Forecasting hourly particulate matter concentrations based on the advanced multivariate methods

Abstract

In this study, several multivariate methods were used for forecasting hourly PM10 concentrations at four locations based on SO2 and meteorological data from the previous period. According to the results, boosted decision trees and multi-layer perceptrons yielded the best predictions. The forecasting performances were similar for all examined locations, despite the additional PM10 spatio-temporal analysis showed that the sites were affected by different emission sources, topographic and microclimatic conditions. The best prediction of PM10 concentrations was obtained for industrial sites, probably due to the simplicity and regularity of dominant pollutant emissions on a daily basis. Conversely, somewhat weaker forecast accuracy was achieved at urban canyon avenue, which can be attributed to the specific urban morphology and most diverse emission sources. In conclusion to this, the integration of advanced multivariate methods in air quality forecasting systems could enhance accuracy and provide the basis for efficient decision-making in environmental regulatory management.