Utilização de aprendizado supervisionado na predição da demanda de energia no processo de produção de cumeno

Abstract

Data Science is a field of knowledge very requested nowadays, specially for digital and technology industries. Despite the increasing investment in this area, also by chemical industries, some elements of data science, such as Artificial Intelligence and machine learning have been used by these industries since 1980s, with artifical neuralnetworks (ANN) being the most used type of algorithms. ANN are great at solving complex non-linear problems but this type of algorithms have some disadvantages such as poor comprehensibility and high hardware dependence. Therefore, this study is about the use of three supervised learning models (linear regression, decision tree and random forest) of high comprehensibility and low hardware dependence for the prediction of total energy and specific energy (energy per mass of cumene) of a cumene producing plant by benzene with propene alkylation. The features are molar flow of propene, molar flow of benzene, temperature and pressure. Data were generated by simulations in Aspen Plus® software, treated and modeled using Python and MS Excel. The models were better at prediction of total energy than specific energy. Even with all the complexity of a chemical process, the models were able to predict the energy required with approximately ±10% errors.