基于人工神经网络的微观交通仿真模型参数灵敏度分析  

Parameter Sensitivity Analysis of Microscopic Traffic Simulation Model Based on the BP Neural Network

作　　者：朱林波[1] 罗霞[1] 

机构地区：[1]西南交通大学交通运输与物流学院,成都610031

出　　处：《武汉理工大学学报（交通科学与工程版）》2014年第2期426-430,共5页Journal of Wuhan University of Technology（Transportation Science & Engineering）

基　　金：四川省科技支撑计划项目资助(批准号:2011FZ0050)

摘　　要：待校正参数集的确定是微观交通仿真模型参数标定的前提与关键,在VISSIM平台上构建快速路出入口匝道仿真模型基础上,为了利用BP神经网络技术研究模型中各参数对主-辅路行程时间这一评价指标的影响,使用拉丁超立方抽样抽取出200组数据作为训练样本及测试样本.经过训练、优化确定出训练误差最小的网络结构为12-11-1并获得各层间权重,改造后的Garson公式可以计算出各输入变量对评价指标的相对敏感度系数.根据相对敏感度系数排序可以确定出对主-辅路行程时间有较高敏感性的关键参数集,只对敏感度系数较高参数进行标定,减轻参数标定工作复杂度.基于BP神经网络对模型参数进行灵敏度分析的结果与实际使用经验较相似.Correction parameter set is determined to be the premise of microscopic traffic simulation model calibration parameters, on the basis for building simulation model of expressway on-ramp and off-ramp. In order to research the impact of model parameters on main-to-side road travel time by BP neural network, using the Latin Hypercube Sampling extracted 200 data sets as training samples and test samples. It can determine the training error minimizing network structure is 12-11-1 and weight between the layers after training and optimization, the modified Garson formula can calculate the rela- tive sensitivity coefficient of the input variables on the evaluation index. According to the sorting of relative sensitivity coefficient can identify the key parameters which have a high sensitivity to main-to-side road travel time, calibrating of these key parameters can reduce the complexity of parameter cali-bration. The results of model parameter sensitivity analysis based on BP neural network are similar with the practical experience.

关 键 词：BP神经网络 微观交通仿真 模型参数 灵敏度分析 

分 类 号：U495[交通运输工程—交通运输规划与管理]