Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (13): 269-275.DOI: 10.3778/j.issn.1002-8331.2006-0175

Previous Articles     Next Articles

Short-Time Traffic Flow Prediction of Graph Convolutional Network Considering Influence of Space and Time

CHEN Danlei, CHEN Hong, REN Anhu   

  1. School of Electronic Information Engineering, Xi’an University of Technology, Xi’an 710021, China
  • Online:2021-07-01 Published:2021-06-29

考虑时空影响下的图卷积网络短时交通流预测

陈丹蕾,陈红,任安虎   

  1. 西安工业大学 电子信息工程学院,西安 710021

Abstract:

To solve the problem that the spatial characteristics of urban road network are difficult to be fully extracted in the process of traffic flow prediction, which leads to the low accuracy of the prediction results, a short-time traffic flow prediction model combining Graph Convolutional Network(GCN) and Gated Recurrent Unit(GRU) is proposed. Based on the advantages of GCN in topological data processing, the spatial arrangement structure of urban road network is transformed into topological relation modeling, and the spatial characteristics of road networks are extracted effectively by solving the problem of topological relation. The GCN model is improved by using GraphSAGE algorithm, and spatial features are aggregated by adding aggregation operator and Graph Attention mechanism(GAT), and the output containing spatial features is used as the input of GRU model to extract time features. Compared with the model without GCN, the result shows that the prediction accuracy of this model is about 8% higher and the mean square error is about 0.010?37 smaller, indicating that the proposed model has relatively high stability and prediction accuracy and can provide an important traffic guidance basis for large urban road networks.

Key words: Graph Convolutional Network(GCN), Gated Recurrent Unit(GRU), GraphSAGE algorithm, Graph Attention mechanism(GAT), urban road network

摘要:

针对交通流预测过程中城市道路路网的空间特征难以充分提取,导致预测结果精度不高的问题,提出图卷积网络(GCN)与门控循环单元(GRU)组合短时交通流预测模型。利用GCN对拓扑结构数据处理的优势,将城市道路路网空间排列结构转换为拓扑关系建模,通过解决拓扑关系问题有效提取出路网间的空间特征。采用GraphSAGE算法改进GCN模型,通过加和聚合算子和图注意力机制(GAT)聚合空间特征,将包含空间特征的输出作为GRU模型的输入提取时间特征。利用真实道路车流量数据进行模型验证,结果表明该模型相较于不具有GCN的模型预测准确率提升约8%,均方误差缩小约0.010?37,说明所提模型具有相对较高的稳定性及预测精度,可以为大型城市路网提供重要的交通诱导依据。

关键词: 图卷积网络, 门控循环单元, GraphSAGE算法, 图注意力机制, 城区路网