Predlog modela za predviđanje koncentracije suspendovanih (PM2.5) čestica u vazduhu

Abstract

Increasing number of studies indicate the negative influence of Particulate Matter on human health. One of the ways to avoid their negative consequences is a timely prediction of airborne PM2.5 concentrations. Knowing hourly PM2.5 concentrations, people could organize their daily activities to reduce exposure to intensive pollution. With the goal to train an optimal predictive model, the predictive performances of three machine learning algorithms were analysed: „Random forest“, „XGBoost“, and „Light gradient boosting machine“. Using mentioned regression algorithms in combination with meteorological and chronological data, the models were trained to predict hourly airborne PM2.5 concentrations with relatively high accuracy. The data about airborne PM2.5 concentrations were collected using the laser sensor in the city of Kragujevac, Serbia. The trained models were evaluated using the coefficient of determination (R2), mean absolute error (MAE), and rootmean-square error (RMSE).