Forecasting tropical cyclone tracks in the northwestern Pacific based on a deep-learning model

Tropical cyclones (TCs) are one of the most severe meteorological disasters, making rapid and accurate track forecasts crucial for disaster prevention and mitigation. Because TC tracks are affected by various factors (the steering flow, the thermal structure of the underlying surface, and the atmospheric circulation), their trajectories present highly complex nonlinear behavior. Deep learning has many advantages in simulating nonlinear systems. In this paper, based on deep-learning technology, we explore the movement of TCs in the northwestern Pacific from 1979 to 2021, divided into training (1979–2014), validation (2015–2018), and test sets (2019–2021), and we create 6–72 h TC track forecasts. Only historical trajectory data are used as input for evaluating the forecasts of the following three recurrent neural networks utilized: recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models. The GRU approach performed best; to further improve forecast accuracy, a model combining GRU and a convolutional neural network (CNN) called GRU_CNN is proposed to capture the characteristics that vary with time. By adding reanalysis data of the steering flow, sea surface temperatures, and geopotential height around the cyclone, we can extract sufficient information on the historical trajectory features and three-dimensional spatial features. The results show that GRU_CNN outperforms other deep-learning models without CNN layers. Furthermore, by analyzing three additional environmental factors through control experiments, it can be concluded that the historical steering flow of TCs plays a key role, especially for short-term predictions within 24 h, while sea surface temperatures and geopotential height can gradually improve the 24–72 h forecast. The average distance errors at 6 and 12 h are 17.22 and 43.90 km, respectively. Compared with the 6 and 12 h forecast results (27.57 and 59.09 km) of the Central Meteorological Observatory, the model proposed herein is suitable for short-term forecasting of TC tracks.