Hierarchical evolutionary construction of neural network models for an Atkinson cycle engine with double injection strategy based on the PSO-Nadam algorithm

Abstract

There is a strong nonlinear relationship between the input and output of the Atkinson cycle engine (ACE), and with the development of artificial intelligence, fitting this nonlinear relationship with the neural network (NN) has become increasingly popular. In this paper, a lot of research has been conducted on constructing a more accurate NN model for ACE. Firstly, an ACE bench test is conducted, and the correlation analysis and dimensionality reduction of 14 parameters are performed by Pearson correlation coefficient (PCC), and five parameters, BSFC, CO2, CO, NOx and PN, are selected to investigate the effects of different injection strategies and EGR on the combustion, thermodynamics and emission performance of ACE. Secondly, the construction of the NN model of ACE is expressed as an optimization problem with constraints. Finally, a hierarchical evolutionary algorithm named PSO-Nadam is proposed, and the NN model built based on rules-of-thumb methods and the NN model built based on the PSO-Nadam algorithm are compared. The results show that there is a strong nonlinear relationship between BSFC, CO2, CO, NOx and PN with the injection strategy and EGR, and the PSO-Nadam-based NN model is better than the rules-of-thumb-based NN model in terms of both prediction performance and generalization ability in fitting this nonlinear relationship. It is worth mentioning that the rules-of-thumb-based model is overfitted in the prediction of BSFC, while the MSE of the PSO-Nadam-based model are reduced by 13.1%, 0.2%, 91.4%, 42.1% and the R2 are improved by 3.9%, 0.05%, 6.5%, 44.0% in the prediction of CO2, CO, NOx, and PN.