Hybrid XGBoost-RF-MLP model and PSO optimization for performance and emissions of CI engine using waste cooking biodiesel blends

Transesterification was used to create methyl ester from waste cooking oil (WCO). Diesel oil and biodiesel blends in 25, 50, 75, and 100% were developed and authorized by ASTM. The primary contribution of this study lies in integrating experimental WCO biodiesel data with a novel hybrid machine learning and Particle Swarm Optimization (PSO) framework. A hybrid model, combining XGBoost, Random Forest, and MLP, was developed to predict engine performance and emissions. The core novelty is the use of base model predictions as meta-features for a final meta-learner, createing a superior stacked ensemble. This hybrid model was then coupled with PSO to identify optimal engine operating conditions. Key experimental results revealed that pure biodiesel (B100) reduced CO, HC, and smoke emissions by 25%, 43%, and 45%, respectively. However, increased NOx emissions by 23% and brake-specific fuel consumption by 22% were shown compared to diesel at full load. Crucially, the hybrid model demonstrated exceptional predictive accuracy, achieving a significantly lower Mean Squared Error (MSE in the order of 10⁻⁷) across all 13 output parameters compared to the individual MLP (MSE ~ 10⁻³), RF (MSE ~ 10⁻⁴), and XGBoost (MSE ~ 10⁻⁶) models. The PSO algorithm successfully converged to an optimal solution of 86% engine load and 26% biodiesel blend (B26), maximizing the defined fitness function that balanced performance and emissions. The results unequivocally demonstrate that the proposed hybrid modeling approach offers a robust and highly accurate framework for engine optimization, establishing WCO biodiesel as a viable alternative fuel when used in optimal blends.