Understanding meteorological impacts on ambient PM2.5 concentrations using random forest models in Beijing
Date
2022
Authors
Brehmer, Collin, author
Carter, Ellison, advisor
Bond, Tami, committee member
Carlson, Kenneth, committee member
Pierce, Jeffrey, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Policymakers and non-governmental organizations have been implementing policies and interventions designed to reduce exposure to hazardous air pollution. Having knowledge of how non-policy related factors (i.e., meteorology) impact air pollution concentrations in a given study area can better inform longitudinal studies of the effects of the policy on air pollution and health. In this study, we apply a random forest machine learning approach to evaluate how meteorological factors including temperature, relative humidity, wind speed, wind direction, and boundary layer height influence daily PM2.5 concentrations in rural Beijing villages during heating months (January and February of 2019 and 2020). Ten-fold cross validation indicated good model performance with an overall r2 of 0.85 for season 1, and 0.93 for season 2. The models were able to identify variables that were the most important for predicting PM2.5 concentrations both field seasons (relative humidity) and variables that had changes in relative importance between seasons (temperature and boundary layer height). Additionally, examination of one and two-way partial dependence plots as well as interactions through Friedman's H-statistic granted insight into how meteorology variables influence PM2.5 concentrations. Findings from this work provide a basis for adjusting for meteorological variability in important indicators of air quality like PM2.5 concentrations in an ongoing real-world policy evaluation of a province-wide ban on household use of coal for space heating in Beijing, which is critical for isolating (to the extent possible) changes in measured pollutant concentrations attributable to the policy.
Description
Rights Access
Subject
machine learning
air pollution
PM2.5