Soot Mass Concentration Prediction at the GPF Inlet of GDI Engine Based on Machine Learning Methods

• Published in: Energies (Basel) Vol. 18; no. 14; p. 3861
• Main Authors: Hu, Zhiyuan, Liu, Zeyu, Shen, Jiayi, Wang, Shimao, Tan, Piqiang
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.07.2025
• Subjects: Accuracy; Air pollution; Air quality management; Algorithms; Data collection; Datasets; Diesel engines; Energy consumption; Engines; Gasoline; GDI engine; GPF; Machine learning; Methods; Neural networks; prediction; soot mass concentration
• DOI: 10.3390/en18143861
• ISSN: 1996-1073

To improve the prediction accuracy of soot load in gasoline particulate filters (GPFs) and the control accuracy during GPF regeneration, this study developed a prediction model to predict the soot mass concentration at the GPF inlet of gasoline direct injection (GDI) engines using advanced machine learning methods. Three machine learning approaches, namely, support vector regression (SVR), deep neural network (DNN), and a Stacking integration model of SVR and DNN, were employed, respectively, to predict the soot mass concentration at the GPF inlet. The input data includes engine speed, torque, ignition timing, throttle valve opening angle, fuel injection pressure, and pulse width. Exhaust gas soot mass concentration at the three-way catalyst (TWC) outlet is obtained by an engine bench test. The results show that the correlation coefficients (R2) of SVR, DNN, and Stacking integration model of SVR and DNN are 0.937, 0.984, and 0.992, respectively, and the prediction ranges of soot mass concentration are 0–0.038 mg/s, 0–0.030 mg/s, and 0–0.07 mg/s, respectively. The distribution, median, and data density of prediction results obtained by the three machine learning approaches fit well with the test results. However, the prediction result of the SVR model is poor when the soot mass concentration exceeds 0.038 mg/s. The median of the prediction result obtained by the DNN model is closer to the test result, specifically for data points in the 25–75% range. However, there are a few negative prediction results in the test dataset due to overfitting. Integrating SVR and DNN models through stacked models extends the predictive range of a single SVR or DNN model while mitigating the overfitting of DNN models. The results of the study can serve as a reference for the development of accurate prediction algorithms to estimate soot loads in GPFs, which in turn can provide some basis for the control of the particulate mass and particle number (PN) emitted from GDI engines.