Linear Parameter Varying Modeling of a High-Purity Distillation Column

More Info
expand_more

Abstract

One of the main reasons for the success story of Proportional-Integral-Derivative (PID) control in process industries has been the difficulty and complexity of modeling chemical, thermal, and physical phenomena in these systems. Linear Time-Invariant (LTI) identification has been found incapable to accurately capture the dynamics in these applications over the entire operating region, while nonlinear identification methods still often result in over-laborious and expensive process modeling tools with a too complex model for control synthesis. The concept of data-driven Linear Parameter-Varying (LPV) modeling offers an in-between-solution over LTI and nonlinear identification by describing the signal relations in a linear manner which vary with the operating point of the system. Furthermore, the LPV modeling and control framework is considered to have the potential to become the long-waited solution for the modeling and control problems of the process field. Through the case study of Propane-Propene splitter, which is a commonly used high-purity distillation column, this thesis explores the potential of LPV modeling and identification for process systems. Importantly, the PP-splitter represents a particularly challenging process system due to its Multiple-Input Multiple-Output (MIMO) nature and significant nonlinear behavior in the high-purity operating region. The results presented in this thesis shows that, although such a system could impose problems for local LPV identification approaches (identification w.r.t. constant operating conditions), the application of global LPV identification (identification w.r.t. varying operating conditions) has found to be able to identify the dynamics of the system even under severe noise conditions. Moreover, the global methodology presented in this thesis, which is called the LPV Least-Square Support Vector Machine (LS-SVM), belongs to an emerging trend of novel approaches which adopt machine learning theories in the system identification. This method is the current state-of-the-art among such learning approaches in the LPV identification framework.