Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Modelling the dynamic operation of a magnetic refrigerator via recurrent neural networks
Submission Author:
Guilherme Fidelis Peixer , SC , Brazil
Co-Authors:
Pedro Miola Silva, Yan Azeredo, Anderson Lorenzoni, Guilherme Fidelis Peixer, Rodolfo C. C. Flesch, Jaime Lozano, Jader Riso Barbosa Jr.
Presenter: Guilherme Fidelis Peixer
doi://10.26678/ABCM.COBEM2023.COB2023-1603
Abstract
Environmental concerns and governmental policies have increased pressure on the refrigeration industry recently. Various technologies have been explored to overcome this problem, with magnetic refrigeration being one of the most promising. Despite recent prototype developments, this technology is still not competitive enough with vapor compression, especially due to high energy consumption. To improve this issue, it is possible to implement a suitable control technique such as Model Predictive Control, based on optimization that allows the reduction of power consumption. This work is focused on the development of two Neural Network models, focusing on Model Predictive Control applications. The outlet temperature of the cold manifold and the total power consumption of the magnetic refrigerator are the output parameters. Through a Design of Experiments methodology, the excitation signals for the identification experiment were proposed. Hence, 5 actuation signals, namely the magnetization frequency, pump power supply, hot side fan power supply, cold side fan power supply and blow fraction, and 3 disturbances, namely the hot chamber temperature, cold chamber temperature and outlet cold side manifold temperature, were selected to compose the models. The excitation signal length was chosen as 20 hours of maximum duration, which consisted of 80 excitation points, placed inside the input space with a Latin Hypercube distribution. The shape of the excitation utilized an Amplitude modulated Pseudo-Random Binary Signal, that excites both frequency and amplitudes of the variables. Two other experiments were executed to be used as validation and test data sets. A three-step approach was used to select the best parameters and architecture of the Nonlinear Autoregressive with Exogenous Inputs Neural Networks. The temperature model achieved an R² of 0.942 with a 0.27 °C mean residual for one-step-ahead prediction and an R² of 0.852 with a 0.42 °C mean residual for 90-step-ahead predictions. The power supply model achieved an R² of 0.987 with a mean residual of 22.76 W for one-step-ahead predictions and an R² of 0.898 with a 68.91 W mean residual for 90 steps-ahead predictions.
Keywords
magnetic refrigeration, machine learning, Recurrent Neural Networks, energy efficiency
