Perspectives of data-driven LPV modeling of high-purity distillation columns

This paper investigates data-driven, Linear-Parameter-Varying (LPV) modeling of a high-purity distillation column. Two LPV modeling approaches are studied: a local approach, corresponding to the interpolation of Linear Time-Invariant (LTI) models identified at steady-state purity levels, and a global Least-Square Support Vector Machine (LSSVM) approach which offers non-parametric estimation of the system w.r.t. data with varying operating conditions. In an extensive simulation study, it is observed that the global LS-SVM approach outperforms the local methodology in capturing the dynamics of the high-purity distillation column under study. The simulation results suggest that the global LS-SVM approach provides a reliable modeling tool under realistic noise conditions.