Energies 2014, 7, 6930-7016; doi:10.

3390/en7116930
OPEN ACCESS

energies
ISSN 1996-1073
www.mdpi.com/journal/energies
Review

A Review of Methodological Approaches for the Design and

Optimization of Wind Farms
JoséF. Herbert-Acero 1,*, Oliver Probst 1, Pierre-Elouan Réthoré2, Gunner Chr. Larsen 2
and Krystel K. Castillo-Villar 3

1
Chair for Wind Energy, Physics Department, Tecnológico de Monterrey, Eugenio Garza Sada 2501 Sur,
64849 Monterrey, N.L., Mexico; E-Mail: oprobst@itesm.mx
2
Department of Wind Energy, Technical University of Denmark, Frederiksborgvej 399,
4000 Roskilde, Denmark; E-Mails: pire@dtu.dk (P.-E.R.); gula@dtu.dk (G.C.L.)
3
Department of Mechanical Engineering, the University of Texas at San Antonio,
One UTSA Circle, San Antonio, TX 78249, USA; E-Mail: Krystel.Castillo@utsa.edu

* Author to whom correspondence should be addressed; E-Mail: jf.herbert.phd.mty@itesm.mx;

Tel.: +52-81-8358-2000 (ext. 4637); Fax: +52-81-8358-2000 (ext. 4632).

External Editor: Frede Blaabjerg

Received: 30 May 2014; in revised form: 13 September 2014 / Accepted: 8 October 2014 /
Published: 29 October 2014

Abstract: This article presents a review of the state of the art of the Wind Farm Design
and Optimization (WFDO) problem. The WFDO problem refers to a set of advanced
planning actions needed to extremize the performance of wind farms, which may be composed
of a few individual Wind Turbines (WTs) up to thousands of WTs. The WFDO problem has
been investigated in different scenarios, with substantial differences in main objectives,
modelling assumptions, constraints, and numerical solution methods. The aim of this paper
is: (1) to present an exhaustive survey of the literature covering the full span of the subject,
an analysis of the state-of-the-art models describing the performance of wind farms as well
as its extensions, and the numerical approaches used to solve the problem; (2) to provide
an overview of the available knowledge and recent progress in the application of such
strategies to real onshore and offshore wind farms; and (3) to propose a comprehensive
agenda for future research.
Energies 2014, 7 6931

Keywords: wind farm design; micro-siting; wind farm layout; wind farm performance
modeling; nonlinear optimization; metaheuristic optimization

1. Introduction

The number, size and complexity of wind farms have been steadily increasing over the past few
years as consequence of the ambitious target of producing a larger share of energy from the wind [1–3].
This has resulted in a remarkable development in the design of large-scale wind energy conversion
systems, which in turn have improved the wind energy economies of scale. Proper wind farm planning
has major positive economic impacts for both wind farm investors and developers. However,
the problem is not only complex from the technological point of view, but is also subject to an
undesirable degree of uncertainty that affects the potential profitability of the project during its expected
operational lifetime. As a consequence of the inherent risk, both the required investments and the
payback period might be too large to the point of economic unviability. Therefore, detailed planning,
modeling and optimization are required to guarantee both technical feasibility and financial
competitiveness compared to other conventional forms of energy conversion.
In the early 1970s, important factors such as the worldwide social pressure for denuclearization
spurred governments, utilities and the scientific community to explore the potential of wind power for
large-scale electricity generation. The first scientific works related to the interactions of Wind Turbines
(WTs) in arrays appeared in 1974 [4–10], marking the beginning of successive paradigm shifts in WT
wake (see Section 2.2.1 for a detailed description) modeling and WT array layout optimization.
The interactions between WTs through wakes were initially modelled as an equivalent increase in
surface roughness which modified the Earth’s boundary layer due to the WT-induced drag. Based on
this coarse wake-interaction modeling, it was concluded that quite large separations between WTs were
required to avoid significant wake-based energy losses. In the subsequent years, staggered
equally-spaced (or grid-like) siting schemes were slowly adopted as a rule-of-thumb by wind farm
designers [11–17]. In the late 1970s and the early 1980s, just after the construction of the world’s first
electricity-generating wind farm in New Hampshire (USA) in 1980, the first engineering wake
models [18–24], mostly based on flow momentum conservation and linearized far wake expansion
assumptions (with the notable exception of the Larsen [25] and Ainslie [26] models), appeared, laying
the groundwork for the first study on Wind Farm Design and Optimization (WFDO) [27], which aimed
at optimizing the spacing between WTs in a linear array by using a generalized reduced gradient
method that maximized the total power conversion of the array. The maximum achieved increase in
power conversion for the given generic study case was 1.29% compared to an equally-spaced linear
array. At first glance, this might appear to be a poor improvement, however, a mere one percent
improvement in electricity output would result in large additional revenues over the typical 20-year
lifetime of wind farms. Seven years later, the classic work of Mosetti et al. [28] on WFDO appeared,
leading to significant improvements in both the description of the wind farm performance (the
minimized objective metric was the cost of power conversion—see Section 4.5 for misconceptions)
and the numerical approach to tackle the problem (a population-based metaheuristic; the Simple
Energies 2014, 7 6932

Genetic Algorithm or SGA). After that, many authors have performed a significant amount of
contributions of variable complexity along two main tracks: (1) a track related to improving the
performance modeling of wind farms (i.e., by considering the inclusion of complex sub-models for
financial prospection, aerodynamics and airflow physics, structural fatigue and degradation, electrical
system and grid interconnection, environmental impact and uncertainty among others); and (2) a track
related to the development and testing of different numerical solution procedures. However, individual
approaches in the literature do not completely respond to the needs of the wind farm designers, since
these tracks were treated and developed separately. In this regard, the interested reader might be
disoriented, since it is still unclear which models, objectives and numerical schemes are best suited for
application in the treatment of the WFDO problem. To overcome this issue, this work reviews and
examines the most promising, effective, recurring models and solution methods, highlighting their
strengths and weaknesses, and proposes a research agenda. The present work is structured as follows:
Section 2 summarizes the most important concepts involved in the wind farm design process. Section 3
analyses the most prominent literature related to the WFDO problem, including the state-of-the-art
modeling approaches and solutions methods. Section 4 outlines the research needs and trends in the
WFDO problem. Finally, Section 5 presents the overall conclusions.

2. Overview of the Basic Concepts of Wind Farm Design and Optimization

The design and optimization of wind farms is a complex multidisciplinary topic that can proceed
in different ways. The starting point are the existing guidelines and international standards, which give
a clear overview of the important factors that affect the decision making of the wind farm designer.
The present section describes these factors, the design variables and the formal definition of the
WFDO problem.

2.1. Wind Farm Planning—Guidelines and Standards

In general, it is recommended to specify and design wind farm projects in accordance with the
state-of-the-art standards of the International Electrotechnical Commission (IEC) [29] and the technical
guidelines proposed by international wind energy associations. Some examples include, but are not
limited to:
- The European Wind Energy Association (EWEA): European Best Practice Guidelines for
Wind Energy Development [30].
- The American Wind Energy Association (AWEA): Wind Energy Siting Handbook [31],
the AWEA fact sheets and the standards for wind energy industry [32].
- The British Wind Energy Association (BWEA or RenewableUK): Best Practice Guidelines for
Wind Energy Development [33].
- Global Wind Energy Council (GWEC): Global Wind Reports (2006–2012) [34].
- Sustainable Energy Authority of Ireland (SEAI): How to develop a Wind Farm [35].
- Wind Energy Handbook, Chapter 9: Wind-turbine Installations and Wind Farms [36].
Four general stages for wind farm planning and development are briefly summarized next:
Energies 2014, 7 6933

(1) Wind Farm Site Selection. The siting process [37–41] refers to the investigation of an
area/region/district with the purpose of selecting the best location (site) for a wind farm. Wind farms
are very versatile and can be integrated successfully into many types of landscapes, from agricultural
fields to mountain ridges, among others. The selection of the wind farm location is commonly based
on a hierarchy scheme including three general factors: (1) the available infrastructure (i.e., accounting
for the current capacity of the electrical, civil engineering and communication infrastructure and
its expansion plans); (2) the environmental suitability (e.g., accounting for restricted areas such as
natural protected areas or bird conservation areas); and (3) the wind resource. The wind resource
quantification [42–45] is based on long-term records (each record being a 10-min averaged measurement
obtained from sampling with a frequency of 1 Hz or higher) of atmospheric variables exhibiting temporal
and spatial variations: wind speeds [m/s] at different heights, wind directions [°] at different heights,
solar radiation [W/m2], temperatures [°C] at different heights, pressure [Pa], relative humidity [%], rain
fall [mm], and sometimes even air quality. Since atmospheric variables are considered to be random on
nature, further analyses are performed to study its statistical characteristics (e.g., determine its probability
density functions), which allow the wind farm designer to quantify expected values at the potential
wind farm site. Additional information that helps the designer to dimension the loads and stresses
the potential WT will undergo while operating at the site can be obtained from statistical-dynamical
analyses, which include the following measured variables: local ambient turbulence intensity [%],
power and turbulence spectra and coherences, mean and maximum gusts [m/s], rain/ice/snow/hail [%],
chemical effects (related to corrosion by salts and/or soiling), and atmospheric stability cycles [36].
In engineering practice, wind farm designers usually calculate the local 3D wind field of the site by
means of commercial packages [46–52] that implement and numerically solve the governing equations
of fluid mechanics. Different levels of sophistication exist for such calculations, with the approach that
considers the Reynolds Averaged Navier-Stokes (RANS) equations, complemented with turbulence
models, being widely used. It has been shown that the RANS-based models’ predictions provide
reasonable agreement with the experimental data measured at sites with complex topography [53].
The micro-scale flow modeling is usually based on high-resolution maps representing the topography,
obstacles and roughness of the region. The output wind field is convoluted numerically with the
resulting statistical distributions of the 10-min averaged wind speed measurements at different heights
and wind directions, resulting in averaged wind resource maps at different heights, which are the base
for the wind farm planning process. Other sources of information such as Mesoscale modeling and the
Measure–Correlate–Predict (MCP) techniques [36,42] could also provide valuable information of the
local atmospheric behavior and the wind resource potential.
(2) Wind Farm Design. The wind farm sizing refers to the process of designing the size of the wind
farm in terms of the constrained land area and its potential capacity. The wind farm size will be a
function of many factors including the expected power demand from the energy consumers. The wind
farm layout design or micro-siting process refers to the design of: (1) a configuration or arrangement
of the individual WTs (specified by locations on spatial coordinates (x, y, z) down to 1 m in resolution);
(2) their electrical interconnections; and (3) the auxiliary communication and civil engineering
infrastructure within the selected site. The optimal design thus results from an optimization given all
economic, environmental, technical and social conditions and constraints by using a mono- or
multi-objective criteria. During the wind farm micro-siting process, economic simulations, accounting
Energies 2014, 7 6934

for all available incentives, are performed with the aim of quantifying the expected economic profits of
the project while assuming all energy and environmental regulations to be satisfied. Once the economic
feasibility of a preliminary project is determined, land owners are contacted and asked for participation
in the project by hosting WTs on their land. Land owners receive incomes (royalty fees) from renting
their land and usually receive a percentage of the gross income generated by the project, and extra
compensations if roads or other infrastructure are built on their terrain. Depending on the landowner’s
participation [54], a wind farm redesign could be required. The legal, leasing and planning application
processes are not straightforward and may take a significant amount of time. In addition, wind energy
best practices for community engagement and public consultation encourage wind farm designers to
develop community outreach programs. In this sense, a significant amount of money is spent during
the wind farm design process even without any guarantee that the project can be built.
(3) Energy Regulatory Framework. The energy regulations depend on the country where the project
is to be developed. The wind farm designer must be aware of the legislation regarding the production,
transmission, transformation and marketing of the electricity, which in turn define the potential
profitability of the project. In addition, the wind farm designer must act in accordance with the local
and national development plans and policies and the electricity market.
(4) Financing and Funding. Once a complete design and organization is performed, the wind farm
developer must have access to funds through banking institutions or special financial organizations.
The acceptance/rejection of the funding will depend on several factors such as the results from the
wind resource assessment, the Power Purchase Agreements (PPAs) for the energy to be converted and
supplied, the full planning permissions, feasibility studies, electrical interconnection and construction
agreements, banking/economic climate and insurance against natural disasters among other factors [36].
In this sense, the uncertainty on the project revenue is one of the most important factors determining a
successful financing.
The present work focuses on describing in detail the approaches used to develop the second stage of
the wind farm planning process, where technical and optimization challenges are commonly encountered.
It should be noted that topics regarding the design and optimization of WTs and the optimal control of
WTs and wind farms were disregarded during the development of this work and should be considered
as important problems to be treated. A comprehensive description of the most important factors influencing
the wind farm design process is presented in the next section.

2.2. Adverse Factors and Conflicting Objectives

Determining the best wind farm design involves many design trade-offs [55,56] that must be
balanced. The set of factors that most affect the micro-siting process are briefly described next, while
the state-of-the-art modeling approaches to account for their side effects are presented in Section 3.

2.2.1. Wind Turbine Wakes

The wake of a WT [57–59] is a volume of affected flow that is characterized by a reduced

streamwise velocity, high vorticity and increased levels of turbulence compared to the free stream
airflow. The consequences of operating WTs in the wakes of obstacles (e.g., WTs, meteorological
masts or buildings) have been reported based on field experience: (1) Overall power losses up to 41% may
Energies 2014, 7 6935

occur, where the precise loss figures depend on the inter-WT distances, wind direction, site complexity
(e.g., onshore, offshore), turbulence intensity and WT type, among other relevant factors [60–68];
(2) Enhanced fatigue loads of up to 80% [58] that drastically reduce the lifespan of the WT and
significantly increase the Operation and Maintenance (O&M) costs. In particular, operating a WT
under partial wake conditions (i.e., only a fraction of the rotor of the downwind WT operates in the
wake of upstream WTs) results in higher fatigue loads in contrast with operating a WT under full wake
conditions as pointed out by Seifert et al. [69], where the blade root bending flatwise, the main shaft
torque and the tower head bending moment-based fatigue loads are the most affected. In addition,
the WT induced turbulence intensity may not only reduce the power output of the downstream WTs
but can also cause dynamic loading and increased levels of noise which, in the long run, lead to blade
damage and higher O&M costs. Moreover, wind farm wakes produce important disturbances in the
atmospheric boundary layer, thus they can potentially affect neighboring wind farms [70–73].
WT wakes have special relevance for the wind farm design process, as it is the single factor that
accounts for the largest energy losses [58,60,61,63,68]. Consequently, the adequate positioning of each
WT within the wind farm site is one of the factors that most influences the profitability of a project
throughout its operational lifetime.

2.2.2. Wind Farm Infrastructure

One of the most important prerequisites for a wind project is that there should be enough electrical,
civil engineering and telecommunication infrastructure, so the project will not incur in large additional
expenses [74–76] by building auxiliary infrastructure, which in turn might lead to economical unviability.
The wind farm electrical infrastructure is mainly composed of substations containing step-up
transformers, Medium Voltage Alternate Current (MVAC) collection lines and High Voltage Alternate
Current (HVAC) or High Voltage Direct Current (HVDC) transmission lines. Typically, each WT
contains a transformer that increases the generated voltage (e.g., 220 V~3 kV) to the MVAC collection
system voltage (e.g., 3 kV~60 kV). The MVAC collection system is in turn connected to a large
capacity transformer that raises the voltage to the HVAC or HVDC transmission system voltage
(e.g., 69 kV~161 kV) [77].
In practice, several WTs can be connected to the HVAC or HVDC transmission system by
sharing a common MVAC collection line circuit. This strategy has proven to reduce the cable length
requirements at the expense of increasing the cable size (cross sectional area) in order to meet capacity
constraints [77–81]. The MVAC collection system layout depend on the wind farm layout, where WTs
can usually be interconnected in a modular fashion, while accounting for the power lines capacity,
to produce a combined power output which can be either increased or decreased through the addition
or subtraction of WTs at the different stages of the electric network implementation. For security
reasons, all the electrical components are typically over dimensioned (e.g., by 30% above the expected
rated capacity) since they are very sensitive to variations in the local climate and to electrical
instabilities derived from the main electrical system, among other factors, which in turn can produce
significant impacts on their performance [82]. The cable size is typically determined based on an
economic analysis that in turn depends on the amount of current the cable can transport without
overheating, which is a consequence of the Joule-based power losses. Fixed costs are higher for
Energies 2014, 7 6936

large cross sectional cables, but Joule losses are lower since they have lower electric
resistance [55,77–81,83].
The size and location of the wind farm might lead to the problem of finding suitable electrical
interconnection points, which must have enough capacity to handle the energy delivered from the wind
farm. In cases where the distance between the electrical interconnection point and the wind farm is
large (e.g., around 100 km [55,84]) HVDC transmission lines may become more technologically and/or
economically convenient than the conventional HVAC transmission lines [85]. However,
the decision of using HVDC transmission lines is stressed by the fact that it is generally difficult to get
permission and funds to build new overhead lines and/or substations and, although losses are low for
HVDC transmission lines, there are added losses due to the Voltage Source Converter (VSC) components,
which are typically difficult to quantify. The decision of using one type of transmission line or another
(AC or DC) [85] is mostly based on the amount of reactive power generation and on the amount of
Joule-based power losses, which depends on the transmission line length, transmission voltage and the
rated power among other factors [78,80,82,85–91]. Therefore, interconnections between the wind farm
and the main electric system and interconnections between WTs within the wind farm must be planned
with the aim of reducing reactive power generation, power losses and total costs.
The available civil engineering infrastructure should provide route access (e.g., through highways,
paved/unpaved roads, tunnels, bridges, railways, etc.) to the wind farm site. It is also desirable to
have airports and/or seaports nearby to improve logistic operations during the construction process.
Additionally, it should provide a way (e.g., radio frequency-based, telecommunication, Internet-based,
wired connection-based or satellite connection-based) through which important information (e.g., wind
resource and wind farm real-time operation) can be consulted and monitored, however it mostly
depends on the coverage of the local network service providers. Civil engineering and telecommunication
infrastructure is commonly not encountered within the wind farm potential areas, thus the wind farm
designer usually incur additional expenses [74,76] by building auxiliary infrastructure (e.g., suitable
auxiliary roads), which depends exclusively on the wind farm layout, local orography and restricted areas.

2.2.3. Environmental Impacts

Although the World Health Organization (WHO) recognizes wind energy as the power conversion
technology with the lowest impacts (during construction and operation) on human health and the local
environment [92–96] precautions must be taken during the wind farm design process. Environmental
impacts can be defined as all the potential impacts the wind farm will produce in the environment of
the site during its construction and while operating. Environmental Impact Assessments (EIAs), which
must be in compliance with existing guidelines proposed by national Environmental Affairs Agencies
(EAA), should in general include the description and study of the impacts on the air, water and soil.
In addition, EIAs should study the effect of possible pollution emissions and their distributions in
space and time, noise generation, shadow flicker issues, visual impact and the impacts on wildlife and
biodiversity. Examples of EIA guidelines are the United Nations Environment Programme (UNEP)
guidelines [97] and the International Finance Corporation (IFC) performance standards on
environmental and social sustainability [98]. The environmental factors that most affect the micro-siting
process are briefly described next:
Energies 2014, 7 6937

Wildlife and Biodiversity

Wind farms may produce side effects during their construction and while operating on: (1) the local
vegetation, by inducing local micro-climate changes (e.g., a spatial and temporal turbulence-induced
impact on the heat flux, temperature and moist concentrations of the atmospheric surface layer) [99,100]
and (2) the local wildlife, with special emphasis on the short and long-term loss, degradation and
fragmentation of bird, bat and marine life (e.g., for offshore wind farms) habitats [36,101–107]. Most
of the potential impacts of wind farms on migrating birds/bats can be reduced to acceptable levels
through a smart micro-siting design (e.g., bird electrocution fatalities can be dramatically reduced by
burying internal wind farm power lines underground while bird-WT collision fatalities can be reduced
by properly separating WTs from sensitive zones such as migration corridors) and by performing
site-specific mitigation processes. Identifying zones (e.g., natural protected areas) of possible concern
should be done as early as possible in the wind farm planning stage. In high-risk zones the ultimate
mitigation solution could be to implement control strategies on selected WTs during bird or bat migration.
Decision and assessment of such possible actions will require detailed site-specific studies.

Visual Impact

Compared to other electricity-producing technologies, wind farms produce relatively small visual
impacts. Visual impacts are site-specific and can be determined from the size and relative proximity
between the observer and the WTs, the spread or extend of WTs, the WTs view orientation, local
topography, the position of nearby buildings, and the local vegetation. Global public surveys have
reported strong support for having WTs installed close to urban areas [108]. However, local opposition
to wind farms may arise depending on the wind farm location and the relatively closeness of WTs to
human settlements, rural and/or urban zones. Visual Impact Assessments (VIA) [109–113] essentially
consist in documenting the baseline conditions of the potential site, perform accurate visualizations of
the proposed wind farm layout with a photo simulation or a three-dimensional planning software, and
analyze the changes compared to the baseline situation. The metrics used within the VIA process includes a
measure of the visual contrast that WTs introduce to the landscape. Proper micro-siting helps to improve
the visual integration of WTs, cables, substations and access roads in the landscape [31]. However,
the decision whether WTs intrude on the landscape is highly subjective. Therefore, variations on VIA
are commonly encountered depending on human perception of visual effects, and sometimes it is not
simply a function of size, number, color or distance between WTs and observers.
WT rotors are prominent objects whose rotational movement induces attention. In special situations,
WT blades may interrupt the sunlight or moonlight producing a flicker of light that can affect illumination
in nearby areas. This effect is commonly known as shadow flicker or stroboscopic effect. WT shadow
flicker has the potential to induce photosensitive epilepsy seizures [114], however, the risk is minimized
with a proper wind farm layout planning. Shadow flicker is predictable and can occur only in specific
conditions, locations and time of day. The recommended shadow flicker setback distance between WTs
and residences is 1 km [94]. However, greater setback distances may be required when WTs are sited on
elevated ridges, as the shadow flicker can be cast over larger distances.
Energies 2014, 7 6938

Noise

The total noise emitted from a WT can be decomposed into: (1) mechanical noise and
(2) aerodynamic noise. The mechanical noise [36,115] is generated from the rotating machinery of
the WT (e.g., gear box, generator and yaw system among other sound-emitting components).
The aerodynamic noise [36,116] is due to the fact that WT rotors are not 100% efficient when
converting kinetic energy from the wind into mechanical energy. Therefore, some of the kinetic energy
turns into noise and heat, the former being dominant. The aerodynamic noise is a function of many
variables [117] including the relative spatial position between the WTs and the receptor, turbulent
inflow conditions (including wakes), WT blade velocity, local wind shear condition and WT blade
shape among other variables. The total WT noise intensity decays with distance from the source in an
anisotropic way [36,118], however, it can be combined incoherently with other sounds coming from
different sources leading to increased levels of noise. Due to the aerodynamic nature of both energy
conversion and noise generation, these are conflicting factors. Therefore, appropriate distances should be
maintained between WTs and sensitive receptors during wind farm layout design in order to avoid
noise nuisance [93,113,119–121]. Once the wind farm is in operation, control mitigation techniques
can be adopted for particular scenarios, as suggested in [36,122]. However, control mitigation
techniques can lead to important reductions in the energy conversion. This problem is stressed when
considering cases where wind resources are found close to human settlements, rural and urban zones.
The wind farm designer must perform Noise Impact Assessments (NIA) or follow any of the
existing national regulations for noise generation or other regulations or standards such as (1) the
Occupational Safety & Health Administration (OSHA) standards [123], which specify permissible
noise exposures in terms of a time-weighted average sound level or a daily noise dose; and/or (2) the
International Energy Agency (IEA) recommended practices [36,124,125] to reduce impacts at potentially
sensitive locations.

Electromagnetic Interference

Any static or dynamic mechanic system located in the vicinity of a transmitter or receiver of
electromagnetic signals (e.g., FM broadcast, UHF television broadcasts, Doppler Very High Frequency
Omni Range (DVOR), Long Range Navigation (LORAN), microwave links, satellite communications,
etc.) may act as a potential source of Electromagnetic Interference (EMI). WTs and power lines can
passively and actively produce EMI by: (1) the near electromagnetic fields emitted by the electric
generator, electronic devices and components located inside the WT nacelle; (2) diffraction by reflection
of a part of a signal, which is a consequence of obstructing the signals path; and (3) forward and
backward scattering, in which obstacles absorb signals and act to produce and transmit scattered
(modified) signals, which in turn are interpreted as EMI by transmitters and receivers [126–130].
As a result of the signal scattering from the WT blades and tower, the signal at the receiver is generally
both amplitude- and phase-modulated, the former being dominant. Signal modulation can adversely
affect the performance of electromagnetic systems. The nature and degree of the resulting EMI effects
will depend on the signal transmission and reception methods used by the communication systems,
the climate, the electromagnetic scattering characteristics of the WT and other electrical components
Energies 2014, 7 6939

(e.g., WT blade composite materials, the shape, size and material of the WT tower and nacelle,
transformers, etc.), the signal frequency and the geometric factors such as the relative position between
the WTs and the communication systems. The basic parameter used to measure EMI effects on a given
communication system is the ratio of the amplitude of the scattered or secondary signal to the amplitude
of the desired ambient or primary signal. Any possible EMI related problems can be prevented by
properly designing the wind farm layout. However, other technical solutions or mitigation techniques
such as the installation of amplifiers, re-transmitters, transmitting antennas or even repeater stations
could reduce the amount of unwanted EMI. EMI can be minimized by ensuring that all equipment
complies with relevant electromagnetic compatibility standards, by following any national legislations
and by performing EMI assessments [131,132].

2.2.4. Uncertainty and Risk Management

The uncertainty associated to a process can be defined as a state of limited knowledge, in which it
is not possible to describe with accuracy and/or precision the actual state and/or a future state and its
outcome(s), if any. Uncertainty can be cataloged as epistemic, when a lack-of-knowledge on the process
exists, or as random, when parameters vary stochastically during the process development. The metric
used to measure uncertainty is the possible deviation in the prediction of the actual and/or future events
based on the actual knowledge of the process. The risk of a process, therefore, can be defined as a state
of uncertainty where the possible outcomes are undesired. Uncertainties associated to the planning
process of wind farms can be decomposed in three main tracks:
(1) Uncertainty related to the wind resource estimation. Uncertainties in wind resource are divided
in four general categories [133]: (1) uncertainty due to the measurement processes; (2) uncertainty in
the short- and long-term variations of atmospheric variables; (3) uncertainty due to wind resource
variability; and (4) uncertainty due site assessment. The large-scale atmospheric effects and the long-term
variation of atmospheric variables imposes the major challenge [134,135], as they are considered to be
the largest sources of uncertainty during the micro-scale wind resource estimation.
The certainty and quality of the measured wind resource is crucial during: (1) spatial wind mapping
and during (2) WT wake-based power losses quantification [136], both inducing additional uncertainty
due to their modeling assumptions, inaccuracies and limitations (e.g., the induced uncertainty related to
the limited resolution of topographic and roughness maps, wind direction uncertainty, flow modeling
limitations, etc. [106]). The objective of the wind farm designer is to increase the accuracy and precision
of both the measurements and predictions of atmospheric variables values in space and time.
(2) Uncertainty related to the WT technology. Uncertainties in WT performance may be divided
in three: those related to (1) the WT manufacturing process; (2) the WT power conversion model
(e.g., the power output as a function of atmospheric variables and other undesirable effects such as
wakes, WT blade soiling and degradation and mechanical/electrical losses among others); and (3) those
related to the WT technical availability (e.g., O&M scheduling). The objective of the WT designer is
to reduce the performance deviations from the theoretical expectations of the constituent mechanical
and electrical systems of a WT.
Energies 2014, 7 6940

(3) Uncertainty related to economic prospections. The objective of the wind farm designer is to
increase the accuracy and precision of the profitability projections of the project accounting for the
evolution of the energy prices as well as all relevant variable costs (O&M, WT degradation, civil work,
decommissioning, etc.).
All uncertainty tracks are related and each one depends on the previous one as naturally listed
before. Therefore, the total economic prospection uncertainty might be high if no proper uncertainty
mitigation processes are performed, thus affecting the final potential funding decisions of financial
institutions. To improve the economic prospection certainty, the wind farm designer must follow
existing guidelines and standards to mitigate uncertainty at all levels. Some examples include, but are
not limited to:
- IEC 61400-12, Annexes D, E, F, G, I, J and K [29]. Anemometry, calibration procedures,
meteorological mast geometry and induced flow affectation.
- Wind Turbine Design Cost and Scaling Model from the National Renewable Energy Laboratory
(NREL) [137].
The wind farm designer can considerably reduce the profitability risk and increase the financing
possibilities of the project by performing a smart micro-siting (e.g., avoiding WT wake effects, locating
WT in zones where wind resource certainty is higher, etc.), given some uncertainty conditions, while
providing the expected minimum and maximum levels of profitability [138].

2.3. The Wind Farm Design and Optimization Problem Definition

The WFDO problem is defined as the set of advanced planning actions needed to extremize the
performance of wind farms, while satisfying all constraints, given partial or complete information
regarding: (1) the wind farm location, the wind farm size, the topography complexity and land-use;
(2) the short- and long-term meso- and micro-climate and wind resource; (3) the available
infrastructure (e.g., electrical, civil engineering and telecommunication infrastructure); (4) the
available technologies (e.g., WT type, size, efficiency and cost); (5) the local environment (e.g., wildlife,
bird migration paths, closeness to human settlements, etc.); (6) the national and international
regulations, directives and conventions on energetic resources (e.g., government and private sector
agreements, historical energy pricing, main electric system condition and capacity); and (7) the
potential issues (e.g., cyclic weather phenomena and/or adverse natural phenomena). The definition of
performance is given by the objectives to be extremized, which will be covered in Section 3.1.

2.4. The Wind Farm Design Variables and Constraints

The WFDO problem aims at defining the best value of the following set of variables: (1) WTs
emplacement locations; (2) WTs type (e.g., the technology including the control/operation strategy);
(3) WTs size (e.g., capacity); (4) number of WTs; (5) WTs hub heights; (6) type of tower, sub-structure
and foundation; and (7) type, size and capacity of auxiliary infrastructure (e.g., auxiliary roads, power
collection systems (i.e., transformer number/type/size/capacity and other electric devices), electrical
interconnection layout and capacity (i.e., power lines type and length), repeater stations, etc.).
The impact of these variables on the expected economic profit is difficult to assess either due to the
Energies 2014, 7 6941

uncertainty in the estimation of the involved costs or due to the random nature of some of the adverse
factors influencing the variables. Any assumptions and/or constraints imposed on the decision
variables will naturally facilitate the optimization process but will limit the design space for the
optimal solution, thus possibly resulting in important opportunity costs. Therefore, during the
micro-sitting optimization, those variables should not be limited or constrained without a reasonable
justification. All the factors described in Section 2.2 will affect the design variables.
All variables are naturally bounded and constrained by physical factors: (1) the WTs emplacement
locations are bounded by the considered wind farm site area and constrained by local terrain aspects
and environmental and climatological characteristics (e.g., land use, restricted areas, geotechnical
capacity, setback constraints, obstacles, etc.); (2) the WTs type is constrained by the available technology;
(3) the size and (4) the number of WTs are constrained by the actual power system capacity and
the expected demand from the energy consumers. In addition, the theoretical maximum number of
WTs can be coarsely calculated as the ratio of the wind farm site area to the individual WT area πR2,
assuming that two or more WTs cannot be located within an area of πR2, otherwise blade collisions
would occur. In practice, the maximum number of WTs is much lower due to geometrical and
geographical constraints; (5) The minimum and maximum WT heights are constrained by the rotor
radius (R), the national aviation regulations and the available technology, respectively; (6) The type
of foundation is a function of the water depth (in the case of offshore wind farms), the bearing capacity
of the terrain, the expected loading and the available technology; and (7) the auxiliary infrastructure
design, often considered an embedded optimization problem, is constrained by all described variables
and the current electric, civil engineering and communication system capacity.

2.5. The Computational Complexity of the WFDO Problem

The computational complexity theory [139,140] defines and classifies problems according to their
type and degree of difficulty. The motivation of categorizing the complexity of computational problems
is two-fold: on the one hand, it allows to dimension the computational resources needed to solve the
problem and, on the other hand, it provides the basis for the selection or development of appropriate
solution procedures.
Although optimization problems are the most treated problems in the literature (i.e., for practical
reasons), other different types of problems exist (e.g., functional problems or decision problems) and
their complexity is determined and categorized differently. In mathematics, the term optimization
refers to the study of problems with the aim of minimizing and/or maximizing mono- or
multi-objective functions by systematically choosing feasible values of the variables describing the
mathematical model(s) being extremized. The complexity of an optimization problem model can be
measured in terms of the computational requirements (e.g., the memory space or the execution time)
needed to compute the objective metric value of a potential solution or by its inherent mathematical
logic, among other metrics [139,141–143]. A problem is regarded as inherently difficult if its evaluation
requires significant time and/or space resources. In order to properly express the complexity measures
of an optimization problem model, such as space and time, Turing machine models [144,145], which
are hypothetical devices representing computing machines that can simulate the logic and language of
any algorithm, are commonly employed. The time complexity of an optimization problem model
Energies 2014, 7 6942

is related to the number of elementary operations needed to compute the objective metric value of a
potential solution and may be expressed as a mathematical function T(m), where m is the input data
size to the problem. Elementary operations consume a fixed amount of time when performed, thus
the total amount of time taken and the total number of elementary operations performed differ at most
by a constant factor (γ). Computational time complexity classes [139,141,143,146,147] are defined as
the type of functions T(m) can assume; e.g., if T(m) is a polynomial, then the problem model is said to
have a polynomial time evaluation as a function of the input size, and the corresponding complexity
class will be Polynomial Optimization (PO) or Non-Deterministic Polynomial Optimization (NPO) if
the optimization problem model is solved by a deterministic or by non-deterministic Turing machine
respectively. Additional denominations (e.g., hard or complete) are given to a problem depending on
the relative complexity it exhibits when contrasted with other problems belonging to the same
complexity class. Stating that a problem is the hardest one in a class is equivalent to stating that it is
complete for that class with respect to a suitable reduction (e.g., the NPO-complete class encompasses
a subset of problems, belonging to the NPO class, such that any other problem belonging to the NPO
class can be reduced to them). A reduction [146,148,149] is a method that compares the complexity of
two problems and determines if a problem is more difficult than another (i.e., a reduction can be
understood as an algorithm for transforming one problem into another problem). A problem A is
reducible to a problem B (e.g., 𝐴 ≤𝑅 𝐵) if an algorithm for solving B could also be used to solve the
problem A, thus A cannot be harder than B. Different types of reduction exist (e.g., mapping reduction
(≤m) or polynomial (or Turing) reduction (≤P), among others). The concepts of time and space
complexity, reducibility and completeness were first proposed by Cook [150] to define the hardness of
a problem, and have been extended progressively in the last two decades [146,147]. Many problems
have been shown to be NPO-complete [147,151]. A common characteristic of such problems is that
they do not have a known efficient solution procedure.
Although some authors, based on the terminology proposed by the computational complexity theory,
have claimed that the WFDO problem falls into the NP-hard (see Section 4.5 for misconceptions) class
problems (problems that are at least as hard as any NP problem) [152–154] no formal proof has been
given. Therefore, this section provides a formal derivation of the NPO-completeness [146,147,155] of
the WFDO problem.
Independently of the considered instance, the WFDO problem design variables, although bounded,
describe an infinite search space. However, depending on the assumptions made (e.g., discretization of
the available WT positions within the wind farm site as a function of a predefined minimum allowed
distance between WTs), the WFDO problem may fall into the class of problems known as combinatorial
optimization problems. Let us suppose the simplest version of the WFDO problem (the Linear Wind Farm
Design and Optimization (LWFDO) problem [27,156,157]) where WTs are located in a flat discretized
straight line, having n discretized segments, which is parallel to unidirectional wind speed. Note that the
domain discretization can be uniform or non-uniform as long as the minimum safety distances between
WTs are respected. WTs are considered to be of the same height, size, type, and operate in steady state
conditions with constant thrust and power coefficients without regard to the inflow condition, but
respecting the Betz-Lanchester theoretical limit [158]. No auxiliary infrastructure (e.g., roads or power
lines) is to be built, and no other economic or environmental aspects are taken into account. Constant,
isotropic free-stream air properties (e.g., density) and wind speed in a neutral atmosphere are assumed.
Energies 2014, 7 6943

Uniform surface roughness is also assumed. The mono-objective function to be maximized is the total
instantaneous mechanical power conversion (see Sections 3.1.1 and 3.1.2 for details) of the wind farm
(defined as the sum of the power conversion of all the WTs, which can be determined using Equation (2)),
while considering wake effects (e.g., the Jensen-Katic wake model, Equations (15) and (39), and the
wake superposition by sum of squares of velocity deficits method, Equation (41), are assumed).
The design variables ( 𝑥𝑖 ∈ {0,1}, ∀𝑖 ∈ {1,2, … , 𝑛} ) represent if a WT is located or not on each
discretized space.
By summing up all the elementary operations needed to quantify the instantaneous power conversion
of a linear wind farm, consisting of 2 ≤ m ≤ n WTs, (i.e., employing Equations (2), (15), (39), and (41)),
an expression that quantifies the total computational time required to evaluate the performance of the
linear wind farm was derived, as shown in Equation (1):
𝑇LWFDO (𝑚) = (5.5𝑚2 + 3.5𝑚 − 5)𝛾 (1)
where γ is a factor representing the time required to perform an elementary operation. The LWFDO
problem lies within the NPO class, since for a given configuration of m WTs, the mono-objective
function can efficiently (in polynomial time) quantify the total instantaneous mechanical power
conversion of the wind farm (considering the simplicity of the physical models and their limitations).
Considering other currently available engineering wake models and/or wake superposition methods
will modify the resultant coefficient values of Equation (1) but will not affect the maximum order of
the polynomial.
The LWFDO problem can be easily transformed into a decision problem (LWFD) by translating the
main objective (i.e., obtaining the maximum power conversion) into the objective of determining if
concrete questions are true or false (e.g., given a specific instance, is there a linear array of WTs that
can convert “at least/at most/more than/less than” k amount of mechanical power from wind?).
The LWFD decision problem lies within the NP class, as it has the same elementary operation
requirements as the LWFDO problem. Both the LWFDO and the LWFD problems, as presented, are
non-convex, Integer Nonlinear Programming (INLP) problems. It has been shown by Karp [151] that
the general 0–1 integer programing decision problems are NP-complete (problems which are both in
the NP and in NP-hard subsets). The demonstration consisted in performing an explicit
polynomial-time Turing reduction (or Cook reduction [150]), showing that the Boolean satisfiability
problem (e.g., 3-SAT) is reducible to the general 0–1 integer programing problem. Therefore,
the Boolean satisfiability problem can be successfully reduced to the LWFD decision problem, resulting
in its NP-completeness. When the decision version of an optimization problem is proved to belong to the
class of NP-complete problems, then the optimization version is NPO-complete [146,147].
If the LWFDO problem considers a discretized domain with n discretized locations, then there are
2𝑛 − 𝑛 − 1 non-trivial possible configurations to be evaluated. An exhaustive enumeration algorithm
applied to the LWFDO problem will have a time complexity of O(2n); therefore it is considered to be an
intractable, exponential time solution procedure and falls into the NEXPTIME complexity class [139–141],
one of the most time-consuming optimization procedures. Optimal solutions to the LWFDO problem
can be confirmed only for small instance problems (e.g., instances allowing 30 discretized locations
will have a search space of 1.0737 × 109 non-trivial possible solutions). For a particular (non-trivial)
instance of the LWFDO problem, there is only one optimal solution [156].
Energies 2014, 7 6944

The LWFDO problem model cannot be completely described in an analytical form (i.e., providing
a closed form analytical expression for the Annual Energy Production [159]) and due to the
non-differentiable, non-convex and non-linear behavior of the mono-objective function, calculus-based
optimization approaches such as linear/quadratic programming, gradient methods, Newton-Raphson, etc.
cannot be used to solve the problem. To attempt solving optimization problems of such complexity,
heuristic and metaheuristic approaches are commonly employed, yielding near-optimal solutions as
discussed in Section 3.3.
The LWFDO problem can be further extended to two different problems: (1) the two-dimensional
Linear Wind Farm Design and Optimization (2DLWFDO) problem in which WTs are allowed to have
different hub heights [138,156,160,161]; and (2) the two-dimensional Wind Farm Design and
Optimization (2DWFDO) problem in which fixed hub height WTs are located in a finite discretized
regular/irregular planar space (which was considered in most of the reviewed works). For both cases,
the amount of elementary operations will increase in a polynomial way by considering the partial wake
operation of partial wake-affected WTs and by the different inter-distances quantification method.
Furthermore, additional elementary operations will be required if multi-directional, anisotropic
free-stream wind, realistic WT efficiency and control operation as a function of the inflow condition
and other complex objective formulations are considered. However, the total amount of elementary
operations can still be described as a polynomial function of the amount of WTs in the
two-dimensional array. Nevertheless, depending on the approach taken for the computation of the wind
field and the wake interactions (e.g., by using CFD techniques [14,162]) the elementary operations
may grow in an exponential or hyper-exponential way. Therefore, the 2DLWFDO and the 2DWFDO
problems are NPO-complete as well as their respective decision problem versions. An enumerative
algorithm will have the same time complexity, O(2n), as in the LWFDO case. Two important
differences must be noted for the 2DWFDO problem in contrast with the LWFDO and 2DLWFDO
problems; the 2DWFDO problem may have multiple optimal solutions (multimodal space) and
typically contains more decision variables, thus being harder to solve.
The 2DWFDO problem can be further extended to the generalized WFDO problem. The generalized
WFDO problem includes all the mixed-integer variables described in Section 2.4. In addition, it considers
the decision variable of designing auxiliary infrastructure (e.g., roads and/or electrical infrastructure),
which can be viewed as an embedded NPO-complete problem (e.g., the auxiliary road and/or electrical
infrastructure layout design problem can be viewed as a Minimum Spanning Tree problem [135,163–166]
or as a Steiner tree problem [167]; both problems have been categorized as NPO-complete
problems [147,151]) or as an embedded NPO-hard problem (e.g., the wind farm array cable layout
design problem can be viewed as a vehicle routing problem [168], which has been categorized as a
NPO-hard problem). Currently, there is no defined computational complexity category for embedded
NPO-complete/NPO-hard problems, and current solution procedures rely in multi-step, nested and
hybrid approaches, which will be discussed in Section 3.3.4. Finally, all WFDO problem versions have
non-parallelizable objective functions (e.g., it is not possible to parallelize the evaluation of a wind
farm configuration due to the multiple wake affectation between WTs) and, therefore, its evaluation
proceed in a sequential way. This, however, does not necessarily imply that certain numerical solution
procedures cannot parallelize the evaluation of multiple candidate solutions or that cannot parallelize
both the objective function and the solution procedure of the embedded optimization problems.
Energies 2014, 7 6945

3. Relevant Research in Wind Farm Design and Optimization

The present section describes the evolution of the research in WFDO. A total of one hundred and
fifty works were reviewed, consisting of two Ph.D., four M.Sc., and one B.Sc. thesis, seventy journal
articles, sixty three conference proceedings, five technical reports, and five book chapters, respectively.
Our exhaustive search for literature works, up to 20 October 2013, considered more than fifteen search
engines, scientific databases, scientific journals and scientific magazines. As observed in Figure 1, the
evolution of the research on this topic has been overwhelming in the last decade. Not unexpectedly,
most of the research works were published in journals or conference proceedings related to wind
energy science. However, a notable amount of works were published in journals with different scopes.
This is mainly due to the multidisciplinary nature of the problem.

Figure 1. Timeline of Scientific Works Related to the WFDO Problem.

Timeline of Type of Works
30
Ph.D. & M.Sc. Thesis
Journal Papers
25 Conference Proceedings
Technical Reports
Book Chapters
20
No. of Works

15

10

0
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Year

Three review articles on the WFDO problem were identified in the literature [169–171]. The present
work goes beyond the existing review papers by: (1) performing a deep analysis of the models describing
the performance of wind farms and the sub-models describing the adverse factors, presented in
Section 2.2; and by (2) identifying and reviewing the most prominent published works on the WFDO
problem, with particular emphasis on both their shortcomings and on future research opportunities.

3.1. Objective Functions

Five different performance metrics, each one showing different complexity, were identified in the
literature, as shown in Figure 2. The most widely used metrics are the wind farm Annual Energy
Production (AEP), the instantaneous power conversion and the Cost of Energy (CoE).
Four metrics equivalent to the AEP and the instantaneous power conversion were found in the
literature: (1) the wind speed reaching each WT in the wind farm (to be maximized); (2) the wind
Energies 2014, 7 6946

speed deficit at each WT (to be minimized); (3) the wind farm capacity factor (to be maximized); and
(4) the wind farm efficiency (to be maximized), defined as the ratio of the wind farm power conversion
to the ideal wind farm power conversion (if no wake and turbulence effects are taken into account).
Two equivalent definitions to the CoE were identified in the literature and are discussed in
Section 3.1.2: (1) the Levelized Cost of Energy (LCoE) and (2) the Levelized Production Costs (LPC).
In addition, different definitions of Financial Balance (FB) (or economic profits) were identified in the
literature. The main difference between FB models reside in which costs are taken into account. Some
works did not implement any performance metric, as their objective was to treat a specific issue of the
WFDO problem (e.g., review works, derivation of mathematical approximations, etc.).

Figure 2. Timeline of Objective Functions used in the WFDO Problem.

Timeline of Objective Functions
30
NA
Max AEP
25 Max Power
Min CoE, LPC
Max FB
20 Max NPV
No. of Papers

15

10

0
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Year

3.1.1. Power Conversion and Annual Energy Production (AEP)

The WT mechanical/electrical power conversion and the WT Annual Energy Production (AEP) are
the basic metrics used to evaluate the performance of wind farms. More than one third of the reviewed
works used one of these metrics as an objective function [11–15,27,156,157,159,163,172–213]. Most
of the non-linear behavior of the wind farm planning process comes from the quantification of both
metrics. The mechanical power of uniform, isotropic and steady wind (𝑈∞ [m/s]) crossing the swept
rotor area (𝐴 [m2]) of a WT, derived from the first law of thermodynamics, is the product of the wind
2
mass flow rate (𝜌𝐴𝑈∞ [kg/s]) times the kinetic energy of the wind (𝑈∞ /2 [m2/s2]), where ρ [kg/m3] is
the air density. The instantaneous mechanical power conversion of a WT, as shown in Equation (2),
is the product of the power conversion potential times the WT power coefficient (𝐶𝑃 ), a variable which
summarizes both the aerodynamic and the electromechanical properties of the WT under consideration
and is often considered as the WT efficiency metric. The power coefficient is a complex function
of many variables including the free-stream wind speed, Turbulence Intensity (TI [%]), wind shear,
yawing angle misalignment (Δθ [°]), the blade aero-structural condition, blade pitch setting, and the
Energies 2014, 7 6947

Tip Speed Ratio (TSR, 𝜆W = 𝛺𝑅/𝑈∞ , where 𝑅 [m] is the rotor radius and 𝛺 [rad/s] is the rotational
speed), among other variables [29,214,215]. Its mathematical representation is strongly influenced by
the adopted control strategy (e.g., stall regulation or pitch regulation) and is commonly reported by WT
manufacturers as an empirical function of the wind speed at hub height (𝑈Hub [m/s]), while assuming
standardized inflow conditions. Due to the complex requirements of the empirical characterization
procedures, it is hard to express the power coefficient as an analytical expression of the previously
mentioned variables. Therefore, the power coefficient is often considered to be a function of the
assumed homogeneous, isotropic, free-stream wind speed at the rotor (𝑈Hub = 𝑈∞ ), while accounting
for the selected control operation strategy, which in turn accounts for the TSR and other
electromechanical variables:
1 3
𝑃Mech (𝑈∞ ) = 𝜌𝐴𝑈∞ 𝐶𝑃 (𝑈∞ ) (2)
2
Since wind speed and density (which is a function of other thermodynamic variables such as pressure,
temperature and humidity) are time- and space-dependent, as mentioned in Section 2.1, they are often
considered stochastic variables that in most cases show well-defined statistical distributions; often these
distributions are assumed to be Weibullian [216] and Normal [217] distributions, respectively [43,218].
Each statistical distribution is defined in terms of statistical parameters, namely the Weibull scale (𝑐)
and shape (𝑘) parameters, and the normal mean (𝜇) and standard deviation (𝜎) parameters, respectively.
The wind direction is also considered a stochastic variable with an unknown general statistical distribution.
However, it has been shown in [43] that the von Mises distribution, depending on the concentration
(𝑘𝑐 ) and mean (𝜇𝑀 ) modulation parameters, might fit the statistical wind direction behavior well. In
order to properly estimate the mechanical power conversion from the wind, considering the stochastic
nature of the atmospheric variables involved, expected values of the power conversion are calculated.
The general definition of the expected mechanical power conversion of an isolated WT, assuming
anisotropic wind at the rotor, is given in Equation (3):
𝑈Max 𝜌Max TIMax 180
1 3
E[𝑃Mech ] = ̅̅̅̅̅̅̅̅
𝑃Mech = ∫ ∫ ∫ ∫ ∫ 𝜌𝑈 𝐶𝑃 (𝑈, 𝜌, TI, Δ𝜃) 𝑝0 (𝑈, 𝜌, TI, Δ𝜃) dΔ𝜃 dTI d𝜌 d𝑈 d𝐴 (3)
2
𝐴 0 0 0 0

where 𝑈Max , 𝜌Max and TIMax are the maximum wind speed, the maximum density and the maximum
turbulence intensity observed at the WT location in a differential segment of area within the WT swept
rotor area. 𝑝0 (𝑈, 𝜌, TI, Δ𝜃) is a multivariate probability density function used to describe, at least
approximately, the probability of occurrence of the correlated, real-valued random variables 𝑈, 𝜌, TI
and Δ𝜃. Therefore, 𝑝0 represents the probability of occurrence of a wind condition defined by a wind
speed 0 ≤ 𝑈 ≤ 𝑈Max , density 0 < 𝜌 ≤ 𝜌Max , turbulence intensity 0 ≤ TI ≤ TIMax at yaw angle
misalignment 0 ≤ Δ𝜃 ≤ 180 in a differential segment area within the WT swept rotor area.
Unfortunately, it is almost impossible to define 𝑝0 and 𝐶𝑃 as a function of such stochastic variables
in practice. Therefore, simplifications in the expected mechanical power conversion quantification are
performed; it is commonly assumed that: (1) WTs operate in steady state conditions; (2) the atmospheric
variables are uniform within the rotor swept area (e.g., 𝑈 = 𝑈∞ = 𝑈Hub ); (3) the air density is constant
at the rotor swept area and has no effect on 𝑝0 ; (4) TI has no direct effects both on 𝐶𝑃 and 𝑝0 ; and (5) yaw
misalignment effects are completely neglected as it is assumed that WTs can instantaneously realign
Energies 2014, 7 6948

themselves with wind direction. The simplified relation for the quantification of expected mechanical
power conversion is given in Equation (4):
𝑈Max
1
̅̅̅̅̅̅̅
E[𝑃Mech ] = 𝑃 Mech = ∫ 𝜌𝐴𝑈 3 𝐶𝑃 (𝑈, 𝜌) 𝑝1 (𝑈) d𝑈 (4)
0 2
where 𝑝1 (𝑈) is a univariate probability density function that represents the probability of occurrence
of a wind condition with wind speed 0 ≤ 𝑈 ≤ 𝑈Max at hub height and 𝐶𝑃 (𝑈, 𝜌) is the WT power
coefficient corrected for the given air density 𝜌. Even after simplifications, closed-form analytical
solutions for Equation (4) cannot be obtained [159], so its value is approximated numerically using
Riemann discrete summations, Monte Carlo integration or Gaussian quadrature integration, among
other numerical integration methods. The expected mechanical energy conversion of an isolated WT
over a time period T [h] (ignoring down time or technical availability) is given in Equation (5):
𝑇 𝑇 𝑈Max
1
̅̅̅̅̅̅̅
E[𝐸Mech ] = ∫ 𝑃 Mech d𝑇 = ∫ ∫ 𝜌𝐴𝑈 3 𝐶𝑃 (𝑈, 𝜌, 𝑇) 𝑝1 (𝑈, 𝑇) d𝑈 d𝑇 (5)
0 0 0 2
Equation (5) presumes that the WT power coefficient might change in time, as a consequence of
WT degradation. It also implies that variations in wind resource might occur during the period time T,
which in turn affect the probability density function modulation parameters. It is common in practice
to neglect time variations of the power coefficient, since appropriate maintenance is assumed to be
performed during the period of time T. In addition, wind resource characterization campaigns are
commonly performed in short periods of time (e.g., one or two years), and the measured data is not
always reliable due to unexpected issues (sensor calibration, sensor malfunction, etc.). Thus, it is
difficult to perform accurate long-term extrapolations of the local wind resource. Therefore, it is
common in practice to assume that the 𝑝1 modulation parameters will not vary within the period of
time T and are well characterized based on the experimental information obtained in the characterization
campaign. The standard AEP (or expected annual energy production [kWh/y]) of an isolated WT can
be calculated using Equation (6), where 8766 is the number of hours in a year:
8766 𝑈Max
1
AEPWT = ∫ ̅̅̅̅̅̅̅
𝑃 Mech d𝑇 = 8766 ∫ 𝜌𝐴𝑈 3 𝐶𝑃 (𝑈, 𝜌) 𝑝1 (𝑈) d𝑈 (6)
0 0 2
For a wind farm composed of N equal WTs, wind direction cannot be neglected during the AEP
calculation [136], since WT wakes will reduce the AEP expectations of the wake-affected WTs.
Therefore, the AEP of the wind farm (AEPFarm ) is calculated by multiplying the expected mechanical
power conversion of the wind farm with 8766 h/y, as proposed by [55,159,166,199,203,219]:
360 𝑈Max
𝑁 1
AEPFarm = 8766 ∑ ∫ ∫ 𝜌𝐴𝑈 3 𝐶𝑃 (𝑈, 𝜌) 𝑝𝑚,𝐻 (𝑈, θ) d𝑈 dθ (7)
𝑚=1 0 0 2
where 𝑝𝑚,𝐻 (𝑈, θ) is a bivariate probability density function that represents the probability of
occurrence of a wind condition defined by a wind speed 0 ≤ 𝑈 ≤ 𝑈Max from a direction 0 ≤ θ ≤ 360
at the position of the m WT and having a hub height 𝐻. To the wind farm designer, the mathematical
form of 𝑝𝑚,𝐻 is unknown a priori. However, two different approaches that help to approximate the
form of 𝑝𝑚,𝐻 were identified in the literature. The first approach [55,159,219] presumes: (1) that the
statistical distributions of wind speed are Weibullian at all the possible WT locations and WT heights
within the wind farm; and (2) that the Weibull modulation parameters for each of the WT locations
Energies 2014, 7 6949

(𝑘(𝜃, 𝑚) and 𝑐(𝜃, 𝑚)) can be described as continuous functions of the wind direction (i.e., functions
that adequately fit the wind speed distribution for each wind direction at each of the WT positions and
WT heights). This assumption has been shown to have experimental evidence [138,219]. Therefore,
the expected annual energy production of the wind farm (AEPFarm ) can be approximated as shown in
Equation (8) [55,219]:
𝑘(𝜃,𝑚)−1 𝑘(𝜃,𝑚)
360 𝑈Max 𝑈
𝑁 1 𝑘(𝜃, 𝑚) 𝑈 −( )
AEPFarm = 8766 ∑ ∫ ∫ [ 𝜌𝐴𝑈 3 𝐶𝑃 (𝑈, 𝜌)] 𝑝𝜃 (𝜃, 𝑚) [( )( ) 𝑒 𝑐eff (𝜃,𝑚) ] d𝑈 d𝜃 (8)
𝑚=1 0 0 2 𝑐eff (𝜃, 𝑚) 𝑐eff (𝜃, 𝑚)

where 𝑝𝜃 (𝜃, 𝑚) is the wind direction probability density function (e.g., von Mises) and the subscript
eff in the Weibull scale parameter refers to the effective scale parameter (e.g., accounting for wake
losses) used to calculate the AEP of the mth WT. In order to include wake effects in the AEPFarm
calculation, previous works [159,199,219] assumed that the Weibull scale parameter can be scaled,
which is equivalent to scaling wind speed values, while employing any desired wake model. It is often
assumed that WT wakes do not affect the Weibull shape parameter (𝑘(𝜃, 𝑚)) value.
The second approach to describe 𝑝𝑚,𝐻 [203] considered a Multivariate and Multimodal Wind
Distribution (MMWD) model [218,220], which is based on a non-parametric, multivariate Kernel
Density Estimation (KDE) method, used to determine the probability density function of random correlated
variables. The MMWD model is more general than the one presented previously, as it has additional
flexibility to adequately fit the statistical behavior of the correlated atmospheric variables. However,
the inclusion of wake effects during the AEPFarm is not straightforward, but requires a iterative
recalculation of the bivariate probability density function for each wind direction, which in turn might
considerably increase the computational requirements during the wind farm design process.
It should be noted that the AEPFarm considers the transformation of kinetic energy in the wind
into mechanical energy in the WT rotor. It is common in practice to include additional constant value
factors describing the net mechanical-to-electrical efficiency. However, both the mechanical and electrical
efficiencies will often be a non-linear function of the WT operation mode and electric system condition
(e.g., TSR, drivetrain efficiency, electric generator efficiency, effective electric load, power line
losses, etc.). Therefore, the effective expected electrical energy of the wind farm can be quantified with
the inclusion of an electromechanical model, which accounts for the control mechanisms of the WTs
integrating the wind farm. In order to simplify the calculations of the effective expected electrical
energy, WT manufacturers usually provide empirical electrical power curves (𝑃Elec (𝑈∞ )), which are a
function of the free-stream wind speed at hub height and account for all possible mechanical and
electrical losses. The empirical electrical power curve replaces 𝑃Mech in Equations (2)–(8) during the
AEPWT and AEPFarm calculation, resulting in the effective expected electrical energy conversion of the
WT and the wind farm respectively.
The WT availability represents the fraction of time that the WT will be operating and supplying
electricity. The availability can be broken into the inherent availability and the operational (or technical)
availability. The WT inherent availability excludes preventive maintenance downtime, logistics time,
and administrative downtime, but it includes corrective maintenance downtime. The WT operational
availability includes logistics time, ready time, administrative downtime, and both preventive and
corrective maintenance downtimes. Both types of availability are generally determined based on statistical
metrics such as the Mean Time to Failure (MTTF), the Mean Time to Repair (MTTR), the Mean Time
Energies 2014, 7 6950

Between Failure (MTBF), and the Mean Downtime (MDT), which in addition provide information for
WT maintenance scheduling [221–225].
In wind energy, the operational availability measure and the concept of reliability are extended by
including external elements to the wind farm (e.g., the main electrical system), which are controlled by
logisticians (e.g., main electrical system operators). Therefore, issues related to electrical instabilities
and energy curtailments, among other external factors, are often included in the availability and reliability
definitions. It is generally acknowledged that energy curtailments produce important impacts on wind
economics [226,227]. Curtailments occur for a number of reasons including electrical grid congestion,
energy oversupply, and operational issues. Each type of curtailment occurs with different frequency
and depends on the regional and the local system’s energy generation and its electrical characteristics.
Electrical system operators typically curtail wind-based energy to maintain the system’s balance and
reliability. Curtailments can be regarded as electrical system-dependent and country-dependent issues
of operating wind farms. The modeling and prediction of curtailments requires specific knowledge of
the wind resource, wind farm design, wind farm control system, main electrical system condition, PPAs,
and the energy regulatory framework (which, for example, states if any excess in energy conversion
may be purchased by the energy supplier/operator or not), among other relevant factors. The WT and
the wind farm total availability are often included as constant factors multiplying Equations (6) and (7),
respectively. It is common to assume conservative values for the availability factor ranging from 95%
to 98% [55,56,135]. However, the WT availability and reliability can be estimated and tracked more
accurately by following the IEC 61400-26 availability standard for WT and wind farms [29] and other
procedures found in the literature [188,228,229].

