Cost-Effective Downscaled Wave Forecasting Using Machine Learning

In the offshore energy environment, accurate and timely metocean forecast data is essential for informed operational decision making. There is a growing demand from small and medium-sized enterprises (SMEs) for more detailed data that is freely available or at low cost. Improving near-shore forecasting can significantly improve decision making, particularly in marginal conditions, thereby expanding the safe operational window for offshore activities. This improvement depends on better data accuracy to improve forecasting and risk assessment.

Ensemble forecast models provide a comprehensive overview of metocean conditions by accounting for uncertainties and offering probabilistic insights. This makes them particularly useful for robust risk assessment and decision-making under uncertainty. However, ensemble forecasts are computationally demanding as they are run at frequent intervals with multiple perturbations. As a compromise, these models use lower resolution domains for computational efficiency, allowing probabilistic modelling with global coverage at the expense of deterministic accuracy.

This work presents a novel machine learning (ML) framework that has been developed to correlate and correct the deterministic error between low-resolution forecast models and high-resolution physics-based models or in-situ measurements. By integrating this ML framework into the forecast workflow, the model-based ensemble forecast results are adjusted at run time to improve deterministic accuracy. This approach enables the production of forecasts with downscaled accuracy, minimising production time and cost without compromising accuracy.

The work is based on an extensive database of calibrated high-resolution hindcast models and metocean measurements. The ML model is trained to learn non-linear downscaling functions that map low-resolution outputs to corresponding high-resolution wave models or observational data at predefined locations. The ML framework includes various models such as long short-term memory (LSTM), linear regression, random forest, gradient boosting, and dense neural networks. All the models share a standardised architecture optimised for metocean forecasting. Designed as a modular, plug-and-play solution, the framework enables rapid deployment, testing, and integration into the forecast workflow.

Results from a case study in the Baltic Sea are presented. With the ML integrated forecast approach a 65% reduction in RMSE for significant wave height is achieved compared to the unadjusted wave forecast data.