Attention based convolutional networks for traffic flow prediction

Real-time and accurate prediction of traffic flow plays an important role in intelligent transportation systems. However, short-term traffic flow forecasting is extremely challenging due to the highly nonlinear nature of the traffic system and the dynamic spatial and temporal correlation. Although various methods, including deep learning based ones, have been proposed, most of them still suffer from problems such as spatial nonstationarity and thus cannot achieve good prediction performance. Inspired by the recent superior performance of attention mechanism, we introduce it into the model for traffic flow prediction with regular grided input. To be specific, we propose a novel deep learning framework, Spatial-Temporal Attention Based Convolutional Networks (STAtt-Net), for accurate forecasting of citywide traffic flow. First, we model the traffic data as a two-dimensional matrix with two channels. Each cell in the matrix represents the traffic in the corresponding region. Taking into account the temporal correlation and dependence of traffic system, the periodic patterns contained in traffic data are modeled by three major components for weekly trend, daily periodicity, and hourly closeness respectively. Then, STAtt-Net employs a STBlock as the basis unit to learn temporal dependence and spatial dependence of traffic flow, taking advantage of attention mechanism. We conduct extensive experiments to evaluate the performance of our model on three real-world datasets (TaxiBJ, BikeNYC, TaxiSZ), with the results revealing better prediction accuracy and efficiency of the proposed model against existing ones.