交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (5): 309-319.DOI: 10.16097/j.cnki.1009-6744.2026.05.026

• 系统工程理论与方法 • 上一篇    下一篇

融合多头交叉注意力的城市轨道交通短时客流预测

巩亮a,b ，朱欣雨a,b ，许得杰*a,b ，徐蕾a,b ，陈子豪a,b   

  1. 兰州交通大学，a.交通运输学院；b.高原铁路运输智慧管控铁路行业重点实验室，兰州730070）
  • 收稿日期:2026-06-17 修回日期:2026-07-22 接受日期:2026-08-21 出版日期:2026-10-25 发布日期:2026-10-21
  • 作者简介:巩亮（1980—），男，甘肃天水人，副教授。
  • 基金资助:
    国家自然科学基金(72261025)；中国国家铁路集团有限公司重点研发计划项目 (N2023X007)。

Short-Term Passenger Flow Prediction for Urban Rail Transit with Multi-Head Cross Attention Fusion

GONG Lianga,b, ZHU Xinyua,b, XU Dejie*a,b, XU Leia,b, CHEN Zihaoa,b   

  1. a. School of Traffic and Transportation; b. Key Laboratory of Railway Industry on Plateau Railway Transportation Intelligent Management and Control, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Received:2026-06-17 Revised:2026-07-22 Accepted:2026-08-21 Online:2026-10-25 Published:2026-10-21
  • Supported by:
    National Natural Science Foundation of China (72261025)；Key Research and Development Program of China State Railway Group Co. Ltd. (N2023X007)。

PDF

摘要： 针对城市轨道交通短时客流预测中时间依赖与空间关联耦合紧密，现有模型难以充分实现时空异构特征深度交互的问题，本文提出一种融合长短期记忆网络（Long Short-Term Memory,LSTM）、图卷积网络（Graph Convolutional Network, GCN）和多头交叉注意力（Multi-Head Cross Attention, MHCA）的短时客流预测模型。模型采用LSTM提取站点客流序列的时间演化特征，利用GCN学习轨道交通网络拓扑下的空间依赖关系，在此基础上，构建双向多头交叉注意力机制，实现时间编码特征与空间增强表示之间的深层交互融合，并结合非对称参数映射和空间掩码机制增强异构特征表达能力，降低远距离弱相关节点带来的干扰。基于杭州地铁80个站点AFC（Automatic Fare Collection）刷卡数据开展实验，并与LSTM、LSTM+GCN、STGCN（Spatio-Temporal Graph Convolutional Network）、ASTGCN（Attention Based Spatial-Temporal Graph Convolutional Network）、DCRNN（Diffusion Convolutional Recurrent Neural Network）和 Graph WaveNet（Graph WaveNet for Deep Spatial-Temporal Graph Modeling）等模型进行对比，结果表明，所提模型在全网15 min客流预测任务中RMSE（Root Mean Square Error）、MAE（Mean Absolute Error）和MAPE（MeanAbsolute Percentage Error）分别达到37.85、30.94和22.68%，整体性能优于各对比模型；在不同预测步长、不同客流类型和高峰时段预测场景下均表现出较好的预测精度和稳定性。消融实验进一步验证了GCN与MHCA的协同增益作用，以及空间掩码和非对称参数映射对模型性能的提升效果。研究表明，所提模型能够有效增强城市轨道交通复杂时空依赖关系的建模能力，为轨道交通客流预测及智能运营管理提供参考。

关键词: 城市交通, 客流预测, 深度学习, 轨道交通客流, 多头交叉注意力

Abstract: Given the strong coupling between temporal dependencies and spatial correlations in short-term passenger flow prediction for urban rail transit, the existing models have difficulties in fully achieving the deep interactions among heterogeneous spatiotemporal features. To address this issue, this paper proposed a short-term passenger flow forecasting model which integrates Long Short-Term Memory (LSTM), Graph Convolutional Network (GCN), and Multi-Head Cross Attention (MHCA). The model employs the LSTM to extract the temporal evolution patterns of station-level passenger flow sequences and uses the GCN to learn spatial dependencies under the metro network topology. Based on these representations, a bidirectional multi-head cross-attention mechanism is constructed to achieve the deep interaction and information fusion between the temporal encoding features and spatially enhanced representations. The mechanism further incorporates asymmetric projection matrices and a spatial masking mechanism to enhance heterogeneous feature representation, and reduce the interference from weakly correlated distant stations. Experiments are conducted by using the Automatic Fare Collection (AFC) data collected from 80 stations of the Hangzhou Metro. Compared with the LSTM, LSTM+GCN, STGCN (Spatio-Temporal Graph Convolutional Network), ASTGCN (Attention Based Spatial-Temporal Graph Convolutional Network), DCRNN (Diffusion Convolutional Recurrent Neural Network), and Graph WaveNet (Graph WaveNet for Deep Spatial-Temporal Graph Modeling), the proposed model achieves the best overall performance, with RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and MAPE (Mean Absolute Percentage Error) of 37.85, 30.94, and 22.68%, respectively, in the network-wide 15-minute forecasting task. A superior performance is also observed under different forecasting horizons, passenger flow types, and peak-hour scenarios. The ablation studies further confirm the effectiveness of the proposed architecture and demonstrate the positive contributions of the spatial masking strategy, asymmetric projection mechanism, and the synergistic interaction between GCN and MHCA. Research findings indicate that the proposed model can effectively enhance the capability to model complex spatiotemporal dependencies in urban rail transit, and provide a reference for passenger flow prediction and intelligent operation management of urban rail transit.

Key words: urban transportation, passenger flow prediction, deep learning, rail transit passenger flow, multi-head cross attention

中图分类号: