【作者】 黄平；

【导师】 王郴平； 彭其渊；

【作者基本信息】 西南交通大学 ， 交通运输规划与管理， 2020， 博士

【摘要】 解析高速列车晚点传播机理对估计晚点的影响、提高高速铁路智能调度指挥水平及运输服务质量具有重要作用。本文基于我国高速列车运行实绩,宏观与微观研究相结合,运用大数据、人工智能等技术,从晚点分布规律出发,分别研究晚点时间维传播机理及模型、晚点空间维传播机理及模型以及考虑晚点时空传播效应的故障宏观和微观影响预测模型。预期成果将能够丰富高速列车晚点传播理论和调度指挥理论,为实现高速铁路调度指挥智能化提供理论支撑。主要工作和结论如下:(1)基于已有研究和公开数据对列车晚点影响因素以及世界主要国家列车正点率进行统计分析,并基于我国高速列车运行实绩分析高速列车运行特征、晚点时空分布、晚点时长分布、列车区间运行时间分布及车站停站时间分布等特点,以了解我国高速列车运行基本信息及晚点宏观规律。(2)基于统计分析和机器学习方法研究列车晚点时间维传播机理及预测模型,从统计结果和初始晚点影响模型两个方面揭示晚点时间维传播机理。首先,基于列车运行实绩解析了晚点时间维传播机理;然后,将初始晚点根据其自身特性以及运行图参数进行聚类研究,得到具有不同特性的初始晚点类别。接着,建立各类别初始晚点的影响列车数以及晚点总时间概率分布模型以实现初始晚时间维传播预测。最后,利用测试数据验证所建立模型,其结果表明所建模型具有较强的实际应用能力。(3)基于统计分析和机器学习方法研究列车晚点空间维传播机理及预测模型,从统计结果、恢复模型和增晚模型三个方面揭示晚点空间维传播机理。首先,对列车晚点增加(增晚)和晚点减小(恢复)情况进行基本统计分析。然后,建立考虑区间及车站缓冲时间布局的晚点恢复预测随机森林模型,并基于晚点恢复数据验证晚点恢复模型,其结果表明:随机森林模型能够很好地拟合高速列车初始晚点恢复数据,对高速列车初始晚点恢复时间具有很高预测精度。最后,针对机器学习离线学习和在线学习在实时调度指挥预测中的缺点,基于支持向量机模型以及滤波技术建立了列车增晚预测模型;对模型的验证结果表明:滤波技术可以有效地解决机器学习模型在线学习耗时过多和离线学习无法准确预测增晚等问题。晚点恢复模型和增晚预测模型可以实现列车晚点空间维传播预测。(4)同时考虑列车晚点时空传播效应,基于贝叶斯网络建立故障宏观影响预测模型以揭示晚点时空传播机理。首先,确定初始晚点时间、影响列车数及影响总晚点时间三个指标以度量故障对高速列车运行的影响;然后,在考虑实时调度指挥需求基础上,结合领域(专家)知识以及数据驱动方法确定高速铁路故障影响预测的贝叶斯网络模型结构;最后,基于两条高速铁路线路列车运行实绩数据对模型测试结果表明:贝叶斯网络模型精度优于其它广泛应用的预测模型,并可以适用于不同速度等级的高速铁路线路。(5)同时考虑列车晚点时空传播效应,基于深度学习方法建立故障微观影响(列车晚点时间)实时预测模型以揭示晚点时空传播机理。首先考虑列车间相互作用关系提出一种基于深度学习的列车晚点预测模型(FCL-Net);该模型结合了全连接神经网络(FCNN)和长短记忆神经网络(LSTM)两种特定的结构以分别考虑晚点影响因素的列车运行相关变量和非列车运行相关变量;利用我国高速列车运行实绩对该模型的预测精度、扩展性和效率进行测试,并选择常用列车晚点预测模型作为基准模型与所提出模型进行预测精度对比;结果表明:相比于基准模型,提出的FCL-Net模型由于考虑了列车间相互作用关系具有较高的预测精度。然后,由于FCL-Net模型只考虑了列车的互相关关系,但列车运行不仅依赖于其前行列车,而且依赖于其自身过去的状态,故提出一种基于卷积神经网的深度学习模型(FCC-Net)来预测列车晚点;该模型同样结合两种神经网络(CNN和FCNN)以分别处理列车运行相关因素和非列车运行相关因素;FCC-Net模型将列车运行图视作图像,利用CNN模型的图像处理能力,同时考虑列车运行的自相关和互相关性,以进一步提高列车晚点模型预测精度;模型测试结果表明:由于同时考虑了列车运行的自相关和互相关关系,FCC-Net模型具有比FCL-Net模型更高的预测精度。更多还原

