An Enhanced Light Gradient Boosting Regressor for Virtual Sensing of CO, HC and NOx

Abstract

The present study introduces a novel methodology that utilizes Light Gradient Boosting Regressors to predict engine-out emissions of NOx, HC, and CO. The accuracy of the proposed models is evaluated on different types of homologation cycles. The dataset used in this study is derived from a set of 47 experimental driving cycles, including RDE, WLTC, NEDC, ECE, US06, and HWFET. The experimental driving cycles are performed on a roll bench using a spark-ignited, naturally aspirated, V12 engine-equipped vehicle. A three-second sliding window is incorporated in the models to capture the dynamic behavior of pollutant emissions. The performance of the LightGBR models is assessed using the mean absolute percentage error (MAPE) on the total pollutant mass, which is found to be 5.2% for CO, 5.7% for HC, and 6.8% for NOx. The results demonstrate the efficacy of the proposed methodology, which can be used to estimate the impact of powertrain calibration changes on pollutant emissions in a virtual environment, thereby reducing the number and the cost of the experimental tests.