Research On Precipitation Nowcasting Based On 3D Unet Structure

Precipitation nowcasting is very important for rainstorm warning,traffic safety and so on.Due to the rapid change of convective weather,accurate prediction has always been challenging.Precipitation nowcasting mainly includes the traditional numerical weather prediction(NWP)and the method based on radar echo extrapolation.The NWP method mostly uses supercomputers and takes several hours to complete,which requires a lot of computing resources and time.Methods based on radar echo extrapolation include the traditional strom cell identification tracking,tracking radar echoes by correlation,optical flow and deep learningbased extrapolation method,which is widely used optical flow method only uses the shallow motion vector features of the pixels of the radar echo map at the previous few historical moments,and the intensity change is difficult to predict,which limits the improvement of its accuracy.With the rise of artificial intelligence methods and the development of hardware resources such as GPUs,radar echo extrapolation based on deep learning shows great potential and has become a hot topic of research.With the support of National key R&D Program of China,the paper studies precipitation nowcasting method based on the 3D Unet structure.Firstly,the meteorological dataset is constructed,including the radar echo image dataset and the NWP mode dataset,and the data is preprocessed separately,including anomaly and vacancy value processing,denoising,normalization,etc.Then,aiming at the problems of gradient dissipation and high GPU memory requirements of the training network based on the precipitation nowcasting method based on RNN network,a encoder-decoder deep residual attention precipitation prediction network(ED-DRAP)based on 3D Unet structure is proposed,which integrates global and local residual structures to achieve a trainable deep residual prediction network,and integrates sequence and spatial attention mechanism to achieve residual prediction from high to low levels.The test results show that compared with the precipitation nowcasting method based on RNN,the proposed method can greatly reduce the resources required for training,and can also significantly improve the prediction accuracy of the full-convolutional network structure in precipitation prediction.Finally,in order to further improve the prediction accuracy of precipitation nowcasting,considering the mining and fusion of multi-source meteorological heterogeneous data,a precipitation prediction method(MS-DRAFP)based on multi-source residual attention fusion prediction is proposed,and the attention-based spatio-temporal diffusion module(ASTD)is used to effectively extract feature information from sparse spatio-temporal tensor.On the other hand,the decoder achieves multi-source heterogeneous fusion precipitation prediction through scale-wise attention fusion module(SWAF)and residual spatiotemporal attention block mechanisms(RSTABs).The test results show that the proposed MS-DRAFP network can effectively mine multi-source meteorological heterogeneous data information for fusion precipitation prediction,which can further improve the prediction accuracy of precipitation nowcasting compared with the method of using only a single data of radar echo.