江苏省级雷暴大风智能预警信号生成技术及其在2023年汛期的应用评估

Objective Warning Signal Generation Method for Thunderstorm Gale in Jiangsu and Its Application for the 2023 Rainy Season

关键词:  雷暴大风，临近预报，深度学习，预警信号

Key words:  thunderstorm gale  nowcasting  deep learning  warning signal

摘要:

为了实现江苏地区雷暴大风智能预警信号的自动生成以提升雷暴大风临近预报预警能力，在建立分钟级、公里级网格风场数据集的基础上，区分雷暴大风、系统性大风和混合性大风过程，并结合生成对抗网络（Generative Adversarial Network, GAN，针 对 雷 暴 大 风 建 模）和 具 有 物 理 约 束 架 构 的 时 空 卷 积 神 经 网 络（Physical Dynamics Network, PhyDNet，针对系统性大风和混合性大风建模），发展了基于深度学习的江苏地区0~2 h 雷暴大风临近预报模型（Blending），实现面向区县的客观、分级预警信号自动生成技术，并对比探究了Blending 和 PhyDNet_ALL（不区分大风类型，直接采用 PhyDNet 建模）两种深度学习方法分别生成的客观预警信号与主观预警信号在 2023年汛期的表现。结果表明：（1）相较于主观预警信号，由深度学习方法生成的客观预警信号能有效提升预警信号提前量；（2）深度学习方法可以提前预报出强对流大风天气及其演变过程；（3）由于 Blending针对雷暴大风单独建模，确保对流系统强度以及中小尺度特征的维持，因此 Blending可以更好地描述极端雷暴大风的演变特征，在橙色和红色预警信号提前量上显著高于PhyDNet_ALL。

Abstract:

To achieve the automatic generation of objective warning signals for thunderstorm gales in Jiangsu and to enhance nowcasting capabilities, a minute-scale and kilometer-scale wind field gridded dataset was established. This dataset distinguishes between different types of wind. By integrating a generative adversarial network for thunderstorm gale modeling and a PhyDNet for the modeling of wind associated with weather systems and mixed-type wind, we developed a deep-learning-based 0-2 hour nowcasting model (Blending) for thunderstorm gales in Jiangsu. Then, we compared the subjective and objective warning signals generated by the PhyDNet_ALL (which uses PhyDNet modeling without distinguishing wind types) and Blending for the 2023 rainy season. The results show that: (1) Compared to subjective warning signals, objective warning signals generated by deep learning methods effectively improved the lead time of warning signals. (2) Deep learning methods can predict thunderstorm gales and their evolution process in advance. (3) Blending, which models convective gales separately, ensures the maintenance of convection intensity and small-scale features, allowing it to better describe the evolution characteristics of extreme convective gales and significantly outperform PhyDNet_ALL in terms of improving the lead time of orange and red alerts.