Traffic speed prediction network based on multi-view spatio-temporal graph convolution network

Traffic speed prediction is the theoretical basis of building intelligent transportation. In urban road network data, the evolution pattern of traffic characteristics fluctuates and is implicit. Moreover, most models that utilise attention mechanisms overlook local information and easily accumulate irrelevant features, decreasing prediction accuracy. To enhance the ability of traffic speed prediction models to learn the complex spatiotemporal characteristics of urban road networks, a Dynamic Temporal Attention Network (DTAN) is proposed. Firstly, the Multi-View Temporal Decomposition Layer is constructed to transform global node correlations, long-term temporal features, stable fluctuations, and anomalous fluctuations into embeddable graphs, capturing multi-view characteristics. Secondly, a recurrent gating mechanism is employed to more effectively assist dilated causal convolutions in filtering out irrelevant information, helping the model mitigate the issue of gradient explosion during multiple iterations. Finally, the Multi-Graph Fusion Graph Convolution is developed, which utilises the previously generated embeddable graphs in combination with an adaptive adjacency matrix to enrich the feature information accessible to the graph convolution. This compensates for the feature information missed by attention mechanisms due to noise, effectively integrating various potential traffic feature information. Experiments on public datasets METR-LA and PEMS-BAY demonstrate the model's validity, showing that the model has high prediction accuracy over a long prediction range. The model code can be found on https://github.com/ccy123q/code .