Evaluation of machine learning approaches for estimating thermodynamic properties of new generation refrigerant R513A

Abstract

In this study, direct thermodynamic properties of R513A refrigerant such as temperature, pressure, enthalpy, and entropy were estimated using polynomial regression and GradientBoosting, XGBR, KNNRegressor, and Decision Tree as machine learning algorithms. The low-order polynomial regression results for the enthalpy and entropy of saturated vapor reached an R2 value of 0.9939 and below, while the 4th-degree polynomial regression reached an R2 value of 1.0 for the pressure of saturated vapor. Compared to other machine learning algorithms, the KNNRegressor reached an R2 value of 0.9947 and it gave better results than other methods for the saturated vapor. For enthalpy calculations of superheated vapor, R2 = 0.9984 result was reached in KNNRegressor, while in entropy calculations, XGBR and KNNRegressor gave 0.9960 and 0.9998 for R2 values, respectively. KNNRegressor gives better results in modeling with an R2 value of 0.9946 and above for the saturated liquid properties of R513A. According to the results obtained in this study, polynomial regression and machine learning methods achieve success in estimating properties of R513A refrigerant with an R2 value of 0.99 and above. The consequence of this study showed that machine learning methods can be used successfully in complicated processes such as predicting the thermodynamic properties of refrigerants.