Spatial-Temporal Multi-Head Attention Networks for Traffic Flow Forecasting

Abstract

Traffic flow forecasting plays an important role in the intelligent traffic system, which is the basis for traffic control and traffic management. However, due to the complex spatial-temporal dependence, traffic flow forecasting has always been a difficulty in the field of intelligent traffic. In order to select a suitable spatial-temporal forecasting method and solve the problem that recurrent neural architecture is not conducive to parallel computing, we construct a spatial-temporal forecasting model by using multi-head attention models. Use graph attention networks with multi-head attention mechanism to capture spatial features, and use the scaled dot product attention with positional encoding like Transformer to capture temporal features. Experimental results on two real-world datasets demonstrate that the forecasting error of our method is lower than baseline methods.