A Granger causality-guided multi-graph transformer framework for traffic flow forecasting

Abstract

Accurate traffic forecasting is challenging because of the difficulty of capturing time-varying propagation and multi-scale spatio-temporal interactions. Most existing deep learning models only learn correlations rather than directional dependencies, which limits interpretability and robustness under dynamic traffic conditions. To address these limitations, this study proposes a Granger Causality-Guided Multi-Graph Transformer (GC-MGT) framework that incorporates Granger causality inspired directional predictive dependencies into spatio-temporal modeling. The framework consists of four key modules: (1) a hierarchical dynamic dependency module that combines a Granger predictive dependency prior with nonlinear neural correction to model directional influence; (2) an adaptive multi-graph fusion mechanism that unifies physical topology, semantic similarity, and directional dependency priors through learnable state-dependent weights; (3) a graph-guided multi-scale Transformer that improves both local and global temporal dependency learning; and (4) an intervention assessment conducted using the Simulation of Urban Mobility (SUMO) microscopic traffic simulator, which evaluates whether the learned dependencies are consistent with traffic propagation. Experiments on three real-world datasets demonstrate that GC-MGT consistently reduces MAE and RMSE compared with state-of-the-art baselines, while improving directional dependency structures aligned with simulated propagation. At the 60-min forecasting horizon, relative to PDFormer, the strongest baseline, GC-MGT reduces MAE by 13.3, 21.3, and 12.0%, and RMSE by 3.2, 24.5, and 19.2% on three datasets, respectively. These findings demonstrate that GC-MGT links data-driven forecasting with interpretable directional dependency modeling. This framework could be integrated with digital twins or traffic simulators for proactive traffic management and reinforcement learning control strategies.