验PhyDNet和通用生成对抗网络试验PhyGAN均有明显提升，通过修正预报频率，使得TS增益幅度随预报时效增加，进而提升了较长时效的短时强降水TS；从各方法在“复杂场景”所展示出的预报性能看，深度学习较传统外推方法能体现强降水生消演变，PhySGAN和PhyMGAN较PhyDNet和PhyGAN表现出更好的预报性能，前者对强降水形态、强度等局部细节刻画能力更好，后者对强降水雨带的整体轮廓和位置表征更好；结合汛期典型强降水实例应用发现，PhySGAN和PhyMGAN在系统性强降水和局地强降水案例中均提前预报出降水增强过程，对临灾预警起到有效指导。此外，PhyMGAN对50 mm·h-1以上极端降水强度有一定指示作用，PhySGAN则能够更好体现雨带形态和位置的转变。
中文关键词: 生成对抗网络，江淮地区，强降水临近预报，典型案例
Abstract:This paper proposes two deep learning-based short-term heavy precipitation nowcasting methods for the Jianghuai Area by optimizing generative adversarial network (GAN), namely PhySGAN (combining PhyDNet and SGAN) and PhyMGAN (combining PhyDNet and MGAN), so as to provide precipitation forecasts in the next 3 hours for the Jianghuai Area. Based on the forecast skill score, the performance assessment in “complex scenarios” and the analysis of typical application examples, this paper analyzes the forecast performance of the two methods in the short-time heavy precipitation (≥20 mm·h-1) forecasts in Jiangsu Province during the flood season of 2024. The results show that the TS scores of short-term heavy precipitation within 3 hours in different verification periods of PhySGAN and PhyMGAN are significantly improved compared with those of the basic experiment PhyDNet and the general GAN experiment PhyGAN. The two new methods can correct the low frequency problem of short-term heavy precipitation forecasts by PhyDNet and PhyGAN, so that the TS score increases with the increase of the forecast lead time, thereby effectively extending the nowcasting lead time of short-term heavy precipitation. Judged from the forecast performance shown by each method in “complex scenarios”, deep learning can reflect the evolution of the generation and dissipation of heavy precipitation relative to the traditional extrapolation methods. PhySGAN and PhyMGAN show better forecast performance than PhyDNet and PhyGAN. The former has a better ability to depict local details such as the shape and intensity of heavy precipitation, while the latter has a better representation of the overall contour and position of the heavy precipitation rain band. Combined with the application of typical heavy precipitation cases during the flood season, both PhySGAN and PhyMGAN can forecast the precipitation enhancement process in advance in both systematic heavy precipitation and local heavy precipitation cases, effectively guiding the early warning of disasters. In addition, PhyMGAN has a certain indicative effect on extreme rainfall intensities above 50 mm·h-1, while PhySGAN can better reflect the changes in the shape and position of the rain band.
文章编号： 中图分类号： 文献标志码：
基金项目:中国气象局能力提升联合研究专项（22NLTSZ001、24NLTSQ015）、江苏省气象局科研项目（KM202520）、中国气象局揭榜挂帅项目（CMAJBGS202512）、江苏省社会发展重大科技示范项目（BE2023766）和中国气象局重点创新团队（CMA2022ZD04）共同资助
引用文本：
庄潇然,代刊,曾康,徐珺,王啸华,刘梅,2026.利用生成对抗网络提升江淮地区强降水临近预报性能[J].气象,52(5):552-565.
ZHUANG Xiaoran,DAI Kan,ZENG Kang,XU Jun,WANG Xiaohua,LIU Mei,2026.Using Generative Adversarial Network to Improve Heavy Precipitation Nowcasting in the Jianghuai Area[J].Meteor Mon,52(5):552-565.
庄潇然,代刊,曾康,徐珺,王啸华,刘梅,2026.利用生成对抗网络提升江淮地区强降水临近预报性能[J].气象,52(5):552-565.
ZHUANG Xiaoran,DAI Kan,ZENG Kang,XU Jun,WANG Xiaohua,LIU Mei,2026.Using Generative Adversarial Network to Improve Heavy Precipitation Nowcasting in the Jianghuai Area[J].Meteor Mon,52(5):552-565.
