Spatio-temporal graph modeling is an important task in analyzing the spatio-temporal correlation of traffic flow prediction models, and although many existing methods optimize traffic flow modeling by constructing adaptive graphs, these methods use adaptive graphs in the training phase and do not effectively learn the traffic flow data used in the testing phase. To address the above problems, the Progressive Graph Convolutional Network Traffic Flow Prediction Model based on Spatio-Temporal Self-Attention (PGCN-STSA) is proposed. Specifically, the spatio-temporal self-attention mechanism layer is constructed to perceive the local context information and extract the nonlinear temporal features together with the dilated convolution module, which is favorable for long-term prediction. The PGCN-STSA model constructs an asymptotic adjacency matrix by learning the trend similarity between graph nodes and builds an asymptotic static graph convolution module based on it, which is combined with a graph convolution network to fully capture the dynamic spatial features of the traffic flow. In addition, multiple spatio-temporal layers are stacked to increase the extraction capability of spatio-temporal features of the model for prediction through the output layer. The experimental results show that, compared with the existing optimal baseline method ST-MetaNet, the MAE and RMSE of PGCN-STSA model on METR-LA data set are improved by 0.74%, 0.77%, 2.58% and 2.87% respectively in the 15-min and 30-min predictions.
Liu et al. (Wed,) studied this question.