Research On Traffic Flow Forecasting Based On Adversarial Deep Learning

With the development of social informatization,Intelligent Transportation System(ITS)has become an important tool for urban traffic management and transportation.Traffic flow forecasting,as a key link in ITS,is one of the important methods to make traffic management strategy quickly and accurately,to improve traffic efficiency and to reduce traffic accidents.Traffic flow forecasting has become a hot topic in traffic field.So far,researchers have developed a number of traffic flow prediction models using deep learning techniques,and have achieved good prediction results.This paper focuses on the research of traffic flow prediction based on deep learning related techniques.The traffic flow forecasting problem is studied from a new perspective in this dissertation.Starting from the two dimensions of space and time,and based on adversarial deep learning as the technical basis,the relevant models of traffic flow spatial prediction and traffic flow time prediction are established.First of all,the traffic flow data set used,the theory of generating adversarial networks,and the evaluation criteria are introduced.And uses Pearson correlation coefficients and statistical analysis to analyze the characteristics of the traffic flow data.Next,the traffic flow forecasting problem from the spatial dimension is studied in this dissertation.This thesis uses the methods of BICUBIC interpolation and Convolutional Neural Network to make spatial predictions.Then,based on residual learning and sub-convolution structure,a specific generative adversarial network for spatial prediction of traffic flow is established.The results show that the Convolutional Neural Network and the Generative Adversarial Network are significantly better than the interpolation method as a whole.Compared with the Convolutional Neural Network,the Generative Adversarial Network reduces the average NRMSE(Normalized Root Mean Squared Error)by 4.06% and improves the structural similarity by 2.34%.Finally,the time series ARIMA(Autoregressive Integrated Moving Average Model),the Deep Residual Network model,and the deep multi-scale Generation Adversarial Network model are used to solve the traffic flow time prediction problem respectively.According to the prediction results,it is found that compared with other models,the Generated Adversarial Network model shows better performance,which reduces the NRMSE numerical error by 17.85% compared with the classical residual learning model.In summary,the traffic flow prediction problem based on the spatial and temporal dimensions are studied through knowledge and technologies such as deep learning and generative adversarial networks in this dissertation.The research in this thesis provides a new application prospect of adversarial deep learning in the field of transportation,and provides a scientific theoretical basis for the research and development of traffic flow prediction.