Traffic Flow Prediction Method Based on Multi-scale Feature Learning and Long-range Dependency Capturing

Aiming at spatio-temporal heterogeneity in traffic flow data that leads to inadequate periodic feature extraction and difficulty in capturing long-distance dependence among nodes in current graph convolutional network-based methods, this paper proposes a traffic flow prediction model based on multi-scale graph convolutional network and Transformer (TransMAG). It combines Transformer with multi-scale graph neural network for traffic flow prediction. Firstly, in this model, the hybrid convolutional pyramid network (HCPN) uses 1D convolution and Dilated convolution to extract temporal features within and between cycles respectively. Secondly, the constrained adaptive graph learner (CAGL) is designed to reduce training parameters. Additionally, Transformer is integrated with GCN to enable long temporal dependency capture through global attention mechanisms, while localized detailed correlations are learned based on graph structures. Finally, based on the above multi-scale periodicity feature and long distance dependence, TCN is used for traffic flow prediction. Experiment results on the real traffic flow datasets PeMS04 and PeMS08 show that TransMAG has better accuracy in dealing with the highly dynamic and complex dependence of urban traffic flow prediction scenarios.