Site-Specific Parameterization of Ocean Spectra for Power Estimates of Wave Energy Converters

TL;DR

This paper demonstrates that using four-parameter spectra, especially a novel machine learning autoencoder, significantly improves the accuracy of wave energy converter power estimates over traditional two-parameter models, emphasizing the importance of detailed spectral characterization.

Contribution

The study introduces a site-specific, machine learning-based four-parameter spectral model that enhances wave energy prediction accuracy compared to traditional two-parameter models.

Findings

Autoencoder-based spectra achieve around 1% error in power prediction.

Two-parameter spectra show less consistent performance with errors of -8% and 1%.

Four-parameter spectra better capture spectral variability and improve power estimates.

Abstract

Estimating the mean annual power of a wave energy converter (WEC) through the method of bins relies on a parametric representation of all possible sea states. In practice, two-parameter spectra based on significant wave height and energy period are ubiquitous. Two-parameter spectra have been shown insufficient in capturing the range of spectral shapes that can occur in an actual ocean environment. Furthermore, through sensitivity analysis, these two-parameters have been shown to be insufficient for predicting power performance of WECs. Four parameter spectra, which expand the parameter space to include two additional shape parameters have been shown sufficient in capturing sea state variance, but their effect on mean power estimates has not been presented. This work directly looks at the effects of incorporating 4-parameter spectra into annual power estimates compared to using the traditional 2-parameter spectra. We use two different 4-parameter spectra: one from the literature and a novel machine learning-based autoencoder, presented here. Both are shown to improve the information retained when parameterizing spectra. The site-specific autoencoder performs consistently the best across two case studies of mean annual power prediction, achieving an error around 1% in each instance. The 2-parameter spectra resulted in less consistent predictive performance, with errors of -8% and 1% in the two case studies. For the case study where all three models performed well, it is shown that the low error in the 2-parameter model is attributable to a symmetrical distribution of large errors whereas both 4-parameter spectra result in relatively low errors throughout the parametric space. These results highlight the need for more sophisticated resource characterization methods for estimating the power performance of WECs and suggest site-specific machine learning-based spectra are an adequate option.