Enhancing the accuracy of metocean hindcasts with machine learning models

To improve the quality of wind and wave forecasts post-processing algorithms are developed to bias-correct the variables of interest using machine learning models trained with buoy data. Two types of deep neural networks are developed and studied to improve the accuracy of ERA5 hourly data of 10-m wind speed (wsp) and significant wave height (swh) at two locations with distinct metocean conditions. The algorithms are the multilayer perceptron (MLP) and the long short-term memory (LSTM) neural networks. The latter is built based on a range of previous time steps included in the input, evaluating the optimal value. Several metrics such as bias, scatter index, root mean square error and correlation coefficient are used to evaluate model performances. After extensive examination and experimentation regarding feature selection, filtering windows, and model architecture, the results show better improvements for swh than wsp, and that LSTM outperforms MLP significantly. The original correlation coefficients between ERA5 and observations for swh have been improved and the ERA5 RMSE has been reduced with the LSTM post-processing model.

Automated summary

Machine learning models trained with buoy data are used to develop deep neural networks for improving the quality of ERA5 wind speed and wave height data, with LSTM outperforming MLP significantly. Multiple metrics are used to evaluate the model performance.