Energy Conversion and Management 90 (2015) 154–165

Contents lists available at ScienceDirect

Energy Conversion and Management

journal homepage: www.elsevier.com/locate/enconman

Operating cycle optimization for a Magnus effect-based airborne wind

energy system
Milan Milutinović ⇑, Mirko Čorić, Joško Deur
Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Zagreb, Croatia

a r t i c l e i n f o a b s t r a c t

Article history: The paper presents a control variables optimization study for an airborne wind energy production
Received 11 June 2014 system. The system comprises an airborne module in the form of a buoyant, rotating cylinder, whose
Accepted 30 October 2014 rotation in a wind stream induces the Magnus effect-based aerodynamic lift. Through a tether, the
Available online 26 November 2014
airborne module first drives the generator fixed on the ground, and then the generator becomes a motor
that lowers the airborne module. The optimization is aimed at maximizing the average power produced
Keywords: at the generator during a continuously repeatable operating cycle. The control variables are the genera-
Airborne wind energy
tor-side rope force and the cylinder rotation speed. The optimization is based on a multi-phase problem
Magnus effect
Optimal control
formulation, where operation is divided into ascending and descending phases, with free boundary con-
Produced energy maximization ditions and free cycle duration. The presented simulation results show that significant power increase can
Energy production analysis be achieved by using the obtained optimal operating cycle instead of the initial, empirically based oper-
ation control strategy. A brief analysis is also given to provide a physical interpretation of the optimal
cycle results.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction An extensive overview of the current AWE concepts and present

As the ground-based wind turbine systems (see [1] for a concise
overview of their design aspects) approach their technological lim-
its, while also suffering from the intermittent nature and relatively
low speed of near-surface winds, a number of airborne wind
energy (AWE) systems have been proposed within the last decade.
The development of AWE systems is motivated by significantly
higher available wind power at high elevations. It is shown in [2]
that the increase in wind power density with height in the region
of 80–500 m is about 0.25 kW/m2/m for median winds (present
50% of the time or more), and that there is a several fold potential
increase in mean wind power density if the height is increased
from 100–150 m to 500–1000 m. The study presented in [3]
concentrates on altitudes accessible to the current and planned
AWE concepts and finds that locations with particularly strong
and relatively steady winds at altitudes below 3000 m are more
common than previously thought. The preliminary calculations of
the same study indicate that the energy production potential of
the AWE systems exploiting such winds can surpass the 2012
global electricity demand by a factor of three.
⇑ Corresponding author.
E-mail addresses: milan.milutinovic@fsb.hr (M. Milutinović), mirko.coric@fsb.hr
(M. Čorić), josko.deur@fsb.hr (J. Deur).
challenges is available in [4], while a summary of the field is given
in [5]. One of the main distinctions between the systems is in the
location of the generator. There are two basic arrangements: on-
board generators (OBG), which are attached to the airborne unit,
or ground-level generators (GLG), which are fixed on the ground
[5]. For example, the ascending-descending kite-based ‘‘Ladder-
mill’’ [6] and ‘‘Kite-Gen’’ [7] systems are examples of GLG systems,
and use flexible wings to generate lift. A rigid stationary tethered
rotorcraft system, an example of OBG system, is discussed in [8].
A lighter-than-air flexible stationary system that also uses an OBG
is presented in [9]. For an OBG system that uses a rigid, glider-like
flying structure that moves in a crosswind direction, see [10].
The airborne wind energy system considered in this paper
[11,12], consists of an airborne module (ABM) connected by a sin-
gle tether (rope) to the winch-generator system located on the
ground (Fig. 1a). A wind stream produces the Magnus effect-based
aerodynamic lift on a rotating cylindrical balloon, thereby driving
the generator in the ascending phase. The generator drives the
winch in the motoring mode to pull the ABM down during the
recovery phase. The rotation of the cylinder is accomplished using
an electric drive attached to the ABM (the tether has a dual
function of both the mechanical and the electrical link with the
ABM). Previously conducted theoretical studies [12] and, in
particular, preliminary proof-of-concept experimental results [11]

