Near real-time predictions of tropical cyclone trajectory and intensity in the northwestern Pacific Ocean using echo state network

Abstract

A recurrent neural network model for predicting the trajectories and intensities of tropical cyclones (TCs) in the northwestern Pacific Ocean basin is described in the present study. By training an echo state network with a meteorological dataset, the recurrent neural network model, named Reservoir Computing for Tropical Cyclone Prediction (RCTCP) was developed to predict four attributes of TCs, i.e., latitude, longitude, maximum sustained wind speed, and minimum sea level pressure. To comprise the dataset, best track data of the TCs occurred in the basin from 1945 to 2017 were obtained from the U.S. Joint Typhoon Warning Center. The CERA-20C and ERA-Interim reanalysis datasets were used along with dynamic time warping to compensate for the unreported pressure information in the best track data. The data for each TC were then attached end-to-end based on the order of occurrence to form a single sequential dataset. The dataset was interspersed with artificial dummy data to strengthen its regularity. After training, the model was able to yield every six hours the likely position and intensity of a TC 6 to 24 h into the future. Validation of the model’s 6-h forecasting produced mean absolute errors for distance, wind speed, and pressure of 32.73 km, 3.84 kn, and 3.12 mbar, respectively. The accuracy of the 6 to 24-h forecasts by the model was comparable or better than climatology and persistence (CLIPER), statistical typhoon intensity forecast (STIFOR), and existing AI-based approaches. The RCTCP thus represents an attractive option to aid in prompt decision making in relevant fields due to its computational efficiency.