Application of gap metric for model selection in linear parameter varying model-based predictive control

Abstract

Selecting the minimum number of linear models required to characterize a nonlinear dynamic system in the local approach to linear parameter varying (LPV) model identification remains a challenge. This study proposes the use of the gap metric for such selection. Subsequently, the identified LPV model is proposed for use in the design of LPV-based model predictive control (MPC). The results of the validation experiment reveal that the identified gap-based local and global LPV models provide a good fit, which is superior to that of any individual linear model, as measured by the computed mean squared error values. Furthermore, the closed-loop results show that the performance of the LPV-MPC closely matches that of the full-blown nonlinear MPC and multi-MPC, while LPV-MPC clearly outperforms Linear MPC. Additionally, the proposed LPV-MPC is found to be robust to process model mismatches and measurement noise, and is much less computationally intensive than nonlinear MPC.