关键词: 长短时记忆网络, 雷达回波, 注意力, 多尺度

Abstract: Aiming at the problem of low prediction accuracy of short-approaching weather based on radar echo images,this paper proposes a multi-scale mixed attention long short-term memory network model.On the basis of long short-term memory network,on the one hand,an auxiliary branch is introduced to enhance the extraction of global image information.On the other hand,a mixed attention feature extraction module is designed to extract the fine-grained and coarse-grained information of the image data.The experimental results show that the proposed network model is superior to 9 models such as Conv-LSTM,Pred-RNN and RAP-Net on the two indexes of HSS and CSI.Especially,compared with those of the Pred-RNN model in the case of 5 dBz,20 dBz and 40 dBz,the HSS index of the proposed model is improved by 1.02%,2.46% and 7.94% respectively,while the CSI index is improved by 0.54%,2.29% and 4.91%,indicating obvious improvement.