Speaker
Description
Accurate forecasting of significant wave height is crucial for offshore safety, marine operations, and risk assessment in the Norwegian North Sea. ERA5 reanalysis provides long, spatially continuous wave records but shows systematic bias relative to in-situ observations. This study proposes a unified, multi-station hybrid deep learning approach that integrates bias correction, multi-horizon forecasting, and extreme-event detection. We construct a hybrid dataset combining ERA5 reanalysis with CMEMS in-situ buoy observations from three offshore platforms in the Norwegian North Sea, Sleipner-A, Troll-A, and Gullfaks-C, covering 2015 to 2022 at hourly resolution. Five bias-correction methods are systematically compared using a two-step nonparametric hypothesis-testing procedure. Two hybrid architectures, CNN-LSTM and CNN-Transformer, are evaluated for forecasting and extreme-event detection across four horizons of 1 h, 6 h, 12 h, and 24 h, using station-specific 95th-percentile thresholds. Both models improve the ERA5 baseline at the shorter horizons. CNN-Transformer reduces RMSE by 22.0% at 1 h, and CNN-LSTM achieves a 15.7% reduction at 12 h. For extreme-event detection, both reach an F1 score of 0.767 at 1 h, but scores decline to around 0.30 at 24 h. Explicit bias correction offers limited benefit, as the hybrid models implicitly correct the ERA5 bias during end-to-end training.
| Topic Area | Sustainable Ocean Structures |
|---|