Spatio-temporal aware graph convolution network for urban traffic speed prediction

Haitao Cheng, Xiaoying Tan, Ke Wang

Abstract: Urban traffic speed prediction is a fundamental task in Intelligent Transportation Systems. However, due to the highly nonlinear and dynamic spatio-temporal dependencies inherent in traffic flow, achieving timely and accurate traffic speed prediction, particularly for long-term forecasts, remains an open challenge. To address the simultaneous modeling of spatial heterogeneity and temporal variation in road network speed prediction, we propose a Spatio-Temporal Aware Graph Convolutional Network (STAGCN). Spatially, STAGCN integrates Markov random walk–based positional encodings with raw traffic speed features to enhance the representation of network connectivity, and employs a multi-head spatial self-attention mechanism and residual convolutional layers to capture both global node interactions and local spatial patterns. Temporally, STAGCN introduces a time window–aware position encoding matrix to adaptively reweight temporal embeddings and input sequences, enhancing the model’s ability to capture periodic traffic patterns and distinguish peak-hour dynamics. A gated fusion mechanism and temporal self-attention further refine the model’s ability to distinguish evolving traffic trends. Experimental results demonstrate that STAGCN achieves state‑of‑the‑art performance in long‑horizon traffic speed prediction, validating its effectiveness in modeling spatiotemporal dependencies.

External IDs:dblp:journals/esi/ChengTW25