Modelo machine learning de um trocador de calor tipo placa via modelo gerado no Aspen

Abstract

Mathematical optimization methods tend to be increasingly applied, both due to the greater capacity of computers, as well as the greater knowledge of the behavior of the systems and consequent knowledge of the variables to be modified, in order to obtain an optimized response. In the search for opportunities for improvement in chemical processes, such as predicting mass flow of refrigeration streams in the beverage industry, artificial intelligence and machine learning (ML) have been commonly used, which have been used in this industry since the 1980s, as they are sufficiently rigorous models and quick to activate control devices effectively. Therefore, this work has the final objective of building a mathematical model using machine learning, which calculates the mass flow rate of the ethanol input stream (cooling fluid) in the plate heat exchanger at the brewery. From measurements of temperature, pressure and flow of the real currents of the company's heat exchangers (AMBEV), a model was developed in Aspen HYSYS®, to represent the behavior of these exchangers. From a mesh of inputs, the model was used to simulate the outputs. These outputs were used to build an empirical model that represented the process. Three types of empirical models were used for supervised learning: linear regression, multilayer perceptron and decision tree. To optimize and evaluate the models, data were separated into cross-validation and testing. The decision tree model was able to better predict the mass flow rate of the Ethanol stream with a standard deviation of 1.92×10^(-08) and 0 Kg/s, for the respective heat exchangers.