基于GSK-AdaBoost-LightGBM的交通事故死亡人数预测研究  

Prediction of Traffic Accident Death Toll Based on GSK-AdaBoost-LightGBM

作　　者：纪俊红[1] 昌润琪 温廷新[2] JI Junhong;CHANG Runqi;WEN Tingxin(School of Safety Science and Engineering,Liaoning Technical University,Huludao 125105,China;System Engineering Institute,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学安全科学与工程学院,辽宁葫芦岛125105 [2]辽宁工程技术大学系统工程研究所,辽宁葫芦岛125105

出　　处：《安全与环境工程》2021年第1期24-28,共5页Safety and Environmental Engineering

基　　金：辽宁省教育厅基金项目(LJ2019JL016)。

摘　　要：交通事故预测研究有助于降低交通事故发生的概率。为提高交通事故预测模型的精度,首先应用网格搜索法获得LightGBM模型的最优超参数,并进行5折交叉验证提升模型的抗拟合能力,建立优化的LightGBM(GSK-LightGBM)模型;然后基于AdaBoost算法训练多个GSK-LightGBM模型,加权组合得到AdaBoost-LightGBM增强集成模型,并采用网格搜索法结合5折交叉验证,实现了AdaBoost-LightGBM模型的参数优化,构建了GSK-AdaBoost-LightGBM模型;最后基于采集到的1953—2018年我国道路交通事故死亡人数相关样本数据训练模型,得到基于GSK-AdaBoost-LightGBM的交通事故死亡人数预测模型,并引入均方误差、平均绝对误差和平均绝对百分误差3项评估指标评估了模型的预测性能,探究了模型的优化效果。结果表明:GSK-AdaBoost-LightGBM模型的3项评估指标值分别为0.014、0.00035和0.077,低于LightGBM模型、GSK-LightGBM模型和AdaBoost-LightGBM模型的评估指标值,说明该模型的预测精度较高,且明显优于LightGBM模型、GSK-LightGBM模型和AdaBoost-LightGBM模型。Traffic accident prediction research can reduce the probability of traffic accidents.To improve the accuracy of the traffic accident prediction model,first of all,this paper applies the Grid search to optimize the hyperparameters of LightGBM model,improves the anti-fitting of the model by using the 5-fold cross-validation,and establishes the optimized LightGBM model(GSK-LightGBM);and then trains multiple optimized LightGBM model based on AdaBoost,and obtains the AdaBoost-LightGBM model enhanced integration model by weighted combination;next,the paper optimizes the parameters of AdaBoost-LightGBM and constructs the GSK-AdaBoost-LightGBM model by combining the Grid search with 5-fold cross-validation.The paper trains the model based on the statistical index values of road traffic accidents in China from 1953 to 2018,and obtains the traffic accident fatalities prediction model based on the GSK-AdaBoost-LightGBM.Fianlly,the paper evaluates the prediction performance of the model by three evaluation index,i.e.,average absolute error,mean square error and average absolute percentager error.The results show that the average absolute error,mean square error,and average absolute percentage error of the model based on GSK-AdaBoost-LightGBM are 0.014,0.00035,and 0.077,respectively,which are lower than the index values based on the LightGBM,GSK-LightGBM,and AdaBoost-LightGBM models.This shows that the prediction accuracy of the traffic accident fatality prediction model based on GSK-AdaBoost-LightGBM is higher,and it is significantly better than that based on the LightGBM,GSK-LightGBM,and AdaBoost-LightGBM model.

关 键 词：交通事故 死亡人数预测 ADABOOST LightGBM 交通安全 

分 类 号：X928[环境科学与工程—安全科学] X951