摘要
为获取完整的交通流数据集,提出了一种交通流数据修复方法.结合多源数据的互补特性,基于深度学习模型构建了时空关联特征提取方法,将高速公路交通流数据缺失情况分为3类,并基于随机森林算法建立修正模型.模型以平均绝对误差最小为优化目标,基于测试集和选择集优化了模型的参数.利用高速公路固定检测器和浮动检测技术获取的多源数据,对比分析了单一数据源与多源数据的修正精度.结果表明:多源数据修正模型明显优于单一数据源修正模型,在点缺失、线缺失和面缺失3种情况下,MAPE的平均值分别提高了24. 87%,39. 87%和52. 93%.此外,随着缺失比例的增加,较单一数据源模型,多源数据修正模型精度更为稳定,在点缺失、线缺失和面缺失3种情况下,其MAPE的方差仅为0. 01,0. 03和0. 08,证明其具有较好的鲁棒性.
To obtain a complete traffic dataset,an imputation method for traffic flow data was proposed.Based on the complementary of different types of data and deep learning method,the spatial and temporal features of traffic flow data were extracted.The missing values were divided into three types and an imputation model was established based on a random forest algorithm to estimate the missing values.With the testing and selecting dataset,the model parameters were optimized by minimizing the mean absolute error.Finally,the models were validated using two real-world datasets from camera and fix detector installed in the highway.The results indicate that the proposed model based on multi-source data is better than the models based on single source data,especially when the missing rate is higher.When the missing pattern is point,line and blook,the average mean absolute percent error(MAPE)is improved by 24.87%,39.87%and 52.93%.Moreover,the variances of MAPE are 0.01,0.03 and 0.08.The proposed model performs steady with the increase of the missing rate,thus it has better robustness.
作者
李林超
曲栩
张健
王永岗
李汉初
冉斌
Li Linchao;Qu Xu;Zhang Jian;Wang Yonggang;Li Hanchu;Ran Bin(School of Transportation,Southeast University,Nanjing 210096,China;Department of Civil and Environmental Engineering,University of Wisconsin-Madison,Wisconsin 53705,USA;School of Highway,Chang an University,Xi’an 710064,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2018年第5期972-978,共7页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(6161001115)
江苏省自然科学基金资助项目(BK20160685)
东南大学优秀博士学位论文基金资助项目(YBJJ1736)
关键词
交通工程
数据修复
深度学习
随机森林
多源数据
自编码网络
traffic engineering
data imputation
deep learning
random forest
multiple data
auto-encoder
