Posters

Abstract

INTRODUCTION Performance monitoring of operational wind assets requires accurate and time-consistent measurements of the environment and operational status of the wind turbines. Despite this, many signals, such as wind speed and absolute wind direction exhibit time-variant bias and variable inaccuracy, leading to improper performance analysis and costing asset owners lost energy yield. These signals therefore must be calibrated by models or external sources. Nacelle anemometer wind speed and nacelle direction both exhibit time-variant bias due to instantaneous step changes (e.g. due to software changes) as well as low-frequency drift (due to e.g. degradation). A calibration signal, such as a model output or an external source (e.g. reanalysis data) will have a non-zero variance (noise) which we want to exclude, whilst leveraging lower frequency information from the calibration signal. This can be achieved by using low-pass filters or rolling averages, but perform badly in the presence of step changes. METHOD The presented method decomposes the calibration signal into parts that are then selected according to use-case: * A raw calibration factor is calculated as the ratio between the measured and calibration signals. * A changepoint detection algorithm is employed to find step changes, and the median value of the samples between each changepoint, is used to create a new component signal which is removed from this calibration factor. * The “extra-seasonal trend” is calculated by taking the 12-month moving average and removed. * A seasonal trend is calculated by fitting a Fourier series to the data, and removed. * An “intra-seasonal” component is calculated using  a low-pass filter with a configurable cut-off period (e.g. 30 days), and removed. * The remainder is the noisy residual, and is generally discarded. This leaves a calibration factor decomposed into four parts which can be combined based on the application. APPLICATION We apply this algorithm to the calibration of anemometer wind speed using the output from a machine learning model trained using data from neighbouring wind turbines. We keep only the changepoint and intra-seasonal components of the calibration factor. This identifies both sudden changes in bias, and long-term drift. Critically, this does not include the seasonal component (potentially an artifact of the model), nor the extra-seasonal trend, likely a result of the slow performance degradation of neighbouring turbines. The calibrated signal is used in existing analytics processes to calculate power curve efficiency in partial load. With an uncalibrated signal, significant deviations in efficiency are observed, the majority being false positives. In contrast, using the calibrated time-consistent signal for power curve efficiency, there is a much smaller deviation in power curve efficiency observed in cases with no known performance issues. This leads to a lower false positive rate and therefore higher confidence in the analysis leading to faster and more decisive action. Conclusion This work demonstrates the problematic impact of time-variant biased signals on performance analysis and the importance of calibration. It presents methods to increase reliability and trust in measured signals, enabling faster corrective action, and maximising the energy generation of assets.