Spatiotemporal fuzzy-graph convolutional network model with dynamic feature encoding for traffic forecasting

Accurate traffic forecasting is challenging, owing to the complex spatial dependence of traffic networks and the dynamic time dependence of traffic patterns. In this study, a novel spatiotemporal fuzzy-graph convolutional network model with dynamic feature encoding is proposed to realize accurate traffic forecasting. The proposed model combines graph convolutional network and long short-term memory network to extract complicated spatiotemporal dependence features of traffic data. Furthermore, a new graph generation method based on the fuzzy C-means clustering is designed to enhance the representation ability of the spatial dependencies between stations in a traffic network. Moreover, to make the graph convolutional network fully consider both global and local spatiotemporal dependency relationship between the stations in the process of convolution operation, a new node feature construction method is proposed. Finally, the forecasting performance of the proposed model is verified on three real-world traffic datasets. The experimental results demonstrate that the proposed model outperforms other baseline models in terms of both spatiotemporal feature extraction and long short-term forecasting.