【Abstract】 Analyzing delay propagation patterns plays an important role in reducing the impact of delays and improving the level of intelligent dispatching and the quality of transport services in high-speed railway(HSR).Based on the operation records of high-speed trains in China,this thesis examined the delay distribution characteristics,delay propagation at time direction,delay propagation at space direction,and the macro and micro effects of disruptions on train operation considering delay propagation at time and space directions by using big data and artificial intelligence technologies.The expected results can enrich the theories of train delay propagation and train dispatching and commanding,and provide theoretical support for the realization of intelligent dispatching in high-speed railway.The main jobs and achievements in the thesis include:1)The influencing factors of train delays and the train punctuality in other countries were first examined from existing literatures and public data.In addition,based on the real-world data of high-speed trains in China,the operational features of high-speed trains,the spatiotemporal and duration distribution of delays,and the running time and dwelling time distribution of trains were investigated,which could provide the railway managers,passengers,and train dispatchers a macro understanding of the basic information of high-speed train delays.2)Patterns and predictive model for train delay propagation at time direction(the influence of the primary delay train on its succeeding trains)were studied with statistical methods and machine learning models to reveal the delay propagation mechanism at time dimension from the statistical results and the primary delay influencing models.First,the delay propagation mechanism at time direction was clarified.Then,the primary delays were clustered according to their own characteristics and timetable parameters.After clustering,probabilistic distribution models were established to examine the influence of primary delays of each category,including the number of affected trains(NAT)and the total delay times(TDT).Finally,the model validation results based on the future data showed that the model could well fit the data about NAT and TDT,and had strong application ability in the future prediction task.3)Patterns and predictive model for train delay propagation at space direction were studied with statistical methods and machine learning models to reveal the delay propagation mechanism at space dimension from the statistical results,recovery model and delay increase model.Firstly,the delay increasing and delay decreasing(recovery)characteristics in(at)section(station)were analyzed.Then,a predictive random forest model that considered the supplement times in section and at station and primary delay severity was established to predict the delay recovery of high-speed trains once they were delayed;the predictive results of the training and testing dataset showed that random forest model could well fit the delay recovery data,and had high predictive accuracy compared against other machine learning models.Finally,in view of the shortcomings of machine learning models by offline learning and online learning in real-time dispatching,a hybrid prediction model for delay increase was established combining support vector machine and filtering technology.The verification results of the hybrid model showed that filtering technology could effectively solve the shortcomings of machine learning model,including much time-consuming by online training and incapability of delay increase situation by offline training,and the proposed model could significantly improve the accuracy of a traditional machine learning model in delay increase prediction.Accurate prediction of train delay propagation at space direction was realized by combing the recovery model and delay increase model.4)Considering the spatiotemporal propagation of train delays,a real-time prediction model for the macro-effect of disruptions was established to reveal the spatialtemporal propagation mechanism of train delays with Bayesian networks(BN).First,the indexes for measuring the influence on disruptions on train operations were determined,including the primary delay time(L),the number of affected trains(N),and the total delay time(T).Then,the BN structure was determined by considering the real-time dispatching requirements,the domain(expert)knowledge,and the structure learned from data using a heuristic algorithm.Next,the BN model was tested using data from two high-speed railway lines,which showed that the proposed BN model,with superior performance compared against other prediction models,had high prediction accuracy about the factor L,N,and T,and it could be applied to high-speed railway lines with different operational features.5)Considering the spatiotemporal propagation of train delays,real-time prediction models for the micro effect of disruptions,i.e.,the train delay,were established to reveal the spatialtemporal propagation mechanism of train delays with deep learning methods.Firstly,considering the interaction between trains,a deep learning train delay prediction model,named FCL-Net,was proposed,which combines two specific neural network structures,i.e.,fully-connected neural networks(FCNN)and long short-term memory(LSTM),to separately consider the operational and non-operational factors of delay influencing factors.The prediction accuracy,expansibility and efficiency of the model were tested using the operation records of two HSR lines in China,and some common train delay prediction models were selected as the benchmark of the proposed model.The results showed that,compared against the benchmark model,the proposed FCL-Net model,which took train interactions into account,outperformed other widely used train delay prediction models on two HSR lines with different operational features,in terms of both regression and classification model evaluation metrics.However,the FCL-Net model only considered the correlation between trains,as the train operation depends not only on its preceding trains,but also on its own past states.To address this drawback,a deep learning model based on convolutional neural network(CNN),named,FCC-Net,was proposed to predict train delays.The model also combines two kinds of neural networks,i.e.,the FCNN and CNN,to deal with train delay influencing factors with different attributes separately.The FCC-Net model took timetables as images,and used the image processing ability of CNN model to capture the self-correlation and cross-correlation of train operations,for further improvement of train delay prediction.The model testing results showed that the FCC-Net,because of considering the self-correlation and cross-correlation of train operation,had higher prediction accuracy than FCL-Net model on data from two HSR lines.更多还原