基于ConvLSTM的改进雷达回波外推方法研究
2023,31(9):166-173
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，从输入数据格式和损失函数两方面对ConvLSTM进行改进，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1h雷达数据预测未来1h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。
关键词：海洋强对流天气; 海上大风; 雷达回波外推; 自编码; ConvLSTM
Research on an Improved Radar Echo Extrapolation Method Based on ConvLSTM
Abstract：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 the 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, improve ConvLSTM from two aspects of input data format and loss function, ConvLSTM network based on self coding is constructed, Training the model by four years historical radar echo data of Cangzhou,the radar echo extrapolation model that can predict the future 1h radar echo intensity and shape using the historical 1h radar echo is obtained. Test set and case test results show that improved model has better extrapolation effect in strong echo prediction.
Key words：severe marine convective weather; sea gale; radar echo extrapolation; self encoding; ConvLSTM
收稿日期：2023-02-20
基金项目：国家自然科学基金(41675046)；环渤海区域协同创新基金(QYXM202013)；天津市气象局科研项目(202107ybxm02)
