A Model and Data Dual-driven Approach to Traffic State Estimation on Freeways

Traffic state estimation can be briefly categorized into two types: model-driven and data-driven, while each suffers from either model deficiencies or data insufficiency. To mitigate these limitations, we contribute a model and data dual-driven approach, called physics-informed deep learning aided by model source term estimator (PIDL-MSTE), which can be applied to traffic state estimation of multi-ramp freeways. The framework introduces a neural network-based model source term estimator (MSTE), which implicitly learns the conservation law and roadway network topology to realize ramp flow imputation. In order to validate the effectiveness of the PIDL-MSTE, a comprehensive evaluation was conducted using traffic data from the inductive loop detectors of real-world expressway. The experimental results show that (i) the mean absolute percentage error for estimated speed of the PIDL-MSTE is only 8.13%, which is better than other methods; (ii) the PIDL-MSTE method exhibits low sensitivity to the number of available detectors.