Recurrent Neural Networks for air-quality forecast models in the city of Rio de Janeiro

Resumo

The tropical climate of the metropolitan region of Rio de Janeiro is especially susceptible to air
pollutants such as Ozone and Particulate Matter, which are directly connected to serious cardiopulmonary
illnesses. The goals of the present work were: to explore the local meteorological data to find useful patterns
among the information and to exam the performance of an ensemble model of Recurrent Neural Networks on the
prediction of daily maximum pollutant levels. The analyzed dataset is provided by the Rio de Janeiro local
government and it is composed by hourly-levels for pollutants and meteorological features from eight different
locations. The Spearman correlation test among the variables of different stations showed that adjacent locations
have similar data, with values up to 95% of correlation depending on the examined variable. The experiments
showed that the ensemble model has superior performance to simpler models in 3 out of 4 studied scenarios.