A comparison of machine learning based surrogate models for multi-site performance assessment of wave energy converters

Wave energy is emerging as a promising renewable resource. Nonetheless, the levelized cost of energy for current Wave Energy Converters (WECs) still exceeds that of mature technologies like wind and solar. Bridging this gap demands an extensive assessment of possible designs and deployment sites. Therefore, we developed a machine learning (ML) based surrogate model that accelerates WEC performance evaluation. Although the framework is broadly applicable, its effectiveness is demonstrated here with the Pendulum Wave Energy Converter (PEWEC). The surrogate model receives the parameters describing single WEC setups as input and outputs the device’s general site-independent power matrix. The surrogate architecture comprises two main components: a classification and a regression algorithm. The first determines whether the parameter combination yields a feasible device setup, while the latter predicts the site-independent power matrix. For the classifier, we compared the performance of five ML algorithms: Logistic Regression, Random Forest (RF), Support Vector Classifier, Extreme Gradient Boosting, and Feedforward Neural Network (FNN). For the regressor, we compared four ML algorithms: RF, Gaussian Process Regression, Radial Basis Function Neural Network, and FNN. The FNN algorithm showed the best results for the classification and regression tasks. Lastly, we assessed the performance of the FNN-based surrogate model for WEC performance prediction in four representative deployment sites. Regarding annual productivity, the surrogate achieved relative errors below 7.1% at all four sites, limited the absolute deviation from the reference numerical model to under 2.1 MWh, and cut the computational time per device assessment from 222 s to 0.11 s.