Incorporating Connectivity Information in Spatiotemporal Graph Convolution Model for Traffic Flow Prediction

Abstract Traffic flow prediction is extremely important for intelligent transportation systems. Accurate traffic flow prediction not only depends on the traffic flow data of the predicted node itself, but also on the traffic flow data of other nodes. These nodes can be mainly divided into the adjacent nodes of the predicted node, the multi-order adjacent nodes corresponding to different prediction durations, and other nodes. In this paper, by constructing an incidence matrix to integrate these nodes' information, the strong correlation between the adjacent nodes and the predicted node is first verified by the Pearson correlation coefficient. Secondly, the multi-order adjacent nodes corresponding to different prediction durations are calculated based on the distance and speed between different nodes. To address the influence of other nodes, a global correlation measurement function is used to evaluate the overall traffic information in Euclidean distance and Manhattan distance, enhancing the learning ability of global information. The incidence matrix is constructed by integrating the information of adjacent nodes, multi-order adjacent nodes, and global information. Finally, as a supplement to the multi-order adjacent information, traffic speed is added as external information. In this paper, a spatiotemporal graph convolution model constructed by Graph Convolutional Network (GCN) and Gate Recurrent Unit (GRU) is used to model the spatiotemporal correlation of data. The experimental results on the London M25 highway dataset show that the proposed model achieves good performance compared with other baseline methods.