Downloaded from orbit.dtu.dk on: janv.

19, 2025

A Mixed-Integer Linear Programming approach to wind farm layout and inter-array

cable routing

Fischetti, Martina; Leth, John-Josef; Borchersen, Anders Bech

Published in:
Proceedings of 2015 American Control Conference

Link to article, DOI:

10.1109/ACC.2015.7172266

Publication date:
2015

Document Version
Publisher's PDF, also known as Version of record

Link back to DTU Orbit

Citation (APA):
Fischetti, M., Leth, J.-J., & Borchersen, A. B. (2015). A Mixed-Integer Linear Programming approach to wind
farm layout and inter-array cable routing. In Proceedings of 2015 American Control Conference (pp. 5907-5912).
IEEE. https://doi.org/10.1109/ACC.2015.7172266

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright
owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

 Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
 You may not further distribute the material or use it for any profit-making activity or commercial gain
 You may freely distribute the URL identifying the publication in the public portal

If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
2015 American Control Conference
Palmer House Hilton
July 1-3, 2015. Chicago, IL, USA

A Mixed-Integer Linear Programming approach to wind farm layout

and inter-array cable routing
Martina Fischetti1 , John Leth2 and Anders Bech Borchersen2

Abstract— A Mixed-Integer Linear Programming (MILP) energy production due to the reduction in wind speed and
approach is proposed to optimize the turbine allocation and to the increase in turbulence intensity [10]. In practice, a
inter-array offshore cable routing. The two problems are turbine that is downstream of multiple turbines is affected
considered with a two steps strategy, solving the layout problem
first and then the cable problem. We give an introduction to by all upstream turbines simultaneously, and the overall
both problems and present the MILP models we developed to effect is a nonlinear function of individual wakes. There are
solve them. To deal with interference in the onshore cases, we different analytical equations to describe the superposition
propose an adaptation of the standard Jensen’s model, suitable of multiple wakes, some being closer to the physical reality
for 3D cases. A simple Stochastic Programming variant of our than others [3]. It is however very difficult to incorporate
model allows us to consider different wind scenarios in the
optimization. For the inter-array cable routing, we propose a the more accurate wake equations into a mathematical
new MILP model able to deal with different constraints arsing programming model due to their nonlinearity: currently,
in practical application, such as capacity limitations, substation only heuristics [14] [17] [12] [2] are able to take accurate
limitations and non-crossing constraints. Computational results wake models into account. In our optimization we are
on real-world instances prove the practical viability of the therefore considering Jensen’s model [10], adapting it to the
approach.
onshore case, and we assume pairwise interference between
I. I NTRODUCTION sites (i.e., assuming the interferences as cumulative). We
use Stochastic Programming to consider different wind
Two topics of great interest in the wind energy sector are scenarios in our optimization.
investigated in the present paper: wind turbine allocation Another problem considered is the inter-array cable routing
and their optimal connection among cables. Our models for optimization. Different cables, with different capacities and
the wind farm allocation fit both onshore and offshore cases, costs, exist and a correct use of them can lead to large
while in the cable routing optimization we focused on the savings, since the cost of cables constitute a significant
offshore case (the so called ”inter-array” cable connection part of the establishment costs of a wind turbine park. This
problem). In the offshore layout optimization the model we optimization problem has been studied in [1] and [16].
propose is also taking costs of foundation into account. Here we developed a different model able to deal with
different constraints appearing in practical application, such
Wind farm layout optimization problems deal with the as non-crossing constraints.
optimal placement of turbines in a wind farm field. Currently,
metaheuristics and greedy approaches have been used. Some Hence the contributions from the present work is the
existing heuristic methods [14] [17] and Mixed-Integer development of original MILP models to solve the two
Linear Programming models [4] [16] [18] [20] [13] have problems with specific constraints. In the present paper,
explored the problem with discretization, while a continuous for both models, some interesting tests are comparing how
approach has been used in [12] [11]. The continuous models the optimal (from a wind-resources or cable-price point of
are highly nonconvex and turn out to be intractable from a view) wind farms should look like, compared with the ones
computational viewpoint when considering real-world cases, actually build nowadays. A detailed use of MILP-based
therefore we preferred a discrete programming approach. heuristics to speed up the current model in difficult test-cases
The MILP approach with some ad-hoc heursitics is able is presented in [13]. An original adaptation of the classical
to solve large-instances problems as shown in [13]. An Jensen’s model is here presented.
interesting feature of the wind farm optimization problem
is that of dealing with the aerodynamic interaction among The present paper is organized as follows: In Sect. II we
multiple turbines. In a simple scenario with only two introduce the MILP model we used for the layout optimiza-
turbines, the turbine downstream is said to be in the wake tion problem. The interference between turbines is taken into
region of the upstream turbine, and it experiences a loss in account in the model, and Sect. III explains how we adapted
1
the Jensen’s model to work in a 3d onshore scenario. In Sect.
Operations Research, DTU Management, Produktionstorvet,
426 B, DK-2800 Kgs. Lyngby [email protected] and Vattenfall IV we report some tests for layout optimization. The cable
BU Renewables Oldenborggade 25-31, 7000 Fredericia, Denmark routing optimization is treated in Sect. V, where we present
[email protected] 2 Automation and our MILP model. Sect. VI compares the cable layout from
Control, Dept. of Electronic Systems, Aalborg University, Fredrik
Bajers Vej 7C, 9220 Aalborg Denmark [email protected] and our tool to existing ones, showing how the use of our tool
[email protected] can lead to large savings. Finally Sect. VII summarizes the

