Analysis of the Blade Element Momentum Theory

Jeremy Ledoux, Sebastián Riffo, Julien Salomon

To cite this version:

Jeremy Ledoux, Sebastián Riffo, Julien Salomon. Analysis of the Blade Element Momentum Theory.
SIAM Journal on Applied Mathematics, Society for Industrial and Applied Mathematics, 2021, 81
(6), pp.2596-2621. �10.1137/20M133542X�. �hal-02550763v2�

HAL Id: hal-02550763

https://hal.archives-ouvertes.fr/hal-02550763v2
Submitted on 1 Jul 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 ANALYSIS OF THE BLADE ELEMENT MOMENTUM THEORY
JEREMY LEDOUX∗ , SEBASTIÁN RIFFO∗ , AND JULIEN SALOMON†

Abstract. The Blade Element Momentum theory (BEM) introduced by C.N.H. Lock et al.
and formulated in its modern form by H. Glauert provides a framework to model the aerodynamic
interaction between a turbine and a fluid flow. This theory is either used to estimate turbine efficiency
or as a design aid. However, a lack of mathematical interpretation limits the understanding of some of
its issues. The aim of this paper is to propose an analysis of BEM equations. Our approach is based
on a reformulation of Glauert’s model which enables us to identify criteria to guarantee the existence
of solution(s), analyze the convergence of usual and new (and more efficient) solution algorithms and
study turbine design procedures. The mathematical analysis is completed by numerical experiments.

Key words. Turbine design, Blade Element Momentum theory, Computational Fluid Dynamics,
Geometry Modeling, Wind Turbine Aerodynamics, Fluid–Structure Interaction

AMS subject classifications. 76G25, 76M99, 65Z05

1. Introduction. Initially introduced to study propellers, the Blade Element

Momentum (BEM) theory is a model used to evaluate the performance of a propelling
or extracting turbine on the basis of its mechanical and geometric parameters as well
as the characteristics of the interacting flow. This model results from the combination
of two theories: the Blade Element Theory and the Momentum Theory. The former
was introduced by William Froude [19] in 1878 to study the turbines from a local
point of view. In this framework, the turbine blade is cut into sections, the blade
elements, each of them being approximated by a planar model. This approach results
in expressions of the forces exerted on the blade element, as functions of the flow
characteristics and blade geometry. The fundamental quantities of this model are
two experimental coefficients (usually denoted by CL and CD ), called lift and drag
coefficients, which account for the forces in the cross-section as functions of the angle
of attack, i.e. the relative angle between the rotating blade and flow. The results are
then integrated along the blade to obtain global values.
The Momentum Theory, also known as Disk Actuator Theory or Axial Momen-
tum Theory, was introduced by William J. M. Rankine in 1865 [39] and is, unlike the
Blade Element Theory, a theory that adopts a macroscopic point of view to model the
behavior of a column of fluid passing through a turbine. This approach was later taken
up independently first by Nikolay Joukowsky [26, 27] (see [50]), Frederick W. Lanch-
ester [28] and Albert Betz [5] to formulate Betz-Joukowsky Limit, which gives the
theoretical optimal efficiency of a thin rotor, see [49] and [37] for a description of the
three derivations obtained by these authors. A combination of these two approaches
was carried out in 1925 by C.N.H. Lock et. al. [29] and formalized in its modern
form in 1926 by Hermann Glauert [22], who also refined the Momentum Theory by
including the rotation of the fluid induced by its interaction with the turbine.
The resulting Blade Element Momentum theory is thus based on two decompo-
sitions: (i) a radial decomposition of the blades and the fluid column, considered as
concentric rings that do not interact with each other, and (ii) a decomposition of
the fluid/turbine system into a macroscopic part via Momentum Theory and a local

∗ CEREMADE, CNRS, UMR 7534, Université Paris-Dauphine, PSL University, France

([email protected],[email protected]).
† INRIA Paris, ANGE Project-Team, 75589 Paris Cedex 12, France and Sorbonne Université,

CNRS, Laboratoire Jacques-Louis Lions, 75005 Paris, France ([email protected]).

1
planar part via Blade Element Theory. Such a description of Glauert’s theory is given
in the monographs [7, 11, 23, 32, 40, 45, 46], see for example [23, p.56] :
“The Blade Element Momentum method couples the momentum theory with the
local events taking place at the actual blades. The stream tube introduced in the 1-D
momentum theory is discretized into N annular elements”
Though old, Glauert’s model is still currently used to evaluate turbine efficiency,
as indicated in [45, p.23]:
“Although a variety of correction models have been developed since then [...], the
momentum theory by Glauert still remains one [of the] the most popular.”
This longevity can be partly explained by the relative simplicity of the approach com-
pared to the complex phenomenon that develops in the coupled turbine/fluid system.
This time dependent 3D-fluid/structure interaction problem is a major modeling chal-
lenge, which BEM reduces to 0D computations with the help of 2D static data, namely
the above mentioned lift and drag coefficients. Indeed, these coefficients are obtained
by solving (2D-)partial differential equations, typically stationary Navier-Stokes, or,
more often than not, of using experimental data from wind tunnel profile tests. Let
us add that the numerical efficiency of this method is all the more crucial as tur-
bine models are mostly implemented as part of design procedures, through iterative
optimization loops. In practice, this means that equations have to be solved many
times and only simple formulations such as BEM allows the designer to carry out the
computations to a satisfactory point, or at least to provide a good enough initial guess
for a finer design, as pointed out in [34, p.2]:
“Blade element momentum theory continues to be widely used for wind turbine
applications such as initial aerodynamic analysis, conceptual design, loads and stability
analysis, and controls design.”
Note that other models have been proposed, based on the pioneering work of
Joukowsky [25], where the axial wake velocity is assumed to be constant. We refer
to [36, 48, 50] for an extensive presentation of this model with an historical perspective.
The significant increase in computational power as well as the theoretical ad-
vances obtained in the field of fluid-structure interaction simulation now also make it
possible to simulate 3D models based on the Navier-Stokes equation [3, 4, 24]. Al-
ternatively, BEM model has been combined with other approaches, as Lagrangian
stochastic solvers [6], multiple vortex cylinder models [8] or adapted scaling leading
to grid-based variant BEM for large rotors [30]. In view of numerical solving, BEM
equations have been reduced to one scalar equation in [35]. A summary of this ap-
proach is given in Section 3.1. Note finally that similar theories have been developed
to model vertical axis turbines (of Darrieus type) [15].
The purpose of this article is to analyze the Blade element momentum theory from
a mathematical point of view. The results we obtain in the course of this analysis
elucidate issues related to the well-posedness of the model, the numerical solution
of its equations and the optimality of a blade design. They concern two versions of
Glauert’s model, which we call Simplified model and Corrected model. The former
allows us to illustrate the main features of our approach, whereas the latter includes
some corrections usually considered to remedy the mismatch between the simplified
model and experimental observations.
The paper is structured as follows: Section 2 provides a brief exposition of the
derivation of the model. We then focus on the resulting algebraic system. The key
point of our analysis is to look deeper into Glauert’s macroscopic-local decomposi-
tion to reformulate these equations into a single equation containing two very distinct
terms: a universal term, independent of the turbine under consideration and asso-
2
ciated with the macroscopic part of the model, and an experimental term, which
depends on the characteristics of the blades and is associated with the local part
of the model. In this context, we show that solving the equations associated with
Glauert’s model actually means finding an angular value that makes these two terms
equal. This result agrees with some implicit conclusions reported before that were
not formalized mathematically and rather used for pedagogic purpose, such as in [32,
Figure 3.27, p.126]. In contrast, our analysis gives rise to new theoretical and numer-
ical results. In Section 3, our reformulation enables us to identify explicitly which
assumptions related to the turbine parameters can guarantee the existence of a solu-
tion. In addition, we obtain a classification of multiple solution cases based on the
modeling assumptions. In Section 4, we present the usual solution algorithms and
derive from our approach more efficient procedures. As already mentioned, BEM is
also used in turbine design, where the simplified model is included in a specific op-
timization method. We recall the details of the resulting procedure in Section 5 and
describe an optimization algorithm for the corrected model. We finally present some
numerical experiments in Section 6.
Our results are based on assumptions related to physical parameters, e.g., on
the coefficients CD and CL . We do not claim that these assumptions are necessary.
However, we treat CD and CL as generic functions (endowed with general properties),
hence it is often possible to find examples which make our assumptions optimal.
In what follows, we denote by R+ and R− the sets of positive real numbers and
non-positive real numbers, respectively.

2. The blade element momentum theory. In this section, we present the

model proposed by Glauert to describe the interaction between a turbine and a flow.
After having introduced the relevant variables, we recall the main steps of the rea-
soning leading to the equations of the model. We then detail the two versions of the
model considered in this paper.

2.1. Variables. The blade element momentum theory aims to establish alge-
braic relations that characterize the interaction between a flow and a rotating blade,
named turbine in what follows. In this way, Glauert’s model couples two descriptions:
a global macroscopic model that describes the evolution of fluids rings crossing the
turbine, and a local one, that summarizes in 2D the behavior of a section of a blade,
a blade element, under the action of the fluid.
The flow is supposed to be constant in time and incompressible. The latter
assumption implies that the flow velocities in the left and right neighborhoods of
the turbine have a same value U0 . We denote by U−∞ and U+∞ the upstream and
downstream velocities, respectively. Though not considered in this paper, tangential
velocity can also be studied. In particular, a jump of this variable caused by the
actuator disk is often reported and can be modeled in the framework of the momentum
theory [7, Chap. 9]. As the BEM model does not take into account interactions
between blade elements and assumes that Ω and U∞ are constant, we consider in this
paper a fixed blade element and a fixed value of the local speed ratio λr := UΩr
−∞
. Here,
r is the distance of the element to the rotation axis, with r ≤ R, where R is the radius
of the blade. In practical cases, the turbine works at constant Tip Speed Ratio (TSR):
Ω is indeed often controlled through the torque exerted by a generator in such a way
that the ratio TSR:= UΩR −∞
is kept constant for various values of U−∞ . It follows that
the value of λr associated with one element only depends on r. In the sequel, we
consequently use the variable λr to describe the location of a blade element.
3
 2.1.1. Macroscopic variables and BEM unknowns. Glauert’s model ulti-
mately consists of a system which links together three variables a, a0 and ϕ associated
with a ring of fluid. The two former are usually called the axial and angular induction
factor, respectively. They are defined by

U−∞ − U0 ω
(2.1) a := , a0 := ,
U−∞ 2Ω

where ω is the rotation speed of the considered ring of fluid. The angle ϕ is the
relative angle deviation (see [32, p.120]) of the ring, meaning that:

