Biodiesel composition based machine learning approaches to predict engine fuel properties

Abstract

Recently data-driven machine learning approaches received considerable attention in several applications, including developing models to predict engine fuel properties of biodiesel. Multilinear regression (MLR) is the most straightforward method among the available approaches in the literature to predict biodiesel properties. However, a nonlinear correlation between biodiesel composition and properties cannot be modeled using the MLR approach, resulting in poor predictability. Artificial neural network (ANN), the most explored nonlinear regression approach to predict biodiesel properties, is prone to poor reproducibility. Support Vector Machine (SVM) regression is a machine learning approach that can be applied to develop biodiesel property prediction models whose favorable features include nonlinear data modeling. Models are developed to correlate biodiesel composition with calorific value, viscosity, and cetane number in the present work using ANN and SVM regression. Seventy biodiesels with varying compositions are used for calibrating the models and 30 other biodiesels for validation. Both the nonlinear regression approaches perform well in predicting the biodiesel properties, among which SVM results in better prediction than ANN. SVM resulted in models with a MAPE of 0.26%, 1.07%, and 1.69%, respectively, for calorific value, viscosity and cetane number of biodiesels which are considerably lower than those predicted using the literature models.