基于可解释机器学习的信号交叉口左转冲突风险建模与量化分析OA

[Background]At signalized intersections,left-turn-related traffic conflicts exhibit dynam-ic response characteristics and have a significant impact on the safe operation of intelligent transpor-tation systems.Typical left-turn traffic conflicts are mainly categorized into those between pedestri-ans and left-turning vehicles,and opposing through-traffic and left-turning vehicles.[Objective]To explore potential factors that influence left-turn traffic conflicts and to clarify the underlying mecha-nisms through which these factors affect both types of conflicts.[Data]Observational data collected from September 13～19,2019,at six representative signalized intersections in Bellevue,Washington,USA.[Method]Multiple machine-learning models were employed to model the left-turn traffic con-flict risk(conflict frequency),and Bayesian Optimization was applied to determine the optimal hy-perparameters of each model.Among these,the Extreme Gradient Boosting(XGBoost)model dem-onstrated superior performance in risk modeling both types of left-turn traffic conflicts.Shapley Ad-ditive Explanations(SHAP)and Accumulated Local Effects(ALE)were used to interpret XGBoost outputs,systematically quantifying the mechanisms through which traffic-flow characteristics,signal-control features,and other potential factors influence left-turn traffic conflicts.[Conclusion]Larger traffic volumes at signalized intersections are associated with a higher left-turn traffic conflict risk;the effect of the traffic volume is highly nonlinear.Different signal phase-control strategies play markedly distinct roles in moderating the conflict intensity;among these,protected phase-control mechanisms can effectively suppress the occurrence of conflicts.Moreover,when the interval of the Flashing Don't Walk(FDW)signal exceeds 20 s,it can result in higher pedestrian volumes—signifi-cantly amplifying the risk of left-turn traffic conflicts.[Application]These findings can provide a reference for optimizing signal-control strategies and geometric intersection designs to enhance the safety of urban signalized intersections.