动态交通流量预测的时空注意力图卷积网络

Attention-based spatio-temporal graph convolutional network for dynamic traffic flow prediction

针对现有交通流预测方法大多忽略时空耦合相关性、时空变化性以及外部特征对预测结果准确性的影响,提出一种动态交通流量预测的时空注意力图卷积网络(attention-based spatio-temporal graph convolutional network,ATST-GCN)模型。提出基于注意力的双向门控循环单元结构,从动态空间序列中提取时间相关性;构建带残差链接的多层图注意网络(graph attention network,GAT)卷积模块,深入挖掘动态空间相关性;融合时变特征与时常特征,充分利用外部静动态特征的共同作用。采用PeMS数据集对交通流量预测的准确度进行验证,试验结果表明:本研究方法能够有效提高交通流量预测精度,优于现有的多数先进方法。在PeMS08和PeMS03数据集上,本研究方法相对STSGCN模型分别提高13.44%和10.96%,相对T-GCN模型分别提高21.41%和21.32%,相对STGCN模型分别提高8.04%和6.55%,相对DMSTGCN模型分别提高3.23%和2.80%,相对Trendformer模型分别提高2.29%和2.00%。

Most existing methods ignored the impact of spatio-temporal coupling correlation,spatio-temporal variability,and external features on the accuracy of prediction results.In response to the above problems,this paper proposed a spatio-temporal attention graph convolution network model(attention-based spatio-temporal graph convolutional network,ATST-GCN)for dynamic traffic flow prediction.An attention-based bidirectional GRU structure was proposed to extract temporal correlation from dynamic spatial sequences.A multi-layer GAT(graph attention network,GAT)convolution module with residual connection was constructed to deeply extract the dynamic spatial correlation.Time-varying features and constant features were integrated to make full use of the joint effect of external static and dynamic features.The PeMS dataset was used for verification of the accuracy of traffic flow prediction using PeMs dataset.The experiment results showed that the method proposed in this paper could effectively improve the accuracy of traffic flow prediction and was better than most existing advanced methods.On the PeMS08 and PeMS03 datasets,the method of this sdudy improved 13.44%and 10.96%relative to the STSGCN model,21.41%and 21.32%relative to the T-GCN model,8.04%and 6.55%relative to the STGCN model,3.23%and 2.80%relative to the DMSTGCN model,2.29%and 2.00%respectively relative to the Trendformer model.