Machine-learning-improved tide–wave simulations and joint extreme analysis for wave energy assessment in the Taiwan Strait

Abstract

High-accuracy environmental characterization is essential for reliable mooring system design and safety assessment of wave energy converters (WECs). This study employed the Finite-Volume Community Ocean Model (FVCOM) and Simulating Waves Nearshore (SWAN) to simulate tidal and wave conditions over 2005–2024. Holland-type typhoon wind fields were constructed for 197 historical typhoons during this interval. Bagged Regression Trees (BRT) machine learning model corrected modeled surface current velocity, significant wave height (H s ), and mean wave period (T m ) to enhance fidelity, reducing RMSEs by 45.5%, 32.5%, and 52.7%. Extreme environmental parameters were estimated using univariate statistics for current velocity and bivariate copula models for wave-period and wave-wind dependence. A floating oscillating water column (FOWC) was modeled in ANSYS/AQWA to evaluate the maximum mooring tension under a comprehensive range of environmental loading scenarios. Results indicate that the BRT model demonstrates a corrective effect in both cross-typhoon and cross-station validations, as well as in multi-year typhoon “leave-one-typhoon-out” cross-validation. Compared with univariate approach, copula-based modeling reduced 50-year return values by 0.32 m and 0.30 s for wave-period and by 0.52 m and 1.92 m/s for wave-wind. Furthermore, the environmental condition corresponding to the highest joint probability density does not necessarily yield the maximum mooring response.