Adaptive recurrent neural network for uncertainties estimation in feedback control system

Abstract

In this paper, a recurrent neural network (RNN) is used to estimate uncertainties and implement feedback control for nonlinear dynamic systems. The neural network approximates the uncertainties related to unmodeled dynamics, parametric variations, and external disturbances. The RNN has a single hidden layer and uses the tracking error and the output as feedback to estimate the disturbance. The RNN weights are online adapted, and the adaptation laws are developed from the stability analysis of the controlled system with the RNN estimation. The used activation function, at the hidden layer, has an expression that simplifies the adaptation laws from the stability analysis. It is found that the adaptive RNN enhances the tracking performance of the feedback controller at the transient and steady state responses. The proposed RNN based feedback control is applied to a DC–DC converter for current regulation. Simulation and experimental results are provided to show its effectiveness. Compared to the feedforward neural network and the conventional feedback control, the RNN based feedback control provides good tracking performance.