Research On Radar Echo Map Extrapolation Based On Deep Sequence Prediction

Weather forecast refers to the detailed monitoring of current weather conditions and the extrapolation forecast of future weather conditions within 2 hours.The key work of near weather forecast is to forecast typhoon,heavy rainfall,hail,thunderstorm and other weather,which is of great value to today's society.At present,near weather forecast is mainly realized through the analysis of radar echo map and the prediction of future radar echo map,so the core problem of near weather forecast is radar echo extrapolation.Since the 1950 s,people have studied the radar echo extrapolation technology and gradually formed a series of traditional methods,among which optical flow method has been widely used because it can extrapolate the weather field with strong change.However,the three assumptions of optical flow method are not tenable in the field of radar echo image,resulting in the low accuracy of its prediction image,which cannot meet the prediction demand.Secondly,it cannot use the large number of existing meteorological data.With the development of deep learning in recent years,a series of radar echo map extrapolation model based on deep learning has been proposed.Although the prediction metrics has been improved,it still has the disadvantages of fuzzy and distorted on generated image and detail disappearance.To solve these problems,this paper mainly completed the following two aspects of research content:Firstly,this paper proposes three improvement strategies for the existing sequence extrapolation deep learning framework.Through the analysis of the existing framework,it points out that the main problems lie in two aspects: the insufficient ability of recurrent neural network unit to model spatiotemporal information;and the loss of features in extracting spatial features of radar images.Therefore,this paper proposes three improvement strategies: reducing the information loss of sampling layer;enhancing the performance of sampling layer;enhancing the performance of recurrent neural network unit.Through comparative experiments,this paper proves that the three improved strategies can improve the accuracy of prediction and improve the quality of the generated image.Secondly,after improving the current framework based on the three improvement strategies,the results still have the problems of lack of clarity and detail.Therefore,this paper proposes the radar echo map extrapolation model based on EBGAN,which introduces the discriminator based on energy and generator based on the three improved strategies.This model regards the discriminator as a learning objective function.The model optimizes the quality of the image generated by radar extrapolation using the advantages of GAN network in the field of image generation,so as to solve the problem of insufficient definition and distortion of the predicted image.Through experiments,this paper proves the validity of the structure.