Abstract
Waves of large size can damage offshore infrastructures and affect marine facilities. In coastal engineering, reliable estimates of the most extreme wave height expected during a structure's design lifetime are essential for safe and economical design. Yet direct measurement of extreme wave events is inherently limited — the rarest, most destructive waves are, by definition, those least represented in observational records.
This study, published in the Journal of Earth System Science (Vol. 132, Article 51, 2023), addresses this challenge by combining machine learning (ML) wave prediction with Generalized Extreme Value (GEV) statistical theory. The research employs multiple ML techniques — linear regression, artificial neural networks (ANN), and support vector machines (SVM) — to predict significant wave height from meteorological and oceanographic input parameters. The study found that support vector machines delivered the highest prediction accuracy among the models tested.
The predicted wave height time series is then fed into a GEV framework to estimate return period wave heights — the 50-year, 100-year, and design-life extreme wave heights that offshore structures must be engineered to withstand. This two-stage approach leverages the pattern recognition strength of ML for prediction while maintaining the statistical rigour of extreme value theory for design applications.
For consulting practice, this research is directly applicable to offshore platform design, coastal protection planning, and port infrastructure development. Traditional wave prediction relies on numerical wave models (SWAN, WAVEWATCH III) driven by wind field inputs, which are computationally intensive and sensitive to forcing data quality. ML models offer a fast, complementary prediction pathway that can be trained on historical buoy or reanalysis data and deployed for rapid screening of design wave conditions at new sites.
The combination with GEV theory is particularly valuable: rather than treating ML predictions as point estimates, the statistical framework provides confidence intervals and exceedance probabilities — the language that structural design codes require.
Key Findings
- Support vector machines (SVM) provided the highest prediction accuracy for significant wave height among the ML models tested.
- ML predictions coupled with GEV theory produce return period wave height estimates suitable for structural design.
- The two-stage approach combines ML pattern recognition with statistical extreme value rigour for offshore design applications.
Methodological Approach
Multiple ML models (linear regression, ANN, SVM) trained to predict significant wave height, with outputs fed into Generalized Extreme Value (GEV) statistical framework for return period estimation.
Implications for Hydraulic Practice
Provides offshore and coastal engineers with a rapid, ML-driven wave prediction tool coupled with extreme value statistics — enabling faster screening of design wave conditions at prospective offshore sites.
Related Insights & Publications
Interactive: Low-Carbon Innovation Building (LCIB) Thermodynamic Modeling and Digital Twins
Explore a live, browser-based thermodynamic simulator that models building energy performance, thermal comfort, and CO₂ footprint in real time. Built for architects, sustainability engineers, and building managers evaluating low-carbon design strategies.
Estimating pier scour depth under combined waves and current using boosting machine-learning models
Bridge pier scour under combined wave-current loading is a highly nonlinear phenomenon that traditional empirical equations struggle to capture. This study deploys ensemble boosting algorithms — AdaBoost, XGBoost, CatBoost, and LightGBM — to predict equilibrium scour depth, offering consulting engineers a data-driven alternative to conservative rule-of-thumb design factors.