With the continuous promotion of the carbon peak emissions and carbon neutralization strategies, higher demands are placed on engine economic performance. Virtual sensors as an online information collection technology can be used to control various performance indicators of engines. Here is an example of ISFC to represent the engine performance prediction. In this paper, the feasibility of three machine learning methods, Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Regression (SVR), for predicting fuel consumption applications are explored. Firstly, a calibrated engine one-dimensional (1D) model is constructed. Then, the 1D model generates a dataset with engine load, engine speed and spark time, and indicative specific fuel consumption (ISFC) as an output, for the training of machine learning methods. The performance of different algorithms was compared using the coefficient of determination (R2), the root-mean-square error (RMSE), and the mean absolute percentage error (MAPE) as evaluation metrics. By comparing test dataset prediction and map prediction, RF has a large prediction error at boundary operation conditions and ANN sometimes has a relative error of more than 10%. SVR performs well in each statistical index and map prediction, and therefore it is an algorithm that can be used by virtual sensors.
Keywords carbon peak emissions; carbon neutralization; virtual sensor; machine learning methods