铁道科学与工程学报2026，Vol.23Issue(3)：1111-1124,14.DOI:10.19713/j.cnki.43-1423/u.T20250757
多维兴趣点驱动的地铁客流多通道集成预测方法
A multi-channel integrated prediction method for metro passenger flow driven by multi-dimensional point of interest
摘要
Abstract
Points of Interests(POIs)within the service radius of metro stations embody diverse spatiotemporal characteristics and play a crucial role in shaping passenger mobility.During specific periods,POIs attract substantial interactive passenger flows.To accurately capture the mechanisms through which POIs influence metro ridership,this study proposed a multidimensional POI-driven multi-channel integrated prediction framework for metro passenger flow.First,this paper had comprehensively considered three cases.(I)the attraction of station attributes on passenger flow,(II)the dynamic variations in POI attractiveness across different time periods,and(III)the distribution and proximity of POIs within each station's service area.It constructed a station passenger attraction function,a POI temporal weighting model,and a POI spatial interaction matrix.Subsequently,this paper developed a Multi-Channel Attention Spatio-Temporal Neural Network(MCASTNN)that adopted a three-branch architecture and incorporates multiple attention mechanisms to mine the deep correlations between multidimensional POIs and metro ridership.A multi-channel attention integration module was then employed to fuse features,enabling the model to effectively capture complex station,temporal,and spatial dependencies in metro passenger flow.Finally,experiments were conducted using Hangzhou metro AFC smart card data.The results demonstrate that,compared with classical machine learning and deep learning models,MCASTNN can achieve superior prediction accuracy across residential-dominated,work-dominated,and mixed-commercial stations under various forecasting horizons.Specifically,relative to the Transformer model,MCASTNN can reduce the root mean square error and mean absolute error in single-step prediction tasks by an average of 3.09 and 3.11,respectively.Ablation studies on POI features and model components further verify the effectiveness of the proposed framework and each feature extraction branch.This approach can provide a valuable reference for optimizing the allocation of metro station resources and formulating rational train operation strategies.关键词
城市交通/多维POI综合应用/注意力机制/地铁客流预测/时空关联特征Key words
urban traffic/multi-dimensional POI comprehensive application/attention mechanism/metro passenger flow forecast/spatial-temporal correlation features分类
交通工程引用本文复制引用
车秉泽,钱名军,王全能..多维兴趣点驱动的地铁客流多通道集成预测方法[J].铁道科学与工程学报,2026,23(3):1111-1124,14.基金项目
甘肃省教育厅高等学校创新基金项目(2020A-038) （2020A-038）
甘肃省教育厅双一流重大科研项目(GSSYLXM-04) （GSSYLXM-04）
