基于多时空图卷积网络的交通流预测

交通流预测在交通管理和城市规划的应用中具有重要意义,然而现有的预测方法无法充分挖掘其潜在的复杂时空相关性,为进一步挖掘路网道路网络数据的时空特性以提高预测精度,提出一种多时空图卷积网络(multi-spatial-temporal graph convolutional network,MST-GCN)模型。首先,利用切比雪夫图卷积(ChebNet)结合门控循环单元(GRU)构建时空组件以深度挖掘节点的时空相关性;其次,分别提取周相关、日相关、邻近时间的序列数据,输入三个时空组件以深度挖掘不同时间窗口间的时间相关性;最后,将时空组件与编码器—解码器网络结构(encoder-decoder)融合组建MST-GCN模型。利用加利福尼亚州交通局(Caltrans)性能评估系统中高速公路数据集PEMS04和PEMS08进行实验,结果表明新模型的性能明显优于门控循环单元模型和最近提出的扩散卷积循环神经网络(DCRNN)、时间图卷积网络(T-GCN)、基于注意力机制的时空图卷积神经网络(ASTGCN)和时空同步图卷积网络(STSGCN)模型。

Traffic flow prediction based on multi-spatial-temporal graph convolutional network

Traffic flow forecasting is of great significance in the application of traffic management and urban planning.However,the existing forecasting methods cannot fully exploit the potential complex spatio-temporal correlations.In order to further explore the temporal and spatial characteristics of road network data to improve the prediction accuracy,this paper proposed an multi-spatial-temporal graph convolutional network(MST-GCN)model.Firstly,by using Chebyshev graph convolution(ChebNet)combined with gated recurrent unit(GRU)to construct spatio-temporal components to deeply mine the spatio-temporal correlation of nodes.Secondly,it extracted weekly,daily,and recent time sequence data separately,and entered three spatio-temporal components to deeply explore the time correlation between different time windows.Finally,it combined the spatio-temporal component and the encoder-decoder network structure to form the MST-GCN model.Experiments were conducted using the highway datasets PEMS04 and PEMS08 in the California Department of Transportation(Caltrans)performance evaluation system.The results show that the new model has significantly better performance than the gated recurrent unit model and the recently proposed diffusion convolutional recurrent neural network(DCRNN),temporal graph convolutional network(T-GCN),attention based spatial-temporal graph convolutional networks(ASTGCN)and spatial-temporal synchronous graph convolutional network(STSGCN)models.