1−a
(2.2) tan ϕ = .
λr (1 + a0 )

For the sake of simplicity, and to emphasize their role of unknowns in Glauert’s model,
we omit the dependence of a, a0 , ϕ (and α in what follows) on λr in the notation.
2.1.2. Local variables. Let us denote by Urel the relative fluid speed (also
called apparent fluid speed ) perceived from this blade element while rotating. By
definition of ϕ, we have:

U0
(2.3) Urel = .
sin ϕ

This variable is not defined when ϕ = 0 and, as an intermediate quantity, will not
appear in the final model. However, the limit case ϕ → 0 is discussed in Section 2.4.
For a given blade profile, the lift and drag coefficients CL and CD are defined by
1 2 1 2
(2.4) dL = CL (α) ρUrel cλr dr, dD = CD (α) ρUrel cλr dr,
2 2
where ρ is the mass density of the fluid, dL and dD are the elementary lift and drag
forces applying to a blade element of thickness dr and of chord cλr .The parameter α
is called angle of attack and defined as the angle between the chord and flow direction,
hence satisfies the relation

(2.5) α = ϕ − γλr ,

where −π/2 < γλr < π/2 is the twist (also called local pitch) angle of the blade. The
parameters associated with a blade element are summarized in Figure 1.
The coefficients CL and CD correspond to the ratio between the lift and drag
forces and the dynamic force, i.e., the force associated with the observed kinetic
energy. They are determined by the profile of the blade. Once this one is fixed, the
main design parameters are cλr and γλr whose optimization is discussed in Section 5.
The coefficients CL and CD are assumed to be known as functions of α and
occasionally of Reynolds number (Re). The latter case is indeed rarely considered in
the monographs, where Re is assumed to be constant with respect to α as soon as
U−∞ , λr , Ω, cλr are fixed. For the sake of simplicity, we also neglect the changes in
Re in this paper. However, our results can be extended to non-constant Reynolds
numbers [46, p.374-375], i.e. in situations where the functions (α, Re) 7→ CL (α, Re)
and α 7→ CD (α, Re) have to be taken into account. In practice, Re is either assumed
to be known a priori and used to select the corresponding CL and CD , or dealt with
4
 Urel

dL α ϕ
dD
γλr

line
C hord

U0 Ωr

Figure 1: Blade element profile and associated angles, velocities and forces.

iteratively together with CL and CD to get a more accurate result. Examples of

variations of CL and CD with respect to Re are given in [11, p.169].
Though changing from one profile to another, the behaviors of CL and CD as
functions of α can be described qualitatively in a general way. The coefficient CL
usually increases nearly linearly with respect to α up to a given critical angle αs ,
with 0 < αs < π/2, where the so-called stall phenomenon occurs: CL then decreases
rapidly (see, e.g., [7, p.93-94] and [46, p.375]), causing a sudden loss of lift. For CD
is associated with a drag force, it is always positive and defined for all angles. This
coefficient usually slightly increases with α up to α = αs , and then becomes very
large. Though most designs do not prevent the inner part of the blade to generate
stall, the condition ϕ − γλr < αs is often considered in the blade design phase. Note
finally that the greater part of our analysis applies for angles of attack such that CL
is positive. The properties of CL and CD required for our analysis are summarized
in the following assumption.
Assumption 1. For some β ∈ R+ , the function α 7→ CL (α) is continuous on
Iβ := [−β, β], and positive on Iβ ∩R+ . The function α 7→ CD (α) is defined, continuous
and non-negative on R.
As a consequence, CL (0) is assumed to be positive, which is equivalent in practice to
the fact that the angle of attack corresponding to zero lift is negative. This assumption
is true for usual designs.
2.2. Glauert’s modeling. For the sake of completeness, we now shortly recall
the reasoning proposed by Glauert to model the interaction between a turbine and
a flow. We refer to [11, Chap. 3], for a more extended presentation of this theory.
We denote by dT and dQ the infinitesimal thrust and torque that apply on the blade
element of thickness dr under consideration.
2.2.1. Macroscopic approach. The first part of the model is related to the
Momentum Theory and deals with the macroscopic evolution of a ring of fluid. It
aims to express dT and dQ in terms of a, a0 and ϕ.
Denote by p− and p+ the fluid pressures on the left and right neighborhoods
of the blade, respectively. Applying Bernouilli’s relation between −∞ and 0− and
between 0+ and +∞ gives rise to p− − p+ = 21 ρ(U−∞2 2
− U+∞ ). Considering then the
rate of change of momentum on both sides of the turbine, we get a second expression
for the variation in the pressure, namely p− − p+ = ρ(U−∞ − U+∞ )U0 . Combining
5
the two previous equations and using (2.1), we obtain U+∞ = (1 − 2a)U−∞ . Since
dT = (p− − p+ )2πrdr and dQ = ωρU0 2πr3 dr, we finally get

2
(2.6) dT = CT (a)U−∞ ρπrdr,
(2.7) dQ = 4a0 (1 − a)λr U−∞
2
ρπr2 dr.

dT
where CT (a) := 1 2 = 4a(1 − a), is the local thrust coefficient [52].
2 U−∞ ρ2πrdr

2.2.2. Local approach. Another set of equations can be obtained via the Blade
Element Theory, where local expressions for infinitesimal thrust and torque are consid-
ered. The reasoning consists in combining the elementary lift and drag (2.4) expressed
in the rotating referential with (2.3). This gives

(1 − a)2 2
(2.8) dT = σλr (CL (α) cos ϕ + CD (α) sin ϕ) U−∞ ρπrdr,
sin2 ϕ
(1 − a)2 2
(2.9) dQ = σλr (CL (α) sin ϕ − CD (α) cos ϕ) U−∞ ρπr2 dr,
sin2 ϕ

Bcλr
where σλr := 2πr , with B is the number of blades of the turbine.

2.2.3. Combination of local and global approaches . To get a closed system

of equations, Glauert combined the results of the two last subsections. More precisely,
equating (2.6) and (2.7) with (2.8) and (2.9), respectively, using (2.5) and dividing
both resulting equations by 4(1 − a)2 gives

a σλr
(2.10) = (CL (ϕ − γλr ) cos ϕ + CD (ϕ − γλr ) sin ϕ) ,
1−a 4 sin2 ϕ
a0 σλr
(2.11) = (CL (ϕ − γλr ) sin ϕ − CD (ϕ − γλr ) cos ϕ) .
1−a 4λr sin2 ϕ

The system obtained by assembling (2.2), (2.10) and (2.11) is the basis of Glauert’s
Blade Element Momentum theory.

2.3. Simplified model. In the monographs devoted to aerodynamics of wind

turbines, the contribution of CD is sometimes set to zero. This point is discussed
in [45, p.135], where it is particular stated:
“Since the drag force does not contribute to the induced velocity physically, CD is
usually omitted when calculating induced velocities.”.
In the same way, Manwell and co-authors mention in [32, p.125]:
“In the calculation of induction factors,[...] accepted practice is to set CD equal to
zero [...]. For airfoils with low drag coefficients, this simplification introduces negligible
errors.”
This assumption is actually justified in many cases, since the procedures used to design
profiles minimize their drag. As a matter of fact, the usual blade design procedure
starts by selecting a twist angle γλr minimizing the ratio C CL , see Section 5.1.
D

We also consider the case where CD = 0, and referred to as simplified model in the
following. In view of (2.2), (2.10) and (2.11), it corresponds to the three equations:
6
 1−a
(2.12) tan ϕ = ,
λr (1 + a0 )
a 1
(2.13) = µL (ϕ) cos ϕ,
1−a sin2 ϕ
a0 1
(2.14) = µL (ϕ),
1−a λr sin ϕ
σ
where we have introduced the dimensionless function µL (ϕ) := λ4r CL (ϕ−γλr ), which
is defined on Iβ,γλr := [−β + γλr , β + γλr ], by virtue of Assumption 1.
2.4. Corrected model. To get closer to experimental results, many modifi-
cations of the model (2.2,2.10,2.11) have been introduced, see e.g. [45, Chapter 7].
Hereafter, we present three important corrections, namely non-zero drag coefficient
CD , tip loss correction and a specific treatment of large values of a. The first and the
last will modify significantly the analysis developed for the simplified model.
2.4.1. Slowly increasing drag. In addition to consider CD strictly positive,
we shall assume in some parts of the analysis a slow increasing of this parameter from
0 up to the occurrence of the stall phenomenon.
2.4.2. Tip loss correction. The equations of Momentum Theory are derived
assuming that the turbine can be modeled as an actuator disk. Such a framework cor-
responds to a rotor with an infinite number of blades. However, in real life situations,
a modification of the flow at the tip of a blade has to be included to take into account
that the circulation of the fluid around the blade must go down (exponentially) to
zero when r → R, where R is the turbine radius. In this way, Glauert (see [21, p.268])
introduced an approximation of the Prandtl tip function Fλr [38] (see also [7, p.240]):
!
B/2(1 − λrΩR U−∞  
2 ) 2 B/2(1 − r/R)
Fλr (ϕ) := arccos exp(− λr U−∞ ) = arccos exp(− ) ,
π ( ) sin ϕ π (r/R) sin ϕ
ΩR

as a supplementary factor in (2.6) and (2.7). This modification gives rise to

2
(2.15) dT = 4a(1 − a)Fλr (ϕ)U−∞ ρπrdr,
(2.16) dQ = 4a0 (1 − a)Fλr (ϕ)U−∞ ρπr3 Ωdr.

Note that some authors include the tip loss factor only in (2.6), see [13]. The results
of our paper can readily be adapted to this version of the model. Further models
of tip loss correction have been introduced in between. We refer to [7, Chap. 13],
and [41] for reviews of other models. An alternative approach based on Extended
vortex theory has also been proposed in [53].
2.4.3. Correction for high values of a. For induction factors a larger than
about 0.4 (see [46, p.297]), a turbulent wake usually appears, and it is broadly con-
sidered that momentum theory does not apply. This fact was already reported by
Glauert (see [20]), who proposed to modify CT (a) in (2.6) when a becomes larger
than a given threshold ac . Subsequently, many other expressions have been proposed
to fit better with experimental data, see [7, Section 10.2.2]. All these variants lead to
a new expression for dT that reads
2
(2.17) dT = 4 (a(1 − a) + ψ ((a − ac )+ )) Fλr (ϕ)U−∞ ρπrdr.
7
where (a − ac )+ := max{0, a − ac } and ψ is a given function defined on R+ . Some
corrections are presented via the function ψ in Table 1. Glauert’s empirical correction
is obtained by combining experimental data and the constraints CT (ac ) = 4ac (1 − ac )
and CT (1) = 2. This leads to a discontinuity at a = ac when Fλr (ϕ) 6= 1. Buhl
proposed in [10] a modification to fix this issue.

