Predicting Distillation Properties of Gasoline Fuel Blends using Machine Learning

Distillation properties of gasoline are regulated to ensure the safe and efficient operation of SI-engines. Blending various gasoline components affects the distillation values in a non-linear fashion, making the prediction of these properties challenging. Furthermore, the rise of renewable components necessitates the development of new property prediction methods. In this work, a variety of Machine Learning models were created to predict the distillation points of gasoline blends based on the blending recipe. As input data, real industrial data from a refinery was used together with data from blends created for R&D purposes. The predicted properties were the evaporated volume at the 70 and 100 °C distillation points (E70 and E100). Altogether four different machine learning models were trained, cross-validated and tested using seven different pre-processing methods. It was found that Support Vector Regression (SVR) was the most effective at predicting the distillation points. It achieved a mean absolute error (MAE) of 1.5 and 1.1 respectively for E70 and E100, compared to a widely used linear model which had MAEs of 12.2 and 3.27. These findings underline that the SVR model was substantially more effective. In addition, the SVR model maintained its performance when used on oxygenated blends for which the linear model was highly inaccurate. It was also found that higher distillation points behave more linearly and therefore there's less need for complex non-linear models near the higher end of the distillation curve. The methodology presented in this work can be further used to predict other distillation points or fuel properties.