Metamodelagem Kriging dinâmica aplicada em trocadores de calor.

Abstract

Due to the recent technological advances in the industrial segment with the development of the foundations of industry 4.0, such as big data additive manufacturing, simulation and internet of things (IoT), artificial intelligence concepts and techniques became widely known. The machine learning (ML) presents itself as a branch of artificial intelligence, which had its development propelled along with the rise in storage capabilities and data analysis. The gaussian process regressor method (Kriging) is one of the ML methods recommended for industrial processes applications due to its ability to deal with highly non-linear system responses, aside from the fact that it provides estimates regarding statistical error by modeling it as gaussian processes. Aiming to contribute to the advances in the industrial sphere by using ML techniques, the present work sets as goal to study Kriging modeling by the means of a methodology developed to determine the process variables dynamic behavior. As case study, it was used a counterflow heat exchanger simulated in the Aspen Plus DynamicTM software in order to obtain dynamic data of the highest coupled variables to the equipment. The model was built with the PythonTM programming language by using both stationary correlation models (radial basis function (RBF), rational quadratic (RQ) and matern) and non-stationary (dot product) to evaluate which one provided the best prediction. The model training (one hundred cases) and validation (ten cases) data sets were generated using the latin hypercube sampling (LHS) technique. The model was validated by evaluating the mean squared error (MSE), the explained variance score (EVS) and the determination coefficient (R2) applied to each correlation model. Ultimately, a wavelet filter coupled to a moving average filter was proposed to smoothen the behavior predicted by the models. The results indicate that for the two model outputs (output temperature of the hot and cold fluid) only the dot product correlation presented significant results, yielding behavior inside the range of ± 4% deviation in the validation scenario selected for appraisal. The MSE, R2 and EVS tests confirm the observed results for the predicted behavior for both fluid temperatures. The addition of filters made it possible to confirm the applicability of themselves while coupled with the resultant model from dot product correlation.