Attention based spatial-temporal graph convolutional networks for traffic flow forecasting

Forecasting the traffic flows is a critical issue for res and practitioners in the feld of transportation. Ho is very challenging since the traffic flows usually sh nonlinearities and complex patterns. Most existin flow prediction methods, lacking abilities of modelin namic spatial-temporal correlations of Traffic data, t not yield satisfactory prediction results. In this p propose a novel attention based spatial-temporal gr volutional network (ASTGCN) model to solve tra forecasting problem. ASTGCN mainly consists of dependent components to respectively model three ral properties of Traffic flows, i.e., recent, daily-peri weekly-periodic dependencies. More specifically, ea ponent contains two major parts: 1) the spatial-tem tention mechanism to effectively capture the dynami temporal correlations in Traffic data; 2) the spatial-t convolution which simultaneously employs graph tions to capture the spatial patterns and common convolutions to describe the temporal features. The o the three components are weighted fused to genera nal prediction results. Experiments on two real-world datasets from the Caltrans Performance Measurement System (PeMS) demonstrate that the proposed ASTGCN model outperforms the state-of-the-art baselines.