Predição de descargas atmosféricas utilizando Machine Learning para prevenção de acidentes

Abstract

The occurrence of atmospheric discharges poses risks to the the company operations and workers in open-air activities. Due to this, this paper aimed to cluster lightning data, simulating real- time monitoring of storms for three different target regions. In addition, storms information were used to predict, 15-minutes earlier, the probability of a lightning strikes these areas. Using a multi-source database from ELAT/INPE, different clusterization methods were evaluated in terms of the Calinski Harabasz, Davies Bouldin and Silhouette metrics. Overall, the best one was the MeanShift which cluster the data in 3-5 storms. Number of storms, density and distance were used into a classification machine learning model to generate warning alerts. The Extreme Gradient Boosting and Support Vector Machines achieved the best results in terms of precision and recall, important metrics to evaluate true and false alerts in this context. Both the false alerts, which implies in inactivity of operations and failure rate were equal to or lower than 40%.