基于XGBoost模型的城市道路实时交通事故风险预测研究

Study on Predicting Real-time Traffic Accident Risk of Urban Road Basedon XGBoost Model

我国城市道路的交通事故数量呈上升态势,在产生大量伤亡的同时,也对整个城市交通造成重要影响。基于卫星单点定位数据提取并评价了实时事故安全替代指标,建立了面向全路段和分路段的城市道路实时事故风险预测模型,用于动态事故风险预测。基于实时事故安全替代指标构建了XGBoost的全路段和分路段的事故风险预测模型。经过模型对比发现,SMOTE方法能够有效提高模型的预测准确率。为了排除测试集和训练集划分对模型效果的干扰,并测试模型的鲁棒性,采用200次划分不同测试集和训练集的方法对模型的鲁棒性进行了检验。经过验证,XGBoost的鲁棒性较好。结果表明:事故上游路段的安全替代指标比下游的重要程度高,表明上游路段的替代指标能够更好地用于实时事故风险的预测;在时间维度上,事故发生前5~10 min的单点定位轨迹特征数据,相比于前10~15 min的重要程度高,说明事故前5~10 min的数据能够更好地解释实时事故风险;基于模型训练集的预测效果,在模型内部进行风险阈值的选取,风险阈值是模型内部的参数;测试集较好的预测结果也证明了阈值选取的合理性。XGBoost模型能够进一步拓展公安交通警察日常应用的主动防控系统的开发,实现对重点道路的实时风险预测和安全管控,为今后的实时主动安全管理技术的发展奠定基础。

The number of traffic accidents on urban roads in China is increasing,which not only causes a large number of casualties,but also has great influence on the entire urban traffic.Based on the satellite single-point positioning data,the real-time accident safety substitute indicators are extracted and evaluated,and the real-time accident risk prediction model for whole road section and sub-road section is established for dynamic accident risk prediction.Based on the real-time accident safety substitution indicators,the accident risk prediction model for the whole road section and sub-road section is established by using XGBoost.By modelling comparison,it is found that the prediction accuracy of the model can be improved effectively by using SMOTE method.In order to eliminate the interference of the division of testing sets and training sets on the model effect and test the robustness of the model,the robustness of the model is tested by using the method of dividing different testing sets and training sets for 200 times.It is verified that XCBoost has good robustness.The result shows that(1)the safety substitute indicators of the upstream section of the accident are more important than those of the downstream section,indicating that the substitute indicators of the upstream section can be beter used for real-time accident risk prediction;(2)in terms of time dimension,the characteristic data from single-point positioning track of 5-10 min before the accident are more important than those from 10 to 15 min before the accident,indicating that the data of 5-10 min before the accident can better explain the real-time accident risk;(3)based on the prediction effect of the model training set,the risk threshold is selected inside the model,and the risk threshold is the internal parameter of the model;(4)the rationality of the threshold selection is proved by the better prediction result of the test set.The development of the active prevention and control system for the daily application by the public security traffic police can be further extended by using XGBoost model,the real-time risk prediction and safety control on key roads can be realized,and the foundation for the development of real-time active safety management technology in the future can be laid.