2025-04-17 中国科学院（CAS）
＜関連情報＞
- https://english.cas.cn/newsroom/research_news/earth/202504/t20250417_1041473.shtml
- https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2024JH000507
ハイブリッド風速場を用いた台風による高潮予測のための機械学習技術 Machine Learning Techniques for Predicting Typhoon-Induced Storm Surge Using a Hybrid Wind Field
Changyu Su, Bishnupriya Sahoo, Miaohua Mao, Meng Xia
Journal of Geophysical Research: Machine Learning and Computation Published: 10 April 2025
DOI:https://doi.org/10.1029/2024JH000507
Abstract
Accurate and timely storm surge prediction is critical information in coastal zone management and risk reduction strategies. The Bohai Sea, a semi-enclosed bay in the Northwest Pacific that used to be less prone to typhoon disasters, has been witnessing a paradigm shift in typhoon activities in the recent past. Since there have been limited typhoon-induced storm surges in the Bohai Sea, an innovative prediction system is warranted to address frequent and intense typhoon-induced impacts. Four Machine Learning (ML) models (Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), CNN-LSTM, and ConvLSTM) were built to predict storm surges and significantly improve prediction when combined with a three-dimensional Finite Volume Community Ocean Model (FVCOM), that is, FVCOM-ML. In this study, the FVCOM-ML model was driven by a hybrid wind field that superimposed the Holland wind and the reanalysis wind field. The ML models were trained via Advanced Circulation Model simulations to compensate for the limited in-situ observations. The prediction performances were analyzed for both spatial (e.g., single and multiple sites) and temporal (e.g., single and multiple steps) scale variability. ML is trained to overcome the residual error of the FVCOM, effectively reducing the inherent uncertainty of traditional methods. FVCOM-ML offers a significant advantage over standalone FVCOM or ML while better incorporating realistic physical constraints and improving the accuracy of storm surge forecasts.
Key Points
- Hybrid winds generated by superimposing reanalysis winds, typhoon tracks, and the Holland model better characterize the typhoon wind field
- The integrated models (CNN-LSTM and ConvLSTM) predict storm surges with 18% greater accuracy than the individual models (Long Short-Term Memory and Convolutional Neural Networks)
- FVCOM-ML improves storm surge prediction accuracy and reduces uncertainty by addressing data heterogeneity and uncertainty
Plain Language Summary
Accurate storm surge prediction is essential in preparedness measures and nullifying causalities. Numerical models with mathematical equations and physics-based formulae have complexities and limitations when dealing with real-world problems. Machine Learning (ML) models offer a potential solution but require accurate wind data as a driver. Hence, we improved the quality of the typhoon wind by using a hybrid wind composed of two wind data. We combined ML, namely, LSTM and CNN into combined models (CNN-LSTM and ConvLSTM), which can learn the factors influencing the storm surges with space and time by performing operations on time and space scale. The composite models show excellent performance for predictions. Although ML methods are more accurate in predicting storm surges, the lack of physical mechanisms results in poor interpretability. Therefore, we introduced an approach that combines a three-dimensional Finite Volume Community Ocean Model (FVCOM) and ML model, that is, FVCOM-ML, to improve storm surge predictions. ML is used to predict and optimize the residual error of FVCOM. The framework exploits the complementary strengths of both models: the expertise in physical mechanisms based on FVCOM and the fact that ML makes nonlinear modeling easier and computationally faster.
