Desenvolvimento de estratégias para predição de incrustação em trocadores de calor baseadas em técnicas de aprendizado de máquina

Abstract

One of the main equipments used to heat, cool or condense fluid streams is the heat exchanger and the fouling formation in these equipments is one of the most critical problems in the industry, causing a decrease in the exchange efficiency and obstruction of the fluid passage. Thus, fouling leads to excessive costs, undesired shutdowns, and total loss of material, making it necessary to stop the process to remove the fouled material in order to achieve good thermal efficiency of the equipment. An alternative to determine the degree of fouling in a heat exchanger is in the form of virtual sensors, which are predictive models that can describe a parameter or variable as a function of other easily measured ones, enabling the ability to monitor, analyze possibilities, predictions, and can be built using phenomenological modeling, statistics or based on intelligent systems, more specifically with the application of machine learning (ML) techniques such as ANN and SVM. In this sense, the objective of this work was to develop strategies for the prediction of fouling in heat exchangers, in order to allow better planning of maintenance stops and, therefore, improve process efficiency. Thus, the process was modeled and simulated from its physical and chemical characteristics in Aspen software to obtain and analyze data and thus apply ML techniques in the form of computational tools in Python language to assist in maintenance planning, providing previous information on fouling levels in the exchanger in operation. Thus, based on the defined strategies applied with ANN and SVM, the results showed a great performance with R2>0.99 for the model studied, in which the behavior was repeated in both techniques, regardless of their structure, being able to predict the variables involved in the heat exchanger with excellent accuracy.