Research On Precipitation Nowcasting Based On Adversarial LSTM And Time Series Residual Network

Precipitation caused by strong convective weather evolves rapidly and destructive,which seriously threatens people's lives and property.Precipitation nowcasting technology provides strong support for the forecast and early warning of heavy precipitation.With the development of meteorological technology,using of radar echo maps for short-term precipitation prediction has become a widespread concern.At present,the main prediction methods are based on the fusion of radar echo extrapolation and multiple regression,but the accuracy is still insufficient to meet business requirements.In order to improve the accuracy,this thesis uses deep learning related technologies to improve the accuracy of short-term precipitation prediction by researching radar echo extrapolation and multiple regression algorithms.The research work of this thesis is as follows:(1)The existing echo extrapolation algorithms based on convolutional long short-term memory(Conv LSTM)have the following two disadvantages:(a)The ordinary convolution operation has limitations in processing images with location-variant features;(b)Conv LSTM uses unsupervised learning in the extrapolation process.The loss function is often defined as mean squared error(MSE),ignoring the distribution similarity between the extrapolated image and the original image.To solve the above defects,an improved echo extrapolation algorithm based on the adversarial LSTM is proposed.In order to solve the first drawbacks,the optical flow method is used to track the local features,which breaks through the limitation of the convolution filter on the location-invariant.In order to solve the second drawbacks,using deep convolutional generative adversarial networks(DCGAN)and extrapolated model to compose a game system to realize the fitting of the extrapolated image distribution to the original image distribution.Finally,compared with mainstream deep learning methods and meteorological business methods under four different radar reflectivity intensities.Experimental results show that the proposed algorithm has higher accuracy than other algorithms.(2)The existing multiple regression algorithms based on neural network have the following two disadvantages:(a)Lack of optimization of the network structure,only relying on the stacked method to expand the model's hypothesis space.(b)The learning of the feature of meteorological samples is insufficient,and the generalization ability of the model is weak.In order to solve the above two drawbacks,an improved multiple regression algorithm based on time-series residual network is proposed.Aiming at the first shortcoming,using residual network(Res Net)to encapsulate Conv LSTM to optimize the network structure.Aiming at the second shortcoming,enhance learning with bidirectional LSTM.Unidirectional Conv LSTM is further encapsulated into a bidirectional,so as to realize the full learning of the samples.Finally,the proposed algorithm is used to perform the precipitation regression prediction experiment on the result of(1),and the accuracy is compared with the existing deep learning algorithms and traditional integrated learning algorithms.The results show that the accuracy of the algorithm in nonlinear complex modeling has been greatly improved.