ABSTRACT
Using statistical models, the average hourly ozone (O3) concentration was predicted from seven meteorological variables (Pearson correlation coefficient, R = 0.87–0.90), with solar radiation and temperature being the most important predictors. This can serve to predict O3 for cities with real time meteorological data but no pollutant sensing capability. Incorporating other pollutants (PM2.5, SO2, and CO) into the models did not significantly improve O3 prediction (R = 0.91–0.94). Predictions were also made for PM2.5, but results could not reflect its peaks and outliers resulting from local sources. Here we make a comparative analysis of three different statistical predictor models: (1) Multiple Linear Regression (MLR), (2) Support Vector Regression (SVR), and (3) Artificial Neuronal Networks (ANNs) to forecast hourly O3 and PM2.5 concentrations in a mid-sized Andean city (Manizales, Colombia). The study also analyzes the effect of using different sets of predictor variables: (1) Spearman coefficients higher than ± 0.3, (2) variables with loadings higher than ± 0.3 from a principal component analysis (PCA), (3) only meteorological variables, and (4) all available variables. In terms of the O3 forecast, the best model was obtained using ANNs with all the available variables as predictors. The methodology could serve other researchers for implementing statistical forecasting models in their regions with limited pollutant information.