Severe Convective Weather Forecast Using Machine Learning Models

This work developed models, based on machine learning, for severe convective weather forecasts characterized by remotely sensed atmospheric discharge (AD) in the approaching landing region of airports in the vicinity of São Paulo. In the training and testing of machine learning algorithms, classical thermodynamic indices (input), derived from the atmospheric profiles of the Marte-São Paulo soundings, and ADs were used to characterize the convective severe weather (output), considering the period 2001–2017. The statistical distribution defined the locations, times, and severity of the convective events. The POD, 1-FAR, BIAS, kappa, and f-measure statistics were used to evaluate the 5-h prediction of convective event detection (and in parentheses for whether it is severe when the occurrence of lightning is greater than or equal to 1000), yielding values of 0.91 (0.85), 0.95 (0.94), 0.92 (0.89), 0.74 (0.77), and 0.88 (0.95), respectively. The results of applying the model to 30 days (hindcast), show that it is effective since it hit 96.7% of occurrence and 86.7% if they are severe. The detection errors of the model are presented and discussed.