DeepBSTN: A Deep Bidirection Network Model for Urban Traffic Prediction

For modern intelligent transportation systems, traffic prediction is the foundation of the proactive traffic management solutions. Beyond its necessity, how to find an available prediction method to achieve accurate traffic prediction is still full of challenges and problems. The complexed relationships, increasing data volume and other difficulties behind the traffic data have promoted a number of researches focusing on feature extraction and prediction modeling methods. Despite the successful achievements, most of the existing research is failed to provide a deeper view on mining the spatio-temporal relations among the traffic data due to being shallow in considering the backward temporal dependence. Motivated by such problems, we propose a DeepBSTN based on LSTM model. In this paper, DeepBSTN aims to use bidirectional LSTM to capture both of the forward and backward dependence among historical traffic data. Furthermore, the stacked deep bidirectional LSTM network is used in our model to obtain high-level representation. Finally, we achieve a deeper view of non-linear time dependence in our model. Based on two real large-scale public datasets from New York City, the abundant experimental results show that DeepBSTN has superior performance.