Lightning Warning Prediction with Multi-source Data

In this paper we describe a new methodology for generating real-time lightning warning prediction by using a reliable multi-source data. To do so, it was used two years of data covering 50km radius over three regions in Brazil. For 5-minutes intervals, it was evaluated three approaches: a rule-based model that monitors an area of radius greater than the protected area, a machine learning model that considers the amount of lightning that hit small nearby regions, and an integrated approach that combines the two above. The results achieved, on average, about 80% of false alarm ratio, when the model generated an alert but no lightning strikes the area, 14% of failures, opposite to the previous one, had lightning without alert, 1% of the total time operations had to be stopped because of alerts, and 9 minutes of lead time between the generation of the alert and there is a lightning strike. A multi-criteria decision method was used to rank the best method for each location. Rule-based and Integrated models were preferred according to the importance of each criterion for stakeholders. Each methodology has its advantages and they can be extended to other areas according to business needs.