From data to energy: Machine learning-driven predictions of biodiesel higher heating values under optimization constraints

The transportation industry still depends on fossil fuels, driving the need to transition to sustainable alternatives such as biofuels. Determining the viability of various biofuels as transportation fuels requires thorough knowledge of their fuel properties under different conditions. Conventional experimental methods of determining these properties are expensive and time-consuming, highlighting the need for reliable predictive techniques. This study is devoted to the precise prediction of the higher heating value of pure biodiesels through multiple linear and non-linear regressions, as well as machine learning techniques (artificial neural networks and support vector machines). The models are developed using density, kinematic viscosity, and cetane number data, and trained and validated over 241 biodiesel samples. The relative errors vary between 0.0039%–4.9195%, 0.0041%–8.9644%, 0.0062%–5.6850%, and 0.0018%–7.0050% for multiple non-linear, multiple linear, artificial neural networks, and support vector machines approaches, respectively. The average relative errors are computed as 0.9999%, 1.2612%, 0.9283%, and 0.9272% for multiple non-linear, multiple linear, artificial neural networks, and support vector machines approaches, respectively. The root mean square error values are computed as 0.5783 MJ/kg for multiple non-linear correlation, 0.7212 MJ/kg for multiple linear correlation, 0.5353 MJ/kg for artificial neural networks technique, and 0.6088 MJ/kg for support vector machines technique. The correlation coefficient value is computed as 0.9999 and 0.9998 for the multiple non-linear correlation and multiple linear correlation. The multiple non-linear regression model achieves the highest accuracy in predicting higher heating value of pure biodiesels. The study establishes also acceptable minimum ranges for higher heating value by solving a constrained optimization problem using the Newton method and Hessian matrix. The determined limited are approximately 35 MJ/kg and 33 MJ/kg, respectively, for EN 14214 and ASTM D6751 standards. Advanced machine learning methods will be evaluated to improve model accuracy in the future studies by incorporating a wider range of fuel properties.