Interpretable deep learning network significantly improves tropical cyclone intensity prediction accuracy

Accurate prediction of tropical cyclones (TCs) is critical for disaster mitigation and public safety. Although the prediction accuracy of TC trajectories has improved significantly in recent decades, progress in predicting TC intensity remains limited. In recent years, deep learning methods have shown great potential in TC intensity prediction. However, they still face challenges such as limited interpretability, unwieldy feature engineering, and unreliable real-time operational predictions.

To overcome these limitations, a research team led by Professor Wei Zhong from the Institute of Advanced Interdisciplinary Research, University of National Defense Technology, China, proposes a new TC strength prediction framework that integrates Kolmogorov-Arnold network (KAN) and dynamic predictor pruning optimization module, or TCI-KAN. It consists of three modules: a predictor pruning optimization module, a neural network optimization module, and a prediction module. Research results recently Atmospheric and Oceanic Science Letters.

Test results show that the predictor pruning optimization module can effectively select the 15 most influential predictors from 317 predictors. TCI-KAN achieves excellent accuracy with a mean absolute error (MAE) of 2.85 kt for 6-hour intensity predictions. TCI-KAN significantly outperforms the highest referenced MAE by 31%, 13%, and 6% compared to official operational forecasts, single deep learning models, and hybrid deep learning models, respectively.

Moreover, TCI-KAN is suitable for different basins and TC categories. Although the eastern Pacific region has higher accuracy and lower uncertainty than other regions, forecast errors and uncertainties increase as the TC intensity increases.

“This study extends the application of interpretable deep learning networks to TC intensity prediction and significantly improves the prediction accuracy. This not only provides a new technological route for TC intensity prediction, but also facilitates the development of a prediction paradigm that integrates data-driven and physical mechanism-based methods,” said Professor Wei Zhong, corresponding author of the paper.

The paper's lead author is Keyun Li, a master's student at the National Defense Technology University.