Octane optimization with a combined machine learning and optimization approach

Abstract

Refinery operations are always sensitive to optimization, and due to the increasingly adverse effects of COVID-19 on energy sectors, its importance had increased significantly. This thesis aims to analyze the reactor temperature that yields a higher RON (octane measurement) value in isomerate product using all available information in the isomerate production network. The main explanatory variables that can affect the RON value can be divided into three categories: feed impurities, isomerization reactor operations, deizohexanizer column operations. Isomerate feed network is quite complex and fed by different crude distillation units and cracker units. Various reactions occur in the isomerization reactors, and depending on the feed content, the reaction mechanism changes. This thesis applies machine learning algorithms to build a model that can capture the relationship between RON and reactor temperature with the other explanatory variables. We implemented a number of machine learning algorithms to assess their performance on the problem, specifically Linear Regression, Decision Tree, Random Forest, XGBoost, Support Vector Regression, and KNN. Comparing with the linear regression, we achieved 0.82 decreases in the mean absolute error. The mean absolute error of the XGBoost model is 0.08 RON. We find a temperature value with the selected model that yields a higher RON number by trying different temperature values while keeping the same values for the other variables. If we used the suggested temperature by our model, we predict that we could obtain a 0.2 RON increase in the validation zone resulting in an annual profit increase of 528 000 USD Dollar.