This content is not included in your SAE MOBILUS subscription, or you are not logged in.
A Study on Machine Learning Algorithms to Support Production of Alternative Fuels for SI Engines
ISSN: 0148-7191, e-ISSN: 2688-3627
To be published on September 15, 2020 by SAE International in United States
Biofuels and synthetic fuels represent an immediate alternative that can effectively help decarbonize the transport sector, but in order to become commercially competitive they need to meet fuel quality standards. Research Octane Number (RON), among other auto-ignition related properties, is a primary indicator of the grade of spark-ignition (SI) fuels, especially relevant for modern highly boosted direct injection SI engines. However, the non-linear behavior of such properties in blends still constitutes a challenge for their prediction in the final product. This study compares popular Machine Learning algorithms and evaluates their potential to develop state-of-the-art models able to predict key SI fuel properties. At the same time, a simple and systematic methodology is established. Typical gasoline composition was simplified and represented by a palette of seven characteristic molecules, including five hydrocarbons and two oxygenated species. Ordinary Least Squares (OLS), Nearest Neighbors (NN), Support Vector Machines (SVM), Decision Trees (DT) and Random Forest (RF) algorithms were trained, cross-validated and tested using a data base containing 243 gasoline-like fuel blends with known RON. Linear methods proved to perform better using molar compositions while predictions on a volumetric basis required non-linear algorithms for satisfactory accuracy. Best results were obtained with the SVM algorithm using a non-linear kernel, able to reproduce synergistic and antagonistic molecular interactions. For this model, the Mean Average Error (MAE) on the test set was equal to 0.9 octane numbers. Moreover, the algorithm maintained its predictive accuracy when alterations and restrictions were performed on the training data set, proving its robustness and suitability for the application. Developed models enhance the understanding of the blending behavior of different hydrocarbons and oxygenates and promote increasing flexibility in advanced gasoline production and blending. Moreover, these models serve as a screening tool for preliminary fuel selection and property evaluation in engine-performance related studies.