CRDI Engine Emission Prediction Models with Injection Parameters Based on ANN and SVM to Improve the SOOT-NOx Trade-Off

Abstract

Artificial Neural Network (ANN) and Support Vector Machine (SVM) have been widely used to solve non-linear problems. In the current study, based on 112 groups of experimental data, ANN and SVM models were established and compared to improve the trade-off relationship between SOOT and NOx emissions of a Common Rail Diesel Injection (CRDI) engine fueled with Fischer-Tropsch (F-T) diesel under different operating conditions and injection parameters. The model parameters for the different predictive targets were selected by evaluating the mean square error (MSE) and determination coefficient. Compared to the number of network iterations, the number of implied nodes had a greater effect on the MSE of the ANN model. Compared to the penalty parameter, the width coefficient had a weaker impact on the SVM performance. A comparative analysis showed that the SVM had better predictive accuracy and generalization ability than the ANN, with a maximum error not exceeding five percent and a determination coefficient of over 0.9. Subsequently, the optimal SVM model was combined with the NSGA-II algorithm to determine the optimal injection parameters for the CRDI engine, resulting in solutions to simultaneously decrease the SOOT and NOx emissions. The optimized injection parameters resulted in a 3.7–7.1% reduction in SOOT emission and a 1.2–2.6% reduction in NOx emissions compared to the original engine operating conditions. Based on limited experimental samples, SVM is inferred to be a useful tool for predicting the exhaust emissions of engines fueled with F-T diesel and can provide support for optimizing injection parameters.

Keywords

• crdi engine
• f-t diesel
• machine learning algorithm
• emission
• model optimization