Transformer-based tropical cyclone track and intensity forecasting

Abstract

Tropical cyclone (TC) is one of the most destructive natural disasters, and hence it is very crucial to predict attributes of TC accurately. In this paper, a transformer-based method for short-term prediction of the track and intensity of TCs, including the central latitude, the central longitude, the minimum sea level pressure and the maximum sustained wind speed, in the Northwest Pacific Basin is proposed. The data set used in this study ranging from 1980 to 2021 was released by China Meteorological Administration (CMA). A data preprocessing step including feature augmentation is adopted, which can make full use of the historical information of TCs. The transformer networks can model the complex dynamic relationships between the time series data and dig the relationship between different attributes based on the multi-head self-attention mechanism. The transformer-based method proposed in this study exhibits excellent performance in the prediction task. Meanwhile, compared to the traditional recurrent neural network models including gated recurrent units model and long short-term memory model, the transformer model has a better performance in predicting all attributes. This method provides a practical guidance of using advanced AI-based approaches to predict TC attributes.