Research On Traffic Flow Forecasting Based On Spatial-Temporal Data

The task of urban traffic flow prediction is an important part of the intelligent transportation system.Accurate prediction results can provide information support for traffic managers in vehicle scheduling and traffic control,as well as provide more convenient services for users to travel,reduce traffic congestion,reduce safety hazards,and promote the healthy and stable development of urban transportation.The current traffic flow prediction models mainly use static graph convolution to construct the spatial dependence of traffic flow.However,static graph convolution cannot capture the multi-level spatial dependence of traffic flow adaptively.At the same time,the existing traffic flow prediction models mainly target large-scale data such as urban road networks,the priori features of traffic flows in local and hotspot areas are not sufficiently applied.In this paper,research on traffic flow prediction algorithms based on spatial-temporal data is carried out to address the above problems,as follows:(1)Aiming at the problem that static graph convolution could not adaptive capture multi-layer spatial dependence of traffic flow,a spatial-temporal prediction model based on dynamic adaptive graph convolution is proposed.In this model,the adjacency matrix of graph convolution is initialized by obtaining the similarity between time series through Dynamic Time Warping algorithm(DTW),and then a dynamic adaptive graph convolution module is designed to change the adjacency matrix of graph convolution in different layers according to the input traffic flow data.The adjacency matrix of adjacent levels is connected by the hidden correlation between different levels of graph structure.Finally,an end-to-end network is constructed,the final prediction results are given and the validity of the model is verified on PEMSD4 and PEMSD8 datasets.(2)Aiming at the problem that the existing traffic flow prediction model is difficult to be applied to some local and hot spot area traffic flow prediction,a traffic flow prediction model suitable for urban hot spot station is proposed.Because the traffic flow of urban hot spot station has a certain correlation with the characteristic information of the station itself.A lightweight meta-learner module based on matrix decomposition is designed to learn the feature information of the site,and it is integrated with the spatial features extracted by the adaptive graph convolution module,and the time features are extracted by gated causal convolution,so as to obtain the final prediction results.Finally,Beijing Railway Station is taken as a hotspot site and the schedule data of Beijing Railway Station is taken as the site characteristic information to analyze and verify the algorithm effect.(3)Aiming at the existing urban traffic flow data,a set of urban traffic flow visualization platform is developed.The system is divided into three main functional modules,data display module were shown to urban traffic flow data,and carries on the visualization of the traffic flow data,traffic flow prediction module designed the data analysis module,can influence the external characteristics of traffic flow visualization analysis,and integrates the prediction model is proposed in this paper,shows the predicted results and the prediction error,It provides a user management module for managers,which can add and delete users.