《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (12): 3941-3949.DOI: 10.11772/j.issn.1001-9081.2021101718

• 前沿与综合应用 • 上一篇    

基于深度聚合神经网络的网约车需求时空热度预测

郭羽含, 田宁()   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 收稿日期:2021-10-08 修回日期:2022-01-13 接受日期:2022-01-14 发布日期:2022-01-24 出版日期:2022-12-10
  • 通讯作者: 田宁
  • 作者简介:郭羽含(1983—),男,黑龙江哈尔滨人,副教授,博士,CCF会员,主要研究方向:智能搜索算法、车辆调度问题、供应链优化问题
  • 基金资助:
    辽宁省自然科学基金资助项目(2019?ZD?0048);辽宁省教育厅基础研究项目(LJ2019JL012)

Spatio-temporal heat prediction of online car‑hailing demand based on deep aggregated neural network

Yuhan GUO, Ning TIAN()   

  1. School of Software,Liaoning Technical University,Huludao Liaoning 125105,China
  • Received:2021-10-08 Revised:2022-01-13 Accepted:2022-01-14 Online:2022-01-24 Published:2022-12-10
  • Contact: Ning TIAN
  • About author:GUO Yuhan born in 1983, Ph. D., associate professor. His research interests include intelligent search algorithm, vehicle scheduling problem, supply chain optimization problem.
  • Supported by:
    Natural Science Foundation of Liaoning Province(2019-ZD-0048);Basic Research Project of Educational Department of Liaoning Province(LJ2019JL012)

摘要:

为解决服务车辆与乘客间的供需不平衡问题,提升服务车辆的运营效率和利润,同时降低乘客等待时间并改善其对服务平台的满意度,针对差异化结构的多维时空数据,提出一种深度聚合神经网络(DANN)模型用于对网约车需求进行预测。首先,通过综合考虑时间、空间和外部环境等多维影响因素,提出了基于周期的时空变量和基于图像点值的空间变量划分方法;其次,依据数据特点构建了不同的子神经网络结构来分别拟合时间变量、空间变量和环境变量与需求间的非线性关系;然后,提出了多种异类子神经网络的聚合方法以同时捕捉不同结构时空数据的隐含特征;最后,分析了聚合权重的设置方法以获得网络模型的最优性能。实验结果表明,在三个真实数据集上所提模型的R2平均误差仅为9.36%,与卷积长短时记忆网络(FCL-Net)和混合深度学习神经网络(HDLN-Net)模型相比,所提模型的R2分别平均提升了4.6%和5.22%,均方误差(MSE)分别平均降低了27.01%和26.6%。因此,DANN在实际应用中能较大幅度地提升需求预测的准确性,可以作为网约车需求预测的有效手段。

关键词: 城市交通, 需求预测, 时空数据, 深度神经网络, 网约车

Abstract:

To solve the supply-demand imbalance between service vehicles and passengers, improve the operational efficiency and profit of service vehicles, and reduce passengers' waiting time as well as improve their satisfaction with the service platform at the same time, a Deep Aggregation Neural Network (DANN) model was proposed for predicting the demand of online car-hailing aiming at the multi-dimensional spatio-temporal data with differentiated structures. Firstly, a period-based spatio-temporal variable classification method and a spatial variable classification method based on image point values were proposed by considering multi-dimensional influencing factors such as time, space, and external environment comprehensively. Secondly, different sub neural network structures were constructed to fit the nonlinear relationships between temporal, spatial, environmental variables and the demand respectively based on data characteristics. Thirdly, an aggregation method of multiple heterogeneous sub neural networks was proposed to simultaneously capture the implicit features of spatio-temporal data with different structures. Finally, a method of setting aggregation weights was analyzed to obtain the optimal performance of the network model. Experimental results show that the proposed model has the average error of R2 on three real-world datasets of 9.36%, and compared with the Fusion Convolutional Long Short-Term Memory Network (FCL-Net) and Hybrid Deep Learning Neural Network (HDLN-Net) models, the proposed model has the R2 increased by 4.6% and 5.22% on average respectively, and the Mean Square Error (MSE) reduced by 27.01% and 26.6% on average respectively. Therefore, DANN can greatly improve the accuracy of demand prediction in practical applications and can be used as an effective means of demand prediction for online car-hailing.

Key words: urban traffic, demand prediction, spatio-temporal data, deep neural network, online car-hailing

中图分类号: