Modeling Dynamic Spatio-Temporal Correlations for Urban Traffic Flows Prediction

Abstract

Prediction of traffic crowd movement is one of the most important component in many applications ' domains ranging from urban management to transportation schedule. The key challenge of citywide crowd flows prediction is how to model spatial and dynamic temporal correlation. However, in recent years several studies have been done, but they lack the ability to effectively and simultaneously model spatial and temporal dependencies among traffic crowd flows. To address this issue, in this article a novel spatio-temporal deep hybrid neural network proposed termed STD-Net to forecast citywide crowd traffic flows. More specifically, STD-Net contains four major branches, i.e., closeness, period volume, weekly volume, and external branches, respectively. We design a residual neural network unit for each property to depict the spatio-temporal features of traffic flows. For various branches, STD-Net provides distinct weights and then combines the outputs of four branches together. Extensive experiments on two large-scale datasets from New York bike and Beijing taxi have demonstrated that STD-Net achieves competitive performances the existing state-of-the-art prediction baselines.

Keywords

• Deep learning
• urban crowd flows
• neural networks
• long short-term memory
• convolutional neural network