Abstract: Accurate traffic prediction is pivotal for intelligent transportation systems (ITS). However, due to complex time and space dependencies, the existing traffic flow prediction methods have failed to effectively capture the spatiotemporal characteristics of the road network and also ignored the strong dynamic nature of the correlation of road network traffic data in both spatial and temporal dimensions. In order to furtherly improve the prediction accuracy, a spatiotemporal graph neural network model based on multi-head attention was proposed. Firstly, the proposed model constructed an adaptive graph structure learning component that could effectively capture the dynamic spatiotemporal correlation of the graph structure. Secondly, the proposed model respectively designed a temporal multi-head attention module and a spatial multi-head attention module, which were based on the attention mechanism. The designed spatiotemporal multi-head attention module could effectively extract the spatiotemporal features of the road network. Finally, the stacked spatiotemporal convolutional layers were used to predict the future traffic conditions. The results, derived from open-source datasets, demonstrate that the proposed model performs well in spatiotemporal feature extraction and long-term prediction, and achieves more accurate prediction results than baseline methods.

Key words: traffic engineering; traffic prediction; intelligent transportation system; spatiotemporal multi-head attention; graph neural network; adaptive graph structure

摘要： 准确的交通预测对于智能交通系统(ITS)至关重要。然而，由于复杂的时间和空间依赖关系，现有的交通流预测方法未能有效捕获路网的时空特征，并且忽略了路网交通数据的相关性在空间维度和时间维度上表现出的较强动态性。为了进一步提高预测精度，提出了一种基于多头注意力的时空图神经网络模型。首先，该模型构造了一个自适应图结构学习组件，该自适应图结构学习组件可以有效地捕获图结构的动态时空相关性。其次，该模型基于注意力机制分别设计了时间多头注意力模块和空间多头注意力模块，所设计的时空多头注意力模块可以有效地对路网的时空特征进行提取。最后，利用堆叠的时空卷积层对未来的交通状况进行预测。在开源数据集上的实验结果表明：该模型在时空特征提取以及长期预测方面表现优异，并且比基线方法取得了更精确的预测结果。

关键词: 交通工程；交通预测;智能交通系统;时空多头注意力;图神经网络; 自适应图结构