Data-driven evolutionary optimisation for the design parameters of a chemical process: a case study

Abstract

The optimisation of Chemical plant design and operation has proven to be challeng ing due to the complexity of real-world systems. The complexity translates into high com putational costs from the mathematical representations and simulation models for these systems. Research has illustrated the benefits of using surrogate models as substitutes for computationally expensive models. This research investigates two main concepts. The first one being the resource cost reduction when implementing surrogate assisted genetic algorithms for the optimisation of computationally expensive simulation models repre senting chemical systems. The second component focuses on determining the robustness of these algorithms towards stochastic and multi-objective systems. Two main algorithms were developed to optimise four different chemical plant systems. The Chemical Plant System - Basic (CPS-B) is a stochastic chemical process including buffer tanks, process ing units, and a tank with multiple feed streams. The Chemical Plant System - Parallel (CPS-P) and Chemical Plant System - Feedback (CPS-F) are more complex variants of the CPS-B introducing additional complexities in the form of parallel and feedback loop sys tems respectively. The Surrogate Assisted Genetic Algorithm (SA-GA) was used to opti mise these three systems. The SA-GA algorithm was adapted for multi-objective optimi sation. The new adapted algorithm called the Surrogate Assisted NSGA-II (SA-NSGA) algo rithm was tested on a popular literature case, the Pressure Swing Adsorption (PSA) system. The optimisation results for all the chemical systems illustrated that using Genetic Algo rithms as an optimisation framework for complex stochastic, single and multi-objective chemical plant systems results in significant computational benefits. Introducing Ma chine Learning Surrogate models as substitutes for computationally expensive simulation models into a Genetic Algorithm framework yielded significant computational efficiency improvements. The optimisation of CPS-B, CPS-P, CPS-F, and PSA achieved 1.8, 1.74, 1.95, and 5 times speedup of the total elapsed run time, despite the increased complexity in the systems. It is worth noting that the SA-GA and SA-NSGA algorithms implemented in this research yielded results confirming both their flexibility and robustness towards more complex stochastic, single and multi-objective systems.