Urban Traffic Flow Prediction Based On Hybrid Deep Learning

In recent years,with the acceleration of urbanization,the total quantity of vehicle in our country is increasing day by day,which has brought the convenience for people to travel.At the same time,traffic congestion,traffic noise,air quality deterioration and other problems have seriously restricted the development of cities,and become a common problem faced by all cities in the country.The emergence of intelligent transportation and various modern technologies provide a new idea for improving traffic conditions effectively.Traffic flow prediction,as the key core technology of intelligent transportation system,has become a hot research topic in traffic field.Based on the theory and method of neural network and graph neural network,this paper establishes two kinds of mixed deep learning models for traffic flow prediction.The specific research contents include the following two aspects:(1)Aiming at the characteristics of traffic flow such as nonlinearity,temporal dynamics and spatial correlation,a CNN-Bi GRU model based on attention mechanism traffic flow prediction model is proposed.In this model,convolutional neural network(CNN)is used to capture the spatial correlation features of traffic flow.The bidirectional gated cycle unit(Bi GRU)acquires information from both forward and backward propagation,and fully extracts the time-dependent characteristics of traffic flow.The introduction of the attention mechanism enables the most important features of traffic flow to be allocated to the maximum degree of attention,so as to better improve the ability of CNN network and Bi GRU network to extract temporal and spatial features and achieve the purpose of improving the prediction accuracy of the model.The results of comparison and ablation experiments on real data sets show that the CNN-Bi GRU traffic flow prediction model based on attention mechanism has certain advantages in capturing temporal and spatial correlation,and has the advantage of improving the prediction performance compared with other baseline models.(2)Aiming at the spatial dynamics and multi-scale temporal dependence of traffic flow,a traffic flow prediction model combining Cheb Net and Bi GRU based on spatio-temporal attention(STAM-CN-Bi GRU)is proposed,which is composed of multi-layer spatio-temporal residual modules.Firstly,Cheb Net was used to learn the spatial dependence of traffic flow caused by road network topology,and Bi GRU combined with CNN to obtain the multi-scale temporal correlation of traffic flow.Then,the mixed spatio-temporal attention module is introduced into the traffic flow prediction model to describe the different importance of the spatio-temporal dependence of each node.The forecast results of three different time granularity traffic flows(the recent segment,the daily-periodic segment,the weekly-periodic segment)were weighted and fused to get the final forecast results.Finally,on the real data set,the prediction performance of one-step prediction and multi-step prediction of the proposed model is extensively evaluated,compared with other baseline models.the results show that the proposed model can significantly improve the prediction accuracy compared with other baseline models,especially in the aspect of multi-step prediction.