Prediction of instantaneous particle number for light-duty gasoline vehicles under real driving conditions based on ensemble learning

Abstract

With increasingly stringent emission regulations for particle number (PN) from real driving emissions (RDE), accurate prediction and evaluation of instantaneous PN emissions from RDE test is becoming increasingly important. However, there is still a lack of reliable PN emission prediction models for RDE test due to the complex nonlinear relationship between various affecting factors and PN emissions. Therefore, this study conducted RDE tests on a light-duty gasoline vehicle to acquire 10,646 sets of data, and four ensemble learning algorithms (randomforest, xgboost, catboost, and lightGBM) were introduced to construct prediction models of instantaneous PN emissions. Based on these models, the importance of 9 features affecting PN emissions was discussed using the Spearman correlation analysis and Shapley Additive Explanation (SHAP) algorithm. Among them, carbon dioxide share of exhaust (CO2,ratio), acceleration, vehicle specific power (VSP) and engine speed were identified as the critical input features. The catboost model has the optimal performance in prediction accuracy among the four models, with an average R2 and RMSE of 0.832 and 0.00253 respectively for the 10-fold cross-validated. Further analysis of the PN prediction results on different types of roads showed that the catboost model has high accuracy for instantaneous PN emissions on urban, rural and motorway trips. Moreover, the SHAP summary plot was employed to reveal the prediction mechanism of catboost algorithm through quantifying the feature contribution. The prediction model for instantaneous PN emissions provided a methodological reference for constructing vehicular emission models for RDE test.