Multi-objective optimization for combustion, thermodynamic and emission characteristics of Atkinson cycle engine using tree-based machine learning and the NSGA II algorithm. (15th June 2023)

• Record Type: Journal Article
• Authors: Sun, Xilei; Xie, Mingke; Zhou, Feng; Fu, Jianqin; Liu, Jingping
• Is Part Of: Fuel. Volume 342(2023)
• Publication Date: 2023-06-15
• Subjects: ACE Atkinson cycle engine -- AdaBoost Adaptive Boosting -- ANN Artificial Neural Network -- BSFC brake specific fuel consumption -- CART classification and regression tree -- CFD computational fluid dynamics -- CO carbon monoxide -- COV coefficient of variation -- DL deep learning -- EGR exhaust gas recirculation -- HEV hybrid electric vehicle -- IDT ignition delay time -- GA genetic algorithm -- GDI gasoline direct injection -- LIVC late intake valve closing -- MCP maximum combustion pressure -- MPRR maximum pressure rise rate -- MSBA mutable smart bee algorithm -- MSE mean squared error -- NOx Nitrous Oxides -- PCC Pearson correlation coefficient -- PF Pareto front -- PN particle number -- PSO particle swarm optimization -- RF Random Forest -- SA spark timing angle -- SOC start of combustion -- SVM support vector machine -- THC total hydrocarbon -- THS Toyota Hybrid System -- VVT-I intake variable valve timing -- VVT-E exhaust variable valve timing; Atkinson cycle engine -- Adaptive Boosting -- Machine learning -- Multi-objective optimization -- NSGA II algorithm; Fuel -- Periodicals; Coal -- Periodicals; Coal; Fuel; Periodicals; 662.6
• DOI: 10.1016/j.fuel.2023.127839
• Languages: English
• ISSNs: 0016-2361

Abstract

Graphical abstract: Highlights: Effects of injection timing, injection pressure and EGR rate on characteristics of ACE are investigated. Three tree-based machine learning models for ACE are built and compared. Multi-objective optimization is carried out based on the AdaBoost model and NSGA II algorithm. Non-dominated solutions play a good optimization effect relative to the original engine. Abstract: In this study, the effects of injection timing, injection pressure and exhaust gas recirculation (EGR) rate on combustion, thermodynamic and emission characteristics of the Atkinson cycle engine (ACE) were investigated experimentally, and the multi-objective optimization was performed so as to improve its overall performance. Firstly, the engine bench test was conducted, and a lot of associated characteristic parameters were collected, which were used as the data basis for the following model building. Secondly, three tree-based (classification and regression tree (CART), Random Forest (RF) and Adaptive Boosting (AdaBoost)) machine learning models for ACE were developed, in which RF and AdaBoost were built respectively by the parallel method and the serial method with CART as the base learner. Finally, the multi-objective optimization for combustion, thermodynamic and emission characteristics of ACE was carried out based on the AdaBoost model and the NSGA II algorithm. Research result show that three tree-based models have high prediction performance and generalization ability (AdaBoost is the best, followed by RF and then CART), and the serial method (AdaBoost) is better than the parallel method (RF) for the data set of this study. Both the second (Optimal BSFC) and fourth (Optimal PN) non-dominated solutions play a good optimization effect relative to the experimental data of the original engine, with CO, BSFC, NOx and PN reduced by 31.9%, 2.3%, 39.1%, 61.6% and 32.3%, 2.2%, 40.0%, 62.6%, respectively.