Gaussian process based model predictive control to address uncertain milling circuit dynamics

Abstract

Model predictive control performance rests heavily on the accuracy of the available plant model. To address (possibly) time-variant model uncertainty, a nominal nonlinear state-space model is combined with an additive residual model that takes the form of a Gaussian process. With sufficient operational data the Gaussian process model is able to effectively describe the residual model error and reduce the overall prediction error for effective model predictive control. The efficacy of the method is illustrated using a milling circuit simulator.