A Novel Learning Approach for Citywide Crowd Flow Prediction

With the accelerating pace of urban life, population movements are changing dynamically in both space and time. However, the ensuing traffic congestion and potentially dangerous problems caused by crowd flow and increasing uncertain traffic data cannot be ignored. Crowd flow prediction is of critical importance for reducing traffic congestion and eliminating public safety risks in smart city. In this paper, we propose a Spatial Temporal Long-and Short-Term Network (STLSN) model employing deep learning method to predict crowd flow in high precision. The proposed STLSN considers both short and long term temporal dependencies, and it employs local Convolution Neural Network (CNN) to capture spatial correlations between regions. Specifically, we employ the Long Short-Term Memory (LSTM) to capture short term dependency while using recurrentskip architecture which utilizes the periodic characteristic of flow data to capture the long term temporal dependencies. Moreover, the weighting Point of Interest (POI) is applied to differentiate the importance of function categories. Finally, we conduct experiments on practical dataset and the results demonstrate the effectiveness of our model is over the compared methods.