Research On Weather Radar Echo Extrapolation Method Based On Deep Spatiotemporal Series Predictio

Short-term extreme precipitation events often lead to secondary disasters such as flooding,posing great risks to people’s lives and property.At present,the meteorological field often uses weather radar echo extrapolation-based methods for nowcasting,to estimate the range and intensity of precipitation in local area within a short period of time,then release timely warning notices of severe precipitation events to the public.Exploring methods to further improve the accuracy of extrapolation has always been a hot issue in the field of nowcasting.In recent years,many scholars have tried to apply the deep learning technology emerging in the field of artificial intelligence to the radar echo extrapolation task.Compared with traditional methods such as optical flow,deep learning-based extrapolation methods are applicable to more conditions and can find potential patterns in a large amount of historical data.However,these methods still face the dilemma of inaccurate prediction of radar echo evolution trends and underestimation of high-intensity echoes,which hinder its further application in nowcasting.In response to the above problems,this paper collects domestic weather radar echo data and generates corresponding datasets to carry out the following research:(1)In order to increase the accuracy of the existing models for predicting the holistic evolution trend of radar echoes,a radar echo extrapolation method based on decoupling and modeling for spatial and temporal features and Self-Attention mechanism is proposed.It constructs the FDSA-Net extrapolation network,which consists of FDSA-LSTM units.First,the dual branches for decoupling and modeling spatiotemporal features are established in the unit,so that the model can extract spatiotemporal features of radar echoes in a more refined manner and model its dynamics of variations accurately.Then,the Self-Attention Module is used to strengthen the model’s capture of the global features.Finally,Depthwise Separable Convolutions are introduced to reduce the amount of model parameters and improve training efficiency.Experimental results show that,this network can more accurately predicts the holistic evolution trends of radar echoes than previous representative methods,in terms of generation,dissipation,and movement.(2)To further improve the prediction ability of existing models for high-intensity echoes,a radar echo extrapolation method based on multi-source features and mixed-domain attention mechanism is presented.The method establishes an MMST-Net extrapolation network with the basic component unit of MMST-LSTM.First,the model avoids forgetting features related to high-intensity echoes by embedding the Multi-source Features-based Contextual Correlations Enhancement Module in the unit.Then,the Mixed-domain Attention-based Feature Update Module is used to enhance the model’s ability to capture and learn high-intensity echo features.Finally,the loss function commonly used for model training is optimized by adding a weight term to prevent extrapolation results from tending to medium and low intensity values.Experimental results show that the method has improved prediction ability in terms of high-intensity echo values and their boundaries.