Prediction of High-ozone Events Using GAM, SMOTE, and Tail Dependence Approaches in Texas (2005–2019)

Abstract

We test three methods for ozone prediction in the El Paso (ELP) and Houston-Galveston-Brazoria (HGB) regions of Texas from 2005–2019: (1) a Generalized Additive Model (GAMs) approach; (2) a GAM approach with the addition of the Synthetic Minority Over-sampling TEchnique (SMOTE) and (3) a tail dependence modeling approach based in extreme value theory (EVT). We also compare the feature selection capabilities of the tail dependence approach to other feature selection methods. We find that the GAM+SMOTE model outperformed the GAM-only model when predicting ozone values for the root mean square error metric, particularly with regard to the above-threshold ozone values, which may be of particularly useful for extreme ozone event prediction. In addition, we find that the improvement of above-threshold MDA8 O3 prediction for the GAM+SMOTE method tends to come at the cost of below-threshold prediction, which is particularly important if MDA8 O3 trends are of interest. We also find that the tail dependence approach is capable of predicting extreme ozone events, but algorithmic stability and configuration complexity can make this approach difficult to operationalize on a broad scale and that the selection of the threshold needs to be carefully considered. Finally, the feature selection via the tail dependence method performs comparably to other forms of machine learning-based feature selection and we find that there are multiple parameter sets that can predict MDA8 O3 with equal success.