Wind turbine wake detection and characterisation utilising blade loads and SCADA data: a generalised approach

Abstract

Large offshore wind farms face operational challenges due to turbine wakes, which can reduce energy yield and increase structural fatigue. These problems may be mitigated through wind farm flow control techniques, which require reliable wake detection (recognising the presence of a clear wake) and characterisation (parametric description of a wake's properties) as prerequisites. This paper presents a novel three-stage framework for generalised wake detection and characterisation. First, a regression model utilises blade loads, pitch and rotor rotational speed data to estimate the wind speed distribution across the rotor plane. Second, a convolutional neural network undertakes pattern recognition analysis to perform wake detection, classifying rotor-plane wind estimates as "fully impinged", "left impinged", "right impinged"or "not impinged". Third, where wake impingement is detected, two-dimensional Gaussian fitting is undertaken to provide a parametric wake characterisation, providing outputs of the wake centre location and wake lateral width. The framework is trained and tested in a simulation environment incorporating the Mann turbulence model, the dynamic wake meandering (DWM) model for generating wakes and an industrial-grade aeroelastic solver. The testing is undertaken under a wide range of wind conditions, with mean ambient wind speeds from 5-15 m s-1, turbulence intensities from 3 %-9 % and a full range of wind directions. Results show high accuracy of wind field estimation, with the mean RMSE across all test cases being 3.7 % when normalised by mean ambient wind speed. The wake detection model accurately identifies the presence of a wake for approximately 77 % of tested wind fields, with subpar performance observed for more extreme conditions or those at the limits of the training data used. The final wake characterisation stage is shown to flexibly adapt to changing wind conditions, successfully tracking the wake's position even for more turbulent conditions. The proposed framework therefore demonstrates strong potential as a generalised approach to wake detection and characterisation.