面向NSM的高速公路大区段事故风险预测方法

A Crash Prediction Method for Long Segments on Freeways Based on Road Network Traffic Safety Management

为探究路网交通安全管理(NSM)中的事故风险预测方法,以国内高速公路的大区段路段为研究对象,首先分别采用系统聚类、k-means动态聚类和二阶聚类方法对路段进行聚类,确定最优聚类方法和聚类数量,然后对"同质性路段"分别建立负二项回归、贝叶斯负二项回归、随机或固定效应的负二项回归和多层混合效应负二项回归4种模型,通过精度评价指标选择出最优的事故预测模型,最后计算出相应路段的事故风险大小并识别出事故多发路段。结果表明:选择最优的聚类方法和聚类数量相较于未聚类的情况将有效提高事故预测的拟合精度,其均方方差下降了64%。当选择二阶聚类方法且聚类数量为3时,"同质性路段"负二项回归的事故模型拟合精度最高,其模型的赤池信息量AIC为464.79,贝叶斯信息量BIC为476.98,均方方差为99.22。在4种事故预测模型中,负二项回归具有良好的预测精度,其预测结果的均方方差最小,为108.64。采用统计学方法识别"同质性路段"的事故多发路段,共识别出辽宁省22条事故多发路段。

In order to study crash prediction methods in road network traffic safety management（NSM）,long segments on freeways in China are regarded as objects.Methods of hierarchical clustering,k-means dynamic clustering,and two-stage hierarchical clustering are used,then the optimal clustering method and clustering number are obtained.Four models use negative binomial regression,Bayesian negative binomial regression,negative binomial regression with random or fix effects,and multilevel mixed-effects negative binomial regression of are developed for homogeneous road segments.The optimal model for crash prediction is determined by comparing precision evaluation indices of those four models.The risk of crash on freeway segments is calculated,and the segments with high frequency of crashes are identified.The results show that the optimal clustering method and cluster number can effectively improve the accuracy of the crash prediction model,which MSE decreases by 64% when compare with non-cluster circumstance.When two-stage clustering method is used and the number of cluster is 3,the fitting accuracy of homogeneous road segments is the highest,and its AIC,BICand MSEare 464.79,476.98,99.22,respectively.Among the four models,the negative binomial regression model has good prediction accuracy,its MSEis 108.64.Statistical methods can be used to identify crash-prone segments in homogeneous road segments,and 22 crash-prone segments on freeways in Liaoning province are identified.