A Deep Learning Framework Based On Spatio-Temporal Attention Mechanism For Traffic Prediction

Traffic prediction is a significant unit of intelligent transportation systems in smart cities. For instance, a precise vehicle prediction model can not only help the city allocate resources in advance to meet travel needs but also reduce traffic congestion and energy waste caused by roadside vehicles. For traffic prediction tasks, how to model complex spatial dependencies and temporal dynamics is still an important issue. Although these two factors are considered in the modeling, the currently existing work mainly captures the spatial dependence by extracting the spatial features of the region, which ignores the mutual influence between the regions. Besides, they usually assume that temporal dynamics are strictly periodic. However, the mutual influence between regions has an impact on the prediction, and the temporal dynamics can be affected by some short-term mutation information (e.g., weather conditions), which will cause temporal dynamics to be perturbed from one period to another. To solve these two problems, this paper proposes a novel Spatio-Temporal Attention Mechanism-Based Dynamic Network (STADN), where via combining spatial attention mechanism and flow gating mechanism to learn dynamic spatial dependencies, via combining periodic shifting attention mechanism and a new transformation-gated LSTM (NTG-LSTM) to solve the long-term periodic temporal shifting caused by short-term mutation information. We prove the effectiveness of this method through the experimental results on real-world traffic datasets.