Research On Spatio-temporal Hypergraph Neural Networks For Traffic Flow Forecasting

Traffic flow forecasting is an important research topic in the field of intelligent transportation,which plays an irreplaceable role in downstream intelligent applications,such as traffic management control,vehicle scheduling,and route planning.Accurate traffic flow forecasting can provide effective decision support for traffic management departments,help them carry out reasonable flow control and scheduling,so as to reduce traffic congestion,improve traffic efficiency and passengers’ travel experience.However,in real scenarios,there are usually complex local spatio-temporal relationships in traffic data,including temporal correlations,spatio correlations and the correlations across the spatio-temporal dimensions.Therefore,accurate traffic flow forecasting faces great challenges.Most existing related works focus on modeling local spatio-temporal relationships from spatio and temporal dimensions respectively,which can not intuitively model the correlations across the spatio-temporal dimensions.In addition,local spatio-temporal relationships usually involve traffic data of multiple adjacent locations over continuous time steps.It is difficult for existing models to effectively mine and exploit these complex relationships simultaneously,which limits their forecasting accuracy.Therefore,there is still a lack of an effective method to model the local spatio-temporal relationships.To this end,this paper innovatively proposes a spatio-temporal hypergraph modeling framework and designs a method that can synchronously model the three correlations in the local spatio-temporal relationships.Based on these,two traffic flow forecasting models are constructed,which be summarized as follows.Firstly,to solve the problem that the existing methods can not model the local spatiotemporal relationships accurately and comprehensively,a spatio-temporal hypergraph convolutional networks model(STHGCN)is proposed.Based on the traffic network,a group of adjacent stations constituting the local spatio-temporal relationships are modeled in a hyperedge artificially,and then a spatio-temporal hypergraph is constituted to comprehensively model the local spatio-temporal relationships in STHGCN.Moreover,a sliding window mechanism and hypergraph convolution modules are designed to effectively capture the spatio-temporal features of traffic data.Secondly,considering the limitations of artificially constructed spatio-temporal hypergraph based on prior knowledge,the forecasting ability of the model will be affected to some extent.Therefore,based on the STHGCN,an adaptive spatio-temporal hypergraph convolutional networks with time encoding model(ASTHGCN-TE)is further proposed.By designing an adaptive spatio-temporal hypergraph learning component,the model can automatically generate more accurate hypergraph structure in a data-driven way.In addition,time information encoding modules of multiple dimensions are introduced into the model to help model the periodic correlations in the traffic flow series,so as to further improve the forecasting accuracy of the model.Finally,in order to verify the effectiveness and rationality of the proposed models,experiments are carried out on the public traffic flow data that collected from two scenarios of subway and freeway.The experimental results show that compared with the existing traffic flow forecasting models,the proposed models effectively improve the accuracy of traffic flow forecasting,and the computational efficiency is relatively high.