摘要
为挖掘客流的复杂时空耦合关系,解析建成环境影响下的轨道交通客流出行规律,本文提出一种考虑城市建成环境的时空双层超图神经网络模型(Spatial Temporal-Double Hypergraph Neural Network,STDHGNN)。模型分为双层超图神经网络和时间序列模块,双层超图神经网络模块用于挖掘轨道交通线路站点间的高阶连通关系和相邻同类建成区域站点的集群关系,时间序列模块用于表征历史客流数据的时间依赖关系。同时,以建成环境和线路作为变量构造新的损失函数,旨在剖析建成环境的影响,提高模型的预测性能。最后,以武汉轨道交通数据为例开展实证研究。研究结果显示:考虑建成环境和轨道站点高阶连通关系对客流预测精度的提升效果显著,本模型均方根误差(RMSE)和平均绝对误差(MAE)值分别为52.04和29.32,比基线模型降低了22%以上,性能显著优于基线模型;通过消融实验验证了融合轨道高阶联通关系和建成环境对模型性能的贡献,其中,单步预测任务中,考虑这两种因素使模型性能分别提升了6%和9%,多步预测任务中,分别提升了4%和12%;构造的融合建成环境因素的可解释损失函数,提高了模型的预测性能,同时,使模型具备更好的科学性和可解释性。研究成果为城市轨道交通的客流管理和列车调度提供了技术支持。
To analyze the spatiotemporal coupling relationships of passenger flow and the passengers travel patterns in urban rail transit,this paper proposes a spatial-temporal double-hypergraph neural network model considering the urban built environment,which is called ST-DHGNN(Spatial Temporal-Double Hypergraph Neural Network).The model is consisted of a doublehypergraph neural network module and a time series module.The double-hypergraph neural network module is designed to uncover high-order connectivity among rail transit line stations and clustering relationships among neighboring stations within similar built-up areas.The time series module is utilized to represent the temporal dependencies in historical passenger flow data.Additionally,a new loss function is developed with the built environment and lines as variables,aiming to dissect the impact of the built environment and enhance the model's predictive performance.An empirical study is conducted using Wuhan Metro data as an example.The results indicate that:(1)Considering the built environment and high-order connectivity among rail stations significantly improves passenger flow prediction accuracy.The proposed model achieves Root Mean Square Error(RMSE)of 52.04 and Mean Absolute Error(MAE)values of 29.32.The performance was improved by over 22%compared to baseline models.(2)The ablation experiments have verified the contribution of integrating high-order connectivity relationships of movement trajectories and the built environment to the model's performance.Specifically,in the single-step prediction task,considering these two factors improves the model's performance by 6%and 9%,respectively.In the multi-step prediction task,the improvements are 4%and 12%,respectively.(3)The constructed interpretable loss function incorporating built environment factors enhances the model's predictive capability while imparting better scientific rigor and interpretability.The study can provide technical support for passenger flow management and train scheduling in urban rail transit.
作者
李之红
郄堃
王健宇
许晗
陈金政
LI Zhihong;QIE Kun;WANG Jianyu;XU Han;CHEN Jinzheng(Beijing Laboratory for General Aviation Technology,Beijing University of Civil Engineering and Architecture,Beijing 100044,China;School of Architecture and Urban Planning,Beijing University of Civil Engineering and Architecture,Beijing 100044,China;Beijing Transport Institute,Beijing 101160,China)
出处
《交通运输系统工程与信息》
北大核心
2025年第1期160-172,共13页
Journal of Transportation Systems Engineering and Information Technology
基金
国家自然科学基金(52402377)
北京市自然科学基金(9234025)
北京市属高等学校高水平科研创新团队建设支持计划项目(BPHR20220109)。
关键词
智能交通
客流多步预测
超图时空网络
城市轨道交通
建成环境影响
可解释损失函数
intelligent transportation
multi-step passenger flow forecasting
hypergraph spatiotemporal networks
urban rail transit
built environment impact
interpretable loss function
