Performance Analysis Among Predictive Models of Lightning Occurrence Using Artificial Neural Networks and SMOTE

Lightning represent a potential threat to various society activities, such as damage to telecommunication systems and the distribution of electric power, as well as injury or death of humans beings. Predicting the occurrence of lightning can help in making decisions about the actions that must be taken to minimize the risks of this natural phenomenon. In this study, data from air temperature profiles, dew point temperature and historical lightning data were used to obtain two predictive models of lightning occurrence. The models were obtained by using an artificial neural network. The first model was obtained through unbalanced data and the second one with data balanced with Synthetic Minority Over-sampling Technique (SMOTE). The model performance was tested in five different classes of lightning predictions: ABSENCE, LOW, MODERATE, VERY and SEVERE, considering five prediction periods: case 1 (one hour), case 2 (two hours), case 3 (three hours), case 4 (four hours) and case 5 (five hours). It was observed that the use of the Synthetic Minority Over-sampling Technique improved accuracy in the recognition of atmospheric patterns that lead to the incidence of lightning in the five classes used in the five prediction cases.