Traffic Spatial-Temporal Transformer for Traffic Prediction

Abstract

Accurate traffic prediction can help administrators better plan and manage urban traffic, alleviating the traffic pressure. Local spatial-temporal dependency is the strongest and most direct dependency within traffic data. However, recent researches in traffic prediction using stacked or coupled fusion methods to combine temporal and spatial learning networks have not fully captured local spatial-temporal dependency in traffic data. This paper introduces a recurrent neural network structure that captures local spatial-temporal dependency by considering the spatial relationship of each time with its current, past, and future time simultaneously. Additionally, a period enhanced attention mechanism is introduced to capture long-term temporal dependency. Finally, the two modules are combined to construct a Traffic Spatial-Temporal Transformer for traffic prediction. Experimental results demonstrate that the proposed transformer outperforms baselines in terms of prediction accuracy.