Nonlinear Dynamic Data Reconciliation Based on Recurrent Neural Networks and its Application

An adaptive gradient descent algorithm for training simplified internally recurrent networks (SIRN) is developed and a new method of reconciling nonlinear dynamic data based on SIRN is proposed. It can reconcile measurements of nonlinear dynamic process and dispenses with the need for accurate process model and prior information about statistical characteristics of noises involved. Simulation research and its application in a continually stirred tank reactor has demonstrated its fairly good performance.