交通运输系统工程与信息

交通运输系统工程与信息 ›› 2023, Vol. 23 ›› Issue (3): 153-160.DOI: 10.16097/j.cnki.1009-6744.2023.03.017

• 系统工程理论与方法 • 上一篇    下一篇

基于集成学习和居民属性数据的出行方式预测模型

苏跃江*1,2，温惠英1，袁敏贤2，吴德馨2，漆巍巍1   

  1. 1.华南理工大学，土木与交通学院，广州510641；2.广州市交通运输研究院有限公司，广州510635
  • 收稿日期:2023-02-04 修回日期:2023-04-06 接受日期:2023-04-10 出版日期:2023-06-25 发布日期:2023-06-23
  • 作者简介:苏跃江(1983-)，男，贵州六盘水人，高级工程师，博士生
  • 基金资助:
    国家自然科学基金面上项目 (52072131);广州市科技计划项目 (202206010056)

Travel Mode Prediction Model Based on Ensemble Learning and Resident Attribute Data

SU Yue-jiang*1,2, WEN Hui-ying1, YUAN Min-xian2, WU De-xin2, QI Wei-wei1   

  1. 1. School of Civil Engineering and Transportation, South China University of Technology, Guangzhou 510641, China; 2. Guangzhou Transport Research Institute Co. LTD, Guangzhou 510635, China
  • Received:2023-02-04 Revised:2023-04-06 Accepted:2023-04-10 Online:2023-06-25 Published:2023-06-23
  • Supported by:
    National Natural Science Foundation of China (52072131);Guangzhou Science and Technology Planning Project (202206010056)

PDF

摘要： 传统的效用理论模型以整体效用最大化为目标，忽略了影响居民出行方式选择肢间可能存在相关性以及部分效用相互补偿性。基于集成学习的居民出行方式预测模型可以有效表达出行效用的个体偏好、差异性和补偿性，解决了不同出行特征影响因素对于不同人群有不同效用表征问题。本文依托大样本居民属性数据，针对不同个体对不同影响因素的感知差异，结合居民出行调查的个人属性、出行属性及环境属性等相关影响因素进行综合分析，构建个体级居民的出行特征向量；研究构建基于集成学习方法的居民出行方式预测模型；以广州市为例进行实证分析，通过准确率、精确率、召回率和F1值这4个指标，对模型预测结果进行综合评价；针对部分出行属性相似导致细分的出行方式判别精度不高，按照大类方式、慢行交通、公共交通、个体机动化交通这4个类别构建层次化LightGBM(Light Gradient Boosting Machine)模型，并对GBDT(GradientBoosting Decision Tree)Random Forest、LightGBM、层次化LightGBM这4个模型预测结果进行综合对比。分析结果表明：LightGBM模型预测慢行交通、公共交通、个体机动化等三大类出行方式的平均准确率约为87%，其中慢行交通方式预测准确率最高，为92%；预测精细化出行方式平均准确率约为78%，其中步行方式预测准确率最高，为89%；LightGBM模型适用于城市级的居民出行方式预测，是一种有效的居民出行方式选择的预测方法。

关键词: 城市交通, 集成学习, 居民属性数据, 特征向量, 出行方式预测模型, LightGBM

Abstract: The traditional model of utility theory aims to maximize the utility on a global level. However, the traditional method may disregard some influencing factors of residents' travel mode and the mutual compensation of utilities. The individual preference, difference and compensation of individual travel utility could be expressed effectively by resident travel mode prediction model based on integrated learning. The problem of different utility representation was solved due to different travel characteristics on different groups. First, the travel feature vector of individual residents was constructed by analyzing the factors of personal attributes, travel attributes and environmental attributes. The difference in the perception of utility of different individuals to different influencing factors was considered. The travel mode prediction model based on ensemble learning method was then developed. Taking Guangzhou as an example, the prediction results were evaluated comprehensively by four parameters: accuracy rate, precision rate, recall rate and F1 value. Due to the similarity of some travel mode attributes, the hierarchical LightGBM models were developed and optimized, including all-mode, slow traffic, public transport and individual motorized traffic. The results were compared with other models, including GBDT, Random Forest, and hierarchical LightGBM. The results indicated that the weighted average accuracy of LightGBM model to predict travel mode is respectively 87% and 78%. The highest accuracy of slow traffic mode is 92% in the big class model, and the highest accuracy of walking mode is 89% in the subdivided class model. The model would be suitable for urban level residents' travel mode prediction and play an important role to predict the residents' travel mode.

Key words: urban traffic, ensemble learning, resident attribute data, feature vector, travel mode prediction model; LightGBM

中图分类号: