Dual-Stage Attention Based Spatio-Temporal Sequence Learning for Multi-Step Traffic Prediction

Abstract

Traffic prediction has great significance including but not limited to mitigating traffic congestion, reducing traffic accidents, and reducing waiting time. At the same time, traffic prediction, especially multi-step prediction, faces many difficulties including temporal correlations and spatial correlations. We propose a dual-stage attention based spatio-temporal sequence learning for multi-step traffic prediction which can not only express temporal correlation and spatial correlation, but also can adaptively learn the contribution weights of different related roads and historical moments. More specifically, for spatial dependencies, we first generate the input vector for each historical moment considering the information of relevant road segments by the method of spatial region of support and further add the first-stage attention termed spatial attention to automatically determine the weight of each relevant road segment for each historical moment. For temporal dependencies, we use LSTM based encoder-decoder networks to fully learn the temporal characteristic and make multi-step prediction considering temporal correlation between multi steps. We further add the second-stage attention termed temporal attention in the decoder part to automatically learn the contribution of different historical moments to each prediction moment. In addition, we consider external factors including weather and holidays and characterize their impacts using fully connected networks. Finally, the effectiveness of the proposed method is evaluated using traffic data in Hangzhou, China.