Deep probabilistic modeling of environment-dependent tropical cyclone intensity evolution using Flow generative models

Abstract

In this study, Flow-based deep generative models are developed to simulate the environment-dependent tropical cyclone (TC) intensity evolution. Leveraging Flow’s unique capability to tractably model complex probability distributions, a conditional randomization test was first implemented to reduce the TC intensity dependency from a large pool of fifteen candidate variables to five significant ones. Then, a conditional Flow model was developed to guide the over-ocean environment-dependent TC intensity evolution, and an enhanced land decay model was developed for over-land evolution. 963 storms from China Meteorological Administration (CMA) dataset during 1980 to 2022 were used to develop and evaluate the Flow-based simulation framework in the Western Northern Pacific (WNP) basin. Overall, the proposed approach showed an excellent fit with historical observations in terms of hindcasting performance, simulated TC key statistics, TC spatial distributions, and the probability distributions of 6-h and 24-h intensity change. Flow-based models also competed favorably over existing stochastic linear regression models or Hidden Markov Models. Finally, a site-specific simulation validation along the China coastline showcased its potential for TC wind disaster assessment.