基于时空特征深度学习模型的路径行程时间预测  

Route travel time prediction on deep learning model through spatiotemporal features

作　　者：李先通[1] 全威 王华[1] 孙鹏程 安鹏进 满永兴 LI Xian-tong;QUAN Wei;WANG Hua;SUN Peng-cheng;AN Peng-jin;MAN Yong-xing(School of Transportation Science and Engineering,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学交通科学与工程学院,哈尔滨150001

出　　处：《吉林大学学报（工学版）》2022年第3期557-563,共7页Journal of Jilin University:Engineering and Technology Edition

基　　金：哈尔滨市科技创新人才青年后备人才项目(2017RAQXJ093)。

摘　　要：针对路径行程时间预测问题,提出了一种时空特征深度学习网络模型,结合卷积神经网络(CNN)和长短时记忆网络(LSTM),并考虑了路段空间依赖、时序依赖以及粗颗粒度中的时间飘移问题,对下一时隙的路径行程时间进行了预测。利用哈尔滨市出租车轨迹数据为实验数据,比较发现基于时空特征深度学习网络模型比差分自回归移动平均模型在多个评价指标上均有提升,其中平均绝对误差(MAE)和R;指标分别提升了18.6%和22.46%。结果表明,本文模型的预测精度可达到90%以上,且效率处于同类模型的领先水平。To process the important problem of path travel time prediction,this paper proposes a deep learning network model on spatiotemporal feature.It is combined with long-short-term memory network(LSTM)and convolutional neural network.At the same time,it considers the spatial dependence of road sections,timing dependence,and time drift in coarse granularity to predict the path travel time of the next time slot.In the experiments of this paper,the dataset of Harbin taxi trajectory is taken as the test dataset.The comparison results show that spatiotemporal characteristics based deep learning network model outperforms that based on the machine learning model with multiple evaluation indicators.On the indicators of MAE and R;,the performance of the algorithm in this paper is better than other algorithms by 18.6%and 22.46%respectively.At last,the prediction accuracy of the model proposed in this paper is over 90%,and the efficiency is at the leading level among similar algorithms.

关 键 词：交通运输系统工程 时空特征 卷积神经网络 长短期记忆网络 注意力机制 

分 类 号：U491[交通运输工程—交通运输规划与管理]