宏观交通流模型参数标定方法

Methodology of Parameter Calibration for Macroscopic Traffic Flow Models

为了提高交通流模型整体估计精度,对交通流模型参数估计方法进行了研究.针对密度-速度、密度-流量以及速度-流量模型之间的关联性以及交通流观测数据分布特征对模型估计精度的影响,提出了联合模型参数估计方法,并给出了联合模型参数估计优化目标函数的表达形式及约束条件.以Castillo-Benítez和Van Aerde模型为例,基于北京市二环快速路实测数据对联合模型参数估计方法可行性及参数估计效果进行了验证;构建了加权判定系数,并结合平均绝对百分比误差(mean absolute percentage error,MAPE)和平方根误差(root mean squared error,RMSE)评价联合模型估计效果.结果表明,对Castillo-Benítez模型而言,由单一模型计算的速度、流量估计MAPE分别是19.8%和18.7%,基于联合模型计算的速度、流量MAPE分别下降为10.0%和10.0%,模型总体判决系数由0.913变化为0.910;对Van Aerde模型而言,由单一模型计算的密度、流量估计MAPE分别为16.4%和16.3%,基于联合模型计算的密度、流量MAPE分别为14.2%和14.2%,模型总体判决系数由0.732变为0.749.

To improve the coherent whole precision of traffic flow stream models,the methodology of parameter calibration was studied.The joint-prediction model parameter estimate method was developed considering the correlation among density-speed,density-volume and speed-volume relationship,and the optimal functions were constructed in the light of principle of minimum sum of squares of estimation deviation and the roles of different dependent and independent variables in the traffic flow stream models.To demonstrate the feasibility and efficiency of joint-prediction model calibration method on the estimate precision,the Castillo-Benítez's model and Van Aerde's model were selected and the field data on the Beijing 2nd ring were used to calibrate the model parameters.The weighted determination coefficient was suggested and combined with mean absolute percentage error(MAPE)and root mean squared error(RMSE)were used to evaluate the model parameter estimate efficiency.Results show that for the Castillo-Benítez's model,the MAPEs of speed and volume calculated from single prediction model are 19.8%and 18.7%,respectively.The values calculated from joint-prediction model are 10.0%and 10.0%,respectively,and the determination coefficients change from 0.913 to 0.910.For the Van Aerde's model,the MAPEs of density and volume calculated from single prediction model are 16.4%and 16.3%,respectively.While the values calculated from joint-prediction model are 114.2%and 14.2%,respectively,and the determination coefficients change from 0.732 to 0.749.