一种用于交通预测的注意力时空图神经网络

城市化的发展使得交通预测在交通规划和城市管理等应用中发挥着重要作用。然而在交通预测任务中,捕获交通数据的高度非线性和复杂的时空依赖关系仍具有很大的挑战性。为了更好地捕获交通数据的时间依赖性和全局空间相关性以及同时满足长期和短期的预测任务,设计了一种用于交通预测的注意力时空图神经网络。首先通过引入注意力机制来调整邻近道路与非邻近道路的重要性,整合全局空间信息;然后再通过图卷积网络和带有扩展因果卷积的门控线性单元来共同捕获时空相关性。在两个真实数据集PeMSD7(M)和PEMS-BAY上的实验结果表明,该网络模型可以较大地提高短期和长期的交通预测精度。

Attention spatial-temporal graph neural network for traffic prediction

With the development of urbanization,traffic prediction plays an important role in the application of traffic planning and urban management.However,in the task of traffic prediction,it is still a great challenge to capture the highly nonlinear and complex spatio-temporal dependencies of traffic data.In order to effectively capture the time dynamics and global spatial correlation of traffic data and satisfy both long-term and short-term prediction tasks,an attention based spatial-temporal graph neural network for traffic prediction is designed.First,the attention mechanism is introduced to adjust the importance of adjacent roads and non-adjacent roads,which is beneficial to integrating global spatial information.Then,the spatial-temporal correlations are captured by graph convolutional networks and gated linear units with extended causal convolution.Experimental results on two real data sets PeMSD7(M) and PEMS-BAY show that the network model can improve the accuracy of both long and short-term traffic prediction.