| 摘 要: 为更全面地建模交通数据中的空间相关性,实现更准确地预测高速公路流量,提出一种自适应时空图卷积循环神经网络。利用扩散图卷积和自适应邻接矩阵整合成的自适应图卷积网络建模交通数据中的空间相关性,采用门控循环单元学习交通数据中的时间相关性,实现多时间步车流量预测。基于桂林市高速公路网真实交通数据集的实验结果表明,相比用于对比的最优方法,该方法的三个误差评价指标,即平均绝对误差、均方根误差、平均绝对百分比误差分别降低了约17.6%、18.6%和10.8%,优于用于对比的方法,该方法可以更准确地预测高速公路的流量。 |
| 关键词: 智能交通;流量预测;自适应图卷积网络;时空相关性 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 北京市自然科学基金(L201015) |
|
| Adaptive Spatio-Temporal Graph Convolutional Recurrent Neural Network for Expressway Traffic Flow Prediction |
|
RUAN Hongzhu1, WANG Jinbao1, DU Menghui2
|
(1.Yunnan Transportation Development Center, Kunming 650031, China;
2.Beijing Jiaotong University, Beijing 100044, China)
qtzrhz@163.com; 583454013@qq.com; 21140020@bjtu.edu.cn
|
| Abstract: In order to more comprehensively model the spatial correlation in traffic data and achieve more accurate expressway traffic flow prediction, this paper proposes an adaptive spatio-temporal graph convolutional recurrent neural network. The adaptive graph convolutional network which is integrated by the diffusion graph convolution and the adaptive adjacent matrix is used to model the spatial correlation in traffic data. Gated recurrent unit is adopted to learn the temporal correlation in traffic data, and the multi-time step traffic flow prediction is realized. The experimental results based on the real traffic dataset of Guilin expressway network show that, compared with other optimal methods used for comparison, the three error evaluation indicators of the method, namely, the average absolute error, the root mean square error, and the mean absolute percentage error, have decreased by about 17.6% , 18.6% , and 10.8% respectively, which is superior to methods for comparison. Therefore, this method can achieve more accurate expressway traffic flow prediction. |
| Keywords: intelligent transportation; flow prediction; adaptive graph convolution network; spatiotemporal correlation |
