Guidelines and optimization criteria of a machine learning-based methodology for mixture design in ORC systems

Abstract

orking fluid selection in organic Rankine cycles (ORC) is a critical issue in the system design for specific applications. In most cases, pure fluids are employed as working medium, and the choice among commercial products is fundamentally based on the heat source temperature level, on fluid cost and safety requirements. Mixtures of organic fluids present some interesting features. First of all, zeotropic blends can help reducing the irreversibility associated to the isothermal heat transfer of pure working fluids in the evaporator and in the condenser, with potential improvement of the overall conversion efficiency from heat sources with finite capacity. Moreover, the possibility of using fluids with relatively high global warming potential, blended with low-GWP fluids, allows reducing the equivalent carbon emissions (according to the European F-gas regulation), helping the full transition to cleaner compounds. However, there is a lack of organic mixtures specifically optimized for ORC systems, particularly because different applications require fluids with different properties, and the available solution may not be the optimal for the specific case. Due to the large number of molecules available to be combined, the use of numerical algorithms is mandatory to accomplish the mixture formulation task (type, number and amount of each component). In this work, an optimization tool based on a Bayesian statistical inference methodology is implemented and tested, to define the composition of mixtures of different groups of organic fluids (HFC, HFO, HC). The algorithm adopts a mixed exploration-exploitation searching strategy on a solution domain made by the different pure molecules commonly employed. The tool is programmed to fulfil the maximization of an objective function, built from a set of key physical properties and performance indexes of the mixture. The choice of the properties accounted in the optimization (with relative weight) is discussed, and the algorithm is applied to a case study to demonstrate its capability to define a low-GWP blend replacing HFC 134a, maintaining and possibly enhancing the performance.