Comparison and evaluation of advanced machine learning methods for performance and emissions prediction of a gasoline Wankel rotary engine

In order to improve the performance, reduce the emissions and enhance the calibration efficiency of a gasoline Wankel rotary engine (WRE), three advanced machine learning (ML) methods, including artificial neural network (ANN), support vector machine (SVM), and Gaussian process regression (GPR), were applied to develop the prediction model of the torque, fuel flow, nitrogen oxide, carbon monoxide, and hydrocarbon. The effect of feature numbers was examined using the recommended parameters of the ANN, SVM, and GPR models. It was concluded that using speed, manifold absolute pressure, and air fuel ratio as input parameters to build the prediction model performed best. The generalization ability of the three ML models was compared on the interpolative and extrapolative data sets using the extended recommendation parameters. The results showed that the GPR model performed the best generalization ability in scarce data sets and was simpler to train compared with ANN and SVM. The response surfaces constructed using the GPR model were very smooth and accurate, in which the coefficient of determination for all the predicted parameters was more than 0.99. It is strongly proposed that the GPR approach is a universal approach which will be an essential direction for WRE system control and surrogate model modeling.

Automated summary

In this study, machine learning methods were applied to develop a prediction model for a gasoline Wankel rotary engine. Artificial neural network, support vector machine, and Gaussian process regression were compared, and the best parameters for the prediction model construction were determined. The results showed that the Gaussian process regression model had the best generalization ability in scarce data sets.