A data-driven air quality assessment method based on unsupervised machine learning and median statistical analysis: The case of China

Abstract

Because the equation for calculating the air quality index (AQI) only considers the most serious pollutants, it remains controversial, and a comprehensive AQI has been a focus of subsequent studies. The present research transformed 22,504,440 pollution characteristics per year into datasets of a Louvain community detection clustering analysis. Unsupervised machine learning was applied to classify 367 cities across China into seven categories. The seven representative cities were selected from seven categories, and the augmented Dickey-Fuller test was used to detect their stationarity. A situation-based composite AQI calculation method was proposed. To verify the method, we clarified the difference between it and the AQI. We also compared its relative error with the total AQI. The relative error values in the seven representative cities were 8.9%, 12.16%, 12.91%, 8.75%, 13.42%, 11.41%, and 11.27%. These values were smaller than the total AQI. The median values of IAQICO,IAQINO2,IAQISO2,IAQIPM2.5,andIAQIPM10 remained relatively stable. Only the changes in IAQIO3 were dramatic. High Thour values are generally encountered in more polluted cities. When there is at most one chief pollutant (CP), then Thour=1. When there are several CPs, Thour>1. The findings of this research provide the public with an intuitive understanding of air pollution and guidance on effective assessments of outdoor air quality.