Ensemble Machine Learning and Simulated Annealing for Behaviour Prediction of a Diesel Engine Fuelled with Dimethyl Ether

Abstract

This study introduces a novel hybrid approach based on ensemble machine learning prediction and simulated annealing for the optimisation and minimisation of emissions from dimethyl ether (DME)-diesel fuel combustion compression ignition engines. Experimental data were obtained from a commercial, four-cylinder, direct injection diesel engine (Model 4113) under different load conditions, with varying blends of DME and diesel. Key emissions (NOx, CO, and HC), the brake mean effective pressure and the blend ratio were measured with high precision by AVL analysers. Multiple regression models were built and compared, including support vector regression (SVR) and XGBoost, with rigorous hyperparameter tuning. SVR yielded better performance for NOx (test R2 = 0.820) and HC (test R2 = 0.854), although both models achieved good accuracy for CO (SVR test R2 = 0.854; XGBoost test R2 = 0.731). A correlation analysis showed that there was a strong positive link between engine load and NOx, and good HC reduction with higher DME content. The trained ensemble models were implemented as objective functions in a simulated annealing algorithm to achieve multi-objective optimisation of critical engine parameters. The proposed approach has a high level of predictive reliability and effective global search capability, and provides a computationally efficient pathway for emission control in alternative fuel engines. These findings provide support for the broader use of DME as a clean, renewable fuel for compliance with stringent future emission standards.