Facing the challenge of worsening highway traffic congestion, precise real-time forecasting is crucial for intelligent traffic management.However, traditional models struggle to effectively capture the complex spatio-temporal dependencies and dynamic propagation delays inherent in traffic data.To address this, this paper proposes a hybrid architecture that integrates graph neural networks with transformers.Through a dynamic graph attention mechanism and a delay-aware module, it significantly enhances the modelling capabilities for long-range spatial correlations and temporal propagation effects.Experiments on public datasets such as performance measurement system 04 and performance measurement system 08 demonstrate that the proposed model reduces the mean absolute error by 6.2%-9.2%compared to existing state-of-the-art methods within the 15-60 minute prediction window, with particularly notable performance improvements during peak congestion periods.The framework presented here has the potential to provide a more reliable technical pathway for traffic state prediction, holding significant practical application value.
Cheng et al. (Thu,) studied this question.