BULGARIAN ACADEMY OF SCIENCES

CYBERNETICS AND INFORMATION TECHNOLOGIES  Volume 17, No 1

Sofia  2017 Print ISSN: 1311-9702; Online ISSN: 1314-4081
DOI: 10.1515/cait-2017-0007

A Two-Stage Placement Algorithm with Multi-Objective

Optimization and Group Decision Making
Daniela Borissova1,2, Ivan Mustakerov1
1Institute
of Information and Communication Technologies, BAS, 1113 Sofia, Bulgaria
2Universityof Library Studies and Information Technologies, 1784 Sofia, Bulgaria
Emails: [email protected] [email protected]

Abstract: A two-stage placement algorithm with multi-objective optimization and

group decision making is proposed. The first stage aims to determine a set of design
alternatives for objects placement by multi-objective combinatorial optimization. The
second stage relies on business intelligence via group decision-making based on
solution of optimization task to make a choice of the most suitable alternative. The
design alternatives are determined by means of weighted sum and lexicographic
methods. The group decision making is used to evaluate determined design
alternatives toward the design parameters. The described algorithm is used for wind
farm layout optimization problem. The results of numerical testing demonstrate the
applicability of the proposed algorithm.
Keywords: Placement algorithm, multi-objective alternatives determination,
business intelligence, group decision making, wind farm layout design.

1. Introduction
The ultimate measure of decision quality expresses the degree to which the decision
has led to the best business outcome. An advanced direction is using of business
intelligence to access and analyse information to improve and optimize decisions and
performance [21]. The complexity of decision making process is approached by
technologies and best practices to deliver the knowledge to make the right decisions
quickly and with confidence [13, 2]. Business decisions are made on the base of the
information for whole system incorporating financial and environmental aspects and
rely on experts with capabilities to evaluate decision alternatives. The problem of
determination of variety of alternatives for placement of objects and choosing the
best one by group decision making arises in many practical problems. One such
problem is designing of wind farm layout to maximize the energy output and
considering wind conditions, technical requirements and other different restrictions
[1]. The optimal placement of wind turbines that ensures maximum efficiency is
essential in designing of wind farm project in both cases of onshore and offshore
wind farm. Therefore, the proper optimization is imperative in designing renewable
87
energy system for maximizing its cost effectiveness. Designing of renewable energy
system considers the problem of placement of turbines with known dimensions in the
farm area to minimize the cost and maximize the system efficiency. Many research
articles are subject of Wind Farm Layout Optimization Problem (WFLOP) and
different approaches and various optimization algorithms are proposed as: genetic
algorithms [4, 15, 24, 26, 40], pattern search algorithm [7], evolutionary algorithm
[25, 35, 39]. An alternative approach to WFLOP is using of the mathematical
programming optimization methods [3, 5, 6, 19, 27, 34]. Due to the complex and
multi-disciplinary nature of WFLOP many conflicting objectives are to be involved
during the design process. In this sense, multi-objective optimization modelling can
more accurately reflect real life situations and are therefore more suitable tool to deal
with WFLOP [4, 20, 38]. From the managerial point of view it is hard to take decision
on the basis of single alternative. More substantiated decision would be reached if
different design alternatives conforming to different design requirements are
compared to get to the best design solution [19]. The design of wind farm is
multidimensional process and involves different aspects – economic, technical,
environmental, social, etc. This means that different design alternatives are to be
defined and evaluated by a group of experts to make the most suitable business
decision. Aggregating different experts’ evaluations over set of alternatives to select
a single collective alternative can be done by Group Decision Making (GDM) [12].
The GDM is based on evaluations of multiple alternatives against multiple criteria by
variety of experts to provide better decisions [2]. In many cases different points of
view about the criteria and their importance leads to conflicting evaluations of
alternatives which complicates the GDM. The further complication arises when
importance of the opinions of experts themselves are to be considered according their
positions in the project management. When the number of the criteria and alternatives
is finite, and the alternatives are given explicitly are known as Multi-Attribute
Decision Making (MADM). The problems MADM can be tackled by different
models and methods [29-33]. An important part of modern decision science is Multi-
Attribute Group Decision Making (MAGDM) where the process of decision is made
by multiple decision makers as assessment of alternatives over a range of attributes.
Two main directions in the MADM methods can be distinguished – methods based
on the Multi-Attribute Utility Theory (MAUT) and outranking methods [12, 16]. In
this article MAUT principles are realized by combinatorial optimization formulation
that takes into account the different experts’ opinions about the importance of criteria
(wind farm parameters), DMs’ evaluations of alternatives and weights of DMs.
The main contribution of the article proposes two-stage algorithm which
combines multi-objective optimization as design simulation tool and GDM based on
single-objective integer linear programming model, to get numerically reasoned
optimal design alternative. The rest of the article is structured as follows: Section 2
describes in details the proposed two-stage algorithm, Section 3 describes the
algorithm application for WFLOP; Section 4 illustrates the applicability of proposed
approach by numerical testing. In Section 5 analysis of the numerical testing results
and discussions are presented, and conclusions are given in Section 6.

88
2. A two-stage placement algorithm with multi-objective optimization
and group decision making
This section focuses on description of the proposed two-stage algorithm for
determination of design alternatives and choice of most suitable alternative (Fig. 1).

Fig. 1. Flowchart of the proposed two-stage algorithm

The two-stage algorithm for placement consists of: 1) determination of a set of

design alternatives conforming to different design scenarios; 2) choice of the best
alternative by GDM. It is assumed that overall project management is done by a Supra
Decision Maker (SDM). The SDM is authorized to organize the overall decision
making process, to gather a group of different experts (DMs) relevant to the project
design aspects and to collect experts’ evaluations of different alternatives. The SDM
is usually a manager who is responsible to propose a final decision at a higher level
of management or has the right to make the final decision by himself.
The responsibilities for input data about the parameters, objectives and their
importance are assigned to SDM on the Step 1.1 (Fig. 1). The goal of multi-objective
optimization problem is to identify Pareto-optimal solution using additional
information from DM about the objectives importance. On the Step 1.2 the experts
in Operation REsearch (OREs) have to formulate and to solve multi-objective
optimization tasks. They choose and implements proper multi-objective solution
method and express objectives importance. A common approach for dealing with
multi-objective optimization problems is to use scalarization techniques [8, 14, 23]
due to their simplicity and effectiveness. As a result of multi-objective task solution,
a Pareto-optimal design alternative is determined. Then, on Step 1.3, design
alternatives are filtered to get a set of unique alternatives.

89
 The input data for GDM is implemented in Step 2.1. SDM determines the
evaluation criteria and group of experts (DMs) from different knowledge domains
and defines some weight coefficients for each of DMs. Weights of DMs play a very
important role in MAGDM because they reflect the expertise of each particular
expert. The range of weights for experts could be limited within interval of 0 to 1,
where 1 means the most experienced and trusted level of expertise [12].
The essence of Step 2.2 is determination of weighted decision matrix. A key
point in decision making is determination of relevant evaluation criteria because they
significantly influence on the selection process. Usually these criteria are given by
high level management. Regardless of who sets the criteria and alternatives, these
two parameters are mandatory elements in WDM.
On Step 2.3 a single criterion optimization modelling is used to formulate an
optimization task for selection of the best alternative. This solution is presented to
SDM for a decision – to accept the determined alternative or to repeat stages of
algorithm.

3. Application of two-stage algorithm for design of WFLOP

Accordingly to the first stage of the described algorithm, the set of alternatives that
define various wind farm layouts are generated by using of multi-objective
optimization as follows:
Step 1.1. Collecting of the input data – set of different types of wind turbines
and characteristics of wind site and wind conditions. SDM plays a crucial role in
negotiating of the phrasing of the design objectives and in determination of their
importance.
Step 1.2. The multi-objective combinatorial optimization is used as an
analytical simulation tool for different wind farm layout design alternatives. For the
goal a combinatorial multi-objective optimization model is proposed. It allows
simultaneously determination of turbines’ type and number, and their placement
positions. The overall objective of a wind farm project accepted by many researchers
is to increase the Annual Energy Production (AEP) while reducing the costs. These
two objectives are used to simulate different wind farm layout design scenarios by
multi-objective combinatorial optimization:
(1) max AEP = hyNPwt
2 1 2
min Costs = 𝑁 ( + 𝑒 −0.00174𝑁 ) ,
3 3
subject to
(2) N = NxNy,
(3) Nx = (Lx/SDx) + 1,
(4) Ny = (Ly/SDy) + 1,
(5) SDx = kxDwt,
(6) SDy = kyDwt,
(7) k ymin ≤ ky ≤ k ymax , ky > 0,
(8) k xmin ≤ kx ≤ k xmax , kx > 0,

90
(9) 𝑃wt = ∑𝑚 𝑖
𝑖 𝑥𝑖 𝑃wt ,
𝑚 𝑖
(10) 𝐷wt = ∑𝑖 𝑥𝑖 𝐷wt ,
𝑚
(11) ∑𝑖 𝑥𝑖 = 1, 𝑥 ∈ {0, 1},
where the AEP is represented by the number of hours over the year (hy), coefficient
of the nominal wind power utilization (η) [27], number of turbines (N) and turbines’
rated power (Pwt). The second objective is non-dimensional costs per year
2 1 2
𝑁( + 𝑒 −0.00174𝑁 ) [15, 24, 36] expressed as a function of total turbines number
3 3
(N) while Nx is the number of turbines in rows and Ny is turbines number in columns.
𝑖
Rated power of i-th wind turbine is denoted by 𝑃wt with corresponding rotor diameter
𝑖
𝐷wt . The notations Lx and Ly are used for dimensions of wind farm area. The decision
variables (unknowns) xi  {0, 1} are used in the model (1)-(11) to choose the type of
turbines. Coefficients of turbines separation distances in rows and cols kx and ky are
used to determine the turbines separation distances respectively SDx and SDy. The
negative influence of the WE is taken into account by introduction of separation
coefficients kx and ky as variables. These coefficients are used to define the separation
distances (as number of turbine’s diameter) needed to avoid the influence of WE. The
separation coefficients kx and ky are limited to some upper and lower boundaries
accordingly the recommendations [5, 15, 24, 37] or can be calculated for particular
wind farm site with given wind conditions.
It should be noted here that application of the proposed two-stage algorithm is
not restricted to these two objectives and other formulations for the objectives can be
used.
Step 1.3. On this step the determined alternative is stored for later comparison
and assessment by group of DMs. The SDM contributes in simulation of other design
scenarios by repeating the Step 1.1 with changed preferences about the objectives
importance or by requesting for another multi-objective solution method. When
different design scenarios are simulated, the determined alternatives are filtered, i.e.,
identical alternatives are considered as one unique alternative and all unique
alternatives are stored for later evaluation on Stage 2.
Step 2.1. On this step the SDM determines the wind farm parameters as
evaluation criteria and selects a group of experts that represent different design
aspects of wind farm project – engineers, economists, ecologists, etc. SDM defines
also DMs’ weighted coefficients according their expertise toward the project goals.
Step 2.2. This step is based on MADM described by a Weighted Decision
Matrix (WDM). This WDM indicates both the set of alternatives and the set of criteria
being considered in a problem. The structure of WDM proposed in the article
includes: DMs’ weights for the importance of the design parameters (criteria); DMs’
scores towards design alternatives’ parameters; and weighted coefficients for DMs.
Each of the DMs estimates alternatives independently, without negotiations with
other DMs and accordingly to his specific point of view. For example, a financial
expert probably will put more weight on financial parameters of the design while
engineers would be more interested in wind farm energy production. The SDM
assigns weighting coefficients for each DM according to their hierarchy positions in
the management of WFLOP.

91
 In general case, the different DMs’ points of view about alternatives can be
conflicting because they reflect different preferences toward importance of design
parameters. In WDM the higher values of evaluations mean a better performance and
the final goal is to maximize the outcome of decision about the best alternative.
Step 2.3. On this step, single criterion optimization modelling is used to
formulate an optimization task as:
(12) max ∑𝑀 𝐾 𝑘 𝑘 𝑘
𝑖=1 ∑𝑘=1 𝛼 𝑤𝑖 𝐴𝑖 ,
subject to
(13) ∀𝑖 = 1, 2, … , 𝑀: (∀𝑘 = 1,2, … , 𝐾: 𝐴𝑘𝑖 = ∑𝐽𝑗=1 𝑎𝑖,𝑗
𝑘
𝑥𝑗 ),
(14) ∑𝐽𝑗=1 𝑥𝑗 = 1, 𝑥𝑗 ∈ {0, 1},
(15) 𝛼 𝑘 ∈ (0, 1),
where i = 1, 2, …, M are indexes of design parameters against which the design
alternatives are to be evaluated; j = 1, 2, ..., J are indexes of design parameters to be
evaluated; k = 1, 2, …, K are indexes of group of DMs involved into decision making
process; 𝑤𝑖𝑘 are weighting coefficients representing relative importance of design
parameters as evaluated by different DMs; 𝑎𝑖𝑗𝑘 is evaluation score of k-th DM for
performance of alternative j against parameter pi; coefficients k represent the weight
of opinion of the k-th DM and xj are binary integer decision variables used to perform
choice of a single alternative.

4. Numerical testing
In order to show the applicability of the proposed two-stage algorithm for WFLOP a
number of numerical examples are tested and analysed.
4.1. Input data for WFLOP
This section provides the input data for numerical testing of the proposed algorithm
in Section 2. A set of 30 different wind turbines are used to select the most appropriate
type of turbine (Table 1).
Table 1. Wind turbines parameters
Rated Rotor Rated Rotor
# Wind turbine type # Wind turbine type
power, kW diameter, m power, kW diameter, m
1 Enercon E-33 330 33.4 16 SWT-2.3-113 2300 113.0
2 Enercon E-48 800 48.0 17 SWT-2.3-108 2300 108.0
3 Enercon E-53 800 52.9 18 SWT-2.3-93 2300 93.0
4 Vestas V52 850 52.0 19 Enercon E-92 2350 92.0
5 Vestas V60 850 60.0 20 C96 Clipper Windpower 2500 96.0
6 Enercon E-44 900 44.0 21 Vestas V100 2600 100.0
7 GAMESA AE-61-1320 1320 61.0 22 Enercon E-82 3000 82.0
8 Vestas V82 1650 82.0 24 Vestas V90 3000 90.0
9 M Torres 1650 77.0 25 Vestas V112 3000 112.0
10 Vestas V100 1800 100.0 23 Enercon -101 3050 101.0
11 Vestas V80 2000 80.0 26 SWT-3.6-120 3600 120.0
12 Enercon E-82 2000 82.0 27 SWT-3.6-107 3600 107.0
13 Enercon E-70 2300 71.0 28 SWT-6.0-154 6000 154.0
14 SWT-2.3-82 VS 2300 82.4 29 Vestas V164-7.0 7000 164.0
15 Enercon E2-82 2300 82.0 30 Enercon E-126 7580 127.0

92
 The solution of the multi-objective problem formulated on Step 1.2 determines:
Pareto-optimal wind turbines type and rotor diameter Dwt; values of separation
coefficients kx and ky; number of turbines consistent with separation distances;
placement of turbines within wind farm site in rows and columns taking into account
wind direction. This solution defines wind farm Pareto-optimal layout. Three wind
direction cases are tested: a) for uniform wind direction (Fig. 2a) with boundaries for
separation coefficients 𝑘𝑥min = 𝑘𝑦min = 4.5, 𝑘𝑥max = 𝑘𝑦max = 5.5; b) for prevailing
wind direction as in Fig. 2b with 𝑘𝑥min = 1.5, 𝑘𝑥max = 3, 𝑘𝑦min = 8, 𝑘𝑦max = 12;
for prevailing wind direction as in Fig. 2c with 𝑘𝑥min = 8, 𝑘𝑥max = 12, 𝑘𝑦min = 1.5,
𝑘𝑦max = 3. The wind farm annual energy production can be represented as
AEP = 2628NPwt using hy = 8760 hours over year and wind power utilization
coefficient η = 0.3. A rectangular wind farm with area of 4 km2 (Lx = 4 km and
Ly = 1 km) is considered with three different cases for wind directions as shown in
Fig. 2.

Fig. 2. Orientations of wind toward the shape of wind farm

4.2. Generation of alternatives

The formulated multi-objective optimization problem on Step 1.2 is solved by
Weighted Sum (WS) and Lexicographic Method (LM). They are based on priori
articulation of the DM preferences but the solution information by both methods can
be used also with posterior preferences handling to generate different alternatives.
These two methods are chosen due to their easy establishment of DM preferences
toward the formulated objectives.
The weighted sum method is one of the most common multi-objective
optimization methods based on scalarization techniques [9, 14]. It aggregates the
different objectives as a weighted linear sum of their normalization. The DM
preferences are expressed by imposing weight coefficients wj about the importance
of different objectives. These weight coefficients are assigned to normalized form of
objectives. The normalization scheme used here yields to non-dimensional objective
function with values between 0 and 1 regardless of their original range [22]. The
maximized objective for AEP and the minimized objective for costs are normalized
by expressions:
min max min −1
(16) 𝑓1∗ = (𝑓AEP − 𝑓AEP )(𝑓AEP − 𝑓AEP ) ,
∗ max max min −1
(17) 𝑓2 = (𝑓costs − 𝑓costs )(𝑓costs − 𝑓costs ) ,
min max min max
where 𝑓AEP , 𝑓AEP , 𝑓costs , 𝑓costs are the lowest and the greatest values of maximized
93
objective for AEP and minimized objective for costs defined via solutions of the four
single objective tasks:
 min AEP = hyNPwt s.t. (2)-(11),
 max AEP = hyNPwt s.t. (2)-(11),
2 1 2
 min Costs = 𝑁 ( + 𝑒 −0.00174𝑁 ) , s.t., (2)-(11),
3 3
2 1 2
 max Costs = 𝑁(3 + 3 𝑒 −0.00174𝑁 ), s.t., (1)-(11),
The implementation of the WS method transforms the initial multi-objective
problem (1)-(11) into single objective problem as
(18) max ∑2𝑗=1 𝑤𝑗 𝑓𝑗∗ ,
subject to constraints (2)-(11) plus additional constraints for the weighting
coefficients:
(19) ∑2𝑗=1 𝑤𝑗 = 1,
(20) 𝑤𝑗 > 0, 𝑗 = 1, 2.
The different wind farm design scenarios are simulated by using of 2 different
sets of weighting coefficients and three different wind direction cases. The solution
results are shown in Table 2.
Table 2. Design alternatives by weighted sum method solutions
Input data Solution results
Task Wind Weight Turbine Turbines Separation AEP,
Costs Alternative
direction coefficients type number coefficients MW.h per 1 year
wp = 0.9
WS-1 #13 52 kx = 4.69; ky = 4.69 314,309 34.824 A-1 (WS)
wc = 0.1
Uniform
wp = 0.1
WS-2 #27 27 kx = 4.67; ky = 4.67 255,442 20.531 A-2 (WS)
wc = 0.9
wp = 0.9
WS-3 Predominant #27 50 kx = 1.56; ky = 9.35 473,040 33.548 A-3 (WS)
wc = 0.1
on Y axis
(Pre-1) w p = 0.1
WS-4 #26 46 kx = 1.52; ky = 8.33 435,197 31.053 A-4 (WS)
wc = 0.9
wp = 0.9
WS-5 Predominant #22 63 kx = 1.52; ky = 8.13 496,692 42.021 A-5 (WS)
wc = 0.1
on X axis
wp = 0.1
WS-6 (Pre-2) #30 24 kx = 1.57; ky = 10.50 478,086 18.936 A-6 (WS)
wc = 0.9

The WS method allows easy and intuitive definition of different preferences for
the objectives. Simulation of different design scenarios can be done by posterior
articulation of preference information, i.e., by altering the objectives’ weights to yield
different Pareto-optimal points.
Another way to handle multi-objective optimization problems by apriori
articulation of the DM preferences is the lexicographic method. Using of this method
requires ranking of objectives by the DM and optimization them in order one at a
time. The general description of this method can be summarized as solution of
sequence of single-objective optimization problems:
(21) min Fj(x), x  X, j = 1, 2, ..., k,
subject to
(22) Fj(xj) ≤ j Fj(xj*), j = 1, 2, ..., i–1, i >1,
where j represents the objective function position in the sequence, Fj(xj*) is the
optimum of the j-th objective function found on the j-th iteration, and j is tolerance

94
determined by the DM to expand the feasible region [23]. The solution of each single
objective problem gives a limiting measure for that objective which is used to define
a proper restriction on the next step when the next objective is optimized and so on.
The Pareto-optimal solution is defined on the last step of the described optimization
procedure. Two cases of lexicographic ordering are numerically tested.
C a s e 1. Wind farm AEP as the foremost objective and the optimization
procedure for that case is as follow:
Step 1.1. Solving of the optimization task
(23) max AEP = hyNPwt,
subject to (2)-(11).
Step 1.2. Solving of the optimization task
2 1 2
(24) min Costs = 𝑁 (3 + 3 𝑒 −0.00174𝑁 ),
subject to (2)-(11) and additional restriction regarding the energy output
(25) AEP = hyNPwt ≥ AEPmax.
The coefficient  represents the degree of proximity to the optimal value of
AEP = (hyNPwt)max calculated on the first step. Two values  = 0.7 and  = 0.9 are
used to define two different design alternatives.
C a s e 2. Costs are considered as more important than AEP. The corresponding
optimization procedure is as follow:
Step 2.1. Solving of the optimization task
2 1 2
(26) min Costs = 𝑁 ( + 𝑒 −0.00174𝑁 ),
3 3
subject to (2)-(11).
Step 2.2. Solving of the optimization task
(27) max AEP = hyNPwt,
subject to (2)-(11) and additional restriction regarding the Costs
2 1 2
(28) Costs = 𝑁 (3 + 3 𝑒 −0.00174𝑁 ) ≤ 𝜀Costsmin ,
where the coefficient  represents the degree of proximity of the costs to the optimal
2 1 2 min
value 𝑁 (3 + 3 𝑒 −0.00174𝑁 ) calculated on the first step. The values of  used for
simulation of different design scenarios are  = 1.3 and  = 1.1.
The results of the lexicographic method implementation are shown in Table 3.
Table 3. Design alternatives by lexicographic method solutions
Input data Solution results
Task Turbine Turbines Separation
Wind Direction  type number coefficients AEP, MW.h per 1 year Costs Alternative
a) lexicographic ordering → AEP, Costs
Lp-1 0.7 # 27 27 kx = 4.67; ky = 4.67 255,442 20.531 A-1
Uniform
Lp-2 0.9 # 13 52 kx = 4.69; ky = 4.69 314,309 34.823 A-2
Lp-3 Predominant 0.7 # 26 36 kx = 1.96; ky = 8.33 340,589 25.253 A-3
Lp-4 on Y axis (Pre-1) 0.9 # 26 46 kx = 1.52; ky = 8.33 435,197 31.053 A-4
Lp-5 Predominant 0.7 # 30 20 kx = 1.97; ky =10.50 398,405 16.657 A-5
Lp-6 on X axis (Pre-2) 0.9 # 30 24 kx = 1.58; ky =10.45 478,086 18.936 A-6
b) lexicographic ordering → Costs, AEP
Lc-1 1.3 # 27 27 kx = 4.67; ky = 4.67 255,442 20.531 A-1
Uniform
Lc-2 1.1 # 27 27 kx = 4.67; ky = 4.67 255,442 20.531 A-2
Lc-3 Predominant 1.3 # 27 36 kx = 2.20; ky = 9.35 340,589 25.258 A-3
Lc-4 on Y axis (Pre-1) 1.1 # 27 28 kx = 2.87; ky = 9.35 264,902 21.052 A-4
Lc-5 Predominant 1.3 # 30 24 kx =10.50; ky = 1.57 478,086 18.936 A-5
Lc-6 on X axis (Pre-2) 1.1 # 30 16 kx =10.50; ky = 2.62 318,724 14.083 A-6

95
4.3. Filtering of identical alternatives
Comparison of design alternatives from Table 2 and Table 3 shows that there are
identical alternatives. For example, in case of uniform wind the identical alternatives
are A-1(WS) and A-2(Lp), A-2(WS), A-1(Lp), A-1(Lc) and A-2(Lc). For
predominant wind (case Pre-1) the identical alternatives are A-3(Lp) and A-3(Lc),
A-4(WS) and A-4(Lp) and for opposite direction (case Pre-2) the identical
alternatives are A-6(WS), A-6(Lp) and A-5(Lc). The identical design alternatives are
considered as one unique alternative as shown in Table 4.
Table 4. Unique design alternatives as result of execution of Stage 1
AEP,
Unique alternatives Turbine type (from Table 1) Turbines number SDx, m SDy, m Costs
MW.h per 1 year
Uniform wind direction
A-1 (uni) # 13 52 333.4 333.4 314,309 34.824
A-2 (uni) # 27 27 500.0 500.0 255442 20.531
Predominant wind direction (Pre-1)
A-1 (Pre-1) # 27 50 166.7 1000.0 473,040 33.55
A-2 (Pre-1) # 26 46 181.8 1000.0 435,197 31.05
A-3 (Pre-1) # 26 36 235.3 1000.0 340,589 25.25
A-4 (Pre-1) # 27 28 307.7 1000.0 264,902 21.05
Predominant wind direction (Pre-2)
A-1 (Pre-2) # 22 63 666.7 125.0 496,692 42.02
A-2 (Pre-2) # 30 24 1333.4 200.0 478,086 18.94
A-3 (Pre-2) # 30 20 1333.4 250.0 398,405 16.66
A-4 (Pre-2) # 30 16 1333.4 333.4 318,724 14.08

The identical alternatives for different wind directions are visualized as

overlapping circles (Fig. 3).

Fig. 3. AEP-Costs diagram for different design alternatives: for uniform wind direction (a); for two
cases (Pre-1 and Pre-2) of predominant wind direction (b) and (c)

After filtering not identical alternatives of Table 4 are stored for use on the
Step 2.3 of the algorithm.

4.4. Determination of optimal alternative by group decision making

Three wind farm design parameters shown in Table 5 are used and evaluated by group
of three DMs. The WDM representing WFLOP via GDM considers the case Pre-2
data which is characterised by four unique design alternatives (Table 5). Three
different experts – engineer (DM-1), a financial expert (DM-2) and ecologist (DM-
3) are involved in the process of alternatives evaluation toward three wind farm

96
design parameters (energy output, costs and number of turbines). For each of these
experts a corresponding weighting coefficient is assigned to represent its opinion
importance.
Table 5. Input WDM data for testing of GDM
Parameters’ Alternative A-1 Alternative A-2 Alternative A-3 Alternative A-4
Design
weights evaluations evaluations evaluations evaluations
parameters
DM-1 DM-2 DM-3 DM-1 DM-2 DM-3 DM-1 DM-2 DM-3 DM-1 DM-2 DM-3 DM-1 DM-2 DM-3
AEP MW.h per 1 year 10 4 3 10 8 4 8 6 7 6 5 6 4 3 5
Costs 4 10 3 5 2 4 6 7 5 7 6 6 8 5 7
Turbines number 5 8 10 3 2 2 4 6 5 5 7 7 6 6 8

The solutions results based on solution of single-objective integer optimization

task (12)-(15) for 4 sets of values of DMs’ weights α are shown in Table 6.
Table 6. GDM results for different weights of DMs
Weighting coefficients of DMs
Tasks Best alternative
α1 α2 α3
GDM-1 0.5 0.5 0.5 A-3
GDM-2 0.9 0.2 0.2 A-2
GDM-3 0.2 0.9 0.2 A-2
GDM-4 0.2 0.2 0.9 A-4

The graphical illustration of the optimal layout design alternatives obtained as

result of execution of Stage 2 are shown in Fig.4.

Fig. 4. Graphical representation of: Alternative-3 (a); Alternative-2 (b); Alternative-4 (c)

5. Results analysis and discussions

The formulated optimizations tasks are solved by Lingo V. 12 system
(http://www.lindo.com). The solution algorithm is branch-and-bound and solution
times for the described numerical examples are about 1 s. Lingo’ status windows for
some examples of the formulated discrete combinatorial optimization tasks are shown
in Fig. 5.
The nonlinearity of the defined optimization tasks leads to local optimums. This
nonlinearity is caused from using of widely accepted by many researches nonlinear
formulation of costs [15, 25, 36]. If linear discrete formulation of costs is used the
solutions of the formulated in the article tasks will define global optimums.
The used two solution methods – weighted sum and lexicographic method show the
benefits of using multi-objective combinatorial optimization of as an analytical
simulation tool for determination of different design alternatives.