http://dx.doi.org/10.1016/j.enconman.2014.10.066
0196-8904/Ó 2014 Elsevier Ltd. All rights reserved.
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165
have shown that the energy production based on the proposed
concept is viable. To achieve a smooth grid power delivery during
the entire operating cycle (Fig. 1b) despite the inherently intermit-
tent power production response of the described system, the
ground station unit should be equipped with an electric storage
subsystem [13]. The storage subsystem stores the energy produced
during the ascending phase, and supplies it to the grid and the gen-
erator during the descending phase, thereby providing steady grid
power over the whole operating cycle.
An important objective for the overall system development is to
design a control system required to facilitate continuous cyclical
operation while maximizing the produced mechanical power
transferred by the rope to the winch. Based on the ABM control-
oriented 2D dynamics model initially outlined in [12], it is useful
to conduct open-loop (off-line) control optimization studies, which
can be used as an insightful basis and a benchmark for the devel-
opment of realistic, closed-loop control systems. Such an optimiza-
tion study is the main subject of this paper, noting that similar
studies have thus far been conducted mainly for GBG AWE systems
that are kite or wing-based, e.g. [14–17], which are aerodynami-
cally quite different from the Magnus effect-based concept pre-
sented herein, and for which it is well known that a crosswind
motion is preferable [18].
The cost function to be minimized relates to the negative value
of the average power produced at the generator during a single,
repeatable operating cycle. The control variables are the rope force
at the winch and the cylinder rotation speed. The problem formu-
lation and related optimization results are presented in a gradually
refined form. The results of simpler formulation stages serve as a
guide on how the optimization problem formulation could be
improved to better suit the goal of maximizing energy production
during continuous operation, i.e. over an arbitrary number of con-
secutive, repeatable operating cycles. Between different stages of
problem formulation, the constraints, boundary conditions and
even the cost function were readjusted according to this goal.
The final formulation is based on multi-phase optimization, by
dividing the time horizon into ascending and descending phases,
with free initial and final conditions on ABM state and control vari-
ables, as well as free cycle duration. Periodicity constraints are
imposed to ensure repeatability of the obtained cycle. The problem
is solved numerically using the TOMLAB/PROPT software tool [19],
which is an optimal control platform intended for integration with
the Matlab package.
Lift force
Air current (high- Airborne
Rotation module
altitude wind)
(cylindrical
Ascending: power vasc balloon)
production (Pr,asc) Descending:
Generator + inverter vdes
power consumed
(Pr,des)
G/M Winch
~ P
Ascending Pr,asc
~ dis
= Charging
consumption generation
Energy storage +
DC link
+ - power converter
PDClink
= Discharging
~ 0
Power
Electrical
~ power grid
Power delivered to
a electrical grid (Pgrid) b
Fig. 1. Schematic of the system (a) and illustration of its idealized power cycle (b).
155
The results obtained in Tomlab have been verified by means
of numerical simulations, using the optimized control variables
as inputs of the ABM simulation model. The optimal cycle power
production results are compared with those obtained by the
existing heuristic control system [12] to analyse the potential
for future improvements by means of control strategy enhance-
ment. An algebraic analysis of the optimal operating cycle is also
presented, in order to provide its physical interpretation. This
knowledge provides insight into the optimal system behaviour
and is required to determine a control strategy aimed to
approach the optimization results in a robust manner. Therefore,
the main contribution of the paper is in determining consider-
ably different and more productive operation compared to that
obtained by the existing control strategy [12], as well as a better
understanding of optimal behaviour of the novel Magnus effect-
based AWE system.
Even though the presented optimization study has been applied
to the particular airborne wind energy system, it illustrates how
advanced numerical optimization methods and tools, combining
control variable and parameter optimization, can be used to gain
insights into optimal behaviour of complex energy conversion
and management systems in general, such as those considered in
[20], where power production of a conventional wind turbine
was maximized by means of appropriate turbine speed control,
and [21], where overall efficiency of an IC engine-powered vehicle
was maximized by optimizing the energy management of its elec-
tric energy storage system.
2. Process model
A model used for the purpose of ABM control system design,
and in particular for numerical optimization of control variables,
should be as simple as possible. At the same time, it should reflect
the basic dynamics of the system. To satisfy these requirements, a
simple 2D ABM state-space dynamics model based on the
approach introduced in [12] is used throughout this paper. The
presented model has been refined into a more complete model in
[22], to account for many effects that were disregarded here for
the sake of computational efficiency, such as elasticity and inertia
of the rope, inertia of the winch and aerodynamic drag of the rope.
Such a model would be suitable for final control system design and
simulation purposes.
Airborne module/generator
Power to the electrical grid
Energy storage
Discharging
P
Constant
Power
output power Pgrid
Tasc Tdes
Tcyc
t
Pchg
Pr,des
Charging
Descending
156 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165

The model is illustrated in Fig. 2. The forces acting on the ABM where mABM is the total ABM mass, and g is standard gravity.
include: the aerodynamic drag force Fd, the aerodynamic lift force The rope force components are obtained from a straight-line
Fl, the buoyancy force Fb, the ABM weight force Fg, and the ABM- rigid rope model with neglected aerodynamic drag (see Fig. 3a,
side rope force Fr. The drag and lift forces form the resultant aero- and cf. [22] for a more precise rope model):
dynamic force Fad. The force components are considered positive x
when oriented as in Fig. 2, and negative for opposite orientation. F r;x ¼ F r;w  cos b ¼ F r;w  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð10Þ
x2 þ z 2
The velocities include: the ABM velocity ~ v , the wind velocity ~v w,
and the relative velocity between the wind and the ABM, z
v rel ¼ ~
~ vw  ~v . In the case of velocities, positive directions of their F r;z ¼ F r;w  sin b þW r ¼ F r;w  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ W r ð11Þ
|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} x2 þ z 2
components always correspond to positive directions of the coor- F r;w;z
dinate axes. Reference direction of the cylinder angular velocity
xcyl is chosen so that for the given reference direction of wind where the winch-side rope force Fr,w is assumed to be one of the
velocity ~ v w , the Magnus effect causes a lift force Fl that points control variables. The rope weight Wr is calculated as weight-per-
upward (i.e. that has positive z component). metre times the variable rope length:
The corresponding model equations are: pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
W r ¼ wr  lr ¼ wr  x2 þ z 2 ð12Þ
v rel;x ¼ v w;x  v x ð1Þ
Another important quantity, the rope speed at the winch vr, is
v rel;z ¼ v w;z  v z ð2Þ modelled as the projection of ABM velocity ~ v on the rope line of
the straight line rope model, i.e. as the radial component of velocity
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi v , as shown in Fig. 3b:
~
v rel ¼ v 2rel;x þ v 2rel;z > 0 ð3Þ
x z
The drag and lift forces are calculated in their x and z direction
v r ¼ v x cos b þ v z sin b ¼ v x  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
þ v z  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
ð13Þ
x þz x þz
components as:
The state vector includes the ABM coordinates, x and z, and the
F d;x ¼ C d  qair  rcyl  lcyl  v rel;x  v rel ð4Þ components of ABM velocity, vx and vz.

F d;z ¼ C d  qair  r cyl  lcyl  v rel;z  v rel ð5Þ x ¼ ½x z v x v z T ð14Þ

The control variables vector includes the rope force at winch,
F l;x ¼ signðxcyl Þ  C l  qair  r cyl  lcyl  v rel;z  v rel ð6Þ Fr,w, and the cylinder rotation speed, xcyl:
 T
F l;z ¼ signðxcyl Þ  C l  qair  rcyl  lcyl  v rel;x  v rel ð7Þ u ¼ F r;w xcyl ð15Þ
where Cd and Cl are the drag and lift coefficients, qair is the air den- The ABM dynamics is further described by the following state
sity, and rcyl and lcyl are the cylinder radius and length, respectively. equations as functions of the calculated forces:
The buoyancy force acting on the ABM cylinder and the weight
of the entire ABM (including the gas inside the cylinder) are given x_ ¼ v x ð16Þ
by the following expressions:
z_ ¼ v z ð17Þ
F b ¼ qair  r 2cyl  p  lcyl  g ð8Þ
F x ðx; uÞ
F g ¼ mABM  g ð9Þ v_ x ¼ ð18Þ
mABM

F z ðx; uÞ
v_ z ¼ ð19Þ
mABM
Z'
where:
Fl,z
Fl Fad F x ¼ F d;x þ F l;x  F r;x ð20Þ
vrel
v
Y'
F z ¼ F l;z  F d;z  F r;z þ F b  F g ð21Þ
vw
Values of the model parameters are given in Appendix A. The
Fb
aerodynamic drag and lift coefficients, Cd and Cl, functions of the
cylinder circumferential speed to relative air speed ratio, X:
Fl,x Fd,x X'  
Fr,x Fd,z X ¼ xcyl r cyl =v rel ð22Þ
Fd
Fg The Cd(X) and Cl(X) maps are taken from [23] and conveniently
initially = z0

approximated by polynomials (see Appendix A).

