Identifiability, observability and model-based control of nonlinear systems

  • Identifizierbarkeit, Beobachtbarkeit und modellbasierte Regelung nichtlinearer Systeme

Joy, Preet Joseph; Mitsos, Alexander (Thesis advisor); Mönnigmann, Martin (Thesis advisor)

Aachen : RWTH Aachen University (2022)
Dissertation / PhD Thesis

In: Aachener Verfahrenstechnik Series 23
Page(s)/Article-Nr.: 1 Online-Ressource : Illustrationen

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2022


Mechanistic process models are used in many applications to improve process performance. To adequately describe the process, model parameters are often estimated from experimental data. The concept of identifiability addresses uniqueness of the parameter estimates. Identifiability is a global concept and there are many methods and tools to test identifiability of linear models. Identifiability analysis for nonlinear models is significantly more complex and the existing methods are not tractable, except for small-scale models. Thus, for medium and large-scale nonlinear models, methods to test local identifiability are used. In this thesis, we investigate an optimization-based approach for global identifiability analysis of nonlinear models, wherein we formulate and solve a problem of semi-infinite programming (SIP). For online model-based applications like state estimation, in addition to identifiability, the concept of observability is of importance. Observability addresses uniqueness of the state estimates. As in the case of identifiability, there are well-established methods for observability analysis of linear models, but in the nonlinear case, the methods are tractable only for small-scale models. Observability is also a global concept and for medium and large-scale nonlinear models the existing approaches can only test for local observability. In this thesis, we present an optimization-based method to analyze global observability in nonlinear systems. The formulation of the SIP for observability is an extension of the SIP formulation for identifiability. In model-based control applications like NMPC and dynamic real-time optimization (DRTO) both identifiability and observability are important for successful implementation. Within the scope of this thesis, we discuss the implementation of NMPC for a continuous polymerization process and we show that the NMPC is able to successfully maintain the desired polymer composition and monomer conversion in the presence of a process disturbance. In this thesis, we also investigate the optimal operation of a semi-batch epoxidation process using offline dynamic optimization, NMPC and DRTO. With these tools, we achieve a reduction in process time of over 75%.