Dynamic data rectification by recurrent neural networks vs. Traditional methods

Recurrent neural networks are used to demonstrate the dynamic data rectification of process measurements containing Gaussin noise. The performance of these networks is compared to the traditional extended Kalman filtering approach and to published results for model‐based nonlinear programming techniques for data reconciliation. The recurrent network architecture is shown to provide comparable, if not superior, results when compared to traditional methods. The networks used were trained using conventional nonlinear programming techniques in a batch fashion.