交通运输系统工程与信息 ›› 2020, Vol. 20 ›› Issue (5): 128-134.

• 系统工程理论与方法 • 上一篇    下一篇

基于混合深度学习的地铁站进出客流量短时预测

赵建立1a,石敬诗1a,孙秋霞*1b,任玲2,刘彩红2   

  1. 1. 山东科技大学 a. 计算机科学与工程学院,b. 数学与系统科学学院,山东 青岛266590; 2. 青岛地铁集团有限公司,山东 青岛266000
  • 收稿日期:2020-05-06 修回日期:2020-08-14 出版日期:2020-10-25 发布日期:2020-10-26
  • 作者简介:赵建立(1977-),男,山东莱芜人,教授.
  • 基金资助:

    2019年度青岛市社会科学规划研究项目/Qingdao Social Science Planning Research Project (QDSKL1901121).

Short-time Inflow and Outflow Prediction of Metro Stations Based on Hybrid Deep Learning

ZHAO Jian-li1a, SHI Jing-shi1a, SUN Qiu-xia1b, REN Ling2, LIU Cai-hong2   

  1. 1a. College of Computer Science and Engineering, 1b. College of Mathematics and Systems Science, Shandong University of Science and Technology, Qingdao 266590, Shandong, China; 2. Qingdao Metro Group Co., Ltd, Qingdao 266000, Shandong, China
  • Received:2020-05-06 Revised:2020-08-14 Online:2020-10-25 Published:2020-10-26

摘要:

针对城市轨道交通多站点短时客流量预测问题,本文提出一种将卷积神经网络 (CNN)与残差网络(ResNet)相组合的预测模型(ResNet-CNN1D).模型将原始客流量数据作为输入,利用二维 CNN 与 ResNet 组成深层神经网络,捕捉站点间的空间特征,同时利用一维 CNN捕捉客流量的时间依赖.最后,基于参数矩阵,将时间和空间特征进行加权融合,完成对目标时段中多个站点进出客流量的同时预测.采集青岛市地铁3号线刷卡数据,对模型进行验证. 结果表明,相比现有传统的预测模型(ARIMA,SVR,LSTM,CLTFP,ConvLSTM),本文 ResNet-CNN1D模型具有更好的预测精度.

关键词: 城市交通, 短时客流量预测, 深度学习, 地铁刷卡数据, CNN, ResNet

Abstract:

This paper proposes a prediction model (ResNet- CNN1D) combining convolutional neural network (CNN) and residual network (ResNet) for multi- station short- term passenger volume prediction of urban rail transit. The original passenger volume data is used as input of the model. The deep network composed of twodimensional CNN and ResNet is used to mine the spatial features between the stations. The one-dimensional CNN is used to mine the temporal features of the passenger flow. Based on the parametric matrix, the temporal and spatial features are weighted to obtain the multi-station inflow and outflow during the research period. The model is verified by the card-swiping data of the Qingdao No.3 metro line. Compared with existing traditional prediction models (ARIMA, SVR, LSTM, CLTFP, ConvLSTM), the proposed ResNet-CNN1D model in this paper has the best prediction accuracy.

Key words: urban traffic, short-time passenger volume prediction, deep learning, metro card-swiping data, CNN, ResNet

中图分类号: