Forecast of daily PM2.5 concentrations applying artificial neural networks and Holt–Winters models

Fine particulate matter (PM2.5) has been considered one of the most harmful atmospheric pollutants to the health. PM2.5 has as its main origin vehicular emissions, a characteristic source in megacities. In order to predict pollution episodes in different areas (rural, industrial, and urban), two models were applied, Holt–Winters (HW) and artificial neural network (ANN), using PM2.5 concentration time series. PM2.5 samples were collected using Hi-Vol samplers during a period of 24 h, every 6 days, from January 2011 to December 2013, in Rio de Janeiro, Brazil. Meteorological data was also obtained for use in the models. The PM2.5 dataset was the longest obtained for this megacity and the Holt–Winters (HW) model was used, for the first time, to predict air quality. The results of the PM2.5 data series showed daily concentrations ranging from 1 to 65 μg m−3. The root mean square error (RMSE) was calculated for each model for the three sites. The HW model best explained the simulation of PM2.5 in the industrial area, since it presented the lowest RMSE (5.8 to 14.9 μg m−3). The ANN was the most appropriate model for urban and rural areas with RMSE between 4.2 to 9.3 μg m−3. Overall, both forecast models proved accurate enough to be considered useful tools for air quality management and can be applied in other world regions.