Interpretable deep learning network significantly improves tropical cyclone intensity forecast accuracy

Predicting tropical cyclones (TCs) accurately is crucial for disaster mitigation and public safety. Although the forecasting accuracy of TC tracks has improved substantially in recent decades, progress in the forecasting of TC intensity remains limited. In recent years, deep  learning methods have shown great potential in TC intensity prediction; however, they still face challenges, including limited interpretability, cumbersome feature engineering, and unreliable real-time operational forecasts.

To overcome these limitations, the research team led by Professor Wei Zhong from the College of Advanced Interdisciplinary Studies, National University of Defense Technology, China, proposes a novel TC intensity prediction framework that integrates Kolmogorov–Arnold networks (KANs) with a dynamic predictor pruning optimization module—namely, TCI–KAN. The overall architecture of TCI–KAN is illustrated in Fig. 1. It consists of three modules: the predictor pruning optimization module, the neural network optimization module, and the prediction module. The research results were recently published in Atmospheric and Oceanic Science Letters.

Testing results demonstrate that the predictor pruning optimization module can effectively select 15 high-impact predictors from 317 predictors. TCI–KAN achieves superior accuracy in 6-h intensity forecasts, with a mean absolute error (MAE) of 2.85 kt. TCI–KAN significantly outperforms the referenced best records by 31%, 13%, and 6% in MAE compared to the official operational forecast, single deep-learning models, and hybrid deep learning models, respectively. In addition, TCI–KAN is suitable for different basins and TC categories. It exhibits higher accuracy and lower uncertainty in the eastern Pacific region compared to other regions, but the prediction error and uncertainty escalate with the increasing intensity of TCs.

"This work extends the application of interpretable deep learning networks to TC intensity prediction and significantly improves the forecasting accuracy. It not only provides a novel technical pathway for TC intensity prediction, but also promotes the development of the forecasting paradigm integrating data-driven and physical mechanism based methods," states Professor Wei Zhong, the corresponding author of the paper.

The first author of the paper is Keyun Li, a master’s student at the National University of Defense Technology. This study was supported by the National Natural Science Foundation of China [grant numbers 42075011 and 42192552].