Downscaling Global Wave Ensemble Forecasts With Machine Learning Techniques in Application to Billia Croo Test Site

Abstract

Abstract Accurate marine weather forecasts are essential for operational planning which could lead to significant cost savings on vessel, personnel, and equipment stand-by. There are many global forecasting products available on the market, providing free wind and wave forecasts on a global scale. However, the spatial resolutions are too coarse to fully resolve local effects. Several methods are available for forecast downscaling to represent local conditions: physical downscaling with high resolution mesoscale models, with global model providing the boundary conditions; bias correction of the global model forecast by an experienced forecaster with knowledge of local conditions; and statistical downscaling such as simple regressions or more sophisticated machine learning algorithms. The statistical downscaling approach allows fast, computationally inexpensive downscaling (as opposed to high resolution local model runs) without the need for continuous, operational support of a forecaster. Another advantage of this method is the possibility of constant improvement through continuous training as more data becomes available. The European Marine Energy Centre's (EMEC) Billia Croo test site for wave energy converters (WEC) is situated at a location dominated by high seas. This makes the operations and maintenance planning for WECs a challenge, which accurate forecasts could help to partially resolve. Datawell Waverider buoys provide in-situ wave measurements in real-time, with data also being collected and stored. This dataset allows the implementation of machine learning algorithms to downscale the global marine forecast products. In this study, an attempt is made to downscale operational ocean wave ensemble forecasts from NOAA/NWS/NCEP model runs, with forecast lead times of +12, +24, +36, +48, +72, +96 and +120 hours. After preliminary machine learning model training, an on-line training algorithm is put in place to assess its efficiency and capacity to continuously improve the accuracy. The utilisation of the wave ensemble forecasts as features in the machine learning model allows for consideration of the uncertainty of initial conditions of the forecast model run. A timeseries of in-situ observations in past hours are used as additional features, representing the preceding wave conditions.