DSTGCS: an intelligent dynamic spatial-temporal graph convolutional system for traffic flow prediction in ITS

Accurate traffic prediction is indispensable for relieving traffic congestion and people's daily trips. Nevertheless, accurate traffic flow prediction is still challenging due to the traffic network's complex and dynamic spatial and temporal dependencies. Most existing methods usually ignore the dynamicity of spatial dependencies or have limitations, as using the self-attention mechanism for capturing dynamic spatial dependencies is computation forbidden in large networks. In addition, there are both short- and long-range dynamic temporal dependencies, which are not well captured. To overcome these limitations, we propose an intelligent dynamic spatial and temporal graph convolutional system for traffic flow prediction. First, we propose a dynamic spatial block to capture the complex and dynamic spatial dependencies, which is computation-friendly. Next, we propose a dynamic temporal block to capture the complex and dynamic temporal dependencies, which well balances the short- and long-range dynamic temporal dependencies. We validate and analyze the performance of the proposed method through extensive experiments on two traffic datasets. Analysis of results demonstrates that our proposed model has better prediction performance than the state-of-art baselines. Compared with the best contrast methods, the proposed method improves by 2.28% and 8.01% in terms of the mean absolute error on PEMS04 and PEMS08 datasets.