Posters

Abstract

Accurate and reliable measurement of local wind speed is critical to wind energy performance analysis. However, as nacelle anemometers are subject to calibration transfer functions which may be changed over time, they are subject to time-variant bias. This is a challenge to the interpretation of nacelle anemometer power curves (NAPC's) for the purposes of tracking wind turbine performance. This work presents a machine learning method to calibrate the turbine local wind speed measurements from more accurately measured operational SCADA signals from neighbouring turbines. Following cleaning of the data by removing invalid samples, the method is not dependent on any explicit identification of turbine operational modes nor wind speed measurements for inference. For a given turbine, a Random Forest model is trained to predict measured wind speed from a 12-month cleaned dataset consisting of rotor speed, blade pitch angle and active power output from up to eight neighbouring turbines, with the power being split into "east" and "west" components according to the measured wind direction. This effectively uses the neighbouring turbine rotors as large anemometers, from which the model may be able to learn the relationship to the local wind speed including wake and terrain effects. The measured nacelle anemometer wind speed is then recalibrated by fitting and applying a rolling linear regression between the measured and ML-predicted wind speed over two-week windows. This corrects for time-variant bias of the target without importing high-frequency variance from the model. The recalibrated local wind speed can then be used to generate new NAPC's which are consistent over time. The model was validated by looking at the difference between measured and predicted wind speed aggregated into weekly means in order to evaluate the low-frequency variance of the model. The model was verified on all turbines in six different European sites of varying levels of terrain complexity each over a hold-out period of two years that was distinct from the training period. In all cases the interquartile range (IQR) of weekly mean wind speed errors was < 0.25 m/s, even for sites with complex terrain. Furthermore, the model was verified for relative performance in waked and wake-free conditions. In the majority of cases, the IQR of weekly mean wind speed errors stayed within 0.25 m/s even for turbines in the middle of a farm that are generally always operating in waked conditions. Time-variant bias of locally measured wind speed of operating turbines is a significant issue for quick and reliable identification of turbine performance anomalies. The presented approach uses readily available SCADA data to calibrate local wind speed measurements and establish a consistent baseline for NAPC's. The verification of the approach across many sites of varying terrain complexity shows that the approach can be widely adopted for fleet performance analysis thus leading to more security in performance monitoring. The approach highlights machine learning as a powerful tool to reduce the uncertainty of local wind conditions, the primary pain point in performance monitoring.