GraphCast-Lite Model for Lane-Level Congestion Spillback Prediction in Urban Arterial Road Networks

Abstract

Accurate lane-level congestion spillback prediction is critical for proactive traffic management in urban arterial networks, yet existing traffic forecasting methods often rely on link-level states and fail to capture the heterogeneous interactions among lane storage, turning movements, signal timing, and downstream capacity. This study proposes a GraphCast-Lite model for lane-level spillback prediction by constructing a directed lane graph in which each lane is represented as an operational node. Multi-source traffic information, including queue length, occupancy, speed, arrival pressure, remaining storage, turning movement, and signal-phase context, is integrated through a lightweight spatio-temporal graph propagation mechanism to estimate future queue evolution and spillback risk. Experiments were conducted on a simulated arterial network containing 18 signalized intersections, 246 lane nodes, 412 directed movement edges, and 30-second traffic-state intervals. Compared with traditional forecasting models and advanced graph-based baselines, the proposed model achieved the highest spillback prediction performance within the 15-minute horizon, increased the average warning lead time to 6.8 minutes compared with 4.3 minutes for STGCN and 5.1 minutes for Graph WaveNet, and reduced false alarms under isolated saturation and downstream blockage scenarios. The model maintained 91.2% of the original F1 score in cross-corridor transfer experiments and achieved inference times of 18.4 ms on a workstation and 46.7 ms on an edge controller. The proposed framework provides an efficient and interpretable solution for lane-level traffic risk prediction and supports real-time urban traffic control applications.