Previsão de MP10 através de redes neurais: estudos de caso no Estado de São Paulo

Abstract

Air pollution is an ever-present and increasingly worrying health problem. and well-being of the population, in particular, the Mp10 particulate matter has become a target of studies because of the problems that their exposure in large amounts or prolonged time can generate in public health. To monitor in order to control and understand more about origins and behavior, a multilayer neural network of Perceptrons (MLP) with two layers of 40 neurons and 4 analysis parameters to predict the concentration of the pollutant particulate matter (PM). This configuration was used in order to increase the efficiency and accuracy of the data. The four chosen parameters are related to meteorology, as they have a strong influence on the dispersion of pollutants, being these: atmospheric pressure, wind speed, relative humidity and temperature environment. The neural network was trained from the PM data between January 1, 2017 until January 1, 2022, where the program used data from previous days as a basis to predict the values ​​of the following days. The regions of Parque Dom Pedro II, Guarulhos, Santos and Jaú were analyzed with data of the air quality stations collected by the Environmental Company of the State of São Paulo (CETESB), available through the QUALAR platform. from the results generated, an average percentage error between the values ​​provided and the values predicted by the neural network of, respectively, 30.93%, 27.77%, 25.12% and 24.69% for the MP10, demonstrating results consistent and close to the real value, with a greater precision for less populated areas, as there are fewer human interactions affecting the concentration of particulate matter. The same procedure was performed for MP2.5 in the regions of Parque Dom Pedro II and Guarulhos, obtaining different results and not very conclusive, however, equally accurate in relation to the average percentage error of 30.66% and 32.94% respectively, showing that MP2.5 does not have much relationship with meteorological data as well as the MP10. In general, the results obtained were satisfactory for the precision and the information about the neural networks and the action cycles of the PM, showing that MLPs are reliable in predicting MP values ​​in general for values daily.