Spatiotemporal multi-graph convolutional networks with synthetic data for traffic volume forecasting

Reliable traffic volume forecasting remains a significant challenge due to the unstable variation patterns of traffic volume and the irregular spatial structure of traffic network. Most of previous works have neglected the implicit variation patterns of traffic volume, failing to model its spatiotemporal dependencies comprehensively. Although graph convolutional networks have been widely applied to traffic volume forecasting, most are constructed using simple graphs with limited representations, which may restrict their ability to extract deep spatial features. Therefore, a novel deep learning model is proposed in this study to address these deficiencies for improving the forecasting accuracy of traffic volume. First, a new spatiotemporal dependencies generative adversarial network is proposed to generate synthetic data (i.e., synthetic traffic volume) by using traffic speed and occupancy to construct latent variables (i.e., the input of generator in generative adversarial network), which is beneficial for capturing the implicit variation patterns of traffic volume. Then, a new multi-graph convolutional network is proposed to explore and model the heterogeneous correlations among sensors from multiple perspectives, so as to facilitate the effective extraction of deep non-Euclidean spatial features. Finally, experiments on two real-world datasets demonstrate that the proposed model makes accurate traffic volume forecasting and outperforms other state-of-the-art methods.