3.1.2. Cost of Energy (CoE) and Levelized Cost of Energy (LCoE)

The Cost of Energy (CoE) [$/kWh] is defined as the cost per kWh of energy (i.e., effective expected
electrical energy) converted from the wind and represents the price of energy at which the wind farm
project won’t have neither economic profits nor economic losses over its operational lifetime (i.e., the
breakeven price of energy). The general mathematical expression used to quantify CoE is shown in
Equation (9):
CoE = Cost Farm /AEPFarm (9)
Several definitions of CoE were identified in the literature, differing in the way the total cost
of the wind farm ( Cost Farm ) is quantified. More than one third of the total reviewed works
(representing more than 80% of the works that adopted the CoE as the wind farm performance
metric) [28,54,55,84,137,138,161,181,219,230–271] considered the Mosetti’s cost function, in which
the total cost of the wind farm is assumed to be an exclusive function of the number of WTs (N)
constituting the wind farm, as shown in Equation (10):
2
Cost Mosetti = 𝑁(2/3 + 1/3 𝑒 −0.00174𝑁 ) (10)
It should be noted that Equation (10) is dimensionless and represents only a fraction of the total
capital costs of wind farms, while disregarding all other variable costs. Mosetti et al. [28], who first
proposed the approximation, assumed a dimension of cost/year [$/y], which is inconsistent. In addition,
it is very unlikely that WT manufacturers adopt this simplified WT cost function, in which for N > 35
Energies 2014, 7 6951

the function value converges to 2/3N. The reason for using Equation (10) derives primarily from the
wish of the authors of the reviewed works to compare their results with the results of the work of
Mosetti et al. In such works, the main objective was to develop better numeric solution procedures
instead of improving the wind farm performance modeling. Although the acquisition of WT represent
the major part of the total capital cost of the wind farm [55,135], other fixed and variable costs cannot
be neglected [3,75,76]. More complex cost models, which include additional fixed and variable costs,
are discussed in Section 3.2.6.
Equivalent measures of the CoE are given by the Levelized Cost of Energy (LCoE) or the Levelized
Production Cost (LPC) [$/kWh]. Both the LCoE and LPC are similar in definition to the CoE, but both
make a clear distinction between fixed and variable costs affecting the profitability of the wind farm
and might include the time variation of the value of money, which in turn includes the concepts of
interest (e.g., related to the additional payments required to pay loans) and the inflation rate. In the
novel approach of Elkinton et al. [55,235], within the scope of the OWFLO framework, the LPC was
defined as:
𝐶Inv 𝐶O&𝑀
LPC = + (11)
𝑎 𝐸𝑎 𝐸𝑎
𝑁
𝐸𝑎 = 𝐴𝐹 [∑ (𝐸WT,𝑖 − 𝐸Loss,wake,𝑖 − 𝐸Loss,col,𝑖 ) − 𝐸Loss,trans ] (12)
𝑖=1

where 𝐸𝑎 [kWh] is the net effective expected electrical energy of the wind farm, accounting for wake
losses and power lines (MVAC and HVAC) losses, 𝐶Inv = 𝐶RNA + 𝐶SS + 𝐶Elect + 𝐶Decom [$] is the
total investment cost (or capital cost), 𝐶RNA is the cost of the WT Rotor-Nacelle Assembly (RNA), 𝐶SS
is the cost of the support structure (including tower), 𝐶Elect is the cost of electrical interconnection
(collection and transmission power lines), 𝐶Decom is the on-time decommissioning cost, 𝐶O&M are the
annual costs for maintenance and operation of the wind farm, 𝑎 = (1 − (1/(1 + 𝑟𝑖 ))𝑇 )/𝑟𝑖 is the annuity
factor, 𝑇 [y] is the economic lifetime of the wind farm and 𝑟𝑖 [%] is the interest rate. Details of the cost
components are given in Section 3.2.6. Other LCoE models were found in the literature [137,272,273]
but are still not implemented for the WFDO problem. A very comprehensive review of the economics
of wind energy and the concepts of CoE and LCoE can be found in [274–276]. In addition, the works
of Elkinton [55] and González et al. [277,278] present typical costs for both onshore and offshore wind
farms, which have been tailored to be used in the WFDO problem.

3.1.3. Sales Revenue, Profit and Financial Balance (FB)

The sales revenue or turnover is the income that a company receives from its business activities
(or productive activities), usually from the sale of goods and services to intermediary companies or
end customers at a certain instant in time. The economic profit is defined as the net capital goods a
company earns from its productive activity. The maximization of economic profits in time corresponds
to the accumulation of the largest possible amount of capital goods, which is the driving force behind
economic activities within the capitalist scheme, while considering the time value of money. The role
of the wind farm designer is to balance the fixed and variable investments such that the economic profits
from the energy sales revenues are maximized during the wind farm operational time.
Energies 2014, 7 6952

The effective profit and the Financial Balance (FB) [$] are equivalent in definition. A FB model
may include: (1) time variations in electricity prices; (2) taxes; (3) payments during construction;
(4) residual economic values of the wind farm; and (5) O&M costs that vary over the lifetime of the
project and are related to the WT technical availability. A simple generic FB model [166,279], which
considers only relevant costs to the WFDO problem in a relative basis, is given in Equation (13):
𝑟𝑐 − 𝑟𝑖 𝑇(𝑛𝐿 )
FB = 𝑊𝑃 − 𝐶D − 𝐶O&𝑀 − (𝐶F + 𝐶G ) (1 + ) (13)
𝑛𝐿
where 𝑊𝑃 is the sales revenue of the expected electrical energy conversion over the wind farm
operational lifetime, 𝐶D is the accumulated cost of components degradation, 𝐶O&𝑀 is the cost of overall
operation and maintenance, 𝐶F and 𝐶G represent variable investment costs of foundations and electrical
infrastructure, 𝑟𝑐 [%] is the interest rate, 𝑟𝑖 [%] is the inflation rate, 𝑛𝐿 is the number of times interests
of loans have to be paid in a year, and 𝑇 is the wind farm expected operational time in years. Detailed
explanations on the calculation of cost components are given in Section 3.2.6.

3.1.4. Net Present Value (NPV)

The Net Present Value (NPV) is a widely used financial metric in discounted cash flow analysis and
is a standard method implementing the idea of the time value of money to appraise long-term projects,
such as wind farms. The NPV is similar in definition to the FB. It is defined as the sum of present values
of individual cash flows over finite periods of time. The NPV is a measure of the excess or shortfall of cash
flows, in present value terms, above investments. The general mathematical form of the NPV is given
in Equation (14), as proposed by Castro Mora [56] and González et al. [135,171,278]:
𝑇 𝑁𝐾
NPV = ∑ − 𝐶Inv + 𝑉R (14)
𝑘=1 (1 + 𝑟𝑖 )
𝑘

where 𝑁𝑘 represents the net incomes (or profit, which is the value of the energy sales revenue
minus all variable costs such as O&M costs) produced by the wind farm during the kth year, 𝑟𝑖
represents the equivalent discount rate or interest rate [%], which is an indicator of the opportunity cost
of capital goods (e.g., the higher the discount rate, the lower the present value of the future cash flows),
𝐶Inv represents all the present capital investments during the expected operational lifetime, and 𝑉R
represents any present residual income/outcome after the lifetime of the project. Note that the first
term of Equation (14), commonly known as the cumulative present value of future cash flows,
translates the net value of future revenues into present values and proposes a clear difference between
the NPV/FB and the standard CoE. In addition, the NPV model works with absolute costs in contrast
with the FB model described in Equation (13). A higher NPV is desirable since it means that a proper
balance between the minimization of investments and the maximization of the net cash flows is
achieved. Each term included in Equation (14), with the exception of the interest rate, is a function of
the wind farm layout. The corresponding mathematical expressions for each variable, as proposed by
González et al. [135,278,280], will be discussed in Section 3.2.6.
Energies 2014, 7 6953

3.2. Sub-Modeling

The complete mathematical model describing the performance of the wind farm must be as accurate
as possible, while encompassing all the individual adverse effects described in Section 2. Sub-models
are complex mathematical functions that translate the implications of the adverse effects into the
performance measures described in Section 3.1 (e.g., the WT wake effects are translated into energy
losses and monetary losses through the use of wake models and cost models, respectively). However,
some adverse effects (e.g., total noise, visual impact or electromagnetic interference, etc.) cannot be
easily translated into performance losses. In addition, the restriction on computational resources, being
the processing time the common limiting resource, might influence the scientific approaches applicable
for the formulation of both the various sub-models (e.g., wake and turbulence) and the optimization
procedures. Therefore, many authors have decided to maintain sub-models as simple as possible at the
expense of modeling accuracy. However, since wind farm projects are often huge projects in terms of
economic investments, wind farm designers may be willing to spend time and additional resources to
get both more accurate and near-optimal solutions to its particular projects, thus potentially enhancing
the financing conditions and the expected economic profits. This fact has motivated the research for
the development of more complete and complex sub-models. The most important sub-models, used to
complement the previously described performance metrics, are analyzed next.

3.2.1. Wind Turbine Wakes and Turbulence Modeling

Wake and turbulence modeling [57–59,281] are two of the most challenging research topics in the
field of fluid mechanics and are considered unsolved problems in classic physics. However, a significant
amount of contributions have been performed in the field during the last four decades resulting in two
different types of WT wake modeling approaches. The first modeling approach is based on Engineering
Wake Models (EWMs) [18,19,21,23,282–284], which, in the present context, are analytical functions
derived from empirical regressions, correlations, or simplified Newtonian-based laws of mass flow and
momentum conservation, which describe in a simplified way the WT energy conversion and the WT
wake evolution processes (e.g., the evolution of the streamwise wind speed at any radial position
measured from the WT hub height) while making certain assumptions regarding the near and far wake
expansion, the wake speed recovery, and the wake merging process, which in turn are a function of the
WT aerodynamics and operation characteristics and the local atmospheric condition. The second modeling
approach consist of advanced wake models that are based on the Reynolds-Averaged Navier-Stokes
(RANS) equations or Large Eddy Simulations (LES), both complemented with turbulence models and
subgrid-scale models respectively, and are combined with the concepts of actuator line and actuator
disk approaches for the WT rotor modeling [25,26,57,285–288].
The second approach (based on either the RANS or the LES approaches) has shown improved
accuracy when describing the evolution and the side effects of WT wakes (in both steady and unsteady
conditions) but requires a significant amount of computational resources, thus is commonly not adopted
by wind farm designers nor by most of the authors of the reviewed works. The first approach (based on
EWMs), although simple, physically limited and computationally inexpensive, has proven to describe
the wind speed deficit in the far wake zone with reasonable accuracy [60,65–68]. However, EWMs
Energies 2014, 7 6954

typically overpredict wake effects [58,60,66,281,289]. Moreover, some commercial software packages
(e.g., WindFarmer, WindFarm, WindPro, OpenWind) complement EWMs with Engineering Turbulence
Models (ETMs) [288,290–295], which describe the net TI evolution within the WT wake. Nonetheless,
none of the reviewed works considered the inclusion of ETMs, since TI effects on the energy
conversion process are often neglected.
The basic mathematical structure of the EWMs is shown in Equation (15):
𝑈W (𝑥, 𝑟) = 𝑈∞ (1 − 𝑈Def (𝑥, 𝑟)) (15)
where 𝑈W (𝑥, 𝑟) is the horizontal wind speed in the wake at some axial (𝑥 [m]) and radial (𝑟 [m])
position downwind from the WT hub, and 𝑈Def (𝑥, 𝑟) is the velocity deficit, which represents the
fraction of the free-stream wind speed that was affected during the energy conversion process. The
wind speed in the wake exhibits a natural recovery due to the turbulent and viscous interactions of
the wake with the unaffected neighboring free stream. The wake recovery is commonly modeled
through the use of wake decay coefficients, which are assumed to be exclusively dependent on the
ambient turbulence intensity levels [19,23,282,283].
The differences between common EWMs are essentially the way they approximate 𝑈Def . The first
wake model, categorized as a Kinematic wake model [59], was proposed by Lissaman in
1977 [19,296,297], and later was updated by him in 1982 [24]. He assumed that the wake of a single
WT can be decomposed into two different main regions: the near and the far wake regions. For each
region, self-similar wind speed profiles were assumed. Between these two regions is a transition zone,
in which the flow profile changes from one self-similar profile to another. The wake is assumed to
grow by turbulent entrainment at its edge, which introduces free stream momentum as well as mass
into the wake, so these quantities are not conserved in the wake. Nonetheless, the momentum deficit in
the wake, considering an idealistic constant pressure field, is conserved. The momentum deficit is
defined by the integral shown in Equation (16) across the wake:
𝑈W 𝑈W
𝐷=∫ (1 − ) 𝑟d𝑟 (16)
𝑈∞ 𝑈∞
where the free-stream wind speed (𝑈∞ ) was assumed constant and homogeneous. Equation (16), along
with the wake growth (d𝑅W /d𝑥), and the wake profile at a radius r [m], measured from the wake
centerline and given by a function of the form 𝐹(𝑟/𝑅W ) = 𝑈W /𝑈∞ (where 𝑅W is the effective outer
radius of the wake), are the most important mechanisms modeling the wake. The momentum deficit of
the wake can be computed from the initial state of the wake, which can be directly related to the thrust
force, or the thrust coefficient (𝑐𝑇 ), of the WT rotor and its output power. The thrust coefficient of the
WT represents the normalized axial force produced by the wind on the WT rotor, or conversely, of the
rotor on the wind. The thrust coefficient can be viewed as a basic parameter representing the aero-structural
loading of the WT and is dependent on the same variables as the power coefficient (e.g., 𝑈, 𝜌, TI and
Δ𝜃), as discussed in Section 3.1.1. Similarly to the power coefficient, the thrust coefficient is strongly
influenced by the adopted control strategy and is commonly reported by WT manufactures as an
empirical function of the wind speed at hub height, while assuming standardized inflow conditions and
disregarding any other dependence on the previously mentioned atmospheric variables.
Energies 2014, 7 6955

Lissaman assumed that turbulence near the edge of the wake controls the lateral growth. This
turbulence is a functional combination of: (1) the ambient turbulence intensity; (2) the turbulence generated
by the shear due to the wind speed gradients within the wake itself (or wake shear); and (3) the
mechanically added turbulence generated by the viscous interactions of the free-stream flow and the
WT rotor (i.e., the mechanically generated turbulence). These terms have a different relative importance
in the different regions of the wake. The ambient turbulence (assumed isotropic and uniform in the free
stream) controls the wake growth after the mechanically added turbulence and the turbulence generated
by the wake shear have decayed. Thus in the near wake region, the three turbulent components drive
the wake growth, and in the far region only the ambient component drives the wake growth.
At the beginning of the near wake region, the radius of the wake is given by 𝑟0 = 𝑅√(𝑚W + 1)/2
and at the end of the near wake region is given by 𝑟𝑐 = 𝑟0 /(0.214 + 0.144𝑚W )1/2 , where R [m] is the
WT rotor radius and 𝑚W = 𝑈∞ /𝑈CW is the inverse of the normalized wake speed at the wake core
(𝑈CW ). In the near wake region, the transversal wake profile has a potential core of uniform velocity
(𝑈CW ) and radius 𝑟′. Outside the wake core (e.g., 𝑟′ < 𝑟 < 𝑅W ), the transversal wake profile is defined
by a characteristic shear layer, which in turn defines the wake velocity deficit, as shown in Equation (17):
𝑈∞ (𝑚W − 1) 1.5 2
𝑈Def = ( ) (1 − (1 − 𝜂W ) ) (17)
𝑚W

where 𝜂W = (𝑅W − 𝑟)/(𝑅W − 𝑟′), so that 𝑅W − 𝑟′ is the width of the turbulent mixed zone in the
transversal wake profile. The shear layer expands outward (i.e., 𝑟 > 𝑅W ) and inward (i.e., 𝑟 < 𝑟′) from
the edge of the wake as it develops in the streamwise direction (𝑥 [m]). It is considered that the near
wake region terminates when the outer shear layer meet the centerline of the wake, so the center
velocity deficit starts to decrease with increasing distance from the WT. The downstream extent
(𝑥𝐶 [m]) of the near wake can be computed using Equation (18):
1
2 2 2 −
d𝑅W d𝑅W d𝑅W 2
𝑥𝐶 = 𝑟0 (( ) +( ) +( ) ) (18)
d𝑥 𝑎 d𝑥 𝑚 d𝑥 𝜆W

where (d𝑅W /d𝑥)𝑎 = 𝐼𝑎 /0.36, (d𝑅W /d𝑥)𝑚 = 0.27(0.214 + 0.144𝑚W )1/2 (𝑚W − 1)/(𝑚W + 1), and
(d𝑅W /d𝑥)𝜆W = (𝑐𝑇 𝜎L 𝑚W 𝜆3W /(6𝑚W + 6))1/2 are the wake growth contributions due to ambient
turbulence, due to the turbulence generated by the wake shear, and due to the mechanically added
turbulence from the WT rotor. Moreover, 𝜎L is the rotor solidity, 𝐼𝑎 [%] is the ambient turbulence
intensity, 𝑐𝑇 is the thrust (or drag) coefficient of the blade profile, and 𝜆W is the tip speed ratio.
The wake growth rate is assumed linear and given by d𝑅W /d𝑥 = (𝑟𝑐 − 𝑟0 )/𝑥𝐶 .
In the transition region, the velocity deficit is given by a linear combination of profiles at the
extremes of the transition region. At the beginning of the transition region, the radius of the wake is 𝑟𝑐
and at the end of the transition region is 𝑟𝐹 = 𝑟0 /(0.134 + 0.124𝑚W )1/2 . The extent of the transition
region can be calculated using Equation (19):
1 1
𝑥𝑐 (0.214 + 0.144)2 (1 − (0.134 + 0.124𝑚W )2 )
𝑥𝐹 = 1 1 (19)
((0.134 + 0.124𝑚W )2 ) (1 − (0.214 + 0.144)2 )
Energies 2014, 7 6956

The wake growth rate at the transition region is the same as in the near wake region, given by
d𝑅W /d𝑥 = (𝑟𝑐 − 𝑟0 )/𝑥𝑐 .
In the far wake, the velocity deficit is given by Equation (20):
1
𝑟 2 2
−0.258𝑚W + ((0.258𝑚W + 0.536(1 − 𝑚W ) (𝑅0 ) )
)2 𝑟 1.5
2
CM W
𝑈Def = 𝑈Def (1 − ( ) ) (20)
0.268(1 − 𝑚W ) 𝑅W
( )
CM
where 𝑈Def is the velocity deficit at the wake core. The wake growth rate in the far region is governed
only by the ambient turbulence intensity, resulting in d𝑅W /d𝑥 = 𝐼𝑎 /0.36. Thus the radius of the far
wake at any point downstream (𝑥 ≥ 𝑥𝐹 ) can be calculated as 𝑅W = 𝑟𝐹 + (𝐼𝑎 /0.36)(𝑥 − 𝑥𝐹 ).
The ground has the effect of reducing vertical downward wake growth, both by suppression of the
turbulence near the ground and by the direct presence of the ground. The effects of the ground were
modeled by classical reflection techniques (or imaging techniques) [281], in which an imaginary WT is
added below the ground (in a symmetrical fashion), so the wakes of both WTs expand beyond the
ground and then their velocity deficits are added, thus the momentum deficit of the original wake is
conserved. However, as pointed out by Crespo et al. [281], this approach is not completely valid as the
wake velocity deficit is not actually conserved because of the friction (viscous) effects at the surface.
To overcome this issue, Crespo et al. [281] considered an antisymmetric WT wake, so that velocity
deficits are subtracted and result in a zero perturbation at the ground. However, although this alternative
procedure eliminates the previously mentioned issue, it is not clear if it is valid throughout the wind
field, where surface effects might not be important. Finally, the tower wake was introduced by treating
the tower as a drag producing cylinder, which can be regarded as an approximately two-dimensional
problem and has appropriate two-dimensional growth rates (i.e., the wake of the tower expands only in
the plane parallel to the ground).
Another interesting wake model, categorized as a parabolic field model [59], was developed
by Ainslie [26] in 1988. The model considers that the wake is axisymmetric, fully turbulent, having nil
circumferential velocities, and stationary with time. In addition, the pressure gradients in the co-flowing
fluid outside the wake were assumed negligible. The model is based on the Reynolds-Averaged
Navier-Stokes equations, which are complemented with an eddy viscosity model as turbulence closure.
The eddy viscosity is predicted by a simple analytical form of the Prandtl’s free shear layer model and
includes contributions by the ambient turbulence. The set of equations were simplified by adopting the
thin shear layer approximation and by disregarding the viscous terms. The complete Ainslie far wake
model is shown in Equations (21) and (22):
∂𝑢̅̅̅𝑥 𝜕𝑢
̅̅̅𝑥 1 𝜕(𝑟𝑢̅̅̅̅̅̅
𝑥 𝑢𝑟 )
′ ′
𝑢𝑥
̅̅̅ + ̅̅̅
𝑢𝑟 =− ( ) (21)
𝜕𝑥 𝜕𝑟 𝑟 𝜕𝑟
𝜕𝑢̅̅̅𝑥
̅̅̅̅̅̅
−𝑢 ′ ′
𝑥 𝑢𝑟 = 𝜀
𝑆 (𝑥)
𝜀 = 𝑙W (𝑥)𝑈W + 𝜀𝑎 (22)
𝜕𝑟
where ̅̅̅
𝑢𝑥 [m/s] is the axial mean velocity, ̅̅̅𝑢𝑟 [m/s] is the radial mean velocity, x [m/s] is the axial
distance coordinate (downwind from the WT position), r [m] is the radial distance coordinate (measured
from the wake centerline), 𝑢̅̅̅̅̅̅ 2 2 2
𝑥 𝑢𝑟 [m /s ] is the Reynolds stress cross-correlation, 𝜀 [m /s] is the eddy
′ ′
Energies 2014, 7 6957

𝑆
viscosity, 𝑙W and 𝑈W are suitable length and velocity scales describing the wake shear layer, and
2
𝜀𝑎 [m /s] is the ambient turbulence contribution to the eddy viscosity.
The length and velocity scales were taken to be proportional to the wake width 𝑏W [m] and
𝐶 𝐶
the velocity difference 𝑈∞ − 𝑈W [m/s] across the wake shear layer (where 𝑈W is the centerline
wake velocity), respectively. 𝜀𝑎 is given by the atmospheric eddy diffusivity of momentum
𝜀𝑎 = 𝑘M = 𝑘𝑢∗ 𝑧/𝜙m (𝑧/𝐿), where k is the Von Karman constant (𝑘 = 0.4), 𝑢∗ [m/s] is the friction
velocity, z [m] is the height above the ground, L [m] is the Monin-Obukhov length, and 𝜙m (𝑧/𝐿) is a
similarity function reflecting the influence of the thermal stability condition on the turbulent mixing
process (e.g., 𝜙m = 1 for neutral conditions). The value of the friction velocity depends on the
atmospheric surface shear layer, which can be approximated with the use of the logarithmic laws.
For neutral conditions, the logarithmic law states that the axial wind speed (𝑢 ̅̅̅𝑥 [m/s]) at any height
within the atmospheric surface layer (e.g., up to 100 m) is governed by the roughness length (𝑧0 [m])
𝑢𝑥 = (𝑢∗ /𝑘) ln(𝑧/𝑧0 ).
and the friction velocity as given by ̅̅̅
A modification is required in order to use the eddy viscosity equation in the near wake region
(which length was taken to be approximately five rotor diameters), since a lack of equilibrium between
the mean velocity field and the turbulence field was observed. A filter function 𝐹𝐴 , shown in Equation (23),
was proposed to modulate the build-up of turbulence in the shear layer of the wake:
1
𝑥 − 4.5 3
𝐹𝐴 = {0.65 + ( ) 𝑥 < 5.5 (23)
23.32
1 𝑥 > 5.5
The resultant eddy viscosity closure, incorporating the previous approximation, results in Equation (24):
𝐶)
𝜀 = 𝐹[𝑘1 𝑏W (𝑈∞ − 𝑈W + 𝑘M ] (24)
where 𝑘1 is a constant, which is a property of the shear layer and largely independent of the ambient
turbulence and the WT characteristics. Ainslie conducted wind tunnel experiments for different inflow
conditions at low ambient turbulence intensities and found that the value k1 = 0.015 was adequate.
The solution of the Ainslie model should be calculated considering downstream distances of more
than two rotor diameters, where pressure gradients no longer dominate the flow. The wake transversal
profile was defined using a Gaussian profile of the form 1 − ̅̅̅/𝑈 𝑢𝑥 ∞ = 𝐷M exp(−3.56(𝑟/𝑏W )2 ) ,
where 𝐷M = 𝑐𝑇 − 0.05 − (16𝑐𝑇 − 0.5)𝐼𝑎 /1000, 𝐼𝑎 [%] is the ambient turbulence intensity, 𝑐𝑇 is the
thrust coefficient of the WT, and 𝑏W = (3.56𝑐𝑇 /(8𝐷M (1 − 0.5𝐷M )))1/2 .
To complement the wake model, Ainslie provided a simple model for the treatment of the wake
meandering [59,281], a phenomenon in which the wake meanders relative to an observer fixed on
the ground due to fluctuations in wind direction. The wake meandering results in a reduced centerline
velocity deficit measured by the fixed observer when compared with the centerline velocity deficit of a
stationary case, because the measurement point sweeps across a region of the wake profile during the
averaging period of the measurement. When meandering occurs (e.g., in neutral and unstable atmospheres),
the deficit measured by a fixed observed can be determined using Equation (25):
1
−2
𝜎𝜃 𝑥 2
̂
𝑑 = 𝑑0 (1 + 7.12 ( ) ) (25)
𝑏W
Energies 2014, 7 6958

where 𝑑0 is the uncorrected centerline velocity deficit and 𝜎𝜃 is the standard deviation of wind
direction fluctuations, measured over the same averaging time as the wake measurements (but not
including very short time timescale fluctuations). Finally, the Ainslie model do not account for ground
effects, tower effects, and variations in turbulent properties across the wake in its basic formulation.
Another relevant wake modeling approach was proposed by Larsen, who derived two similar wake
models in 1988 [25], which were updated by him later in 2009 [287]. The models are based on the
turbulent boundary equations (a simplified version of the Reynolds-Averaged Navier-Stokes equations),
from which closed-from analytical solutions were obtained for the wake width as well as for the mean
wind speed profile within the WT wake. Similarity assumptions and the Prandtl’s mixing length theory
for the description of the turbulent stresses were considered. Both models have different levels of
sophistication; the first model only takes into account the dominant terms of the turbulent boundary
equations, whereas the second model takes into account the full system of equations. The free-stream
airflow was considered turbulent, homogeneous, incompressible and stationary. Larsen simplified the
original set of parabolic differential equations, expressed in cylindrical coordinates, by assuming an
axisymmetric wake, by neglecting some terms (e.g., the pressure terms and distance terms of the form
𝑥 −7/3 ), by conducting analyses of order of magnitude (e.g., the thin shear approximation), and by
using similarity assumptions for the description of the wake profile evolution, resulting in the first order
model. The analytical solution to the first order model, together with the semi-analytical calibration of
the involved constants, is shown in Equations (26)–(28):
3 2
𝑈∞ 1 3 1 35 10 1
𝑢𝑥 (𝑥, 𝑟) = − (𝑐𝑇 𝐴𝑥 −2 )3 [𝑟 2 (3𝑐12 𝑐𝑇 𝐴𝑥)−2 − ( ) (3𝑐12 )−5 ] (26)
9 2π

3 2
𝑈∞ 1 5 3 1 35 10 1
𝑢𝑟 (𝑥, 𝑟) = (𝑐𝑇 𝐴)3 𝑥 −3 𝑟 [𝑟 2 (3𝑐12 𝑐𝑇 𝐴𝑥)−2 −( ) (3𝑐12 )−5 ] (27)
3 2π
1
35 5 1 1
(28)
𝑅W (𝑥) = ( ) (3𝑐12 )5 (𝑐𝑇 𝐴𝑥)3
2π
where 𝑢𝑥 [m/s] is the axial wind speed magnitude at some axial (x [m]) and radial (r [m]) position within
the wake, 𝑢𝑟 [m/s] is the radial wind speed magnitude at some axial and radial position within the wake, 𝑐𝑇
−5/6
is the thrust coefficient of the WT, A [m2] is the rotor swept area, 𝑐1 = (𝐷/2𝛼1 )5/2 𝑥0 , D [m] is the
−1
rotor diameter, 𝛼1 = (105/2π)1/5 (𝑐𝑇 𝐴)1/3 , 𝑥0 = ((𝐷/2𝛼1 )−3 (𝛽1 /(𝑈∞ − 𝑈𝑚 ))3/4 − 1) Δ𝑥 , 𝑈𝑚 =
−4/5
𝑈∞ (1 − 𝛽1 𝑐1 (𝑥0 + Δ𝑥)−2/3 ) is a measured axial velocity that depends on Δ𝑥 [m], which represents
a distance measured from the WT rotor, and 𝛽1 = (1/9)(3)−2/5 (35/2π)3/5 (𝑐𝑇 𝐴)1/3 . The second order
model of Larsen included the previously neglected distance terms of the form 𝑥 −7/3 and its analytical
solution, together with additional assumptions, resulted in a very large model that is not described here
for the sake of brevity.
The updated version of the Larsen model [287] included two major improvements: an approach to
merge wakes, which is necessary to quantify the performance of wind farms, and a proper calibration
of the model with full-scale experiments. The updated first order model is shown in Equations (29)
and (30):
Energies 2014, 7 6959

3 2
𝑈∞ 1 3 1 35 10 1
𝑢𝑥 (𝑥, 𝑟) = − (𝑐 𝐴(𝑥 + 𝑥0 )−2 )3 [𝑟 2 (3𝑐12 𝑐𝑇 𝐴(𝑥 + 𝑥0 ))−2 −( ) (3𝑐12 )−5 ] (29)
9 𝑇 2π
1
35 5 1 1
(30)
𝑅W (𝑥) = ( ) (3𝑐12 )5 (𝑐𝑇 𝐴(𝑥 + 𝑥0 ))3
2π
where 𝑐1 = (105/2𝜋)−1/2 (𝑑1 𝐷/2)5/2 (𝑐𝑇 𝐴𝑥0 )−5/6 , 𝑥0 = 9.6𝐷/((2𝑅9.6 /𝑑1 𝐷)3 − 1) ,
1/2
𝑑1 = ((1 + 1/√1 − 𝑐𝑇 )/2) , 𝑅9.6 is the wake radius at 9.6 rotor diameters downstream from the
WT, which was calculated empirically form the Vindeby offshore wind farm data and expressed by
𝑅9.6 = 𝑎1 exp(𝑎2 𝑐𝑇2 + 𝑎3 𝑐𝑇 + 𝑎4 )(𝑏1 𝐼𝑎 + 1)𝐷, where the 𝑎𝑖 coefficients are resumed in [287].
In order to calculate the effective wind speed of merged wakes in wind farms, Larsen considered
the calculation of the mean effective wind speed reaching the rotor of a WT, given by either
2
𝑈eff = 1/𝐴 ∫ 𝑈d𝐴 or 𝑈eff = 1/𝐴 ∫ 𝑈 2 d𝐴 , where the wind shear was assumed to be described by
the logarithmic laws. The effective wind speed within the wind farm was calculated using a linear
superposition of wakes, as shown in Equation (31):
𝑁𝑢
eff
𝑈W = 𝑈eff − ∑ 𝑢 𝑥𝑖 (31)
𝑖=1

where 𝑁𝑢 is the number of upstream WT. A second approach was suggested by using the effective
kinetic energy reaching the WTs, as shown in Equation (32):
2
1 𝑁𝑢
eff
𝑈W = √ ∫ (𝑈eff − ∑ 𝑢𝑥𝑖 ) d𝐴 (32)
𝐴 𝑖=1

Finally, like the Ainslie wake model, the Larsen models do not account for the ground effects or the
tower effects in its basic formulations.
Another wake modeling approach was suggested by Frandsen et al. in 2006 [283]. For the
description of single WT wakes, they considered expressions derived from the Betz-Lanchester theory,
which link the thrust and power coefficient of the WT to the flow speed deficit in the wake.
A cylindrical control volume with constant cross-sectional area equal to the wake area, with horizontal
axis parallel to the mean wind velocity vector, and with no flow across its surface was used to perform
a momentum balance between the free-stream flow and the WT wake, considering a simplified integral
version of the momentum equation, to determine the momentum deficit in the wake. The simplified
momentum equation disregarded pressure gradients (e.g., the control volume extends upwind,
0.5𝐷 < 𝑥 < 𝐷, and downwind, 2𝐷 < 𝑥 < 3𝐷, far enough for the pressure to become equal to the
free-stream pressure), the gravity force, viscous forces (e.g., shear forces in the cylinder surface were
neglected), rotational effects, and tower effects. The free-stream wind speed was considered constant,
homogeneous, stationary, and incompressible. The wake was considered axisymmetric and its transversal
profile constant (i.e., a top-hat profile was considered). The model defines the wake wind speed (𝑈W )
as shown in Equation (33):

𝑈W (𝑥) 1 1 𝐴
= ± √1 − 2 𝑐 (33)
𝑈∞ 2 2 𝐴W (𝑥) 𝑇
Energies 2014, 7 6960

where 𝐴 [m2] is the swept area of the WT rotor, 𝐴W (𝑥) is the area of the wake, which is assumed to be
a function of the downstream distance (x [m]), and “+” in Equation (33) applies when 0 ≤ 𝑐𝑇 ≤ 0.75
and “−” when 𝑐𝑇 > 0.75. The wake area after it initially expands (assumed to happen at the rotor
position, so 𝐴W (𝑥 = 0) = 𝐴𝐼W ) is given by 𝐴𝐼W = 𝛽𝐴, where 𝛽 = (1 + √1 − 𝑐𝑇 )/(2√1 − 𝑐𝑇 ).
The evolution of the wake diameter downstream the WT was defined as shown in Equation (34):
𝑘𝑠 1/𝑘𝑠
𝐷W (𝑥) = (𝛽 2 + 𝛼𝑠 𝑠) 𝐷 (34)
𝑘𝑠 𝑘𝑠
where 𝛼𝑠 = 𝛽 2 ((1 + 2𝛼𝑛𝑜𝑗 𝑠) − 1) 𝑠 −1 is a wake decay factor, 𝛼𝑛𝑜𝑗 ≈ 0.05, 𝑠 = 𝑥/𝐷, 𝑘𝑠 = 2, and
𝐼
the initial wake initial wake diameter is 𝐷W (𝑥 = 0) = 𝐷W = √𝛽𝐷. Nonetheless, 𝛼𝑠 must be determined
depending on the study case; for small 𝑐𝑇 and large 𝑠, the decay factor 𝛼𝑠 is of the order 10 𝛼𝑛𝑜𝑗 .
The model was used to predict the performance of regular arrays of WTs. Three different scenarios
were considered; in the first scenario, a row of WTs, with each WT exposed to multiple wakes, was
studied and an analytical link between the wake expansion and the asymptotic flow speed deficit was
derived. The wake-affected wind speed reaching a WT was derived from the momentum balance
within the control volume, resulting in the expression given in Equation (35):
𝑛+1 𝑛+1 2
𝑈W 𝑈W 𝐴𝑛W 𝑈W𝑛 𝑛
𝑈W 1 𝐴 𝑈W 𝑛
(1 − ) = 𝑛+1 ( ) (1 − )+ ( ) 𝑐𝑇 (35)
𝑈∞ 𝑈∞ 𝐴W 𝑈∞ 𝑈∞ 2 𝐴𝑛+1
W 𝑈∞
𝑛+1
where 𝑈W is the wake speed reaching a WT 𝑛 + 1, 𝐴𝑛+1 W is the wake area at the position of the WT
𝑛 𝑛 (𝑠) 𝑛
𝑛 + 1, 𝐴W = 𝐴W = 𝐴W (𝑛𝑠𝑟 ), 𝑠𝑟 = 𝑥𝑟 /𝐷, and 𝑥𝑟 is a constant separation distance among the WTs.
The effects from boundaries, such as the ground, were included implicitly through the wake area
growth d𝐴𝑛W = 𝐴𝑛+1 𝑛
W − 𝐴W . For an infinitely large wind farm, it was assumed that there is an
𝑛 𝑛+1
asymptotic, non-zero wake flow speed, thus (𝑈W /𝑈∞ ) − (𝑈W /𝑈∞ ) ≈ 0 when 𝑛 → ∞. In such cases,
the wake speed and the wake area are given by Equations (36) and (37):
𝑛 𝑛 2
𝑈W 𝑈W 𝐴𝑛W 𝑈W𝑛 𝑛
𝑈W 1 𝐴 𝑈W 𝑛
(1 − ) = 𝑛+1 ( ) (1 − )+ ( ) 𝑐𝑇 (36)
𝑈∞ 𝑈∞ 𝐴W 𝑈∞ 𝑈∞ 2 𝐴𝑛+1
W 𝑈∞
𝑈𝑛
1 ( 𝑈W )
∞
𝐴𝑛+1 𝑛
W − 𝐴W = 𝐴 𝑐𝑇 (37)
2 𝑈𝑛
(1 − 𝑈W )
∞

The right hand side of Equation (37) is a constant value, thus the wake cross-sectional area expands
linearly with increasing downstream distance. The wake decay factor when 𝑛 → ∞ is given by
𝑛 𝑛
𝛼𝑠 = 𝑐𝑇 (𝑈W /𝑈∞ )/(2𝑠𝑟 (1 − (𝑈W /𝑈∞ ))).
In the second scenario, the wakes from neighboring rows of WTs merge and the resultant wake
can only expand vertically after that point. Since the area must, asymptotically, expand linearly
(e.g., Equation (37)), the height of the wake must increase linearly. Therefore, the growth (h) of the
internal boundary layer in the limit for 𝑛 → ∞ is given by Equation (38):
𝑈
( 𝑈W ) π𝑐𝑇
∞
ℎ= ( ) (𝑥 − 𝑥0 ) + ℎ0 (38)
𝑈W 8𝑠𝑟 𝑠𝑓
(1 − )
𝑈∞
Energies 2014, 7 6961

where 𝑠𝑓 is the dimensionless distance between neighboring rows, 𝑥0 and ℎ0 are the axial distance and
wake height when the neighboring wakes start to merge.
The third scenario is when the wind farm is infinitely large and the wake-affected flow is in balance
with the atmospheric boundary layer. The interested reader is referred to [283,298] for a complete
description of the model, which links the surface wind speed to the geostrophic wind speed by defining
two flow layers divided at the WT hub height and by introducing the geostrophic drag law.
The most widely used EWM is the Jensen-Katic model [21,23,282], as shown in Figure 3.
The Jensen-Katic wake model defines the velocity deficit as shown in Equation (39):
1 − √1 − 𝑐 𝑇
𝑈Def (𝐶𝑇 , 𝑥) = (39)
𝐼𝑎 𝑥 2
(1 + )
𝐷
where 𝐼a [%] is the ambient turbulence intensity, representing the effect of turbulent forces on the WT
wake recovery and wake expansion, 𝑐𝑇 is the thrust coefficient, and D [m] is the WT rotor diameter. 𝐼𝑎
is sometimes written as 𝐼𝑎 = 2𝛼, where 𝛼 denotes the momentum entrainment constant (or wake decay
coefficient) modulating both the wind speed recovery and the wake expansion, as first proposed
in [23,298] and as shown in Equation (40):
1
𝐼𝑎 = 2𝛼 = 𝑧
ln ( ) (40)
𝑧0
where 𝑧 is the height above ground [m], and 𝑧0 is the roughness length [m], representing the height
above ground level at which a zero wind speed is found. Equation (40) implies that the wake recovery
and expansion is lower with increasing height, since lower ambient turbulence intensities are found far
from the surface. The wake expansion is commonly assumed to be proportional to the downwind
distance (x [m]) from the WT hub such that 𝑅W = 𝛼𝑥 + 𝑅, where 𝑅W [m] is the radius of the wake and
𝑅 [m] is the rotor radius.

Figure 3. Timeline of Wake Models used in the WFDO Problem.

Timeline of Wake Models
30
NA
Custom Function
25 Vermeulen
Lissaman
Katic-Jensen
20 Frandsen
No. of Papers

Larsen
CFD-Based
15

10

0
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Year
Energies 2014, 7 6962

Like Frandsen’s wake model, the Jensen-Katic wake model disregards pressure gradients, the gravity
force, viscous forces (e.g., shear forces in the edge of the wake were neglected), rotational effects,
ground effects, and tower effects. The free-stream wind speed was considered constant, homogeneous,
stationary, and incompressible. Moreover, the wake was considered axisymmetric and its transversal
profile constant (i.e., a top-hat profile was considered). The Jensen-Katic wake model relies on
momentum conservation in a fully developed wake, thus its validity is limited to the far wake region.
However, conflicting views about which is the axial length at which the wake can be considered to be in
the far region arise in the literature. It is often considered that the far wake starts at an axial distance of
about three to four rotor diameters downstream from the WT. A formal definition for such a distance,
however, should be based on the necessary distance required for the reestablishment of the axial and
radial pressure field, turbulent kinetic energy and angular momentum values, which in turn are a
function of the WT operation mode, as previously described.
It has been observed that the total velocity deficit, composed of several merged wakes, depends
mostly on the closest WT [59,299,300]. However, the interaction of multiple wakes is not fully
understood and is subject of many studies in the fluid mechanics field. Different methods for calculating
the total effect of merged WT wakes have been proposed in the literature [23,269,300]. The wake
merging process is based on the concept of constructive superposition, where several wakes can be
combined to produce a stronger wake. One effective and recurring way to calculate the effective value
eff
of the velocity deficit (𝑈Def ) in merged wakes is to sum the kinetic energy deficits of all WTs
contributing to the wake (i.e., the sum of the square of the velocity deficits) as shown in Equation (41):
2
𝑁𝑢 𝑈 𝑖 (𝑥, 𝑟)
eff
𝑈Def (𝑥, 𝑟) = √∑ (1 − W ) (41)
𝑖=1 𝑈∞
𝑖 (𝑥,
where 𝑈W 𝑟) is the velocity deficit contribution of the upstream WT i, and 𝑁𝑢 is the number of
upstream WTs. The effect on the energy conversion of a WT operating under partial wake conditions
is commonly modeled through the use of a geometric factor [156,178,301] that affects the value of the
effective wind speed deficit (i.e., Equations (39) and (41)). The geometric factor is commonly defined
as the ratio of the wake affected rotor area to the WT rotor area (𝐴Overlap /𝐴 ). Therefore, the modeled
partial wake operation is less harmful, in terms of power conversion, than full wake operation. The
assumption, however, does not account for the increased fatigue loading the WT will undergo while
operating in partial wake conditions, which is important as denoted by the experimental findings of
Seifert et al. [69]. Moreover, the tower and nacelle contributions to the wake are usually not considered
explicitly by EWMs. Nevertheless, its effects can be included indirectly through the use of empirical
values of 𝑐𝑇 , which may contain contributions from all passive and active parts of the WT.
New semi-empirical approximations that complement the Jensen-Katic wake model have been
proposed to describe: (1) the wake expansion and recovery as a function of the net TI (composed of
both the ambient and the WT mechanically added turbulence intensities); (2) the effects of yaw
misalignment and unsteady conditions in wake development; and (3) the effects of operating WTs
under partial wake conditions [282].
Energies 2014, 7 6963

3.2.2. Structural Fatigue and Degradation Modeling

Most of the published works on the WFDO problem have ignored both the environmental and
the mechanical factors that produce unfavorable effects on the energy conversion efficiency such as
the WT aero-structural loading and structural fatigue degradation (which is commonly related to
O&M costs). Although the effects of TI and unsteady aerodynamics on structural fatigue loading have
been observed and have been the subject of several studies [59,290,302], a general model to account
for its overall effects has not yet been developed and validated. The reason is probably related to the
limited state of knowledge on their effects and the impossibility of describing them through simple
mathematical expressions.
The only existing approach to model structural fatigue and degradation impacts in the wind farm
financial balance was proposed by Larsen [166,279]. Larsen’s model rely on the accurate estimation of
the mean accumulated equivalent fatigue load (𝐿𝑒𝑞 [Pa]) for each WT in the wind farm and for each
of the WT structural members, which can be calculated, as shown in Equation (42), by summing all
possible contributions from individual load cases and while taking into account the probability of
occurrence of each load case and the total number of equivalent load cycles the WT will undergo
during its expected operational lifetime:
1
1 𝑛𝑊𝐷 𝑛𝑊𝑆 𝑚 𝑚
𝐿𝑒𝑞 = [ ∑ ∑ 𝑅𝑒𝑞 (𝑈𝑖,𝑗 ) 𝑛𝑒𝑞 𝑝(𝑈𝑖,𝑗 )𝑛 𝑇 Δ𝑛𝑊𝐷 Δ𝑛𝑊𝑆 ] (42)
𝑛𝑒𝑞,𝐿 𝑗=1 𝑖=1

where 𝑅𝑒𝑞 [Pa] is the equivalent load for the load case [279], defined by the mean wind speed i in the
direction j (𝑈𝑖,𝑗 ), 𝑚 is the Wöhler exponent [29] for the WT member considered, 𝑛𝑒𝑞 is the number of
600 s cycles corresponding to 𝑅𝑒𝑞 (𝑈). 𝑝(𝑈) is the probability for the load case, 𝑛 𝑇 is the number of
10-min cycles corresponding to the wind farm expected lifetime, 𝑛𝑒𝑞,𝐿 is the lifetime equivalent
number of cycles. 𝑛𝑊𝐷 and 𝑛𝑊𝑆 are defined as the number of discretized wind directions and wind
speeds during the calculation. Δ𝑛𝑊𝐷 and Δ𝑛𝑊𝑆 are defined as wind direction and wind speed differentials.
As suggested by Réthoréet al. [166], a simplified approach can be implemented, in which an extensive
database of generic individual load cases is pre-calculated using aero-elastic simulations [303] and
subsequently used during the WFDO process. The database should encompass several different
situations in which the WT under consideration operate under the wake of another WT (e.g., different
free-stream inflow conditions, turbulence intensities, WT spacings, etc.). The computed structural
degradation can be used during the calculation of the expected O&M and fatigue degradation costs, as
discussed in Section 3.2.6.

3.2.3. Civil Engineering and Electrical Infrastructure

The layout of the civil engineering infrastructure of a wind farm can be considered as an embedded
problem, where the main objectives are: (1) minimizing the distance, or costs, of the roads accessing
the WTs and (2) minimize the costs associated to WT foundations. The wind farm road design problem
has been treated previously, within the scope of WFDO problem, by translating the problem into a
minimum spanning tree problem [135,278] and into a Steiner tree problem [167]. The assumptions in
both approaches were that the wind farm was planar (e.g., flat terrain), and that the cost per kilometer
Energies 2014, 7 6964

of road was constant and independent of the WT positions. The Steiner tree problem implementation
allowed to find shorter road paths involving nodes (Steiner points), that were not necessarily active
WT locations, when compared with the output of the minimum spanning tree approach. However,
the Steiner tree problem is significantly more difficult to solve than the minimum spanning tree
problem, requiring specialized solution procedures [167]. Although several algorithms have been
developed to solve the minimum spanning tree problem, the most widely used are the Prim–Jarník and
the Kruskal greedy algorithms [304,305].
The problem of deciding the kind of WT foundation to use is significantly more complex for
offshore wind projects than for onshore wind projects in both economic and technologic terms. For
offshore cases, the problem has been treated by Elkinton et al. [55,234], within the OWFLO framework,
Réthoré et al. [166], within the TOPFARM framework [306], and by González et al. [307–310].
In all cases, cost functions that depend on the WT individual locations, as discussed in Section 3.2.6,
were developed to model cost of foundations due to seabed depth. The most robust approach was
proposed by Elkinton [55], who considered the selection of the kind of tower, type of sub-structure and
type of foundation during the optimization process while accounting for wind and wave loads and
validated cost functions obtained from different real offshore wind farms. A detailed description of the
commonly used offshore foundation technologies and its costs can be found in [55,83]. For onshore
cases, the problem has been treated by Fagerfjäll [167], and González et al. [135,278,280]. Both
approaches assumed that foundation costs are an exclusive function of the soil bearing capacity.
In addition to the civil engineering infrastructure, the wind farm electrical grid layout, which is
mainly composed of substations and power lines of different types and capacities, must be designed
with the aim of minimizing the overall costs, as mentioned in Section 2.2.2. The electrical layout
design is significantly different between offshore wind farms and onshore wind farms. For offshore
cases, the problem has been treated by: (1) Elkinton [55], who considered a validated model, based on
transmission line theory, to quantity the HVAC transmission cable losses and the MVAC collection
cable losses and while accounting for its finite capacity; (2) by González et al. [309,310], who proposed
a capacitated internal collection network design based on a minimum spanning tree procedure while
accounting for Joule-based power losses; (3) by Réthoré et al. [166], who proposed a deterministic
heuristic procedure based on clustering of non-capacitated power lines, which is similar to the
minimum spanning tree approach; (4) by Nandigam and Dhali [84], who developed models accounting
for the cost, loss and reliability of the transmission and collection power lines and proposed an
optimization strategy for the electrical layout design based on geometric programing; and (5) by Bauer
and Lysgaard [168], who proposed a vehicle route problem approach, which minimized the total cable
cost subject to capacity and physical constraints (e.g., cables did not cross each other in the final optimized
electric implementation). For onshore cases, the problem has been treated by: (1) Fagerfjäll [167], who
proposed a Mixed-Integer Non-Linear Programming (MINLP) model for capacitated cables based on a
Steiner tree formulation; (2) by González et al. [278], who restricted the capacity of the MVAC
collection lines, the HVAC transmission lines and the substations while considering a minimum
spanning tree approach; and (3) by Tran et al. [163], who proposed a multi-objective optimization
procedure in which the electrical layout was designed through a minimum spanning approach while
considering non-capacitated power lines. The best power line losses model was provided by Elkinton [55],
as it accounts for most of the involved physics of electricity transmission, while the best optimization
Energies 2014, 7 6965

procedure was provided by Fagerfjäll [167] by implementing a Steiner tree approach. In all the
reviewed works, the common assumptions were that the wind farm site is flat, so the transmission and
collection power lines length can be computed by summing the distances between WTs based on the
resultant electrical layout of the minimum spanning tree or Steiner tree implementation. In addition,
it was assumed that no additional accessories/technologies are required to introduce the power lines in
the environment (sea cables, underground cabling, etc.), and that no maintenance is required during the
wind farm operational lifetime.

3.2.4. Environmental Impacts

Most of the reviewed works did not include in their optimization strategies any of the environmental
adverse effects described in Section 2. However, different approaches to model such adverse effects
were found in the literature and are summarized next.

Visual Impacts (VI)

None of the reviewed works incorporated VI in their optimization approaches. However, models
and methods for visibility analysis are presented in [109,110]. A conceptual model for assessing visual
effects can be described as a procedure that calculates shadow flicker for both sun and moon induced
exposure on the entire neighboring of the wind farm based on the actual location of the WTs and all
other factors as described in Section 2.2.3. Conservative assumptions might be used to ensure maximum
protection against adverse human health effects and minimize risks. However, a standardized shadow
flicker dose-response has not yet defined, primarily due to the lack of experimental results and published
studies, and thus remains an important topic of research.

Electromagnetic Interference (EMI)

In none of the reviewed works the wind farm EMI effects were considered. However, other
literature works have reported different models and methods for calculating the EMI effects produced
by wind farms [36,311]. Most approaches assume that signal scattering is the most important source of
EMI from WTs [36,312], among the factors described in Section 2.2.3. The common procedure to
determine the overall EMI effects is to calculate all the possible attenuations of original signal and
compare the resultant signal with the original signal. The amplitude of a useful electromagnetic signal
reaching a receiver (𝐶𝑆 ) can be approximated using Equation (43):
𝐶𝑆 = 𝑃T − 𝐴TR + 𝐺TR (43)
where 𝑃T [dB] is the transmitter power, 𝐴TR [dB] is the attenuation between the transmitter and
receiver, and 𝐺TR [dB] is the receiver antenna gain in the direction of transmitted signal. A WT
interfering signal, (𝐼𝑆 ) can be calculated using Equation (44):
4π𝜎𝑆
𝐼𝑆 = 𝑃T − 𝐴TW + 10 log10 ( ) − 𝐴WR + 𝐺WR (44)
𝜆2
where 𝐴TW [dB] is the attenuation between the transmitter and the WT, 𝐴WR [dB] is the attenuation
between the WT and the receiver, 𝐺WR [dB] is the receiver antenna signal gain in the direction of the
Energies 2014, 7 6966

reflected signal, 10 log10 (4π𝜎𝑆 /𝜆2 ) [dB] is the contribution to the signal scattering by the WT,
𝜎𝑆 [m2] is commonly known as the radar cross section (and can be considered a function of the WT
geometry, WT dielectric properties and the incident signal wave length) and 𝜆 [m] is the signal
wavelength. The difference between the useful and the interfering signals is:
4π𝜎𝑆
𝐶𝑆 − 𝐼𝑆 = 𝐴TW − 10 log10 ( ) + 𝐴WR − 𝐴TR + 𝐺TR – 𝐺WR (45)
𝜆2
If a forward scattering case is considered, where a WT is between transmitters and receivers, and
if the distance between the transmitter and receiver (𝑑TR ) is much larger than the distance between
the WT and the receiver (𝑑WR ) then 𝐴TW ≈ 𝐴TR . Furthermore, the free space loss is commonly
approximated as 𝐴WR = 20 log10 (4π𝑑WR /𝜆), resulting in:
𝐶𝑆 − 𝐼𝑆 = 10 log10 (4π) + 20 log10 (𝑑WR ) − 10 log10 (𝜎𝑆 ) + 𝐺TR − 𝐺WR (46)
Equation (46) suggests that the difference between the useful signal and the interference signal may
be improved by: (1) increasing the distance between WT and receivers; (2) reducing the radar cross
section (𝜎𝑆 ); (3) enhancing the signal gain between the transmitter and the receiver; and (4) reducing
the signal gain between the WT and the receiver. Some of the suggested improvements, however, have
strong impacts in the WT energy conversion (e.g., separating WT from key locations) and induce the
need to spend more capital costs for mitigation (e.g., by incorporating amplifiers, re-transmitters or
additional transmitting antennas). Equation (46) can be rearranged to define a forbidden zone, in terms
of the distance 𝐹𝑧, within which a WT may not be located if an adequate interference level (𝐼𝑆,𝑅𝑒𝑞 ) is
to be maintained:
𝐹𝑧 = 20 log10 (𝑑TR ) = (𝐶𝑆 − 𝐼𝑆,𝑅𝑒𝑞 ) + 10 log10 (𝜎𝑆 ) − 𝐺TR + 𝐺WR − 11 (47)
The term 𝐹𝑧 is highly dependent of the radar cross section (𝜎𝑆 ), which, unfortunately, is not easy to
calculate and is dependent on the direction the signal is transmitted. Different approaches in literature
have suggested how to calculate 𝜎𝑆 [36,129,312–314], where the method of moments is applied to
calculate the EMI contributions of both the WT tower and the WT rotor, assumed to be composed of
perfect electric conductor materials (which, however, is not the case in practice). It is suggested that, in
the back-scatter region, reflection is caused only by the metallic parts of the blades, thus the blade
dimensions are used to calculate the radar cross section and typically results in a smaller 𝐹𝑧, in contrast
with the forward-scattering 𝐹𝑧 in which the whole WT is considered to contribute to the electromagnetic
signal modulation. A common accepted relation for 𝜎𝑆 [dBm2] was proposed by Hall [314], as shown
in Equation (48):
𝑓(𝜃)
10 log10 (𝜎𝑆 ) = 20 log10 (𝐴𝐵 ) + 11 − 𝐶 (48)
𝜆

where 𝐴𝐵 [m2] is the superficial area of one blade, 𝑓(𝜃) is a function describing the amplitude of the
scattered signal in the direction 𝜃, and 𝐶 is a calibration constant with an approximated value of 15 dB.
Other complex approaches have been proposed [36,311], however, significant further work is required
to develop more precise techniques to calculate the radar cross section area and the overall EMI effects
of WTs and wind farms. In particular, techniques to determine the effects of irregular terrain and the
effects of the blade shape and structural materials have yet to be developed and validated.
Energies 2014, 7 6967

Noise

Only two works incorporated noise concerns into the WFDO problem. The first work, developed by
Fagerfjäll [167], considered noise as a constraint with a certain maximum threshold value in the
optimization framework. A noticeable improvement was introduced by Kwong et al. [199,202] who
implemented a multi-objective strategy, based on the Non-Domination Sorting Genetic Algorithm
(NSGA-II), in which the total noise at the boundaries of the wind farm was minimized while the
AEPFarm was maximized. Not unexpectedly, a comparison between the results of single-objective
optimizations of noise and energy (i.e., when considered these objectives individually) showed that the
WT layouts and the inter-WT distances distributions were different. The noise calculation was based
on the ISO-9613-2 standard [315], which proposes a general model for isotropic noise calculation.
According to the ISO-9613-2, the sound pressure level (LP) of an isotropic sound source can be
calculated as follows:
𝐿P = 𝐿W + 𝐷C − 𝐴𝑓 (49)
where 𝐿W is the octave-band sound power emitted by the source, 𝐷C is the directivity correction for
sources that are not omni-directional, 𝐴𝑓 is the octave-band attenuation, which is composed of
different attenuation effects (e.g., geometrical divergence, atmospheric absorption, ground effects, and
sound barrier, among others [315]). Because of the acoustic response of the human ear, a logarithmic
scale is often used based on reference levels that correspond to the limit of human hearing. Therefore,
it is common to weight the noise measurements (e.g., tone control) to reflect the response of the human
ear with frequency. In practice, A-weighting noise calculation is performed to measure hearing risk
annoyance based on a daily noise dose, which is in compliance with OSHA and MSHA regulations
that specify permissible noise exposures. The equivalent continuous A-weighted downwind sound
pressure level (or average sound pressure level, 𝐿PAvg ) at a specific location can be calculated from the
summation of the contributions of each point sound source at each octave band (𝑓) as shown in
Equation (50):
𝑛𝑆 8
𝐿PAvg = 10 log (∑ ∑ 100.1[𝐿P(𝑖,𝑗)+𝐴𝑓(𝑗)] ) (50)
𝑖=1 𝑗=1

where 𝑛𝑠 is the number of sound sources, 𝑗 is the index representing one of the eight standard
octave-band mind-band frequencies from 63 Hz to 8 kHz, and 𝐴𝑓 (𝑗) are the standard A-weighting
coefficients [315].
More complex noise models were found in the literature [36,124,125,316]. In such models,
two different measures are used to describe WT noise: the sound power level of the source (𝐿W [dB])
and the sound pressure level (𝐿P [dB]) at the location. A noise source is described in terms of its sound
power level as shown in Equation (51):
𝑊
𝐿W = 10 log10 ( ) (51)
𝑊0
where 𝑊 [W] is the total sound power level emitted from the source and 𝑊0 is a reference value of
10−12 W. The sound pressure level (𝐿P ) is calculated by replacing 𝑊 and 𝑊0 by 𝑃 [Pa], the RMS
Energies 2014, 7 6968

value of the sound pressure, and 𝑃0 = 2 × 10−5 Pa, a reference pressure value, respectively. The total
sound pressure (𝐿P,T ) of 𝑛𝑠 sound pressure levels can be quantified using Equation (52):
𝑠=𝑛𝑠
𝐿P,T = 10 log10 ∑ 10𝐿P(𝑠)/10 (52)
𝑠=1

Equation (52) implies that by adding two sound pressure levels of the same magnitude an increase
of 3 dB can be obtained over the base level. An equivalent sound pressure level 𝐿eq,T is the value of a
continuous steady sound that, within the specified time interval (𝑇), has the same mean square sound
pressure level as the sound under consideration, which varies with time [124,125]:
1 𝑇 𝑃2
𝐿eq,T = 10 log10 ( ∫ 2 d𝑡) (53)
𝑇 0 𝑃0
A similar calculation can be performed using A-weighted values to give 𝐿Aeq,T :
1 𝑇 𝑃𝐴2
𝐿Aeq,T = 10 log10 ( ∫ 2 d𝑡) (54)
𝑇 0 𝑃0
where 𝐿Aeq,T is the equivalent continuous A-weighted sound pressure level determined over time 𝑇
and 𝑃A is the instantaneous A-weighted sound pressure. The sound intensity (𝐼 = 𝑊/𝐴) is the sound
power transmitted through and unit area, A. The sound intensity far from the source, assuming a
uniform noise flux, can be approximated as 𝐼 = 𝑃2 /𝑍0 , where 𝑃 is the RMS value of the sound
pressure level, and 𝑍0 is the characteristic acoustic impedance. The sound pressure level may be
expressed in terms of sound intensity as shown in Equation (55):
𝐼
𝐿P = 10 log10 ( ) (55)
𝐼ref
where 𝐼ref is a reference intensity (e.g., 10−12 W/m2). In addition, the sound may spread in an isotropic
or anisotropic way. In the former case, a spherical or a hemispherical spreading can be assumed. For a
spherical spreading, the intensity will decrease with distance from the source of sound as
𝐼 = 𝑊/4π𝑟𝑠2 , where 𝑟𝑠 is the radial distance from the source. Hence, under conditions of ideal
spherical spreading, the sound pressure level can be approximated as shown in Equation (56):
𝑊
𝐿P = 10 log10 ( ) = 𝐿W − 10 log10 (4π𝑟𝑠2 ) (56)
4π𝑟𝑠2 10−12
Equation (56) implies that for each doubling of distance the sound pressure level is reduced by
6 dB. In addition, it implies that the terrain surface and all the nearby objects will act as perfect noise
absorbers, otherwise sound reflections may occur and the sound pressure level may not decay with the
same rate. Equation (56) can be rewritten to consider hemispherical spreading (with 𝐼 = 𝑊/2π𝑟𝑠2 ),
which in turn implies that the terrain surface and all nearby objects will act as perfect noise reflectors.
Hence, the determination of the reflecting/absorbing properties of the terrain and the nearby objects is
important during noise intensity computations.
The International Energy Agency (IEA) [124,125] provides a method to estimate the sound pressure
level of a single WT or a group of WTs, assuming that they are located in flat and open terrain. The
calculation is based on hemispherical spreading and considers a correction for atmospheric absorption:
𝐿P (𝑟𝑠 ) = 𝐿W − 10 log10 (2π𝑟𝑠2 ) − Δ𝐿𝑎 (57)
Energies 2014, 7 6969

where Δ𝐿𝑎 is the atmospheric absorption and can be calculated from Δ𝐿𝑎 = 𝑟𝑠 𝛼𝑠 , where 𝛼𝑠 is a
coefficient for sound absorption for each octave band, and 𝑟𝑠 is the Euclidean distance between the WT
hub and the key location (e.g., a household). An alternative approach is to consider the value
αs = 0.005 dB/m [317] and 𝐿W specified as a single, broadband sound power level. If there are several
WTs which influence the sound pressure level, the individual contributions are calculated separately
and summed, as previously shown in Equation (52). The IEA method ignores any effects of
meteorological gradients [36,316] (e.g., the sound waves can be curved by density gradients, which are
a function of temperature, pressure and relative humidity). The IEA study proposes that a rather
straightforward spherical propagation model should be assumed to quantify the total noise of wind
farms, with additional terms included to account for directivity, air absorption and topographical effects.
More noise models of increased complexity, which account for additional noise mechanisms linked
to the WT rotor geometry and turbulent inflow conditions in quasi-steady states (e.g., trailing edge
noise, tip noise, stall effects, blunt trailing edge noise, surface imperfections, etc.), were found in the
literature [117]. However, these models are not yet implemented in the WFDO problem, since they are
currently being improved and validated.

3.2.5. Uncertainty and Risk Modeling

The works presented by González et al. [135,310,318] were the first to introduce the concept of
uncertainty and risk within the scope of the WFDO problem. González et al. pointed out that energy
future prices, future variable costs, discount rate (e.g., accounting the time value of money) and wind
resource estimation are the main factors contributing to the overall uncertainty in the profitability of
wind farm projects. González et al. studied the effects of the uncertainty on the wind resource (i.e., the
wind direction and wind speed). They developed a probabilistic optimization methodology, which
attempted to mimic the effects of both short and long-term variations in wind resource, where a set of
possible scenarios and their probability of occurrence were taken into account during the calculation of
the project expected NPV, as shown in Equation (58):
𝑚
E[NPV] = ∑ NPV𝑞 𝑝𝑞 (58)
𝑞=1

where E[NPV] is the expected net present value of the wind farm, and 𝑝𝑞 denotes the probability of
occurrence of the scenario 𝑞. An additional criterion was proposed, incorporating concepts of utility
theory (UT), in which a utility function, as shown in Equation (59), was used to express the level of
risk the wind farm designer is willing to assume given the different possible scenarios:
NPV(𝑥𝐹 )−NPVmin
−
1−e 𝜌 𝑅

NPVmax −NPVmin , 𝑖𝑓 𝜌R ≠ ∞
−
u(NPV(𝑥𝐹 )) = 1−e 𝜌
𝑅 (59)
NPV(𝑥𝐹 ) − NPVmin
, 𝑖𝑓 𝜌R → ∞
{ NPVmax − NPVmin
where 𝑢 is the utility value corresponding to NPV(𝑥𝐹 ), 𝑥𝐹 is the wind farm configuration, and 𝜌R is the
parameter of risk tolerance, which allows modeling the risk attitude of the wind farm designer
(e.g., 𝜌R < 0 increases risk level). NPVmin is the minimum value of NPV that the wind farm designer
Energies 2014, 7 6970

would be willing to accept to undertake the project. Therefore, the expected value of the utility function
can be calculated using Equation (60):
𝑚
E[u(NPV)] = ∑ u(NPV𝑞 )𝑝𝑞 (60)
𝑞=1

González et al. [135] concluded that by increasing the risk attitude, the optimized wind farm layouts
are more sensitive to changes in wind resource, however, the expected profitability increases. In
addition, they concluded that the risk analysis (considering conservative risk attitude levels) led to wind
farm configurations that are less sensitive to uncertainty in contrast with common deterministic approaches.
Another interesting approach was provided by Afanasyeva et al. [134], who considered the effects
of the uncertainty on the WT power curve and the uncertainty on the wind resource (wind speed
and direction) during the calculation of the AEPFarm . The uncertainty in the AEPFarm was translated
into uncertainty in the NPV of the project, which was the maximized objective metric during the
optimization procedure. They employed an embedded Monte Carlo simulation method to obtain the
probability distribution of the NPV of the project (i.e., the approach maximized the NPV for a chosen
certainty level). The input variables were given as probability density functions instead of expected
values, which were assumed uncorrelated, to the objective function during the optimization process.
The standard uncertainty of the experimental WT electrical power curve was assumed to follow a
normal distribution along the discretized wind speed bins. In addition, the discretized wind direction
was assumed to change in the range (−45°, 45°) with a uniform probability distribution to emulate its
stochastic nature. The wind speed probability density function was assumed lognormal. The approach
maximized the NPV probability distribution of the project. However, one of the shortcomings of the
proposed approach is that the model requires a large number of Monte Carlo-based samples to evaluate
the objective function, which in turn is evaluated several times during the optimization process, thus
being a very time consuming approach.
A major contribution, including three different sources of uncertainty within the scope of the
WFDO problem, was proposed by Chen and MacDonald [54,138] in which uncertainties in land
availability due to landowner participation, wind shear, and economies of scale costs-reduction were
studied. Landowner participation was represented with an uncertain utility model of willingness-to-accept
monetary compensation (i.e., a minimum annual compensation per MW installed in their lands).
An uncertain wind shear parameter, based on a wind shear power law, and a cost model incorporating
uncertainty with an economies-of-scale cost reduction parameter (representing a cost reduction for
purchasing multiple WT) were introduced and developed. The goal of the work was to propose a
comprehensive framework, in which both the normalized mean and normalized standard deviation of
the CoE were minimized during the WFDO process. The study considered a Latin Hypercube method
for the quantification of the uncertainty propagation during the different stages of the CoE calculation
while considering different scenarios with different probabilities of occurrence, as equivalently
proposed by González et al. [135]. The results of the work demonstrated that even in uncertain
environments, wind farm designers can predict the viability of the project with an estimated CoE while
providing measures of CoE deviations.
Models and methods describing and quantifying additional sources of uncertainty are presented
in [133,137]. However, several sources of uncertainty, as pointed out by Elkinton [55], are still not
Energies 2014, 7 6971

studied/considered: technical availability, Rotor Nacelle Assembly (RNA) cost, foundations cost, HVAC
and MVAC cable cost, tower cost, installation cost, decommissioning cost, wake losses, HVAC, MCAC,
and substation electrical losses, among others. In this regard, it is recognized that significant research
is needed to encompass all sources of uncertainty during the wind farm design.

3.2.6. Economic Modeling of Wind Farms

The total cost of a wind farm can be decomposed into two different kind of costs: fixed (or capital)
and variable (or recurrent) costs. Fixed costs include the acquisition and installation of WTs, cost of
the land, foundations, power lines, transformers, auxiliary roads and additional infrastructure as pointed
out previously. Variable costs include O&M costs, costs of WT component degradation, costs for
unexpected problems mitigation and taxation. In many approaches variable costs, such as O&M costs,
has been considered constant over the optimization procedure. Only two works [166,279] tried to
model the effect of fatigue degradation in the O&M costs. In addition, the energy price evolution,
which plays an important role in the economic feasibility of the wind farm project, has been considered
constant, or predefined with a linear evolution, over the project lifetime, with the exception of the
works of González et al. [278] and the TOPFARM framework [306]. Some costs such as taxes have
not been considered because of the different taxation legislation found in different countries.

Cost of Energy (CoE)

Several major European projects have looked at the cost scaling of wind farm components.
The Structural and Economic Optimisation of Bottom-Mounted Offshore Wind Energy Converters
Study (Opti-OWECS) [319,320] investigated the state of the art of offshore wind technologies from
1996 to 1997 and endeavored to determine methods by which to lower the CoE from offshore wind
farms. The Opti-OWECS study covered the economics of offshore WTs, support structures, electrical
interconnection, installation, siting, layout, and O&M. The Offshore Wind Energy Cost and Potential
(OWECOP) project was focused on the development of a software tool to model offshore wind farm
costs. The Dutch Offshore Wind Energy Converter (DOWEC) project, conducted from 1997 to 2003,
looked at ways of improving offshore WT design and increasing their cost-effectiveness. The project
identified approaches to design and build better, more reliable large WTs (e.g., 5~6 MW) for large
wind farms, and further examined the electrical, O&M, support structure, and rotor-nacelle assembly
(RNA) costs as well as WT wakes. The major findings of such works are summarized and adapted in the
work of Elkinton [55], who proposed the definition of Levelized Production Costs (LPC), as shown in
Equation (11). The cost components of Equation (11) are briefly summarized here, while the full
derivation and validation can be consulted in [55]. All cost functions are reported in 2006 U.S. dollars.
The Rotor-Nacelle Assembly (RNA) cost [$] consist of an empirical function of the rated WT
power (𝑃rated [kW]), as shown in Equation (61):
𝐶RNA = 1170 𝑃rated (61)
The cost of the support structure (𝐶𝑆𝑆 ), which consists of the tower, sub-structure, and foundation,
depends on the type of support technology considered. The cost of the WT tower (𝐶tower ) is given next:
Energies 2014, 7 6972

2.8
𝑧H 1.7 𝑃rated 0.6
𝐶tower = 1.5 𝑚tower = 1.5 (𝑐tower 𝐷 ( ) ( ) ) (62)
𝐷 𝐴

where 𝑚tower [kg] is the mass of the tower, 𝑐tower = 0.016 is a constant determined in a validation
study, 𝐷 [m] is the rotor diameter, 𝑍H [m] is the hub height and 𝑃/𝐴 [W/m2] is the rated power density
of the WT. Three different support/foundation technologies were studied: the gravity ( 𝐶GB ), the
monopile (𝐶pile ), and the tripod (𝐶tripod ) bases, and their costs are summarized in Equations (63)–(65)
respectively. The cost of gravity bases is given by:
𝐶GB = 𝑚GB 𝑐𝑐concrete (1 + 𝑐𝑙GB ) (63)
where 𝑚GB [kg] is the mass of the gravity base, which is obtained based on an iterative procedure,
which includes the soil bearing capacity, maximum horizontal forces (created by wind and wave
loads), buoyant forces and downward forces [55]. 𝑐𝑐concrete [$/kg] is the price of the material used to
manufacture the gravity base (e.g., reinforced concrete), and 𝑐𝑙GB is the labor factor including the cost
of installation. The cost of monopile bases is:
𝐶pile = 𝑚pile 𝑐𝑐steel (1 + 𝑐𝑙MP ) (64)
π 2
𝑚pile = 𝜌steel 𝐿pile [𝐷pile − (𝐷pole − 2𝑡wall )] (65)
4
where 𝑚pile [kg] is the mass of the monopile, which depend on the density of the material used
( 𝜌steel [kg/m3]), the diameter of the monopile ( 𝐷pile [m]), the thickness of the monopile wall
(𝑡wall = 0.011 𝐷pile [m]) and the length of the pile (𝐿pile [m]), which depend on the expected axial and
lateral loads [55]. 𝑐𝑐steel [$/kg] is the price of the material used to manufacture the monopile (e.g.,
steel), and 𝑐𝑙MP ≈ 1.8 is the labor factor including the cost of installation. The cost of tripod bases is:
𝐶tripod = 𝑚tripod 𝑐𝑐steel (1 + 𝑐𝑙tripod ) + 3𝐶pile (66)
where 𝑚tripod [kg] is the mass of the tripod, which depends on the expected loads, and 𝑐𝑙tripod ≈ 1.5
is the labor factor including the cost of installation.
The electrical interconnection cost model is decomposed in two separated models; the first model
quantifies the costs of collection power lines (MVAC), and the second model quantifies the costs of
transmission power lines (HVAC).
The MVAC collection system cost (𝐶MVAC [$]) is a function of the total cable length (𝐿MVAC [m])
and the rated voltage (𝑉MAC [V]) as shown in Equation (67):
𝐶MVAC = 𝐿MVAC (𝑎1 𝑉MAC + 𝑏1 ) (67)
where 𝑎1 = 0.0105 and 𝑏1 = 95. The HVAC collection system cost (𝐶HVAC [$]) depends on the wind
farm power (𝑃𝐹 [W]), the cable voltage (𝑉HAC [V]) and the cable length (𝐿HVAC [m]). The number of
cables required to transmit the wind farm power is estimated using Equation (68):
𝑁cable = ceil(6.03 × 10−4 𝑃𝐹 /𝑉HAC + 0.464) (68)
where ceil is a function that rounds its argument towards infinity. The cost of the switch gear
(𝐶switch [$]), and the transformer (𝐶𝑥frm [$]), used to rise the voltage of the collection network, is
given in Equations (69) and (70):
Cswitch = 0.668 𝑉HAC + 36,000 (69)
Energies 2014, 7 6973

𝐶𝑥frm = 1.035𝑃𝐹0.751 (70)

The cost of the HVAC transmission system is given in Equation (71):
𝐶HVAC = 𝑁cable [(𝑐𝑐cable + 𝑐𝑐laying )𝐿 + 𝐶switch + 𝐶𝑥frm ] (71)
where the price per unit length of the cable itself is 𝑐𝑐cable = 0.00168 𝑉HAC + 1280 and
𝑐𝑐laying = 100. The decommissioning cost model depend on the type of foundation and can be
expressed in terms of cost per WT or cost per MW installed, as shown in Equation (72):
(𝑐𝑐WT 𝑁WT )(1 + 𝑐𝑐cable )
𝐶decom = { (72)
(𝑃𝐹 𝑐𝑐MW )(1 + 𝑐𝑐cable )
where 𝑐𝑐WT = 200,000 [$] is the decommission cost for each WT in the wind farm, 𝑐𝑐MW = 55,000 [$]
is the cost for decommission per MW installed, and 𝑐𝑐cable = 12 [$] is the cost of decommissioning of
power lines (i.e., for both collection lines and transmission lines). The O&M cost was considered
constant over the wind farm operational time and calculated as a percentage of the initial capital cost
(𝐶Inv = 𝐶RNA + 𝐶SS + 𝐶Elect + 𝐶Decom ).

Net Present Value (NPV)

The most comprehensive NPV model was defined by Castro Mora [56] and González et al. [135,278,310].
The model considers that wind farms require initial investments or capital costs ( 𝐼WF [$], which
represent the WT acquisition and infrastructure costs) to operate. Once in operation, the wind farm
delivers a stream of both financial benefit, which represent the sales revenue of the energy, and O&M
costs during its expected operational lifetime (𝐿𝑇 [y]) from which net cash flows 𝑁𝐾 of the kth year,
representing the net income (energy sale minus O&M costs), are obtained. At the end of the
operational lifetime of the wind farm additional present value expenses arise for the installation
decommissioning (𝐶D [$]), and an additional income might come from present residual values (𝑉R [$])
as shown in Equations (14) and (73):
𝐿𝑇 𝑁𝑘 (𝑥F )
NPV(𝑥F ) = −𝐼WF (𝑥F ) − 𝐶D (𝑥F ) + 𝑉R (𝑥F ) + ∑ (73)
𝑘=1 (1 + 𝑟)
𝑘

where 𝑥F is a wind farm configuration, and 𝑟 is the equivalent discount rate or real interest rate (annual
depreciation of money), which can be calculated using the Fisher equation, which accounts for the
nominal interest rate (𝑖) and the expected inflation rate (𝑝), shown in Equation (74):
1+𝑖
𝑟= ≅𝑖−𝑝 (74)
1+𝑝
The previous model was extended by González et al. [278] in order to include both the energy
price and the O&M costs evolution in time. In addition, the costs of civil engineering and electric
infrastructure as well as the energy losses were taken into account during the NPV calculation as
shown next:
NPV(𝑥F ) = −𝐼WT − 𝐼CW (𝑥F ) − 𝐼EI (𝑥F ) − 𝐶D (𝑥F ) + 𝑉R (𝑥F )
𝑘+1
𝐿𝑇 (𝐸 (𝑥 ) − 𝐸 (𝑥 ))𝑘 𝑝
G F EL F AV kWh (1 + Δ𝑝kWh ) − 𝐶O&𝑀 (𝑥F )(1 + Δ𝐶O&𝑀 )𝑘+1 (75)
+∑
𝑘=1 (1 + 𝑟)𝑘
Energies 2014, 7 6974

where 𝐼WT [$] is the initial investment for the acquisition of WTs, 𝐼CW [$] is the civil infrastructure
cost, 𝐼EI [$] is the electrical infrastructure cost, 𝐸G [kWh] is the gross converted energy, 𝐸EL [kWh] is
the energy losses in the electric system, 𝑝kWh [$/kWh] is the energy sale price, 𝑘AV is the availability
factor [%], Δ𝑝kWh [%] is the annual increase of the wholesale price of energy, 𝐶O&𝑀 [$] is the cost of
operation and maintenance, and Δ𝐶O&𝑀 [%] is the annual increase of 𝐶O&𝑀 .
The civil engineering infrastructure cost (𝐼CW ) is composed of the WT foundations costs and costs
derived from the construction of auxiliary roads, which are assumed proportional to the total length of
the auxiliary road. The foundation costs are modeled considering the bearing capacity of the land.
The electrical infrastructure of the wind farm has three main components: the internal MVAC
collection network, substations, and the HVAC transmission lines. Given a certain electrical configuration
𝑦(𝑥𝐹 ), the total present cost of the execution and commissioning of the electrical infrastructure can be
calculated using Equation (76):
∑𝐿𝑇
𝑘=1 𝐶EL,𝑘 (𝑦(𝑥F ))
NPV(𝑦(𝑥F )) = −𝐼E (𝑦(𝑥F )) − (76)
(1 + 𝑟)𝑘
where 𝐼E [$] is an initial capital investment. The electrical infrastructure generates running costs
related to electrical losses ( 𝐶EL,𝑘 [$]), which are modeled as simple Joule-based power losses,
depending on the power line’s resistance and the amount of current flowing through the conductor.
Equation (76) can be further decomposed to describe the NPV contribution of each one of the three
main components of the electrical infrastructure:
NPV(𝑦(𝑥F )) = NPV (𝐶MVAC (𝑦(𝑥F )) + 𝐶SB (𝑦(𝑥F )) + 𝐶HVAC (𝑦(𝑥F ))) (77)
where 𝐶MVAC [$], 𝐶SB [$], 𝐶HVAC [$] are the costs of the internal distribution network, cost for
substations and the cost of the transmission lines, respectively. The present value of the total cost of
the internal distribution network (𝐶MVAC [$]) is quantified as the summation of the effective value of
the power to be transported by each section of the power line conductor 𝑃MVAC RMS𝑗 [W], the length of
each power line 𝐿MVAC𝑗 [m], its respective cost per unit length 𝐶F MVAC𝑗 [$/m], and the unit variable
costs per unit length and power 𝐶V MVAC𝑗 [$/mW]:
MVAC Lines
𝐶MVAC (𝑦(𝑥F )) = NPV (∑ (𝐶F MVAC𝑗 + 𝐶V MVAC𝑗 𝑃MVAC RMS𝑗 (𝑦(𝑥F ))) 𝐿MVAC𝑗 (𝑦(𝑥F ))) (78)
𝑗=1

In an equivalent way, the present value of the total cost of the HVAC transmission lines (𝐶HVAC )
can calculated as follows:
HVAC Lines
𝐶HVAC (𝑦(𝑥F )) = NPV (∑ (𝐶F HVAC𝑗 + 𝐶V HVAC𝑗 𝑃HVAC RMS𝑗 (𝑦(𝑥F ))) 𝐿HVAC𝑗 (𝑦(𝑥F ))) (79)
𝑗=1

The total present value of the cost of the substations (𝐶SB ) can be calculated as a summation of the
annual effective value of the power to be carried by each substation 𝑃SB RSM𝑗 [W], their respective unit
fixed costs 𝐶F SB𝑗 [$], and the unit variable costs per unit power 𝐶V SB𝑗 [$/W], as shown in Equation (80):
Subs
𝐶SB (𝑦(𝑥F )) = NPV (∑ 𝐶F SB𝑗 + 𝐶V SB𝑗 𝑃SB RMS𝑗 (𝑦(𝑥F ))) (80)
𝑗=1

Equations (78)–(80) assume that the cost of the electrical infrastructure increase linearly as a function
of the power transmitted. This however, might not always be true as shown in Equations (67)–(71).
The presented NPV model accounts for most fixed and variable costs. However, the inclusion of
Energies 2014, 7 6975

additional recurrent costs (e.g., landowner royalty fees, WT structural degradation, taxes, etc.) might
be necessary to complement the model.

