Weather Radar Echo Extrapolation Based On Recurrent Neural Network

Weather radar echo extrapolation,which is to predict the appearance,intensity and distribution of future echoes based on historical echo observations,is one of the most fundamental means for weather nowcasting.Accurate and effective extrapolation has an important significance in promoting the quality of nowcasting.Recent years,researchers have applied deep learning to the extrapolation problem and made great progress compared with the traditional extrapolation methods.However,there are still three limitations which should be considered.First,current methods are usually difficult to modeling and predict the evolutionary process of echoes accurately;Second,for deep learning models using general image reconstruction losses as training objective,are easy to suffer from the problem of blurry prediction;Third,the effective extrapolation time of existing methods is usually short and thus cannot well satisfy the need of actual nowcasting application.To address the above three problems,this paper has studied the methods of radar echo extrapolation based on the Recurrent Neural Network(RNN),the main work is as follows:(1)To extrapolate the echo evolution and address the blurry prediction problem,an extrapolation method based on Multi-level Correlation Long Short-term Memory(MLC-LSTM)is proposed.The model modeling and predicts the echo evolution by exploiting the spatiotemporal correlation between multi-level echoes,and avoids the blurry prediction problem by utilizing adversarial training.The experimental results show the method can effectively predict the complex patterns of echo motion and evolution and improve the extrapolation accuracy,while the extrapolated echoes are also realistic and sharp and the blurry prediction problem is alleviated.(2)To make long-term extrapolation,a Hierarchical Prediction Recurrent Neural Network(HPRNN)is designed,which is mainly composed by the long-range and short-range RNN and a refinement module.HPRNN employs both a hierarchical prediction strategy and a recurrent coarse-to-fine mechanism.The experimental results demonstrate that the HPRNN outperforms other models in terms of long-term extrapolation,which can well meet the needs for the application.