Novel Approaches in Robust Multistage Nonlinear Model Predictive Control

Abstract

A powerful tool that has been widely used for control and optimization in the chemical process industry is model predictive control (MPC), thanks to its capacity for handling multivariable constrained control problems. However, process plants are operated under a wide variety of operating conditions, product specifications and safety limits, often following highly complex dynamics. A key challenge in process optimization is that a large majority of modern real-world processes lack perfect system information and have to contend with significant uncertainties. To ensure optimal performance under uncertainty, a class of methods classified under robust MPC has received widespread attention in recent years.

In this thesis, I investigate and extend one such prominent method, known as the robust multistage nonlinear MPC, where the uncertainty is described in form of a scenario tree. I propose novel approaches and algorithms to this framework that focus on two important aspects: 1) selecting scenarios in the scenario tree that better describe the uncertainty, and 2) improving its computational efficiency for solving large problems. To this end, the first part of the thesis shows a novel data-driven scenario selection strategy that leverages the interdependencies within the uncertainty data for given applications. The second part of the thesis demonstrates, through the use of nonlinear optimization theory and sensitivity analysis, a framework that drastically shrinks the size of the scenario tree, offers robustness in terms of performance and stability, and is computationally fast.