Defense
Student: Gabriel Elias Mandanda
Program: Meteorology
Title: "Predictability of Thunderstorms in South America"
Advisor: Profa. Dra. Maria Assunção Faus da Silva Dias
Judging Comitee:
- Profa. Dra. Maria Assunção Faus da Silva Dias - Presidente e Orientadora - IAG
- Profa. Dra. Rachel Ifanger Albrecht - IAG
- Dr. Ernani de Lima Nascimento - CPTEC/INPE (por videoconferência)
- Prof. Dr. Antonio José Homsi Goulart - ESALQ/USP (por videoconferência)
- Prof. Dr. Vagner Anabor - UFSM (por videoconferência)
Abstract:
South America is a region prone to thunderstorms, making the early prediction of their occurrence essential to reduce their impacts. This research aims to forecast lightning discharges based on atmospheric parameters, targeting spatial and temporal prediction horizons of up to 24 hours. For this purpose, data from the GLM (Geostationary Lightning Mapper) onboard the GOES-16 (Geostationary Operational Environmental Satellite), ERA5 reanalysis data (Fifth Generation of ECMWF Atmospheric Reanalyses of the Global Climate), and forecasts from the GFS (Global Forecast System) were used. The methodology was based on the analysis of the diurnal and annual cycle of lightning discharges, using the K-means algorithm to identify homogeneous regions in South America. For each region, probabilistic and regression-based artificial intelligence models were trained to forecast accumulated lightning discharges over 6, 12, and 24 hours, using atmospheric parameters derived from ERA5 reanalyses and GLM lightning data. The models were trained, validated, and tested on independent periods and subsequently evaluated using GFS forecasts to ensure their operational applicability. The most relevant parameters for forecasting lightning discharges in South America include the 2-meter air temperature (T2M), identified as the main predictor of electrical activity in most tropical regions. CAPE (Convective Available Potential Energy) showed the greatest contribution in regions with higher lightning incidence in South America (La Plata Basin, northern Colombia, and Lake Maracaibo), also standing out in all other regions, highlighting its central importance in forecasting electrical activity. Additionally, the TT (Total Totals Index) also performed well. Tests with ERA5 indicated good performance of the probabilistic models across all forecast periods (6, 12, and 24 hours), with emphasis on the 24-hour forecasts, which presented a mean POD (Probability of Detection) above 70%. The regression model exhibited NMAE (Normalized Mean Absolute Error) below 40% in most regions and stations, whereas the 6-hour forecasts showed high false alarm values (FAR, False Alarm Ratio), which decrease with increasing forecast horizon. In the operational evaluation with the GFS, the POD reached mean values above 75% in the 24 hour forecasts, while the 6-hour forecasts presented the highest errors, decreasing with longer forecast horizons. The regression model outperformed the product of CAPE and Precipitation (CAPE×P) in representing lightning discharge patterns across most regions and seasons of the year, demonstrating robustness for operational application.
Keywords: Lightning forecasting; Artificial intelligence; South America.