Please login first
Analysis of changes in pollutant concentrations levels using a meteorological normalization technique based on a machine learning algorithm
* 1 , 2
1  Istituto Superiore di Sanità
2  INAIL-DIT, Via del Torraccio di Torrenova 7, 00133, Rome, Italy
Academic Editor: Daniele Contini (registering DOI)

There is growing awareness that the development of optimal strategies to prevent health damages associated with the exposure to the atmospheric pollution requires the assessment of the weather influence on the attainment of air quality objectives. Meteorological conditions, in fact, affect the link between emissions and air pollution over multiple scales in time and space so masking the real trends in the observed pollutants concentrations.

An emerging approach to afford the problem consists in developing machine-learning (ML) based ‘meteorological-normalization’ algorithms establishing the relationship between local meteorology and air pollutants surface concentrations.

In this study, a technique of meteorological-normalization, based on a random forest (RF) ML algorithm, is developed to assess changes in the nitrogen oxides (NOx and NO2) and sulfur dioxide (SO2) time series in a rural area affected by anthropic sources of air pollutants.

The RF model was trained on air pollutants and meteorological parameters daily data, acquired over the period 2013-2019. Several variables representing time predictors were added to the training data to capture seasonal and weekly pollutants variations.

Thanks to the interpretability of the RF model, the functional relationships between each input explanatory variable (i.e. meteorological parameters and time variables) and each response variable (the pollutant concentration) of the model can be provided, pointing out the role of local meteorological processes in the observed pollutants concentrations.

Throughout the metadata publicly available, some hypothesis on the potential link between the change points of the normalised time series and the pollutants sources existing in the area are also discussed. Overall, our findings show that the developed approach proves to be a powerful technique to correctly detecting variations in pollutant concentrations discriminating the contribution of meteorology from those of source’s emissions; this represents a crucial information for the implementation of effective strategies to prevent health impact of air pollution.

Keywords: air pollution, machine learning, meteorological normalization, random forest, change points analysis