CAST-LSTM:一种用于雷达回波外推的时空LSTM模型

CAST-LSTM:A Spatio-Temporal LSTM Model for Radar Echo Extrapolation

基于循环神经网络的雷达回波外推算法的预报结果随时间逐渐模糊失真,同时难以预报强回波区域。针对上述问题,提出一种上下文融合和注意力机制的时空长短期记忆网络模型。该方法通过上下文融合模块充分提取雷达图像不同尺度的短期上下文信息;通过注意力模块拓宽预测单元的时间感受域,使模型感知更多的时间动态。以2019—2021年4—9月江苏省气象雷达数据为样本,通过试验对比分析,基于上下文融合和注意力机制的时空长短期记忆网络取得了更好的预测性能。在外推60 min,阈值为10、20、40 dBz的条件下,临界成功指数和HSS分别达到0.7611、0.5326、0.2369和0.7335、0.5735、0.3075,有效提高了预测精度。

The forecast results of radar echo extrapolation algorithm based on recurrent neural network are gradually blurred and distorted with time,and it is difficult to forecast the severe echo area.To solve the above problems,this paper proposes a spatio-temporal long short-term memory network model based on context fusion and attention mechanism.The method fully extracts the short-term context information of different scales of radar image through the context fusion module.The attention module broadens the time perception domain of the prediction unit,so that the model perceives more time dynamics.Taking the weather radar data of Jiangsu Province from April to September in 2019-2021 as a sample,the spatio-temporal long short-term memory network based on context fusion and attention mechanism achieves better prediction performance through experimental comparison and analysis.Under the conditions of 60 min extrapolation and the thresholds of 10,20 and 40 dBz,the critical success index(CSI)and heidke skill score(HSS)reach 0.7611,0.5326,0.2369 and 0.7335,0.5735,0.3075,respectively,which effectively improved the prediction accuracy.