Abstract:

Accurately predicting taxi demand can reduce the waiting time of taxi driver and passengers, and alleviate traffic in cities. However, most of the current research work ignores the impact of taxi historical inflows and interregional potential spatial dependency on taxi demand. In view of this, this paper considers multiple factors affecting taxi demand and establishes Multi-Factors Spatio-Temporal Graph Convolutional Network（MFSTGCN） in attempt to improve the accuracy of taxi demand prediction. Specifically, MFSTGCN designs four components to model the temporal dependence of taxi demand on adjacent time for demand sequence, daily demand sequence, historical inflow sequence and the potential spatial dependence between different regions. These components can capture the potential spatio-temporal representations of different factors through spatio-temporal convolutional block. To prove the performance of the proposed MFSTGCN, this paper compares it with five reference models, which are commonly used in traffic predication. Three metrics of RMSE, MAE and MAPE are applied for model evaluation. It demonstrates MFSTGCN’s effectiveness.

Key words: taxi demand, potential spatial dependency, spatio-temporal convolutional block

摘要：

精确地预测出租车需求，有助于缩短司机和乘客的等待时间，缓解交通拥堵状况。然而，目前的研究工作大都忽略了出租车历史流入量以及不同区域间潜在的空间依赖关系对出租车需求的影响。鉴于此，综合考虑多个影响出租车需求的因素，建立多因素时空图卷积网络模型（MFSTGCN），期望精确地预测出租车需求。具体地，MFSTGCN模型设计4个组件，分别建模出租车需求对临近时刻需求序列、日需求序列、出租车历史流入量序列的时间依赖性以及不同区域间的潜在空间依赖性。组件通过时空卷积块捕捉不同因素下的潜在时空表达。为了证明MFSTGCN模型的有效性，与交通预测常用的五种基准模型进行比较，并利用RMSE、MAE和MAPE三个指标进行评估。实验结果表明，融合多因素的MFSTGCN模型能够更精确地预测出租车需求。

关键词: 出租车需求, 潜在空间依赖性, 时空卷积块