摘要：AbstractControl of real driving soot emissions in diesel vehicles requires accurate predictive models for engine-out soot emissions. This paper presents an innovative modeling approach that combines a physics-based model and a black-box model to predict soot from a 4.5-liter compression ignition engine under varying load and speed conditions. The physical model is based on an experimentally validated 1D engine model in GT-power. In contrast, the black-box model is designed by investigating different machine learning approaches, including a Bayesian neural network (BNN), support vector machine (SVM), regression tree, and an ensemble of regression tree. The experimental data from running the engine at 219 load and speed conditions are collected and used for training and testing the soot model. The least absolute shrinkage and selection operator (LASSO) feature selection method is used on the GT model outputs to find the most critical parameters in soot prediction. The grey-box modeling results are compared with those from the black-box as well as the physical model. The results show that the grey-box SVM and black-box single hidden layer BNN method provide the best performance with a coefficient of determination (R2) of 0.95. For most cases, grey-box models outperform the black-box models with the same Machine Learning (ML) algorithm by comparing R2of the test data, but this difference becomes negligible when a single hidden layer neural network is used.
