基于时空多图卷积网络的网约车乘客需求预测

随着智能手机的普及,网约车成为常用的出行替代方式。网约车运营平台因此成为智能交通系统的主要组成部分,在满足公众出行需求中发挥重要作用。乘客需求预测是网约车系统需要解决的核心问题,现有文献中提出的模型忽略了长期时间相关性及多种空间相关性,本文针对现有研究成果存在的局限性,在充分考虑网约车乘客出行需求时空相关独特性的基础上,提出一种融合全局特征的时空多图卷积网络(Spatio-Temporal Multi-Graph Convolutional Network Fused With Global Features,GST-MGCN)模型。该模型遵循临近性、周期性和趋势性(Closeness, Period and Trend,CPT)范式,利用时序信息拟合时间依赖关系;通过识别多种空间语义相关性构建对应的关系图结构、建立多图卷积模型;模型中的全局特征融合模块,使用门控融合和总和融合方法分别捕捉乘客需求的突变和渐变。以海口市数据集为样本的实验结果表明,本文提出的GSTMGCN模型MAE、RMSE和MAPE指标的值分别是2.269、3.917、21.447,优于其他同类主流模型。本研究证明提出...

Prediction of Passenger Demand for Online Car-hailing based on Spatio-temporal Multi-graph Convolution Network

With the popularization of smartphones, online car-hailing has become a common travel alternative and plays an important role in meeting public travel demand. Therefore, online car-hailing operation platforms have been a major component of Intelligent Transportation Systems in which passenger demand prediction is one of the core problems to be solved. However, models proposed in the existing literature usually ignore the long-term temporal correlation and multiple spatial correlations. This paper presented a Spatio-Temporal Multi-Graph Convolutional Network Fused With Global Features (GST-MGCN) to address the limitations of existing research achievements, taking full account of the unique spatiotemporal correlations of the travel demand of online car-hailing passengers. Following the Closeness, Period, and Trend (CPT) paradigm, the model fitted temporal dependencies with time series information. By identifying multiple spatial semantic correlations, the corresponding relational graph structure was constructed, and a multi-graph convolutional model was built in which the global features fusion module employed gated fusion and sum fusion methods to capture sudden and gradual changes of passenger demand, respectively. Taking the Haikou city dataset as an example, our experimental results show that the values of the three indicators, MAE, RMSE, and MAPE of the GST-MGCN model proposed in this paper were 2.269, 3.917, and 21.447, respectively, which were lower than those derived from other similar mainstream models. This study demonstrated that the proposed model GST-MGCN can effectively mine the spatio-temporal pattern of online car hailing passenger travel demand, extract the impact of global features, and accurately predict it.