Hybrid mechanistic data-driven modeling for the deterministic global optimization of organic rankine cycles

  • Hybride mechanistisch-datengetriebene Modellierung für die deterministische globale Optimierung von Organic Rankine Cycles

Huster, Wolfgang Raphael; Mitsos, Alexander (Thesis advisor); Karellas, Sotirios (Thesis advisor)

Aachen (2020)
Book, Dissertation / PhD Thesis

In: Aachener Verfahrenstechnik series - AVT.SVT - Process systems engineering 10 (2020)
Page(s)/Article-Nr.: 1 Online-Ressource (XX, 200 Seiten) : Illustrationen, Diagramme

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2020


An approach for the deterministic global optimization of organic Rankine cycles (ORC) with accurate thermodynamic models embedded is presented. Two different fields of application are considered: waste heat recovery for diesel trucks and geothermal power generation. Design decisions have a tremendous impact on the economics of energy processes throughout their lifecycle. Thus, identifying a globally optimal design is highly desirable, but computationally expensive. In preceding studies, it was demonstrated that reduced space optimization is advantageous for deterministic global process optimization, as it can drastically reduce CPU times compared to a full-space formulation. Still, a relevant challenge for global flowsheet optimization is the consideration of accurate thermodynamic models. First, a validated dynamic model for an ORC in a diesel truck is presented. Based on this model, an approach for quasi-stationary process optimization is developed, in which the thermodynamic equation of states (EoS) of the working fluid are learned using artificial neural networks (ANN). Compared to accurate multiparameter EoS, ANNs have desirable properties for deterministic global optimization, namely tight relaxations and explicit evaluations. The resulting hybrid mechanistic data-driven process model can be solved efficiently to guaranteed global optimality. The method is extended to the automated data generation using thermodynamic libraries, thus it is possible to identify optimal working fluids from a wide range of candidates. For the geothermal application, it is first demonstrated how simple thermodynamic models result in low computational demand, but lead to false design decisions. The surrogate models are extended to supercritical properties and a working fluid selection for a transcritical ORC is performed. Furthermore, ANNs are used to learn the properties of zeotropic binary working fluids, allowing to find the globally optimal mixture composition for geothermal power cycles. Finally, additional structural options in the flowsheet are considered via the optimization of an ORC superstructure.