Research On Radar Echo Extrapolation Method Based On Attention Mechanism And Deep Spatiotemporal Fusion Network

Precipitation nowcasting is an important and challenging worldwide problem.Due to the highly nonlinear,random and complex characteristics of precipitation itself,the forecast accuracy of traditional methods is not enough to meet the needs of meteorological operations.At present,the main method of precipitation nowcasting is the radar echo extrapolation technology based on weather radar data.The radar echo extrapolation based on the deep learning method has received extensive attention from researchers.Deep learning methods have the ability to model nonlinear complex systems,and have surpassed traditional methods in terms of accuracy.However,there are still some deficiencies in the radar echo extrapolation model proposed by the researchers.In order to further improve the accuracy of precipitation nowcasting,this paper studies the radar echo extrapolation method.The main innovations are as follows:(1)This paper proposes an improved radar echo extrapolation model for the current sequence-to-sequence extrapolation framework.Through the analysis of the existing encodingdecoding framework,it is pointed out that the main problems lie in two points: information loss in the process of encoding and compression;no primary and secondary information in decoding and extraction.Therefore,this paper designs an encoder-decoder architecture based on a special attention module.The improved structure does not need to compress a lot of spatiotemporal information,and can use the attention module to learn to extract important information with obvious motion features.Through comparative experiments,the encoding-attention-forecasting architecture(AttEF)obtained by improving the encoding-decoding architecture in this paper can effectively improve the accuracy of prediction and the quality of images generated by the AttEF model.(2)Aiming at the lack of clarity and lack of details caused by the difficulty of modeling the high-order non-stationarity of the radar echo sequence with the AttEF model,this paper designs a spatio-temporal fusion neural network(STUNNER)based on the characteristics of radar data.In this paper,the idea of difference is introduced to construct a temporal-difference network to extract the deterministic trend features of high-order non-stationary data of radar sequences;the idea of dynamic convolution is introduced to construct a spatiotemporal trajectory network to learn the trajectory characteristics of the non-rigid motion of radar echoes over time.And through the two-stream fusion strategy,the trend feature and the trajectory feature are fused to realize the high-order non-stationary modeling of radar data.Through experiments on simulated data and real data sets,this paper verifies that this structure can effectively increase the extrapolation time and improve the extrapolation blur problem.