Air pollution prediction using machine learning

Prediction of air pollution with particulate matter is a critically important task for developing effective strategies to improve the environmental situation. Despite the large number of predictive machine learning models, insufficient attention has been paid to investigating the effectiveness of pollution prediction in different ranges of microparticle concentrations. The paper proposes models for forecasting atmospheric pollution with particulate matter up to 2.5 microns in size (PM2.5) based on the Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), and Random Forest algorithms taking into account meteorological and spatio-temporal data obtained by the developed air quality monitoring system. Particular attention was focused on studying the dependence of forecasting accuracy on the level of atmospheric pollution. It was found that the proposed models successfully predict the PM2.5 content in the air at low and medium levels of pollution but underestimate the predicted values as their concentration increases. Based on the analysis of the concentration dependences of absolute and relative errors, it was found that the Random Forest method demonstrates the highest prediction accuracy in a wide range of the PM2.5 concentration with a relative error of 6–9 % despite deviations for some peak values. Models based on the XGBoost and LSTM methods are characterized by errors of 9–11 and 11–14 %, respectively. A decrease in forecast accuracy and a significant increase in the variance of predicted values were found with an increase in the concentration of the particulate matter in the air. The LSTM method demonstrates the worst results for high levels of air pollution. The decrease in the effectiveness of predictive models with increasing atmospheric pollution may be due to the small number of records with a high concentration of particulate matter in the dataset and the random appearance of additional pollution sources unrelated to meteorological conditions and spatio-temporal characteristics. An integral assessment of the accuracy of the developed models using the metrics Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination R² confirms the high efficiency of predicting the PM2.5 concentration in the air.