Quantification of COVID-19 impacts on NO2 and O3: Systematic model selection and hyperparameter optimization on AI-based meteorological-normalization methods

ABSTRACT

Since the unprecedented outbreak of the COVID-19, numerous meteorological-normalization techniques have been developed in lockdown-imposed regions to decouple the impacts of meteorology and emissions on the atmospheric environment. However, the application of normalization techniques in regions without lockdown is limited. In this study, we propose a novel research framework to investigate the observed and meteorological-normalized concentrations of nitrogen dioxide (NO2) and ozone (O3) across 62 cities in Taiwan. Four meteorological-normalization techniques, namely, the generalized additive model (GAM), generalized linear model (GLM), gradient boosting machine (GBM), and random forest (RF), were developed, optimized, and compared using meteorological and temporal variables. The models were optimized using a systematic trial-and-error approach for data distribution type in GAM and GLM and a grid-search approach for tree numbers in GBM and RF. Based on the findings, for GLM, the optimal data distribution for both NO2 and O3 modeling was Gaussian, whereas for GAM, the optimal data distribution for NO2 and O3 simulation was quasi- Gaussian and Poisson, respectively. In contrast, for RF and GBM, the optimized number of trees varied significantly by site, ranging from 10 to 6310. The simulation performance of RF and GBM was better than that of GAM and GLM across Taiwan and the best-performing optimized model was selected to identify changes in NO2 and O3 concentrations during COVID-19. Throughout 2020, even in the absence of a lockdown, the daily mean meteorological-normalized NO2 and O3 levels across Taiwan decreased by 14.9% and 5.8%, respectively, providing novel insights for sustainable air quality management.Copyright © 2023