基于图注意力机制的交通流预测模型

A Traffic Flow Prediction Model Based on Graph Attention Mechanism

针对现有交通流预测模型在预测精度上的不足,提出一种基于注意力机制的图模型。首先,利用多头注意力机制在交通图中编码高阶邻域结构,提取交通网络中的高阶空间特征。然后,嵌入长距离时间结构注意力机制提取长期性的历史周期信息。模型采用注意力机制替代传统的局部卷积核结构,可以有效提取长距离时空依赖关系。在METR-LA(洛杉矶路网)、PeMS-BAY(加州湾区路网)、PeMS-S(加州小型路网)三个真实的交通数据集上进行实验证明,模型在预测未来60 min的交通流精度上较传统深度学习方法,RMSE(均方根误差)平均降低3.1%、3.9%和1.8%,表明所提模型的长时间预测能力优势明显。

This paper proposes a graph model based on attention mechanism to address the accuracy issues of existing traffic flow prediction models.Firstly,the multi-head attention mechanism is used to encode high-order neighborhood structures in the traffic map and extract high-order spatial features in the traffic network.Then,a long-distance time structure attention mechanism is embedded to extract long-term historical cycle information.The model uses attention mechanism to replace traditional local convolutional kernel structure,which can effectively extract long-distance spatio-temporal dependencies.Experiments are carried out on three real traffic data sets,METR-LA(Road Network of Los Angeles),PeMS-BAY(Bay Area Road Network of California),and PeMS-S(Small Road Network of California).Experimental results show that the RMSE(Root Mean Square Error)of the traffic flow prediction accuracy of the proposed model in the next 60 minutes is 3.1%,3.9%,and 1.8%lower on average than that of traditional deep learning method,which indicates that the proposed model has obvious advantages in long-term prediction ability.