Effects of meteorology on PM10 concentrations: a comparative assessment of machine learning methods

Administrative decisions regarding the application of measures to address air quality issues have to rely both on present observation and future predictions of the concentration of various pollutants. Since PM10 is one of the most critical pollutants, the ability to provide accurate forecasts for its concentration, when required, is crucial in order to enforce the necessary measures at the right time. Together with the pattern of emission sources which is present in a geographical area, meteorological conditions can significantly affect the concentration of pollutants in air, since they can favour the dispersion or, on the other hand, the build-up of those compounds. It is possible then to predict (at least partially) the concentration of PM10 in air using meteorological variables as predictors. In fact, various statistical models have been proposed for accomplishing similar tasks on a number of geographical regions and urban areas, with varying results. The set of meteorological variables that have been considered in those cases included various predictors, measured both in the day of interest and in the previous ones. Sometimes also some non-meteorological descriptors (e.g. time-related variables) that are grossly related to the variation of the emission patterns have been considered as input variables for those models. In this work an analysis of the relationship between meteorology-related variables and PM10 concentration levels in the capitals of the provinces of Emilia-Romagna has been performed in order to understand how the meteorological conditions affect PM10 concentration. Then the considered meteorological variables have been input as predictors to statistical regression models based on machine learning in order to obtain predictions for the daily mean value of PM10 concentration.