Error Correction for Wave Modelling

Abstract

Installation and maintenance strategies regarding offshore wind farm operations involve extensive logistics, since the main focus is the right temporal and spatial placement of personnel and equipment, while taking into account forecasted meteorological and hydrodynamic conditions. In order for these operations to be successful weather windows characterized by certain permissive wave, wind and current conditions is of enormous importance, whereas unforeseen events result in high cost and risk in terms of safety.
For that purpose, Deltares created Meteo Dashboard, an integrated software system that collects, stores, computes and presents measured and forecasted meteorological and hydrodynamic data for decision making of maintenance or installation activities in an offshore wind farm. The wind speed, as well as the air and water temperatures, result from a meteorological model and serve as an input for the numerical modelling (e.g. SWAN or Delft3D) of waves, water levels and current related parameters. To account for the inherited uncertainty, several error modelling techniques, such as Artificial Neural Networks (ANN), Copulas, Stochastic Interpolation (SI), ARMA models, and Linear Regression (REG), already run operational on Meteo Dashboard and can be implemented in order for the numerical model forecasts to be corrected. A number of the aforementioned techniques require training using historical or present time data, while others can be incorporated forthwith.
In this research, a fully automated ARIMA model and different kinds of Bayesian Network (BN) models are incorporated in order to enhance the accuracy of the significant wave height (Hs) predictions even further. Both techniques are implemented using packages provided by the free software environment of R, namely the bnlearn and forecast. The implemented BN models differ in terms of training and structure, and provide overall the most satisfying accuracy in comparison to the rest of the error correction techniques, when tested with data retrieved from stations deployed in the Irish Sea (adjacent to the Gwynt-y-Mor and Rhyl Flats offshore wind farms) corresponding to the whole year of 2017 (from January 2017 – to January 2018).
Supplementary, it is also shown that the BN models illustrate even more advantages when compared to the rest of the error correction techniques, since they provide information about the incorporated variables dependence relationship through their structures, while producing estimates for the underlying uncertainty of the phenomenon, by means of 95% confidence intervals extracted by the significant wave height (Hs) conditional distribution.
Finally, all error correction models are tested in operational (online) mode, with real-time data from the aforementioned locations, with the newly implemented BN models producing results of enhanced accuracy, even in the absence of measurements