SCADA-Data-Based Static Yaw Misalignment

Estimation for Wind Turbines

Bojian Du**, Yoko Furusawa**, Yoshiaki Narusue****, Nozomu Nishihara***, Kentaro Indo***,
Hiroyuki Morikawa**** and Makoto Iida**

1. Introduction
For upwind wind turbines (WTs), yaw systems are Actual wind direction
used to align rotors to face into wind flow to maximize
energy capture. However, yaw misalignment is
inevitable as the measured wind direction values are
different from the free flow values due to the flow
distortion behind the rotor (Fig. 1). Yaw misalignment Measured
Wind vane wind direction
can lead to average energy losses up to 5% and
increasing in loading and structural fatigue [1]. The
Figure 1 Wind flow in front of and behind the rotor.
sketch map of the relation between yaw misalignment
angle and power loss is as shown in Fig. 2.
Employing additional sensors such as nacelle-
mounted lidar to directly measure the free flow wind
speed and direction in front of the rotor is an effective
solution [2]. While mounting sensors for all the WTs in a
wind farm requires significant cost. To compensate this,
modeling-based correction approaches have been
studied to estimate the actual yaw angle by using the
free flow data as reference [3][4]. However, these
approaches are time-costly as it takes at least several
Figure 2 Power loss against yaw misalignment angle,
months for the additional sensor to collect training data
assuming the examples of square-cosine (red) and cubic-
for every WT.
cosine (blue) relation [2].
To correct yaw misalignment efficiently at low cost, we
propose a machine learning-based approach to estimate
The proposed approach is evaluated on a real WT
the static yaw misalignment angle (i.e., the constant
SCADA data by comparing with nacelle-mounted lidar
error in yaw angle) by only using the historical
data. The result shows that the proposed approach can
Supervisory Control And Data Acquisition (SCADA)
estimate the static yaw misalignment angle of 1m/s
data. SCADA is a remote monitoring system installed in
interval wind speed bins with errors of no more than 1°,
WTs by default, which requires no additional investment.
and the effectiveness of the wind speed correction
Our proposed approach consists of two parts. Firstly,
technique is verified.
in order to reduce the effect of flow distortion on the wind
2. Methodology
speed measurements, we develop a wind speed
correction technique with a linear transformation model 2.1 Wind speed correction
which minimizes the dispersion of power against the
Besides wind direction, wind speed measurements
corrected wind speed. Secondly, we construct an artificial
are influenced by flow distortion as well, which will
neural network (ANN)-based model to predict wind
result in degradation of accuracy of the yaw
power and then leverage an argmax optimization to
misalignment estimation. Therefore, it is of great value
calculate the constant yaw misalignment angle.
to be able to correct the wind speed measurement.
* November 19, 2021 43th Wind Energy Symposium A linear transformation model-based wind speed
** Member, The University of Tokyo correction technique has been proposed by Mittelmeier
*** Nonmember, Eurus Technical Service Corporation et al. [3]. The free flow wind speed 𝑣free can be expressed
**** Nonmember, The University of Tokyo
as a function of the real yaw angle 𝛾real and the The basic idea of the SCADA-data-based yaw
measured wind speed 𝑣mea as the following equation: misalignment estimation is that, if all the other
𝑣free = 𝛽0 + 𝛽1 𝑣mea + 𝛽2 𝛾real + 𝜀, (1) operation conditions remain the same, the active power
where 𝜷 is the estimated regression coefficients, and 𝜀 is reaches the maximum value when the real yaw angle is
the random error. 0°.In this case, the yaw misalignment angle is just equal
In this study, 𝛾real and 𝑣free are unavailable since no to the measured yaw angle.
additional sensors are used. To address this, we develop To implement this idea, accurate modeling of WT
the wind speed correction technique by solving an behavior is needed. However, constructing the physical
argmin optimization problem to estimate the coefficients model of WT behavior is highly complex, as WT behavior
without using 𝑣free , where the measured yaw angle depends on environmental factors and mechanical
𝛾mea is used as an approximation to 𝛾real . Specifically, characteristics which may vary with different WTs.
the estimation of coefficient 𝜷 ̂ is determined as the Therefore, we use an autoregressive single hidden
arguments which can minimize the standard deviation layer ANN to construct a model to predict active power.
of the difference between 𝑃𝑤𝑠𝑐 , the power against Specifically, features that are able to affect the power
corrected wind speed (i.e., the estimated free flow wind generation are selected as the model inputs, including
speed) 𝑣̂free , and median power curve 𝑃𝑚𝑒𝑑𝑖𝑎𝑛 at the wind speed (here, we use 𝑣̂free ), ambient temperature,
same wind speed 𝑣: yaw angle, pitch angle, and turbulence intensity (the
̂ = argmin(𝜎(∆𝑃)),
𝜷 (2) ratio of the standard deviation of fluctuating wind
𝜷 velocity to the mean wind speed).
where In preprocessing, we select data from operating region
∆𝑃 = 𝑃wsc (𝑣) − 𝑃median (𝑣). (3) 2 in the segmentation according to the WT operating
𝑃𝑤𝑠𝑐 at 𝑣̂free is the same value as the actual power 𝑃 at regime [5]. In this region, the wind speed values are
𝑣mea : between cut-in speed and rated speed, where the power
𝑃wsc (𝑣̂free ) = 𝑃(𝑣mea ). (4) is sensitive to the change of yaw angle. To ensure that
The sketch map is as shown in Fig. 3. the model is constructed with data under normal
Though the optimal 𝑣̂free is not theoretically equal to operation conditions, we set thresholds for all the
𝑣free , the fluctuation caused by the distortion of wind variables involved in the modeling to eliminate outliers.
speed measurement can be reduced by the wind speed In model construction, the number of nodes in the
correction. hidden layer is determined by repeated evaluation with
2.2 Static yaw misalignment estimation the minimum average Root-Mean-Square Error
(RMSE), because the optimal node number is data-
dependent.
After the model is constructed, we fix all the other
input variables of the existing data records while
changing the yaw angle with a small interval (0.2°) to
search for the maximum output active power. And the
estimated static yaw misalignment angle 𝛾̂static is equal
to the average of input yaw angle 𝛾in corresponding to
the maximum output active power 𝑃out in the argmax
optimization:
1
∆𝑃 𝛾̂static = 𝑛 ∑(argmax(𝑃out )), (5)
𝛾in

where 𝑛 is the number of data records.

𝑣mea 𝑣̂ free 3. Evaluation

3.1 Evaluation setting and data description

Figure 3 Wind speed correction based on 𝑃wsc (the red
scatters) and 𝑃median (the blue line). E.g., position of the We evaluate our approach at a 1.75 MW WTs with 1-
green point changed after wind speed correction. min-interval SCADA data where one month's nacelle-
 yaw misalignment estimation occurs on 9-10m/s wind
speed bin, which is 1.00°.
Similar to Fig. 4, Fig. 5 shows the evaluation results of
static yaw misalignment but without wind speed
correction. The average value of the estimated yaw
misalignment angles is 1.84°, where the error is larger
than the estimated error of that with wind speed
correction.
4. Conclusion
In this paper, we propose a SCADA-data-based static
Figure 4 Evaluation result of static yaw misalignment yaw misalignment estimation approach and develop a
estimation with wind speed correction. wind speed correction technique to improve the
performance. The evaluation results show that our
proposed approach can achieve a close estimation of
static yaw misalignment angle compare to nacelle-
mounted lidar measurement. And the SCADA-data-
based wind speed correction technique is proved to be
effective.
Reference
[1] G. Cortina, V. Sharma, and M. Calaf. "Investigation
of the incoming wind vector for improved wind turbine
yaw-adjustment under different atmospheric and wind
Figure 5 Evaluation result of static yaw misalignment
farm conditions." Renewable Energy 101 (2017): 376-
estimation without wind speed correction.
386.
mounted lidar data is available for verification. The cut- [2] P. A. Fleming, et al. "Field-test results using a nacelle-
in speed and the rated speed of the WT are 6m/s and mounted lidar for improving wind turbine power
13m/s, respectively. The number of nodes in the hidden capture by reducing yaw misalignment." Journal of
layer of the ANN model is set as 10 by a preliminary Physics: Conference Series. 524.1 (2014): 012002.
repeated evaluation. The training period is set to one [3] N. Mittelmeier, and M. Kühn. "Determination of
year in order to reduce seasonality. optimal wind turbine alignment into the wind and
Fig. 4 shows the static yaw misalignment for the six detection of alignment changes with SCADA data."
respective 1m/s interval wind speed bins between 6- Wind Energy Science 3.1 (2018): 395-408.
13m/s in the evaluation. Where the blue points represent [4] D. Choi, et al. "Static and dynamic yaw
the actual mean yaw misalignment angles for the wind misalignments of wind turbines and machine learning
speed bins, red points with error bars represent the based correction methods using lidar data." IEEE
estimated static evaluation, the gray scatters represent Transactions on Sustainable Energy 10.2 (2018): 971-
the wind speed against yaw angle of all data records. 982.
Where the blue points represent the actual mean yaw [5] K. Kim, et al. "Use of SCADA data for failure
misalignment angles for the wind speed bins, red points detection in wind turbines." Energy Sustainability. Vol.
with error bars represent the estimated static yaw 54686. 2011.
misalignment angles with 95% confidence intervals. The
average value of yaw misalignment angles based on
lidar measurement and our proposed approach
estimation are 1.37°and 1.42°, respectively, with a very
small error. The maximum error of the wind speed bin