Analyzing extreme muti-vehicle rear-end collision risks in adverse weather through Generalized Pareto Regression Trees.

Abstract

Objectives: The study seeks to explore rear-end collision risks in multi-vehicle car-following scenarios under adverse weather conditions by proposing an integrated framework.

Methods: The integrated framework is applied to a case study of three-vehicle car-following scenario in Norway without loss of generality. For identifying car-following groups with extreme collision risks, the collision risk of each group in the raw dataset is evaluated using an extended probabilistic driving risk field. Quantitative collision risks are analyzed to fit the Generalized Pareto distribution, and high-risk scenarios screened via mean residual life plots and threshold stability plots. To determine risk-contributing factors, Generalized Pareto Regression Trees (GPRT) are constructed to pinpoint significant influences on rear-end collision risks. By integrating the classification and regression trees with extreme value theory, the GPRT discards data assumptions and covariate continuity requirements of most extreme value analysis (e.g., extreme quantile regression). Moreover, the GPRT not only identifies the hierarchical structure of variables affecting rear-end collision risks but also determines risk-impact thresholds for covariates, offering superior interpretability and engineering applicability.

Results: The results show that revealed risks conform well to the Generalized Pareto distribution, allowing for the formulating Generalized Pareto regression trees. Compared to the Generalized Additive Model (GAM) and Negative Binomial Regression (NBR) methods, the GPRT approach demonstrates superior performance in balancing risk fitting accuracy and model complexity. Vehicle speeds, weights, and headways emerge as critical factors for collision risks under clear, rainy, and snowy conditions. As weather conditions deteriorate from clear to rainy or snowy, the influence of vehicle speed and weight diminishes, while the influence of headway and road surface conditions becomes more pronounced. Collision risks are high on sunny days, regardless of whether the middle vehicles of three-vehicle groups are light or heavy vehicles.

Conclusions: The integrated evaluation framework developed in this study provides a tool for car-following safety assessment under extreme weather conditions.