Análise e previsão de materiais particulados em Minas Gerais utilizando modelos estatísticos, redes neurais e abordagens híbridas

Abstract

The human impact on the planet is evident, and the need to mitigate these effects is becoming increasingly urgent. Atmospheric pollution, for example, is intrinsically linked to the environment and human health. Particulate matter with diameters smaller than 10 and 2.5 micrometers, called PM10 and PM2,5, respectively, can severely impact human health due to their ability to penetrate the respiratory system. This study focuses on predicting the concentrations of PM10 and PM2,5 at monitoring stations in the Brazilian state of Minas Gerais. In the pursuit of better predictive performance, hybrid models were developed that combine parametric, non-parametric, and neural network approaches for particulate matter forecasting. The performance of these models is evaluated across different forecasting horizons. The results indicate that the proposed hybrid models exhibit competitive performance at various time horizons for both pollutants when compared to the use of the models in isolation. These findings highlight the importance of combining different modeling approaches, leveraging the strengths of each method to capture complex patterns and improve the accuracy of predictions. Furthermore, the integration of parametric, non-parametric, and neural network techniques allows for a more robust and flexible model, capable of adapting to different temporal patterns and data characteristics. The methods and insights obtained in this study can be applied to research on other pollutants in different locations, thus contributing to a broader understanding of air quality in various contexts and supporting more effective public policies in air pollution management.