Home Search Collections Journals About Contact us My IOPscience

Prediction and Reduction of Noise from a 2.3 MW Wind Turbine

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2007 J. Phys.: Conf. Ser. 75 012083

(http://iopscience.iop.org/1742-6596/75/1/012083)

View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 87.72.249.156
The article was downloaded on 25/12/2011 at 21:33

Please note that terms and conditions apply.

The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

Prediction and Reduction of Noise from a 2.3 MW Wind Turbine

G Leloudas1,2, W J Zhu1, J N Sørensen1, W Z Shen1, S Hjort3

1
Technical University of Denmark
2
Dark Cosmology Centre, Copenhagen University
3
Siemens Wind Power A/S

E-mail: [email protected]

Abstract. We address the issue of noise emission from a 2.3 MW SWT-2.3-93 wind turbine
and compare simulations from a semi-empirical acoustic model with measurements. The noise
measurements were taken at the Høvsøre test site for large wind turbines. The acoustic model
is based on the Blade-Element Momentum (BEM) technique and various semi-empirical
acoustic relations. The comparison demonstrates a generally good agreement between
predicted and measured noise levels. The acoustic model is further employed to carry out a
parametrical study to optimize the performance/noise of the wind turbine by changing tip speed
and pitch setting. We show that it is possible to reduce the noise level up to 2 dB(A) without
sacrificing too much the power yield.

1. Introduction
Wind turbines are viewed with sympathy by most people. However, one of the main concerns for
neighbors to planned wind farms is noise. This is not a surprise: experience has shown that people are
worried about noise, more than anything else, when it comes to the creation of anything new, ranging
from roads and shopping centers to night clubs, in their surrounding area.
In the planning stage of a wind farm, it is in most cases sufficient to utilize simple empirical
prediction models, such as those included in most commercial packages, to predict the noise level.
However, there have been cases where, for one or another reason, these predictions have failed leading
to complaints (e.g. see [1]).
Although the emission of noise from wind turbines in most cases is smaller than that from other
environmental noise sources, such as roads, airports and construction machinery, wind turbines are
usually placed in rural environments, where the background noise typically is low. Thus, wind turbine
noise is of great concern since it may be the only major noise source in rural districts.
Today, machinery noise is reduced efficiently by well-known engineering techniques, such as
proper insulation of the nacelle. Traditionally, aerodynamic noise has been controlled by lowering the
tip speed to a maximum of about 60 m/s, as the tip speed is the most significant parameter affecting
aerodynamic noise. However, in recent years the biggest development of wind turbines has taken place
offshore, with the result that the latest generation of wind turbines operate at tip speeds up to 80 m/s,
indicating that noise again may be a problem with respect to public acceptance.
Studying noise from wind turbines is not a new field. Recently, however, its importance has grown
so that it even needs its own dedicated conferences: the first conference on wind turbine noise took
place in October 2005 in Berlin with a follow-up scheduled for 2007 in Lyon.


c 2007 IOP Publishing Ltd 1
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

Through the years, several models have been proposed to explain and predict wind turbine noise.
Some of the models are somewhat simplistic (e.g. [2-4]), whereas others make use of complex CFD
solvers that have not yet matured to be applied in practice (e.g. [5]). As a compromise between
computing speed and accuracy, the most commonly used models are based on semi-empirical
relations. As basis, most models employ the experimental results on airfoil self-noise by Brooks, Pope
and Marcolini [6]. These data, that can be directly used to predict wind turbine noise, are based on
wind tunnel experiments of NACA 0012 airfoils.
In a previous work, we have developed a semi-empirical noise prediction model [7,8] using the
scaling laws given in Brooks, Pope and Marcolini [6], together with the turbulence inflow model
proposed by Amiet [9]. The model was tested successfully against a Bonus 300 kW wind turbine and
was used in a parametrical study yielding many useful and interesting results.
Similar work, based on the results of [6], has been conducted by e.g. Fuglsang and Madsen [10],
Moriarty and Migliore [11] and in the SIROCCO project [12]. This demonstrates that these prediction
laws are widely accepted as the best available prediction tools.
The main goal of this paper is to validate the model against measurements of a MW wind turbine
and to couple it with an optimization tool in order to optimize the operational conditions of wind
turbines with respect to both performance and noise.
In section 2, we give a short description of the noise prediction model, including a summary of the
main ingredients of the model. In section 3, we briefly report on some recent noise measurements and
we validate our predictions against the obtained data. In section 4 we present the results from a
parametrical optimization study, and in section 5 we summarize our conclusions.

