Abstract

Two state-of-the-art models (deterministic: Weather Research and Forecast model with Chemistry (WRF-Chem) and statistic: Artificial Neural Networks: (ANN)) are implemented to predict the ground level ozone concentration in Sao Paulo (SP), Brazil. Two domains are set up for WRF-Chem simulations: a... Show moreTwo state-of-the-art models (deterministic: Weather Research and Forecast model with Chemistry (WRF-Chem) and statistic: Artificial Neural Networks: (ANN)) are implemented to predict the ground level ozone concentration in Sao Paulo (SP), Brazil. Two domains are set up for WRF-Chem simulations: a coarse domain (with 50 km horizontal resolution) including whole South America (D1) and a nested domain (with horizontal resolution of 10 km) including South Eastern Brazil (D2). To evaluate the spatial distribution of the chemical species, model results are compared to the Measurements of Pollution in The Troposphere (MOPTIT) data, showing that the model satisfactorily predicts the CO concentrations in both Dl and D2. The model also reproduces the measurements made at three air quality monitoring stations in SP with the correlation coefficients of 0.74, 0.70, and 0.77 for O-3 and 0.51, 0.48, and 0.57 for NOx. The input selection for ANN model is carried out using Forward Selection (FS) method. FS-ANN is then trained and validated using the data from two air quality monitoring stations, showing correlation coefficients of 0.84 and 0.75 for daily mean and 0.64 and 0.67 for daily peak ozone during the test stage. Then, both WRF-Chem and FS-ANN are deployed to forecast the daily mean and peak concentrations of ozone in two stations during 5-20 August 2012. Results show that WRF-Chem preforms better in predicting mean and peak ozone concentrations as well as in conducting mechanistic and sensitivity analysis. FS-ANN is only advantageous in predicting mean daily ozone concentrations considering its significantly lower computational costs and ease of development and implementation, compared to that of WRF-Chem. (C) 2016 Elsevier Ltd. All rights reserved. Show less