Differentiating Periodic Drivers of Air Quality Changes: A Two‐Step Decomposition Approach Integrating Machine Learning and Wavelet Analysis

Abstract

Air quality time series exhibit significant periodic patterns, which are linked to a diverse array of emission sources and atmospheric processes. To discern and distinguish these periodic drivers, we have devised a two‐step decomposition approach that integrates a machine learning‐based model for weather normalization with Morlet wavelet analysis. This approach was applied to a 7‐year data set encompassing six regulated air pollutants across eight Chinese cities. Our analysis revealed distinct periodicities in weather‐normalized concentrations of primary air pollutants: a dominant annual cycle with periodicity around 365 days, which accounts for over 50% of the variance on average and is primarily driven by recurrent winter heating activities; daily cycles characterized by regular diurnal patterns attributable to combustion sources such as traffic; and periodicities exceeding 512 days that associated with long‐term regulatory policies targeting SO2. Particularly notable was the significant drop in the strength of the annual cycle in northern cities following the implementation of China's clean heating policies in 2017/2018, affirming the success of these initiatives. Additionally, diurnal dispersion and photochemistry, large‐scale atmospheric circulation, and synoptical weather patterns are likely responsible for the observed daily cycle (accounting for over 40% of the variance), annual periodicities, and intra‐monthly variations in the meteorologically driven concentrations, respectively. The varying power of these periodic drivers across time and locations implies the heterogeneity of emission rates and region‐specific climates. This work highlights the efficacy of this decomposition approach in air quality research, warranting attention for its potential value for enhancing our understanding of air pollution dynamics.