Regularized PCA-Based Prediction of Wind Turbine Underwater Noise From Few Unbalanced Observations

Recently, a method has been proposed that, by performing a principal component analysis (PCA) of the noise spectra recorded near an offshore wind turbine, followed by a supervised regression procedure, computes the prediction of the noise power spectral density as wind speed changes. This article addresses two problems of practical importance that the above method fails to adequately solve. In the former, the operating condition of the turbine is defined by the combination of wind speed and rotational speed. The spectra available are few and are recorded during sea campaigns in which operating conditions occur with different probabilities. As a result, the data set is highly unbalanced. In the second, the noise spectrum near a given turbine is predicted on the basis of numerous balanced recordings taken near a second turbine, nominally identical. To enhance the prediction, very few unbalanced recordings taken near the first turbine are available. Here, an original method is proposed that combines weighted PCA, where different importance is given to the recorded spectra, with a regularized technique to estimate the scores that finally combine the PCA results. A wide range of metrics (both general and focused on harmonics or on performance variation with respect to operating condition) is also adopted to assess the similarity between the predicted and actual spectra, thanks to which the advantage of the proposed method in solving the two problems is demonstrated.