Research On Traffic Flow Prediction Algorithm Based On Adaptive Graph Convolution

Accurate and timely traffic flow prediction can help people plan driving routes in advance,help relevant departments reasonably allocate traffic resources and reduce road congestion.However,due to the complex nonlinear temporal and spatial dependence between traffic flow data and the influence of various unexpected traffic conditions,traffic flow prediction has always been considered as a challenging research topic.The existing traffic flow prediction models are difficult to fully mine the hidden spatiotemporal characteristics between traffic flow data,and incomplete in modeling the temporal and spatial relationship of data.To solve this problem,this thesis proposes two traffic flow prediction methods based on adaptive graph convolution.(1)This thesis proposes a spatial and temporal embedding-based adaptive graph convolution traffic flow prediction model.Existing graph convolution-based traffic flow prediction models usually set the traffic network as a fixed graph structure,however,the spatial dependencies of traffic nodes will change dynamically over time due to the influence of morning and evening peaks.In order to solve this problem,this model models the structure of the traffic network as a time-varying directed graph,learns the dynamic correlation between traffic nodes from node features and temporal features by means of spatiotemporal embedding,and uses the bidirectional graph convolution structure to process the spatial information of inflow and outflow nodes.For the modeling of temporal relationships,this thesis designs a long short-term temporal convolutional network,which can learn the short-term and long-term dependencies of traffic flow data in the temporal dimension.In addition,a feature transformation module based on attention mechanism is designed to reduce the error propagation problem of long-term prediction by establishing a direct relationship between future data and historical data.Comparing the proposed model with several classic models in traffic flow prediction algorithms,the experimental results show that the proposed model achieves the best prediction accuracy on four large public traffic flow datasets.(2)This thesis proposes an attention-based adaptive graph convolution traffic flow prediction model.The above model only considers the influence of the traffic network structure on the prediction results when modeling the spatial relationship,while ignoring the impact of traffic conditions on the spatial dependencies of nodes.To solve this problem,this thesis adds a spatial attention module to the above model,which adaptively learns the influence of traffic conditions on the spatial relationship of nodes from the hidden features of the data through a self-attention mechanism.For modeling temporal relationships,the proposed model designs a multi-kernel temporal convolutional network.By stacking multiple convolution kernels in the temporal convolution of each layer,the model can not only learn the short-term and long-term temporal dependencies of the data,but also capture the temporal patterns of different scales implied in data.The experimental results on four large public traffic flow datasets show that the prediction accuracy of the model is further improved compared with the above proposed spatial and temporal embedding-based adaptive graph convolution traffic flow prediction model.