Abstract
Accurate weather forecasting is crucial for various industries,
including offshore wind, which is vital for global net zero
energy goals. Machine learning models, trained on historical
data, offer a new opportunity by replacing traditional physicsbased
equations. These models can learn patterns not always
represented by physical equations, potentially increasing the accuracy
and efficiency of weather forecasting compared to traditional
Numerical Weather Prediction (NWP).
This study applies a machine learning framework (MaLCOM)
to offshore wind forecasting in the Celtic Sea. It uses an
attention-based LSTM recurrent neural network to learn temporal
patterns and a random forest-based spatial nowcasting model,
trained on ERA5 data, for spatiotemporal predictions. Winds
derived from wave spectra measured by buoys are integrated,
showing the framework’s value even with imperfect data.
Validation with independent observations from floating lidar
units in 2023 confirms the framework’s suitability for regional
wind prediction. This work extends previous machine
learning-based predictions of ocean conditions to wind forecasting,
demonstrating a new approach to metocean data. These
lightweight, data-driven predictions can be run on standard computers,
improving real-time decision-making for offshore planning.
including offshore wind, which is vital for global net zero
energy goals. Machine learning models, trained on historical
data, offer a new opportunity by replacing traditional physicsbased
equations. These models can learn patterns not always
represented by physical equations, potentially increasing the accuracy
and efficiency of weather forecasting compared to traditional
Numerical Weather Prediction (NWP).
This study applies a machine learning framework (MaLCOM)
to offshore wind forecasting in the Celtic Sea. It uses an
attention-based LSTM recurrent neural network to learn temporal
patterns and a random forest-based spatial nowcasting model,
trained on ERA5 data, for spatiotemporal predictions. Winds
derived from wave spectra measured by buoys are integrated,
showing the framework’s value even with imperfect data.
Validation with independent observations from floating lidar
units in 2023 confirms the framework’s suitability for regional
wind prediction. This work extends previous machine
learning-based predictions of ocean conditions to wind forecasting,
demonstrating a new approach to metocean data. These
lightweight, data-driven predictions can be run on standard computers,
improving real-time decision-making for offshore planning.
| Original language | English |
|---|---|
| Journal | 44th International Conference on Ocean, Offshore & Arctic Engineering |
| Publication status | Published - 2025 |
