Increasing the prediction efficacy of the thermodynamic properties of R515B refrigerant with machine learning algorithms using SMOGN data augmentation method

Melike Siseci Cesmeli,
Increasing the prediction efficacy of the thermodynamic properties of R515B refrigerant with machine learning algorithms using SMOGN data augmentation method,
International Journal of Refrigeration,
Volume 179,
2025,
Pages 44-59,
ISSN 0140-7007,
https://doi.org/10.1016/j.ijrefrig.2025.07.026.
(https://www.sciencedirect.com/science/article/pii/S014070072500297X)
Abstract: Thermodynamic properties of refrigerants are very important parameters affecting the performance of refrigeration and air conditioning systems. In this study, various machine learning algorithms are used to predict the enthalpy (h), entropy (s), and specific volume (vv) thermodynamic properties of low Global Warming Potential R515B refrigerant. Due to the limited data, data augmentation was first performed using the Synthetic Minority Over-sampling Technique for Regression with Gaussian Noise (SMOGN) method, and then machine learning algorithms were used. Machine learning predictions are compared with and without augmented data including intermediate values. According to the results obtained, machine learning methods with data augmented with the SMOGN method successfully predicted the thermodynamic properties of R515B refrigerant compared to classical data. In the predictions of h, s, and vv values of superheated vapor, the R² value of the Categorical Boosting model applied to SMOGN was 0.9990 for h, 0.9983 for s, and 0.9999 for vv. In the analysis for saturated vapor states, the R² value of the K-Nearest Neighbors model with SMOGN applied was 0.9997 for hsv, 0.9956 for ssv, and 0.9981 for vvsv. According to the analysis results, when the results for saturated liquid states are analyzed, the most successful results are found when SMOGN data augmentation is applied. In the analysis performed by applying SMOGN, the Random Forest (RF) method generally showed successful results for the saturated liquid. The R² value of the RF model was obtained as 0.9997 for hliq, 0.9997 for sliq, and 0.9996 for vvliq.
Keywords: Data augmentation; Machine learning; New generation refrigerant; R515B; SMOGN