Short-term lightning prediction in the Amazon region using ground-based weather station data and machine learning techniques

Abstract

Lightning occurrence is a real threat to human beings and animals over the Amazon region. Lightning is also responsible for economic losses in electric, telecommunication, and other sectors, but its prediction remains a challenging task. Lightning prediction can contribute to minimizing the risks and losses caused by this natural phenomenon. In this work, we have used data from ground-based weather stations, including air temperature, humidity, pressure, and wind speed to predict lightning occurrence within one hour. Forecasts are made for a region up to about 30 km from each of the nine capital cities of the nine states of the Legal Amazon region in Brazil. We use GLD360 data to validate predictions and train the machine learning algorithm. We used a database of 6 years of observation (2015 to 2020) to test and validate the prediction models. The model for Belém - Pará showed the highest F1-Score, 0.34, and the highest Area Under the ROC Curve, 0.836. Overall, the accuracy of the models for each city is higher than 71%. This approach can be used in regions of the Amazon in which only ground-based weather station data is available.