Adaptive Generalized Predictive Control Based on Just-in-Time Learning in Latent Space

Abstract

An adaptive generalized predictive control approach based on just-in-time learning(JITL) in latent space is proposed to deal with the problems associating with multivariate, nonlinearity and time-varying characteristics in industrial process systems. To begin with, the latent variable space is constructed by the partial least squares algorithm, thus the complicated multivariable control problem can be decomposed into univariate ones, subsequently the local model of each SISO subsystem can be established online by JITL at every sampling instant in latent space, where the generalized predictive control is implemented to these subsystems. To improve the real-time performance of modeling, the similarity measure will be utilized to determine whether or not to update the current local model at each sampling instant. The proposed approach not only can obtain the satisfactory control results for nonlinear and multivariate system, but also can solve the unstable problem caused by model mismatch. The proposed adaptive predictive control approach is applied to a pH neutralization process. Simulation studies are presented to verify the advantage of the proposed approach.