Attention-based spatial-temporal synchronous graph convolution networks for traffic flow forecasting

Accurate traffic flow forecasting is crucial for urban traffic control, planning, and detection. Most existing spatial-temporal modeling methods overlook the hidden dynamic correlations between road network nodes and the time series nonstationarity while synchronously capturing complex long- and short-term spatial-temporal dependencies. To this end, this paper proposes an Attention-based Spatial-Temporal Synchronous Graph Convolutional Network (AST-SGCN) to capture complex spatial-temporal correlations over long and short terms. Specifically, we design a self-attention mechanism that utilizes spatial-temporal synchronous computation to efficiently mine dynamic spatial-temporal correlations with changes in traffic and enhance computational efficiency. Then, we construct a residual adaptive adjacency matrix, which includes historical data and node vectors, to stimulate the information transfer of spatial-temporal graph nodes and mine the hidden spatial-temporal dependencies through the graph convolution layer. Next, we establish a Fourier transform layer (FTL) to handle the nonstationary data. Finally, we develop a spatial-temporal hybrid stacking module for capturing complex long-term spatial-temporal correlations, within which two layers of graph convolution and one layer of self-attention are deployed. Extensive experimental results on three real-world traffic flow datasets demonstrate that our AST-SGCN model outperforms the comparable models.