Zone-specific real-time traffic conflict risk modeling for freeway tunnels: a CrossTabNet approach

Abstract

This study proposes a zone-specific, real-time traffic conflict risk modeling framework specifically designed for freeway tunnels. The framework integrates traffic conflict analysis, refined tunnel segmentation, and interpretable deep learning to address limitations in traditional collision data. Vehicle trajectory data are utilized to derive surrogate safety measures based on traffic conflicts. A refined five-zone tunnel classification—pre-entrance, entrance, interior, exit, and post-exit—is adopted by extending existing zoning frameworks. This facilitates more precise spatial attribution of risk patterns in real-time conflict analysis. To model complex, interdependent risk factors, a CrossTabNet architecture is developed. This innovative structure combines a feature interaction layer with a TabNet encoder, enabling the model to capture high-order nonlinear relationships between traffic variables while maintaining interpretability through sparse attention mechanisms. The proposed model demonstrates superior predictive performance compared to established machine learning and deep learning methods. Notably, zone-specific models significantly outperform a global model trained on all data, emphasizing the necessity of localized modeling for effective tunnel safety assessment. Global sensitivity analysis reveals that the standard deviation of upstream traffic flow consistently contributes positively to conflict risk across all zones, highlighting the critical role of flow variability. Other significant features vary by tunnel segment, reflecting distinct local dynamics. These findings provide valuable insights for implementing adaptive, zone-targeted traffic safety interventions in freeway tunnel environments.