Forecasting Ozone and PM2.5 Pollution Potentials Using Machine Learning Algorithms: A Case Study in Chengdu

WANG Xinlu1, HUANG Ran1,†, ZHANG Wenxian1, LÜ Baolei2, DU Yunsong3, ZHANG Wei3, LI Bolan3, HU Yongtao4

1. Hangzhou Aima Technologies, Hangzhou 311121; 2. Huayun Sounding Meteorological Technology Company, Ltd., Beijing 102299; 3. Sichuan Bio-environmental Monitoring Center, Chengdu 610091; 4. School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA 30332; † Corresponding author, E-mail: ranhuang2019@163.com

Abstract Potential forecast models have been developed for air pollution of summertime (Apr.–aug.) ozone and wintertime (Nov.–feb.) PM2.5 in Chengdu using the multiple linear regression (MLR), back-propagation (BP) neural network (NN) and random forest (RF) algorithms. The key predicting factors for each of the models are selected from various potential factors that may impact the spatiotemporal distribution of pollutions. The models are trained and established with 2016–2018 datasets and evaluated with a data-withheld method and further with independent 2019 dataset. The results show that the MLR, NN and RF models are all capable to accurately predict O3 and PM2.5 pollution potentials in short lead-time (1–3 days) in Chengdu. The models are also found having quite stable performances in medium- and long-term (7–15 days lead time) forecasts. Among the three models, the MLR model performs the best in prediction of O3, while RF model performs the best for PM2.5. Key words multiple linear regression; BP neural network; random forest; medium- and long-term air pollution potential forecast

环境空气质量的好坏对公众健康有着显著影响, 不论是极端重污染事件还是长期暴露于低浓度空气污染环境中, 均会直接增加人体心血管和呼吸[1–2]系统等多种疾病的发病率 。近年来, 我国大多数城市的空气质量持续改善, 尤其是秋冬季细颗粒物(PM2.5, 空气动力学直径小于或等于2.5 μm的气