Detection of cup anemometer anomalies using machine learning

Abstract

Past research has successfully increased the accuracy and quality of horizontal wind speed measurements made by cup anemometers. Errors are introduced in a variety of ways: when operational conditions deviate from calibration, due to flow distortion and as a result of the mechanical properties of an instrument changing during its lifetime. Specific investigations on the latter source of errors have resulted in means to detect such damages, e.g. icing and bearing degradation. These methods, however, often require impractical measurement set-ups. A data-driven approach using expert knowledge of cup nemometer performance that uses two operational anemometers at the same altitude is suggested as a novel means of anomaly detection. The proposed anomaly detection model uses regression to predict the next residual between both instruments. Evaluation between measurements and predictions through confidence intervals reveal whether or not a flag is placed on the measurements. Data is obtained at the Light Mast North (LMN) in Østerild National Test Centre for Large Wind Turbines, Denmark. Wind speed measurements are taken at 10Hz from 2016-02-10 15:00 (UTC+2) until 2020-01-29 00:00 (UTC+2). The training set, used to develop the regression models, is comprised of normal data. The collocated cups show high correlation after designated filters, at both a 10-min and 10Hz scale, have removed the errors described above. Measurements of wind direction are required as well, since the residual between both instruments is dependent on where the wind is coming from. Directions where either cup is downstream of the mast are removed as well, since correlation becomes lower in these regions. Three regression models of increasing complexity are trained on a subset of data. A comparison is made between a Stochastic Gradient Descent (SGD), ARIMA and MultiLayer Perceptron (MLP) model. The MLP performed best in terms of false positive rate and mean squared error on a test set due to its adaptive confidence intervals and ability to model non-linearities. The trained MLP model is employed on regular time series and validation sets that contain periods of icing and structural damage. The former reveals that the number of flags is sensitive and proportional to the wind speed. At high wind speeds, the limitation of the model become clear, as it reflected a large rise in supposed anomalies. Large prediction uncertainties made it harder to detect failure at lower wind speeds, but sharp increases in the number of flags right before and after cup anemometers froze over were observed. Periods where structural damage affected the measurements stood out from normal behaviour, in terms of density of anomalies and extremeness. The promising capabilities of the MLP model are evident in that faulty events were labelled correctly, but its dependency on the magnitude of wind speed under normal operation indicates that sound user judgement is still required to interpret whether or not certain data can be deemed anomalous.