一种用于交通预测的自适应时空图神经网络

An adaptive spatio-temporal graph neural network for traffic prediction

针对以往复杂的神经网络构架在空间维度上所采用的预定义图结构未包含完整交通数据空间信息,且在时间维度上不能很好地捕获交通数据长期依赖关系的问题,提出一种新的时空图神经网络。通过自适应图卷积网络(AGCN)自动捕获节点的特定状态以及自动推断不同节点之间的相互依赖关系,提取更完整的交通数据空间特征,再通过时空长短期记忆网络(ST-LSTM)中的时间记忆模块来提取交通数据的时间特征,捕获短中长期的时间依赖关系。在PeMSD4和PeMSD8数据集上进行了验证,实验结果表明,所提网络相比基线模型能够更好地提升交通预测性能。

For the problem that the neural network can't capture the long-term traffic information in the spatial dimension,the new neural network structure proposed in the past can't capture the complex traffic data in the spatial dimension.Through adaptive graph convolutional network,the specific state of nodes is automatically captured and the interdependence between different nodes is automatically inferred to extract the complete spatial features of traffic data.Then,the time characteristics of traffic data are captured by the time memory module in the spatio-temporal short-term memory network,and the short,medium and long-term time dependence is simulated.