Dynamic Graph Convolutional Networks For Traffic Flow Forecasting And Impact Prediction Of Road Network Changes

For researchers in the transportation field,accurate traffic flow prediction is a key research direction,as it can improve the safety,stability,and efficiency of intelligent transportation systems.Despite years of research,predicting traffic flow is still very challenging and has room for improvement,including but not limited to the following reasons: While traffic flow data exhibits nonlinear and complex patterns in time and space dimensions,there are also dynamically changing spatial-temporal dependencies;Failure to make full use of various heterogeneous signal data in the field of transportation;The impact of changes in road traffic status on traffic flow is not considered.To address these challenges,this paper proposes an attention-based spatiotemporal graph convolutional ordinary differential equation network（HASTGODEN）supporting heterogeneous data input for road network traffic flow and road network change impact prediction.In addition to completing the task of traffic forecasting,try to apply the traffic flow change patterns learned by the model to the analysis of key nodes in the road network to fully tap the potential of the traffic forecasting model.Specifically,the main work of this paper includes the following aspects:First,in the problem of traffic flow prediction,in order to make full use of various valuable heterogeneous signal information,such as road network attribute information,GPS trajectory sequence,etc.,this paper presents a set of processing solutions.Combining the advantages of knowledge graph representation learning and factorization machine,suitable embedding representation methods are used for various heterogeneous data to fully mine and use these data.Turning these data into data usable by the neural network increases the upper limit of the theoretical prediction ability of the model.Second,in terms of the main model architecture,an encoder-decoder-based spatial-temporal graph convolutional network structure is adopted to capture temporal features and spatial patterns using standard convolution and graph convolution operations,respectively.On this basis,the self-attention mechanism and the change update and change diffusion unit of the road network are added to deal with the dynamic change characteristics of the spatial and temporal dependencies and the road network structure,so that the model can change the signal dynamics according to the change trend of the traffic flow and the road network.Adjust the parameters of the graph convolution.Then,extensive experimental studies are conducted on real traffic datasets in Shenzhen and Changchun.The results show that the traffic prediction model based on dynamic graph convolutional neural network proposed in this paper can obtain more accurate prediction results than other classical methods.Finally,this paper proposes a critical path analysis algorithm.On the basis of the node deletion method,the road network change score predicted by the network model that can learn the spatiotemporal dependence of traffic flow replaces the original topology network structure feature score.The characteristics of road network topology,traffic flow characteristics and traffic flow change patterns are comprehensively considered,and more reasonable and effective conclusions are drawn.In addition,on the basis of GPS trajectory data,this paper designs a set of simulation data generation algorithm based on real-time road conditions and shortest path planning.This algorithm comprehensively considers the user’s driving intention,road attribute information and the current driving condition of the road,so that the generated data can match the real situation as much as possible.And based on the simulation data set and other critical path analysis algorithms on the Changchun map,the experimental comparison is carried out.