基于XGBoost的高速公路事故类型及严重程度预测方法

高速公路事故频发，而以往研究未能充分揭示交通流动态特性对事故类型与严重程度的影响。为此研究了基于动态交通流数据的高速公路事故类型与严重程度的预测方法。从高速公路门架数据中提取流量、密度、速度等交通流数据，同时考虑时间特征以及时间和空间不均匀性特征的数据，与事故数据相匹配构成全样本。建立了基于极端梯度提升树（extrem Gradient Boosting，XGBoost）算法的预测模型，预测事故是否发生、事故类型以及事故严重程度。分别考虑追尾事故和其他事故2种事故类型、有人员伤亡和仅财产损失2种事故严重程度，模型的结果表明：①上下游速度差大、低速、路段车流量大且频繁分流、合流条件下交通事故风险较高；②低速、路段车辆多且合流、分流交通量大、上下游速度差大的情况下发生追尾事故的风险更高；③路段车流量较少且追尾事故发生于周末或夜间可能会增大事故严重程度。将常用机器学习算法与XGBoost算法的预测效果进行对比，XGBoost事故类型预测模型与事故严重程度预测模型的ROC曲线下面积（Area Under Curve，AUC）分别达到了0.76和0.88——相比于序列Logistic、高斯朴素贝叶斯、线性SVM、随机森林以及神经网络等其他常用算法，平均分别提升了0.08和0.24。这表明基于XGBoost建立的模型具有较好的预测性能。研究结果为高速公路路段实时交通流状态预警提供了可靠手段，进而可以提升高速公路行车安全。

A Method for Predicting the Type and Severity of Freeway Accidents Based on XGBoost

Freeway accidents are frequent, and previous studies have failed to adequately reveal the effect of dynamic traffic flow on accident type and severity. This study focuses on a prediction method for types and severity of freeway accidents based on real-time traffic flow data. Traffic flow characteristics, including volume, density, and speed, are extracted from freeway gantry data. Simultaneously, temporal features and spatiotemporal non-uniformity features are considered. These data are then matched with accident data to constitute the full dataset for modeling. The model based on the extreme gradient boosting tree (XGBoost) algorithm is developed to predict the occurrence of accidents and accident types, and also to assess accident severity. Two types of accidents (i.e., rear-end collisions and other types of accidents) are considered and two levels of accident severity (i.e., injury or fatal accidents and proper-ty-damage-only accidents) are distinguished. The results indicate that: ①a higher risk of traffic accidents is associated with significant speed difference between upstream and downstream traffic, low speeds, high traffic volumes with frequent merging and diverging conditions; ②rear-end accidents are more likely to occur in situations with lower speeds, high traffic volumes with merging and diverging flows, and significant speed difference between upstream and downstream traffic; ③accidents involving rear-end collisions may result in higher severity when they occur on road segments with lower traffic volumes or occur during weekends or nighttime. The Area Under Curve (AUC) of the XGBoost-based models for accident types prediction and accident severity prediction reached 0.76 and 0.88 respectively. Compared with other commonly used algorithms such as Sequential Logistic, Gaussian Naive Bayes, Linear Support Vector Machine (SVM), Random Forest, and Neural Network, the XGBoost-based model demonstrates an average improvement of 0.08 and 0.24 in AUC values for predictions of accident types and accident severity. These results indicate that the XGBoost-based model exhibits better prediction performance. The research findings provide a reliable way for state warning of real-time traffic flow on freeway segments, which could be useful for improving driving safety.