Traffic state variables estimating and predicting with neural network via extended Kalman filter algorithm with estimated parameters as offline

Developing mathematical models and estimating their parameters are fundamental issues for studying dynamic behaviors of traffic systems. METANET model is one of the most applicable models in traffic modeling in which the parameters have plenty of effects on the model behavior. In this paper, the effects of the model parameters on the model behavior and the estimation quality of the system states in the undetermined parameters are described. The extended Kalman filtering (EKF) algorithm instead of the error back-propagation (BP) algorithm is used to train artificial neural networks (ANNs) for dynamical traffic networks modeling. The basic idea is to prevent over fitting discrepancy occurrence caused by outliers in the training samples by the EKF. Numerical simulations show that the EKF algorithm is greater to the BP algorithm.