基于UI-LSTM模型的短时降水预测研究OA

Precipitation nowcasting is to predict short-term rainfall in the future.Most existing precipitation forecasting models based on Recurrent Neural Network(RNN)use a single convolution kernel to extract the features of the input and hidden states,which is limited by locality.Thus these models cannot capture complex physical changes in radar echo images,and cannot effectively extract spatiotemporal correlations and make accurate forecasts for heavy rainfall regions.In view of the problems in the existing models,the UI-LSTM model is proposed for precipitation nowcasting,which can effectively extract the spatiotemporal correlation of the radar echo sequence.The proposed model adopts a U-shaped structure and uses skip connections for feature stitching to learn the contextual semantic information of the entire radar echo map and fuse features from the shallow and deep information.In addition,the Inception structure is added to replace the convolution in the ConvLSTM cell structure,and features of the input and the hidden state are effectively extracted through convolution kernels of different sizes.The experimental results show that the UI-LSTM model performs much better than the existing model in terms of HSS,CSI,MAE and SSIM,and the accuracy of heavy precipitation prediction is improved.