数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (5): 113-122.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.05.009

doi: 10.11871/jfdc.issn.2096-742X.2025.05.009

• 专刊：国产算力新力量，助力科学计算应用新发展 • 上一篇    下一篇

基于雷达回波外推的短临降水预报模型TrajCast在国产加速器上的实践

辛昱杭(),王琦祎,孙婧,赵春燕*(),刘雨佳,梁雪,陈杰   

  1. 国家气象信息中心，北京 100081
  • 收稿日期:2025-02-27 出版日期:2025-10-20 发布日期:2025-10-23
  • 通讯作者: 赵春燕
  • 作者简介:辛昱杭，国家气象信息中心先进计算室，工程师，研究方向为人工智能气象应用.
    本文中承担工作为TrajCast方案设计实现与在GPU上的实验。
    XIN Yuhang is an engineer at the Advanced Computing Division of the National Meteorological Information Center. His research direction is the application of artificial intelligence in meteorology.
    In this paper, he is responsible for the design and implementation of the TrajCast scheme and experiments on GPU.
    E-mail: xinyj@cma.gov.cn|赵春燕，国家气象信息中心先进计算室，正高级工程师，研究方向为气象智能应用支撑技术，本文中承担工作为方案及评测方法总体设计。
    ZHAO Chunyan is a senior engineer at the Advanced Computing Division of the National Meteorological Information Center. Her research direction is the supporting technology for meteorological intelligent applications.
    In this paper, she is responsible for the overall design of the scheme and evaluation methods.
    E-mail: zhaocy@cma.gov.cn
  • 基金资助:
    光合基金“基于TrajGRU雷达回波外推算法的短临降水预报模型在国产加速器上的实践”(ghfund202302034726)

Application of Radar Echo Extrapolation Based Model TrajCast on Domestic Accelerators for Short-Term and Imminent Precipitation Forecasting

XIN Yuhang(),WANG Qiyi,SUN Jing,ZHAO Chunyan*(),LIU Yujia,LIANG Xue,CHEN Jie   

  1. National Meteorological Information Centre, Beijing 100081, China
  • Received:2025-02-27 Online:2025-10-20 Published:2025-10-23
  • Contact: ZHAO Chunyan

RichHTML

0

PDF

16

摘要：

【目的】 本文旨在构建一种基于深度学习的短时临近降水预报模型，并探索国产加速器在气象领域的应用，以提高气象防灾减灾建设中的预报准确性，为城市建设和交通等提供更可靠的气象保障。【文献范围】 本文聚焦于短临降水预报技术的发展，特别是基于深度学习的雷达回波外推方法在气象领域的应用。【应用背景】 传统数值预报在前两小时内的降水预报存在不确定性，而基于光流法和交叉相关法等物理模型的雷达回波外推方法在复杂气象条件下预测精度和稳定性受限。【方法】 本文采用中国气象局研发的短时强降水AI应用专项数据集，基于TrajGRU深度学习网络架构改进形成短临降水预报模型TrajCast，并在国产加速器上应用数据并行和混合精度训练技术进行模型训练。【结果】 试验结果表明，模型在降水预报性能上优于传统光流法，且通过计算优化技术实现了国产加速器单卡模型训练性能4.1倍的加速比。【局限】 模型在实际应用中对数据质量和计算资源有一定要求，且在极端天气条件下的表现仍需进一步验证。【结论】 本文的研究为短临降水预报提供了新的方法和技术支持，推动了国产加速器在气象领域的应用，模型已在湖南、浙江、湖北等气象部门进行了示范应用，为实现高分辨率、高时率的气象智能预报产品提供了智能算法支撑。

关键词: 短临降水预报, 深度学习, 国产加速器, 雷达回波外推, 混合精度训练

Abstract:

[Objective] This paper aims to construct a short-term and imminent precipitation forecasting model based on deep learning and explores the application of domestic accelerators in the field of meteorology, with the goal of improving the accuracy of meteorological disaster prevention and mitigation, and providing more reliable meteorological support for urban development and transportation. [Literature Scope] This paper focuses on the development of short-term precipitation forecasting technology, especially the application of deep learning-based radar echo extrapolation methods in the field of meteorology. [Context] Traditional numerical weather prediction has uncertainty in precipitation forecasting within the first two hours, and radar echo extrapolation methods based on physical models, such as optical flow and cross-correlation methods, have limited prediction accuracy and stability under complex meteorological conditions. [Methods] In this study, we employ specialized datasets for short-term intense precipitation of an AI application developed by China Meteorological Administration. Based on TrajGRU, a deep learning network architecture, we establish the short-term and imminent precipitation forecasting model TrajCast. Employing data parallelism and mixed-precision training techniques, the model training is implemented on domestic accelerators. [Results] Experiment results demonstrate that the model outperforms traditional optical flow methods in precipitation forecasting, achieving a 4.1-fold acceleration ratio in model training on a domestic accelerator. [Limitations] The model has certain requirements for data quality and computing resources in practical applications, and its performance under extreme weather conditions still needs further verification. [Conclusions] This study provides new methods and technical support for short-term precipitation forecasting, promotes the application of domestic accelerators in the field of meteorology. The model has been demonstrated and applied in meteorological departments in Hunan, Zhejiang, and Hubei provinces, providing intelligent algorithm support for high-resolution and high-frequency meteorological intelligent forecasting products.

Key words: short-term and imminent precipitation forecast, deep learning, domestic accelerator, radar echo extrapolation, mixed-precision training