Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
APPLICATIONS OF NEURAL NETWORKS INTO HEAT EXHANGERS TYPE PCHE
Submission Author:
Lucas Rangel Freire , ES , Brazil
Co-Authors:
Lucas Rangel Freire, Carolina Palma Naveira Cotta
Presenter: Lucas Rangel Freire
doi://10.26678/ABCM.COBEM2021.COB2021-2257
Abstract
This work involves the study of compact printed circuit heat exchangers. The PCHEs have advantages in several industrial applications when compared to other traditional heat exchangers on the market, features stand out are: compact size, reduced weight and the ability to operate at high pressures and high temperatures. The objective of this work was to apply the neural networks method to predict the impact on the thermo-hydraulic performance of a heat exchanger, due to changes in either the internal geometry of the channels and changes in operating conditions. High computational cost simulations involving the solution of partial differential equations related to the principles of conservation of mass, momentum and energy were employed in the training of the neural network. The results for the Nusselt number and Fanning friction factor were obtained through the artificial intelligence method and compared with the predictions provided by correlations in the literature. One of main advantages of correlations and neural networks compared with CFD simulations is ability to predict fastly, even the results is being less accurate, so is more acceptable the companies incorporate in daily activities neural networks, which generate a result in seconds, once CFD simulation results may take days. Moreover, the artificial neural network (ANN) was trained by backpropagation algorithm, a supervised method responsible for training the weights in a multilayers feed-forward neural network. An attempt was made to select the optimum parameters for the ANN, such as number of neurons; number of hidden layers; activation function; training function… this was possible by an iterative method developed. For each configuration was calculated a mean squared error, and after all, the architecture with minimum error was chose. In addition, other tests have been made with white noise at input in order to measure the distortion at ANN’s output. In conclusion, the predictions from ANN were compared with correlation’s predictions and true labels, for the case in analysis, the ANN’s results overcome the correlation’s results.
Keywords
Printed circuit heat exchanger, artificial neural network, PCHE, artificial inteligence
