The Research On Highway Traffic Flow Prediction Based On Fusion Dynamic GCN And Gated Convolutional Networks

Timely and accurate traffic flow analysis and prediction are of great practical significance to the new road construction and traffic management of smart cities in the new era.However,due to the complexity and non-stationary characteristics of traffic flow data,how to mine the nonlinear spatiotemporal characteristics of traffic flow is a challenging problem in the field of traffic flow forecasting.Graph convolution network had been widely used to represent the spatial correlations between roads.The Laplace matrix defined according to the distance between road sections is used to model the spatial correlations of traffic flow.However,in practical applications,this spatial dependency will change over time,and this change cannot be captured by using the pre-defined Laplace matrix.Therefore,in order to predict highway traffic flow more accurately,this thesis proposes a highway traffic flow forecasting model(DSTGCNN)based on the fusion of dynamic GCN and gated convolution networks.This model can fully mine the spatiotemporal correlations of data in the process of constructing a graph,dynamically extract the spatial features of traffic flow,and consider the potential temporal features.The traffic flow of each section of the highway is effectively predicted.First of all,in the model of extracting dynamic spatiotemporal features,considering the limitation of previous methods using fixed static Laplace matrix,this thesis designs a new Laplace matrix construction method,which adaptively and dynamically extracts the spatiotemporal relationships of traffic flow,and then feeds this relationship into GCN to form dynamic graph convolution networks.Then combined with the gated convolution networks to form a sandwich-style dynamic spatiotemporal convolution block to extract the spatiotemporal features of traffic flow more effectively.Secondly,in the construction method of dynamic Laplace matrix,this thesis proposes a feature sampling method to reduce the traffic flow data dimension of traffic features and improve the running efficiency of the model,and uses GRU network to mine the potential temporal features between Laplace matrices to form a dynamic spatiotemporal Laplace matrix,which can more truly reflect the actual situation of traffic flow and improve the prediction ability of the model.Finally,through the verification and analysis of the proposed model on four real highway data,the experimental results show that the performance of the prediction model proposed in this thesis is better than the benchmark methods,especially the GCN prediction methods based on fixed Laplace matrix.In the future work,the method proposed in this thesis will be extended to other spatiotemporal prediction tasks,and the influence of external conditions on the prediction performance of the model will be considered.