A3T-GCN: Attention Temporal Graph Convolutional Network for Traffic Forecasting

TLDR

Accurate real‑time traffic forecasting is essential for intelligent transportation systems but remains difficult due to complex spatial and temporal dependencies across road networks and varying importance of distant time points. This study introduces the A3T‑GCN, an attention‑based temporal graph convolutional network designed to jointly model global temporal dynamics and spatial correlations in traffic flows. A3T‑GCN combines gated recurrent units to capture short‑term trends, graph convolutional layers to encode road‑network topology, and an attention mechanism that weighs time points to assemble global temporal information. Experiments on real‑world datasets show that A3T‑GCN achieves 2.51–46.15 % and 2.45–49.32 % RMSE reductions, and 0.95–89.91 % and 0.26–10.37 % accuracy gains over baselines on SZ‑taxi and Los‑loop, respectively.

Abstract

Accurate real-time traffic forecasting is a core technological problem against the implementation of the intelligent transportation system. However, it remains challenging considering the complex spatial and temporal dependencies among traffic flows. In the spatial dimension, due to the connectivity of the road network, the traffic flows between linked roads are closely related. In the temporal dimension, although there exists a tendency among adjacent time points, the importance of distant time points is not necessarily less than that of recent ones, since traffic flows are also affected by external factors. In this study, an attention temporal graph convolutional network (A3T-GCN) was proposed to simultaneously capture global temporal dynamics and spatial correlations in traffic flows. The A3T-GCN model learns the short-term trend by using the gated recurrent units and learns the spatial dependence based on the topology of the road network through the graph convolutional network. Moreover, the attention mechanism was introduced to adjust the importance of different time points and assemble global temporal information to improve prediction accuracy. Experimental results in real-world datasets demonstrate the effectiveness and robustness of the proposed A3T-GCN. We observe the improvements in RMSE of 2.51–46.15% and 2.45–49.32% over baselines for the SZ-taxi and Los-loop, respectively. Meanwhile, the Accuracies are 0.95–89.91% and 0.26–10.37% higher than the baselines for the SZ-taxi and Los-loop, respectively.