Random Forests Models of Particulate Matter PM10: A Case Study

Air pollution and high concentrations of fine particulate matter (PM) are extremely harmful to human health, ecosystems and the climate in many countries around the world. To solve this type of air quality problems, a huge number of studies has been devoted. It is important to note that pollution sources have a local character for each urban region and depend on a large number of factors such as meteorological, manufacturing, domestic, transport and other. In the field of mathematical modeling and processing of accumulated measured data for PM with aerodynamic size up to 10 microns (PM10), along with classical statistical methods, powerful, flexible data mining techniques and approaches such as neural networks, fuzzy logic, regression trees, multivariate adaptive regression splines, random forest (RF), and more are adapted. This study explores the possibilities of the random forest method for modeling the concentrations of PM10 in Blagoevgrad, Bulgaria. Average daily data over 9 years (2009-2018) and a large number of input variables as predictors are used - one lagged PM10, meteorological and temporal. The forecasting process is performed by a multi-step procedure repeated five times with 3-day horizon ahead using data not included in the construction of the model. The constructed models show high-performance both in fitting and forecasting the measured data.