基于变量选择-神经网络模型的复杂路网短时交通流预测  

Complex Network Short-Term Traffic Forecasting Based on Lasso-NN Model

作　　者：蒋士正[1] 许榕[1] 陈启美[1] 

机构地区：[1]南京大学电子科学与工程学院,南京210046

出　　处：《上海交通大学学报》2015年第2期281-286,共6页Journal of Shanghai Jiaotong University

基　　金：江苏省科技厅项目(BE2009667);江苏省自然科学基金(BK2010366)资助

摘　　要：针对传统交通流预测模型正在由单断面历史数据处理向多断面、多时刻历史数据处理转变,但在考虑各断面间的影响时,多变的交通状况往往会使预测模型复杂化的问题,引入一种多元线性回归最小绝对收缩和选择算子方法(Lasso),并利用其优秀的变量选择能力,在复杂路网多断面中选出相关性较高的断面;结合神经网络(NN)的非线性特性,提出了Lasso-NN组合模型.结果表明:Lasso-NN模型在路网交叉口对未来15min交通流数据预测的误差率低于9.2%;在非交叉口的误差率低于6.7%,总体优于各自单独使用得出的结果.Traditional traffic forecasting models are transforming from single section historical data processing to multi-section and multi-timing historical data processing. However, when considering the influence between each section, the fickle traffic condition tends to complicate the forecasting model. Therefore, this paper introduced the Lasso method used in multivariable linear regression and utilized its excellent ability of variable selection. It selected partial high correction sections from a complex multi-section road net-work. Combined with the non-linear neural network, a new Lasso-NN model was proposed. The result shows that the Lasso-NN model has an overall lower error rate. In the intersection region of the road net- work, the error rate is less than 9.2% and in the non-intersection region, it is less than 6.7%.

关 键 词：短时交通流预测 最小绝对收缩和选择算子 变量选择 神经网络 

分 类 号：U491[交通运输工程—交通运输规划与管理]