2. The noise prediction model

In this report we treat only aerodynamic noise from wind turbines (i.e. mechanical noise is not
considered). Aerodynamic noise can be divided into airfoil self-noise and turbulence inflow noise. The
former is a result of the interaction of the boundary layer of the airfoil with the trailing edge and the
latter results from the interaction of the existing turbulence in the wind with the airfoil.
In our model, the airfoil self-noise prediction is based on the functions given by Brooks, Pope and
Marcolini [6]. In total, five airfoil self-noise mechanisms were identified and studied separately:

• Turbulent Boundary Layer Trailing Edge noise

• Separation-Stall noise
• Laminar Boundary Layer Vortex Shedding noise
• Tip Vortex Formation noise and
• Trailing Edge Bluntness Vortex Shedding noise

As a result, scaling laws were proposed, yielding the sound pressure level at the observer position
as a function of frequency for the 1/3 octave band spectrum. The scaling laws for the different
mechanisms are all of similar form:

⎛ δ i* M f ( i ) LDh ⎞
SPLi = 10 log⎜⎜ ⎟ + Fi (St ) + Gi (Re )
r2 ⎟ (1)
⎝ ⎠

where δ i* is the boundary layer displacement thickness, M is the Mach number, f (i ) is the
raised power which depends on the particular noise mechanism i , L is the airfoil section semi-span,
Dh is a sound directivity function and r is the distance to the observer. The additional terms Fi (St )
and Gi (Re ) are functions of the Strouhal number St = fδ ∗ U and the Reynolds number Re . The

2
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

nature of dependency is different for each noise mechanism but it is impressive that all the formulas
look so much alike.
For turbulence inflow, a prediction equation based on the work of Amiet [9] has been implemented
in the model:

⎛ ΔL
(
L p = 10 log⎜ ρ 02 c02l 2 M 3 I 2 kˆ 3 1 + kˆ 2 )
−7 3 ⎞ ⎛ Kc ⎞
⎟ + 58.4 + 10 log⎜⎜ ⎟⎟ (2)
⎝ r ⎠ ⎝ 1 + Kc ⎠

where l is a turbulence length scale, I is the turbulence intensity, ρ 0 is the density, c0 the speed
of sound, ΔL is the blade segment semi-span, k̂ is a corrected wave length and K c is a low
frequency correction.
Taking into account all the variable dependencies, the problem of predicting the noise spectrum at a
given observer position for a given airfoil reduces to identifying the following quantities:

• The boundary layer thickness δ ∗ at the trailing edge of the airfoil

• The relative wind speed defining M and Re
• The boundary layer transition type (forced or natural), leading to tripped or un-tripped flow
• Miscellaneous input parameters to the turbulence inflow noise model, such as turbulent length
scale and intensity, in the model reduced to the knowledge of the height from the ground z and
the roughness length zo .

In this paper we do not go into the theory behind the empirical correlations, and for details about
the nature of each of the modelled noise mechanisms we refer the reader to the original work of
Brooks, Pope and Marcolini [6] and Amiet [9].
As mentioned above, an important parameter for the calculation of airfoil self-noise is the boundary
layer thickness at the trailing edge. This is calculated by use of the program XFOIL [13]. It is
important to note that the scaling laws shown above are deduced from experiments based only on the
NACA 0012 airfoil. For this reason, an independent calculation of δ ∗ for each airfoil type is vital.
This was done for a number of different values of Reynolds number and angle of attack and the
computed boundary layer thickness was stored in a database and subsequently determined by
interpolation.
Essentially, the code consists of a ‘traditional’ BEM code (see e.g. [14]), to compute the relative
velocities along each blade element defining the rotor, coupled with the routines by [6] to predict the
noise contribution for each noise source along the span of the rotor blades. In a few words, the
prediction code works as follows. First, the relative velocities seen by the blade elements are
computed, just like in an ordinary BEM computation. Next, a table look up in the boundary layer
thickness database is made and the sound pressure level L p and the noise spectrum at the observer
position is calculated for each noise mechanism and for each blade element. Finally, the sound
pressure levels are added for all elements, all blades and all mechanisms and converted to sound
power levels Lw referring to the hub of the wind turbine.
The main advantage of the semi-empirical model is that it is fast to run, even on a PC, and that it
gives surprisingly reliable results, as will be demonstrated in the following. It is also fairly easy to
couple the prediction code to an optimisation algorithm and use it as a tool to optimize the rotor with
respect to both performance and noise.

3. Noise measurements and model validation

The noise measurements were performed on a Siemens SWT-2.3-93 wind turbine at the Høvsøre Test
Site for Large Wind Turbines. The measurements took place during two days, resulting in

3
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

approximately 11 hours of data. Microphones were placed at four different positions, two downwind
the rotor and two in the rotor plane, and the data were recorded simultaneously by all the microphones.
Further, power production and wind speed, measured at hub height, were recorded in order to set the
reference. The measurements were taken at different tip speeds and pitch settings. In order to check
noise emission at off-design situations, the settings did not necessarily correspond to the settings at
normal operational conditions.
In figure 1 we compare the computed Sound Power Level (SPL) with measurements as a function
of wind speed for a turbine running at normal operational conditions. The measured data points are
displayed together with a best fit and an 2σ interval, giving a 95% confidence. It is observed that the
simulation over predicts the measurements with up to 2 dB, staying, however, within the 95%
confidence interval.

Figure 1: Comparison between computed SPL and Figure 2: Comparison between computed SPL
measurements for the Siemens SWT-2.3-93 wind and measured data as function of rotational
turbine operating at normal tip speed and pitch speed at various wind speeds.
setting.

The rotational velocity is the most important parameter affecting the noise production and it is
therefore vital that the model captures the right ω -dependency. In figure 2, we compare the simulated
SPL against measured data as function of rotational speed at different wind speeds, while keeping the
pitch angle constant. A best fit together with an 2σ interval has been added to the measurements. All
points have been binned with respect to wind speed and the simulation curve is for the mean measured
wind speed.
We make three main observations: 1) There exists a linear relationship between the sound power
level and the rotational speed; 2) For the same rotational velocity, higher wind speeds result in higher
noise levels; 3) The comparison demonstrates that the model is capable of predicting the
measurements.
In order to study the dependency of the noise level on the pitch setting, we present two graphs
showing the SPL at different wind speeds and constant rotational speed. In figures 3 and 4 we show
the results at a rotational speed of 17 RPM and 13 RPM, respectively. All simulations were made for a
mean wind speed derived from the measured values. It should be noted that the tip speed tested in
figure 3 is actually larger than the normal operational limit of the SWT-2.3-93.

4
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

Figure 4: Comparison between computed SPL

Figure 3: Comparison between computed SPL and measurements as function of pitch settings
and measurements as function of pitch settings at at different wind speeds and a medium
different wind speeds and a high rotational speed rotational speed of 13 RPM.
of 17 RPM.

From the figures we observe that the prediction at ω =17 RPM is in very good agreement with
measurements. At ω =13 RPM, however, the measured level is over predicted by around 3 dB. A
likely explanation for this is that the measured data at this particular experiment was dominated by a
high background noise level that may have polluted the filtered data. Nevertheless, in both cases the
noise clearly increases when decreasing the pitch angle. This is due to the increasing importance of the
boundary layer separation noise mechanism, as the angle of attack becomes higher. This difference
can be as high as 2 dB(A) for 5o difference in pitch.
Except for broadband noise, it is important to test how the model predicts the frequency
distribution in the 1/3 octave spectrum for different operational settings. In figure 5 we show a
representative example in which a computed spectral distribution of SPL is compared to a measured
distribution.

Figure 5: Spectral distribution of total

SPL. The experimental spectrum is the
average of all the measurement points
(107 10sec averages in total) which
had 13.5 RPM< ω <14.5 RPM, -
3< θ <-1 and 7 m/s<Vo<9 m/s. The
simulated spectrum is for the averaged
setting i.e. ω =14 RPM, θ =-2 and
Vo=8 m/s. This is a representative
example of agreement between
modelled and measured spectra.

Comparing the two curves, a good agreement is observed for frequencies between 10 and 3000 Hz.
For higher frequencies, however, the simulation systematically over predicts the measurements. In
fact, the measurements show that the noise level is negligible above 10 kHz, while the model predicts

5
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

a level up to 90 dB(A). Analyzing this phenomenon more closely, it was found to be partly due to
bluntness noise, which is tonal and predicts an unphysical high peak, and partly due to turbulence
inflow noise, which turns out to be unreasonably high at these high frequencies. This indicates that the
bluntness mechanism elaborated by Brooks, Pope and Marcolini [6] should be analyzed further and
that there seems to be room for improvements. However, up to 3 kHz the computed distribution is
everywhere in excellent agreement with the measured values.

4. Results of parametrical study

4.1. Simulation results

In figure 6 we show the result of a parametrical study in which predicted noise contours are shown as
function of pitch angle and rotational speed. As expected, the noise level increases as we move to
higher rotational velocities. However, the pitch angle also has an effect. At constant rotational speed,
the noise increases as the pitch angle decreases (the angle of attack increases and the blade goes into
stall). In figure 7 we present computed contour levels of total sound pressure as a function of radial
distance in the rotor plane. The contour levels reflect in which part of the blade most noise is
produced. As expected, most noise is generated at the tip, where the velocities are highest. However,
we also observe a secondary maximum at 50% blade radius, where the blade becomes significantly
blunt.

Figure 6: Computed noise contours (SPL) for the Figure 7: Computed noise contours (SPL) at
SWT-2.3-93 wind turbine operating at different different points in the rotor plane. Each point is
settings. The simulation is for a wind speed of 10 computed by integrating the contribution from
m/sec. all blade elements and all noise sources.

4.2. Optimization of operating conditions

It is our purpose here to optimize the SWT-2.3-93 operation, by changing the rotational velocity and
the pitch angle. As demonstrated by Fuglsang and Madsen [10], different optimization strategies exist.
In the present work we either seek to minimize noise keeping the power at a constant minimum level
or we maximize the power keeping the noise below an a priori defined level. While minimizing the
noise level we try to lower Lw without sacrificing too much the power production. A strict constraint
for the power would be 99% while a more relaxed would be 95% of the maximum power production.
In power maximization, like the name implies, we try to maximize the power production, constraining
however the noise not to exceed a maximum value. In the optimization we only include wind speeds in
the range from 7 m/s to 15 m/s. The reason for this limitation is that at higher wind speeds the wind

6
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

turbine noise gets masked by the increased background noise from trees, bushes, waves, etc, and that
this interval probably is the most sensitive both with respect to power production and noise annoyance.
The tip speed-pitch combination is optimized for each of the considered wind speeds separately. As
optimization tool we employ the function fmincon that is a part of the MATLAB optimization toolbox
and that is used for constrained non-linear optimization.
As in every optimization subject to limitations, the result is basically related to the constraints
imposed to the optimization search. In our case, the optimization results were forced to obey two
constraints: that the power should not exceed 2300 kW and that the driving torque should not exceed a
certain limit, which would increase the damage probability of the gearbox. No other parameters, such
as those related to loads or the control system, were taken into account.

4.2.1. Noise minimization. As explained above, we look for a combination of rotational velocity and
pitch setting that leads to a less noisy operation, by constraining the power loss that we are willing to
trade for this noise reduction. We have performed optimizations for two different constraints in power:
99% and 95% of the power achieved at normal operational settings. In figure 9 we show the reduction
in sound power level SWL, measured in dB(A), for these two cases. We observe that considerable
reduction in noise (2 dB) can be obtained by losing only 1% in power at wind speeds up to 9 m/s. This
gain however exhibits a minimum of 0.4 dB at 11 m/s. Unfortunately, this is also the location at which
the noise level is maximum. In order to obtain a significant noise reduction at wind speeds higher than
11 m/s, more sacrifices in power (5%) are needed.

Figure 9: optimized (minimum) power losses

Figure 8: Noise reduction in dB(A) as a result of for different noise constraints at different wind
optimization at different wind speeds. speeds. The stricter the constraint, the lower the
power production. The maximum losses are at
11 m/s.

4.2.2. Power maximization. In this type of optimization, we force an upper limit in noise and look for
the optimum settings that will maximize power. This might for example be a limit that is imposed by
legislation and that cannot be trespassed (e.g. in combination with the distance to the nearest
dwellings). In a sense, this type of optimization is more realistic, since one might not want to reduce
the power yield at 8 m/sec, and only accept the minimum reduction in power at 11 m/sec. However,
the study in the previous section can still be valuable in its conclusions if all results are given in
relative terms. In figure 9 we depict the power loss (in %) for different wind speeds and different
imposed noise constraints. This reduction in power, however, is only significant for wind speeds
between 10 m/s and 12 m/s. In figures 10 and 11 we show the values of the optimum rotational speed
and the pitch setting that have been utilized to achieve the maximum noise reductions shown in figure

7
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

9. As expected, the most important means for reducing noise is lowering the rotational speed,
combined, however, with some changes in the pitch angle (usually by pitching away from stall).

Figure 10: Modifications of the rotational speed Figure 11: Modifications of the pitch setting
leading to the optimum noise constrained power leading to the optimum noise constrained power
curves. curves.

5. Conclusions
A semi-empirical noise model has been validated by comparing computations with measurements of a
Siemens SWT-2.3-93 wind turbine. The measurements were performed using 4 microphones located
in the vicinity of the wind turbine that was operated at different rotational velocities and different pitch
angles. The comparison shows that the model tends to over predict the noise level at normal operation
conditions of the wind turbine, and that it follows the correct trend when varying the pitch setting and
the rotational speed of the rotor. The computed noise spectrum is in excellent agreement with the
measured one up to frequencies of 3 kHz. For higher frequencies, however, there seems to be a
systematic problem in relation to the expression for blunt trailing edge noise. This excess in the right
side of the predicted spectrum is also responsible for the overall overprediction.
The code was coupled to an optimization tool and used for optimizing the wind turbine's
operational settings. This was done for a series of wind speeds and for two different optimization
strategies. It was found that for some wind speeds considerable reductions in noise can be obtained at
a low cost. As we approach the wind speed at which the noise level is highest, however, the loss of
power increases considerably.
It is our hope that this kind of optimization can lead to a more quiet operation of existing wind
turbines by simply varying the tip speed and pitch setting in a an appropriate manner.

Acknowledgements
We wish to thank Ejler Kristensen at Siemens Wind Power A/S for valuable help with the noise
measurements and for pre-processing the data files.

References
[1] Van den Berg G P 2004 Effects of the wind profile at night on wind turbine sound J. Sound Vib.
277 4-5 pp 955-970
[2] Grosveld F W 1985 Prediction of broadband noise from horizontal axis wind turbines J. Prop.
Power 1 pp 292-299
[3] de Wolf W B 1986 Een predictiemethod voor her aerodynamische geluid van winddurbines met
horizontale AS NLR TR 87018 L pp. 1-55
[4] Wagner, Bareiss and Guidati 1996 Wind Turbine Noise (EUR 16823 Springer)
[5] Shen W Z and Sørensen J N 2001 Aeroacoustic modeling of turbulent airfoil flows AIAA

8
The Science of Making Torque from Wind IOP Publishing
Journal of Physics: Conference Series 75 (2007) 012083 doi:10.1088/1742-6596/75/1/012083

Journal 39 6
[6] Brooks T F, Pope D S and Marcolini M A 1989 Airfoil self-noise and prediction NASA
Reference Publication 1218 National Aeronautics and Space Administration USA
[7] Zhu W J, Heilskov N, Shen W Z and Sørensen J N 2005 Modeling of aerodynamically
generated noise from wind turbines J. Solar En. Engin. 127 pp 517-528
[8] Zhu W J, Sørensen J N and Shen W Z An aerodynamic noise propagation model for wind
turbines Wind Engin. 29 2 pp 129-143
[9] Amiet R K 1975 Acoustic radiation from an airfoil in a turbulent stream J. Sound Vib. 41 pp
407-420
[10] Fuglsang P and Madsen H A 1996 Implementation and verification of an aeroacoustic noise
prediction model for wind turbines Risø National Laboratory publication R-867(EN)
[11] Moriarty P and Migliore P 2003 Semi-empirical aeroacoustic noise prediction code for wind
turbines National Renewable Energy Laboratory publication NREL/TP-500-34478
[12] Schepers J G, Curvers A P W M, Oerlemans S, Braun K, Lutz T, Herrig A, Wuerz W and
Mendez Lopez B 2005 SIROCCO: SIlent ROtors by aCoustic Optimalisation 1st Int. Meet.
on Wind Turbine Noise: Perspectives for Control
[13] Drela M 1989 XFOIL: An analysis and design system for low Reynolds number airfoils Conf.
on Low Reynolds Number Aerodynamics University Notre Dame
[14] Hansen M O L 2000 Aerodynamics of Wind Turbines (James & James)