Vehicle Exhaust Concentration Estimation Based on an Improved Stacking Model

This paper aims to estimate the carbon monoxide (CO) and hydrocarbon (HC) concentrations of vehicle exhaust. For this purpose, an improved Stacking model is designed. Compared with individual estimation models, the improved Stacking model can achieve better concentration estimation performance. That model has a three-layer structure. The first layer is made up of multiple estimation models, which produce intermediate estimation results from the original exhaust data based on the
$K$
-fold cross-validation. The second layer takes these intermediate estimation results as input and trains a statistical learning model which generates preliminary estimation results of the concerned exhaust concentrations. The first two layers actually constitute the Stacking model, which is extended by the additional third layer of our improved one. The third layer implements a weighted summation method. More specifically, the preliminary estimation results generated by the second layer are linearly combined with the concentration estimation results of some strong estimation models, such as XGBoost and LightGBM, to produce the final estimation results in the third layer. Our improved Stacking model is verified through experimental data, which were collected by a Urban Road Network Vehicle Emissions Monitoring System and small weather stations in two Chinese cities, including Beijing and Jiaozuo. Experimental results show that compared with some regression and neural network estimation models, especially the Stacking model and Boosting models, our improved Stacking model achieves higher exhaust concentration estimation accuracy.