Author ORCID Identifier

https://orcid.org/0009-0006-5308-9717

Semester

Spring

Date of Graduation

2024

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Chemical and Biomedical Engineering

Committee Chair

Fernando V. Lima

Committee Member

Fernando V. Lima

Committee Member

Heleno Bispo

Committee Member

Stephen E. Zitney

Committee Member

Yuhe Tian

Committee Member

David S. Mebane

Abstract

This dissertation aims to develop strategies for process systems engineering (PSE) mapping using models, emerging tools and algorithms motivated by process operability analysis research. Such strategies will be employed to ensure simultaneous design and control of large-scale industrial systems. The emerging tools and techniques in this research include supervised machine learning-based (ML-based) operability mapping, automatic differentiation (AD) for implicit mapping, and the development of a systematic mapping approach for control structure selection using operability analysis. Thus far, the developed operability algorithms either recur to nonlinear programming (NLP) solutions which are computationally expensive or to linearizing the underlying modeling task at the expense of losing accuracy. In addition, the combinatorics properties inherent to the control structure selection problem are yet to be addressed using operability analysis. Motivated by these existing gaps, the following contributions are provided in this research: (i) development of a systematic framework for operability analysis using Gaussian Processes (GP), a type of supervised machine learning technique to handle general nonlinear and large-scale process models. This approach focuses on maintaining the representation of inherent nonlinearities of the underlying process model aided by a GP surrogate instead of the computationally expensive, first-principles model. The use of GP regression opens the opportunity of employing real plant data to develop a surrogate that can be used for operability analysis when a first-principles model is not readily available. Lastly, due to GP intrinsic roots in Bayesian inference, the uncertainty of the operability sets obtained can also be quantified; (ii) establishment of an implicit mapping approach using the implicit function theorem and automatic differentiation. Recent innovations and advances in automatic differentiation allow the use of the implicit function theorem and a simple integration scheme to directly obtain the inverse or forward map of a given process. This circumvents the use of NLP solvers, reducing the computational effort while maintaining accuracy; and (iii) development of a framework that systematically uses operability analysis to deal with control structure selection in plantwide systems. This allows the systematic ranking of control structures based on their process operability characteristics. Thus, contributions related to the systematic operability analysis of large-scale, high-dimensional systems are made more tractable with the developed algorithms in this dissertation. The results show that the proposed approach for ML-based mapping has the potential to reduce computational time by up to four orders of magnitude when addressing large-scale systems, while keeping accuracy within 0.3% for the worst-case scenario, when compared to classic NLP mapping techniques. In addition, for the novel AD-based mapping, the accuracy is also guaranteed to be within around 0.02%, while the computational time is reduced from 19.2 times to 111% faster depending on the implementation of the compared nonlinear programming formulation. Lastly, the developed approach for control structure selection using operability analysis is capable of ranking posed control structures according to their operability characteristics using the operability index as a metric. The use of supervised machine learning, automatic differentiation and approaches for efficient evaluation of inherent combinatorics in large-scale process systems are pertinent contributions that are not well-addressed currently in process operability research. This directly impacts the application of process operability techniques in the industry since the algorithms and approaches developed in this dissertation potentially enable the use of operability research in large-scale applications. Lastly, the development of an open-source process operability package in Python, namely opyrability, also provides ease-of-assess of all developed operability algorithms to users in a single-bundle fashion, and in a freely available programming language.

Share

COinS