作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (5): 297-305. doi: 10.19678/j.issn.1000-3428.0061309

• 开发研究与工程应用 • 上一篇    下一篇

面向短时地铁客流量预测的混合深度学习模型

彭桐歆1,2, 韩勇1,2, 王程3, 张志浩1,2   

  1. 1. 中国海洋大学 信息科学与工程学院, 山东 青岛 266100;
    2. 青岛海洋科学与技术国家实验室 区域海洋动力学与数值模拟功能实验室, 山东 青岛 266237;
    3. 青岛市市北区党建引领基层治理推进中心, 山东 青岛 266000
  • 收稿日期:2021-03-29 修回日期:2021-07-07 发布日期:2021-07-13
  • 作者简介:彭桐歆(1996—),女,硕士研究生,主研方向为时空大数据挖掘;韩勇(通信作者),教授;王程、张志浩,硕士研究生。
  • 基金资助:
    山东省自然科学基金面上项目(ZR2020MD020)。

Hybrid Deep-learning Model for Short-term Metro Passenger Flow Prediction

PENG Tongxin1,2, HAN Yong1,2, WANG Cheng3, ZHANG Zhihao1,2   

  1. 1. College of Information Science and Engineering, Ocean University of China, Qingdao, Shandong 266100, China;
    2. Laboratory for Regional Oceanography and Numerical Modeling, Qingdao National Laboratory for Marine Science and Technology, Qingdao, Shandong 266237, China;
    3. Center of Grassroots Governance Led by the Chinese Communist Party in Shibei District, Qingdao, Shandong 266000, China
  • Received:2021-03-29 Revised:2021-07-07 Published:2021-07-13

摘要: 城市交通客流量精准预测是智能交通系统的重要环节,是有效管控交通、规划最佳出行线路的关键。目前城市交通客流量短时预测研究主要集中在利用深度学习模型进行时空特征的提取,忽略了对模型优化的研究。针对短时地铁客流量预测存在的问题,提出一种混合深度学习模型ResGRUMetro,将卷积神经网络、残差单元和门控循环单元相结合,捕获流量数据的时空特征。针对深度学习模型常用的损失函数难以对交通客流量峰值进行精准预测的问题,引入面向短时交通流量预测的加权平方误差,根据交通客流量的大小为预测误差赋予不同权重,并加大对交通客流量峰值处误差的惩罚,使神经网络在反向传播时更加关注峰值处的预测和误差,从而提升交通客流量峰值的预测精度。此外,通过耦合天气、空气质量等外部因子,改善模型的整体预测性能,增强模型的稳定性。实验结果表明,相比LR、PSVR、CNN等典型的预测模型,ResGRUMetro模型有更高的预测精度,能够准确预测交通客流量的峰值。

关键词: 智能交通, 短时客流量预测, 时空特征, 残差神经网络, 门控循环单元, 加权平方误差

Abstract: Accurate prediction of traffic passenger flow is an essential part of an Intelligent Transportation System (ITS), which helps manage traffic and plan the best travel routes.Presently, the short-term prediction of urban traffic passenger flow mainly focuses on using deep-learning models to extract spatiotemporal features, which neglects the model optimization.A hybrid deep learning model, ResGRUMetro, is proposed to predict short-term metro passenger flow.It combines Convolutional Neural Networks (CNN), Residual Units (ResUnits), and Gated Recurrent Units (GRU) to capture the spatiotemporal dependency of flow data.A Weighted Square Error (WSE) with a bias towards peak traffic passenger flow is proposed to address the problem that the classic loss function used in deep learning models cannot capture the peak hours features.The WSE applies different weights to the prediction errors according to the traffic passenger flow, which may increase the penalty for the peak traffic passenger flow error. Additionally, it helps the neural network pay more attention to the prediction and error of the peak traffic passenger flow during the backpropagation period, which may improve the prediction accuracy of the peak traffic passenger flow.In addition, the model incorporates external factors, such as weather and air quality, which improves the overall prediction performance and enhances the model stability.The results show that the hybrid ResGRUMetro model has a more accurate prediction ability than some typical prediction models such as LR, PSVR, and CNN.In addition, hybrid ResGRUMetro can accurately predict the peak value of passenger traffic flow.

Key words: intelligent transportation, short-term passenger flow prediction, spatio-temporal feature, Recurrent Neural Networks(RNN), Gated Recurrent Units (GRU), Weighted Square Error (WSE)

中图分类号: