基于CLTAttention的大型客站多进站口客流协同预测方法OA
CLTAttention-based approach for collaborative passenger flow forecasting across multiple entrances in large railway stations
大型铁路客站普遍采用多进站口布局,但各进站口客流存在显著的空间异质性.现有管理模式仅能通过售票系统获取宏观客流总量,缺乏对细分进站口分时段客流分布的精准预测,致使资源配置呈现静态化特征,进而导致高峰时段局部通道拥堵、旅客进站效率下降等问题.为提高进站口客流预测精度,支持动态资源配置,提出一种多进站口客流协同预测方法.基于多源数据融合,整合历史售票数据(由历史上车人数模拟)、列车开行计划及闸机数据,构建包含时空流量、周期特征及高峰时段特性的综合数据集.设计CLTAttention预测模型,结合卷积神经网络(convolutional neural network,CNN)的局部特征提取能力、长短期记忆网络(long short-term memory network,LSTM)的时序建模优势,以及时间注意力机制的自适应权重分配,实现多进站口客流的时空协同预测.同时,通过动态权重可视化技术增强模型的可解释性.在某枢纽站的实证研究中,CLTAttention模型表现优于单模型及经典组合模型:平均绝对误差(mean absolute error,MAE)较次优组合模型(CNN+LSTM)降低4.0%,均方根误差(root mean square error,RMSE)降低6.4%,拟合优度(R2)提升至0.938,验证了模型的高精度与鲁棒性.动态权重可视化进一步揭示了客流分布的时空规律,提升了模型的可理解性.基于预测结果生成的动态资源配置方案,为客站精细化运营提供了理论依据.提出的多源数据融合与CLTAttention模型能够有效提升大型客站多进站口客流协同预测精度,对缓解高峰拥堵、优化资源利用率具有重要实践价值.
Modern large railway stations commonly adopt multiple entrance configurations,but significant spatial heterogeneity exists in passenger distribution across different gates.Current management systems primarily rely on ticketing data to obtain aggregate passenger volume,lacking precise predictions of time-varying passenger distribution at individual entrances.This can lead to static resource allocation,causing localized congestion during peak hours and reduced passenger throughput efficiency.To address these limitations,this paper proposes a collaborative prediction framework for multi-gates passenger flow that enhances forecasting accuracy and enables dynamic resource optimization.By integrating multi-source data including historical ticket sales(simulated from historical boarding counts),train schedules,and gate transit records,this study constructed a comprehensive dataset encompassing spatiotemporal flow patterns,periodic characteristics,and peak-hour features.The proposed CLTAttention model synergistically combined the local feature extraction capability of Convolutional Neural Networks(CNN),the temporal modeling strength of Long Short-Term Memory(LSTM)networks,and the adaptive weight allocation of temporal attention mechanisms to achieve spatiotemporal collaborative prediction.Dynamic weight visualization was incorporated to enhance model interpretability.Experimental results from a major railway station demonstrate CLTAttention's superior performance.Achieving a 4.0%reduction in MAE,6.4%improvement in RMSE,and 0.938 R2 score compared to benchmark models.The visualization component can effectively reveal underlying spatiotemporal patterns in passenger distribution.The prediction outputs facilitate data-driven dynamic resource allocation strategies,providing both theoretical foundations and practical solutions for station operation optimization.This research can contribute an effective multi-gates passenger flow prediction system that significantly improves forecasting accuracy in large railway stations,with substantial practical implications for congestion mitigation and resource utilization optimization during peak operational periods.
张馨予;史天运;李昊光;李超
中国铁道科学研究院,北京 100081中国铁道科学研究院,北京 100081中国国家铁路集团有限公司,北京 100038中国铁道科学研究院,北京 100081
交通工程
大型铁路客站卷积神经网络长短期记忆网络时间注意力机制多进站口客流预测
major railway stationsconvolutional neural network(CNN)long short-term memory network(LSTM)temporal attention mechanismmulti-gates passenger flow prediction
《铁道科学与工程学报》 2026 (3)
1083-1095,13
中国国家铁路集团有限公司系统性重大项目(P2024X001)
评论