面向高速铁路货运站选址的深度学习模型研究  

Research on Deep Learning Model for Location of High-speed Railway Freight Stations

作　　者：郑倩 甘蜜[1,2,3] 姚竹 魏力飞 ZHENG Qian;GAN Mi;YAO Zhu;WEI Lifei(School of Transportation and Logistics,Southwest Jiaotong University,Chengdu 610031,China;National Engineering Laboratory of Application Technology of Integrated Transportation Big Data,Chengdu 611756,China;National United Engineering Laboratory of Integrated and Intelligent Transportation,Chengdu 611756,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]综合交通大数据应用技术国家工程实验室,四川成都611756 [3]综合交通运输智能化国家地方联合工程实验室,四川成都611756

出　　处：《铁道学报》2024年第6期36-45,共10页Journal of the China Railway Society

基　　金：国家自然科学基金(52372306);四川省科技厅项目(2024YFHZ0315);中央高校基本业务经费(2682023JX007,2682023ZTPY029)。

摘　　要：近年来急剧增长的快递量催生了高速铁路快运模式,高速铁路建设的持续推进为高速铁路快运体系的完善创造了新的发展契机。在规划过程中,高速铁路货运站的选址与特征之间存在复杂的异质性和非线性关系,因此本研究开发了一种基于深度学习的高速铁路货运站选址方法,该方法旨在根据现有的高速铁路快运物流基地建设项目挖掘出高速铁路货运站潜在的选址规律。研究初步构造市场需求、区位交通、发展环境3个维度的选址条件及其下属的10个指标,基于基础特征与深度学习算法构建选址模型,利用遗传算法优化超参数,同时根据各特征的重要度依次对模型进行测试以此确定出最佳的特征体系和模型学习层。实验结果表明,车站节点覆盖率、动车所预留情况、城市快递业务量、公路交通可达性和周边机场数目构成的特征子集在以LSTM和Dense堆叠的深度学习模型中表现出了最佳的选址效果,其F1分数达到0.9444,相比优化前有了较大幅度的提高。进一步以浙江省为例进行实证分析,选址结果也证实了研究所建立的深模型较好地学习了实际高速铁路货运站选址规律,可以帮助更加科学化、结构化的高速铁路货运站选址决策。The rapid growth of express delivery volume in recent years has resulted in the high-speed railway(HSR)express mode.However,in the planning process,there is a complex heterogeneous and non-linear relationship between the site selection of HSR freight stations and logistics demand and facility conditions.In this context,this study developed a deep learning-based method for selecting the location of HSR freight stations.By initially constructing the site selection model based on three dimensions:market demand,geological location and transportation,and development environment,as well as 10 basic features and deep learning algorithm,this study optimized the hyperparameters by using genetic algorithm,and tested the model according to the importance of each feature in order to determine the best feature system and model learning layer.The experimental results show that the feature subset composed of station node coverage,high-speed train reservation,urban express delivery volume,road traffic accessibility,and the number of surrounding airports shows the best site selection performance in the deep learning model stacked with LSTM and Dense,with F 1 score up to 0.9444,a significant improvement compared to that before optimization.This paper further takes Zhejiang province as an example to empirically analyze the optimized deep learning model,and the site selection results also confirm that the deep learning model established by the study learns the actual HSR freight station site selection law sufficiently and can help to make more scientific and structured HSR freight station site selection decisions.

关 键 词：高速铁路货运站 多源数据 深度学习 超参数优化 长短期记忆网络 

分 类 号：U294.51[交通运输工程—交通运输规划与管理]