FD-TGCN: Fast and dynamic temporal graph convolution network for traffic flow prediction

The traffic flow prediction has recently been challenged due to its complicated dynamic spatial–temporal features. In terms of temporal modeling, the dilated convolution used to model the temporal relationship consumes more training time. In terms of spatial modeling, traffic flow prediction results are affected not only by the dynamic connection spatial relationship, but also by the changes of traffic road structure, which is ignored by most methods. In order to address these concerns, we propose a new traffic flow prediction method which is called Fast and Dynamic Temporal Graph Convolution Network (FD-TGCN). FD-TGCN comprises a temporal module and a spatial module. In the temporal module, we propose a Fast Time Convolution Network (FTCN) to reduce the training time. The spatial module improves prediction accuracy by separately modeling dynamic connection spatial relationship and the change in the structure of the road. A series of experiments have shown that compared with the baseline models, our proposed method achieves an average accuracy improvement of 1.3% and 1.85% on two datasets, respectively, while saving an average training time of 293.55%. •Apply a novel graph neural network structures for traffic forecasting.•Design an efficient traffic flow prediction method, named FD-TGCN.•An original approach FTCN is created to save the time consumption while ensuring accuracy.•Propose a dynamic adaptive convolution matrix for learning the influence of road structure changes.