A Joint Traffic Flow Estimation and Prediction Approach for Urban Networks

Classical methods of traffic flow prediction with missing data are generally implemented in two sequential stages: a) imputing the missing data by certain imputation methods such as kNN, PPCA based methods etc.; b) using parametric or non-parametric methods to predict the future traffic flow with the completed data. However, the errors generated in missing data imputation stage will be accumulated into prediction stage, and thus will negatively influence the prediction performance when missing rate becomes large. To solve this problem, this paper proposes a Joint Traffic Flow Estimation and Prediction (JT-FEP) approach, which considers the missing data as additional unknown network parameters during a deep learning model training process. By updating missing data and the other network parameters via backward propagation, the model training error can generally be evenly distributed across the missing data and future data, thus reducing the error propagation. We conduct extensive experiments for two missing patterns i.e. Completely Missing at Random (CMAR) and Not Missing at Random (NMAR) with various missing rates. The experimental results demonstrate the superiority of JTFEP over existing methods.