Previsão de trajectória de ciclones tropicais na Bacia do Sudoeste do Oceano Índico baseado em modelos de aprendizado profundo

Abstract

Tropical cyclones (TCs) are intense, synoptic-scale storm systems that form over tropical oceans and are linked to catastrophic meteorological hazards, including extreme winds, torrential rains, and storm surges, as well as economic losses. Generating rapid and accurate forecasts of their tracks is crucial for developing prevention and mitigation strategies. The movement of TCs is governed by complex, nonlinear interactions between atmospheric variables, oceanic conditions, and broader circulation flows, making their track prediction an intrinsic challenge for both dynamic and statistical models. This study uses Deep Learning (DL)—a subtype of machine learning that employs hierarchical neural networks to autonomously extract spatiotemporal patterns from complex datasets—to develop predictive models of TC trajectories over the next 6–72 hours over the Southwest Indian Ocean Basin (SOI). TC movements over the SOI from 1993 to 2023 were analyzed using data extracted from the International Best Track Archive for Climate Stewardship (IBTrACS) database. From the input data, 11 relevant variables that best explain the changes in TC latitude and longitude over the next 6–72 hours were selected using the Random Forest Regressor algorithm. The selection was based on the ranked importance indices of the variables. Five recurrent neural network (RNN)-based models were developed and trained to predict CT trajectories using the last 24 hours of meteorological data and observed trajectory (of the selected variables) as input. Post-training evaluation, using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and trajectory distance error, revealed good performance in simulating CT trajectories in the SOI using the Conv-GRU, Stacked-LSTM, and Encoder-Decoder models, with average distance errors ranging from 19 to 160 km for the 6–24-hour predictions on the test data. An ensemble learning approach, combining predictions from the best-performing models through stacked ensemble learning, further reduced trajectory deviations in the first few hours. Comparison with operational forecasts from the Regional Tropical Cyclone Center (La Réunion) demonstrated good alignment, particularly in short-term forecasts (<24h). These results highlight the potential of DL-based models to complement existing numerical and statistical methods, offering a more cost-effective and data-driven solution for operational forecasting of tropical cyclone tracks in the SOI Basin(TRADUÇÃO NOSSA)