An Efficient Traffic Prediction Model Using Deep Spatial-Temporal Network

Recently years, traffic prediction has become an important and challenging problem in smart urban traffic computing, which can be used for government for road planning, detecting bottle-neck congestions roads, pollution emissions estimating and so on. However, former data mining algorithms mainly address the problem by using the traditional mathematical or statistical theories, and they were impossible to model the spatial and temporal relationship simultaneously. To address these issues, we propose an end-to-end neural network named C-LSTM to predict the traffic congestion at next time interval. More specifically, the C-LSTM is based on CNN and LSTM to collectively capture the spatial-temporal dependencies on the road network. Inspired by the procedure of handling the image by CNN, the city-wide traffic maps are first converted into a series of static images like the video frame and then are fed into a deep learning architecture, in which CNN extracts the spatial characteristics, and LSTM extracts the temporal characteristics. In addition, we also consider some external factors to further improve the prediction accuracy. Extensive experiments on reality Beijing transportation datasets demonstrate the superiority of our method.