UCLA

Data is widely recognized as a crucial player in the fourth industrial revolution, in which engineers and computers must harness data to enhance the efficiency of industrial processes and their associated control systems. Traditional industrial process control systems rely on linear data-driven models, with parameters fitted to experimental or simulated data. In specific control loops, such as those critical for profit optimization, they may employ first-principles models describing the underlying physico-chemical phenomena but with a few data-derived parameters. Nevertheless, modeling complex, nonlinear processes on a large scale remains an open challenge in process systems engineering. The quality of these models depends on various factors, including model parameter estimation, model uncertainty, the number of assumptions made during model development, model dimensionality, structure, and the computational demands for real-time model solutions [1,2]. This is especially pertinent as process models are integral to advanced model-based control systems, such as model predictive control (MPC) and economic MPC (EMPC). Designing MPC systems that utilize data-driven modeling techniques to account in real-time for large data sets is a new frontier that will impact the next generation of industrial control systems. While a significant body of research has been dedicated to the use of neural networks for nonlinear process modeling and control, in both the theoretical [3] and practical [4] domains, more computationally efficient models that can directly be used in MPC rather than their linearized counterparts, are still an growing area of research that can lead to the design of more robust and efficient control systems.

Motivated by the above considerations, this dissertation presents the use of a computationally efficient data-driven technique known as sparse identification in model predictive control for chemical processes described by nonlinear dynamic models. The motivation and organization of this dissertation are first presented. Then, the use of sparse identification to develop nonlinear dynamic process models to be used in model predictive controllers is presented, specifically addressing the challenges of two-time-scale systems, sensor noise, industrial nonlinearities, and process shifts. The MPC and economic MPC schemes that use sparse identified models are presented in detail with rigorous analysis provided on their closed-loop stability and recursive feasibility properties. Finally, the dissertation closes with an overview of the novelties introduced to overcome the aforementioned challenges and a detailed guide to developing nonlinear process models for complex chemical processes using sparse identification. Throughout the dissertation, the proposed methods are applied to numerical simulations of nonlinear chemical process examples and Aspen Plus simulations of large-scale chemical process networks to demonstrate their effectiveness.

[1] S. S. Ge and C.Wang. Adaptive neural control of uncertain MIMO nonlinear systems. IEEE Transactions on Neural Networks, 15:674–692, 2004.[2] H. W. Ge, Y. C. Liang, and M. Marchese. A modified particle swarm optimization-based dynamic recurrent neural network for identifying and controlling nonlinear systems. Computers & Structures, 85:1611–1622, 2007. [3] Z. Wu, A. Tran, D. Rincon, and P. D. Christofides. Machine learning-based predictive control of nonlinear processes. Part I: Theory. AIChE Journal, 65:e16729, 2019. [4] J. Luo, B. Çıtmacı, J. B. Jang, F. Abdullah, C. G. Morales-Guio, and P. D. Christofides. Machine learning-based predictive control using on-line model linearization: Application to an experimental electrochemical reactor. Chemical Engineering Research and Design, 197:721–737, 2023.