基于多阶段特征优选的高速铁路列车晚点预测模型  

Train Delay Prediction Model of High-Speed Railway Based on Multi-Stage Feature Optimization

作　　者：李建民 许心越[1] 丁忻 LI Jianmin;XU Xinyue;DING Xin(State Key Laboratory of Rail Traffic Control and Safety,Beijing Jiaotong University,Beijing 100044,China;School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学轨道交通控制与安全国家重点实验室,北京100044 [2]北京交通大学交通运输学院,北京100044

出　　处：《中国铁道科学》2023年第4期219-229,共11页China Railway Science

基　　金：中央高校基本科研业务费专项资金资助项目(2020YJS208);国家重点研发计划项目(2018YFB1201403);教育部人文社科基金资助项目(18YJCZH176)。

摘　　要：为克服大规模高维数据集不相关和冗余信息对列车晚点预测模型性能的影响,提出一种融合多阶段(MS)特征优选方法和改进深度神经网络(IDNN)模型的高速铁路列车晚点预测模型(简称MS-IDNN模型)。首先,利用MS特征优选方法,基于列车运行实绩提取影响列车晚点的相关特征,构建初始特征集,并对其进行数据清洗和特征优选,生成最优特征子集;其次,将列车晚点特征映射为IDNN模型的神经元,采取全连接方式提取特征间的交互关系,并叠加多个浅层神经网络以克服深度神经网络反向传播过程中梯度消失的缺陷,实现列车到达晚点的精准预测;最后,以武广高速铁路列车运行实绩为例,验证MS-IDNN模型的有效性。结果表明:相比初始特征集,构建得到的最优特征子集特征维度降低了54.29%;相比6种基线模型,MS-IDNN模型的平均绝对误差和均方根误差分别至少降低4.85%和8.97%,在沿线至少66.66%的车站中表现出更高的预测性能;MS-IDNN模型能够有效剔除数据集中的不相关和冗余信息,提升列车晚点预测精度。In order to overcome the influence of irrelevant and redundant information in large-scale high-dimensional data sets on the performance of train delay prediction models,a train delay prediction model of high-speed railway(MS-IDNN model for short) is proposed,which combines a multi-stage(MS) feature optimization method and an improved deep neural network(IDNN) model.Firstly,the MS feature optimization method is used to extract relevant features that may affect the train delay based on the actual train operation performance,and the initial feature set is constructed.Then,the data cleaning and feature optimization are carried out to generate the optimal feature subset.Secondly,the train delay features are mapped to neurons of the IDNN model,and the interactive relationship between features is extracted through fully connected approach.Moreover,multiple shallow neural networks are superimposed to overcome the defects of gradient disappearance in the process of back propagation of deep neural networks,so as to achieve accurate prediction of train arrival delay.Finally,the effectiveness of MS-IDNN model is verified by taking the actual train operation performance of Wuhan-Guangzhou High-Speed Railway as an example.The results show that compared with the initial feature set,the feature dimension of the optimal feature subset derived from construction is reduced by 54.29%.Compared with the six baseline models,the mean absolute error and root mean square error of MSIDNN model are reduced by at least 4.85% and 8.97%,respectively,showing higher prediction performance in at least 66.66% stations along the line.In short,MS-IDNN model can effectively eliminate irrelevant and redundant information in the data sets and improve the prediction accuracy of train delay.

关 键 词：高速铁路 晚点预测 多阶段特征优选 深度神经网络 反向传播 

分 类 号：U292.4[交通运输工程—交通运输规划与管理]