摘要
准确的地铁客流预测是智能交通系统应对交通挑战、协调运营调度、规划未来建设的重要战略需求。然而,先前将图卷积网络与深度学习模型(如循环神经网络、长短期记忆网络和门控循环神经网络等)相结合的相关研究只能提取基于路网图结构的时间和空间相关性,而忽略了地铁站点之间隐藏的空间相关性和随时间变化的动态时间和动态空间相关性。为了挖掘交通数据中复杂的时空相关性以实现精确的地铁客流预测,提出一种基于自适应扩散图卷积注意力(Adaptive Diffusion Graph Convolution Attention, ADGCA)网络的客流预测方法。此方法的创新点主要包括2个方面:首先,通过构建多图和自适应矩阵,并结合多头注意力机制,能够挖掘地铁站点之间隐藏的空间相关性。这种方法优化了现有方法在提取地铁系统空间信息特征方面的不足,使得ADGCA模型能够更全面地提取地铁系统中的空间信息特征。其次,构建了一种结合因果卷积、自适应扩散图卷积和多头注意力机制的深度学习模型组件。该组件能够在局部和全局层面捕捉地铁客流数据中的动态时空相关性,相比于先前的方法,能够更有效地提取复杂的地铁客流数据特征。在根据上海和杭州地铁自动检票系统的乘客刷卡记录所构建的2个真实数据集上,对模型有效性进行评估。研究结果表明,相较于现有的基线模型,ADGCA模型能够提取更加真实的动态时空相关性,从而有效地降低了预测误差。在所有预测时间步长上,ADGCA模型预测准确度指标均优于基线模型。研究结果为进一步优化城市地铁运营计划和保障地铁安全营运提供了更加精确的数据支持。
Accurate metro passenger flow prediction is an important strategic requirement for intelligent transportation systems to address traffic challenges,coordinate operational scheduling,and plan future developments.However,previous research that integrated graph convolutional networks with deep learning models such as recurrent neural networks,long short-term memory networks,and gated recurrent neural networks,could only extract temporal spatial correlations based on the road network map structure,while ignoring the hidden spatial correlations between metro stations and the dynamic temporal correlations over time.To mine the complex spatial and temporal correlations in the transportation data to achieve accurate metro passenger flow prediction,a method based on Adaptive Diffusion Graph Convolution Attention(ADGCA)network was proposed.The innovations of this method mainly include two aspects:first,by constructing multiple graphs and adaptive matrices combined with multi-head attention mechanisms,it is able to mine the hidden spatial correlations between metro stations.This approach optimized the inadequacy of existing methods in extracting the spatial information features of metro systems,which made the ADGCA model able to extract the spatial information features in the metro system.Second,a deep learning model component combining causal convolution,adaptive diffusion map convolution and multi-head attention mechanisms was constructed.The component can capture the dynamic spatio-temporal correlations in metro passenger flow data at both local and global levels,and is more effective in extracting complex metro passenger flow data features than previous methods.The effectiveness of the model was evaluated on two real datasets constructed from passenger swipe records of the automatic metro ticketing systems in Shanghai and Hangzhou.The research results indicate that the ADGCA model can extract more realistic dynamic spatio-temporal correlations compared to existing baseline models,thereby effectively reducing prediction error.The prediction accuracy indices of the ADGCA model are better than the baseline model in all prediction time steps.The research findings provide more precise data support for further optimizing urban metro operation plans and ensuring the safe and efficient operation of metros.
作者
唐郑熠
黄嘉欢
王金水
邢树礼
TANG Zhengyi;HUANG Jiahuan;WANG Jinshui;XING Shuli(College of Computer Science and Mathematics,Fujian University of Technology,Fuzhou 350118,China;Fujian Provincial Key Laboratory of Big Data Mining and Applications,Fujian University of Technology,Fuzhou 350118,China;Key Laboratory of Hunan Province for Mobile Business Intelligence,Hunan University of Technology and Business,Changsha 410205,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2024年第12期4910-4923,共14页
Journal of Railway Science and Engineering
基金
福建省自然科学基金资助项目(2022J01933)
湖南省重点实验室开放研究基金资助项目(2015TP1002)。
关键词
智能交通
地铁客流预测
自适应扩散图卷积
因果卷积
多头注意力
intelligent transportation
metro passenger flow prediction
adaptive diffusion graph convolution
causal convolution
multi-head attention
作者简介
通信作者:王金水(1981-),男,福建漳州人,副教授,博士,从事交通数据挖掘与处理技术研究,E−mail:wangjinshui@fjut.edu.cn。
