Improvements of response surface modeling with self-adaptive machine learning method for PM2.5 and O3 predictions

Abstract

Quickly quantifying the PM2.5 or O3 response to their precursor emission changes is a key point for developing effective control policies. The polynomial function-based response surface model (pf-RSM) can rapidly predict the nonlinear response of PM2.5 and O3 to precursors, but has drawbacks of overload computation and marginal effects (relatively larger prediction errors under strict control scenarios). To improve the performance of pf-RSM, a novel self-adaptive RSM (SA-RSM) was proposed by integrating the machine learning-based stepwise regression for establishing robust models to increase the computational efficiency and the collinearity diagnosis for reducing marginal effects caused by overfitting. The pilot study case demonstrated that compared with pf-RSM, SA-RSM can effectively reduce the training number by 70% and 40% and the fitting time by 40% and 52%, and decrease the prediction error by 49% and 74% for PM2.5 and O3 predictions respectively; moreover, the isopleths of PM2.5 or O3 as a function of their precursors generated by SA-RSM were more similar to those derived by chemical transport model (CTM), after successfully addressing the marginal effect issue. With the improved computation efficiency and prediction performance, SA-RSM is expected as a better scientific tool for decision-makers to make sound PM2.5 and O3 control policies.