Correntropy based Elman neural network for dynamic data reconciliation with gross errors

Abstract

BackgroundMeasurement information in chemical processes is inevitably corrupted. Dynamic data reconciliation is an effective method to improve the quality of noisy measurement data. However, except for random errors, possibly gross errors are usually present in the measurement data, which may result in the deterioration and even failure of process control and optimization. MethodsConsidering the problem that dynamic mathematical models of some complex and unknown processes are difficult or impossible to be obtained, this paper combines the neural network with the correntropy estimator to design a robust dynamic data reconciliation scheme for unknown dynamic systems corrupted by both random and gross errors. The correntropy estimator is taken as a new objective function to form the robust Elman neural network (ENN). Furthermore, the processes of weight learning and parameter optimization are designed and described in detail based on the correntropy estimator. FindingsThe performance of the proposed robust ENN is demonstrated through its application on the free radical polymerization of styrene. With mixed gross errors, the mean squared error decreases more than 14-fold and falls below one ten-thousandth. Comparisons with other approaches show that the proposed method can effectively suppress the influence of gross errors and greatly improve the dynamic behavior of systems.