You are on page 1of 14

Received: 19 August 2016 Revised: 20 December 2017 Accepted: 19 January 2018

DOI: 10.1002/we.2180

RESEARCH ARTICLE

A Beddoes-Leishman–type model with an optimization-based


methodology and airfoil shape parameters
Abdulqadir Aziz Singapore Wala1,2 Eddie Y. K. Ng1 Srikanth Narasimalu1

1 Nanyang Technological University, Singapore

2 Lloyd's Register Global Technology Centre, Abstract


Singapore
Floating offshore wind turbines operate in a highly unsteady environment; thus, many flow tran-

Correspondence sients occur at the blade cross-sectional level, which affect the rotor aerodynamics. In every rotor
Abdulqadir Aziz Singapore Wala, Lloyd's aerodynamics modelling technique requiring the blade element theory, the blade cross-sectional
Register Global Technology Centre and aerodynamics need to be predicted accurately on the basis of the flow conditions. At reduced fre-
Nanyang Technological University, Singapore.
quencies of 0.01 and greater, the flow unsteadiness can be considered significant and cannot be
Email: qadir.singaporewala@lr.org
treated as quasisteady. Floating offshore wind turbines can be expected to consistently operate
Funding information in some degree of yaw or pitch, which may result in reduced frequencies greater than 0.01 over
EDB-IPP-LR-JIP, Grant/Award Number: most of the blade when operating at rated wind speeds and rotor RPM. The Beddoes-Leishman
RCA-13/056
model is a comprehensive but complex model for predicting unsteady airfoil aerodynamics, con-
taining 8 dimensionless time constants. In the present study, the Beddoes-Leishman model was
compared with experimental results of 10 different airfoil profiles, each performed under a range
of Reynolds numbers, motion frequencies, mean, and amplitudes of angle of attack. An optimiza-
tion was performed for all time constants in the model, the results of which were used to formulate
a simplified model with fewer equations, without any reduction in accuracy. Further, optimizations
were performed against the experimental results of each airfoil, and the optimized constants were
compared with shape parameters of the airfoils, yielding possible correlations, which were then
applied in the simplified Beddoes-Leishman model to yield improved accuracy, measured as a 5%
reduction in accumulated error between experimental and predicted coefficients of lift.

KEYWORDS

Beddoes-Leishman, dynamic stall, floating offshore wind turbines, unsteady aerodynamics,


unsteady airfoil

1 INTRODUCTION

The blade element momentum method can be divided into the rotor forces and flow field momentum balance, which need to be solved simultane-
ously. In most analysis methods, the rotor forces are modelled, by using a lookup table to determine the blade cross-sectional forces based on the
relative wind conditions at the blade station. This method of analysis can be expected to work in quasisteady cases with good accuracy. In float-
ing offshore wind turbines, the unsteadiness arising from floating platform motions due to wave and wind forces would result in periodic motions
of the rotor, meaning that the flow field cannot be considered quasistatic. In unsteady flows, these lookup tables, based on static aerodynamics
experiments, cannot accurately reflect the unsteady airfoil aerodynamics, which would need to be modelled on the basis of its static aerodynamic
behaviour.
Unsteady airfoil aerodynamics comprises the attached flow phenomenon, or Theodorsen effect and the postseparation phenomena. The
attached flow phenomenon is an inviscid flow phenomenon caused by vortex shedding, which increases circulation in the wake. By Kelvin's theorem,
the total amount of circulation does not change; thus, the circulation around the airfoil must decrease, resulting in a reduction of lift. Since the effect
of this change in circulation is not instantaneous, there is a time lag between the change in flow conditions such as flow velocity or angle of attack
..............................................................................................................................................................................................................

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any
medium, provided the original work is properly cited and is not used for commercial purposes.
© 2018 The Authors Wind Energy Published by John Wiley & Sons, Ltd.

590 wileyonlinelibrary.com/journal/we Wind Energy. 2018;21:590–603.


10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 591

and the change in lift coefficient of the airfoil. The postseparation phenomena consist of the change in coefficient of lift due to the change in circula-
tion by Kelvin's theorem and the vortex lift due to the propagation of a leading edge vortex over the airfoil suction surface, in a phenomenon called
dynamic stall.
The Beddoes-Leishman model1 is widely used for the prediction of unsteady airfoil aerodynamics. It predicts the unsteady attached flow lift using
an indicial function with 4 constants. In addition, the impulsive lift component due to the fluid displacement around the moving airfoil is modelled
with a time decay. For the separated flow circulation lift, the position of flow separation is used in the equation for lift on a flat plate in a potential
Kirchoff flow. This is modelled on the basis of the leading edge pressure and a time lag in change of position of flow separation. Finally, the vortex lift
is modelled because of an increase in lift while the vortex propagates over the airfoil and a time decay when it leaves the trailing edge.
The use of 4 constants in unsteady attached flow lift assumes that the vortex shedding leading to the time lag between the physical and effec-
tive angles of attack occurs over 2 timescales. On the basis of a multivariate optimization of the constants in the Beddoes-Leishman model, the
2 timescales were found to be of similar magnitudes. Thus, they were successfully combined to reduce the number of equations required in the
Beddoes-Leishman model while maintaining the model accuracy.
The time constants present in the Beddoes-Leishman model are dependent on airfoil shape; however, few studies have correlated these to shape
parameters. The problem is that these correlations are complex, and difficult to form, since interactions between the time constants are highly
significant, with the result of an equation involving one time constant being used in the next equation. However, beginning with the values of the
constants generally optimized over all experimental conditions, it is possible to methodically develop correlations between the constants and airfoil
shape parameters. In this study, the shape parameters for the suction surface are used, on the basis of a general equation for a round-nosed, sharp
trailing-edged airfoil. 2

2 METHODOLOGY

2.1 Unsteady airfoil experimental results


The unsteady airfoil experiments were conducted by the Ohio State University, and the database of experimental results were available from NREL.3
Sinusoidal pitching motions were introduced to a range of wind turbine airfoils, each at several Reynolds numbers, 3 different frequencies, 2 different
amplitudes, and 3 mean angles of attack. In the present study, experimental data for 10 airfoils were used, each under 54 combinations of Reynolds
numbers, frequencies, amplitudes, and mean angles of attack, with a total of 540 data sets used. The airfoils and the experimental conditions chosen
for the present study are summarized in Table 1.

2.2 Beddoes-Leishman model


A nondimensionalized time, s, is defined for time step n.
2V∞ · t(n)
s(n) = , (1)
c
where V∞ is the freestream velocity, t(n) is the time at time step n, and c is the airfoil chord length.

2.2.1 Unsteady attached flow


The unsteady attached flow is computed by defining an effective angle of attack for the unsteady airfoil.1

𝛼E = 𝛼n − X(s) − Y(s), (2)

where X(s) and Y(s) are indicial functions of nondimensional time, s, defined as
( )
( ) ( ) ( ( )) −b1 · Δs(n)
X s(n) = X s(n−1) exp b1 Δs(n) + A1 · Δ𝛼(n) · exp (3)
2
( )
( ) ( ) ( ( )) −b2 · Δs(n)
Y s(n) = Y s(n−1) exp b2 Δs(n) + A2 · Δ𝛼(n) · exp (4)
2

TABLE 1 Experimental conditions and airfoils used in present study


Parameter Conditions

Airfoil LS(1)-0417MOD, LS(1)-0421MOD, NACA 4415,


S801, S809, S810, S812, S814, S815, S825
Reynolds number Low (∼ 0.75 × 106 ), medium (∼ 1.00 × 106 ), high(∼ 1.25 × 106 )
Frequency, Hz Low (∼ 0.60), medium (∼ 1.18), high (∼ 1.85)
Mean-peak amplitude, ◦ 5.5, 10
Mean 𝛼, ◦ 8, 14, 20
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
592 SINGAPORE WALA ET AL.

Δ refers to the change in the parameter from the previous time step. A1 , b1 , A2 , and b2 are constants. These were proposed by Leishman and
Beddoes1 and since applied in Pereira,4 Minnema,5 and Mert6 as shown in Table 2.
The values of X and Y describe the change in circulation over the airfoil in terms of a change in effective angle of attack. The distribution of X and
Y against angle of attack for the S809 airfoil oscillating at 1.18 Hz, with an amplitude of 10◦ , mean angle of attack of 14◦ , and Reynolds number of
1.0 × 105 , is shown in Figure 1.
The circulatory attached flow lift is then computed using

CLC = CL𝛼 · (𝛼E − 𝛼0 ), (5)

where CL𝛼 is the lift curve slope in the linear region of the lift curve and 𝛼0 is the angle of attack of zero lift. By the Kelvin theorem, the total circulation
must remain constant. This is possible by shed vorticity when there is a change in the angle of attack. The shed vorticity results in a reduction of the
effective change in angle of attack.
The impulsive lift is given by
( )
4 · K𝛼 · c Δup(n) − Δup(n−1)
I
CL(n) = 2
· − D(n) , (6)
V∞ Δt

where K𝛼 is a constant related to Mach number, set to 0.846 for a Mach number of 0.15, as described in Björck.7 up is the plunge velocity given by7

up = V∞ (Δ𝛼) + 𝛼(ΔV∞(n) ), (7)

where ΔV∞(n) is the change in freestream velocity from the previous time step and D(n) is the deficiency function given by
( ) ( ) ( )
−Δt Δup(n) − Δup(n−1) −Δt
D(n) = D(n−1) exp + exp . (8)
K𝛼 · TI Δt 2K𝛼 · TI

The impulsive lift can be described as an added-mass lift, as it is resultant from the reaction of the air accelerating around a plunging airfoil. Its
deficiency function acts as a source for the impulsive lift, with one component being a source term from the plunge velocity and the other a time decay
of the function. The coefficient of impulsive lift and its deficiency function, D, is shown for the S809 airfoil oscillating at 1.18Hz, with an amplitude
of 10◦ , mean angle of attack of 14◦ , and Reynolds number of 1.0 × 105 in Figure 2.
The total unsteady attached flow lift, or potential flow lift, is given by

CLP = CLC + CLI . (9)

TABLE 2 Values for A1 , b1 , A2 ,


and b2 proposed by Leishman
and Beddoes1 and applied in
Pereira,4 Minnema,5 and Mert6
A1 b1 A2 b2

0.3 0.14 0.7 0.53

0.8
X
Y
0.6
Change in effective angle of attack(°)

0.4

0.2

−0.2

−0.4

−0.6
0 5 10 15 20 25
Angle of attack(°)

FIGURE 1 Plot of X and Y against angle of attack ◦ for oscillatory pitching motion of S809 airfoil, at Re=71.0 × 105 , amplitude = 10◦ ,
frequency = 1.18 Hz and mean 𝛼 = 14◦ [Colour figure can be viewed at wileyonlinelibrary.com]
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 593

x 10−4
5 0.1
D
CiL

Induced coefficient of lift( CiL )


Deficiency function D
0 0

−5 −0.1
0 5 10 15 20 25
Angle of attack(°)

FIGURE 2 Plot of CLi and D against angle of attack ◦ for oscillatory pitching motion of S809 airfoil, at Re = 71.0 × 105 , amplitude = 10◦ ,
frequency = 1.18 Hz, and mean 𝛼 = 14◦ [Colour figure can be viewed at wileyonlinelibrary.com]

2.2.2 Unsteady trailing edge separation


The unsteady attached flow lift, as described by Hansen et al,8 requires defining the point of flow separation under steady-state conditions. This is
done using the equation for lift on a flat plate in a potential Kirchoff flow.9 The point of flow separation for an airfoil in steady-state flow is defined
at each angle of attack:
√ 2
⎛ CLst (𝛼) ⎞
f = ⎜2
st
− 1⎟ , (10)
⎜ CL𝛼 · (𝛼 − 𝛼0 ) ⎟
⎝ ⎠

where CLst is the steady-state lift.


fst is set to zero (fully separated flow) at angles of attack where the value of fst derived from Equation 10 does not retrieve the static lift coefficient:
( √ )2
1+ f st (𝛼)
CLst ≠ · (𝛼 − 𝛼0 ) . (11)
2

Using experimental lift curves, CL𝛼 is taken to be the maximum lift curve value from the set of measured lift coefficients:
( st )
CL (𝛼)
CL𝛼 = max . (12)
𝛼 − 𝛼0

The flow separation over an airfoil is dependent on pressure distribution, which is related to lift.8 There is a time lag between this pressure
response and the lift, and this time lag is modelled through a further time lag to the unsteady attached flow lift:

CL′P = CLP − DP . (13)

DP is a deficiency function given by


( ) ( )
−Δs(n) −Δs(n)
DP(n) = DP(n−1) exp P
+ (CL(n) P
− CL(n−1) )exp , (14)
TP 2 · TP

where TP is a time constant.


This time-lagged unsteady attached flow lift is used to find a new effective angle of attack:

CL′P
𝛼f = − 𝛼0 . (15)
CL𝛼

The position of flow separation based on the new effective angle of attack is looked up from the static data table:
( )
f ′ = f 𝛼f . (16)

The time lag between the unsteady pressure response and the position of flow separation is modelled to find a dynamic position of flow separation.

f dyn = f ′ − Df . (17)
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
594 SINGAPORE WALA ET AL.

Df is the deficiency function given by


( ) ( )
−Δs(n) −Δs(n)
Df(n) = Df(n−1) exp + (f FL(n) − f FL(n − 1))exp , (18)
Tf 2 · Tf
where Tf is a constant.
In Equation 17, the dynamic position of flow separation (with respect to the airfoil chord) is modelled with the assumption of a time lag between
dynamic flow separation and the flow separation on the basis of the unsteady pressure response on the airfoil. The flow separation based on the
unsteady pressure response on the airfoil is found using a lookup table of static positions of flow separation with the effective angle of attack derived
from the time-lagged unsteady attached flow lift. The time lag is modelled using the deficiency function in Equation 18 with a time constant of Tf .
The total dynamic lift with trailing edge separation is then found using the equation for lift on a flat plate in a potential Kirchoff flow9 :
( √ )2
1 + f dyn
CLf = · (𝛼E − 𝛼0 ) (19)
2

When f dyn = 0, this is set to


CLf = CLst (𝛼E ) . (20)

The circulation of an airfoil is linearly related to the lift produced, not angle of attack, since the lift and angle of attack do not share a linear
relationship when there is flow separation. Thus, Equations 6 to 25 are used as an initial estimate of CLf , to find the change in circulatory angle of
attack, Δ𝛼 c .
f f
CL(n) − CL(n−1)
Δ𝛼c = (21)
CL𝛼

Δ𝛼 c then replaces Δ𝛼 in Equations 3 and 4, and Equations 6 to 25 are repeated to find CLf .

2.2.3 Vortex lift


Vortex lift refers to the dynamic stall phenomenon, where leading edge separation causes the propagation of a vortex over the suction surface of
the airfoil, resulting in a momentary increase in lift. While the leading edge separation onset criterion has been discussed and has been based on the
potential flow lift exceeding a certain value,4 this criterion does not need to be applied explicitly and can instead be applied when the feed to the
vortex lift exceeds zero.7 The feed for the vortex lift is modelled using the difference between circulatory attached flow lift and dynamic lift with
trailing edge separation:
Δcv = cv(n) − cv(n−1) (22)

cv = CLC − CLf . (23)

The vortex lift then is


( ) ( )
−Δs(n) −Δs(n)
v
CL(n) v
= CL(n−1) · exp + Δcv · exp , (24)
Tv 2 · Tv

where Tv is a constant. This is set to zero if the feed to the vortex lift is negative, the angle of attack is increasing, or exceeds 50◦ .7
The total unsteady lift is then

CLdyn = CLI + CLf + CLv . (25)

The constants TP , Tf , and Tv have been proposed by Leishman and Beddoes1 for thin airfoils at different Mach numbers, starting from 0.3. In
addition, Mert6 performed an optimization of the constants against experimental data and suggested a set of values. Pereira4 proposed values on
the basis of an analysis of both these sets of constants. These are included in Table 3.
The values cited in Mert6 were based on optimizations to reduce the error between experimental and model-predicted coefficients of lift and
drag. The optimizations were performed using the method of moving asymptotes and involved experimental data both from NWTC Informa-
tion Portal3 and other sources. Because of the similarity in approach in performing an optimization to reduce the error between predicted and
experimental coefficients, these values were used as the start point for the optimization described in Section 3.

TABLE 3 Values for TP , Tf , and Tv from various sources


Source TP Tf Tv

Leishman & Beddoes1 (M = 0.3) 1.7 3 6


Mert6 0.8 5 2
Pereira4 1.5 5 6
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 595

2.2.4 Drag and moment components


The unsteady drag is modelled as an additional unsteady component to static drag, on the basis of parameters predicted during the computation of
dynamic lift.8
( √ )2 ( √ )2
( ) ⎛ 1 − f dyn 1 − f st (𝛼E ) ⎞⎟
CDdyn = CDst (𝛼E ) + CLdyn · (𝛼 − 𝛼E ) + CDst (𝛼E ) · ⎜ − . (26)
⎜ 2 2 ⎟
⎝ ⎠

The dynamic moment includes additional components on the basis of impulsive lift.7

dyn
CLI
CM st
= CM (𝛼E ) − . (27)
4

The equations for both coefficients of dynamic drag and dynamic moment consist of static coefficients and components found during derivation
of the dynamic lift coefficient. Thus, analysis of the error produced from predicting the dynamic lift coefficient would be sufficient in determining
the accuracy of the model. This description of the classic Beddoes-Leishman model will for the basis of the work presented in this paper.

3 OPTIMIZATION

The optimization of the Beddoes-Leishman model constants were performed using the Polak-Ribiére non-linear conjugate gradients method.10 This
is a local optimization algorithm, which is used in place of a global optimization to save computational time. The start point of the optimization used
values from Mert,6 as shown in Tables 2 and 3, and the objective function was minimization of the error:


k
( ( ) ( ))2
error = CLsim t(n) − CLexp t(n) , (28)
n=1

where the superscript sim refers to simulation results, exp refers to experimental results, and k is the total number of time steps in all cases.
The parameters optimized were A1 , b1 , A2 , b2 , K𝛼 , TP , Tf , and Tv . The constants proposed in Mert6 were first used to compute a reference error
and compared with the error from the optimized constants, presented in Table 4.
The optimization resulted in an 8.5% reduction in total error computed. The drawback of this form of optimization, over 540 sets of experimental
data, is the lack of completeness in the efficacy of the model. The optimization cannot be guaranteed to produce an improvement in all 540 sets
of data, and indeed, the optimized model shows an increase in error for 124 cases compared with the baseline model. This means that the optimal
model constants are dependent on each case. Furthermore, noise in the experimental results also affected the effectiveness of the optimization,
since the optimization was based on the accuracy of prediction at each data point.
This initial optimization, which will be called the general optimization in this paper, will form the baseline for comparison for further improvements
to the Beddoes-Leishman model presented in this paper.
The accumulated error for each experimental condition under each parameter shown in Table 1 was compared and shown in Tables 5 to 9.

TABLE 4 Model constants for reference and optimized cases


A1 b1 A2 b2 K𝜶 TP Tf Tv Error

Mert6 0.300 0.140 0.700 0.530 0.846 0.800 5.000 2.000 9.35
Optimized −0.200 0.504 0.736 0.431 0.314 0.555 5.747 2.867 8.55

TABLE 5 Accumulated error for reference and


optimized cases for each airfoil
Airfoil Mert6 Optimized Reduction, %

LS(1)-0417MOD 1.28 1.12 12.5


LS(1)-0421MOD 1.04 0.90 14.0
NACA 4415 0.49 0.49 1.8
S801 0.98 0.86 12.3
S809 0.83 0.77 7.0
S810 1.48 1.33 9.8
S812 0.70 0.69 0.3
S814 0.60 0.59 3.0
S815 0.78 0.75 3.8
S825 1.18 1.07 9.6
Total 9.35 8.55 8.5
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
596 SINGAPORE WALA ET AL.

TABLE 6 Accumulated error for reference and


optimized cases for each Reynolds number
Re Mert6 Optimized Reduction, %

Low 4.21 3.83 9.2


Medium 2.91 2.67 8.1
High 2.24 2.06 8.0
Total 9.35 8.55 8.5

TABLE 7 Accumulated error for reference and


optimized cases for each frequency
Frequency Mert6 Optimized Reduction, %

Low 2.81 2.49 11.1


Medium 2.88 2.59 10.2
High 3.67 3.48 5.2
Total 9.35 8.55 8.5

TABLE 8 Accumulated error for reference and


optimized cases for each amplitude angle of attack
Amplitude, ◦ Mert6 Optimized Reduction, %

5.5 4.37 3.95 9.7


10 4.98 4.60 7.6
Total 9.35 8.55 8.5

TABLE 9 Accumulated error for reference and


optimized cases for each mean
Mean, ◦ Mert6 Optimized Reduction, %

8 1.02 0.94 8.0


14 2.83 2.51 11.5
20 5.50 5.11 7.1
Total 9.35 8.55 8.5

The accumulated error for each of the conditions at each experimental parameter reveals that the optimization registered a reduction in error
for all conditions; thus, the optimization is not skewed away from any particular condition, and any distribution of cases where the optimized
constants performed worse than the reference ones would be likely random. It is notable in Table 5 that some airfoils such as LS(1)-0417MOD,
LS(1)-0421MOD, and S801 registered a reduction in error of greater than 10% while others such as NACA 4415, S812, S814, and S815 registered
less than 4% reduction in error. These differences suggest that the optimization might have been more favourable for some airfoils; however, the
airfoils with the lowest reductions in error already had the lowest error as well. This implies that the Beddoes-Leishman model itself is better suited
to those airfoils. These findings were further explored with an optimization of the simplified model described in Section 4.

4 SIMPLIFIED MODEL

The result of the optimization shown in Table 4 shows a small difference between the indicial constants b1 and b2 of 0.07, compared with the original
values, which have a difference of 0.39. This small difference possibly indicates that one timescale is sufficient to simulate the indicial response of
the effective angle of attack due to Kelvin's theorem, as shown in Equations 2 to 4. Equations 2 to 4 can thus be replaced with

( )
( ) ( ( )) −b1 · Δs(n)
𝛼E = 𝛼n − X s(n−1) exp b1 Δs(n) + A1 · Δ𝛼(n) · exp . (29)
2

The constants for the simplified model were optimized, with the optimized constants from the original model summarized in Table 4 used as the
start point. The start point for the value A1 in the simplified model was the sum of A1 and A2 in the optimized original model, and the start point for
b1 in the simplified model was the weighted average of b1 and b2 from the optimized original model, using A1 and A2 as weighting parameters.
The results of the optimization of the simplified model are tabulated in Table 10.
There was no change in error, and there was minimal variation in the constants K𝛼 , TP , Tf , and Tv . This indicates that the simplified model is
in fact similar to the original model. Furthermore, the start point of the optimization of the simplified model, using the weighted average of the
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 597

TABLE 10 Model constants for original optimized and simplified model (Equation (29)) cases
A1 b1 A2 b2 K𝜶 TP Tf Tv Error

Original optimized model −0.200 0.504 0.736 0.431 0.314 0.555 5.747 2.867 8.55
Simplified optimized model 0.570 0.430 … … 0.310 0.554 5.741 2.868 8.55
Weighted average start point 0.536 0.404 … … 0.314 0.555 5.747 2.867 …

TABLE 11 Values of 𝜏, A and 𝜏(A) for each airfoil


Airfoil 𝝉 A 𝝉A

LS(1)-0417MOD 0.299 0.205 0.061


LS(1)-0421MOD 0.350 0.201 0.070
NACA 4415 0.274 −0.182 −0.050
S801 0.222 −0.500 −0.111
S809 0.230 −0.164 −0.038
S810 0.191 −0.576 −0.110
S812 0.232 −0.400 −0.093
S814 0.259 −0.341 −0.089
S815 0.268 −0.326 −0.087
S825 0.227 −0.446 −0.101

optimized original model constants, was close to the final optimized result. This shows that the weighted average approach was an accurate method
of determining the constant, with differences due to the fact that the weighted average is not an exact method in combining exponents of an indicial
response. When comparing each experimental case, the most significant difference in error between the 2 models was 1.2%. This implies that the
simplified model behaves in a similar manner as the original model across all cases.

5 SHAPE PARAMETER–BASED MODEL

Most of the unsteady effects occur over the suction surface of the airfoil, since the suction effect results in most of the lift generated by an airfoil.
It would thus be a logical conclusion that the values of the various constants would depend primarily on the shape of the suction surface. For most
round-nosed, sharp trailing-edged airfoils, the suction surface shape can be described by (for x [chord-wise] and z-coordinates) 2 :
√ ( )
( ) ∑N ( )i
x x x
z=𝜏· · 1− · 1− Ai · , (30)
c c i=1
c

where c is the chord, 𝜏 is a thickness parameter, and Ai is the ith coefficient of the N-polynomial.
Using N = 1, the equation is simplified to

√ ( ) ( ( ))
x x x
z=𝜏· · 1− · 1−A· . (31)
c c c

In Equation 31, A is a curvature parameter that determines the general shape of the curve, while 𝜏 is a thickness parameter that determines the
thickness of the airfoil. Expanding the equation would show that 𝜏 A would also be an important parameter for the airfoil shape, which takes into
account both curvature and thickness. Unless the airfoils were designed using Equation 31, their profile would not precisely match the equation.
Instead, the parameters 𝜏 and A were chosen to obtain the best fitting curves to each airfoil's suction surface. These are summarized in Table 11.
The best-fit curves using Equation 31 for the airfoil suction surfaces are shown in Figure 3 for 3 of the airfoils.
Each airfoil in the experiment was placed under 54 different conditions. The constants K𝛼 , TP , Tf , and TV in the simplified Beddoes-Leishman model
were optimized, with the objective function as the error in Equation 28, using all of the experimental conditions for each airfoil. The values of A1
and b1 were kept constant, while the values of K𝛼 , TP , Tf , and TV were optimized for each airfoil. Then, the values for A1 and b1 were optimized
generally for all airfoils, in a successive optimization approach. This process was repeated until convergence, when the general optimization step
produced an error within 0.0001 of the total error from the airfoil-specific optimization step. This method was used over a more accurate and logi-
cal nested optimization technique where the airfoil-specific optimization is performed at each iteration of the general optimization because of the
lower computational time of a successive optimization technique. The assumption is that the airfoil-specific optimization step would produce a
much more significant reduction in error as compared with the general optimization step. The error for optimization of the constants for each air-
foil (airfoil-specific optimization) is compared with the error from the general optimization of constants (from Section 4) for all the airfoils at once
and included in Table 12.
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
598 SINGAPORE WALA ET AL.

0.1 L421
Approximated
0.05

z
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x
0.1
S809
Approximated
0.05

z
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

0.1 S815
Approximated
0.05
z

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

FIGURE 3 Best fit curves for LS(1)-0421MOD (top), S809 (middle), and S815 (bottom) airfoil suction surfaces using Equation 30 [Colour figure
can be viewed at wileyonlinelibrary.com]

TABLE 12 Accumulated error for each airfoil in generally


optimized and airfoil-specific optimized cases
Airfoil General Airfoil-specific Reduction, %

LS(1)-0417MOD 1.12 1.05 6.0


LS(1)-0421MOD 0.90 0.83 7.3
NACA 4415 0.49 0.45 8.2
S801 0.86 0.85 0.8
S809 0.77 0.72 5.8
S810 1.33 1.12 16.1
S812 0.69 0.65 6.1
S814 0.59 0.51 12.3
S815 0.75 0.73 2.8
S825 1.07 1.04 2.4
Total 8.55 7.96 7.0

1.2
Optimised
Best fit
1

0.8

TP= 72.8τ2− 45.1τ+7.0


TP

0.6
R2= 0.94

0.4

0.2

0
0.2 0.25 0.3 0.35
τ

FIGURE 4 Plot of TP against 𝜏 and quadratic regression [Colour figure can be viewed at wileyonlinelibrary.com]

These comparisons showed that 𝜏 has a strong quadratic correlation to the optimized value of the constant TP and A to TV . The best-fit quadratic
relationship between TP and 𝜏 was found to have a negative minimum point, which was adjusted to be positive to ensure numerical stability. These
are illustrated in Figures 4 and 5.
The values of K𝛼 and Tf were then optimized again for each airfoil, with the values of A1 , b1 from the previous optimization, and the quadratic cor-
relations between A and TP and 𝜏 and TV implemented in the general optimization step, in the successive optimization approach described previously.
The new optimized values of Tf had a good quadratic correlation to 𝜏 A as shown in Figure 6.
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 599

5 TV= 22.7A2+10.1A+2.8
R2= 0.87

Tv
4

2
Optimised
Best fit
1
−0.6 −0.4 −0.2 0 0.2
A

FIGURE 5 Plot of TV against A and quadratic regression [Colour figure can be viewed at wileyonlinelibrary.com]

7.5
Optimised
7 Best fit
6.5

5.5
Tf

4.5 Tf= − 260.7(τA)2− 21.3τA+6.3

4 R2= 0.87

3.5

2.5
−0.15 −0.1 −0.05 0 0.05 0.1
τA

FIGURE 6 Plot of Tf against 𝜏 A and quadratic regression [Colour figure can be viewed at wileyonlinelibrary.com]

0.8
Optimised
0.7 Best fit

0.6

0.5

0.4
Kα= − 64.4(τA)2−2.7τA+0.7
0.3 R2= 0.92 (corrected)

0.2

0.1

0
−0.15 −0.1 −0.05 0 0.05 0.1
τA

FIGURE 7 Plot of K𝛼 against 𝜏 A and quadratic regression, with anomalous points circled in red [Colour figure can be viewed at
wileyonlinelibrary.com]

The values of K𝛼 appeared to have a weak quadratic correlation to 𝜏 A, with an R2 value of 0.60 (before correction). However, 3 points, from the
LS(1)-0417MOD, S814, and S825 airfoils appeared to deviate much more from the best-fit curve than the rest. These points were removed from
the K𝛼 against 𝜏 A and the quadratic relation showed a closer fit to the remaining data points. This is shown in Figure 7, with the 3 points removed
for the best fit circled in red.
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
600 SINGAPORE WALA ET AL.

TABLE 13 Suction surface shape-dependent


values of constants in Beddoes-Leishman
model after final optimization
Constant Value

A1 0.93
b1 0.52
K𝛼 −64.4(𝜏A)2 − 2.7𝜏 A + 0.8
TP 72.8𝜏 2 − 45.1𝜏 + 7.0
Tf −260.7(𝜏A)2 − 21.3𝜏 A + 6.5
TV 22.5A2 + 9.3A + 2.3

TABLE 14 Accumulated error in each airfoil for generally


optimized and shape-based cases
Airfoil General Shape-based Reduction, %

LS(1)-0417MOD 1.12 1.07 4.8


LS(1)-0421MOD 0.90 0.85 5.8
NACA 4415 0.49 0.45 6.8
S801 0.86 0.86 0.03
S809 0.77 0.74 3.8
S810 1.33 1.14 14.3
S812 0.69 0.66 4.4
S814 0.59 0.56 5.3
S815 0.75 0.74 1.3
S825 1.07 1.07 0.01
Total 8.55 8.12 5.0

The quadratic relations for K𝛼 and Tf to 𝜏 A were added to those for TP to 𝜏 and TV to A, to form a suction surface shape-based model. This model,
together with the values for A1 and b1 , was optimized again, generally for all airfoils, to obtain the values and correlations shown in Table 13. This
new model based on the auction surface shape will be referred to as the shape-based model.
It is likely that this is not a true physical representation of the relationship, since the quadratic relationship would result in negative values of K𝛼
and Tf , and extremely large values of TP and TV as A, 𝜏 and 𝜏 A extend beyond the presently used ranges. It can be conjectured that these constants
should asymptotically tend to certain threshold values as the shape parameter for the airfoil increases or decreases, possibly in a Gaussian-type
function as a better representation of this relationship. At present, however, a quadratic relationship is used, since the range of shape parameters
is small.

6 RESULTS AND DISCUSSION

The error associated with the suction surface is compared to the error from the fixed-constant model seen in Table 10, for each airfoil as well as in
total, as shown in Table 14.
An improvement is noticeable in all cases, supporting the hypothesis of airfoil shape dependence of the model constants. The different changes
in model accuracy can be explained by the linear distance between the model constants in the shape-based model and general optimization and
shape-based model and airfoil-specific optimization. The linear distances are normalized against the general optimization parameter values, and
compared with the reduction in error from the general optimization to the shape-based model, and the increase in error from the airfoil-specific
optimization to the shape-based model in Figure 8.
The normalized distance describes the degree of change of the constants between the general optimization case and shape-based case as well as
the airfoil-specific case and shape-based case. The vertical axis describes the change in error between the cases compared for normalized distance.
Generally, the greater the change in constants, the greater the change in error observed. The airfoils whose parameters were closest to the optimum
after the general optimization showed very little reduction in model error. Similarly, parameters from the shape-based model that deviated the most
from the airfoil-specific optimization showed the greatest increase in error comparatively.
The differences between the shape-based model approach and the airfoil-specific optimizations can be attributed to the fact the shape depen-
dence of A1 and b1 was not included in this study; thus, these were generally optimized. In some airfoils, differences between suction surface shape
and best-fit using Equation 31 were significant and the fact that the pressure surface shapes were not considered in this study also contributed to
the error. An approach using both the suction and pressure surface shapes would be ideal, but would not be possible with just 10 airfoils as reference.
For thin airfoils, this study can be performed using the airfoil camber shape instead.
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 601

0.2
general
0.18 airfoil−specific
0.16

0.14

Change in error
0.12

0.1

0.08

0.06

0.04

0.02

0
0 0.5 1 1.5 2
Normalised distance

FIGURE 8 Plot of normalized distance between the model constants in the shape-based model and general optimization and shape-based model
and airfoil-specific optimization against change in error [Colour figure can be viewed at wileyonlinelibrary.com]

1.8 Static
Measured
1.7 General optimised
Shape−Based
1.6
Coefficient of lift (CL)

1.5

1.4

1.3

1.2

1.1

0.9

8 10 12 14 16 18 20
Angle of attack(α)

FIGURE 9 Coefficient of lift for oscillatory pitching motion of LS(1)-0421MOD airfoil, at Re = 7.5 × 105 , amplitude = 5◦ , frequency = 1.85 Hz, and
mean 𝛼 = 14◦ [Colour figure can be viewed at wileyonlinelibrary.com]

2 Static
1.8 Measured
General optimised
1.6 Shape−Based
Coefficient of lift (CL)

1.4

1.2

0.8

0.6

0.4

0.2

0
0 5 10 15 20
Angle of attack(α)


FIGURE 10 Coefficient of lift for oscillatory pitching motion of NACA 4415 airfoil, at Re = 7.5 × 105 , amplitude = 10 , frequency = 1.85 Hz, and
mean 𝛼 = 8◦ [Colour figure can be viewed at wileyonlinelibrary.com]
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
602 SINGAPORE WALA ET AL.

Static
2 Measured
General optimised
1.8 Shape−Based

Coefficient of lift (CL)


1.6

1.4

1.2

0.8

0.6
14 16 18 20 22 24 26
Angle of attack(α)

FIGURE 11 Coefficient of lift for oscillatory pitching motion of S801 airfoil, at Re = 7.5 × 105 , amplitude = 5◦ , frequency = 1.85 Hz, and mean

𝛼 = 20 [Colour figure can be viewed at wileyonlinelibrary.com]

2.5
Static
Measured
General optimised
2 Shape−Based
Coefficient of lift (CL)

1.5

0.5

5 10 15 20 25
Angle of attack(α)

FIGURE 12 Coefficient of lift for oscillatory pitching motion of S814 airfoil, at Re = 7.5 × 105 , amplitude = 10◦ , frequency = 1.85 Hz, and mean

𝛼 = 14 [Colour figure can be viewed at wileyonlinelibrary.com]

The shape-based model is compared with the general optimization for several experimental cases in Figures 9 to 12.
The shape-based model is able to better predict the lift hysteresis because of the oscillatory pitching motion of the airfoils. However, from the
lift predictions of the NACA 4415 in Figure 10 and S801 airfoil in Figure 11, there is still some significant difference between the experimentally
measured and predicted lift; thus, the Beddoes-Leishman model needs to be improved further for wind turbine applications.

7 CONCLUSION

A simplified Beddoes-Leishman model is proposed, using a 1-equation model for the unsteady attached flow circulation around the airfoil compared
with the traditional 2-equation model. It is found that the 1-equation model works with a similar level of accuracy as the 2-equation model; thus, the
number of equations and computational time in large simulations using the Beddoes-Leishman model may be reduced.
The constants of the simplified Beddoes-Leishman model were optimized against the measured coefficient of lift for 10 airfoils in 54 differ-
ent unsteady conditions each. The error between predicted and experimental values for this model was used as a baseline for comparison with
airfoil-specific optimization of the model and the shape-based model proposed. A 7% total reduction in total error for all 10 airfoils was observed
when the constants K𝛼 , TP , Tf , and TV were optimized for each airfoil.
The optimized constants for each airfoil were compared with airfoil suction surface shape parameters, found by using the best fit of Equation (31)
to the airfoil suction surface quadratic relations between K𝛼 and 𝜏(A), TP and 𝜏, Tf and 𝜏(A), and TV and A were found and implemented in a suction
surface shape–based Beddoes-Leishman model. This model had a 5% reduction in error compared with the general optimization case and showed
improvements in accuracy for all airfoils, although these varied generally on the basis of how much the shape-based model constants varied from
the general optimization.
10991824, 2018, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/we.2180, Wiley Online Library on [15/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
SINGAPORE WALA ET AL. 603

The present study explored the shape dependencies of 4 constants in the models and only used the suction surface shape as reference. Quadratic
relationships were found and used because of the limited number of airfoils in this study, although a Gaussian-type distribution is preferred for the
given shape parameters. Future attempts at improving this model should strive to include the pressure surface shape as well.
This study has shown the shape dependence of the different time constants in the Beddoes-Leishman model. To simplify this initial study, experi-
mental results from one database and of similar conditions were used. In addition, the shape equation used was truncated to 1 term for simplicity of
the study. A larger number of terms would improve the fit to the actual airfoil shape, but dilute the effectiveness of the study since many more data
sets would be needed to reasonably observe trends. Using machine learning methods such as support vector machines, larger and more varied sets
of experimental data and more complex shape equations can be included to develop a better understanding of these phenomena.

DISCLAIMER

The views expressed in this paper are those of the authors and do not necessarily reflect those of their affiliated companies and organizations.

ORCID

Eddie Y. K. Ng http://orcid.org/0000-0002-5701-1080

REFERENCES
1. Leishman JG, Beddoes TS. A generalised model for airfoil unsteady aerodynamic behaviour and dynamic stall using the indicial method. In: Proceedings
of the 42nd Annual Forum of the American Helicopter Society; 1986; Washington DC, USA. 243-266.
2. Kulfan BM, Bussoletti JE. “Fundamental” parametric geometry representations for aircraft component shapes. In: 11th AIAA/ISSMO Multidisciplinary
Analysis and Optimisation Conference; 2006; Portsmouth, Virginia, USA. 1-45.
3. NWTC Information Portal - OSU Wind-Tunnel Tests. http://wind.nrel.gov/airfoils/OSU_data/; 2012. Accessed October 28, 2015.
4. Pereira RBS. Validating the Beddoes-Leishman dynamic stall model in the horizontal axis wind turbine environment. Master's Thesis, Universidad Tecnica
de Lisboa; 2010.
5. Minnema JE. Pitching moment predictions on wind turbine blades using the Beddoes-Leishman model for unsteady aerodynamics and dynamic stall.
Master's Thesis, The University of Utah; 1998.
6. Mert M. Optimisation of semi-empirical parameters in the FFA-Beddoes dynamic stall model. Technical Report FFA TN 1999-37, Flygtekniska
Frsöksanstalten, Stockholm, Sweden; 1999.
7. Björck A. DYNSTALL: subroutine and package with a and dynamic stall model. Technical Report FFAP-V-110, Flygtekniska Frsöksanstalten,
Stockholm, Sweden; 2000.
8. Hansen MH, Gauna M, Madsen HA. A Beddoes-Leishman type dynamic stall model in state-space and indicial formulations. Technical Report
Risø-R-1354(EN), RisøNational Laboratory, Roskilde, Denmark; 2004.
9. Thwaites BE. Incompressible Aerodynamics. Oxford University Press, Oxford, UK; 1960.
10. Polak E, Ribiére G. Note sur la convergence de méthodes de directions conjuguées. ESAIM: M2AN. 1969;3(R1):35-43.

How to cite this article: Singapore Wala AA, Ng EYK, Narasimalu S. A Beddoes-Leishman–type model with an optimization-based
methodology and airfoil shape parameters. Wind Energy. 2018;21:590–603. https://doi.org/10.1002/we.2180

You might also like