交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (4): 193-205.DOI: 10.16097/j.cnki.1009-6744.2025.04.018

• 系统工程理论与方法 • 上一篇    下一篇

多特征融合的网约车拼车起讫点需求时空预测

谢秉磊1 ，冯健茜1 ，秦筱然*2   

  1. 1. 哈尔滨工业大学(深圳)，建筑学院，广东深圳518055；2.华南理工大学，土木与交通学院，广州510641
  • 收稿日期:2025-03-07 修回日期:2025-03-26 接受日期:2025-04-09 出版日期:2025-08-25 发布日期:2025-08-25
  • 作者简介:谢秉磊(1975—)，男，陕西凤县人，教授，博士。
  • 基金资助:
    国家自然科学基金 (71974043)。

Spatio-temporal Prediction of Origin-destination Demands for Ride-pooling Considering Multiple Features

XIE Binglei1, FENG Jianxi1, QIN Xiaoran*2   

  1. 1. School of Architecture, Harbin Institute of Technology (Shenzhen), Shenzhen 518055, Guangdong, China; 2. School of Civil Engineering and Transportation, South China University of Technology, Guangzhou 510641, China
  • Received:2025-03-07 Revised:2025-03-26 Accepted:2025-04-09 Online:2025-08-25 Published:2025-08-25
  • Supported by:
    National Natural Science Foundation of China (71974043)。

PDF

摘要： 针对现有拼车需求预测研究中对拼车订单关联性考虑不足的问题，本文提出一种基于时空多图卷积神经网络的拼车起讫点需求时空预测模型。首先，将拼车订单数据处理为各个起讫点之间需求的时序数据，分析并挖掘拼车需求的多个特征信息。在此基础上，根据拼车拼成率、路径相似性和土地功能相似性构建多个反映拼车订单关联的语义图，为拼车需求预测提出新的语义建模视角；同时，构建反映相邻性和距离的地理图，从多维度对地理相关性与行程语义关联进行建模。在模型构建上，本文提出层级化多图信息融合机制，充分捕捉数据中的时空相关性。此外，引入影响拼车需求的外源因素，构建融合多特征的时空多图卷积模型。实验结果表明，拼车拼成率、路径相似性和天气是影响拼车需求的关键因素。相较于多图注意力卷积网络(GMAN)、时空长短期记忆网络(SP-LSTM)和残差多图卷积网络(RMGCN)，本文方法的预测结果均方根误差分别降低了11.44%、7.06%和3.89%，平均绝对误差分别降低了9.45%、10.85%和7.26%，表明本文提出的方法具有较高的预测精度和科学性。

关键词: 智能交通, 需求预测, 多图卷积神经网络, 网约车拼车, 门控循环单元, 起讫点预测

Abstract: To address the insufficient ride-pooling order correlations in existing ride-pooling demand prediction, this paper proposes a spatiotemporal prediction model for ride-pooling origin-destination demand based on a spatiotemporal multi-graph convolutional neural network. First, ride-pooling order data is processed into time-series data representing demand between origin destination points, and multiple characteristic features of ride-pooling demand are analyzed and extracted. Based on this, multiple semantic graphs reflecting ride-pooling order correlations are innovatively constructed using ride-pooling success rate, route similarity, and land-use similarity, with a novel semantic modeling perspective for ride-pooling demand prediction. Meanwhile, geographic graphs are built to capture adjacency and distance relationships, enabling a multidimensional modeling of geographical correlations and travel semantic associations. For model development, this paper proposes a hierarchical multi-graph information fusion mechanism to fully capture spatiotemporal correlations in the data. Additionally, exogenous factors affecting ride-pooling demand are incorporated to develop a spatiotemporal multi-graph convolutional model that integrates multiple features. Experimental results show that ride-pooling success rate, route similarity, and weather are key factors influencing ride-pooling demand. Compared to the Multi-Graph Attention Network (GMAN), Spatiotemporal Long Short-Term Memory Network (SP LSTM), and Residual Multi-Graph Convolutional Network (RMGCN), the proposed method reduces root mean square error by 11.44%, 7.06%, and 3.89%, respectively, and decreases mean absolute error by 9.45%, 10.85%, and 7.26%, respectively. These results demonstrate that the proposed method achieves higher prediction accuracy and scientific validity.

Key words: intelligent transportation, demand prediction, multi-graph convolutional neural network, ride-pooling in ride-hailing; gated recurrent unit, origin-destination prediction

中图分类号: