Measured ten-minute mean turbine performance characteristics are used to characterize the turbulence within wind farms. A neural network is trained to reproduce turbulence at each mean wind speed, given the ten-minute mean power production, blade pitch angle and rotor speed. The predicted turbulence from the neural network is verified using simulated wind turbulence that is input to aeroelastic simulations. The neural network is further trained to predict loads time series, given SCADA input time series. The load predictions are validated using an instrumented turbine. Based on the verified ability to predict wind turbulence and loads, the methodology is extended to reproduce loads on the blade and tower of wind turbines within wind farms in complex terrain. The input to the neural network is reduced to only the ten-minute mean measurements from the wind turbines, which are converted to 10-minute time series using Principal Component Analysis (PCA). The output of the neural network is the load time series, which are processed to damage equivalent loads. The standard deviation of measured power is used as a proxy for loads, to validate the predicted damage equivalent loads. The methodology is shown to provide a useful indication of relative fatigue life consumption within the wind farm, even when only mean SCADA measurements are available.
|Book series||Journal of Physics: Conference Series|
|Number of pages||10|
|Publication status||Published - 2020|
|Event||TORQUE 2020 - Online event, Netherlands|
Duration: 28 Sep 2020 → 2 Oct 2020
|Period||28/09/2020 → 02/10/2020|