978-1-4799-8684-2/$31.00 ©2015 AACC 5907

contributions of the paper.
P
II. L AYOUT OPTIMIZATION max z= i∈V (Pi xi − wi ) (3)
P
This section is based on the original results presented in s.t. NM IN ≤ i∈V xi ≤ NM AX (4)
[6], [13], however here we also include, as a first result, the xi + xj ≤ 1 ∀{i, j} ∈ EI (5)
cost of foundations in the offshore case. P
Iij xj ≤ wi + Mi (1 − xi ) ∀i ∈ V (6)
j∈V
xi ∈ {0, 1} ∀i ∈ V (7)
Our model determines a feasible allocation of turbines that
maximizes power production. The building area (site) and its wi ≥ 0 ∀i ∈ V (8)
resource maps are given on input. The optimizer considers where the big-M term
the following constraints: X
a) a minimum and maximum number of turbines that can Mi = Iij
j∈V
be built; [i,j]6∈EI

b) a minimum separation distance between any pair of is used to deactivate constraint (6) in case xi = 0. The
turbines to ensure that the blades do not physically clash model above follows a recipe of Glover [9] which is widely
(turbine proximity constraints); used, e.g., in the Quadratic Assignment Problem [19], [8].
c) interference between installed turbines; It allows our model to work on larger instances compared
d) cost of foundations in the offshore case. with equivalent models in the literature (we refer to [6] for
Let V ⊂ R2 denote the set of possible positions for a turbine, details).
called “sites” in what follows, and let The definition of the turbine power Pi and of the inter-
• Iij be the interference (loss of power) experienced by ference Iij depends on the wind scenario considered, that
site j when a turbine is installed at site i, with Ijj = 0 greatly varies in time. Let us introduce a new variable zi,j
for all j ∈ V ; equal to 1 if turbines are built both in position i and j, 0
• Pi be the power that a turbine would produce if built otherwise. With respect to (1), zi,j as the same meaning
(alone) at site i; of xi xj . Using statistical data, one can in fact collect a
• NM IN and NM AX be the minimum and maximum large number, say K, of wind scenarios k, each associated
number of turbines that can be built, respectively; with Pik , Ii,j
k
and with a probability πk . Using that data, one
• DM IN be the minimum distance between two turbines; can write a Stochastic Programming variant of the previous
• dist(i, j) be the distance between sites i and j. model where only the objective function needs to be modified
as
In the sequel we assume that a numbering have been chosen K
X X XX
Pik xi − k


among the elements of V and by abuse of notation we also z= πk Iij zij (9)
let i denote the numbering of element i ∈ V . In addition, let k=1 i∈V i∈V j∈V
GI = (V, EI ) denote the incompatibility graph with while all constraints stay unchanged as they only involve
EI = {{i, j} : i, j ∈ V, dist(i, j) < DM IN , i 6= j} “first-stage” variables x and z. It is therefore sufficient to
define
note that {i, j} = {j, i} by convention. Let n := |V | denote K
X
the total number of sites. Pi := πk Pik ∀i ∈ V (10)
In our model, binary variables are defined for each i ∈ V : k=1
 K
1 if a turbine is built at site i ∈ V ; Iij :=
X
k
πk Iij ∀i, j ∈ V (11)
xi =
0 otherwise k=1
The original quadratic objective function (to be maximized) to obtain the same model (1)–(8) as before. The above
X X X model can be solved for large-scale instances (around 20 000
Pi x i − ( Iij xj ) xi (1)
possible positions to locate turbines) in a few minutes on a
i∈V i∈V j∈V
standard PC, using some ad-hoc heuristics and a MILP-based
can be restated as heuristic called ”Proximity Search” [7]. We refer to [13] for
X details.
(Pi xi − wi ) (2)
In the offshore case, and contrary to [6], we here also add
i∈V
the cost of foundation, as the building cost highly depends
where on the sea depth and type of turbine, and this affects the final
optimal turbine allocation (for further details see [15]). To
 P
X
j∈V Iij xj if xi = 1;
wi := Iij xj xi = consider that, the objective function (9) becomes
0 if xi = 0.
j∈V
X p(di )
denotes the total interference caused by site i. Our compact maximize [(Pi − )xi − wi ] (12)
F ACT
model then reads i∈V

5908
 Fig. 2: A simple scheme for our 3D Jensen’s model

Fig. 1: Average wind speed [m/s] at 80m agl (above ground

level) for our onshore test site

Where p(di ) is the price of foundation as a function of the

depth di of the ith position, while F ACT is a factor to
scale the price from e/KW to MW in a typical 20-years
production horizon. All the others constraints of the model
stay unchanged. Fig. 3: Optimal layout solutions imposing maximum number
of turbine equal to 20 (right) and 30 (left)
III. I NTERFERENCE IN ONSHORE CASES
To compute the interference matrix we used the Jensen’s not considering wind flow inclination. That means that the
model [10] , adapting it to a 3D scenario. That means that, interference cone is always horizontal and does not depend
under proper assumptions, Jensen’s model can be as well on the shape of mountains and hills. Due to this assumption,
used in an onshore setting, where the terrain topology plays our 3D wake cone would look like the one in Figure 2.
a role. The offshore case is a 2D case since the sea level can The wind speeds U and V are the average wind speed
be assumed as constant and the wind as equally distributed given on input for points i and j respectively (i.e., we take
in the site. In this framework, the standard Jensen’s formula the values shown in Figure 1 for the real-world test site).
computes the loss in wind speed δV . So if we want to Because of this, value δV computed by (13) cannot be used
compute the loss δV on a turbine located in the position as it is, as this could even produce a negative wind value at
(say) j due to the presence of an upwind turbine in position point j. Therefore a proportion has been used to compute
(say) i, we apply the following formula wind reduction at point j due to the presence of a turbine at
point i. To be more specific, δV is still calculated according
√ 
D
2 to (13) but the wind reduction applied in j is δVU V , i.e., the
δV = U (1 − 1 − Ct ) D+2kX (13) new wind, V 0 , in j is now computed as
where δV
V0 =V − V (14)
U = upwind speed at turbine i U
Ct = thrust coefficient corresponding to Wind direction is also different from point to point. However,
wind speed U considering a wind direction time series for each point would
D = rotor dimension (diameter) be computationally too heavy, so we consider only one of
X = distance between the first coordinates such series to compute the average interference.
of i and j (i.e. on the x-axis)
k = wake decay IV. T ESTS
( constant, typically
0.075 onshore Our model was first tested on an offshore grid of 10 × 10
k= possible positions in a square of 3000×3000 m2 , with DM IN
0.050 offshore
400 m. Wind scenarios used to compute the interference
matrix are taken from the European Wind Energy Associa-
The wind speed at j is computed as V = U − δV . tion (EWEA) data, aggregated in bins according to 24 wind-
Determining interference for onshore problems is actually angles (15 degrees each) and wind speeds in 1m/s per bin.
much more difficult than in the offshore (2D) case, due to The first set of plots (Figure 3 ) shows how the optimal
the presence of nonuniform wind. layout changes when imposing a different maximum number
Due to possible different heights even without any interfer- of turbines. Red circles represent the built turbines, while the
ence, wind is no longer uniform in the site. Figure 1, for color in the background refers to the interference induced by
example, shows how the wind is distributed in average in a those turbines.
real-world onshore site in United Kingdom. It interesting to notice that the optimal solutions have
Some assumptions therefore need be formulated to be able as many turbines as possible on the boundaries: those are,
to extend Jensen’s model to a 3D scenario. First, we are indeed, the positions in which a turbine would cause less

5909
Fig. 4: Allocation of wind turbine in the Horns Rev 3
test case. The colors in the background represent the water
depth. The first plot is the solution non considering cost
of foundations, the second one is the solution considering
cost of foundations. In the second case the turbines are
concentrated in the less deep area Fig. 5: The first plot shows feasible positions while the
second shows the power distribution within the site. All other
interference to the others. This means that, from a wind- plots show our layout solutions when considering a maxi-
resources point of view, turbines should not be allocated mum number of turbines equal to 10 and 15, respectively
regularly, as we are used to see nowadays.
WindPRO (WP) MILP heuristic
If we use the objective function (12) the solution layouts max n. turb. layout production layout production production layout production Improvement
by WindPRO (our obj value) difference (our obj value) wrt WindPRO
change due to the cost of foundations. If we for example 5
(MWh/y)
51 695
(MWh/y)
51 776 0.2%
(MWh/y)
52 123 0.7%
consider the real-world case of Horns Rev 3, which infor- 10
15
99 330
143 727
99 577
144 789
0.2%
0.7%
101 024
146 454
1.5%
1.1%
mation are available in [5], Figure 4 shows how the optimal 20
22
185 285
200 699
185 252
200 412
0.0%
-0.1%
188 782
205 039
1.9%
2.3%
layout would change using the model without and with cost 25
(unlimited) 28
200 699
200 699
200 412
200 412
-0.1%
-0.1%
228 270
247 569
13.9%
23.5%
of foundations. In this test we imposed the construction of
TABLE I: Summary of the comparison between Optimize
20 Vestas V164 wind turbines in the site, minimum distance
and our tool. Note that WindPRO’s Optimize is not able to
of 10 rotor diameters.
locate more than 22 turbines.
In the onshore case, instead, the wind is not homoge-
neously distributed, i.e., a turbine can give different power
productions depending on its position (even without con- • the energy flow in each connection cannot exceed the
sidering any interference). As a consequence, turbines are capacity of the installed cables
concentrated in the areas with the best wind. We considered • different cables, with different capacities and costs, can
a real-world site in United Kingdom, where the wind is be installed
distributed as shown in Figure 1. Some positions (red in the • cable crossing should be avoided since it is expensive
first plot of 5) have not been considered in the optimization (requires an expensive bridge structure, it decreases the
since unavailable for building turbines (due to extreme wind capacity of the crossing cables and drastically increase
speed, extreme turbulance, high inclination and too high the risk of damages)
shear coefficient). The allowed positions turn out to identify • a given maximum number of strings can be connected
3103 potential points (blue area in the first plot in Figure 5). to each substation
The potential power distribution in this area is shown in the
second plot of Figure 5. Let us consider the turbine positions as the nodes of a
We compared our onshore solutions with the solutions directed graph G = (V, A) and all possible connections
given by a commercial software for the same study-case. between them as directed arcs. Some nodes correspond to
This commercial software is using heuristics to determine the the substations that are considered as the roots of the trees,
layout. In the tests we impose for both software a maximum and are the only nodes that collect energy. We can therefore
number of turbine to built in the area (5,10,15,20,22,25,un- indicate with Ph the power production at node, so
limited). The results of the comparison, in Table I, show the (
≥0 if the h-th node is a turbine
effectiveness of our method, since we are able to outperform Ph (h ∈ V )
the competitor in all tests. = −1 if the h-th node is a substation
Suppose that T different types of cable can be used. Each
V. MILP MODEL FOR C ABLE ROUTING O PTIMIZATION
type of cable t has its given capacity kt and its cost cti,j . This
As a second result we now present inter-array cable cost can be defined as cti,j = dist(i, j)ut for each arc (i, j)
optimization. Here we consider the turbine layout as given and for each type t ∈ T , where dist(i, j) is the distance
and we want to find an optimal cable connection between between turbine i and turbine j and ut is the unit cost of
all turbines and the given substations, minimizing the total cable type t.
cable cost. The optimization considers that: We used the continuous variables fi,j ≥ 0 to indicate the
• the energy flow leaving a node must be supported by a flow from i to j, and the binary variables xti,j to define the
single cable cable layout, such that

5910
 ( Cost (e) Strings to substation Crossings Cost saving (e) Cost saving (%)
1 if arc (i, j) is built with the cable type t 18 918 866 8 3 0 0%
xti,j = 17 522 596 8 0 1 396 270 7.38 %
0 otherwise
(i, j ∈ V, t ∈ T ) TABLE II: The first line records the cost of the layout
proposal by Vattenfall (manual solution). Second line refers
Finally, variables yi,j indicate whether an arc between tur-
to the layout found by our tool when considering non-
bine i and j is built (with any type of cable). Consequently,
crossing constraint and maximum number of strings to the
yi,j is a binary variable
( substation equal to 8.
1 if arc (i, j) is built
yi,j = (i, j ∈ V, t ∈ T )
0 otherwise 59
72
39
59

1
72

1
2 1
39 1 3 71
58
38

xti,j ,
2

Note that variables yi,j are related to variables

71

namely 19 1
38 2
58

3
4
19

3
2

37
2

57
70

t 70 3
P
6
18

x
t∈T ij = y ij for all i = j in V . Our model then reads: 18 1
37

2
57

4
69
5 4
1

36
56
3
69

4
56 17
4
36 6 2
17 68
t t 5
P P
t∈T ci,j xi,j
55
min i,j∈V (15) 16
1
3

35
55

6
68 16

1
35

4 54
34
t 2
P
∀i, j ∈ V : j 6= i (16)
15

s.t. t∈T xi,j = yi,j

54 1
34 1 3 5
15 7 7 2 5
5 53
3 33
53 14
P 33
2

i:i6=h (fh,i − fi,h ) = Ph ∀h ∈ V : Ph ≥ 0 (17)

14 3
6 52 2
4 32
52 8 13
32
13 3
4
t 7
P
t∈T kt xi,j ≥ fi,j ∀i, j ∈ V : j 6= i, (18) 5 31
12
31
12
9
51
5
6
51
8 5
11
P 9

∀h ∈ V : Ph > 0 (19)
8 3

y h,j = 1 11
1

Pj:j6=h
9 50
7 30 9
50
10 9
30 9 9 3
10 Drain
9

yh,j ≤ 1 ∀h ∈ V : Ph = 0 (20) 8 9
Drain
9
49
1 9
29
49

Pj:j6=h
9 29 9 9 9
9 9 9
3 3
8 48

∀h ∈ V : Ph < 0 (21)
8

yh,j = 0 8
28
48
8
28
8

P j:j6=h
8 8 67
7 67 2 47
8 47 27
27 7 5 5 2

i6=h yi,h ≤ M S ∀h ∈ V : Ph < 0 (22) 7 7 7

7 7
6 1 66
7 66 46
46 26
26 6 1
6 6 6
6 6

xti,j ∈ {0, 1}
65

∀i, j ∈ V, t ∈ T
5

(23) 5
6
25
45

5
65

5
5
25

5
45

5
4 4 64
5 64 44
44 24

yi,j ∈ {0, 1} ∀i, j ∈ V (24) 4

4
24

3
43
4
63
4 4

3
4
43
4
63
4

23 23
3 3 3
3 3

fi,j ≥ 0 ∀i, j ∈ V, j 6= i
3

(25) 3

2
22
2
42
62

2
2

2
22
3
42

2
62

2
2
1 2
61 1 61
2 41 41
21

The objective function (15) minimizes the total cable 1

1
21

40
1
60
1 1
1
40
1
60
1

20 20

layout cost. Constraints (16) impose that only one type

of cable can be selected for each built arc, and defines (a) Vattenfall’s proposal (b) The layout from our
(18 918 866 e) tool (17 522 596 e)
the yi,j variables. Constraints (17) are flow conservation
constraints: the energy (flow) exiting each node h is equal to Fig. 6: Comparison of the layouts for Sandbank’s cable
the energy entering h plus the power production of that node routing. Our layout (on the right) is 1 396 270 e (7.38 %)
(if Ph ≥ 0). Note that the constraint is not imposed if h is less expensive.
a substation. Constraints (18) ensure that the flow does not
exceed the capacity of the installed cable, while constraints routing design is nowadays carried out manually, so we
(19), (20), and (21) impose that only one cable can exit compared our results with manual solutions. Note that some
a turbine and none can exit the substation (tree structure). crossings are allowed in the manual layouts with the idea to
Finally, constraint (22) imposes the maximum number of avoid them at construction time. Those crossings are easy
strings (M S) that can enter the substation(s). to avoid by just laying cables in curved trajectories. The
Non-crossing constraints are difficult constraints from the tests prove that the use of our automatized tool provides big
optimization point of view. In the model we have to consider savings.
the complete set of all possible connections, so theoretically Table II and Figures 6a-6b compare the cable layouts for
we should impose that each two arcs of the complete graph Sandbank offshore wind farm. This wind farm contains 72
G should not cross. Of course this implies to deal with a wind turbines and one substation. The substation allows up
huge number of constraints. We have, therefore, used an to 8 strings to be connected. Three types of cables are
approach on the fly, as suggested in [1], where the optimizer considered in the optimization:
will consider model (15) - (25) and add on the fly the new • the first type (blue) has a price of 135e/m and can
constraints: whenever two build arcs (i, j) and (h, k) cross support up to 5 turbines
yi,j + yj,i + yh,k + yk,h ≤ 1 (26) • the second type (green) costs 250e/m and can support
up to 7 turbines
Using this approach, the number of non-crossing constraints
• the third type (red) costs 370e/m and can support up
actually added to the model decreases dramatically, making
to 9 turbines
the optimization feasible.
Those significant savings are achieved by a smart use of
VI. C OMPARISON WITH EXISTING LAYOUT the less expensive cables. This results clear from Figures 6a-
We compared our cable routing results with some existing 6b Table III shows how much the use of the expensive cables
layouts, kindly provided us by Vattenfall AB [21]. The cable impacts the final cost of the Sandbank solution.

5911
 Cost (e) tot.length blue cables(m) tot.length green cables(m) tot.length red cables(m) Cost saving (e)
18 918 866
17 522 596
40 493.37
83 611
18 264
4 021
24 016
14 134
0
1 396 270
ACKNOWLEDGMENT
TABLE III: The first line refers to the layout proposed by We thank Michele Monaci (DEI, University of Padova)
Vattenfall for Sandbank. The second line (yellow) refers to who supervised the thesis work of first author regarding
the solution found by our tool. Second, third and fourth wind farm layout, and Jesper Runge Kristoffersen and Iulian
columns record the total length of the three different type Vranceanu from Vattenfall AB who helped in defining the
of cables. cable routing constraints and provided the real-world cases
used in our tests.
Cost (e) Strings to substation Crossings Cost saving (e) Cost saving (%) R EFERENCES
23 515 587 12 2 0 0%
21 490 400 12 0 2 025 187 8.6 %
[1] Bauer J., Lysgaard J.: The onshore wind farm array cable layout
problem: a planar open vehicle routing problem. Journal of the
TABLE IV: The first line records the cost of the layout Operational Research Society (2013)
proposal by Vattenfall (manual solution). Second line refers [2] Chowdhury S., Messac A, Zhang J., Castillo L., Lebron J.: Optimizing
to the layout found by our tool when considering non- the unrestricted placement of turbines of differing rotor diameters in
a wind farm for maximum power generation. In: Proceedings of the
crossing constraint and maximum number of strings to the ASME 2010 IDETC/CIE 2010 Montreal (2010)
substation equal to 12. [3] Renkema D.J.: Validation of wind turbine wake models. Master’s
thesis, Delft University of Technology (2007)
[4] Donovan S.: Wind farm optimization. In: Proceedings of the 40th
Table IV refers to a real-world wind farm containing Annual ORSNZ Conference, pp. 196-205 (2005)
[5] Energinet.dk: Technical project description for the large-scale onshore
around 50 8MW turbines1 and one substation with at most wind farm (400mw) at horns rev 3, document no. 13-93461-267. Tech.
12 connections to the substation. In our tests we considered rep., Energinet.dk (2013)
two types of cables: [6] Fischetti M.: Mixed-integer models and algorithms for wind farm lay-
out optimization and cable routing. Master’s thesis, Aalborg University
• the less expensive type of cable (180.0e/m) can support (2014)
up to 3 turbines [7] Fischetti M., Monaci M.: Proximity search for 0-1 mixed-integer con-
• the most expensive (370e/m) can support up to 5 vex programming. Tech. rep., DEI, University of Padova (submitted)
(2012)
turbines [8] Fischetti M., Monaci M., Salvagnin D.: Three ideas for the quadratic
Table V refers to another real-world wind farm consisting assignment problem. Operations Research 60(4), 954-964 (2012)
[9] Glover F.: Improved linear integer programming formulations of
of around 70 6MW turbines 1 and 1 substation. The maxi- nonlinear integer problems. Management Science 22, 455-460 (1975)
mum number of strings connected to the substation is still [10] Jensen N.: A note on wind generator interaction. Tech. rep., Tech-
12. The two types of cables are considered: nical Report Riso-M-2411(EN), Riso National Laboratory, Roskilde,
Denmark (1983)
• the less expensive cable (180.0e/m) can support up to
[11] Kusiak A, Song Z.: Design of wind farm layout for maximum wind
4 turbines energy capture. RenewEnergy 35, 285694 (2010)
• the most expensive cable (370e/m) can support up to [12] Kwong W.Y., Zhang P.Y., Romero D., Moran J., Michael M., Amon
C.: Multi-objective optimization of wind farm layouts under energy
6 turbines generation and noise propagation. In: Proceedings of the ASME 2012
IDETC/CIE 2012, Chicago (2012)
VII. C ONCLUSION [13] Fischetti M., Monaci M.: Proximity search heuristics for wind farm
In the present paper a Mixed-Integer Linear Program- optimal layout. To appear in Journal of Heuristics (2015), DOI:
10.1007/s10732-015-9283-4
ming (MILP) approach has been used to optimize turbine [14] Mosetti G., Poloni C., Diviacco B.: Optimization of wind turbine
allocation and inter-array offshore cable routing. The use positioning in large windfarms by means of a genetic algorithm.
of Stochastic Programming allowed to take into considera- Journal of Wind Engineering and Industrial Aerodynamics 51(1),
105116 (1994)
tion the variability of the wind in the layout optimization. [15] Nielsen P.: Offshore wind energy projects feasibility study guidelines
Different constraints arsing in practical application have sea wind altener project 4.1030/z/01-103/2001. Tech. rep., EMD
been considered in the optimization. A specific strategy has (2003)
[16] Fagerfjall P.: Optimizing wind farm layout: more bang for the buck
been developed, in particular, to deal with the no-crossing using mixed integer linear programming. Master’s thesis, Chalmers
constraints for cable routing. Computational results on real- University of Technology and Gothenburg University (2010)
world instances prove the practical viability of the approach [17] Grady S.A., Hussaini M.Y., Abdullah M.A.: Placement of wind
turbines using genetic algorithms. Renewable Energy 30(2), 259270
and show how the optimized wind-farm should look like. (2005)
1 the exact number of turbines used in the optimization is not disclosed
[18] Turner S.D.O, Romero D.A., Zhang P.Y., Amon C.H., Chan T.C.Y.:
A new mathematical programming approach to optimize wind farm
due to privacy issues layouts. Renew Energy 63, 674680 (2014)
[19] Xia Y., Yuan Y.: A new linearization method for quadratic assignment
Cost (e) Strings to substation Crossings Cost saving (e) Cost saving (%)
problem. Optimization Methods and Software 21, 803-816 (2006)
25 365 822 12 3 0 0%
23 848 104 12 0 1 517 717 5.98 % [20] Zhang P.Y., Romero D.A., Beck J.C.,Amon C.H.: Solving wind
farm layout optimization with mixed integer programs and constraint
TABLE V: Summary of the results. The first line reports the programs. EURO Journal on Computational Optimization 2 (3), 195-
219 (2014)
cost of the layout found manually. The second line (yellow) [21] Vattenfall, http://corporate.vattenfall.com/
refers to the solution found by our tool when considering
non-crossing constraint and maximum number of strings to
the substation equal to 12.

5912