基于时空多图卷积网络的网约车乘客需求预测
黄 昕(1998—),女,湖北黄梅人,硕士生,主要从事深度学习、交通流预测等研究。E-mail: huang_xin1998@163.com |
收稿日期: 2022-06-09
修回日期: 2022-09-07
网络出版日期: 2023-04-19
基金资助
国家自然科学基金项目(41801324)
国家自然科学基金项目(41701491)
福建省自然科学基金面上项目(2019J01244)
福建省自然科学基金面上项目(2019J01791)
Prediction of Passenger Demand for Online Car-hailing based on Spatio-temporal Multi-graph Convolution Network
Received date: 2022-06-09
Revised date: 2022-09-07
Online published: 2023-04-19
Supported by
National Natural Science Foundation of China(41801324)
National Natural Science Foundation of China(41701491)
General Project of Nat ural Science Foundation of Fujian Province(2019J01244)
General Project of Nat ural Science Foundation of Fujian Province(2019J01791)
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GST-MGCN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。本研究证明提出的模型GST-MGCN可以有效挖掘网约车乘客出行需求的时空模式,提取全局特征的影响,对其进行准确的预测。
黄昕 , 毛政元 . 基于时空多图卷积网络的网约车乘客需求预测[J]. 地球信息科学学报, 2023 , 25(2) : 311 -323 . DOI: 10.12082/dqxxkx.2023.220397
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.
表1 不同方法在海口数据集上的预测结果Tab. 1 Prediction results of different methods on the Haikou dataset |
方法 | MAE | RMSE | MAPE/% |
---|---|---|---|
XGBoost | 6.146 | 8.565 | 32.255 |
GRU | 4.823 | 5.289 | 28.780 |
LSTM | 4.832 | 5.214 | 29.028 |
GCN | 5.741 | 5.887 | 30.038 |
ConvLSTM | 3.449 | 4.123 | 25.297 |
ST-MGCN | 3.038 | 4.036 | 23.180 |
T-GCN | 2.873 | 3.960 | 22.291 |
Graph WaveNet | 2.635 | 3.985 | 23.612 |
GST-MGCN | 2.269 | 3.917 | 21.447 |
表2 时间相关性GST-MGCN模型变体在海口数据集上的预测结果Tab. 2 Prediction results of the time-dependent GST-MGCN model variants on the Haikou dataset |
方法 | MAE | RMSE | MAPE/% |
---|---|---|---|
GST-MGCNh | 3.823 | 4.289 | 24.780 |
GST-MGCNd | 5.832 | 6.214 | 28.028 |
GST-MGCN | 2.269 | 3.917 | 21.447 |
图8 GST-MGCN模型及其空间相关性模型变体在验证集上的不同轮数的性能比较Fig. 8 Performance comparison of the GST-MGCN model and its spatial correlation model variants at different epochs on the validation set |
表3 空间相关性GST-MGCN模型变体在海口数据集上的预测结果Tab. 3 Prediction results of spatially correlated GST-MGCN model variants on the Haikou dataset |
方法 | MAE | RMSE | MAPE/% |
---|---|---|---|
GST-MGCNga-fs | 2.941 | 4.788 | 24.038 |
GST-MGCNga-ci | 2.469 | 4.036 | 23.180 |
GST-MGCNfs-ci | 2.489 | 4.423 | 23.297 |
GST-MGCN | 2.269 | 3.917 | 21.447 |
表4 全局特征GST-MGCN模型变体在海口数据集上的预测结果Tab. 4 Prediction results of global feature GST-MGCN model variants on the Haikou dataset |
方法 | MAE | RMSE | MAPE/% |
---|---|---|---|
GST-MGCNsum | 2.573 | 3.960 | 22.291 |
GST-MGCNgat | 3.335 | 4.985 | 25.612 |
GST-MGCN | 2.269 | 3.917 | 21.447 |
表6 不同实验区域在海口数据集上的预测结果Tab. 6 Prediction results of different experimental regions on the Haikou dataset |
实验区域 | MAE | RMSE | MAPE/% |
---|---|---|---|
实验区域1 | 2.361 | 3.942 | 22.189 |
实验区域2 | 2.482 | 4.131 | 22.931 |
原始研究区 | 2.269 | 3.917 | 21.447 |
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
宋龙泽, 万怀宇, 郭晟楠, 等. 面向出租车空载时间预测的多任务时空图卷积网络[J]. 计算机科学, 2021, 48(7):112-117.
[
|
[8] |
冯宁, 郭晟楠, 宋超, 等. 面向交通流量预测的多组件时空图卷积网络[J]. 软件学报, 2019, 30(3):759-769.
[
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
李雪峰. 混合需求条件下城市网约车需求预测及派单方法研究[D]. 南京: 东南大学, 2021.
[
|
[27] |
|
/
〈 | 〉 |