基于组合深度学习的轨道交通短时进站客流预测模型  

Short-Term Inbound Passenger Flow Prediction of Model Rail Transit Based on Combined Deep Learning

作　　者：李淑庆[1] 李伟 刘耀鸿 马波 LI Shuqing;LI Wei;LIU Yaohong;MA Bo(School of Traffic&Transportation,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学交通运输学院,重庆400074

出　　处：《重庆交通大学学报（自然科学版）》2024年第2期92-99,共8页Journal of Chongqing Jiaotong University(Natural Science)

基　　金：国家自然科学基金项目(52078070);重庆交通大学研究生科研创新资助项目(CYS21355)。

摘　　要：针对轨道交通短时进站客流考虑不充分和特征学习不全面而导致预测精度不高的问题,选取客流特征、天气、空气质量和道路交通拥堵指数等多个因素,提出了一种基于组合深度学习的轨道交通短时进站客流预测模型(CNN-ResNet-BiLSTM)。基于卷积神经网络(CNN)对多因素客流时间序列进行自动提取,在CNN网络中插入多个残差神经网络(ResNet)来加深网络深度,利用双向长短时记忆神经网络(BiLSTM)捕捉前后两个方向的客流时间序列特征并得到预测结果;以杭州市全网80个站点工作日的进站客流为例,验证了该模型的有效性。研究结果表明:与常用的几种模型相比,多因素CNN-ResNet-BiLSTM组合模型的均方根误差(E RMS)至少降低了8.50%,平均绝对误差(E MA)至少降低了6.74%,平均绝对百分比误差(E MPA)至少降低了6.52%。Aiming at the problem of low prediction accuracy caused by incomplete consideration of factors and feature learning in the short-term inbound passenger flow prediction model of rail transit,a combined deep learning model based short-term inbound passenger flow prediction method for rail transit(CNN ResNet BiLSTM)was proposed by selecting multiple factors such as passenger flow characteristics,weather,air quality and road traffic congestion index.The multi-factor passenger flow time series were automatically extracted based on convolution neural network(CNN),and several residual neural networks(ResNet)were added into the CNN network to deepen the depth of the network.The bidirectional long short-term memory neural network(BiLSTM)was used to capture the time series characteristics of passenger flow in bidirectional directions and obtain the prediction results.The validity of the proposed prediction method was verified by a case study on the inbound passenger flow predication of 80 stations in whole network of Hangzhou city on workdays.The research results show that the root mean square error(E RMS)of the multi-factor CNN-ResNet-BiLSTM combined model is reduced by at least 8.50%,and the mean absolute error(E MA)is reduced by at least 6.74%and the mean absolute percentage error(E MPA)is reduced by at least 6.52%,compared with the several commonly used models.

关 键 词：交通工程 短时客流预测 组合深度学习 轨道进站客流 

分 类 号：U293.13[交通运输工程—交通运输规划与管理]