CTSG-Net: A Fine-Grained Intersection Queue Length Prediction Model at a Network Scale

Accurate and robust short-term traffic prediction is critical for the advancement of Intelligent Transportation Systems (ITS). With the accessibility of more fine-grained traffic data, intersection queue lengths prediction with both high spatial and temporal resolution have become feasible, offering significant potential for applications such as dynamic and adaptive traffic management in real-time systems. However, predicting these fine-grained queue lengths is challenging due to their rapid fluctuations and the zero-inflated, long-tailed nature of the data. To address these challenges, this paper proposes CTSG-Net (Cyclic Temporal-Spatial Graph Convolution Network), capable of predicting lane-level, seconds-by-seconds queue lengths at a network scale. CTSG-Net incorporates a cyclic feature extraction module that utilizes signal cycle information to capture periodic traffic patterns, and a graph topology learning module that combines static and dynamic graph layers to model lane-specific spatial dependencies. Additionally, CTSG-Net employs a weighted loss function to further address the zero-inflated nature of queue length data. Experiments have been conducted on real-world traffic data from Yizhuang, Beijing, during both off-peak and evening peak periods. The results demonstrate that the proposed CTSG-Net consistently achieves state-of-the-art performance. Specifically, for off-peak and evening peak scenarios, the proposed CTSG-Net demonstrates substantial improvements of up to 8.04%, 28.23%, 24.68% in MAE and 6.54%, 27.75%, 22.12% in RMSE across dataset of all queue lengths, non-zero queue lengths, and large queue lengths ( $\ge 100$ m), respectively. The source code is available at: https://github.com/cherryh2021/CTSG-Net.git