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): 352-361.DOI: 10.16097/j.cnki.1009-6744.2026.05.030

• 系统工程理论与方法 • 上一篇    下一篇

面向环境扰动的轨道交通短时客流场景自适应门控融合预测

李先旺* ，张凯欣，胡珂，阳紫欣   

  1. 广西大学，广西制造系统与先进制造技术重点实验室，南宁530004
  • 收稿日期:2026-06-10 修回日期:2026-08-10 接受日期:2026-08-20 出版日期:2026-10-25 发布日期:2026-10-22
  • 作者简介:李先旺（1982—），男，湖南郴州人，讲师，博士。
  • 基金资助:
    国家自然科学基金 (U22A2053)。

Scenario-Adaptive Gated Fusion Forecasting of Short-Term Urban Rail Transit Passenger Flow Under Environmental Disturbances

LI Xianwang*, ZHANG Kaixin, HU Ke,YANG Zixin   

  1. Guangxi Key Lab of Manufacturing System and Advanced Manufacturing Technology, Guangxi University, Nanning 530004, China
  • Received:2026-06-10 Revised:2026-08-10 Accepted:2026-08-20 Online:2026-10-25 Published:2026-10-22
  • Supported by:
    National Natural Science Foundation of China (U22A2053)。

PDF

摘要： 针对现有模型难以适应环境扰动下多源信息贡献度的动态变化，导致异常天气下预测精度下降的问题，本文提出一种场景自适应门控融合预测方法。该方法融合进站客流、出站客流、图结构和天气环境这4类信息源，构建基于缩放点积的图注意力模块以克服静态拓扑局限、提取站点间动态空间依赖，并设计环境驱动的门控融合机制，将环境信息作为场景感知信号动态调整4类信息源的分支级融合权重。基于北京地铁276个站点的自动售检票系统数据，在15、30、60min这3种时间粒度下进行实验。在60min粒度下，该方法平均绝对误差较双向长短时记忆网络图神经网络混合模型降低28.5%，优势最为显著；在15min粒度下，平均绝对误差为22.34，均方根误差为41.41，加权平均绝对百分比误差为8.79%，综合精度优于对比模型。异常天气条件下，图结构分支融合权重从0.5996降至0.5308，进站客流分支权重从0.1815升至0.232 1，天气分支权重相应上升。消融实验显示，移除门控融合机制后平均绝对误差上升至34.10，移除图注意力模块后上升至36.47。实验结果表明：场景自适应门控机制能够根据环境扰动强度动态调整多源信息的贡献度，提升复杂环境条件下的预测鲁棒性。

关键词: 城市交通, 客流预测, 图注意力机制, 门控融合, 场景自适应

Abstract: Existing models struggle to adapt to the dynamic changes in the contributions of multi-source information under environmental disturbances, which leads to the degraded forecasting performance under abnormal weather conditions. To address this problem, the study proposes a scenario-adaptive gated fusion method for metro passenger flow forecasting. The proposed method integrates four types of sources: passenger inflow, passenger outflow, graph topology, and environmental information. A scaled dot-product graph attention module is developed to capture the dynamic spatial dependencies beyond static topology, and an environment-driven gating mechanism is introduced to adaptively adjust the fusion weights of the four information sources. Experiments are conducted by using the automatic fare collection data from 276 Beijing subway stations under 15 minutes, 30 minutes, and 60 minutes forecasting intervals. The proposed method reduces mean absolute error (MAE) by 28.5% compared with a bidirectional long short-term memory and graph neural network hybrid model at the 60 minutes interval. At the 15 minutes interval, the proposed method achieves a MAE of 22.34, a root mean square error (RMSE) of 41.41, and weighted mean absolute percentage error (WMAPE) of 8.79%, outperforming all comparison methods. Under abnormal weather conditions, the branch weight of graphs decreases from 0.599 6 to 0.530 8, while the branch weight of inflows increases from 0.181 5 to 0.232 1, and the weather branch weight rises correspondingly. Ablation experiments demonstrate that removing the gated fusion mechanism alone increases MAE to 34.10, while removing the graph attention module alone increases MAE to 36.47. The results show that the proposed scenario-adaptive gating mechanism effectively adjusts the contributions of multi-source information according to the intensity of environmental disturbances, and improves the forecasting robustness under complex weather conditions.

Key words: urban transportation, passenger flow prediction, graph attention mechanism, gated fusion, scene adaptation

中图分类号: