Speaker
Description
Accurate estimates of extreme wave heights are important for the design and operation of coastal and offshore infrastructure. Although long-term wave hindcasts provide valuable spatial and temporal coverage, systematic biases can affect their reliability, particularly in nearshore environments. This study presents a bias-correction framework for significant wave height from the Norwegian Reanalysis (NORA3) using wave-buoy observations from three locations along the Norwegian coast: Fauskane, Fedjeosen, and Vestfjorden. Multiple statistical and machine-learning correction methods were combined using a Mixture-of-Experts (MoE) approach, in which an XGBoost-based gating model assigned input-dependent weights to each correction method. Results showed that NORA3 biases varied across locations, highlighting the need for site-specific correction. Local calibration consistently improved hindcast accuracy, with the largest improvement observed at Fedjeosen, where the MoE approach reduced the root mean square error by 26.6%. In cross-location transfer, the MoE model trained at Fedjeosen reduced the upper-tail root mean squared error at Fauskane by 29.2% but was less effective in other transfer scenarios. The corrected wave series also affected extreme-value estimates, resulting in changes in 50-year return levels of up to 0.57 m. These findings demonstrate the potential of data-driven bias correction to improve hindcast-based assessments of extreme wave conditions.
| Topic Area | Sustainable Ocean Structures |
|---|