Optimal mixture design for organic Rankine cycle using machine learning algorithm

Abstract

This study presents a new design tool for working fluid mixtures in organic Rankine cycles. The proposed tool comprises a blend model for the thermophysical properties of the formulated mixtures, an ORC model to predict the performance of the mixtures in a specific application, and an optimizer based on the Bayesian inference method to identify the optimal mixtures compositions to be assessed. The tool is programmed to optimize an objective function based on predefined optimization targets. Importantly, the targets and their respective weights within the objective function can be adjusted to meet the specific requirements of the application under analysis, making this approach adaptable to diverse research and industrial objectives. The algorithm is applied to a case study to demonstrate its ability to define a low-GWP blend that can replace HFC-134a in a micro-scale ORC with recuperator, while maintaining and potentially enhancing performance. The optimization targets specified for the case study are the net power output, the net efficiency, the GWP and the blend size. Power and efficiency are computed through a validated model of the low-temperature ORC system used as benchmark case. The results showed that the procedure was able to formulate several blends that comply with the targets of the assigned task. Amongst the high-scoring mixtures, the most used pure fluids are R32, R152a, R1234yf, and R1234ze(E). The presence of HCs is limited to fewer mixtures, playing the main role of GWP-limiter. A method to estimate the flammability classification of the blends has been also applied, obtaining that most of them belong to the ASHRAE class 2l, except when an HC is present, in which case the fluid is may result in class 3.