Spatial-temporal bottleneck attention transformer networks for traffic flow forecasting

Accurate traffic flow prediction helps managers in travel and decision making. Most of the existing models are based on graph neural networks (GNN), which solve the problem by capturing the spatial dependencies of fixed graph structures. However, this approach is limited due to the incompleteness of data with dynamic spatio-temporal dependencies. More often than not, with more prediction nodes with increasing time span, these types of models are not ideal in terms of efficiency in prediction. To overcome these limitations, this paper proposes a new model: spatio-temporal bottleneck attention Transformer network (STBAN Transformer) for spatio-temporal relationship modeling and long term traffic prediction. Transformer models sequences through the mechanism of self-attention, and by applying it to traffic flow prediction, it can capture the nodes and other nodes well The spatial correlation between nodes and other nodes can be well captured, which is very suitable for the extraction of spatial features of traffic network. In addition, in temporal correlation modeling, we also design an efficient temporal bottleneck attention module to obtain temporal attention in the global spatio-temporal state with low complexity. Experimental results on two public transportation datasets show that our method achieves state-of-the-art performance.