Predicting the performance of magnetocaloric systems using machine learning regressors

Since refrigeration, air-conditioning and heat pump systems account to 25–30% of all energy consumed in the world, there is a considerable potential to mitigate the Global Warming by increasing the efficiency of the related appliances. Magnetocaloric systems, i.e. refrigerators and heat pumps, are promising solutions due to their large theoretical Coefficient Of Performance (COP). However, there is still a long way to make such systems marketable. One barrier is the cost of the magnet and magnetocaloric materials, which can be overcome by decreasing the materials quantity, e.g. by optimizing the geometry with efficient dimensioning procedures. In this work, we have developed a machine learning method to predict the three most significant performance values of magnetocaloric heat pumps: temperature span, heating power and COP. We used 4 different regressors: ordinary least squares, ridge, lasso and K-Nearest Neighbors (KNN). By using a dataset generated by numerical calculations, we have arrived at minimum average relative errors of the temperature span, heating power and COP of 23%, 29% and 31%, respectively. While the lasso regressor is more appropriate when using small datasets, the ordinary least squares regressor shows the best performance when using more samples. The best order of polynomials range between 3, for the heating power, to 5, for the COP. The worse performance in predicting the three performance values occurs when using the KNN regressor. Furthermore, the application of regressors to the dataset is more adequate to evaluate the temperature span rather than energetic performance values.