Refining the Understanding of Ozone Formation Response Regulations Through Ensemble Machine Learning Analysis in Highly Polluted Areas

Abstract

The comprehensive understanding of the intricate processes of ozone (O3) formation in highly polluted areas is still limited. An ensembled machine learning (ML) approach that integrates random forest (RF) and artificial neural network (ANN) algorithms was proposed with strong prediction performance. Through controlled simulation experiments with high frequency values, a quantified O3 concentration calculation function was established with NO2 concentration, T and RH, as the most important variables. The function exhibits a robust predictive capability of 0.751 under stable atmosphere conditions. Moreover, the incremental reactivity (IR) of NO2 shows limited correlation with its initial concentration, the influence of NO2 concentration on O3 formation was primary-linearly and quadratic-linearly controlled by T and RH, respectively. Additionally, the relative incremental reactivity (RIR) of NO2 would increase with the increasing of the ratio of the concentration of NO2 and O3. However, the influence of both total VOCs and individual VOC concentrations on O3 was found to be weak. The IRs of individual VOC species were relatively small, indicating a weak sensitivity to O3 concentration. The results help to the refinement of the understanding regarding O3 formation in NOx sensitivity and VOC oversaturation areas, thereby providing further support for air quality improvement in similar areas.

Keywords: O3 formation, response regulation, NOx-sensitivity, machine learning (ML), incremental reactivity (IR)