基于数据流集成回归的短时交通流预测

传统的交通流预测技术使用静态和离线算法,无法对模型的参数值和内部结构进行在线调整.然而,交通流变化具有明显的动态性,其内在模式会随时间发生变化,导致构建好的模型准确度下降.针对上述问题,提出了基于数据流集成回归的短时交通流预测模型.将不断产生的交通流数据划分成数据块,每个数据块训练1个基础回归模型,然后加权组合为集成模型.通过不断训练新的基础模型,并置换出集成模型中准确度最差的基础模型,实现在线更新.在实测数据上的对比实验结果表明,与静态离线的BN模型相比,模型的均方根误差降低了19.5%,运算时间降低了48.7%,并能够快速适应交通状况发生明显变化的情况,适用于城市主干道路的短时交通流预测问题. 

An Online Short-term Traffic Flow Prediction Model Based on Data Stream Ensemble Regression

Traditional approaches to short-term traffic flow prediction are static and off-line,whose parameter values and internal structures cannot be adjusted on-line.The intrinsic pattern of the traffic flow usually changes over time,leading to model degradation and drop in accuracy.To solve the problem,a novel approach for short-term traffic flow prediction based on ensemble learning on data streams is proposed.The idea is to train a group of base regression models from sequential chunks of the traffic flow data,and combine them into an ensemble with different weights based on their expected accuracy.New base models are constantly constructed and the base model with the worst accuracy from the ensemble will be replaced.Experiments on real data show that the ensemble model can reduce the RMSE by 19.5% and computation time by 48.7%,compared with the static and off-line model.The ensemble model is adaptive to the evolving environment;therefore,it can be applied in the short-term traffic flow prediction for urban roadways.