Triple Dynamic Graph Convolutional Recurrent Network for Traffic Prediction

Effective traffic prediction is a critical component of traffic management, especially long-term traffic prediction, as it holds significance for urban traffic planning, traffic warning, people’s travel planning, etc. Traffic flow data usually include natural spatio-temporal characteristics, which means the spatial and temporal dependencies are both to be considered for building traffic prediction model. Hence, a triple dynamic graph convolutional recurrent network (TDGCRN) is proposed in this paper, in which a temporal segmentation-based triple spatio-temporal encoder-decoder module is used to capture correlation features of different periods contained in traffic dataset, such as hourly, daily, and weekly cycles; and then a dynamic fusion module is introduced to adaptively learn the weights of different periods to effectively fuse these different cycle features; finally a congestion index-based adjacency matrix update module is utilized to model the topology of graphs to capture the dynamic topology characteristics of the road network. In addition, the congestion index is presented to detect the topology changes caused by traffic congestion to reduce the number of dynamic graphs so as to save computational cost in the training process on our self-collected dataset WH-CN. Experiments on the public dataset (PEMS-BAY and METR-LA datasets) and WH-CN dataset show that our model achieved average improvements of 0.98%, 1.19%, and 1.17% in MAE, RMSE, and MAPE metrics by comparison with the suboptimal baseline, respectively. And on the WH-CN dataset, it also reduced the average training time per epoch by reducing the frequency of dynamic graph generation, thereby reducing the computational cost.