Posters

Abstract

Classification of different turbine operating modes and the highlighting of data quality issues forms the basis of performance monitoring of operational wind turbines. What is considered the normal turbine operating mode consists of the relationship between pitch angle, rotor/generator speed and power. The inherent stochastic nature of signals provided through turbine SCADA systems makes the analysis of this relationship and the identification of anomalous behaviour challenging. This often leads to suboptimal turbine performance going undetected for long periods causing significant lost yield. This work proposes the use of kernel density estimators (KDE's) to increase the sensitivity with which we can identify anomalous data samples, caused by either change in wind turbine control or data quality issues. This method is independent of turbine type and can be used to quickly identify changes in controller behaviour. A feature set is constructed from the 10-minute mean signals of generator speed, generator torque and pitch angle. A training period of one complete year when the turbine control behaviour did not exhibit significant variation is chosen. The dataset is filtered for valid samples and normal non-curtailed operation. The features are then each normalised into a range of zero to one. A KDE with a Gaussian kernel and Euclidean distance is then fitted on the samples to create a three-dimensional density function. The bandwidth of the Gaussian kernel is chosen by using a gradient descent algorithm to maximise the sum of sample densities in a 4-fold cross-validation. The density function is then applied to subsequent time samples beyond the training period to calculate a density time series, where a lower value indicates more anomalous behaviour relative to the training period. The peak probability density of the Gaussian kernel divided by the total training sample count is chosen as a critical density value. The proportion of samples below critical density in each day and week is then monitored. An increase in the proportion of samples indicates a consistent change in behaviour relative to the training period. The method was tested on data from turbines in two wind farms in Europe, where the monitoring team had discovered inconsistencies. In one, the apparent rated rotor speed had dropped by 0.25% and increased in variance. In the other, the mode gain (ratio between torque and speed in partial load) had changed by around 10%. When using the KDE to calculate the proportion of samples in each week below critical density, there was a clear step change at the time that the controller changes occurred. Due to the clarity in the step change, using this approach would have resulted in much quicker identification of anomalous behaviour compared to traditional methods of observed statistics. Controller changes can result in significant loss in annual energy yield when left unchecked. Using a largely automated KDE approach can significantly shorten the time in which controller changes can be picked up, which ultimately leads to a quicker rectification of the issue and a return to full performance, hence avoiding further loss of energy.