Data-driven structural synthesis of supercritical CO2 power cycles

Abstract

Read online

With new materials, objectives or constraints, it becomes increasingly difficult to develop optimal processes using conventional heuristics-based or superstructure-based methods. Hence, data-driven alternatives have emerged recently, to increase creativity and accelerate the development of innovative technologies without requiring extensive industrial feedback. However, beyond these proof-of-concepts and the promise of automation they hold, a deeper understanding of the behaviour and use of these advanced algorithms by the process engineer is still needed. In this paper, we provide the first data-driven solution for designing supercritical CO2 power cycle for waste heat recovery, a challenging industrial use case with lack of consensus on the optimal layout from the field literature. We then examine the issue of artificial intelligence acceptance by the process engineer, and formulate a set of basic requirements to foster user acceptance - robustness, control, understanding of the results, small time-to-solution. The numerical experiments confirm the robustness of the method, able to produce optimal designs performing as well as a set of selected expert layouts, yet only from the specification of the unit operations (turbomachinery and heat exchangers). We provide tools to exploit the vast amount of generated data, with pattern mining techniques to extract heuristic rules, thereby explaining the decision-making process. As a result, this paper shows how the process engineer can interact with the data-driven design approaches, by refocusing on the areas of domain expertise, namely, definition and analysis of the physical problem.

Keywords

• process design
• deep learning
• structural optimization
• pattern mining
• combined cycle
• supercritical CO2