A Physics-informed Deep-learning Intensity Prediction Scheme for Tropical Cyclones over the Western North Pacific

Accurate prediction of tropical cyclone (TC) intensity is challenging due to the complex physical processes involved. Here, we introduce a new TC intensity prediction scheme for the western North Pacific (WNP) based on a time-dependent theory of TC intensification, termed the energetically based dynamical system (EBDS) model, together with the use of a long short-term memory (LSTM) neural network. In time-dependent theory, TC intensity change is controlled by both the internal dynamics of the TC system and various environmental factors, expressed as environmental dynamical efficiency. The LSTM neural network is used to predict the environmental dynamical efficiency in the EBDS model trained using best-track TC data and global reanalysis data during 1982–2017. The transfer learning and ensemble methods are used to retrain the scheme using the environmental factors predicted by the Global Forecast System (GFS) of the National Centers for Environmental Prediction during 2017–21. The predicted environmental dynamical efficiency is finally iterated into the EBDS equations to predict TC intensity. The new scheme is evaluated for TC intensity prediction using both reanalysis data and the GFS prediction data. The intensity prediction by the new scheme shows better skill than the official prediction from the China Meteorological Administration (CMA) and those by other state-of-art statistical and dynamical forecast systems, except for the 72-h forecast. Particularly at the longer lead times of 96 h and 120 h, the new scheme has smaller forecast errors, with a more than 30% improvement over the official forecasts.