摘要
【背景】随着城镇化进程的加快,城市轨道交通压力越来越大,准确的地铁客流预测对于优化列车时刻表、减少高峰时段拥堵、提高地铁系统服务水平具有重要的作用。【目标】综合考虑地铁客流的时空特性,充分利用多时间粒度客流数据,提高较大时间粒度客流预测任务的准确性。【方方法法】分析不同时间粒度的地铁客流数据之间的相关性,确定多时间粒度融合机制。提出一种Resnet-GCN-Transformer模型:利用图卷积神经网络(Graph Convolutional Neural Network,GCN)提取不同站点客流的空间相关性;利用残差块构建深度卷积神经网络,对不同时间粒度的数据从小到大进行聚合,得到多时间粒度的多通道特征图;利用Transformer Encoder对客流数据的长时间依赖特性进行建模,并通过多个由全连接层构成的预测头输出预测结果。同时基于Optuna框架进行超参数优化,得到最优的超参数组合。【数据】对杭州市地铁刷卡数据集进行降噪处理并构建不同时间粒度的地铁客流数据集,基于10 min和30 min的数据集对模型进行验证。【结果】在两组不同目标时间粒度的数据集上,所提模型的平均绝对百分比误差分别为12.62%和10.61%,均优于6个基线模型,表明融合多时间粒度的特征在地铁客流预测任务中的重要性,模型能够充分捕捉多时间粒度的客流特征,地铁站点的连通关系,以及客流数据的时间依赖关系,从而显著提升客流预测效果。
[Background]With the acceleration of urbanization,the pressure on urban rail transit is increasing.The accurate prediction of subway passenger flow plays an important role in optimizing train schedules,reducing congestion during peak hours,improving the service level of subway systems,and providing subway freight.[Objective]To comprehensively consider the spatiotemporal characteristics of subway passenger flows,fully utilize the multi-time-granularity passenger-flow data,and improve the accuracy of larger time-granularity passenger-flow prediction tasks.[Method]The correlation between subway passenger-flow data with different time granularities is analyzed,a multi-time granularity fusion mechanism is determined,and a ResNet-GCN-Transformer model is proposed.Graph convolutional network(GCN)is used to extract the spatial correlation of passenger flows at different stations,and a deep convolutional neural network is constructed with residual blocks.Data with different time granularities are aggregated in the order of small to large time granularities to obtain multichannel feature graphs with multi-time granularity.A transformer encoder is used to model the long-term dependence of the passenger-flow data,and the prediction results are output through multiple prediction heads composed of fully connected layers.Besides,hyperparameter optimization is performed based on the Optuna framework to obtain the optimal combination of hyperparameters.[Data]Noise reduction is performed on the Hangzhou Metro card swipe dataset,and subway passenger-flow datasets with different time granularities are constructed.The model is verified based on 10 min and 30 min datasets.[Result]For the two datasets with different target time granularities,the mean absolute percentage error(MAPE)of the ResNet-GCN-Transformer model is found to be 12.62%and 10.61%,respectively.Both values are lower than the MAPE of the six baseline models,which indicates that the proposed model has higher prediction accuracy.This demonstrates the importance of integrating multi-time granularity features in metro passenger-flow prediction tasks.The model can fully capture these features,and also the connectivity characteristics and time-dependent relationships of subway stations so that the passenger-flow prediction effect can be significantly improved.
作者
杜姿晨
郑长江
郑树康
马庚华
陆野
DU Zichen;ZHENG Changjiang;ZHENG Shukang;MA Genghua;LU Ye(College of Civil and Transportation Engineering,Hohai University,Nanjing 210024,China;College of Environment,Hohai University,Nanjing 210024,China;College of Harbour,Coastal and Offshore Engineering,Hohai University,Nanjing 210024,China)
出处
《交通运输工程与信息学报》
2026年第1期64-79,共16页
Journal of Transportation Engineering and Information
基金
国家自然科学基金项目(72471083)。
