A Study on Urban Traffic Speed Prediction Using Spatiotemporal Attention-Based Convolutional Neural Networks

Abstract

Urban traffic speed prediction requires effective modeling of road-network spatial dependencies and temporal dynamics. This study develops an attention-based spatiotemporal graph convolutional framework (AT-GCN-GRU) for multi-horizon traffic speed prediction. The model combines graph convolution for spatial feature extraction, GRU for temporal modeling, and a temporal attention mechanism for adaptively weighting historical hidden states. Experiments were conducted on the SZ-taxi dataset and further evaluated on METR-LA. On SZ-taxi, AT-GCN-GRU was compared with conventional and modern baselines, including T-GCN, ASTGCN, and AGCRN, over 15, 30, 45, and 60 min forecasting horizons. The proposed model remained competitive with the modern baselines and obtained lower mean RMSE than T-GCN across the four evaluated horizons, with relative reductions of approximately 1.4–2.9%. Additional ablation, missing-data robustness, attention-weight visualization, and computational-cost analyses were conducted. The results suggest that temporal attention can provide incremental improvements to the GCN-GRU framework with relatively limited additional computational overhead.