Predicting Hourly Particulate Matter (PM2.5) Concentrations Using Meteorological Data

Condensations of PM_2.5 have turn into an issue worldwide, specifically in the United States, China and India. Monitoring of PM_2.5 concentrations still in a very early stage and the spatial coverage is limited. These constraints lead to difficulty in assessing the health issues related to PM_2.5 exposure for prior time periods or in certain areas. As PM_2.5 levels are strongly influenced by meteorological conditions, this work employs machine learning models, namely, linear regression (LR), support vector machine (SVM), neural network (ANN) and random forest (RF) to predict PM_2.5 concentrations at hourly time scale, utilizing the readily available meteorological data. The main input variables for the proposed models were hourly measurements of meteorological data such as temperature, wind speed and direction, humidity and pressure. Dataset was collected in five cities in China: Beijing, Guangzhou, Shanghai, Shenyang and Chengdu as a study region and hourly readings for three years were included in this work. Experiments showed that the random forest model achieved higher predictive accuracy than other proposed models in estimating hourly PM_2.5 concentrations with R² value varied between 0.67 and 0.78. This work proposes a promising and affordable approach for robustly predicting PM_2.5 concentrations.