Research On Passenger Flow Spatiotemporal Series Prediction Based On Deep Learning

With the development of transportation and the increase of urban population,traffic congestion occurs from time to time,which not only increases people’s travel time,but also reduces the operation efficiency of transportation.As an important function of intelligent transportation system,passenger flow prediction will detect and manage the information of the whole traffic network and provide decision-making reference for managers.It is not only the best way to alleviate traffic congestion,but also one of the development directions of intelligent cities.Traffic flow data has spatiotemporal correlation,which belongs to a typical spatiotemporal series,therefore,the method of dealing with spatiotemporal series can be used to analyze and predict passenger flow.In view of the shortcomings of existing passenger flow forecasting methods,this thesis proposes two passenger flow forecasting models based on deep learning framework,and analyzes their effects.The main contents of this thesis are as follows:(1)In view of the fact that most of the current traffic passenger flow forecasting methods do not accurately express the correlation of each node in the traffic network at different times,which can not obtain effective dynamic spatial correlation and affect the accuracy of prediction,therefore,a passenger flow forecasting method based on dynamic graph convolution network is proposed in this thesis.First of all,according to the traffic volume in different time intervals in the road network,this thesis calculates the correlation matrix between the nodes at the current time,and the correlation matrix can be obtained from the average of historical traffic volume,it can also be obtained from the traffic volume of the last moment directly as the correlation matrix of the current moment.Secondly,considering that only the correlation matrix may change the node connectivity of the traffic network,the correlation matrix is fused with the adjacency matrix.Finally,the spatiotemporal characteristics of passenger flow data are obtained by using one-dimensional convolution network and graph convolution network based on fusion matrix,which are mapped to the final predicted node passenger flow.The model is tested on three real traffic data sets,and the results show the effectiveness of the model proposed in this thesis.(2)In view of the fact that most of the current passenger flow origin and destination matrix prediction methods can not effectively obtain the dynamic spatial correlation,and the input information can not be fully utilized,this thesis proposes a passenger flow origin and destination prediction method based on attention mechanism and dynamic graph convolution network.Firstly,the dynamic spatial characteristics based on the real road network connectivity are obtained by merging the mean value of the historical passenger flow origin and destination matrix with the adjacency matrix.Then the attention mechanism is used to deal with the dynamic spatial feature matrix and dynamically adjust the influence weight between nodes.At the same time,the passenger flow between the nodes in the historical passenger flow origin and destination matrix is taken as a separate time series,and the time correlation is obtained by using the long-short term memory network.Finally,the spatiotemporal correlation is integrated and mapped to the final predicted passenger flow origin and destination matrix.Through the experimental comparison on three data sets,it is shown that the proposed model is effective in predicting the origin and destination matrix of passenger flow.