cyl
X'Y'Z' II XYZ
Fr Fr,z 3. Optimization
Fl Fd
T
Z
Y The goal of numerical optimization is to find control variables’
time-responses that maximize energy production during continu-
X
initially = x0 ous system operation, which will provide an optimal operating
cycle. The optimization problem needs to be formulated by
Fig. 2. Reference coordinate system, forces and velocities of the basic two- defining: (i) the cost function, (ii) the constraints and (iii) boundary
dimensional ABM dynamics model. (initial/final) values of state variables.
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165
a b
z Fr,z
B B
Straight rope
lr
Constant along
entire rope
β vr
Fr,x A A
Fr,w,z
Fr,w x x
Fig. 3. Straight line quasi-static rope model (a) and calculation of rope speed at winch
3.1. Optimization approach
The optimization problem formulation is gradually refined. On
one hand, this is done as a purely practical matter, since the initial,
simpler formulations are more easily implemented and solved. On
the other hand, the results obtained using simpler formulations
serve not only as a guide as to how to improve the further formu-
lations, but can also provide insight in the behaviour of the system,
and thereby help to physically interpret the final optimization
results.
Four optimization problem formulations were employed, some
of them divided into sub-cases. The formulations are summarized
in Table 1.
All formulations share some common constraints. The mini-
mum allowed value of the altitude z is 32 m. Allowable range of
rope speed vr is between 6 and 4 m/s. The minimum value of
the winch-side rope force Fr,w is set to zero, since the rope cannot
exert a push. The maximum allowable ABM-side rope force Fr is
dictated by the rope strength. The speed ratio X is limited to the
maximum X value for which the maps Cl,d(X) are defined. Finally,
the initial and final states are required to be the same, i.e. x0 = xf,
which is the basic periodicity constraint. This initial/final boundary
state xb differs between the formulations, ultimately becoming a
free parameter that is optimized together with the control vari-
ables (Table 1).
All cost functions feature the power at the generator, which is
given by (Fig. 3; Pr in Fig. 1):
P ¼ F r;w  v r
Depending on the formulation, power is integrated either to cal-
culate the energy produced during a cycle, or the average power
obtained by dividing the energy with the cycle time. A negative
sign appears in the cost function because the cost function is to
be minimized.
The second integral term in any of the cost functions in Table 1
serves to limit derivatives of the control variables, i.e. to limit sud-
den changes of Fr,w and xcyl, which can be unrealistic and can also
cause numerical problems. The level of derivatives suppression is
modulated by the penalty coefficient k, whose value is determined
empirically.
3.2. The TOMLAB optimization platform
TOMLAB optimization platform is the Matlab interface package
for modelling, optimization and optimal control [19]. It consists of
modules for optimal control formulation (PROPT), automatic
differentiation (MAD) and a variety of problem solvers (SNOPT,
KNITRO, CPLEX etc.). The user specifies state equations, which
are considered as equality constraints, as well as initial and final
conditions, constraints and the cost function. The optimization
157
vx
Fr z
vz β
Wr
β v
Fr,w vr
Straight rope
Fr,x
lr
z z
β
x x
vr (b).
approach is based on pseudo-spectral collocation method in which
the optimal control problem is discretized and transformed into a
large-scale nonlinear programming (NLP) problem [24]. The solu-
tion takes polynomial form of order n, where n is the number of
collocation (time grid) points specified by the user, and the polyno-
mial coefficients are parameters to be optimized. The polynomial
solution needs to satisfy the discretized state equations and con-
straints at the collocation points.
In this work the large-scale sparse solver SNOPT is used, which
is based on the sequential quadratic programming (SQP) method.
As SNOPT uses nonlinear function gradients, calculated automati-
cally by MAD, the problem must be smooth (first-order differentia-
ble) around the optimal solution, which is the case for the
particular system, except for the sign functions in Eqs. (6) and
(7), which are replaced in the code by the smooth tanh function.
If the solver claims that the solution is optimal, it satisfies neces-
sary, but not always sufficient, optimality conditions. This means
that solution is local, and hence not guaranteed to be a global opti-
mum. To guarantee that the global optimum is obtained, one must
either solve Hamilton–Jacobi–Bellman (HJB) equation [25] or show
that the problem is convex. The computational effort to achieve
this for the presented problem is excessively high because the
problem is not convex. A more feasible and effective approach is
to use TOMLAB solvers, and eventually perturb the main problem
formulation parameters (e.g. initial solution guess and number of
grid points) for multi-run optimizations to check their sensitivity
to local optima and obtain more accurate result (the one with min-
imum cost function).
ð23Þ
3.3. Formulation 1
Initially, it was envisaged that the cycle was to start from rest
(vx = vz = 0 m/s) at a point close to the winch (at x0 = z0 = 40 m),
so that the initial/final state was xb = [40 40 0 0]T (Table 1). The
cycle time tf was predefined based on the given maximum length
of the rope (lr,max = 800 m) and the aforementioned rope speed
range, which gave approximately tf = 300 s.
Fig. 4 shows the optimization results for the four cases of For-
mulation 1, which differed by the value of the penalty coefficient
k (see last column in Table 1).
Case 1: Small penalty coefficient k allows abrupt changes of the
control variables. While this is favourable from the standpoint of
energy production, it may not be realistic due to various opera-
tional and hardware constraints. It is, though, worth to observe
that the ABM tends to ascend and descend many times during
the designated time tf = 300 s, thus making many small loops of
the trajectory z(x).
Case 2: By increasing k, the time response of the system
becomes much smoother and the number of trajectory loops sig-
nificantly decreases. More importantly, it becomes evident that
158 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165

Table 1
Optimization formulations.

Form. Case Cost function, J Common constraints Initial/final state, Final Asc./desc. Phase Other specific
xb = [xb zb vxb vzb]T time, tf phases constraints constraints
1 1 k = 0.001
2 R tf R tf  _ 2 
[40 m 40 m 0 m/s 0 m/s]T 300 s n/a n/a k = 0.01
J¼ 0 F r;w v r dt þ k 0
_ 2cyl dt
F r;w þ x
3 k = 0.1
4 k = 0.5
z P 32 m
2 1 vz,A > 0
vr 6 4 m/s A–D vz,D < 0 kA = 0.0001
R tf R tf  
xAf = xD0 kD = 0.001
J¼ 0 F r;w v r dt þ k 0 F_ 2r;w þ x
_ 2cyl dt
T
vr P -6 m/s [40 m 40 m free free] 300 s uAf = uD0

2 A1–D1–A2–D2 vz,A1,2 > 0

Fr,w P 0 vz,D1,2 < 0
xA1f = xD10 k = 0.01
uA1f = uD10
X67 xD1f = xA20
uD1f = uA20
3 – R tf R tf   x0 = xf = xb –
F r;w v r dt 0:001 F_ 2r;w þx
_ 2 dt
J¼ 0
þ
0 cyl (xb defined in [40 m 40 m free free]T free vz,A > 0
tf tf
R tf   next column) A-D vz,D < 0
4 1 R tf 0:001 F_ 2r;w þx
_ 2 dt lr < 800 m
F r;w v r dt
2 J¼ 0
þ
0 cyl
xAf = xD0 lr < 1600 m
tf tf
[free free free free ]T free uAf = uD0
3 lr < 800 m
u0 = uf

Case1 Case 2 Case 3 Case 4

Eprod = 6.07 kWh Eprod = 4.77 kWh Eprod = 4.05 kWh Eprod = 3.67 kWh
Pavg Case1
= 72.86 kW PavgCase 2 kW
= 57.20 PavgCase 3 kW
= 48.56 PavgCase 4 kW
= 44.01

500 500 500 500

400 400 400
z [m]

400
300 300 300 300
200 200 200 200
100 100 100 100
-200 0 200 400 -200 0 200 400 -200 0 200 400 -200 0 200 400
x [m] x [m] x [m] x [m]
4 4 4 4
x 10 x 10 x 10 x 10
5 5 5 5
[N]
r,winch

2.5 2.5 2.5 2.5

F

0 0 0 0
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300

40 40 40 40
[rad/s]

20 20 20 20
cyl

0 0 0 0

-20 -20 -20 -20

0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300

5 5 5 5
3 3 3 3
v [m/s]

1 1 1 1
-1 -1 -1 -1
r

-3 -3 -3 -3
-5 -5 -5 -5
-7 -7 -7 -7
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300

8 8 8 8
C
6 l
Cd 6 6 6
C , C [-]
d

4 4 4 4
l

2 2 2 2
0 0 0 0
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300
t [s] t [s] t [s] t [s]

Fig. 4. Optimization results obtained for Formulation 1 for various control variables derivatives’ penalty coefficients k (see Table 1).
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165 159

the loops appear to be rather upright (vertical), as opposed to slant
loops produced by the basic control strategy from [12] (see also
Fig. 10). It should also be noted that the optimized control variable
Fr,w (and xcyl to some extent) has a square-wave-like time
responses, i.e. the optimized control action is of ‘‘bang-bang’’-like
type with the magnitudes corresponding to the control variable
limits. The energy produced is reduced by 22% when compared
with the less constrained Case 1 (see the values at the top of Fig. 4).
Case 3: When k is further increased, the number of subcycles
further reduces (to only two cycles). The trajectory is again mostly
upright, except in its initial and final parts when the ABM is near
the winch. The energy production drops by nearly 15% compared
with Case 2.
Case 4: If the control effort is further suppressed by increasing k
to 0.5, only one cycle remains, whose shape reinforces the conclu-
sion that the optimal result tends to have vertical trajectory in the
middle part of response when the ABM is far away from the winch.
The energy production is 10% lower than in Case 3.
An interesting observation valid for all cases (including subse-
quent Formulations), is that the optimal response of aerodynamic
coefficients Cd and Cl is such that their squared sum is maximized.
The corresponding speed ratio X is about 5.5.
3.4. Formulation 2
In the second formulation the operation is divided into a strictly
defined number of ascending (A) and descending (D) phases to
consider the practical case of repeatable operating cycles. Further-
more, the initial/final velocity of the ABM was not predetermined,
but was instead treated as a free parameter to be optimized (see
Table 1). In this way, the control effort penalty term k can be signif-
icantly decreased (Table 1), thus strictly focusing on getting the
maximum energy production. Formulation 2 includes two cases,
differing by the number of A and D phases, where Case 1 includes
two phases (A–D, producing one ascending-descending cycle), and
Case 2 has four phases (A1–D1–A2–D2, producing two ascending-
descending cycles). The optimization results are shown in Fig. 5.
Case 1: To define a single ascending phase followed by a single
descending phase, the sign of z component of velocity, vz, is con-
strained to be positive during ascending and negative during
descending (see Table 1). For a continuous overall cycle, the state
at the end of the ascending phase is constrained to be equal to
the state at the start of the descending phase, and the same applies
to control variables (Table 1). This reversal-point state and control
values are optimized, as well as the corresponding reversing
instant (the time at which the reversal occurs). Compared to the
results of single-cycle Formulation 1 – Case 4 (Fig. 4), the single-
cycle results obtained in Fig. 5 indicate that the control variable
changes are more emphasized, the ABM trajectory is steeper, and
the produced energy is boosted by 14%. This is explained by the
aforementioned improvement of Formulation 2 in terms of relax-
ing the constraint on control variables’ change.
Case 2: The dual-cycle optimization results shown in Fig. 5,
when compared to the corresponding results of Formulation

Eprod = 4.18 kWh Eprod = 5.17 kWh

Case 1 Pavg = 50.16 kW Case 2 Pavg = 62.06 kW
600 600
z [m]

400 400
200 200

-500 0 500 1000 -500 0 500 1000

x [m] x [m]
4 4
x 10 x 10
5 5
[N]
r,winch

2.5 2.5
F

0 0
0 100 200 300 0 100 200 300

40 40
[rad/s]

20 20
cyl

0 0
-20 -20
0 100 200 300 0 100 200 300

5 5
3 3
v [m/s]

1 1
-1 -1
r

-3 -3
-5 -5
-7 -7
0 100 200 300 0 100 200 300

8 8
C
l
6 6
C , C [-]

Cd
d

4 4
l

2 2
0 0
0 100 200 300 0 100 200 300
t [s] t [s]

Fig. 5. Optimization results obtained for Formulation 2 with one ascending-descending cycle (Case 1) and two ascending-descending cycles (Case 2).
160 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165

1 – Case 3, show the same trends as the above results of Case 1,
with energy production boost of 28%. Furthermore, these results
more clearly reveal that the optimal ABM cycle trajectory is appar-
ently such that it has a vertical shape, and that it tends to be
located far away from the winch.
3.5. Formulation 3
Formulation 3 is introduced to verify the hypothesis that the
apparently optimal, vertical ABM motion cannot be sustainable if
executed in the vicinity of the winch. Here, the previously fixed
final time tf is now ‘‘free’’ and optimized along the control variables
and the initial/final velocity (Table 1). In the presence of free final
time tf, the average power during the two-phase cycle is maxi-
mized instead of the energy (Table 1).
The optimization results are shown in Fig. 6. The produced
power is marginally better than in Formulation 2 – Case 1. It is,
however, important to note that the ABM motion is nearly vertical
over the whole cycle, with the optimal cycle time tf of only 12 s
(down from previously fixed time tf = 300 s). This indicates that
the optimal motion should be vertical even in the vicinity of the
winch, but the cycle would become impractically short. Namely,
in a real system, which has a slower dynamic response of the
winch and ABM drives than the simple model used in the optimi-
zation, the reversals between ascending and descending would
take up a large portion of such a short cycle duration, thereby
decreasing the average power. Also, the number of bending cycles
that the rope undergoes grows as cycle duration is decreased,
thereby reducing the rope lifetime.
3.6. Formulation 4
Formulation 4 is an extension of Formulation 3 such that the
boundary state vector xb is free, i.e. it is optimized. This means that
the initial/final position of the ABM is now optimized, along the ini-
tial velocity. This change requires an additional constraint, namely
the limitation on the rope length lr (Table 1). If there were no such
limitation, the ABM would be allowed to operate farther from the
winch than is realistically possible. Formulation 4 includes three
cases (Table 1), differing in the allowed rope length, and/or the final
periodicity constraint. The results are shown in Fig. 7.
Case 1: The results obtained for lr = 800 m clearly confirm that
the favourable trajectory is nearly vertical, and that its position
in the x–z plane is quite far from the winch, more specifically as
far as possible for the given rope length limit. The average power
is substantially larger than in the previous Formulations 1 through
3, where initial and final positions were preset, instead of being
optimized (the power is increased 46% compared to the next best
result, obtained with Formulation 3).
Case 2: The results are qualitatively similar for lr = 1600 m,
while the average power is only slightly boosted. Since high rope
lengths are generally not desirable due to increased winch inertia,
ABM load, rope stresses, aerodynamic drag and system costs, the
advantage should be given to the lower rope length lr = 800 m.
Case 3: This is a definite formulation for achieving a repeatable
cycle, as the periodicity condition of initial and final control vari-
ables’ equality is added (Table 1). The results obtained for
lr = 800 m are comparable to those of Case 1 (Fig. 7), with the
average power slightly reduced in the presence of the additional
constraint (Pavg = 89.12 kW).
3.7. Additional formulations
The above optimization study is extended in this subsection by
considering the effects of limited cylinder rotation speed xcyl
and sloped (non-horizontal) wind. The analysis includes two
characteristic cases of sloped wind direction and eight levels of cyl-
inder speed limit. The optimization formulations are based on For-
mulation 4 – Case 3.
Limited cylinder rotation speed xcyl: The following constraint is
added to the problem formulation: xcyl < ncyl,maxp/30, where
ncyl,max e {25, 50, 75, 100, 125, 150, 175 and 200} rpm. The opti-
mized trajectories and the most relevant time responses are shown
in Fig. 8. The corresponding dependence of the average produced
power on the cylinder speed limit is plotted in Fig. 9.
The results in Fig. 8 show that it remains beneficial to lift the
ABM vertically even if the cylinder rotation speed is limited. The
shape of other system time responses is not largely influenced by

Pavg = 62.69 kW 4
x 10
5
[N]

60
z [m]

r,winch

2.5
50
F

40 0
20 40 60 0 5 10 15
x [m] t [s]
5
[rad/s]

40 3
v [m/s]

20 1
-1
cyl

0 -3
r

-5
-20 -7
0 5 10 15 0 5 10 15
t [s] t [s]
8
C
l
C , C [-]

6
C
d

4 d
l

2
0
0 5 10 15
t [s]

Fig. 6. Optimization results obtained for Formulation 3, introducing cycle time as parameter to be optimized.
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165 161

Case1 Case 2 Case 3

PavgCase 1 kW
= 91.45 Pavg Case 2 kW
= 93.23 PavgCase 3 kW
= 89.12

550 550
900

z [m]
500 500
450 800 450
400 400
350 700 350
400 600 800 1000 1200 1400 200 400 600 800
x [m] x [m] x [m]
4 4 4
x 10 x 10 x 10
5 5 5
[N]
r,winch

2.5 2.5 2.5

F

0 0 0
0 20 40 60 0 50 100 0 50 100

40 40 40
[rad/s]

20 20 20
cyl

0 0 0
-20 -20 -20
0 20 40 60 0 50 100 0 50 100

5 5 5
3 3 3
v [m/s]

1 1 1
-1 -1 -1
r

-3 -3 -3
-5 -5 -5
-7 -7 -7
0 20 40 60 0 50 100 0 50 100

8 8 8
Cl
6 6 6
C , C [-]

Cd
d

4 4 4
l

2 2 2
0 0 0
0 20 40 60 0 50 100 0 50 100
t [s] t [s] t [s]

Fig. 7. Optimization results obtained for Formulation 4, introducing initial ABM position as parameter to be optimized for ‘‘nominal’’ rope length (Case 1), double rope length
(Case 2), and consistent control variable boundary conditions (Case 3).

the cylinder rotation speed. However, the magnitude of these
responses during the power production phase increases as ncyl,max
grows, and it saturates at the speed of around 150 rpm at which
the winch-side rope force Fr,w becomes saturated for the given sys-
tem (particularly wind) parameters. Correspondingly, the average
produced power also saturates1 at the cylinder speed of approxi-
mately 150 rpm (Fig. 9).
The main discrepancy in the shapes of responses in Fig. 8 is in
the appearance of a triangular form of the ABM trajectory for med-
ium cylinder speeds, and the related occurrence of a power pro-
duction phase at the end of the response. This kind of trajectory
incorporates a somewhat counterintuitive form of lowering the
ABM, where the rope is actually unwound and the power produced
while the ABM is descending. This corresponds to the lower cath-
etus of the triangle-shaped trajectory and the positive power in the
last stage of response.
The results of optimization with cylinder speed limited to
1
The fact that the average power decreases as ncyl,max increases from 150 rpm to
175 rpm (Fig. 9) appears to be due to a failure of the optimization algorithm to find a
precise solution for ncyl,max = 175 rpm (the optimizer apparently got stuck in a local
optimum, see Section 3.2). This may be related to the fact that the optimal triangle-
shaped trajectory occurs for ncyl,max e {100, 125, 150} rpm (triangle markers in Fig. 9),
but not at the apparently sub-optimal result for ncyl,max = 175 rpm (and also higher
speeds).
200 rpm, providing average power of 89.18 kW, have been com-
pared with the corresponding results of initial closed-loop control
strategy described in [12] and [22]. The comparison is shown in
Fig. 10. The particular cylinder speed limit was chosen in optimi-
zation because it roughly equals the observed maximum achieved
speed of the initial control strategy. The initial strategy com-
mands constant rope unwinding speeds during ascending and
descending (vr = 4 m/s and vr = 6 m/s, respectively) [12,22]. Dur-
ing ascending, the cylinder speed is generated so that the speed
ratio X = 4.51 is achieved, which maximizes the difference
between the lift and the drag coefficients Cl and Cd [23]. During
descending, the cylinder speed reference is equal to the scaled
absolute value of the ratio of ABM x and z coordinates, xcylR = Kcyl-
|x/z|, which is a heuristic way to prevent excessive descending of
the ABM (the cylinder speed, and hence the lift force, are
increased as the ratio |x/z| increases). That strategy produced
average power of 40.15 kW, meaning that an increase of 122%
has been achieved through optimal control.
Non-horizontal wind direction: Two possibilities of wind direc-
tion are considered, both for Formulation 4 – Case 3: (i) upward
sloping wind with w = 15°, and (ii) downward sloping wind with
w = 15°, where the absolute wind speed remains equal to the
one used previously (10 m/s). The optimized ABM trajectory is
shown in Fig. 11. The optimized trajectory is consistent with the
results of Formulation 4 in terms of being nearly perpendicular
162 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165

n : 4
cyl,max
25 rpm
50 rpm 2
75 rpm
100 rpm

v [m/s]
0
125 rpm
500 150 rpm

r
175 rpm -2
400
z [m]
200 rpm
300 -4
200
100 -6
0 500 1000 0 20 40 60 80
x [m] t [s]
4
x 10
5 30

20
[N]

[rad/s]
10
r,winch

2.5

cyl
ω
F

-10
0
0 20 40 60 80 0 20 40 60 80
t [s] t [s]

5 200

4 150

3 100
P [kW]
X [-]

2 50

0
1
-50
0
0 20 40 60 80 0 20 40 60 80
t [s] t [s]

Fig. 8. Optimization results for Formulation 4 – Case 3 and different levels of cylinder speed limit ncyl,max.

100 4. Algebraic analysis

90
A brief interpretation of the optimization results by means of an
80
algebraic analysis is concentrated on the ascending (production)
70 part of the operating cycle. The main aim of the analysis is to
60
explain why the optimal ABM trajectory is nearly perpendicular
Pavg [kW]

to the wind direction and executed far away from the winch.
50

40
4.1. Optimal shape of ABM trajectory
30

20 The power P = Fr,wvr produced during the ascending phase is

10
predominantly dependent on the winch-side rope force Fr,w,
because the rope speed is preferably kept at or near its maximum
0 value. The rope force Fr,w is induced mainly by the resultant aero-
0 50 100 150 200 250
dynamic force Fad, which is the dominant force in the system. This
ncyl,max [rpm]
force is determined from the ABM model as (see Fig. 2 and Eqs.
(4)–(7)):
Fig. 9. Optimization results for average produced power Pavg as a function of
cylinder speed limit ncyl,max (the last point, plotted at 225 rpm, corresponds to
unlimited ncyl,max).
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

2
2
ffi
F ad ¼ F 2ad;x þ F 2ad;z ¼ F l;x þ F d;x þ F l;z  F d;z
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ qair  r cyl  lcyl  v 2rel  C 2l þ C 2d ¼ qair  r cyl  lcyl  v 2rel  C ad ð24Þ
to the wind direction. The time responses are omitted in Fig. 11 as
they are comparable to those given in Fig. 7. Similarly, the average Firstly, Eq. (24) explains why the optimization results tended to
produced power is approximately equal in the two cases (horizon- give approximately constant aerodynamic coefficients Cl,d, which
tal and sloped wind). correspond to the speed ratio X = 5.5. Namely, for that value of X,
the total aerodynamic coefficient Cad has a maximum value, thus
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165 163

4
x 10
Initial control 5 30
strategy
500 Optimization 4
20
400

Fr,winch [N]

ωcyl [rad/s]
3

z [m]
300 10
2
200
0
1
100
0 -10
200 400 600 0 100 200 300 0 100 200 300
x [m] t [s] t [s]

5 200
4 Pavg = 40.15 kW
4 150 P = 89.18 kW
2 avg

0 3 100
v [m/s]

P [kW]
X [-]
-2 2 50
r

-4
1 0
-6
0 -50
0 100 200 300 0 100 200 300 0 100 200 300
t [s] t [s] t [s]

Fig. 10. Comparison between the initial control strategy-based results and optimization results for cylinder speed limit ncyl,max = 200 rpm.

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
700 v rel ¼ ðv z Þ2 þ ðv w  v x Þ2
= 15 ° sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
650  
line
= -15 ° vx vr 2
15 ° ¼  þ ðv w  v x Þ2 ð25Þ
600 tan b sin b
550 Which, for the characteristic mode definitions above, reduces to
Pavg = 88.67 kW
500 the following expressions, (note that the wind velocity is assumed
z [m]

to have only x component):

450

400 Pavg = 89.26 kW sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

 
jv x j vr 2
350 v rel;back ¼ þ þ ðv w þ jv x jÞ2 ð26Þ
tan b sin b
-15 °
300 line
250 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

200
200 300 400 500
x [m]
Fig. 11. Optimized trajectory for Formulation 4 – Case 3 and sloped wind (wind
angle w).
maximizing the aerodynamic force Fad. Note that, according to Eq.
(22), the required value of X can be achieved by means of control-
ling the cylinder rotation speed xcyl.
Secondly, for the approximately constant Cad, and since qair is
assumed constant, the only variable term in Eq. (24) is the relative
velocity between the air and the ABM, vrel. This velocity is
described by Eqs. (1)–(3) and it should be maximized for maximal
aerodynamic force Fad. To facilitate further analysis, three charac-
teristic modes of ABM ascending are introduced:
1. Radial lift – ground-rope angle b is constant, thus v = vr.
2. Vertical lift – ABM x coordinate is constant, thus v = vz, while
vx = ax = 0.
3. Backward lift – ABM ascends opposite the wind, i.e. ‘‘back-
wards’’, with vx = const. <0.
Eqs. (1)–(3) and (13) give the general formula for relative
velocity:
vr 2 2
v rel;v ert ¼ þ vw ð27Þ
600 700 800 sin b
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
v rel;rad ¼ ðv r sin bÞ þ ðv w  v r cos bÞ2 ð28Þ
For the same position in the x–z plane, i.e. the same b, it is clear
that vrel,back > vrel,vert. Also, since sin b and cos b are between 0 and 1
and vr < vw, it holds that vrel,vert  vrel,rad. This yields:
v rel;back > v rel;v ert  v rel;rad ) F ad;back > F ad;v ert  F ad;rad
) Pback > Pv ert  P rad ð29Þ
which confirms that it is favourable to lift the ABM backwards or
vertical to the wind, as opposed to sub-optimal radial lift.
4.2. Region of feasible vertical and backward lift
The first step in the vertical lift feasibility analysis is to develop
a kinematic model suitable for describing the three characteristic
cases of ABM lift. In addition to having constant (maximum) rope
unwinding speed vr = 4 m/s (see Table 1 and e.g. Fig. 7), all cases
are characterized by a constant value of the ABM velocity x compo-
nent, vx, implying that ax = 0 (see the above definitions of ascend-
ing modes and the kinematical relationships in Fig. 3b). Taking
the time derivative of kinematic Eq. (13), taking into account that
dvr/dt = 0 and ax = d2x/dt2 = 0, and rearranging yields:
164 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165

a b 1200
1200 Outer border of Outer border of
feasible vertical lift region feasible backward lift region
1000 Border of the gap within 1000 Border of the gap within
feasible vertical lift region feasible backward lift region
800 800

600 600

400 400

z [m]
z [m]

200 200

0 0

-200
-200

-400
-400

-600
-600
200 400 600 800 1000
200 400 600 800 1000
x [m] x [m]

Fig. 12. Regions where (a) vertical lift is possible and (b) backward lift with vx = 0.25 m/s is possible, both for Db = 1°.

2
x_ 2 þ z€z þ z_ 2 ðxx_ þ zz_ Þ near zero. Fig. 12a shows the feasibility analysis results for the
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ 0 ð30Þ given system parameters, vertical lift (vx = 0), and the collinearity
x2 þ z 2 ð x2 þ z 2 Þ
3
condition error margin Db = 1°. Evidently, as the horizontal distance
from the winch (x = x0) is increased, the vertical lift feasibility
Eq. (30) can be rewritten into the following state-space kinematic
region expands externally, and also the prohibitive internal gap
model, with the state variables defined as p1 = x, p2 = z, p3 = vx,
diminishes. This explains the optimization results which showed
and p4 = vz:
that the favourable vertical lift trajectory needed to be farther from
p_ 1 ¼ x_ the winch (see Fig. 7).
p_ 2 ¼ z_ The analysis has been extended to the case of backward lift,
where a small negative velocity vx is prescribed. The feasibility
p_ 3 ¼ 0 ð31Þ
2 3 results, shown in Fig. 12b for vx = 0.25 m/s, indicate that the fea-
pffiffiffiffiffiffiffiffiffiffi
p21 þp22 ðp1 p3 þp2 p4 Þ2 p23 þp24
4 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  pffiffiffiffiffiffiffiffiffiffi5 sibility region becomes significantly narrower externally. This may
p_ 4 ¼ p
2 3 p21 þp22 explain why the optimization results mostly included vertical ABM
ðp21 þp22 Þ
trajectories, as opposed to backward ones, despite the fact that the
When fed by the inputs vr = const. and vx = const., the model latter can give somewhat higher power as shown in the previous
(31) gives a set of simulation results for x(t), z(t), vx,z(t) and ax,z(t). subsection.
However, these simulation results may not be achievable based on
the available set of forces acting on the ABM. In order to verify the
feasibility of ‘‘prescribed’’ motion, the simulation results are used 5. Conclusion
as inputs of the dynamic model described in Section 2. In this
way, time response of the required resultant ABM force compo- The conducted optimization study has shown that the optimal
nents Fx and Fz is obtained by inverted dynamic Eqs. (16)–(19). power production cycle of the presented Magnus effect-based air-
Similarly, the individual ABM forces, other than rope force at borne wind energy system should take place at a position far from
winch, Fr,w, can be calculated from ‘‘prescribed’’ kinematic model the winch, so that a vertical or slightly upstream-inclined trajec-
simulation results (this includes Fb, Fg, Fl, Fd and Wr, see Eqs. (1)– tory can be produced for a reasonably long cycle period. The opti-
(9) and (12)). Subtracting these individual force components from mization results indicate that optimal, near-vertical motion is not
the resultant ones, and taking into account Eq. (11) gives the preconditioned by the possibility to realize large cylinder rotation
winch-side rope force components Fr,w,x = Fr,x and Fr,w,z. speeds, although the influence of larger speeds is beneficial. If the
The kinematic model response will be feasible at any instant at cylinder rotation speed is not limited and for the particular system
which the winch-side rope force, obtained in the above way, satis- parameters that were used, the optimization study has shown that
fies the rope collinearity condition, which states that the rope force the potential increase in power production compared with the pre-
needs to be collinear with the rope direction. That is, the following viously utilized (conventional) control strategy, giving a radial
equation needs to be satisfied: ABM trajectory, can reach over 120%.
F r;w;z x The analysis of the optimization results has shown that the
bF ¼ arctan ¼ b ¼ arctan ð32Þ physical reason for the optimal, near-vertical shape of the trajec-
F r;x z
tory is the possibility of achieving large air-ABM relative velocities
where bF is the angle of rope force at winch. A small deviation from during the ABM lift, which induces large aerodynamic (and conse-
ideal trajectory may be tolerated, so some small error margin may quently rope) forces and thus facilitates higher power production.
be allowed: Db = |bF – b|  b. The lift feasibility simulation tests The kinematic model-based feasibility analysis has confirmed that
can now be performed individually for a succession of ABM initial the optimal vertical trajectory can be achieved for a longer (practi-
x values, i.e. horizontal positions, while initial height z is always cally sustainable) period of time only if it is horizontally quite dis-
 M. Milutinović et al. / Energy Conversion and Management 90 (2015) 154–165
tant from the winch position, especially if the rope speed is to be
reasonably high.
Based on the obtained open-loop optimization results and their
physical interpretation, the next step is the development of a
robust closed-loop control system, such as the one proposed in
[26], where it has been shown that the benchmark optimization
results can be closely matched by a realistic closed-loop control
system.
Acknowledgments
It is gratefully acknowledged that this work has been supported
by the European Commission through the ‘‘High Altitude Wind
Energy’’ FP7 project, Grant No. 256714. The authors would also like
to express their gratitude to the project coordinator Omnidea Lda
for the support extended on the project activities.
Appendix A. Values of model parameters
Fg = mABM, g = 1471.5 N, g = 9.81 m/s2, lcyl = 21 m, mABM = 150 kg,
rcyl = 1.75 m, vw = 10 m/s (horizontally), wr = 2.942 N/m, q =
1.18 kg/m3.
Nondimensional drag and lift aerodynamic coefficients (func-
tions of speed ratio X):
C d ¼ 0:0211  X 3 þ 0:1837  X 2 þ 0:1183  X þ 0:5
C l ¼ 0:0126  X 4  0:2004  X 3 þ 0:7482  X 2 þ 1:3447  X
References
[1] Bansal RC, Bhatti TS, Kothari DP. On some of the design aspects of wind energy
conversion systems. Energy Convers Manage 2002;43(16):2175–87.
[2] Archer CL, Caldeira K. Global assessment of high-altitude wind power. Energies
2009;2(2):307–19.
[3] Archer CL, Delle Monache L, Rife DL. Airborne wind energy: optimal locations
and variability. Renewable Energy 2014;64:180–6. http://dx.doi.org/10.1016/
j.renene.2013.10.044.
[4] Ahrens U, Diehl M, Schmehl R, editors. Airborne wind energy. Springer; 2013.
ISBN 978-3642399640.
[5] Fagiano L, Milanese M. Airborne wind energy: an overview. In: Proceedings of
the 2012 American control conference. Montreal, Canada; 2012. p. 3132–3143.
[6] Lansdorp B, Ockels WJ. The Laddermill – innovative wind energy from high
altitudes in Holland and Australia. Global Windpower ‘06 Conference &
Exhibition. Adelaide, Australia; September 2006.
165
[7] Canale M, Fagiano L, Milanese M. KiteGen: a revolution in wind energy
generation. Energy 2009;34(3):355–61.
[8] Roberts BW, Shepard DH, Caldeira K, Cannon ME, Eccles DG, Grenier AJ, et al.
Harnessing high-altitude wind power. IEEE Trans Energy Convers 2007;22(1):
136–44.
[9] Vermillion C, Grunnagle T, Lim R, Kolmanovsky I. Model-based plant design
and hierarchical control of a prototype lighter-than-air wind energy system,
with experimental flight test results. IEEE Trans Control Syst Technol
2014;22(2):531–42.
[10] http://www.google.com/makani/.
[11] Penedo RJM, Pardal TCD, Silva PMMS, Fernandes NM, Fernandes TRC. High
altitude wind energy from a hybrid lighter-than-air platform using the
magnus effect. In: Ahrens U, Diehl M, Schmehl R, editors. Ch. 29 of Airborne
wind energy. Springer; 2013. ISBN 978-3642399640.
[12] Perković L, Silva P, Ban M, Kranjčević N, Duić N. Harvesting high altitude wind
energy for power production: The concept based on Magnus’ effect. Appl
Energy 2013;101:151–60.
[13] Pavković D, Hoić M, Deur J, Petrić J. Energy storage systems sizing study for a
high-altitude wind energy application. Energy 2014;76:91–103. http://
dx.doi.org/10.1016/j.energy.2014.04.001.
[14] Houska B, Diehl M. Optimal control of towing kites. In: Proceedings of the 45th
IEEE conference on decision and control. San Diego, USA, 2006. p. 2693–97.
[15] Houska B, Diehl M. Optimal control for power generating kites. In: Proceedings
of the 9th European control conference. Kos, Greece; 2007. p. 3560–67.
[16] Houska B, Diehl M. Robustness and stability optimization of power generating
kite systems in a periodic pumping mode. In: 2010 IEEE international
conference on control applications, Yokohama, Japan; 2010. p. 2172–77.
[17] Fagiano L, Milanese M, Piga D. Optimization of airborne wind energy
generators. Int J Robust Nonlinear Control 2012;22(18):2055–83. http://
dx.doi.org/10.1002/rnc.1808.
[18] Loyd ML. Crosswind kite power. J Energy 1980;4(3):106–11.
[19] Rutquist PE, Edvall MM. PROPT – Matlab Optimal Control Software. Users
guide, April 26; 2010.
[20] Aissaoui AG, Tahour A, Essounbouli N, Nollet F, Abid M, Chergui MI. A Fuzzy-PI
control to extract an optimal power from wind turbine. Energy Convers
Manage 2013;65:688–96.
[21] Nguyen A, Lauber J, Dambrine M. Optimal control based algorithms for energy
management of automotive power systems with battery/supercapacitor
storage devices. Energy Convers Manage 2014;83:58–72.
[22] Milutinović M, Kranjčević N, Deur J. Multi-mass dynamic model of a variable-
length tether used in a high altitude wind energy system. Energy Convers
Manage 2014;87:1141–50. http://dx.doi.org/10.1016/j.enconman.2014.04.
013.
[23] White FM. Fluid mechanics. McGraw-Hill; 1998, ISBN 978-0072281927.
[24] Betts JT. Practical methods for optimal control using nonlinear programming.
SIAM 2001. ISBN: 9780898714883.
[25] Bertsekas DP. Dynamic programming and optimal control, vol. 1. Athena
Scientific; 2005, ISBN 9781886529267.
[26] Milutinović M, Čorić M, Deur J. Optimization-based control strategy for a
Magnus effect-based airborne wind power production system. In: 9th
Conference on sustainable development of energy, water and environment
systems digital proceedings, Venice-Istanbul, September 20–27; 2014.