A real-time conflict risk prediction modeling based on causal inference and graph-model: Applied to multi-configuration expressway diversion zones

• A full-chain conflict prediction framework is proposed to reveal risk mechanisms and enable real-time collision prediction. • Conflict risk is quantified using 2D time-to-collision and time spent in conflict metrics across diverse scenarios. • Deep causal inference uncovers four types of traffic variables as direct causes of conflict risks. • A causal graph-enhanced Graph Attention Networks predicts risk by leveraging reconstructed causal dependencies. • Trajectory data from real-world expressway diversion zones validates the framework’s generalizability and accuracy. Integrating causal analysis with conflict risk prediction modeling aids traffic managers in understanding and controlling traffic risks better. This study introduces a real-time conflict risk prediction modeling framework based on causal inference and graph neural network theory, an innovative full-chain approach that integrates causal discovery, data reconstruction, and real-time prediction for conflict risk prediction. The framework employs the structural agnostic modeling (SAM) algorithm to discover causal relationships among conflict risk covariates from traffic data. Based on these causal relationships (represented by the causal graph) and considering the spatiotemporal characteristics of conflict risk, traffic variables are reconstructed into causally structured data. Finally, the graph attention networks (GATs) is employed to process the causal-structured data for real-time conflict risk prediction. The framework was validated in multi-configuration diversion zones of urban expressways in Wuhan, China. Approximately 10 h of spatiotemporal trajectory data were collected from five diversion zones using drone-based vehicle target recognition technology. A 2-D conflict incident extraction method was designed and implemented, with an indicator termed “time spent in conflict (TSC)” used to evaluate the risk. Results indicate that this framework significantly outperforms other models in prediction accuracy and effectively reveals direct causal relationships between four types of traffic variables and traffic conflict risks. Further sensitivity analysis shows that both oversampling techniques and the utilization of prior interval traffic risk as input can significantly enhance model performance. Considering multi-configuration factors also positively impacts model performance. The conflict prediction framework has the potential to be integrated into existing intelligent roadside traffic risk warning systems. This application enhances risk perception on highways and expressways, thereby assisting the warning system in automatically releasing risk information and implementing intervention measures.