Eventos Anais de eventos
ENCIT 2022
19th Brazilian Congress of Thermal Sciences and Engineering
Evaluation of Machine Learning models to predict the thermal performance of Active Magnetic Regenerators
Submission Author:
Marco Antônio Cerutti , SC
Co-Authors:
Marco Antônio Cerutti, Guilherme Fidelis Peixer, Jaime Lozano, Jader Riso Barbosa Jr.
Presenter: Guilherme Fidelis Peixer
doi://10.26678/ABCM.ENCIT2022.CIT22-0135
Abstract
Machine Learning (ML) and data-driven modelling has developed drastically over the past few years with the increase in the ability to generate data from both numerical and experimental means and processing power of modern computers. Novel applications of these techniques are tested con- stantly, thriving when fast and accurate predictions of processes or phenomena with high modelling complexity are involved. Active Magnetic Regenerators (AMR), the heart of magnetic refrigeration systems, are an example of such scenarios. The modelling of these components involves non-linear equations for thermal, hydraulic and magnetic quantities, requiring closure relations, complex ma- terial properties characterization and application of numerical methods for the temporal and spatial solution of the problem. However, numerical solutions are hindered in applications where many sim- ulations are required, such as in optimization problems, or for fast predictions, such as in predictive control. Despite its advantages, AMR modelling through ML methods is not well established in the literature. For that, this work seeks to explore the capability of nine different machine learning tech- niques in the prediction of the cooling capacity produced by AMRs. The evaluated techniques were Linear Regression (LR); Second, Third and Forth order Polynomial Features (PF); Random Forest (RF); K-Nearest Neighbours (KNN); Support Vector Regression (SVR); Extreme Gradient Boosting (XGBoost) and Neural Networks (NN). A data set composed of 533 points from numerical simula- tions was employed, being 80% of it used as a training set, 20% used as a test set. A hyperparameter optimization was developed in the RF, KNN, SVR and XGBoost models by the GridSearch method, and in the NN by the Hyperband algorithm and for these cases 20% of the train set was employed as a validation set. The models were evaluated by the coefficient of determination (R2) associated with the test set. The algorithms which presented the best results were the Third Order PF and NN, with R2 of 0.998 and 0.995, respectively.
Keywords
magnetic refrigeration, machine learning, numerical models
