Building Urban Traffic Flow Prediction Model Using Spatio-Temporal Graph Convolutional Networks

Abstract

To address the inadequacy of existing models in capturing spatial position information and multi-scale periodic patterns in traffic flow, this paper proposes a Position-Aware Spatio-Temporal Graph Convolutional Network (PASTGCN). The model incorporates a spatio-temporal position embedding module that encodes geographic coordinates and topological attributes into spatial embeddings, combined with sinusoidal encoding for periodic temporal representations. A dual graph learning mechanism fuses a static distance graph based on a Gaussian kernel with an attention-based dynamic graph, while spatial-aware dilated causal convolution enables temporally modulated spatial feature injection. Multigranularity periodic branches further enhance the modeling of daily and weekly patterns. Extensive experiments on the METR-LA dataset show that PASTGCN achieves an RMSE of 2.53 (veh/5 min) and an R2 of 0.982, outperforming several contemporary baselines including DCRNN, STGCN, Graph WaveNet, ASTGCN, AGCRN, and STAEformer. Ablation studies confirm the contribution of each module, with the dynamic graph learning alone reducing RMSE by 0.34. The model also maintains high efficiency with an inference latency of 18.7 ms and a throughput of 53.5 samples/second, demonstrating its suitability for real-time urban traffic prediction. Beyond intelligent transportation applications, the proposed position-aware spatio-temporal modeling strategy provides a practical reference for information fusion and dynamic state estimation in large-scale wireless sensing networks and electromagnetic signal-driven monitoring systems, supporting future intelligent perception and communication infrastructures. This high accuracy and low-latency capability also makes PASTGCN particularly promising for supporting intelligent logistics scheduling in time-sensitive industries.