Multi-view Dynamic Graph Convolutional Traffic Prediction with Periodic Offset

The existing traffic prediction methods are often based on strong assumptions, that is, spatial dependence is often based on direct connection relationships between adjacent nodes, and time dependence is strictly periodic. However, in real roads, traffic characteristics are influenced by global spatial and node personalized local spatial correlations, and their temporal dependence is not strictly periodic. To address the above issues, a periodic offset multi-view dynamic graph convolutional network PO-SMVGCN is proposed. Firstly, PO-SMVGCN optimizes the node2vec walk algorithm by associating the walk strategies of different types of nodes with connectivity, generating a local dynamic graph of personalized node effects, and fusing it with the global spatial representation of the calculated correlation coefficient between nodes to construct a multi view graph convolution. Secondly, a periodic offset attention gated recurrent unit is designed to capture the temporal dependence of nodes and their periodic offset features, and the historical spatio-temporal features are converted into future representations by a temporal attention transformation unit to achieve accurate traffic prediction. Finally, comparative experiments were conducted with nine baseline methods on three real-world traffic datasets of PEMS, and the results showed that the proposed PO-SMVGCN method exhibited significant performance advantages compared to other traffic prediction methods, further verifying the effectiveness of this method in achieving accurate traffic flow prediction in complex and changing road environments.