基于LSTM-BF的高速公路交通事故风险模型  

Study on risk model of highway traffic accidents based on LSTM-BF

作　　者：熊晓夏 刘擎超[2] 沈钰杰[2] 蔡英凤[2] 陈龙[2] XIONG Xiaoxia;LIU Qingchao;SHEN Yujie;CAI Yingfeng;CHEN Long(School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang Jiangsu 212013,China;Automotive Engineering Research Institute,Jiangsu University,Zhenjiang Jiangsu 212013,China)

机构地区：[1]江苏大学汽车与交通工程学院,江苏镇江212013 [2]江苏大学汽车工程研究院,江苏镇江212013

出　　处：《中国安全科学学报》2022年第5期170-176,共7页China Safety Science Journal

基　　金：国家自然科学基金青年基金资助(52002154);国家自然科学基金区域创新发展联合基金资助(U20A20331);江苏省交通运输科技与成果转化项目(2021G05)。

摘　　要：为减少高速公路交通事故的发生,综合利用长短期记忆神经网络(LSTM)和贝叶斯滤波(BF)研究高速公路交通事故风险状态预测模型,首先,通过LSTM模块学习历史交通流风险数据中存在的时间依赖关系;然后,通过BF模块融合LSTM预测结果提高实时风险预测效果;最后,利用宁波绕城高速2020年交通事故和沿线卡口数据,基于事故发生前20 min内上下游卡口间的时空范围,采用移动时间窗形式构建多步特征变量,并进行五折交叉验证。结果表明:相比随机森林(RF)算法,LSTM模型具有更高的精确率和召回率;在LSTM模型基础上,增加贝叶斯滤波BF模块可使最终风险预测结果F_(1)值达到0.80水平。In order to reduce highway traffic accidents,a traffic risk state prediction model was studied by using LSTM neural network and BF.Firstly,time dependency existing in historical traffic flow risk data was studied through LSTM module.Then,real-time risk prediction performance was improved by BF module integrating LSTM prediction results.Finally,with accident data and traffic flow data of Ningbo roundabout in 2020 as an example,multi-step feature variables were constructed in the form of migration time window based on spatio-temporal data between upper and lower port stations within 20 min prior to accidents,and 5-fold cross-validation was carried out.The results show that the precision and recall rate of LSTM model are higher than that of random forest(RF)algorithm,and F_(1) score of the final prediction result is close to 0.80 by adding BF module.

关 键 词：长短期记忆(LSTM) 贝叶斯滤波(BF) 高速公路 交通事故 风险预测 交通流 

分 类 号：X928.03[环境科学与工程—安全科学]