ASTDGCN: An Adaptive Spatial-Temporal Diffusion Graph Convolutional Network for Traffic Forecasting

Abstract

Precise traffic flow prediction functions as the fundamental cornerstone for the efficient, safe, and reliable operation of intelligent transportation systems (ITS). It not only provides data-driven support for key applications, for instance, real-time traffic signal regulation, proactive congestion mitigation, and personalized route optimization, but also exerts a critical effect on reducing traffic accidents and improving overall urban travel efficiency. However, the traffic system belongs to a complex system, with spatio-temporal dynamics that are both intricate and variable, ranging from predictable fluctuations during morning and evening peak hours to localized propagation effects caused by accidents, as well as seasonal variations and significant nonlinear characteristics. These factors collectively pose substantial challenges to building accurate and reliable prediction models, creating a long-standing technical bottleneck in this field. With the aim of solving the dilemma that existing methods are hardly able to capture traffic flow’s spatio-temporal dependence effectively, we advance an adaptive spatial–temporal diffusion graph convolutional network (ASTD-GCN) for a traffic flow prediction model that integrates adaptive graph learning, diffusion convolution, and bi-directional long short-term memory network (Bi-LSTM) with attention mechanism. The model dynamically constructs the correlation between the nodes of the transportation network through the adaptive graph learning module and accurately describes the spatial topology. The diffusion convolution module realizes multi-order spatial information diffusion based on graph structure, which realizes the effective extraction of the traffic flow’s spatial dependence features. The Bi-LSTM module incorporating the attention mechanism captures the historical and future context information of traffic flow simultaneously through the bidirectional loop structure and the temporal attention mechanism, and strengthens the key time step features. Experimental results on -world traffic datasets PEMS03, PEMS04, PEMS07, and PEMS08 indicate that our proposed model exhibits better predictive precision in traffic flow forecasting tasks than baseline counterparts.