Multi-horizon prediction of tropical cyclone intensity and its interpretability with temporal fusion transformer

Abstract

Predicting tropical cyclone (TC) intensity is challenging, involving numerous variables and uncertainty, especially for TC with rapid intensification (RI). One of the frequently used operational methods for such a case relies on statistical–dynamical models subjected to several limitations stemming from linear regression approximation to a complex TC system. This study proposes a new approach using a Temporal Fusion Transformer (TFT) to overcome the limitations attributed to the conventional models. Besides improving accuracy, TFT is also known for its result interpretability, as opposed to the general perception that deep learning is purely a black-box model. Furthermore, the probabilistic nature of TFT facilitates prediction uncertainty quantification, an important feature and advancement to the standard deterministic prediction. We train and evaluate our model using the Western North Pacific basin TC observation and reanalysis datasets from 1996 to 2021. The results show that the prediction of TC intensity by TFT reduces the error of the conventional model by approximately 12% on average for all forecast horizons of up to 72 h, along with estimated uncertainty bands. A higher rate of 14% error reduction is attained specifically for TCs undergoing RI, an intractable phenomenon for the traditional modeling procedure.