UrbanFlow-GST: A Dynamic Graph-Based Spatio-Temporal Fusion Model for Real-Time Traffic Speed Forecasting

Abstract

 Real-time and accurate traffic forecasting forms the basis of intelligent transportation systems, especially in densely populated areas where traffic conditions keep varying with time depending on complicated spatio-temporal correlations. Conventional as well as deep learning-based approaches tend to lack simultaneous consideration of spatio-temporal dynamics, thus rendering them inefficient in the practical setting. In light of this problem, this paper presents UrbanFlow-GST, an adaptive graph learning and attention-based modeling approach for traffic forecasting. The proposed system builds upon a graph and implements spatio-temporal transformer, using a dynamic adjacency mechanism that takes into account time-dependent associations among traffic sensors. For capturing temporal dynamics, the authors adopt the transformer architecture, enabling them to learn both shortand long-term traffic characteristics. Finally, a unified method is developed for combining spatial and temporal features to make multi-horizon predictions. To test the effectiveness of the proposed framework, experiments were conducted using the METR-LA dataset consisting of real-world traffic speeds collected via sensors installed at 207 locations on highways of Los Angeles. The model outperformed competing baselines at all horizons under evaluation. In particular, the authors report MAE, RMSE, and MAPE values of 2.52, 4.90, and 8.1%, respectively, for 15-minute prediction, while the same metrics for 60-minute forecasting are 3.20, 6.80, and 10.3%. The findings suggest high reliability of UrbanFlow-GST