LAUSR

Abstract In recent years, support vector machine (SVM) method has been rapidly developed because of its great advantage in solving small sample regression problems. Based on the prediction accuracy of NO x emission, the SVM method is applied to the regression analysis of the steady-state calibration experimental results of a hydrogen enriched compressed natural gas (HCNG) engine in this research article. The effects of the model parameters (penalty factor kernel, function width and insensitive band loss function) and the training sample size on the prediction accuracy of the regression model are studied. Results show that both model parameters and training sample size can influence the prediction accuracy of the SVM model. Additionally, the method of determining the optimal SVM regression model is also summarized. The optimal SVM regression model is obtained by the manifold absolute pressure (MAP) and the fuel equivalence ratio ( θ ) halved sample, with the training sample size of 270 for the experimental data used in this study. Results show that the optimal SVM regression model can decrease the predicted mean absolute percentage error (MAPE) and the maximum relative prediction error (MRE) of the brake specific NO x emission greatly, from 12.54% to 8.32% and 56.66% to 25.89%, respectively. It indicates that the prediction performance can be improved apparently by the method promoted in the paper, which provides a new perspective for the further application of SVM method in the field of automobile engines calibration.