基于ConvLSTM的改进雷达回波外推方法研究  

Research on Improved Radar Echo Extrapolation Method Based on ConvLSTM

作　　者：赵玉娟[1] 李宗飞[1] 陈凯华[1] 王彦 朱男男 李祥海[1] 姜罕盛 ZHAO Yujuan;LI Zongfei;CHEN Kaihua;WANG Yan;ZHU Nannan;LI Xianghai;JIANG Hansheng(Tianjin Meteorological Information Center,Tianjin 300074,China;Tianjin Weather Modification Office,Tianjin 300074,China;Tianjin Marine Meteorological Centre,Tianjin 300074,China)

机构地区：[1]天津市气象信息中心,天津300074 [2]天津市人工影响天气办公室,天津300074 [3]天津海洋中心气象台,天津300074

出　　处：《计算机测量与控制》2023年第9期166-173,共8页Computer Measurement &Control

基　　金：国家自然科学基金(41675046);环海区域协同创新基金(QYXM202013);天津市气象局科研项目(202107ybxm02)。

摘　　要：海上大风及其引发的次生灾害是导致海洋气象灾害的主要因素;雷达观测数据是临近预报主要参考数据之一,准确的雷达外推数据对于提升海上强对流大风临近预报能力极为关键;面向海上大风预报需求和大风天气雷达回波特征,从输入数据格式和损失函数两方面对ConvLSTM进行改进,在损失函数中增加观测真实值作为系数提高强回波影响,构建了基于自编码的ConvLSTM网络,利用4年的沧州历史雷达回波数据对其进行训练,得到了可基于历史1 h雷达数据预测未来1 h雷达回波的雷达回波外推模型;测试集及个例检验结果表明,改进模型在强回波预测方面具备更好效果。Strong winds at sea and the secondary disasters caused by them are the main factors leading to marine meteorological disasters.Radar observation data is one of main reference data for proximity prediction,accurate radar extrapolation data is very important for improving the ability of predicting the approaching strong convective gales at sea.Facing the demand of offshore gale forecast and radar echo characteristics of gale weather,ConvLSTM is improved from two aspects of input data format and loss function,the observed true value is added as a coefficient in the loss function to improve the strong echo effect,a ConvLSTM network based on self-coding is construced,the model is trained by four years historical radar echo data of Cangzhou,the radar echo extrapolation model is obtained that the intensity and shape of the future 1-hour radar echo can be predicted by using the historical 1-hour radar echo data.The test set and case test results show that the improved model has better extrapolation effect in strong echo prediction.

关 键 词：海洋强对流天气 海上大风 雷达回波外推 自编码 ConvLSTM 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]