基于时空超图卷积模型的城市轨道站点客流预测

Passenger flow forecast of urban rail transit stations based on spatio-temporal hypergraph convolution model

城市轨道客流预测是智能交通系统的重要环节,准确的短时客流量预测有助于缓解城市轨道运营压力并提升地铁系统的服务质量。为挖掘交通系统中的时空交互特性并实现客流的精确预测,提出一种基于时空超图卷积模型(Spatio-Temporal Hypergraph Convolutional,ST-HConv)的短时进站客流预测方法。门控卷积层用于提取客流的时间特征,双层超图卷积用于获取站点间的近邻性和路网中的全局性,实现空间特征的提取;时空交互模块由时间门控卷积和空间超图卷积组成,将时空特征融合进而获取时空交互信息。以杭州地铁自动检票系统(AFC)采集的乘客刷卡数据为例,对模型的有效性进行检验。研究结果表明,与传统机器学习模型、传统深度学习模型和图网络模型相比,ST-HConv模型同时考虑时间特征和空间特征,并实现了时空特征的有效融合,使得ST-HConv模型的平均绝对误差和均方根误差都低于其他模型。在图结构性能方面,与时空图卷积模型(Spatio-Temporal Graph Convolutional,ST-GConv)相比,ST-HConv模型中的超图卷积层获得了路网中的局部特征和全局特征,有效地降低了预测误差。在不同的时间间隔(15 min/30 min/45 min/60 min)下,ST-HConv相较于ST-GConv,平均绝对误差分别降低了1.3,1.05,1.51和2.29,均方根误差分别降低了2,1.44,2.48和2.89。由此可见,ST-HConv模型综合考虑了时空交互信息,能够提高客流预测的准确性。

Urban rail transit passenger flow prediction is an important part of intelligent transportation systems.Accurate short-term passenger flow prediction can alleviate the operational pressure of urban rail transit and improve the service quality of the subway system.In order to explore the spatio-temporal interaction characteristics in the transportation system and achieve accurate passenger flow prediction,a short-term inbound passenger flow prediction method based on the spatio-temporal hypergraph convolutional model(ST-HConv)was proposed.The gate convolutional layers were used to extract the temporal features of passenger flow,and two-layer hypergraph convolution was utilized to obtain the local and global connectivity between stations and the road network,thus realizing the extraction of spatial features.The spatio-temporal interaction module was composed of time gate convolution and spatial hypergraph convolution,which integrated the spatio-temporal features to obtain spatio-temporal interaction information.This paper verified the effectiveness of the proposed method by using passenger card-swiping data collected from the Hangzhou metro automatic fare collection(AFC)system.The research results show that compared with traditional machine learning models,traditional deep learning models,and graph network models,the ST-HConv model considers both temporal and spatial features,and effectively integrates spatio-temporal features,resulting in lower mean absolute error and root mean square error than other models.In terms of graph structure performance,compared with the spatio-temporal graph convolutional model(ST-GConv),the hypergraph convolutional layer in ST-HConv obtains both local and global features in the road network,effectively reducing prediction errors.In different time intervals(15 min/30 min/45 min/60 min),the ST-HConv reduces the mean absolute error by 1.3,1.05,1.51,and 2.29 compared with the ST-GConv,and reduces the root mean square error by 2,1.44,2.48,and 2.89,respectively.Therefore,the ST-HConv model comprehensively considers spatio-temporal interaction information and can improve the accuracy of passenger flow prediction.