Learning-based Koopman modeling for efficient state estimation and control of nonlinear processes

Xunyuan Yin
Assistant Professor
Nanyang Technological University
Abstract: Industries are increasingly prioritizing heightened process operation safety, production consistency, efficiency, waste and emissions reduction, and profitability optimization. This new dynamic environment calls for smarter, more efficient, and more flexible integrated automation solutions that comprise efficient and easy-to-use monitoring, control, and beyond. Modern industrial processes exhibit highly nonlinear dynamic behaviors.
Nonlinear first-principles models have been extensively used as the basis of state estimation and advanced control systems development. However, this type of solution has the following limitations: First, the design and analysis of nonlinear estimation and control algorithms are much more challenging as compared to those based on linear models. Second, in the presence of constraints, the use of a nonlinear model for optimization-based estimation and control can incur a much heavier computational load as compared to the linear counterpart. More importantly, it can be expensive or hard to build a high-fidelity first-principles nonlinear model for complex industrial processes in practice.
Considering these challenges, we have made attempts to address optimization-based control and state estimation for nonlinear processes within an alternative framework of Koopman modeling, which aims to construct data-driven linear dynamic models to predict the dynamical behavior of nonlinear processes. This presentation covers the following aspects:
1) A concise overview of the Koopman operator concept, and its connection with nonlinear optimal estimation and control.
2) Koopman modeling for predictive control with manually selected observables.
3) Machine learning-based Koopman modeling for efficient estimation, model predictive control (MPC), and economic MPC of general nonlinear systems.
Additionally, we plan to present our recent findings on integrating machine learning with the data-enabled predictive control (DeePC) framework. Our goals are twofold: a) to facilitate economic operation within the DeePC framework, and b) to reduce or eliminate the need for optimization during online control implementation.
Bio: Xunyuan Yin received the Ph.D. degree in process control from the University of Alberta, Edmonton, AB, Canada, in August 2018. Between August 2018 and November 2021, he worked as a Postdoctoral Fellow at the University of Alberta. Currently, he is an Assistant Professor in the School of Chemistry, Chemical Engineering and Biotechnology at Nanyang Technological University (NTU), Singapore. His research interests include machine learning-based process modeling and control, distributed estimation and control, and process monitoring, and their applications to wastewater treatment, carbon capture processes, and a few other large-scale industrial systems and processes. He is an Associate Editor for Control Engineering Practice, and Digital Chemical Engineering. He is a member of the IEEE Control Systems Society Conference Editorial Board, and is a member of Journal of Process Control Paper Prize Selection Committee.