Research On Deep Learning Model Of Short-term Precipitation Nowcasting Based On Self-attention Mechanism

The precipitation nowcasting refers to the forecast of the precipitation intensity in a specific area in the next few hours(normally in 2 hours),which can be used for the early warning of extreme weather.The new generation of Doppler weather radar owns the advantages of high sensitivity and high resolution that is very suitable for real-time forecasting so that it is widely used in nowcasting.Regarding to the forecasting methods,the extrapolation method based on radar echo,especially the optical flow,is the most typical one.However,due to the challenges such as lacking the full leverage of the observed historical data,difficulty in determining parameters,the performance optical flow method is limited.In recent years,the widely use of deep learning methods such as Convolutional Neural Networks(CNN)and Long Short-Term Memory networks(LSTM)that show great power in areas like image processing and time series data processing,and are now increasingly applied in precipitation nowcasting.In this paper,by summarizing the existing deep learning based methods in precipitation nowcasting,we make the following contributions.(1)Based on the traditional CNN-UNet model,we propose a self-attention UNet network model(Self-Attention UNet).In this model,the self-attention mechanism module is introduced into the encoder-decoder structure of the CNN-UNet model.In this way,the image feature extraction process happening in the encoder could capture information about focused areas in radar echo images.Then the decoder is linked with encoder by Skip-Connection to make full use of the data.Therefore,the model can track the target movement in the key area of the radar image.The experimental results show that compared with the optical flow method,the CNN-UNet model,and the Sma At-UNet model,in most cases,our Self-Attention UNet model has higher precipitation forecast accuracy.(2)Based on the convolutional long short-term memory(Conv LSTM)network model,we propose a long-term short-term memory network(Optical Flow Attention Fusion LSTM,OFAM-LSTM)that integrates self-attention mechanism and optical flow estimation.The optical flow estimation position correction module and the attention memory module are added to the convolutional long-short term memory network to improve the overall prediction performance of the model.In this network model,the input continuous time series images first go through the optical flow estimation position correction module to improve the calculation error accumulation problem caused by the object position offset in the feature image calculation process,and then use the improved result as the convolutional long-short term memory network.Finally,the attention memory module is added in the hidden state and output result calculation process to capture the overall characteristics of the input sequence image.The experimental results show that compared with the convolutional long-term short-term memory network and the other two advanced network models,this method can reach a similar level of precipitation forecast within 0-1 hours,and make more accurate forecasts for heavy precipitation weather.