D-STANet: A delay-enhanced spatio-temporal attention network for traffic prediction

Abstract

With the accelerating pace of global urbanization, accurate traffic flow prediction has become crucial for alleviating congestion and optimizing resource allocation. However, existing methods often fail to effectively capture the complex spatio-temporal dependencies inherent in traffic data, which limits predictive accuracy. To address this challenge, we propose the D-STANet, which is an innovative traffic flow prediction model that integrates spatio-temporal attention mechanisms with a delay-aware module. Specifically, D-STANet leverages the spatio-temporal attention mechanism to adaptively select the most relevant features across different temporal and spatial scales, thereby capturing complex spatio-temporal dependencies. Additionally, the proposed delay-aware module is designed to model the temporal delay effects in traffic flow data, as predictions are not only dependent on current flow data, but also influenced by fluctuations in past traffic states. Furthermore, D-STANet incorporates a graph attention mechanism to enhance its ability to respond to dynamic changes. This module automatically adjusts the weight of each node in the graph based on the degree of association between nodes in the traffic flow data, further improving the model’s ability to capture traffic flow variations. Experimental results demonstrate that D-STANet outperforms all baseline models across multiple metrics, particularly on the HZME dataset, where its superior ability to model spatio-temporal dependencies is evident. Specifically, D-STANet achieves improvements of 31.71%, 20.48% and 5.06% in MAE, RMSE and MAPE, respectively, compared to DMSTGCN. The model’s exceptional performance in sparse traffic networks further underscores its robustness and reliability in complex traffic environments.