Generative Adversarial Nets Model Based on Spatio-Temporal Graph Attention Network for Multi-Step Prediction

The precise forecasting of traffic data holds significant importance for numerous practical applications in the real world. The challenge lies in the nonlinear spatio-temporal dependencies of the data, as well as rapid error accumulations for multi-step predictions. Therefore, we propose a Generative Adversarial Nets model based on a Spatio-Temporal Graph Attention Network (STGAT-GAN). Utilizing spatio-temporal graph attention proves to be an effective method for capturing the spatio-temporal dependencies, while GANs can learn a rich distribution over spatio-temporal data implicitly to produce realistic data adversarially. Experiments are carried out on two public traffic datasets. Comparative experimental analysis reveals that the proposed STGAT-GAN has better performance.