Although tunnels are critical traffic nodes in freeway networks, academic research addressing their real-time traffic safety is noticeably lacking. This study proposes a novel dual-task approach to analyze causal precursors and predict real-time collision risks in freeway tunnels. Unlike traditional models, which often trade off between predictive accuracy and causal depth, the proposed approach achieves both high causal interpretability and predictive performance. The approach utilizes a structural agnostic model (SAM) to discover causal precursors of freeway tunnel collisions using observational data. The collision causal graph data is then constructed based on the causal relationships identified by SAM. Additionally, the Causal Directed Graph Convolutional Networks (CDGCN) model is developed to capture causal relationships in the graph for real-time collision prediction. Utilizing freeway tunnel collision analysis data collected from the Caltrans Performance Measurement System, the approach performs dual tasks: identifying causal precursors of collisions and predicting future collisions in real time. The SAM results reveal five critical causal precursors influencing the likelihood of collisions. Comparative analyses with existing interpretable machine learning models show similarities between the causal precursors of tunnel collision risk revealed by SAM and the important correlation precursors identified by the comparative models. However, correlation is not the same as causation. When tested against current state-of-the-art real-time collision predictive models, the proposed CDGCN model demonstrates superior accuracy, especially on datasets containing causally relevant precursors, highlighting the potential of this approach for feature selection and risk prediction. This advancement not only provides a practical framework for mitigating collision risks in freeway tunnels but also makes a significant contribution to traffic safety research.
Read full abstract