DUTNG: Employing dynamically updating traffic network graph for short-term traffic flow prediction

The complex topology of actual road networks and the interlinked nature of traffic flow with spatiotemporal factors pose challenges to traditional node-static correlation models. Therefore, the DUTNG model—a short-term traffic flow prediction model leveraging dynamically updated traffic network graphs—has been proposed to address these issues. Initially, a parameterized dynamic graph learning module is employed for real-time road condition modeling, enhancing the expressiveness of the short-term traffic network. Subsequently, a dynamic time extraction network is integrated into the dynamic graph to address the extraction of features from traffic flows across various time ranges. Additionally, stacked dynamic spatial extraction network modules resolve challenges related to varying regional influences. Finally, these three modules are sequentially integrated, enhancing the model’s ability to extract spatiotemporal correlations. Experimental studies on real-world datasets (California highway network) demonstrate that this model significantly outperforms machine learning models. Compared to recent benchmark models, its performance improves by 5.3% to 18.7%.