ARTIFICIAL NEURAL NETWORK CORRELATION FOR PREDICTION OF THERMAL CONDUTIVITY AND VISCOSITY OF R32 REFRIGERANT

Abstract

Thermodynamic parameters are often values ​​that are difficult to obtain. Knowing this, numerical methods are tools that become valuable. In this work, artificial neural networks (ANNs) were used, which have been shown to be quite efficient in solving problems. Experimental data of the refrigerant fluid R32, in liquid phase, available in the literature, were used to carry out the training of ANNs, aiming at the calculation of viscosity (μ) and thermal conductivity (λ), at different pressures and temperatures (which were the variables input). Training was performed with both ANNs with one intermediate layer and with two. Different optimization methods, combination of activation functions and number of neurons were evaluated. The best result, in the training, validation and testing stages, was obtained with an ANN of an intermediate layer and 35 neurons. The average percentage error obtained was 0.001% and 0.074% for μ and λ, respectively. Thus, it is concluded that ANNs are a very accurate tool for calculating these properties of the R32 fluid, based on its temperature and pressure.