Prediction Of Road Traffic Flow Based On Deep Spatio-Temporal Graph Neural Networks

Traffic flow prediction is the emphasis of ITS.The accurate prediction results can provide reliable data support for the traffic management departments.It is the prerequisite for traffic guidance and control decisions,and has a positive significance for alleviating traffic congestion.However,traffic flow is a spatio-temporal data with complex spatio-temporal dependence,which poses a great challenge to traffic flow prediction.In order to capture the spatio-temporal correlations of traffic flow data deeply,we construct two models to capture the spatio-temporal dependence of traffic flow.The research contents are as follows:(1)We summarize the relevant research progress of traffic flow prediction,and analyze the results and deficiencies of existing research.The various topological structures of the road network and the spatio-temporal characteristics of the traffic flow are analyzed in detail,and the construction methods of the spatial neighbor graph and the weighted spatial neighbor graph are proposed.The traffic network is regarded as a graph to learn to model the spatio-temporal relationship.Combining the GCN’s ability to extract spatial features and the processing advantages of RNN for time series data,according to the idea of the combined model,the GCN-LSTM model is proposed.It extracts the shallow and deep spatial features through the graph convolutional layer,and extracts the time series features through the recurrent layer.(2)On this basis,we deeply analyze the limitations of the combined model,point out the organic unity of the traffic flow characteristics in the temporal dimension and the spatial dimension,and propose an improved method of spatio-temporal feature fusion modeling.In order to capture the dynamic spatio-temporal dependence of traffic flow,dynamic spatio-temporal graph,spatio-temporal graph convolutional recurrent module(STGCRM)and multi-graph component are introduced to construct the dynamic spatio-temporal graph convolutional recurrent network(DSTGCRN).First,construct a dynamic spatio-temporal relationship graph based on the dynamic correlation of node features,construct static graph component and dynamic graph component,each component is composed of STGCRM,which is constructed by embedding the structure of the GRU by the Chebyshev graph convolution,to extract the spatio-temporal characteristics synchronously,and then weighting and fusing the output components of each graph component as the final prediction result.(3)On the two real-world traffic datasets,the model is subjected to experimental analysis and performance evaluation.Comparing experiments with the baseline model,the results show that the DSTGCRN model has a better performance than the baseline model in terms of MAE,RMSE,and MAPE.In addition,through ablation experiments,the effectiveness of each sub-module in improving the prediction performance is proved.Finally,according to the prediction results of the DSTGCRN model,the application of traffic state discrimination is carried out.According to the different datasets,the methods of cluster analysis and index division are used to distinguish the traffic state into different levels,and the results are intuitively displayed through visual graphs,and the road congestion is visually represented by different shades of colors.The result of the traffic state discrimination can be used to release traffic congestion information,which is helpful for travelers to adjust their travel plans in time,and helps traffic management departments to take traffic guidance and control measures to avoid deterioration of congestion.