计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (12): 293-300.DOI: 10.3778/j.issn.1002-8331.2203-0274

• 工程与应用 • 上一篇    下一篇

基于WCGAN的出租车需求热点预测

王博伟,邓君,吕斌   

  1. 1.兰州交通大学 交通运输学院,兰州 730070
    2.中交城市轨道交通设计研究院有限公司,武汉 430056
  • 出版日期:2023-06-15 发布日期:2023-06-15

Hotspot Forecast of Taxi Demand Based on WCGAN

WANG Bowei, DENG Jun, LYU Bin   

  1. 1.School of Traffic and Transportation, Lanzhou Jiaotong University, Lanzhou 730070, China
    2.CCCC Urban Rail Transit Design and Research Institute Co., Ltd., Wuhan 430056, China
  • Online:2023-06-15 Published:2023-06-15

摘要: 出租车空驶影响交通资源分配,同时影响出租车司机收益,准确的出租车需求热点预测可以有效地指导出租车驾驶员对寻客路线进行优化,针对该问题提出了基于条件生成对抗网络的出租车需求热点预测方法。该方法在Wasserstein条件生成对抗网络(WCGAN)的基础上,利用生成器中的卷积长短时记忆神经网络(ConvLSTM)捕捉时间序列的长期依赖关系,分别利用时间判别器和空间判别器提取乘客历史需求分布时空特性。利用兰州市出租车轨迹数据,将提出的方法与长短时记忆神经网络(LSTM)算法、时空残差网络(ST-ResNet))和BP神经网络(BPNN)三种算法进行对比,平均绝对误差分别降低了17.3%、8.4%和10.3%。

关键词: 需求热点预测, 生成对抗网络, 卷积长短时记忆神经网络(ConvLSTM)

Abstract: Empty taxis affect the allocation of traffic resources and the income of taxi drivers. Accurate taxi demand hot spot prediction can effectively guide taxi drivers to optimize the route for passengers. Aiming at this problem, a taxi based on conditional generative adversarial network is proposed. Demand hot spot prediction method, which is based on the wasserstein conditional generative adversarial network(WCGAN), uses the convolutional long short-term memory neural network(ConvLSTM) in the generator to capture the long-term dependencies of the time series, and uses the temporal discriminator and spatial discriminator respectively. The discriminator extracts the spatio-temporal characteristics of passenger historical demand distribution. Using Lanzhou taxi trajectory data, the method proposed in this paper is compared with three algorithms:long short-term memory neural network(LSTM) algorithm, spatio-temporal residual network(ST-ResNet) and BP neural network(BPNN). The average absolute error is reduced by 17.3%, 8.4% and 10.3% respectively.

Key words: hot spots prediction, generative adversarial networks(GAN), convolutional long short-term memory neural network(ConvLSTM)