The performance of artificial neural networks for modeling daily concentrations of particulate matter from meteorological data.

The use of techniques based on artificial intelligence and machine learning for the simulation of many processes is becoming increasingly important in environmental sciences, with applications in the study of time series of atmospheric properties, such as pollution levels. The present work aimed to evaluate the efficiency of a model based on Artificial Neural Networks (ANN) in the simulation PM10 from meteorological data observed between 2018 and 2019 in Guaíba, southern Brazil, thus also having an estimate of the influence of atmospheric conditions on local air pollution. For this purpose, meteorological and PM10 data obtained from the stations Parque 35, sustained by Celulose Riograndense (CMPC), and A-801, sustained by the National Institute of Meteorology (INMET), were used. The ANN used for the simulation was of the Multilayer Perceptron type, trained by the backpropagation algorithm with cross-validation. The results obtained indicate that the simulation was satisfactory with a Nash-Sutcliffe index (NSE) of 0.64, a linear correlation coefficient (R) of 0.81, a relative error (Er) of 26% and a root mean square error (RMSE) of 7.40 µg/m3. Thus, even with some difficulty in estimating extreme concentrations, the model was suitable for the largest range observed, of 10 µg/m3 to 50 µg/m3. For this dataset, the model proved to be an useful assessment tool and has the potential to be applied operationally to contribute to the monitoring and control of air quality levels both in the study area and in other regions of Brazil and the world.