Authors ac ψ((a − ac )+ )
(a − ac )2+
 
(a − ac )+
Glauert [21],[23, p.53] 1/3 + 2(a − ac )+ + ac
4 ac
Empirical Glauert 
1

2/5 ≈ 2(1−ac ) − ν4 (1 − a) (a − ac )+ with ν = 5.5563
[17, 10]
 2
1 (a − ac )+
Buhl [10] 2/5
2Fλr (ϕ) 1 − ac
Wilson et al.,
1/3 (a − ac )2+
Spera [46, p.302]

Table 1: Various corrections that can be introduced in the relation between the thrust
coefficient and the induction factor (2.15) in case of turbulent state.

2.4.4. On 3D effects. In order to take into account 3D effects of the rotating

blade, correction formulas have been proposed [42, 16, 12] (see the extensive review
in [44]). As an example, the following expression can be used to correct CL :
cλ
CLc = CL + a( r )b ∆CL ,
r
where a ∈ [2, 3] and b ∈ [1, 2] are constants and ∆CL = CL,inv − CL . In the latter
expression, CL,inv is the lift obtained by considering an inviscid flow, i.e., by solving
a Laplace problem. A similar expression can be used for the drag coefficient CD . For
simplicity, we do not consider 3D effects in our study, but all the following results
can be readily adapted by replacing CL and CD by their corrected expressions, as
done in the next section with tip-loss correction. Note finally that the 3D correction
presented in [16] is included in the model considered in Section 6.2.
2.4.5. Corrected system. We repeat the reasoning used to obtain (2.12–2.14),
that is, we equalize (2.17) and (2.16) with (2.8) and (2.9), respectively. We get:
1−a
(2.18) tan ϕ = ,
λr (1 + a0 )
a 1 ψ ((a − ac )+ )
(2.19) = (µcL (ϕ) cos ϕ + µcD (ϕ) sin ϕ) − ,
1−a sin2 ϕ (1 − a)2
a0 1
(2.20) = (µcL (ϕ) sin ϕ − µcD (ϕ) cos ϕ),
1−a λr sin2 ϕ
σ
where we have introduced the dimensionless functions µcL (ϕ) := 4Fλλr(ϕ) CL (ϕ − γλr ),
r
σ
and µcD (ϕ) := 4Fλλr(ϕ) CD (ϕ − γλr ), defined on Iβ,γλr and R, respectively. The cor-
r
rected and simplified models coincide when Fλr (ϕ) = 1, ac = 1 and CD = 0.
Remark 2.1. Combining (2.20) with (2.18), one obtains
a0 1
(2.21) = (µc (ϕ) sin ϕ − µcD (ϕ) cos ϕ),
1 + a0 sin ϕ cos ϕ L
8
which is sometimes considered instead of (2.20) to define a0 .

3. Analysis of Glauert’s model and existence of solution. In this section,

we reduce each of the two previous versions of Glauert’s model to a single scalar
equation. With a view to obtaining existence results, this leads us to formulate
assumptions related to the characteristics of the turbine. To simplify notation, we
introduce the angle θλr ∈ (0, π2 ) defined by

1
(3.1) tan θλr := ,
λr
and the intervals
π π
(3.2) I := Iβ,γλr ∩ (− + θλr , + θλr ), I + := I ∩ (0, θλr ].
2 2
We comment about these definitions in the next section.
3.1. Simplified model. In the setting of the simplified model, a reformulation
of (2.12–2.14) can be obtained after a short algebraic manipulation.
Theorem 3.1. Suppose that Assumption 1 holds and that (ϕ, a, a0 ) ∈ I − {0, π2 } ×
R − {1} × R − {−1} satisfies (2.12–2.14). Then ϕ satisfies

(3.3) µL (ϕ) = µG (ϕ),

where µG (ϕ) := sin ϕ tan(θλr − ϕ). Reciprocally, suppose that ϕ ∈ I − {0, π2 } sat-
isfies (3.3) and define a and a0 as the corresponding solutions of (2.13) and (2.14),
respectively. Then (ϕ, a, a0 ) ∈ I − {0, π2 } × R − {1} × R − {−1} satisfies (2.12–2.14).
Note that (3.3) appears – up to a factor – in [32, p.128, Fig. 3.85a]. Some concrete
examples of graphs of µL and µG are given in Section 6.1.
We have excluded the angles ϕ = π2 and ϕ = 0 for the sole reason that (2.12–
2.14) are not defined for these angle values. However, ϕ = 0 is naturally associated
with the case a = 1, as it appears in (2.13). On the other hand, the value ϕ = π2 ,
that belongs to I if β + γλr > π2 > −β + γλr , is neither a solution of (2.12–2.14)
nor of (3.3): setting this value in the former system leads indeed to a0 = −1, a = 0
and −λr = µL ( π2 ) which corresponds to a negative lift, hence contradicts CL ≥ 0 on
Iβ ∩ R+ in Assumption 1. As a matter of fact, µG is well-defined at these values,
which generally do not give rise to solution of (3.3). Finally, note that the right-
hand side of (3.3) is not defined in the values ϕ = ± π2 + θλr . However, they do not
correspond to any solution of (2.12–2.14): inserting them in the last equations leads
to ∓λr = 0.µL (± π2 ) which contradicts λr = UΩr −∞
> 0. For all these reasons, the
formulation (3.3) will be considered on the whole interval I in the rest of this paper.
Proof. Suppose that (ϕ, a, a0 ) ∈ I−{0, π2 } × R − {1} × R − {−1} satisfies (2.18–
2.20). Eliminating a and a0 in (2.12) using (2.13) and (2.14), we get:

1 + a0
 
−1 cos ϕ 1
tan ϕ = λr = λr 1 + 2 µL (ϕ) + sin ϕ µL (ϕ).
1−a sin ϕ
so that (3.3) follows from the definition (3.1) of θλr . Repeating these steps backward
ends the proof of the equivalence.
This result shows that Glauert’s model – here in its simplified version – essentially
boils down to one scalar equation. Indeed, suppose that ϕ satisfies (3.3), then a and
9
a0 can be post-computed thanks to (2.13–2.14). A similar conclusion has also been
obtained by Ning in [34], but his approach results in another equation that is
sin ϕ cos ϕ
(3.4) − = 0,
1 − a λr (1 + a0 )
where a and a0 read as functions of ϕ by (2.19) and (2.21), respectively. Note that (3.3)
and (3.4) have the same singularities, namely 0 and π/2. However, our approach gives
rise to a specific physical interpretation and mathematical analysis. More precisely,
an important property of (3.3) is that its left-hand side corresponds to the local de-
scription of the problem related to Blade Element Theory, whereas its right-hand side
is related to the macroscopic modeling arising from Momentum Theory. As a conse-
quence, µG reads as a universal function of fluid-turbine dynamics depending only on
θλr and related to Momentum theory. On the contrary, µL reads as a function which
strictly depends on the turbine under consideration, i.e., on its design parameters γλr
or σλr as well as its 2D experimental data, through CL , hence, rather relates to Blade
element theory. In this view, (3.3) is in line with the approach considered by Glauert.
In the same way, the two intervals defining I, namely Iβ,γλr and (− π2 + θλr , π2 + θλr )
play similar roles in the local and in the macroscopic descriptions as they correspond
to the domains of definition of µG and µL , respectively, whereas Iβ,γλr ∩ R+ and
(0, θλr ], whose intersection I + corresponds to positive lift in the two descriptions.
The formulation given in Theorem 3.1 can be used to establish criteria to ensure
existence of solution of (3.3): existence indeed holds as soon as the graphs of µG and
µL intersect. As an illustration, we give a simple condition in the case of symmetric
profiles. We express the assumptions in terms of µL to make it coherent with the
formulation (3.3) ; they can however easily be formulated in terms of CL and σλr .
Corollary 3.2. In addition to Assumption 1, suppose that γλr ≤ θλr , that the
profile under consideration is symmetric with γλr > 0, and that
(3.5) µG (max I + ) ≤ µL (max I + ),
where max I + := min{θλr , β + γλr }, see (3.2). Then (3.3) admits a solution corre-
sponding to a positive lift in [γλr , max I + ]. Moreover, if max I + = θλr , i.e. θλr ≤
β + γλr , then (3.5) is necessarily satisfied.
Proof. Since γλr > 0 and β > 0, max I + is well defined. Because of Assumption 1,
CL and consequently µL are continuous. As we consider a symmetric profile, we have
σ
µL (γλr ) = λ4r CL (0) = 0 whereas µG (γλr ) > 0. Because of Inequality (3.5), the
existence of solution of (3.3) in [γλr , max I + ] then follows from Intermediate Value
Theorem. Since µG ≥ on [γλr , max I + ], the resulting lift is positive.
Suppose finally that max I + = θλr . Assumption 1 guarantees that µL is positive
on [γλr , θλr ]. Since µG (θλr ) = 0, the last assertion follows.
In the case where µL is supplementary assumed to be increasing on [γλr , β + γλr ],
then, the solution defined in Theorem 3.2 is unique.
3.2. Corrected model. We now consider the corrected model defined by (2.18–
2.20), for a given value ac ∈ (0, 1). The algebraic manipulations performed in the
previous section to get Theorem 3.1 cannot be pushed as far as with the simplified
model and the resulting formula still contain the unknown a. Hence, before stating a
reformulation of this model and an existence result, we need to clarify the dependence
of a on the variable ϕ. Again, we express our assumptions in terms of µcL and µcD ,
but the translation in terms of CL , CD , σλr and Fλr (ϕ) is straightforward.
10
 In all this section, we suppose that 0 ∈ I, i.e. |γλr | ≤ β, which means in particular
that I + = (0, min{θλr , β + γλr }].
Lemma 3.3. Assume that Assumption 1 holds and define, for ϕ ∈ I +
µcD (ϕ)
g(ϕ) := tan−1 ϕ tan(θλr − ϕ) + 1 + tan−1 ϕ tan(θλr − ϕ) .


(3.6)
sin ϕ
Let ψ be one of the functions given in Table 1, with Fλr (ϕ) = 1 in the case of Glauert’s
empirical correction. Then, the equation
a sin θλr sin ϕ ψ ((a − ac )+ )
(3.7) + = g(ϕ)
1 − a cos(θλr − ϕ) (1 − a)2
defines a continuous mapping τ : ϕ ∈ I + 7→ a ∈ [0, 1).
Moreover, if g is decreasing and µcD differentiable on I + , then τ is decreasing and
differentiable for all ϕ ∈ I + with a possible exception of one point ϕc , which satisfies
τ (ϕc ) = ac .
Because of Assumption 1, the function µcD is non-negative and defined for all
angles in concrete cases so that g is well defined on I + . The only obstruction for g
to be decreasing would come from this term. Indeed, CD typically increases as angle
increases from zero, hence our assumption in Section 2.4.1. But for usual profiles, its
variations are negligible when compared to the other (decreasing) terms in (3.6).
Proof. For simplicity of notation, let us rewrite (3.7) under the form
(3.8) u(a) + v(ϕ)w(a) = g(ϕ),
sin θ sin ϕ
with u(a) := 1−a a
, v(ϕ) := cos(θλλr −ϕ) , w(a) := ψ((a−a c )+ )
(1−a)2 . Let us first consider the
r
left-hand side of (3.8). We see that u is positive and increasing on [0, 1) as well as w
for any function ψ given in Table 1 (with Fλr (ϕ) = 1 in the case of Glauert’s empirical
correction). In the same way, v is positive and increasing on (0, θλr ], hence on I + .
Fix now ϕ ∈ I + , it is fairly easy to see that the mapping a ∈ [0, 1) 7→ u(a) + v(ϕ)w(a)
is continuous, strictly increasing, strictly positive and goes from 0 to +∞. Since g
is bounded and assumed to be positive on I + , there exists an only a in [0, 1) such
that (3.8) holds. Hence the existence of the mapping τ .
Suppose now that g is decreasing and µcD differentiable. If we set aside the
function w in the point a = ac , all the functions involved in (3.8) are differentiable.
Consider ϕ ∈ I + , such that τ (ϕ) 6= ac . Differentiating (3.8) with respect to ϕ gives
0 0
τ 0 (ϕ) = u0g(τ(ϕ)−v (ϕ)w(τ (ϕ))
(ϕ))+v(ϕ)w0 (τ (ϕ)) . Combining the fact that g is decreasing with the above
properties of v, w, u and their derivatives implies that τ 0 (ϕ) ≤ 0. As a consequence,
the mapping τ is decreasing and differentiable either on I + , or on I + − {ϕc } where
ϕc is the unique value in I + such that τ (ϕc ) = ac . The result follows.
Remark 3.4. The quantity a = τ (ϕ) can generally be computed explicitly provided
that the function ψ is specified analytically as, e.g., in Table 1. In these cases, the
computation consists in solving a low order polynomial equation (in a).
We can now state a result similar to Theorem 3.1 in the case of the corrected model.
Theorem 3.5. Let Assumption 1 hold and ψ be one of the functions given in
Table 1, with Fλr (ϕ) = 1 in the case of Glauert’s empirical correction. Suppose also
that (ϕ, a, a0 ) ∈ I + − { π2 } × [0, 1) × R − {−1} satisfies (2.18–2.20). Then ϕ satisfies
(3.9) µcL (ϕ) − tan(θλr − ϕ)µcD (ϕ) = µcG (ϕ),
11
where
cos θλr sin2 ϕ ψ ((τ (ϕ) − ac )+ )
(3.10) µcG (ϕ) := µG (ϕ) + .
cos(θλr − ϕ) (1 − τ (ϕ))2
Reciprocally, suppose that ϕ ∈ I + − { π2 } satisfies (3.9) and define a and a0 by
(3.11) a =τ (ϕ),
1 − τ (ϕ) c
(3.12) a0 = (µL sin ϕ − µcD cos ϕ).
λr sin2 ϕ
Then (ϕ, a, a0 ) ∈ I + − { π2 } × [0, 1) × R − {−1} satisfies (2.18–2.20).
We refer to Section 6.1 for concrete examples of graphs of µcG and ϕ 7→ µcL (ϕ) −
tan(θλr − ϕ)µcD (ϕ). As was the case with the simplified model, the value ϕ = π2 is
excluded only for the technical reason that (2.18) is not defined for this angle.
Proof. Thanks to Lemma 3.3, τ is well defined on I + . Let (ϕ, a, a0 ) ∈ I + −
0
{ π2 } × [0, 1) × R+ satisfying (2.18–2.20). Because of (2.18), tan−1 ϕ = λr 1+a
1−a =
0 0
a a a a
λr (1 + 1−a ) + λr 1−a . In this equation, the terms 1−a and λr 1−a can be eliminated
thanks to (2.19) and (2.20), respectively. After some algebraic manipulations, we end
cos θλr sin2 ϕ ψ ((a − ac )+ )
up with µcL = (sin ϕ + µcD ) tan(θλr − ϕ) + , hence (3.9).
cos(θλr − ϕ) (1 − a)2
c
Using this formula to eliminate µL in (2.19) gives (3.7). Consequently, Lemma 3.3
implies that a and ϕ satisfy (3.11). Finally, (3.12) follows from (3.11) and (2.20).
Suppose now that ϕ ∈ I + − { π2 } satisfies (3.9), and let (a, a0 ) defined by (3.11–
3.12), meaning that a ∈ [0, 1). Replacing τ (ϕ) by a in (3.12) gives immediately (2.20).
Combining (3.11) with the definition of µcG gives (2.19). Finally, (2.18) is obtained
by introducing a and a0 into (3.9), where ϕ 6= π2 implies that a0 6= −1.
As in the simplified model, Glauert’s model boils down to a scalar equation in ϕ.
However, formulation (3.9) does not completely decompose the terms into a local part
and macroscopic modeling part: much as the left-hand side of (3.9) still only relies
on the turbine the right-hand side now also depends on it via τ , since (3.7) includes
µcD . Before going further, let us give more details about the behavior of τ in ϕ = 0.
Lemma 3.6. Let Assumption 1 hold, µcD (0) 6= 0 and ψ be one of the functions
given in Table
q 1, with Fλr (ϕ) = 1 in the case of Glauert’s empirical correction. Then
ψ(1−ac ) 3/2
τ (ϕ) = 1 − µcD (0) ϕ + oϕ=0 (ϕ3/2 ).
Proof. Thanks to Lemma 3.3, τ is well defined on I + . Let us first prove that
lim
ϕ→0+
τ (ϕ) = 1− . From (3.6), we see that ϕ→0
lim g(ϕ) = +∞. Given ϕ ∈ I + , we have
+

cos θ cos ϕ sin θ sin ϕ

a = τ (ϕ) ∈ [0, 1) and 1 − cos(θλλr −ϕ) = cos(θλλr −ϕ) ≥ 0, so that all the terms of the
r r
left-hand side of (3.7) are positive. As a consequence, the only possibility for the sum
of these terms to go to +∞ is that ϕ→0 lim+ τ (ϕ) = 1− .

Define now ν(ϕ) = 1 − τ (ϕ). Expanding (3.7) in a neighborhood ϕ = 0+ , we get

1 ψ ((1 − ac )+ ) + oϕ=0 (1) µcD (0) 1
−1+(tan θλr .ϕ + oϕ=0 (ϕ)) 2
= tan θ λ r 2
+oϕ=0 ( 2 ),
ν(ϕ) ν(ϕ) ϕ ϕ
which implies:
ν 2 (ϕ) ϕ2
 
(3.13) − ϕ2 − (tan θλr .µcD (0) + oϕ=0 (1)) = − tan θλr ψ ((1 − ac )+ ) .
ϕ3 ν(ϕ)
12
 + +
Let (ϕn )n∈N a sequence satisfying lim lim
n ϕn = 0 , so that n ν(ϕn ) = 0 . Suppose
2/3
ϕ2 ϕ3
 
ν 2 (ϕn )
that lim
n ϕ3 = +∞. Since ν(ϕnn ) = ν 2 (ϕnn ) ν 1/3 (ϕn ), this sequence goes to
n
zero. Back to (3.13), we find a contradiction since the left-hand side goes to +∞
whereas the right-hand side is constant. It follows that, up to a subsequence, we can
ν 2 (ϕn )
assume that lim
n ϕ3 = ` for a certain `. Setting ϕ = ϕn in (3.13) and passing to
n
ψ((1−ac )+ )
the limit n → +∞, we obtain that ` = .
The result follows.
µcD (0)
p
Remark 3.7. If µcD (0) = 0, we obtain τ (ϕ) = 1 − ψ(1 − ac )ϕ + oϕ=0 (ϕ1/2 ).
σ
The quantity µcD (0) = λ4r CD (−γλr ) has no specific physical meaning in the
applications. We have introduced it as a constant (that can be expressed explicitly),
for simplicity of presentation. As a matter of fact, ϕ = 0 is a specific angle from the
macroscopic point of view, as appears when considering µG (that cancels in 0) and
µcG , see the proof of the next result.
We are now in a position to give an existence result about the corrected model.
Corollary 3.8 (of Theorem 3.5). Suppose that Assumption 1 holds and that
(3.14) µcG (max I + ) ≤ µcL (max I + ) − tan(θλr − max I + )µcD (max I + ).
Then (3.9) admits a solution in I + corresponding to a positive lift. Moreover, if g is
decreasing, max I + = θλr and ϕc < θλr , where ϕc is defined in Lemma 3.3, then (3.14)
is necessarily satisfied.
Proof. Because of Lemma 3.6 and Definition (3.10) of µcG , we get µcG (ϕ) ≈ϕ→0+
µcD (0) lim µc (ϕ) = +∞. This implies that there exists a small enough ϕ0 > 0
ϕ , so that ϕ→0 + G
such that µG (ϕ0 ) ≥ µcL (ϕ0 ) − tan(θλr − ϕ0 )µcD (ϕ0 ). Because of the assumption (3.14),
c

the existence of a solution of (3.9) then follows from Intermediate Value Theorem.
The positivity of µcG on I + implies that the resulting lift is positive.
Suppose now that g is decreasing, max I + = θλr and ϕc < θλr . Lemma 3.3
implies that τ is decreasing so that the correction associated with ψ is not anymore
active on [ϕc , θλr ). We then have µcG (max I + ) = µG (max I + ) = µG (θλr ) = 0 whereas
µcL (max I + ) − tan(θλr − max I + )µcD (max I + ) = µcL (max I + ) ≥ 0. Hence (3.14).
Unlike the simplified model, no condition on γλr or µcL (γλr ) is assumed in Corol-
lary 3.8 , but the alternative assumption 0 ∈ I is required. This makes the corrected
model much better posed than its simplified version.
Remark 3.9. In the case µcD (0) = 0, similar reasoning gives µcG (ϕ) ≈ϕ→0+ (1 +
tan θλr )ϕ. As a consequence, µcG (0) = 0, so that, as in the simplified model, one needs
an assumption about, e.g., µcL (γλr ) to get an existence result similar to Corollary 3.2.
3.3. Multiple solutions. The results of the previous sections can be completed
by some additional remarks about cases of multiple solution. More precisely, these
cases can be sorted into three independent categories:
1. Multiple solutions in the simplified model: since ϕ→θlimλr ±π/2
µG (ϕ) = −∞, there
shall be two intersections between the graphs of µG and µL , e.g. in the case
where µL is affine on a large enough interval, CL (0) = 0 and γλr ∈ (0, θλr ].
In this case, one of the two roots gives rise to a negative lift.
2. Multiple solutions caused by stall: as mentioned in Section 2.1.2, the stall
phenomenon is generally associated with a sudden decrease in CL . It follows
that if the stall angle αs satisfies αs + γλr ∈ I, the graph of µL shall cross the
graph of µG at an angle in ϕ ≥ αs + γλr . This fact is reported in [32, p.129]:
13
 “In the stall region [...] there may be multiple solution for CL . Each of these
solutions is possible. The correct solution should be that which maintains the
continuity of the angle of attack along the blade span.”
lim+ µc (ϕ) = +∞, the graph
3. Multiple solutions in the corrected model: since ϕ→0 G
of µcG may no longer be concave on I + when a correction for large values of
a is active. Hence possible multiple solution, e.g. in the case µL is affine.
Concrete examples of these three types of multiple solution are given in Section 6.
4. Solution algorithms. To solve numerically Glauert’s model, a specific fixed-
point approach is often highlighted in the literature. In this section, we recall its main
features and introduce more efficient procedures.
4.1. Standard fixed-point procedure . Solving the simplified or the corrected
model is usually done by a dedicated fixed-point procedure that comes in two versions,
see [7, 23, 32, 43, 45]1 or the early presentation in [52, p.47]. This procedure is given

Algorithm 4.1 Solving BEM system, Standard fixed-point (versions 1 and 2)

Input: Tol > 0, α 7→ CL (α), α 7→ CD (α), λr , γλr , σλr , Fλr , x 7→ ψ(x).
Initial guess: a, a0 .
Output: a, a0 , ϕ.
Set err := Tol + 1.
while err > Tol do 
1−a
(1) Set ϕ := atan λr (1+a0) .
(2) Set a as the solution of (2.19).
(3) Set a0 as the solution of (2.20) (version 1) or of (2.21) (version 2)
1−a
(4) Set err := tan ϕ − λr (1+a 0) .

end while

in Algorithm 4.1, where the stopping criterion is arbitrary and usually not mentioned
in monographs. The convergence of these algorithms is problematic. Instabilities are
often observed in practice, as reported, e.g., in [43]:
“Note that this set of equations must be solved simultaneously, and in practice,
numerical instability can occur.”
“When local angle of attack is around the stall point, or becomes negative, getting
the BEM code to converge can become difficult.”
We also refer to [31] for a specific study of some convergence issues. The analysis of
the algorithm is tedious ; we refer to Appendix for an example of setting where the
convergence of Version 1 is guaranteed.
4.2. Optimized fixed-point procedures. To cure the convergence issues of
observed when using the standard fixed-point procedure, various alternative fixed-
point procedures have been proposed in the last decade. A Newton-Raphson pro-
cedure have been studied numerically in [33]. As usual with Newton’s iteration,
this method outperforms the standard fixed-point in case of convergence. However,
this approach fails to converge in some regimes, which can be described numeri-
cally in terms of λr and σ. Sun et al. proposed to modified the standard fixed-
point by introducing a relaxation
 term,
 i.e., replace the step (1) in Algorithm 4.1 by
1−a
ϕ := (1 − w)ϕ + w atan λr (1+a0 ) , for w ∈ (0, 1]. In case of non-convergence, w is

1 Version 1 is mentionned in [7] and [32].

14
divided by 2. This procedure is tested numerically in [47]. Thanks to our new formu-
lation, we propose now alternative fixed-point procedures whose convergence can be
guaranteed in some cases.
4.2.1. General formulation. In view of (3.9), we consider now optimized fixed-
point procedures based on the iteration

(4.1) ϕk+1 =f (ϕk ),

with f (ϕ) = ϕ + ρ(ϕ)Res(ϕ) , where Res(ϕ) := µcG (ϕ) − µcL (ϕ) + tan(θλr − ϕ)µcD (ϕ)
and ρ(ϕ) > 0 is a given relaxation coefficient. The procedure is summarized in
Algorithm 4.2. The parameter of ρ(ϕ) can be optimized to obtain a robust version of

Algorithm 4.2 Solving BEM system, Optimized fixed-point procedure

Input: Tol > 0, α 7→ CL (α), α 7→ CD (α), λr , γλr , σλr , Fλr , x 7→ ψ(x).
Initial guess: ϕ.
Output: ϕ.
Set err := Tol + 1.
while err > Tol do
Compute a := τ (ϕ), i.e. the solution of (3.7).
Compute µcG (ϕ), using (3.10) and f (ϕ).
Set err := |f (ϕ) − ϕ|.
Set ϕ = f (ϕ).
end while

this procedure or a Newton procedure.

4.2.2. Robust version. Defining
ε
(4.2) ρ(ϕ) = ρε (ϕ) := c 0 + 1 + tan2 θ

c
max {0, −µcG 0 (ϕ)} + max
+
µ L λr µD (ϕ)
I

leads to a robust algorithm, whose convergence is guaranteed when ψ = 0 (or when a

remains below ac ).
Theorem 4.1. Suppose that max I + = θλr , ψ = 0, and that Assumption 1 holds.
Assume also that µcL and µcD are continuously differentiable on I + , with µcL and µcD
non-decreasing and that (2.12–2.14) admit a solution in I + . Given ε ∈ (0, 1) the
sequence (ϕk )k∈N defined by (4.1) with ρ(ϕ) = ρε (ϕ) defined in (4.2) and the initial
value ϕ0 = θλr converges to ϕ? , the largest solution of (3.9) in I + .
Proof. The assumption on µcL guarantees that the denominator in (4.2) is strictly
positive on I + so that ρε (ϕ) is well-defined and positive on this interval. Since µcG =
µG is concave, ϕ 7→ ρε (ϕ) is decreasing on I + . Since Res(ϕ) ≤ 0 on [ϕ? , θλr ], we get:

f 0 (ϕ) =1 + ρε (ϕ) µ0G (ϕ) − µcL 0 (ϕ) − (1 + tan2 (θλr − ϕ))µcD (ϕ) + tan(θλr − ϕ)µcD 0 (ϕ)

+ ρε 0 (ϕ)Res(ϕ)
 
0 c0 2

c
≥1 − ρε (ϕ) max {0, −µG (ϕ)} + max
+
µL + 1 + tan θλr µD (ϕ) = 1 − ε ≥ 0,
I

so that f is increasing on [ϕ? , θλr ]. Since µcL is non-decreasing, we have

f (θλr ) = θλr + ρε (θλr ) (µG (θλr ) − µcL (θλr )) = θλr − ρε (θλr )µcL (θλr ) ≤ θλr .
15
These results and f (ϕ? ) = ϕ? , imply that f ([ϕ? , θλr ]) ⊂ [ϕ? , θλr ]. Since ϕ0 = θλr ,
(ϕk )k∈N is bounded and decreasing, hence converges. The result follows.
In some cases, we can estimate the rate of convergence of (ϕk )k∈N .
Theorem 4.2. In addition to the assumptions of Theorem 4.1, suppose that
(4.3) tan θλr (1 + max
+
µcD 0 ) < min
+
µcL 0 .
I I
k
Then the sequence (ϕ )k∈N defined by (4.1) satisfies
k
µcL 0 − tan θλr (1 + max µcD 0 )

min+ +
I I
|ϕk −ϕ? | ≤ 1 −  |θλr −ρε (θλr )µcL (θλr )|.
max µcL 0 + sin θλr + (1 + tan2 θλr )µcD (θλr )
I+

Proof. For we have already shown in the previous proof that f 0 (ϕ) ≥ 0 on
[ϕ , θλr ], it remains to determine an upper bound for f 0 (ϕ). To do this, we use
?

the bound (4.3) and µ0G (ϕ) ≤ µ0G (0) = tan θλr to get:
f 0 (ϕ) = 1 + ρε µ0G (ϕ) − µcL 0 (ϕ) − (1 + tan2 (θλr − ϕ))µcD (ϕ) + tan(θλr − ϕ)µcD 0 (ϕ)

min
+
µcL 0 − tan θλr (1 + max
+
µcD 0 )
I I
≤1− ,
max µcL 0 + sin θλr + (1 + tan2 θλr )µcD (θλr )
I+

where we have used max {0, −µ0G (ϕ)} ≥ −µ0G (θλr ) = sin θλr to bound ρε from below.
The result is then obtained by induction.

4.2.3. A Newton version. One can actually obtain quadratic convergence, i.e.
k
|ϕk − ϕ? | ≤ δ|ϕ0 − ϕ? |2 for some δ > 0 by using a Newton procedure, i.e., setting
1
ρ(ϕ) := − .
µcG 0 (ϕ) − µcL 0 (ϕ) − (1 + tan2 (θλr − ϕ))µcD (ϕ) + tan(θλr − ϕ)µcD 0 (ϕ)
and by choosing ϕ0 close enough to ϕ? . In this formula, the functions µcL and µcD are
usually only known experimentally, i.e. pointwise. In practice, splines or polynomial
interpolation are used to evaluate CL and CD for any arbitrary angle. In this way, the
derivatives µcL 0 and µcD 0 (of the extension) can be obtained without any supplemen-
tary computational cost. The term µcG 0 (ϕ) can also be computed without significant
additional cost as soon as µcG (ϕ) has been computed, see Remark 3.4.
4.3. Root-finding algorithms. Reducing (2.18–2.20) to a one dimensional
equation allows the application of usual root-finding algorithms such as bisection. In
this way, Ning [34] used Brent’s procedure [9] to solve (3.4). Using (3.3), a new root-
finding approach consists in applying Brent’s procedure to the equation Res(ϕ) = 0.
If a correction for high values of a is considered, then the framework of Corollary 3.8
implies that there exists a solution of the corrected model in I + . In this case, the
convergence of bisection algorithm or Brent’s procedure is guaranteed. Moreover, the
solution found in the case ψ = 0, e.g., by Algorithm 4.2, can be used to bracket the
solution in a finer way than I + .
Lemma 4.3. Keep the assumptions of Corollary 3.8, and denote by ϕ0 a solu-
tion (3.9) where ψ = 0. Then (3.9) admits a solution in (ϕ0 , min{θλr , β + γλr }]
corresponding to a positive lift.
cos θλr sin2 ϕ0 ψ ((τ (ϕ0 ) − ac )+ )
Proof. Since ψ = 0, (3.9) implies Res(ϕ0 ) = − ,
cos(θλr − ϕ0 ) (1 − τ (ϕ0 ))2
so that Res(ϕ0 ) ≤ 0. The Intermediate Value Theorem and (3.14) give the result.
16
 5. Optimization. The BEM model does not only aim to evaluate the efficiency
of a given geometry, but also provides a framework to design rotors, that is, to select
high-performance parameters γλr and cλr . In this way, monographs often consider
a specific maximization procedure of a functional Cp , called power coefficient ([32,
p.129, (3.90a)], [21, p.328]), which corresponds to the ratio between the received
and the captured energy. This quantity is usually defined by Cp (γλr , cλr , ϕ, a, a0 ) =
R λr max
λr min
Jλr (γλr , cλr , ϕ, a, a0 )dλr , where the elementary contribution Jλr reads:

8Fλr (ϕ)λr 3 0
 
CD (ϕ − γλr )
(5.1) Jλr (γλr , cλr , ϕ, a, a0 ) := a (1 − a) 1 − tan−1
ϕ .
λr 2max CL (ϕ − γλr )
in which the variables satisfy the constraints (2.18–2.20). The drag coefficient CD
is consequently taken into account (though partly neglected in the reasoning, as ex-
plained hereafter) as well as the tip loss correction. On the contrary, no correction
related to high values of a is considered. This motivates the description of an op-
timization algorithm for the full corrected model in Section 5.2. In any case, the
contributions are independent. As a consequence, we focus on the optimization prob-
lem associated with one element, i.e., we fix the value of λr and optimize Jλr .
5.1. Simplified model and usual optimum approximation. The usual op-
timization procedure is described in, e.g., [32, p.131-137]. For the sake of completeness,
we recall it in the case where Fλr = 1.
Considering independently each λr on a discretization grid associated with the in-
terval [λr min , λr max ] and the corresponding functional Jλr (γλr , cλr , ϕ, a, a0 ), the proce-
dure starts by determining an angle α which minimizes the ratio C D (α)
CL (α) . In the follow-
C (ϕ−γ )
ing steps, the coefficient CD is neglected: not only the factor 1− CD λr
L (ϕ−γλr )
tan−1 (ϕ) is
set to 1 in (5.1), but CD is also set to 0 in the constraints, which correspond to the sim-
plified model (2.12–2.14) afterwards. Using Theorem 3.1 to replace µL by µG in (2.13–
2.14), a, a0 , and consequently Jλr are expressed exclusively in terms of ϕ, namely
sin ϕ cos(θλr − ϕ) 0 sin ϕ sin(θλr − ϕ) 3 sin2 ϕ sin(2(θ
λr −ϕ))
a = 1− ,a = and Jλr = λ8λ 2
r
sin 2θλr .
sin θλr cos θλr r max

As a consequence, it remains to optimize ϕ 7→ sin2 ϕ sin(2(θλr − ϕ)) on [0, θλr ]. It is

easily seen that the maximum is attained at ϕ∗ = 32 θλr . Finally, γλ∗r := γλr (ϕ∗ ) and
c∗λr := cλr (ϕ∗ ) can be computed from (2.5) and (3.3), which gives
8πrµG (ϕ∗ )
(5.2) γλ∗r := ϕ∗ − α, c∗λr := .
BCL (α)
5.2. A gradient method for the corrected model. We now detail an adjoint-
based gradient method to tackle the optimization of γλr and cλr in the framework of
the corrected model. Throughout this section, CL0 and CD 0
denote the derivatives of
CL and CD . We omit in the notation the dependence of µcL and µcD on ϕ and cλr .
We first recall how the introduction of Lagrange multipliers enables to compute
the gradient of Jλr . Define the Lagrangian of Problem (5.1) by
 
1−a
Lλr (ϕ, a, a0 , p1 , p2 , p3 , cλr , γλr ) = Jλr (ϕ, a, a0 , cλr , γλr ) − p1 tan ϕ −
λr (1 + a0 )
 
a 1 c c ψ ((a − ac )+ )
− p2 − (µ cos ϕ + µ sin ϕ) +
1 − a sin2 ϕ L D
(1 − a)2
 0 
a 1
− p3 − (µc sin ϕ − µcD cos ϕ ,
1 − a λr sin2 ϕ L
17
where p1 , p2 and p3 are the Lagrange multipliers associated with the constraints (2.18–
2.20). The optimality system is obtained by canceling the partial derivatives of Lλr .
Differentiating Lλr with respect to p1 , p2 and p3 and equating the resulting terms to
zero gives the corrected model (2.18–2.20), that can be solved using the algorithms
presented in Section 4. Canceling the derivatives of Lλr with respect to (ϕ, a, a0 ) gives

(5.3) M · p = b,

where p := (p1 p2 p3 )> is the Lagrange multiplier vector, and

 
1
cos2 ϕ vϕ wϕ
 
1 1+ψ 0 ((a−ac )+ ) 2ψ((a−ac )+ ) a0
− +
 
M := 
 λr (1+a0 ) (1−a)2 (1−a)3 (1−a)2 

 
1−a 1
λr (1+a0 )2 0 1−a
∂µcL ∂µcD
2 cos ϕ(µcL tan−1 ϕ + µcD ) − (( ∂ϕ + µcD ) cos ϕ + ( ∂ϕ + µcL ) sin ϕ)
vϕ := 2
sin ϕ
−1 ∂µcD ∂µcL
2 cos ϕ(µcL − µcD tan ϕ) − ((µcL − ∂ϕ ) cos ϕ + (µcD + ∂ϕ ) sin ϕ)
wϕ :=
λr sin2 ϕ
 0
 0 0

a (1−a) CL (ϕ−γλr )CD (ϕ−γλr )−CD (ϕ−γλr )CL (ϕ−γλr ) CD (ϕ−γλr )
+ sin2 ϕ
8Fλr (ϕ)λ3r   CL (ϕ−γλr ) CL (ϕ−γλr ) tan ϕ 
C (ϕ−γλr )
b := 0
−a (1 − CD tan−1 ϕ)

λ2max 
 
L (ϕ−γλr ) 
C (ϕ−γλr )
(1 − a)(1 − CD L (ϕ−γλr )
tan−1 ϕ)
 
CD (ϕ−γλr )
a0 (1 − a)(1 − CL (ϕ−γλr ) tan−1 ϕ)
0 3 
8F (ϕ)λr  
+ λr2 .

0
λmax 


0

Fix now the values of the pair (γλr , cλr ) and set ϕ, a, a0 , p as the corresponding solu-
tions of (2.18–2.20) and (5.3), respectively. The gradient ∇Jλr (γλr , cλr ) reads
 >
∂Lλr ∂Lλr
(5.4) ∇Jλr (γλr , cλr ) = ,
∂γλr ∂cλr

where
∂Lλr 8Fλr (ϕ)λr 3 0 0
CD (ϕ − γλr )CL (ϕ − γλr ) − CL0 (ϕ − γλr )CD (ϕ − γλr )
= a (1 − a)
∂γλr λr 2max CL2 (ϕ − γλr ) tan ϕ
c c
1 ∂µ ∂µD 1 ∂µcL ∂µcD
− p2 2 ( L cos ϕ + sin ϕ) − p3 ( sin ϕ − cos ϕ),
sin ϕ ∂ϕ ∂ϕ λr sin2 ϕ ∂ϕ ∂ϕ
∂Lλr 1 ∂µc ∂µcD 1 ∂µc ∂µcD
= p2 2 ( L cos ϕ + sin ϕ) + p3 2 ( L sin ϕ − cos ϕ).
∂cλr sin ϕ ∂cλr ∂cλr λr sin ϕ ∂cλr ∂cλr

The associated optimization procedure is then formalized with Algorithm 5.1.

6. Numerical experiments. In this section, we test the performance of the al-
gorithms presented in Section 4 on a practical cases and tackle the design optimization
problem considered in Section 5 in the case of an actual wind turbine.
18
Algorithm 5.1 Numerical optimization
Input: Tol > 0, κ > 0, α 7→ CL (α), α 7→ CD (α), λr , x 7→ ψ(x).
Initial guess: γλr , cλr .
Output: γλr , cλr .
Set err := Tol + 1.
while err > Tol do
Set ϕ, a, a0 as the solutions of (2.18–2.20).
Set p as the solution of (5.3).
 ∇J
Compute thegradient  λr (γλr , cλr ) given by (5.4).
γλr γλr
Update = + κ∇Jλr (γλr , cλr ),
cλr cλr
Set err := k∇Jλr (γλr , cλr )k.
end while

6.1. Example of a small river turbine. This example is related to the project
HyFloEFlu, which was devoted to the design of a river turbine adapted to the Garonne
river in Bordeaux, France. We consider a turbine of radius R = 1.1m, consisting
of three blades, designed with a unique profile, namely NACA 4415. In the case we
study, the TSR is close to 3, which corresponds for example to U−∞ =1.5 m.s−1 and
ω = 2/3.2πs−1 . The functions CL and CD have been obtained using truncated Fourier
representations of data provided by the free software Xfoil [14], with Re = 9.105 .
Remark 6.1. Note that for such a value, Xfoil sometimes fails to predict airfoil
lift and drag accurately because a very simple model is then used to describe the tran-
sition from laminar to turbulent flow in the airfoil boundary layer [51]. Hence, this
test must be considered as an experiment to test the solution algorithms rather than
an accurate estimate of the turbine efficiency.
The first step of the usual design procedure presented in Section 5.1 gives α =
0.2215 rad. Plots of CL and CD are given in Figure 2.

Figure 2: Lift and Drag coefficients CL and CD as functions of α, for Re = 9.105 .

We use the correction of Wilson et al and Spera, meaning that ac = 1/3, see
Table 1. We first focus on three different blade elements associated with λr,1 =
0.5, λr,2 = 1.75, and λr,3 = 3, respectively. In these three cases, we either set
(γλr , cλr ) = (γλ∗r , c∗λr ), i.e. the optimal values of the simplified model given by (5.2)
19
 λr γλ∗r γλc r c∗λr ccλr ϕc (γλ∗r , c∗λr ) ϕc (γλc r , ccλr )
λr,1 = 0.5 0.516627 0.520195 1.429701 0.255311 (0.715856,0.996060) 0.707018
λr,2 = 1.75 0.124625 0.123680 0.318397 0.200768 0.343358 0.343095
λr,3 = 3 -0.006972 -0.052160 0.045246 0.066050 0.213839 0.214025

Table 2: Optimal values of the twist γλr and the chord cλr for various values of λr ,
with (γλr , cλr ) = (γλ∗r , c∗λr ) and (γλr , cλr ) =: (γλc r , ccλr ).

or (γλr , cλr ) =: (γλc r , ccλr ), i.e. the optimal values of the corrected model. The former
typically corresponds to the first step of an optimization procedure, where (γλ∗r , c∗λr )
is used as an initial guess. The latter is computed with Algorithm 5.1 and corresponds
typically to one of the last steps of an optimization process. These values, as well as
the associated ϕc are given in Table 2, whereas corresponding graphs of the functions
µcLD : ϕ 7→ µcL (ϕ) − tan(θλr − ϕ)µcD (ϕ), µL , µcG and µG are presented in Figure 3.
For these two blade geometries, I + = (0, θλr ], ϕc < θλr for all elements. However,
the function g is non-decreasing in few cases, as, e.g. when λr = 0.5 and (γλr , cλr ) =
(γλ∗r , c∗λr ). In this case, the last statement of Lemma 3.3 does not apply which explains
the existence of two values ϕ1c and ϕ2c where the graphs of µcG and µG merge or
separate. In the other cases, Corollary 3.8 applies, which is confirmed by the plots.
We observe multiple solution of type (1) (see Section 3.3) in the case of the simplified

Figure 3: Graphs of the functions µL , µcLD , µcG and µG , with (γλr , cλr ) = (γλ∗r , c∗λr )
(top) and (γλr , cλr ) = (γλc r , ccλr ) (bottom), for λr = 0.5 (left), λr = 1.75 (middle),
λr = 3 (right). The values of these parameters are given in Table 2. Note that these
figures are similar to the scheme given in [32, Figure 3.27, p.126].

model. As for the full corrected model, we always have a unique solution.
To compare the efficiency of the solution algorithms we consider the corrected
model and measure the number of iterations (k) required to solve accurately (3.9)
in the sense that Res(ϕk ) ≤ Tol = 10−10 . We use this stopping criterion in all
our tests, instead of the respective definitions of err given in the algorithms. We
test the algorithms presented in Section 4, i.e. the two versions of the standard
fixed point, the robust and Newton versions of the optimized fixed point and the
20
root-finding algorithms. Remark that due to the choice of correction, iterations in
each algorithm have similar computational costs, namely, the solving of second order
polynomials corresponding to (2.19) when applying Algorithm 4.1 or Ning’s algorithm
and to (3.7) when applying Algorithm 4.2 or our new root-finding to solve Res(ϕ) = 0
(see Section 4.3). In this test, root-finding algorithms do not always apply when
ac = 1: such a case gives rise to multiple solution (see Figure 3) implying that Res(ϕ)
have the same sign on both sides of I0 . The initialization is done with ϕ0 = θλr for
Algorithm 4.1 and Algorithm 4.2, whereas the root-finding algorithms are initialized
with the intervals I0 := [10−2 , π/2−10−2 ] for Ning’s algorithm and I0 := [10−2 , θλr ] for
our new root-finding algorithm to solve Res(ϕ) = 0. We set γλr = γλ∗r and cλr = c∗λr
and run our test on the two cases ac = 1 and ac = 1/3. The former case gives rise to a
situation where ψ((a − ac )+ ) = ψ(0) = 0, i.e., µcG = µG , so that Theorem 4.1 applies.
The results are presented in Figure 4. In case of convergence, the obtained limit is

Figure 4: Number of iterations required to solve (3.9) with Algorithm 4.1 (Standard
fixed-point version 1 & 2), Algorithm 4.2 (Optimized fixed-point, robust and Newton
versions) and root-finding approaches (Ning’s algorithm and new root-finding) as
function of λr . Left: ac = 1, right: ac = 1/3. Only convergence cases are reported.

the same with all algorithms. We observe that the robust version of Algorithm 4.2
is the only algorithm that always converges and that its Newton version outperforms
all other algorithms and only diverges in three cases.
6.2. Example of a large wind turbine. We consider the IEA Wind 15-MW
reference wind turbine2 . In this example the blade length is 117 m, decomposed into
50 elements and designed for a TSR equal to 9. The coefficients CL and CD take into
account 3D-effects by using Du-Selig [16] stall delay 3D correction and are specified for
each element. We interpolate them using the Akima algorithm [1, 2] which provides
smooth approximations. A full description of the turbine is given in [18].
In our test, we compute optimal parameters (γλc r , ccλr ) using the Matlab function
fminunc, providing the gradient as in Algorithm 5.1. For the sake of consistency,
with use the same tolerance Tol= 10−7 in the stopping criteria of the optimization
procedure and of the solution algorithm considered to solve (3.3), i.e., the algorithms
stop when k∇Jλr (γλr , cλr )k ≤ Tol and |Res(ϕ)| ≤ Tol, respectively. For the sake of
numerical efficiency, we use a continuation approach, meaning that we optimize the
elements sequentially by starting from the blade tip and initialize the optimization
of the current element with the design obtained for the element previously consid-
ered. This process is itself initialized with optimum approximation applied to the
2 All numerical data related to this example are given in https://github.com/IEAWindTask37.
21
 Figure 5: Graphs of γλ∗r and γλc r , c∗λr and ccλr , Jλr (γλ∗r , c∗λr ) and Jλr (γλc r , ccλr ).

Solution algo. Stand. F-P (v1) Stand. F-P (v2) Opt. F-P (robust)
#Iterations 9105 8646 5045
Solution algo. Opt. F-P (Newton) New root-finding Ning’s alg.
#Iterations 1688 5653 5965

Table 3: Number of iterations of the solutions algorithms during the optimization.

tip element. We then compare it to the actual design, denoted by (γλtrue r

, ctrue
λr ) to
c c
(γλr , cλr ). The results are presented in Figure 5, whereas the corresponding values of
the power coefficient CP are CP (γλtrue r
, ctrue 3 c c
λr ) = 0.4531 , CP (γλr , cλr ) = 0.4628 and
∗ ∗
CP (γλr , cλr ) = 0.0127. We observe that the optimum approximation gives a very bad
design. On the other hand, the optimal design (γλc r , ccλr ) appears to be close to the
actual design (γλtrue
r
, ctrue
λr ), except for small values of λr , i.e. in the neighborhood
of the hub where it only weakly influences the power coefficient. This explains that
these two designs give similar values of this coefficient. Remark that the chords c∗λr
and ccλr are very large for small values of λr , which makes them unrealistic in practice.
Though the results do not depend on the considered solution algorithm (up to the
tolerance Tol), this one impacts the total number of iterations used to solve (3.3) re-
quired during the optimization process, as shown in Table 3. The robust version and
Newton version of Algorithm 4.2 appear to be the fastest procedures in this example.
7. Conclusion. In this paper, we present a new formulation of the BEM model,
which respects the paradigm of this approach in the sense that it decomposes the
model into a macroscopic part, related to the momentum theory, and a local part, re-
lated to the blade element theory. This framework allows us to obtain existence results
and new solution algorithms which outperform the usual algorithms and whose con-
vergence can be analyzed mathematically. We have focused on the case of extracting
turbines and a future work could consist in extending our formulation to propellers,
which was the initial purpose of this theory. Moreover, some work is required to
combine or include this model into modern CFD codes devoted to turbine design. Us-
ing BEM model as coarse solver or preconditionner of the fluid-structure interaction
system could improve the convergence properties of associated PDE solvers.
Acknowledgments. The authors acknowledge support from ANR Ciné-Para
(ANR-15-CE23-0019) and ANR HyFloEFlu (ANR-10-IEED-0006-04). The authors
thank the anonymous referees for their insightful suggestions. J.S. thanks Dylan
Machado for his careful proofreading of the article.
3 The actual design is planned to achieve CP = 0.489, see [18, p.8]
22
 Appendix: a case of convergence of Algorithm 4.1 (version 1). In the
case of the simplified model, Algorithm 4.1 reads as an iterative procedure based on
the formula

(7.1) ϕk+1 := fe(ϕk ),

λr tan−1 x + 1
where fe(x) := π2 −atan (λr + µL (x)h(x)) and h(x) := . This framework
sin x
makes it possible to obtain bounds for this sequence.
Lemma 7.1. Suppose that Assumption 1 holds and that max I + = θλr , with µL
non-decreasing. If

(7.2) µL (θλr ) ≤ µG (γλr ).

and ϕ0 ∈ [γλr , θλr ], then the sequence defined by (7.1) satisfies ∀k ∈ N, ϕk ∈ [γλr , θλr ].
Proof. Assume that for some k ∈ N, ϕk ∈ [γλr , θλr ]. Because of (7.1), we have
−1
tan ϕk+1 := λr + µL (ϕk )h(ϕk ). Combining it with µL ≥ 0 gives ϕk+1 ≤ θλr . Since
µL is increasing and h is decreasing, tan−1 ϕk+1 ≤ λr + µL (θλr )h(γλr ). Because
of (7.2), the latter is bounded by tan−1 γλr . The result follows by induction.
We complete this result by a condition about a contraction property.
Lemma 7.2. Suppose that µL is differentiable and denote by µ0L its derivative.
The derivative of fe satisfies

− sin θλr max µ0L h(γλr ) ≤ fe0 (ϕ) ≤ sin θλr max µL |h0 (γλr )|.
I+ I+

−1
Proof. We have fe0 (ϕ) = 1+(λr +µL (ϕ)h(ϕ))2
(µ0L (ϕ)h(ϕ) + µL (ϕ)h0 (ϕ)). Since ϕ ∈
(γλr , θλr ), µ0L (ϕ)h(ϕ) ≥ 0, µL (ϕ)h (ϕ) ≤ 0, we have 1+(λ +µ−1(ϕ)h(ϕ))2 µ0L (ϕ)h(ϕ)
0
≤
r L

fe0 (ϕ) ≤ 1+(λ +µ−1(ϕ)h(ϕ))2 µL (ϕ)h0 (ϕ). 0

Because h and h are decreasing on (γλr , θλr ),
r L
and since µL (ϕ)h(ϕ) ≥ 0, the result follows.
We are now in a position to obtain a conditional convergence result.
Theorem 7.3. In addition to the assumptions of Lemma 7.1, suppose that µL is
differentiable and satisfies

(7.3) sin θλr max µ0L h(γλr ) ≤ 1, sin θλr max µL |h0 (γλr )| ≤ 1.
I+ I+

Then, if ϕ0 belongs to [γλr , θλr ], the sequence (ϕk )k∈N defined by (7.1) converges to
the unique solution of (3.3).
Proof. As a consequence of Lemma 7.1, the function fe maps [γλr , θλr ] onto itself.
From (7.3) and Lemma 7.2 we deduce fe is contracting. The result follows from the
Banach fixed-point theorem.

REFERENCES

[1] H. Akima. A new method of interpolation and smooth curve fitting based on local procedures.
Journal of the ACM, 17(4):589–&, 1970.
[2] H. Akima. Method of bivariate interpolation and smooth surface fitting based on local proce-
dures. Communications of the ACM, 17(1):18–20, 1974.
23
 [3] Y. Bazilevs, M. C. Hsu, I. Akkerman, S. Wright, K. Takizawa, B. Henicke, T. Spielman, and
T. E. Tezduyar. 3D simulation of wind turbine rotors at full scale. Part I: Geometry
modeling and aerodynamics. Int. J. for Num. Meth. in Fluids, 65(1-3, SI):207–235, 2011.
[4] Y. Bazilevs, M. C. Hsu, J. Kiendl, R. Wuechner, and K. U. Bletzinger. 3D simulation of wind
turbine rotors at full scale. Part II: Fluid-structure interaction modeling with composite
blades. Int. J. Num. Meth. in Fluids, 65(1-3, SI):236–253, 2011.
[5] A. Betz. Das maximum der theoretisch möglichen ausnützung des windes durch windmotoren.
Zeitschrift für das gesamte Turbinenwesen, 26:307–309, 1920.
[6] M. Bossy, J. Espina, J. Moricel, C. Paris, and A. Rousseau. Modeling the wind circulation
around mills with a lagrangian stochastic approach. SMAI comp. math., 2:177–214, 2016.
[7] E. Branlard. Wind turbine aerodynamics and vorticity-based methods: fundamentals and recent
applications. Research topics in wind energy. Springer, 2017.
[8] E. Branlard and M. Gaunaa. Superposition of vortex cylinders for steady and unsteady simu-
lation of rotors of finite tip-speed ratio. Wind Energy, 19(7):1307–1323, JUL 2016.
[9] R. Brent. Algorithm with guaranteed convergence for finding a zero of a function. Computer
Journal, 14(4):422–425, 1971.
[10] J. Buhl, M.L. New empirical relationship between thrust coefficient and induction factor for
the turbulent windmill state. Technical Report NREL/TP-500-36834, National Renewable
Energy Laboratory, Golden, CO, August 2005.
[11] T. Burton, D. Sharpe, N. Jenkins, and E. Bossanyi. The Wind Energy Handbook, volume 1.
John Wiley and Sons, Ltd, 2001.
[12] P. K. Chaviaropoulos and M. O. L. Hansen. Investigating Three-Dimensional and Rotational
Effects on Wind Turbine Blades by Means of a Quasi-3D Navier-Stokes Solver . Journal
of Fluids Engineering, 122(2):330–336, 02 2000.
[13] M. J. Clifton-Smith. Wind turbine blade optimisation with tip loss corrections. Wind Engi-
neering, 33(5):477–496, 2009.
[14] M. Drela. XFOIL: An analysis and design system for low Reynolds number airfoils. In T. J.
Mueller, editor, Low Reynolds Number Aerodynamics, pages 1–12. Springer, 1989.
[15] L. Du, G. Ingram, and R. G. Dominy. A review of H-Darrieus wind turbine aerodynamic
research. Proc. Inst. of Mech. Eng., Mech. Eng. Sci. (C), 233(23-24, SI):7590–7616, 2019.
[16] Z. Du and M. Selig. A 3-d stall-delay model for horizontal axis wind turbine performance
prediction. In 1998 ASME Wind Energy Symposium, 1998.
[17] D. Eggleston and F. Stoddard. Wind turbine engineering design. Springer, New-York, 1987.
[18] G. Evan, J. Rinker, L. Sethuraman, F. Zahle, B. Anderson, G. Barter, N. Abbas, F. Meng,
P. Bortolotti, W. Skrzypinski, G. Scott, R. Feil, H. Bredmose, K. Dykes, M. Shields,
C. Allen, and A. Viselli. Definition of the iea 15-megawatt offshore reference wind. Techni-
cal Report NREL/TP-5000-75698, National Renewable Energy Laboratory, Golden, CO,
March 2020.
[19] W. Froude. On the elementary relation between pitch, slip and propulsive efficiency. Trans.
Roy. Inst. Naval Arch., 19(47):47–57, 1878.
[20] H. Glauert. The analysis of experimental results in the windmill brake and vortex ring states
of an airscrew. London: Aeronautical Research Committee, 1026, 1926.
[21] H. Glauert. Airplane propellers. In W. F. Durand, editor, Aerodynamic Theory, volume 4,
pages 169–360. Berlin: Julius Springer, 1935.
[22] H. Glauert. The Elements of Aerofoil and Airscrew Theory. Cambridge University Press, 1983.
[23] M. O. Hansen. Aerodynamics of Wind Turbines. Taylor and Francis, 2015.
[24] M.-C. Hsu and Y. Bazilevs. Fluid-structure interaction modeling of wind turbines: simulating
the full machine. Comp. Mech., 50(6, SI):821–833, 2012.
[25] N. Joukowsky. Vortex theory of screw propeller, i. Trudy Otdeleniya Fizicheskikh Nauk Ob-
shchestva Lubitelei Estestvoznaniya, 16(1):1–31, 1912.
[26] N. Joukowsky. Windmill of the NEJ type. Trans. of Cent. Inst. for Aero-Hydrodyn., 1920.
[27] N. Joukowsky. Joukowsky NE. Collected papers, volume VI. The Joukowsky Institute for
AeroHydrodynamics, Moscow-Leningrad, Russia: ONTI, 1937.
[28] F. Lanchester. A contribution to the theory of propulsion and the screw propeller. Trans. Inst.
of Naval Arch., 57:98–116, 1915.
[29] C. Lock, H. Bateman, and H. Townend. An extension of the vortex theory of airscrews with
applications to airscrews of small pitch, including experimental results. A.R.C. Research
Reports and Memoranda, 1014, 1925.
[30] H. A. Madsen, T. J. Larsen, G. R. Pirrung, A. Li, and F. Zahle. Implementation of the blade
element momentum model on a polar grid and its aeroelastic load impact. Wind Energy
Science, 5(1):1–27, JAN 2 2020.
[31] D. Maniaci. 49th AIAA Aerospace Sciences Meeting including the New Horizons Forum and

24
 Aerospace Exposition, chapter An Investigation of WT Perf Convergence Issues. Aerospace
Sci. Meetings. American Institute of Aeronautics and Astronautics, 2011.
[32] J. Manwell, J. Mcgowan, and A. L Rogers. Wind Energy Explained: Theory, Design and
Application, Second Edition, volume 30. John Wiley and Sons, Ltd, 2006.
[33] M. McWilliam and C. Crawford. The behavior of fixed point iteration and newton-raphson
methods in solving the blade element momentum equations. Wind Engineering, 35(1):17–
31, 2011.
[34] A. Ning, G. Hayman, R. Damiani, and J. M. Jonkman. Development and validation of a new
blade element momentum skewed-wake model within aerodyn. In Proc. of the 33rd Wind
Energy Symp., 2015.
[35] S. A. Ning. A simple solution method for the blade element momentum equations with guar-
anteed convergence. Wind Energy, 17(9):1327–1345, SEP 2014.
[36] V. L. Okulov, J. N. Sorensen, and D. H. Wood. The rotor theories by Professor Joukowsky:
Vortex theories. Progress In Aerospace Sciences, 73(SI):19–46, FEB 2015.
[37] V. L. Okulov and G. A. M. van Kuik. The Betz-Joukowsky limit: on the contribution to
rotor aerodynamics by the British, German and Russian scientific schools. Wind Energy,
15(2):335–344, MAR 2012.
[38] B. A. Prandtl L. Vier abhandlungen zur hydrodynamik und aerodynamik. Göttinger Klassiker
der Strömungsmechanik, 3:1–100, 2010. (Flüssigkeit mit kleiner Reibung; Tragflügeltheorie,
I. und II. Mitteilung; Schraubenpropeller mit geringstem Energieverlust).
[39] W. J. M. Rankine. On the mechanical principles of the action of propellers. Trans. Roy. Inst.
Naval Arch., 6:13–30, 1865.
[40] A. P. Schaffarczyk. Introduction to wind turbine aerodynamics. Green Energy and Technology.
Springer, 2014.
[41] W. Shen, R. Mikkelsen, J. Sørensen, and C. Bak. Tip loss corrections for wind research turbine
computations. Wind Energy, 8(4):457–475, 2005.
[42] H. Snel, R. Houwink, G. van Bussel, and A. Bruining. Sectional prediction of 3d effects for
stalled flow on rotating blades and comparison with measurements. In Proceedings of
European Community Wind Energy Conference, pages 395–399, 1993.
[43] Q. Song and W. D. Lubitz. Bem simulation and performance analysis of a small wind turbine
rotor. Wind Eng., 37(4):381–399, 2013.
[44] J. Sørensen. Aerodynamic Aspects of Wind Energy Conversion. In Davis, SH and Moin, P, ed-
itor, Annual Review Of Fluid Mechanics, volume 43 of Annual Review of Fluid Mechanics,
pages 427–448. Annual Reviews, 2011.
[45] J. Sørensen. General Momentum Theory for Horizontal Axis Wind Turbines. Springer, 2016.
[46] D. Spera, editor. Wind Turbine Technology: Fundamental Concepts in Wind Turbine Engi-
neering, Second Edition. ASME, New York, NY, 2009.
[47] Z. Sun, W. Z. Shen, J. Chen, and W. J. Zhu. Improved fixed point iterative method for blade
element momentum computations. Wind Energy, 20(9):1585–1600, SEP 2017.
[48] G. van Kuik. The Fluid Dynamic Basis for Actuator Disc and Rotor Theories. Amsterdam:
IOS Press, 2018.
[49] G. A. M. van Kuik. The Lanchester-Betz-Joukowsky limit. Wind Energy, 10(3):289–291, 2007.
[50] G. A. M. van Kuik, J. N. Sorensen, and V. L. Okulov. Rotor theories by Professor Joukowsky:
Momentum theories. Progress In Aerospace Sciences, 73(SI):1–18, FEB 2015.
[51] K. W. Van Treuren. Small-Scale Wind Turbine Testing in Wind Tunnels Under Low Reynolds
Number Conditions. Journal of Energy Resources Technology, 137(5), 09 2015. 051208.
[52] R. E. Wilson, P. B. S. Lissaman, and S. N. Walker. Aerodynamic performance of wind turbines.
final report. ERDA/NSF/04014-76/1, 1976.
[53] D. H. Wood. Application of extended vortex theory for blade element analysis of horizontal-axis
wind turbines. Renewable Energy, 121:188–194, JUN 2018.

25