Research on an Adaptive Traffic Congestion Prediction Method Considering a Time-varying Network Topology

Predicting traffic congestion in traffic systems is beneficial for traffic management and reduces traffic risks. However, owing to traffic control, road construction, bad weather, natural disasters, and other reasons, the topology of road network traffic systems often changes; therefore, congestion prediction methods relying on a fixed road network topology are not effective. To solve this problem, this paper proposes a dual-adaptive graph convolution recurrent network architecture (DAGCRN) to predict traffic congestion in a dynamic time-varying road network topology. In this network, an adaptive auxiliary adjacency matrix is used to learn the static graph structure of a pre-defined road network, and the transferred information among the original connections is optimized dynamically, thereby overcoming the information uncertainty caused by changes in the road network topology. An adaptive embedding adjacency matrix is used to capture the hidden information of a predefined road network. This ensures the dynamic integrity of the road network topology. A gated recurrent unit is used to extract the temporal characteristic information of the traffic flow in road networks. Experimental results demonstrate that (1) the proposed DAGCRN can effectively capture and locate changes in the topology structure of a road network and still ensure the accuracy of congestion prediction when the topology changes; (2) compared with some common prediction models, the predictive accuracy is higher, especially in terms of long-term prediction, while overcoming changes in the structure of the road network; and (3) further results from a dual-adaptive function ablation experiment confirm that the dual-adaptive graph convolution structure with adaptive auxiliary adjacency matrix and adaptive embedded adjacency matrix has a strong adaptive ability to time-varying road networks, and the absence of any adaptive module would lead to a significant decline in model prediction performance. © 2022 Xi'an Highway University. All rights reserved.