Work

Gradient-informed data splitting and model setup for machine learning prediction of storm surge

Year: 2025

Type: article

Abstract: Efficient and accurate prediction of tropical cyclone induced storm surge is critical to coastal disaster emergency response and recovery. Recently, machine learning (ML) methods have been widely applied to storm surge predictions but there is still notable room for improvement. In this paper, we focus on optimizing the generalization and accuracy of storm surge ML models by integrating hydrodynamics in data featuring and model setups, which can be extended to all categories of ML methods like recurrent neural networks and self-attention deep learning architectures rather than a specific model. First, a surge-gradient-informed data splitting strategy is proposed, in which the maximum gradient of surge level is used as a priori indicator to identify the outliers in the adopted tropical cyclones and these outliers are compulsorily included in the training dataset to improve the model generalization. It is shown that, compared with the common random data splitting, the surge-gradient-informed data splitting has a non-negligible effect on reducing the errors in the predicted rapid increases of surge levels. Further, based on shallow-water hydrodynamics in the motion of coastal water, a gradient-target setup of ML models is suggested which takes the temporal gradient of surge level, rather than directly the surge level, as the output variable of ML models. Correspondingly, according to a comparison study, the combination of historical surge gradient and meteorological-oceanographic conditions is recommended for the input features of gradient-target ML models. This gradient-target setup is tested in four categories of ML models, i.e., XGBoost, LSTM, DLinear, and Multi-head MLP. It is shown that, for forecasts with a lead time of 1-2 hours, the gradient-target setup improves the performances of all four ML models to different extents, especially those of XGBoost and LSTM models. For forecasts with a lead time of over 3 hours, the improvement of ML model performances by using the gradient-target setup diminishes with the lead time and the performances of XGBoost and MLP models are even worsen. Nevertheless, the gradient-target models capture the rapid rising and falling of surge heights more accurately than the surge-target models. The surge-gradient informed data splitting and gradient-target model setup proposed in this study provide an alternative view to incorporating physics into ML models for ocean hydrodynamic disasters.

Source: Applied Ocean Research

DOI: 10.1016/j.apor.2025.104873

Authors: Yujia Wang, Dingqi Yang, Linlin Li, Huabin Shi

Institutions: University of Macau, Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Sun Yat-sen University, Sichuan University, State Key Laboratory of Hydraulics and Mountain River Engineering

Language: English

Topic: Tropical and Extratropical Cyclones Research

Subfield: Atmospheric Science

Field: Earth and Planetary Sciences

Domain: Physical Sciences

SDG: Climate action

Open Access status: gold

Estimated APC paid: $2,950