Research On Spatial-temporal Graph Network Based Traffic Flow Forecasting Method

With the rapid rise of electric vehicles,the car ownership ratio is growing rapidly,and the urban transportation system is severely challenged.Reasonable setting of traffic facilities,planning of traffic routes,and implementation of diversion schemes can effectively reduce the time and frequency of road congestion,reduce potential safety hazards,thereby improve residents’ happiness,which has also become an important topic in the Intelligent Transportation System(ITS).Traffic flow prediction,as an important step in intelligent traffic control methods,has been widely concerned by the academic community.Spatial-temporal patterns and randomness of traffic flow data pose challenges for efficient and accurate forecasting.Existing studies have incomplete modeling perspectives on spatiotemporal associations,local spatial patterns or short-term temporal patterns are extracted in these studies.Whereas traffic flow data not only has feature of temporal proximity,but also of temporal periodicity and long-term trends.In order to replenish perspectives in the model,This thesis proposes a novel traffic flow prediction method based on multi-view spatiotemporal graph convolutional networks,which models the global temporal correlation,long short-term local temporal correlation,and global spatial correlation in data respectively.In addition,existing methods use pre-built graphs based on real road networks and graphs obtained by graph embedding,which are difficult to match the graph structure characteristics in traffic scenarios.This thesis proposes a bidirectional heterosexual graph learning module,which adaptively learns the graph structure for the temporal and spatial graph convolution layers through the bidirectional heterogeneity formula.On this basis,This thesis proposes a traffic flow prediction model based on directed graph spatiotemporal recurrent network.The above-mentioned mixed temporal convolution layer needs to heuristically set the size of continuous convolution kernel and hole convolution kernel,which is difficult to adapt to the time characteristics of data in different application scenarios.In this model,the recurrent neural network structure is used instead of the convolutional neural network structure to extract the temporal correlation,and the directed graph convolution based on the first-order and second-order adjacent expressions is introduced.An encoder-decoder network structure composed of directed graph spatiotemporal recurrent units(DGSTRU)is proposed to generate prediction sequences from historical sequences.Finally,This thesis uses two public datasets of expressway vehicles to verify the prediction performance of the proposed multi-view spatial-temporal graph neural network(MSTGNN)and the directed graph spatial-temporal recurrent network(DGSTRN).Results show that proposed MSTGNN has higher prediction accuracy than CNN based baseline models,and proposed DGSTRN has higher prediction accuracy than most traffic flow prediction models.In addition,This thesis uses taxi data in the main urban area of Chongqing to verify the advantages of MSTGNN in adaptive methods that do not require pre-built graphs.