Data-Driven Modeling And Optimization Of Engine Performance Fueled By Karanja Biodiesel: A Comparative Approach

Abstract

The increasing demand for sustainable fuels has prompted the exploration of biodiesel as a viable alternative to conventional diesel. Karanja biodiesel, derived from non-edible oil sources, offers promising properties. However, its impact on engine performance necessitates advanced modeling for prediction and optimization. This study investigates the use of Artificial Neural Networks (ANN), Response Surface Methodology (RSM), and Machine Learning Random Forest (RF) techniques to model and optimize engine performance parameters using Karanja biodiesel blends. Experimental data were collected from a single-cylinder diesel engine by varying inputs—Speed, Load, Fuel Blend (%), and Compression Ratio. Key performance outputs included Indicated Power (IP), Brake Power (BP), IMEP, Brake Thermal Efficiency (BThEff), Specific Fuel Consumption (SFC), Torque, and Mechanical Efficiency. The ANN model (4–10–7 architecture) trained using MATLAB 2014a achieved high prediction accuracy (R²: 0.994 for BThEff, 0.998 for Torque). Optimization using fmincon yielded maximum BThEff of 33.37% at Speed 1454.84 rpm, Load 12.19 kg, Fuel 23.56%, and CR 16.65.RSM analysis using Minitab showed excellent fit (R² = 1.000 for BP, 0.995 for BThEff), and identified optimal parameters through the desirability function. RF performed competitively, with the highest R² of 1.000 for Torque but showed high MAPE for SFC (78.12%). Comparative analysis revealed that RSM had the lowest average MAPE across most outputs. This study demonstrates that ANN and RSM are effective tools for biodiesel engine modeling and optimization, reducing experimental workload and supporting efficient biodiesel utilization.