Evaluating ML models for lightning forecasting in Brazil

Instruments for monitoring severe meteorological phenomena (such as lightning, flooding and landslides) can be used to assist in decision-making by state agencies, in an attempt to mitigate their possible harmful effects. These phenomena usually occur suddenly on a short-term duration, under a limited region, imposing difficulties in being predicted by regular weather forecast models, requiring specific prediction systems. Very short-term weather forecasting systems, on order of a few hours, known as nowcasting, can include numerical models of physical phenomena and machine learning algorithms. This work presents a system for forecasting the incidence of lightning, a common phenomenon in electrically active storms, through the application and evaluation of two machine learning models, an Artificial Neural Network and a Random Forest model, which were able to detect the occurrence of atmospheric electrical discharges from the automatic recognition of patterns obtained from the data generated by the numerical weather forecasts. The Random Forest model presented the best results when trained with the set that includes the ten best correlated variables, reaching 99.77% of accuracy for the case study performed.