Spatial-Temporal Dynamic Graph Convolutional Network for Traffic Flow Forecasting

In recent years, traffic problems have become more and more prominent. In order to solve the traffic congestion problem, this paper proposes a spatial-temporal dynamic graph convolution network (ST-DGCN) that can simultaneously consider the weather and the traffic accidents and can dynamically adjust the prediction time to predict the traffic flow according to the needs: The constructed distance matrix, node correlation matrix, accident matrix, and weather matrix are fused into a dynamic matrix. The spatial-temporal feature extraction module is divided into an attention mechanism module and a convolution module. The graph convolution based on Chebyshev polynomial approximation in the convolution module is used to capture spatial information fragments. The multiscale gated Tanh unit (M-GTU) module is used to capture temporal information fragments. The experimental results based on the four datasets show that the proposed ST-DGCN performs than several other baseline models better.