Traffic Speed Forecasting Based On Spatial-Temporal Graph Attention Network

Traffic forecasting is an important task in current urban traffic management,and accurate traffic prediction is very important for the smooth operation of urban traffic management system.For example,urban traffic supervisors can alleviate urban traffic congestion with the help of accurate traffic prediction.However,the spatial-temporal correlations of traffic data are very complex,which makes it very challenging to achieve accurate traffic prediction.Historical research methods mainly capture the complex spatial-temporal dependencies of traffic data based on the node data characteristics.To achieve more accurate traffic prediction,this paper also takes the features of the edge data as well as the weekly periodicity characteristics of the node data in traffic dataset into account,whose characteristics have not been explored or not fully explored,and designs the corresponding module to realize the extraction of the aforementioned features.In this paper,a traffic speed forecasting model is constructed based on spatial-temporal graph attention network.Considering the efficiency and flexibility of multi-head attention mechanism in modeling dependency,this paper chooses to apply multi-head attention mechanism to capture the complex spatial-temporal dependencies of traffic data.Firstly,this paper proposes a method which can further improve the modeling efficiency of multi-head attention mechanism.Specifically,we propose spatial-temporal embedding and local traffic condition embedding of nodes in the road network,and concatenate these two embedding vectors into the query vector and key vector of multi-head attention mechanism respectively,so as to improve the modeling efficiency of multi-head attention mechanism in terms of spatial-temporal correlation.Then,in order to utilize the features of the edge data in traffic dataset,we construct a digraph-enhanced graph attention network to capture the temporal-dynamic spatial correlations of traffic data,and design a node-edge-node gated graph aggregation mechanism to capture the nonlinear temporal correlations between adjacent moments of traffic data.Furthermore,we integrate the above digraph-enhanced graph attention network,the above node-edge-node gated graph aggregation mechanism and a two-layer FNN into a spatial-temporal aggregation block,which can effectively capture the complex spatial-temporal correlations of traffic data in the range of multiple consecutive moments.In addition,in order to take advantage of the weekly periodicity of node data in traffic dataset,a temporal multi-head attention network is constructed to capture the nonlinear temporal correlations of node data between multiple historical moments and multiple future moments.Finally,the aforementioned spatial-temporal aggregation block and temporal multi-head attention network are seamlessly connected for many times to generate a traffic speed forecasting model based on spatial-temporal graph attention network.The model is an encoder-decoder architecture,in which the encoder is composed of L spatial-temporal aggregation blocks and the decoder is composed of L layers where each layer is integrated with a spatial-temporal aggregation block and a temporal multi-head attention network.Moreover,the traffic speed forecasting model based on spatial-temporal graph attention network predicts the traffic speeds of all nodes in the road network at multiple future moments in a non-autoregressive way,which can effectively alleviate the error propagation.In this paper,we evaluate the proposed model on two public traffic speed datasets,METR-LA and PEMS-BAY,and observe better performance than the state-of-the-art baselines,especially in long-term forecasting horizon.We believe that the traffic speed forecasting model based on spatial-temporal graph attention network captures the edge data features as well as the weekly periodicity of node data in traffic dataset which have not been explored or not fully explored,better than other baselines,on the basis of capturing the structural information of the road network.In addition,although the research work of this paper is evaluated through two traffic speed datasets,all of the research results hold for other spatial-temporal traffic forecasting tasks,such as traffic flow forecasting task.Moreover,the research results of this paper can be extended and applied to the spatial-temporal forecasting tasks in non-transportation field.