Multi-scale Spatio-temporal Attention Network for Traffic Flow Prediction

Traffic flow prediction has important implications for multiple fields, such as urban planning, traffic management and transportation. Accurate Traffic flow prediction helps improve transportation efficiency. At the same time, getting accurate traffic conditions can ensure traffic safety during special times. The key to accurate traffic flow prediction lies in the ability to accurately mine temporal and spatial dependencies, i.e., information on cycles and trends contained in historical time series and correlations between different locations in space. In recent years, a variety of algorithms have been used for traffic flow prediction, but all of them have their own limitations that lead to less accurate predictions in some cases. In this paper, we propose a multi-scale attention graph neural network model for traffic flow prediction, which captures multi-scale spatial dependencies through a multi-scale graph neural network. And a Pyramidal temporal network model is also proposed for mining temporal dependencies progressively from global to local. To validate our proposed method, we conduct extensive experiments on real-world traffic datasets to verify the effectiveness of our method.