| With the rapid growth of car ownership,private travel has become the first choice for more and more people.The problem of urban traffic congestion caused by the growth of car ownership is becoming increasingly serious.The research field of intelligent transportation systems faces more challenges.As one of the key technologies of intelligent transportation system,accurate and efficient traffic flow prediction can provide accurate and reliable dynamic path guidance for travelers.At meanwhile,it can effectively improve the execution capability of ITS,thus alleviating the pressure brought by traffic congestion to the actual road network operation.In the research of traffic flow prediction,how to effectively mine the natural complex spatial-temporal dependent attributes in traffic flow data is the key to solve the problem.The spatial-temporal dependence of traffic flow data originates from the spatial structure relationship of the actual traffic road network due to geographic and regional factors.Therefore,this paper investigates the traffic flow problem based on spatial-temporal graph convolutional networks from the traffic flow data collected by the actual road network.The main work is as follows:(1)Traffic flow data characterization and spatial-temporal graph construction of road network traffic flow.This paper focused on the research of real road network traffic flow data,analyzes its characteristics,and abstracts the actual road network into road network graph data with unstructured attribute characteristics based on spectral graph theory,and the spatialtemporal graph of road network traffic flow data is constructed which laying the foundation for the establishment of traffic flow prediction model.(2)A spatial-temporal graph wavelet convolution network based for traffic flow prediction is proposed(ST-GWCN).In order to solve the problem of localized features in the previous used of graph convolutional networks for traffic flow prediction modeling.In this paper,the wavelet transform is used instead of Fourier transform to define the convolution operation on the graph,which reflects the node-centered information diffusion and improves the localized representation of node features.In ST-GWCN,multiple time step spatial graphs are fused together to effectively integrate graph wavelet convolution and gated time convolution.With the ability to simultaneously learn temporal and spatial correlation features,ST-GWCN can better capture complex spatiotemporal correlation information in spatiotemporal sequences.The experiment was carried out on real public transportation datasets: PEMS-BAY and PEMSD7(M).The comparison results showed that our proposed model outperformed baseline networks on these datasets,which indicated that ST-GWCN could better capture the spatialtemporal correlation information.(3)An attention and wavelet based spatial-temporal graph convolution network for traffic prediction is proposed(ST-AGWCN).It integrated attention and graph wavelet convolution networks to capture local and global spatial information.Meanwhile,in order to solve the timing dependency problem that gated time convolution cannot capture further time,we stacked a gated temporal convolutional network with a temporal attention mechanism to extract the time series information.The comparison results showed that our proposed model outperformed baseline networks on these datasets,which indicated that ST-AGWCN could better capture the spatial-temporal correlation information. |
