基于CNN-GRU模型的中欧班列运到时限预测  

A CNN-GRU hybrid model for travel time prediction of China-Europe Express Railway

作　　者：张永祥 谷丽婷[1,2] 郭经纬 闫旭 冯涛 钟庆伟[5] ZHANG Yongxiang;GU Liting;GUO Jingwei;YAN Xu;FENG Tao;ZHONG Qingwei(School of Transportation and Logistics,Southwest Jiaotong University,Chengdu 610031,China;National United Engineering Laboratory of Integrated and Intelligent Transportation,Southwest Jiaotong University,Chengdu 610031,China;Faculty of Business,City University of Macao,Macao 999078,China;Department of Logistics and Maritime Studies,The Hong Kong Polytechnic University,Hongkong 999077,China;College of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618307,China)

机构地区：[1]西南交通大学交通运输与物流学院,四川成都610031 [2]西南交通大学综合交通运输智能化国家地方联合工程实验室,四川成都610031 [3]澳门城市大学商学院,中国澳门999078 [4]香港理工大学物流及航运学系,中国香港999077 [5]中国民航飞行学院空中交通管理学院,四川广汉618307

出　　处：《铁道科学与工程学报》2024年第10期3989-4001,共13页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(72201218,72201268);四川省科技计划资助项目(2023NSFSC0901);中国国家铁路集团科技研究开发计划(P2022X013)。

摘　　要：随着经济贸易的全球化发展,中欧班列已经成为重要的国际货运方式。由于中欧班列的运到时限受诸多因素影响,现有模型难以充分捕捉运输时间数据的复杂特征关系,因而无法准确预测列车运到时限,影响中欧班列的调度及沿线运力的安排。提出一种基于卷积神经网络(CNN)和门控循环单元(GRU)的中欧班列运到时限预测方法,该方法能有效捕捉运到时间序列的空间及时间特征,从而提高预测精度。所提方法首先利用小波变换技术对中欧班列运到时限历史数据进行降噪处理,再经过最大−最小归一化、多粒度扫描窗及数据划分后,通过一维CNN模块提取输入时间序列的空间特征,GRU模块提取序列的时间特征,最后输出中欧班列运到时限的预测值。在实验部分进行了模型的参数调优、小波变换去噪效果分析及模型对比。结果显示,经小波变换去噪处理后,CNN-GRU模型的均方根误差(RMSE)和平均绝对误差(MAE)分别降低了34.18%和26.77%;模型的RMSE和MAE比单一模型中预测效果表现最好的随机森林(RF)分别降低了17.28%和21.67%,比组合模型CNN-LSTM分别降低了5.68%和15.70%。本文所构建的CNN-GRU模型对于小样本复杂数据的预测性能较高,且模型训练参数较少,轻量化程度较高,可解释性较强。基于该模型的中欧班列运到时限预测方法提高了运到时限预测的准确性,能够为缓解中欧班列路网运力不足等现状提供较为可靠的技术支持。As economic and trade globalization progresses,the China-Europe Express Railway has emerged as a key mode of international freight transport.However,various factors impact the travel time of the China-Europe Express Railway,making the existing models challenging to fully capture the complex characteristic relationship of the travel time data.Consequently,these models struggle to predict train travel times accurately,which in turn affects the scheduling and capacity management of the China-Europe Express Railway.A deep learning approach for predicting train travel times,which utilizes a convolutional neural network(CNN)and a gated cycle unit(GRU),was proposed in this paper.This methodology effectively captures both spatial and temporal characteristics inherent in travel time series,thereby enhancing prediction accuracy.First,a wavelet transform technology was employed to denoise historical travel time data from the China-Europe Express Railway.After performing the Min-Max normalization,multi-grained scanning window and data division,the spatial features of the input time series were extracted by one-dimensional CNN module,and the time features of the data were extracted by GRU module.The model ultimately outputs the predicted travel time for the China-Europe Express Railway.In the experimental section,the study conducted model parameter tuning,analyzed the effectiveness of wavelet transform denoising,and compared various models.The results indicate that the root mean square error(RMSE)and mean absolute error(MAE)of the CNN-GRU model decrease by 34.18%and 26.77%,respectively,followed by wavelet transform denoising.As compared to the random forest model,which previously demonstrated the best predictive performance among single models,the CNN-GRU model shows reductions in RMSE and MAE by 17.28%and 21.67%,respectively.Additionally,in comparison to the combined CNN-LSTM model,the CNN-GRU model exhibits reductions in RMSE and MAE of 5.68%and 15.70%,respectively.In conclusion,the CNN-GRU model demonstrates superior predi

关 键 词：铁路运输 中欧班列 列车运到时限预测 CNN-GRU 小波变换 

分 类 号：U292.4[交通运输工程—交通运输规划与管理]