Dynamic traffic subarea partitioning: a graph convolutional network approach

This study aims to enhance the performance of regional coordinated traffic control and optimize the dynamic partitioning of traffic control subareas. A novel GCN-HAC approach is proposed, which integrates Graph Convolutional Network (GCN) and Hierarchical Agglomerative Clustering (HAC) algorithms. The method utilizes Path–Cycle–State correlation features and road network topology as inputs. Subarea partitioning results are continuously evaluated and adjusted over time based on homogeneity, correlation, and inter-subarea difference metrics. Applied to a case study in a local area of Chengdu, the GCN-HAC method demonstrated superior partitioning effectiveness compared to several benchmark algorithms, including the Newman algorithm, Density Peak Clustering (DPC), Normalized Cut (NCut), HDBSCAN, and Fuzzy C-Means Clustering (FCM). The algorithm also exhibited strong interpretability and broad applicability. Moreover, its operational efficiency supports quasi-real-time application and scalability to large-scale road networks. During evolution analysis across morning peak (08:00–10:00), weekday off-peak (10:00–17:00), and evening peak (17:00–19:00) periods, the number and configuration of subareas dynamically adapted to actual traffic flow characteristics, leading to more reasonable and realistic partitioning. The proposed GCN-HAC framework provides a robust, adaptive solution for dynamic traffic subarea division, with potential to improve real-time traffic management systems and support the development of large-scale intelligent transportation infrastructures.