To use SCADA data for wind turbine performance evaluation, we must clean the data by removing curtailment, downtime, and other anomalies. However, when eliminating data, there is a risk of introducing unwanted biases and losing the statistical characteristics of the evaluated power curve. A new paper by Morrison, Liu, and Lin (2022) investigates the performance of different combinations of data-filtering and Anomaly Detection (AD) techniques and evaluates the method’s prediction error and their ability to maintain the statistical variability of the power curve.

Before data cleaning, any missing or erroneous data must be excluded from the analysis. If these represent less than 5% of the data, their removal can be considered trivial; otherwise, applying gap-filling methods with the risk of introducing biases is necessary. Quality treatment of the SCADA system and its data is a prerequisite for any analysis.

Morrison, Liu, and Lin (2022) suggest applying a pre-filtering step that removes explicit operational anomalies, such as fault alarms, operational time, and large pitch angles, before running AD. Using SCADA data from 20 wind turbines, they compare the efficacy of five different AD techniques with and without pre-filtering. The analysis shows that pre-filtering followed by a Gaussian Mixture Modelling AD-technique significantly reduces the error while maintaining the statistical variability of the power curve.

Rory Morrison, Xiaolei Liu, and Zi Lin. 2022. “Anomaly Detection in Wind Turbine SCADA Data for Power Curve Cleaning.” Renewable Energy 184 (January): 473–86. https://doi.org/10.1016/j.renene.2021.11.118.