Degradation Estimation of Turbines in Wind Farm Using Denoising Autoencoder Model

We propose a method to estimate the power performance degradation in wind turbines (WTs) that arises from damage in the turbine blade and other components. In general, the single anemometer mounted on the nacelle is unable to measure precise wind speed distributions that the WT receives, thus, degradation of the power output is difficult to evaluate. By focusing on the fact that the power output data of adjacent WTs have some correlation although the wake effect on downstream turbines sometimes exists, our method uses the power output data of every WT in a farm to estimate the amount of degraded power performance of each turbine by the introduction of the virtual variable which corresponds to each turbine’s degraded amount. The feature of the correlation among each WT’s non-degradation data was learned by a denoising autoencoder (DAE). The virtual variables along with the power output are fed into the trained DAE model and these variables were updated by minimizing the reconstruction error. Moreover, the proposed method can perform the estimation even when some WTs are down, i.e., due to the periodical maintenance, and can classify between non-degraded and degraded WTs without enforcing diagnostics to set suitable threshold parameters. We demonstrated the superiority of this novel method over traditional methods by using real and artificial data inputs.