Deep Learning Approaches in Predicting Tropical Cyclone Tracks: an Analysis Focused on the Northwest Pacific Region
Peng Hao,Yaqi Zhao,Shuang Li,Jinbao Song,Yu Gao
DOI: https://doi.org/10.1016/j.ocemod.2024.102444
IF: 3.293
2024-01-01
Ocean Modelling
Abstract:In this study, we conducted a comprehensive and integrated test of tropical cyclone track prediction using deep learning technologies, aiming to enhance the efficiency and accuracy of the prediction methods. We employed the Best Track dataset from the China Meteorological Administration's Tropical Cyclone Data Center, which covers the Northwest Pacific region from 1949 to 2023. This dataset provides comprehensive coverage, encompassing critical tropical cyclone details like time, latitude, longitude, and wind speed. Our focus was on evaluating and comparing different deep learning models, including Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), and Gated Recurrent Units (GRU), for their effectiveness in handling complex time series data. Through detailed analysis of various model configurations, including factors such as input-output lengths, hidden size, the number of layers, the implementation of bi-directional networks, and attention mechanisms, we discovered that LSTM and GRU models significantly outperform traditional RNN models in dealing with long-term dependencies and enhancing prediction accuracy. Moreover, the LSTM model, used to forecast key tropical cyclones during the 2023 Pacific tropical cyclone season, achieved mean errors of 21.84 km, 37.56 km, and 26.12 km for Typhoons Mawar, Doksuri, and Saola, respectively. This method also demonstrated high efficiency in rapid response to extreme weather changes, processing each tropical cyclone's forecast in just about 8 s. The results not only illustrate the practical utility of deep learning in tropical cyclone track prediction but also provide new, effective tools for disaster prevention and mitigation efforts.