| 13 Nov 2025
Evaluating Yawed Turbine Transfer Functions from SCADA Data
Abstract. While nacelle transfer functions (NTFs) have been applied to correct free-stream wind speed measurements at the nacelle of turbines steered into the wind, less is known about the relationship between unsteered (non-yawed) and steered (yawed) wind measurements on turbines performing wake steering. As wake steering becomes an important tool for maximizing collective wind farm power, determining and correcting bias caused by prolonged yaw misalignment on wind measurements is critical to improving collective wind farm control and analysis. We propose a new approach for evaluating NTFs using SCADA data. Using SCADA and wake steering controller 1-minute statistics recorded over 3.5 months at a large utility-scale wind plant, we apply several consensus methods to estimate unsteered turbine measurements for steered turbines using neighboring turbines. Then a bagged tree regressor algorithm is trained to predict the unsteered wind direction, wind speed, and generator power using the measured SCADA data during wake steering using the best consensus estimate as the target value. With the NTFs estimated through the ML model, we define experimentally determined non-linear sensor bias in the measured data as a function of yaw angle.
The manuscript develops a machine learning model to predict the wind speed, wind direction, and power production yawed turbines would experience if they were unyawed, with has as inputs only SCADA data. To support this, it also develops a way to estimate these quantities based on data from neighboring turbines. This work is useful, as it addresses a real problem that will become more important as active wake steering becomes more widespread. Furthermore, the use of real SCADA data demonstrates that the method could be implemented in real-world scenarios. However, the developed model itself is not convincing, and the manuscript does not explain it well. A full overview of the issues I have is given below.
Overall, the manuscript needs major changes before being considered for publication. I recommend adding a simpler model for comparison and a major rewrite of some sections, as discussed below. I would be happy to review a revised version with these changes.
Major comments:
It therefore seems to me that a simple linear regression between the measured and predicted wind vane and speed could perform well. Furthermore, literature already contains models relating yawed and unyawed power output as a function of turbine angle, such as a cosine power law. Compared to these simpler approaches, the full ML algorithm developed by the authors seems needlessly complex.
Including such a simple approach and comparing it against the ML model would greatly enhance the value of this manuscript. If not, the authors should argue more strongly and convincingly why a fully non-linear model is needed, building on earlier literature.
Furthermore, it’s not clear whether this is the standard method for estimating turbulence intensity based on SCADA data. If so, please refer to relevant literature.
Technical issues and minor comments/suggestions:
As a suggestion, consider moving this figure to section 3.2, where the filtering is developed.