97
 (a) (b) (c)
Fig. 5. Lingo status window for: task WS-5 (a); two stages of task Lp-6 (b) and (c)

The results of solution by weighted sum method show that stronger preference
on AEP leads to the choice of turbines with a smaller rotor diameter. As the rotor
diameter is decisive for considering of WE by proper separation distances, the smaller
rotor diameter leads to increasing of turbines number and as a consequence – to
increasing of the energy yield. The opposite preference (Costs over AEP) results in
choice of more powerful turbines with a bigger rotor diameter. This leads to reducing
of the turbines number that in turn decreases the costs compared to the first case. The
site orientation towards the prevailing wind direction also affects the Pareto-optimal
design alternatives.
The lexicographic ordering of the objectives is another easy implementation
method to simulate different scenarios for WFLOP. Variety of Pareto-optimal
solutions (design alternatives) can be determined by different lexicographic ordering
of the objectives and by different values of coefficient . For lexicographic ordering
AEP, Costs, when  changes from 0.7 up to 0.9, the turbines’ number increases and
the AEP is approaching closer to its maximum. In lexicographic ordering Costs
before AEP and value of  decreases from 1.3 to 1.1, the result leads to decreasing of
the turbines’ number and Costs are getting closer to their minimum. Changing the
value of  coefficient allows flexible adjustment of DM preferences in terms of trade-
offs between objectives, i.e., provides various design alternatives.
Numerical testing demonstrates that both methods are adequate for the
generating of wind farm layout design alternatives. The weighted sum method has
limitations in regard to representation of compromises that DM is willing to make
among objectives and also requires normalization of the objectives. The
lexicographic method allows a more precise tuning of these trade-offs by considering
the desirable degree of proximity to the “best” values of objectives. The discrete
nature of the turbines’ type choice defines in some cases overlapping Pareto-optimal
alternatives. This overlapping reduces the number of design alternatives to be
evaluated that helps for facilitating the selection of the final alternative.
In contrast to other methods for GDM, the proposed GDM approach considers
simultaneously not only different DMs weights about design parameters (criteria) and
alternatives evaluations but also group members’ opinions weights. Apparently, the
assignment of weights to the group members’ opinions influences on decision process

98
and selection of collective alternative. That is why Alternative-2 (A-2) (Table 4, case
Pre-2) appears to be the best for engineers and economists while Alternative-3
(A-3) is accepted as best alternative if the weights of experts are equal. Unlike other
GDM approaches the described approach provides globally optimal alternative as a
solution of the formulated single-objective integer linear optimization task.
From engineering point of view, it is important to get to the problems solution
by the most direct and effective way. This requires interdisciplinary approach to
modelling of the decision-making processes and using of quantitative optimization
methods. The proposed in the article algorithm combines two approaches –
generating of multiple alternative design solutions for WFLOP and choice of the best
alternative by group decision making. This idea expresses the complexity and
multidimensionality of the WFLOP. In contrast to mass used genetic and other
heuristic algorithms, the described algorithm is based on combinatorial optimization
modelling. This approach allows using of the proven over the years computational
efficiency of the algorithms for linear, nonlinear and integer optimization. The used
modelling approach on Stage 1 of the algorithm defines wind turbines of the same
type. This is preferred case for commercial wind farms because it reduces installation
costs, maintenance cost, etc., [4]. Like many other published results for WFLOP
[10, 19, 26, 38], the testing of the proposed algorithm is done for a rectangular wind
farm site. The rectangular shape could be quite practical for offshore wind farms and
for onshore wind farms located on flat terrain. The placement of turbines in rows and
columns has the advantage that small deviations of prevailing wind direction lead to
increasing of the separation distances between turbines, i.e., to decreasing the
influence of WE (Fig. 6).

Fig. 6. Separation distances between turbines: predominant wind direction (SDx) (a);
small deviations of predominant wind direction (SD∗𝑥 ) (b)

To justify the effectiveness and applicability of the proposed algorithm it is

compared with other WFLOP algorithms. The main idea of the proposed algorithm
– combining the determination of multiple design alternatives for WFLOP and using
of GDM for choice of the best alternative, cannot be compared because it has no
analogue in other publications on WFLOP. The other feature of the algorithm, namely
using of discrete combinatorial optimization modelling, can be compared with other
modelling approaches. Widely used for WFLOP are GA and other metaheuristic
algorithms. Usually that kind of algorithms are used for problems that cannot be
easily solved and they give nearly the right answer or provide a solution not for all
instances of the problem [18]. The widely used GA have some known drawbacks: A
solution is “better” only in comparison to other, presently known solutions; in reality
99
they have no concept of “optimal solution” or way to test whether a solution is
optimal; stop criterion is not always clear in every problem, aside from the runtime
length, or the number of iterations or candidate solutions to explore
(http://www.solver.com/genetic-evolutionary-introduction). Sometimes it is
possible GA to converge not only to local optima but even to arbitrary points. The
GA repeated fitness function evaluation requires essential computation power and is
time consuming.

5.1. Comparison of computing performance of proposed algorithm and other

algorithms for WFLOP
To compare the computing performance of the proposed two-stage algorithm with
other algorithms for WFLOP the task WS-6 is modified and solved for different wind
farm areas as shown in Table 7.
Table 7. Comparison of the proposed algorithm with other algorithms
Algorithm Turbines number Area Runtime Authors
3170.21-3498.25 s
Random search 80 (fixed) 4×5 km [11]
(per run)
16 000 s
Genetic 15 2×2 km [26]
(on CRAY XMP14)
MIP model limited to < 10 <1h [5]
Evolutionary NSGA-II,
30, 50, 70 (fixed) 3×3 km < 15 min [39]
SPEA2, IBEA
Stochastic evolutionary 19.75 h; 70.87 h;
5, 10, 15 (fixed) 3×3 km [38]
algorithm (NSGA-II) 149.75 h
Evolutionary algorithm 4, 5, 20 (fixed) limited to < 50 20×20 grid 90 min [25]
MIP 10 – 100 (fixed) 1.7×2 km limited to 1800 s [6]
within wall-clock
MIP 40, 50, 60, 70 (fixed) 4×4 km [19]
time of 200 s
MINLP (Task WS-6) defined as 63 of Enercon E-82 1×4 km about 1 s present work
MINLP (Task WS-6) defined as 18 of SWT-3.6-120 1.7×2 km about 1 s present work
MINLP (Task WS-6) defined as 21 of SWT-3.6-120 2×2 km about 1 s present work
MINLP (Task WS-6) defined as 48 of Enercon E-126 3×3 km about 1 s present work
MINLP (Task WS-6) defined as 84 of Enercon E-126 4×4 km about 1 s present work
MINLP (Task WS-6) defined as 108 of Enercon E-126 4×5 km about 1 s present work
MINLP (Task WS-6) defined as 540 of Enercon E-126 20×5 km about 1 s present work
about 2 s
MINLP (Task LP-6) defined as 24 Enercon E-126 1×4 km present work
(2 runs of 1 s)
A testing experiment is done with increased wind farm area to 100 km2 to show
the computational time of the task and as it is seen from Fig. 7, the time remain in the
range of one second.
Some remarks can be stated in regard to the application of two-stage algorithm
for WFLOP and usage of mathematical model (1)-(11):
 Most of the WFLOP results are related with preliminary fixed number of
turbines [5, 10, 11, 19, 28, 36, 39]. It is not clear how this number is determined and
if it is optimal for the considered wind site area. In contrast, the proposed algorithm
allows determination of Pareto-optimal number and type of turbines for the given
wind farm area while considering the wind direction and recommendations for
separations distances to avoid the negative influence of the WE.
100
  An advantage of the described combinatorial modelling approach is that the
number of variables of the defined optimization problems depends mainly on the
number of turbines, from which selection is made. The area of the wind farm does
not affect the dimension of the tasks (Fig. 5 and Fig. 7).
 Both stages of the proposed algorithm can be used separately. The generated
on Stage 1 set of design alternatives could be evaluated by other GDM approaches.
On the other hand, if variety alternatives are known or given in advance the described
in the Stage 2 approach for GDM would be quite effective for choice of best
alternative.

Fig. 7. Lingo status windows of task WS-6 for site dimensions 20×5 km

6. Conclusion
In this article a two-stage placement algorithm with multi-objective optimization and
group decision making is proposed. The main idea is to generate multiple Pareto-
optimal placement alternatives and to select one of them via group of experts to get
the most suitable business decision. Using of multi-objective discrete combinatorial
optimization, on the first stage of the algorithm provides different Pareto-optimal
alternatives and can be consider as a flexible analytical tool to simulate different
design scenarios. The different DMs preferences toward importance of used criteria
simulate different design alternatives. All of these alternatives are used on the second
stage of the proposed algorithm to select the most suitable alternative by group
decision making.

101
 The described two-stage algorithm is numerical tested for design of wind farm
layout to show its practical applicability. The runtimes for tested examples are in the
range of a few seconds versus hundreds or thousands of seconds needed for the
execution of other published algorithms. This demonstrates the effectiveness of the
use of combinatorial optimization modelling for wind farm layout optimization
problems.
The proposed two-stage algorithm can be applied for design of other types of
engineering systems provided that their specificity is taken into account in modelling
and formulation of corresponding optimization tasks.

References
1 B o r i s s o v a, D., I. M u s t a k e r o v. Wind Power Plant Layout Design and Assessment
Considering Forbidden Zones for Location of Turbines. – Adv. Modeling and Optimization,
Vol. 19, 2017, pp. 29-38.
2 B o r i s s o v a, D., I. M u s t a k e r o v, D. K o r s e m o v. Business Intelligence System via Group
Decision Making. – Cybernetics and Information Technologies, Vol. 16, 2016, No 3,
pp. 219-229.
3 B o r i s s o v a, D., I. M u s t a k e r o v. A Generalized Combinatorial Optimization Approach to
Wind Power Plant Design. – Cybernetics and Information Technologies, Vol. 10, 2010, No 4,
pp. 62-74.
4 C h e n, Y., H. L i, B. H e, P. W a n g, K. J i n. Multi-Objective Genetic Algorithm Based Innovative
Wind Farm Layout Optimization Method. – Energ. Convers. Manage., Vol. 105, 2015,
pp. 1318-1327.
5 D o n o v a n, S. Wind Farm Optimization. – In Proc. of Annual Conference of the Operations
Research Society, Wellington, New Zealand, 2005.
6 D o n o v a n, S., G. N a t e s, H. W a t e r e r, R. A r c h e r. Mixed Integer Programming Models for
Wind Farm Design. – In: Workshop on Mixed Integer Programming, Columbia University,
New York City, 2008.
7 D u P o n t, B., J. C a g a n. An Extended Pattern Search Approach to Wind Farm Layout
Optimization. – J. Mech. Design., Vol. 134, 2012.
8 E h r g o t t, M. A Discussion of Scalarization Techniques for Multiple Objective Integer
Programming. – Ann. Oper. Res., Vol. 147, 2006, pp. 343-360.
9 E i c h f e l d e r, G. Adaptive Scalarization Methods in Multiobjective Optimization. – Springer,
Berlin Haidelberg, 2008.
10 F e n g, J., W. Z. S h e n. Modelling Wind for Wind Farm Layout Optimization Using Joint
Distribution of Wind Speed and Wind Direction. – Energies, Vol. 8, 2015, pp. 3075-3092.
11 F e n g, J., W. Z. S h e n. Solving the Wind Farm Layout Optimization Problem Using Random
Search Algorithm. – Renew. Energ., Vol. 78, 2015, pp. 182-192.
12 F u l o p, J. Introduction to Decision Making Methods. Working Paper 05-6, 2005.
13 Gartner Research Methodologies.
http://www.gartner.com/technology/research/methodologies
14 G e n o v a, K., L. K i r i l o v, V. G u l j a s h k i. New Reference-Neighbourhood Scalarization
Problem for Multiobjective Integer Programming. – Cybernetics and Information
Technologies, Vol. 13, 2013, No 1, pp. 104-114.
15 G r a d y, S. A., M. Y. H u s s a i n i, M. M. A b d u l l a h. Placement of Wind Turbines Using
Genetic Algorithms. – Renew. Energ., Vol. 30, 2005, pp. 259-270.
16 S. Greco, M. Ehrgott, J. R. Figueira, Eds. Multiple Criteria Decision Snalysis: State of the Art
Surveys. New York, Springer Verlag, 2016.
17 K a t i c, I., J. H o j s t r u p, N. O. J e n s e n. A Simple Model for Cluster Efficiency. – In: Proc. of
European Wind Energy Association Conference and Exhibition, 1986, pp. 407-410.
18 K o k a s h, N. An Introduction to Heuristic Algorithms. 2005.

102
19 K u o, J. Y. J., D. A. R o m e r o, C. H. A m o n. A Mechanistic Semi-Empirical Wake Interaction
Model for Wind Farm Layout Optimization. – Enegry, Vol. 93, 2015, pp. 2157-2165.
20 K w o n g, W.Y., P.Y. Z h a n g, D. R o m e r o, J. M o r a n, M. M o r g e n r o t h, C. A m o n. Multi-
Objective Wind Farm Layout Optimization Considering Energy Generation and Noise
Propagation with NSGA-II. – J. Mech. Design., Vol. 136, 2014.
21 L a r s o n, D.,V. C h a n g. A Review and Future Direction of Agile, Business Intelligence, Analytics
and Data Science. – Int. J. Information Management Vol. 36, 2016, pp. 700-710.
22 M a r l e r, R.T., J.S. A r o r a. Function-Transformation Methods for Multi-Objective Optimization.
– Eng. Optimiz., Vol. 37, 2005, pp. 551-570.
23 M a r l e r, R.T., J.S. A r o r a. Survey of Multi-Objective Optimization Methods for Engineering. –
Struct. Multidisc. Optim., Vol. 26, 2004, pp. 369-395.
24 M a r m i d i s, G., S. L a z a r o u, E. P y r g i o t i. Optimal Placement of Wind Turbines in a Wind
Park Using Monte Carlo Simulation. – Renew. Energ., Vol. 7, 2008, pp. 1455-1460.
25 M o r a, J.C., J.M.C. B a r o n, J.M.R. S a n t o s, M.B. P a y a n. An Evolutive Algorithm for Wind
Farm Optimal Design. – Neurocomputing, Vol. 70, 2007, pp. 2651-2658.
26 M o s e t t i, G., C. P o l o n i, B. D i v i a c c o. Optimization of Wind Turbine Positioning in Large
Windfarms by Means of a Genetic Algorithm. – J. Wind. Engand. Ind. Aerod., Vol. 51, 1994,
pp. 105-116.
27 M u s t a k e r o v, I, D. B o r i s s o v a. Wind Turbines Type and Number Choice Using
Combinatorial Optimization. – Renew. Energ., Vol. 35, 2010, pp. 1887-1894.
28 O s t e r g a a r d, P. A. Reviewing Optimisation Criteria for Energy Systems Analyses of Renewable
Energy Integration. – Energy, Vol. 4, 2009, pp. 1236-1245.
29 P e n e v a, V., I. P o p c h e v, Fuzzy Multi-Criteria Decision Making Algorithms. – Compt. Rend.
Acad. bulg. Sci., Vol. 63, 2010, pp. 979-992.
30 P e n e v a, V., I. P o p c h e v. Models for Decision Making by Fuzzy Relations and Fuzzy Numbers
for Criteria Evaluations. – Compt. Rend. Acad. bulg. Sci., Vol. 62, 2009, pp. 1217-1222.
31 P e n e v a, V., I. P o p c h e v. Models for Fuzzy Multicriteria Decision Making Based on Fuzzy
Relations. – Compt. Rend. Acad. bulg. Sci., Vol. 62, 2009, pp. 551-558.
32 P e n e v a, V., I. P o p c h e v. Multicriteria Decision Making Based on Fuzzy Relations. –
Cybernetics and Information Technologies, Vol. 8, 2008, pp. 3-12.
33 P e n e v a, V., I. P o p c h e v. Multicriteria Decision Making by Fuzzy Relations and Weighting
Functions for the Criteria. – Cybernetics and Information Technologies, Vol. 9, 2009,
pp. 58-71.
34 P e r e z, B., R. M i n g u e z, R. G u a n c h e. Offshore Wind Farm Layout Optimization Using
Mathematical Programming Techniques. – Renew. Energ., Vol. 53, 2013, pp. 389-399.
35 S a a v e d r a-M o r e n o, B., S. S a l c e d o-S a n z, A. P a n i a g u a-T i n e o, L. P r i e t o,
A. P o r t i l l a-F i g u e r a s. Seeding Evolutionary Algorithms with Heuristics for Optimal
Wind Turbines Positioning in Wind Farms. – Renew. Energ., Vol. 36, 2011, pp. 2838-2844.
36 S h a k o o r, R., M. Y. H a s s a n, A. R a h e e m, Y. K. W u. Wake Effect Modeling: A Review of
Wind Farm Layout Optimization Using Jensen’s Model. – Renew. Sust. Energ. Rev., Vol. 58,
2016, pp. 1048-1059.
37 S m i t h, G., W. S c h l e z, A. L i d d e l l, A. N e u b e r t, A. P e n a. Advanced Wake Model for
Very Closely Spaced Turbines. – In: Proc. EWEC 2006, Athens. 2006.
38 S o r k h a b i, S. Y. D., D. A. R o m e r o, G. K. Y a n, M. D. G u, J. M o r a n, M. M o r g e n r o t h,
C. H. A m o n. The Impact of Land Use Constraints in Multi-Objective Energy-Noise Wind
Farm Layout Optimization. – Renew. Energ., Vol. 85, 2016, pp. 359-370.
39 T r a n, R., J. W u, C. D e n i s o n, T. A c k l i n g, M. W a g n e r, F. N e u m a n n. Fast and Effective
Multi-Objective Optimisation of Wind Turbine Placement. – In: Proc. Genetic and
Evolutionary Computation, 2013, pp. 1381-1388.
40 W a n, C., J. W a n g, G. Y a n g, X. L i, X. Z h a n g. Optimal Micro-Siting of Wind Turbines by
Genetic Algorithms Based on Improved Wind and Turbine Models. – In: Proc. of 48th IEEE
Conf. on Decision & Control and the 28th Chinese Control Conference, 2009.

103