基于Stacking集成学习的机场线短时客流预测研究  

Study on short-term passenger flow prediction for a subway airport line based on Stacking ensemble learning

作　　者：杨安安 韩星玉 田旷 刘泽远 明玮 YANG An’an;HAN Xingyu;TIAN Kuang;LIU Zeyuan;MING Wei(Beijing Intelligent Transportation Development Center(Beijing Automotive Regulation and Management Service Center),Beijing 100161,China;Beijing Metro Operation Administration Co.,Ltd.,Beijing 100068,China;Beijing Jingcheng Metro Co.,Ltd.,Beijing 100082,China)

机构地区：[1]北京市智慧交通发展中心(北京市机动车调控管理事务中心),北京100161 [2]北京市轨道交通运营管理有限公司,北京100068 [3]北京京城地铁有限公司,北京100082

出　　处：《山东科学》2024年第4期112-120,共9页Shandong Science

基　　金：北京市自然科学基金(L191023)。

摘　　要：地铁机场线客流具有高度时变性,受机场航班影响使得精准的短时客流预测具有挑战性。综合考虑机场航班信息和机场线路历史客流,构建了一种以随机森林(RF)、LightGBM(light gradient boosting machine)、梯度提升决策树(GBDT)和逻辑回归算法作为集成学习器,基于叠加(Stacking)集成模型的机场线路短时客流预测模型。以北京地铁大兴机场线为实例进行验证,并与Informer和长短时记忆神经网络(long short-term memory,LSTM)两种基线模型进行对比。结果表明,考虑航班信息和机场线历史客流的双通道预测效果明显优于仅考虑机场线历史客流的单通道预测;Stacking模型在各项指标中均表现出优越的性能,其中,在96步长(24 h)下的预测效果最好,预测进站客流的平均绝对误差为7.66,预测出站客流的平均绝对误差为4.67;分析航班信息特征对预测模型的影响,发现离港航班信息重要性不如到港航班,这与离港旅客提前到达机场时间差异较大有关。The highly dynamic nature of subway airport line passenger flows and their susceptibility to the influence of airport flight schedules present challenges for accurate short-term forecasting of passenger flow.This study integrates airport flight information and historical passenger flow data from airport lines to construct a short-term passenger flow forecasting model based on a stacking ensemble model.The model incorporates random forest(RF),LightGBM(light gradient boosting machine),gradient boosting decision tree(GBDT),and logistic regression algorithms to act as ensemble learners.The proposed model is validated using data from the Beijing Subway Daxing Airport Line and is compared against two baseline models,namely informer and long short-term memory(LSTM)networks.The results indicate that the dual-channel prediction,which considers flight information and historical passenger flows,outperforms the single-channel prediction solely based on historical passenger flows.The results also indicate that the stacking model demonstrates superior performance across all metrics.Particularly,the best prediction performance is achieved at a 96 step(24 h)forecast horizon,with mean absolute error of 7.66 and 4.67 for inbound and outbound passenger flow predictions,respectively.Analysis of the impact of flight information characteristics on the prediction model reveals that departure flight information is of relatively lower importance than that of arrival flights,which is attributed to large differences in advance arrival times for departing passengers.

关 键 词：机场线 短时客流预测 Stacking集成模型 航班信息 

分 类 号：U121[交通运输工程]