Traffic Flow Prediction Based On Global Spatial-temporal Graph Attention Network

With the rapid development of urbanization and modernization,the stock of automobiles in my country has grown rapidly,and traffic congestion has become a severe challenge facing urban management.Intelligent Transportation System(ITS)helps to improve the traffic efficiency of the traffic road network.It can not only provide people with traffic information services,but also help promote smart management of urban traffic.The wide application of this technology will effectively alleviate the problem of urban traffic congestion.Accurate and effective traffic flow forecasting is the core component of the intelligent transportation system,which can provide a data-driven basis for intelligent traffic decision-making,thereby optimizing traffic scheduling and reducing road congestion.Due to the complex and dynamic temporal and spatial correlations between traffic road networks,this makes traffic flow forecasting challenging.On the one hand,the traffic network in the three-dimensional world has a complex spatial structure,and the traffic flow of the road network is directional and exhibits non-Euclidean correlation;on the other hand,the change of the road network traffic flow over time is nonlinear and non-linear.Stable is a typical time series data.Therefore,the core of traffic flow forecasting is to efficiently model the potential correlations and influences between temporal and spatial data.In previous studies,Convolutional Neural Networks(CNN)and Graph Convolutional Networks(GCN)were used for spatial correlation modeling,and time series were used to model the temporal dependence of traffic flow.However,the non-Euclidean correlation in the road network space reduces the modeling effect of the convolution operator.In order to improve the above problems,this paper proposes a spatial-temporal attention-based long short-term memory network(STALSTM).First,the complex transportation network is abstracted into a spatiotemporal graph structure,and then it is expanded and decomposed into a set of spatiotemporal influence factors.Finally,the spatiotemporal attention-based long-term memory network is derived from the factor graph representation of the spatiotemporal graph.Since the degree of influence of each traffic node is different at different times,we introduce an attention mechanism to capture the different degree of influence factors between traffic nodes.In addition,only considering the interactive impact of traffic around the point of interest,which oversimplifies the impact between the traffic network.On the basis of STALSTM,this paper proposes a global spatiotemporal graph attention network(GST-GAT)to predict traffic flow.This model uses the "global interaction + node query" method to model dynamic traffic spatiotemporal correlation.In the encoder,the LSTM component flexibly transforms the traffic dynamic spatial-temporal map into feedforward and differentiable feature codes.The global traffic interaction is proposed to summarize the changes in the traffic network context,and integrate all node features through a forward calculation at each moment.Then,the impact of global traffic interaction on a single traffic node is calculated,and the spatial-temporal interaction information is merged through the gated fusion mechanism.Finally,the encoder-decoder structure is used to train rich mixed feature codes to generate traffic predictions for each node.Experiments on the public transportation data set show that both STALSTM and GST-GAT have excellent prediction performance,and GST-GAT is better than previous work in terms of accuracy and reasoning speed.