Comparison and implementation of machine learning models for predicting the combustion phases of hydrogen-enriched Wankel rotary engines

Abstract

Combustion phases, such as the development period (CA0-10) and flame propagation period (CA10-90), are the critical parameters for hydrogen-enriched Wankel rotary engines. An accurate simulation model and a suitable engine management system are required to control combustion phases. In this paper, five machine learning (ML) models, including the linear regression (LR), regression tree (TR), ensembles of trees (EnTR), support vector machine (SVM), and Gaussian process regression (GPR), are initially applied to predict combustion phases. Experiments were performed with variations of the main fuel types (gasoline and n-butanol), loads (idle and part load), ignition timing, hydrogen volume fraction, and excess air ratio. The sample data were divided into training and testing data set, and the normalization method, 5-fold cross-validation, and Bayesian optimization algorithm were used for data processing and model optimization. Among five ML models, the training speed of the LR model was the fastest; the generalization ability of the TR model was the worst. The minimum leaf size of the TR model significantly influenced regression and generalization ability. On this basis, the EnTR model improved the regression ability, but required more training time. The GPR model showed the best generalization ability among the above model, while SVM performed well in a certain data set. For CA0-10, the coefficient of determination (R2) of the best LR, TR, EnTR, SVM and GPR models was 0.9910, 0.9912, 0.9985, 0.9984 and 0.9994, respectively; for CA10-90, the R2 was 0.9348, 0.8974, 0.9873, 0.9916 and 0.9975, respectively. It is highly recommended to apply the GPR model to the combustion phases prediction and control system modeling.