Financial Balance (FB)

The most comprehensive FB model, shown in Equations (13) and (81), was provided by Larsen [279],
which considers only relevant cost to the WFDO problem in a relative basis (i.e., fixed costs such as
the price of WT, costs of planning and projecting, cost of the land, etc. are not taken into account since
it is assumed that they don’t influence the global optimum during the WFDO process):
𝑟c1 − 𝑟𝑖 𝑇𝑛𝐿 𝑟c2 − 𝑟𝑖 𝑇𝑛𝑎
FB = 𝑊𝑃 − 𝐶D − 𝐶M − 𝐶O − 𝐶 (1 + ) − 𝑝𝑠𝑓 𝐶 (1 + ) (81)
𝑛𝐿 𝑛𝑎
where 𝑊𝑃 is the sales revenue of the expected electrical energy conversion over the wind farm
operational lifetime (𝑇 [y]), 𝐶D is the accumulated cost of components degradation, 𝐶M is the cost of
overall maintenance. 𝐶O = 𝑝t 𝑊𝑃 + 𝑝lp 𝑊𝑃 is the cost of taxes and land payments, which are assumed
proportional to the sales revenue. 𝑝t and 𝑝lp are the rate of taxation and rate of the land payment
respectively. 𝑟c1 [%] is the interest rate that the wind farm consortium has to pay for loans (i.e., price
of money in banks or by other investors), 𝑟c2 [%] is the investment rate (i.e., the interest rate that the
consortium can obtain if investing their money alternatively in financial instruments/projects with
identical risk profile compared to the wind farm), 𝑟𝑖 [%] is the inflation rate, 𝑛𝐿 is the number of times
interests of loans has to be paid in a year, 𝑛𝑎 is the number of times additional potential revenues from
alternative investments are to be received in a year. 𝑝𝑠𝑓 describes the split of the total (variable)
investment on consortium loans and consortium financial assets and is defined as the ratio between
the consortium financial assets and 𝐶. If no secondary investments are to be made, and no taxes and
land payments are to be considered (e.g., offshore wind farms), Equation (81) can be simplified to
Equation (13) as proposed by Réthoréet al. [166].
The variable part of the total wind farm investment costs (𝐶) is expressed in Equation (82):
𝐶 = 𝐶F + 𝐶I + 𝐶T + 𝐶E (82)
where 𝐶F = ∑𝐼𝑖=1 𝑛𝑖 𝐶𝑇𝑖 is the total cost of foundations and 𝑛𝑖 is the number of WTs erected on
locations having foundation costs 𝐶𝑇𝑖 . 𝐶I = ∑𝐽𝑗=1 𝑛𝑗 𝐶𝐼,𝑗 is the total cost of WT installation, and 𝑛𝑗
is the number of WTs having installation costs 𝐶𝐼,𝑗 . 𝐶T = ∑𝐾 𝑘=1 𝑛𝑘 𝐶𝑇𝑘 is the total cost of civil
engineering infrastructure, and 𝑛𝑘 is the total length of the road having construction cost 𝐶𝑇𝑘 .
𝐶E = ∑𝐿𝑙=1 𝑛𝑙 𝐶𝐶𝑙 + ∑𝑀 𝑚=1 𝑛𝑚 𝐶𝑃𝑚 is the total cost of electrical infrastructure (excluding transformer
costs and the costs of connecting the wind farm with the main electrical grid, since these costs are
considered independent of the wind farm topology). For the electrical infrastructure costs, two classes
of grid elements are defined; one associated with cabling and the other associated with an overhead
pole setup. 𝑛𝑙 is the total length of the power grid having a cost per meter 𝐶𝐶𝑙 (including cables,
trenching and laying of these, etc.), and 𝑛𝑚 is the total electrical elements with associated cost 𝐶𝑃𝑚
(including the price of overhead poles, the erection of these, cables, etc.).
Fatigue degradation is a stochastic process driven by stochastic loading of the WTs and is modeled
as a linear process. For a WT structural member 𝑆, the mean cost of fatigue load degradation is assumed
to be proportional to the mean accumulated equivalent fatigue load, associated with a suitable component
Energies 2014, 7 6976

hot spot, caused by turbine operation during the lifetime of the wind farm. The mean relative
degradation of the structural component 𝑆 is given by:
𝐿𝑒𝑞,𝑆
𝐷𝑆 = (83)
𝐿𝑑,𝑆
where 𝐿𝑒𝑞,𝑆 (e.g., Equation (42)) and 𝐿𝑑,𝑆 are the lifetime accumulated and the design equivalent
fatigue loads, respectively. The cost of degradation of the structural member 𝑆 is defined as:
𝐶𝐷𝑆 = 𝑃𝑆 𝐷𝑆 (84)
where 𝑃𝑆 is the price of the structural component 𝑆 . The cost of degradation of all structural
components of all the WTs (N) of the wind farm can be calculated as:

𝐶𝐷 = ∑ ∑ 𝐶𝐷𝑆 (85)
𝑁 𝑆

The O&M costs were formulated by applying a probabilistic failure criterion [279]. Assuming that
the component equivalent fatigue load is a stochastic variable with a log-Gaussian distribution [321],
the cost of O&M, related to the structural component 𝑆, is approximated as:
𝑅max
𝐶𝑀𝑆 = 𝑁R 𝑃𝑆𝑟 𝐹(𝐷𝑆 ; 𝜇𝑆,(𝑅max +1) , 𝜎𝑆,(𝑅max +1) ) + 𝑃𝑆𝑟 ∑ 𝐹(𝐷𝑆 ; 𝜇𝑆,𝑗 , 𝜎𝑆,𝑗 ) (86)
𝑗=1

where 𝑃𝑆𝑟 denotes the replacement cost (i.e., the cost of the physical component replacement and
additional expenses originating from derived energy conversion loss), 𝑅max is the maximum number
of allowable component replacements defined by the designer, and 𝑁R is a weight factor large enough
to assure that more than 𝑅max component replacements is unfavorable for the optimal wind farm
layout. 𝐹(𝐷𝑆 ; 𝜇𝑆,𝑗 , 𝜎𝑆,𝑗 ) is the log-Gaussian cumulative distribution function with modulation
parameters 𝜇𝑆 and 𝜎𝑆 . The jth set of modulation parameters is related to the component expected
lifetime equivalent fatigue load, E [𝐿𝑒𝑞,𝑆 ] , and the component equivalent fatigue load variance,
VAR [𝐿𝑒𝑞,𝑆 ], through the respective mean and variance of the relative degradation measure as:

VAR[𝐿𝑆𝑟 ]
𝜎𝑆,𝑗 = √ln ( + 1) (87)
𝑗 2 𝐿2𝑑,𝑠

2 (88)
𝜇𝑠,𝑗 = ln(𝑗) − 𝜎𝑆,𝑗

The cost of operation and maintenance of all structural components of the WTs can be calculated as:

𝐶𝑀 = ∑ ∑ 𝐶𝑀𝑆 (89)
𝑁 𝑆

3.3. Solution Methods

The solution methods adopted for the WFDO problem can be categorized in three different groups:
(1) Mixed-Integer Nonlinear Programing (MINLP) models and Calculus-based methods; (2) Heuristic
methods; and (3) Metaheuristic methods. Their usage to provide a solution to the different versions of the
WFDO problem is illustrated in Figure 4. Descriptions of such methods are provided in Section 3.3.1.
The most widely used solution methods are non-deterministic population metaheuristics based on the
Energies 2014, 7 6977

Genetic Algorithm. Nonetheless, is common to find combinations of methods in the literature, which
are discussed in Section 3.3.4. In addition, the interested reader is directed to the review
works [153,154,235,322], which present a clear and comprehensive comparison between the most used
metaheuristics for the solution of the WFDO problem.
The evaluation of the performance metrics, discussed in Section 3.1, can be performed in a
single-objective or in a multi-objective strategy. In the single-objective strategy, the first level of
sophistication aims to include adverse effects (described in Section 2.2) as constraints and/or as
penalties to the objective performance metric, which in turn affects the decision variables values (described
in Section 2.4) during the optimization procedure. The second level of sophistication, adopted by most
of the reviewed works, aims to include adverse effects during the performance metric quantification
with the use of sub-models (as described in Section 3.2). However, as pointed out previously, the
development of sub-models used to translate the implications of certain adverse factors (e.g., noise,
EMI or visual impacts) into the performance metric (e.g., the cost of energy) is difficult, thus hindering
extremization of each objective/factor simultaneously.

Figure 4. Type of solution methods used in the WFDO problem.

Timeline of Solution Procedures
30
NA
MINLP & Calculus-Based
25 Heuristic
Metaheuristic

20
No. of Papers

15

10

0
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Year

In the multi-objective strategy, the inclusion and combination of performance metrics and adverse
effects is easier and, therefore, both can be extremized simultaneously during the optimization procedure.
Therefore, the main advantage of the multi-objective strategies is that there is no need to formulate
complex sub-models that translate physical variables values into performance measure values, since
both kind of variables can be considered as performance metrics of the problem. The WFDO problem
is naturally categorized as a multi-objective problem, where individual design objectives are generally
in conflict. An optimal solution to a multi-objective problem is one of a set of feasible solutions that
satisfies the objectives to an acceptable level (i.e., where the trade-off between two or more objectives
is the best). Only a few works in the WFDO problem [154,159,163,202,209,250,264] have implemented
Energies 2014, 7 6978

multi-objective approaches. In addition, as pointed out in [153,154], many different multi-objective

formulations have not been implemented and/or tested for the solution of the WFDO problem.

3.3.1. Calculus-Based Optimization

Calculus-based optimization (or exact optimization) encompass all deterministic and non-deterministic
algorithms and iterative methods that rely on the first and second order derivative of the objective
function to explore the search space with the aim of finding the optimal solution(s). Calculus-based
algorithms require that the objective function is Lipschitz continuous (i.e., it satisfies the Lipschitz
continuity condition) and is differentiable. In addition, it is required that the search space is
convex/concave and bounded to guarantee convergence to the optimal solution. The most common
calculus-based optimization algorithms are the gradient-based, Newton-based, branch-and-bound,
branch-and-cut, Extended Cutting Plane (ECP), Generalized Benders Decomposition (GBD),
Outer-Approximation (OA), and backtracking, among others [323–327]. The most basic optimization
techniques perform an explicit enumeration of all admissible solutions to obtain the optimal solution.
However, for most problems this is computationally expensive. Local search calculus-based algorithms,
such as the gradient descent method or Newton-based methods cannot solve non-convex problems to
optimality, since they don’t have the required capacity to overcome local extreme points. To overcome
this issue, procedures based on global search [328,329] and multi-start procedures have been proposed
in the literature to improve its performance. These approaches perform a local search from multiple
feasible starting points and sample multiple basins of attraction (i.e., sets of initial starting points that
converge to the same local extreme point) to assess if an optimal solution is obtained. The limitations
of such approaches are that the objective function must be continuous, differentiable, bounded, and
must not be too complex, otherwise a significant amount of starting points will be required to properly
sample the search space. Different approaches in the literature have addressed the solution of small and
simplified versions of the WFDO problem using gradient-based algorithms [27,180,219]. However,
it is recognized that the capabilities of such algorithms to find optimal solutions to WFDO problem are
limited [219].
Improved exact techniques, such as tree search-based procedures (e.g., branch and bound algorithms),
perform an implicit enumeration of all the admissible solutions (i.e., all the admissible solutions are
considered and implicitly evaluated but are not explicitly constructed) [324,326,328,329]. Although
tree-search based procedures have shown improved computational efficiency to find optimal solutions,
these methods perform poor when highly constrained models and non-convex search spaces are
considered. To overcome this limitation new improvements based on relaxation procedures are being
developed in recent research [330–332]. Relaxation consists of transforming the original problem into
another equivalent problem of reduced complexity (e.g., having less and/or different restrictions,
different type of variables, objective functions with reduced order, etc.), which can be solved with
other commonly used solution procedures. Relaxation denominations depend on the kind of relaxation
performed (e.g., the Linear Programming (LP) relaxation consist of transforming an integer-type
problem into a problem with continuous variables by disregarding the constraint that the variables
must be integer). The solution of the relaxed problem can be used to gain information about the
solution to the original problem. However, relaxation procedures can significantly affect the original
Energies 2014, 7 6979

problem and thus its optimal solution (i.e., depending on the problem solved and the adopted
relaxation procedure, both the shape of the solution and the solution quality of the relaxed problem can
be significantly different that the optimal solution shape and quality of original problem). Therefore,
the most important challenges are to improve relaxation procedures and to develop effective
optimization algorithms to solve nonlinear and non-convex problems.
Depending on the type of objective function (e.g., convex, nonlinear, etc.), the type of constraints
and the type of variables (e.g., Boolean, integer, continuous, mixed-integer, etc.) different denominations
can be given to the complete optimization model being solved (e.g., a model having a nonlinear objective
function, nonlinear constraints and involving both discrete and continuous variables is referred as
a Mixed-Integer Nonlinear Programing Problem (MINLP)). The common operations performed by
commercial software (e.g., CPLEX, LINDO, KNITRO, etc.) to solve MINLP problems with calculus
based-algorithms are: (1) relax (e.g., linearize the objective function) and (2) search (e.g., in an
exhaustive fashion using tree search-based algorithms). However, as mentioned in Section 2.5, even
the simplest versions of the WFDO problem are recognized as non-analytically-differentiable,
non-continuous and non-convex MINLP models with NPO-complete complexity. This fact limits the
applicability of the calculus-based procedures to provide a solution to the WFDO problem.
One of the first approaches using MINLP models in the WFDO problem was proposed by
Donovan et al. [172–175], who formulated a wind farm production model based on the generalized
vertex packing problem (GVP), seeking to maximize the power generated in accordance with the
constraints based on the number of WTs, WT proximity and WT interference. However, the adopted
wake modeling approach is not validated, and the method is nonviable for sophisticated wake models
since the MINLP formulation does not allow the inclusion of non-differentiable objective functions.
An extended formulation, which is partially based on the works of Donovan, was proposed by
Fagerfjäll [167] who developed a MINLP model for wind farm power conversion and a MINLP model
for civil engineering and electrical infrastructure design, which was treated as Steiner tree problem.
Although the approach shows the same weakness when modeling wakes (and possibly other complex
adverse effects as described in Section 3.2), an important improvement in the auxiliary infrastructure
design was achieved, thus denoting the importance of MINLP models and tree search-based solution
procedures in the WFDO problem. To overcome the wake modeling limitation, a new approach was
suggested [186] in which common engineering wake models can successfully be integrated into the
MINLP formulation. Additional works formulating MINLP models of different complexity were found
in the literature [84,186,197,198,205,213,248,258]. In most works, the performance metrics and
adverse factors were oversimplified to reduce the complexity of the optimization model. In this regard,
it is recognized that significant research is needed to encompass all the factors and variables described
in Section 2, in order to improve such MINLP formulations.

3.3.2. Heuristic Optimization

Since obtaining an optimal solution to the WFDO problem may not be possible with a reasonable
amount of computational resources, as pointed out in Section 2.5, or because of the natural limitations
of the calculus-based solution procedures, the attention is turned to approaches that compute near-optimal
solutions. Essentially, there is an interest to explore the trade-off between the final solution quality and
Energies 2014, 7 6980

the amount of computational resources spent. While many problems may not have polynomial time
solutions, it may indeed be possible to find a solution that is optimal or very close to the optimal. Heuristics
(commonly known as approximate algorithms, derivative-free algorithms or black-box methods) are
deterministic or non-deterministic solution procedures based on semi-empirical rules that are used to
find near-optimal solutions to complex problems. Generally, these approaches are very fast (fast enough
to be used in practice), and have improved solutions over traditional calculus-based optimization
techniques [140,327]. However, they need to be executed several times to guarantee convergence to a
high quality solution. Heuristics do not necessarily require the computation of gradients during
optimization and do not necessarily require that the objective function is continuous. Heuristics are
classified in two different groups: constructive and iterative. Constructive heuristics build a complete
solution by performing multiple sequential deterministic or non-deterministic assemblies of the involved
variables while considering all defined constraints (i.e., the value of each variable is decided based on
deterministic or non-deterministic rules). Iterative heuristics attempt to improve a complete solution
(that can be obtained from a constructive heuristic) by doing a controlled evaluation of the local search
space of each of the involved variables. Constructive heuristics are faster than iterative heuristics, but
in most cases converge to solutions of unacceptable quality. For highly constrained problems, constructive
techniques may even fail to find a feasible solution.
There are several general heuristic frameworks that have been demonstrated to perform well for
various types of problems (e.g., the Traveling Salesman Problem (TSP) [333]). Determining which
heuristic to use for solving a particular problem is one of the most common (and most difficult to
solve) questions in heuristic search. It is common to suspect that some heuristics might perform better
than other heuristics when considering a set of different optimization problems, but unfortunately the
No Free Lunch Theorems (NFLT) [327,334] suggest that this is not the case. A summary of the NFLT
states “for all possible performance measures, no algorithm is better than another when its performance
is averaged over all possible optimization problems” [334]. The most basic heuristic search algorithm,
Random Search (RS), randomly samples a number of feasible solutions and determines the best solution
from the sample. Apparently, any heuristic that evaluates problems more cleverly should always perform
better than RS, but as the NFLT suggests, their performance will be the same when averaged over all
possible optimization problems. While it may not be possible to perform better than RS over all
optimization problems, research has shown that it is possible for heuristics to perform much better than
RS for specific problems [335–337]. Such heuristics usually take advantage of the structure of the
specific problem to obtain near-optimal solutions. This structure may include the constraints of the
problem, how a feasible solution is represented (e.g., binary permutation, real-valued string, etc.),
or even the instance of the problem at hand. Since any heuristic is equivalent to RS when averaged
over all possible optimization problems, any fine-tuning of the driving parameters of the considered
heuristic algorithm(s) will inherently create weaknesses with respect to other optimization problems.
Therefore, careful tuning of the heuristic driving parameters must be performed when considering a
specific optimization framework and should not be based on the suggested values obtained from
different optimization frameworks.
Different heuristic strategies have been used to solve the WFDO problem. The first heuristic
strategy was provided by Ozturk and Norman [152], who implemented a deterministic greedy constructive
algorithm used to solve simplified wind farm scenarios. The reason behind the implementation and
Energies 2014, 7 6981

use of the heuristic procedure was to overcome the natural complications of the adopted non-linear
mathematical model, which, however, showed modest complexity as compared with the models
resumed in this work. Greedy heuristics have been retaken recently as shown in the works of
Chen et al. [266] and Yin and Wang [265] and have demonstrated the potential to solve complex
versions of the WFDO problem. In addition, other approaches in the literature have successfully
adopted different heuristics such as the Monte Carlo method [183,187–189,238,239], pattern search
algorithms [244], viral algorithms [255,256], Powell’s method [257], particle filtering [204], bionic
optimization [208], and customized local search algorithms [210,271]. The set of heuristics, however,
have never been compared systematically to determine the best performer, thus resulting in an important
topic of research.

3.3.3. Metaheuristic Optimization

For some problems, there are known heuristic algorithms which are practically useful, but for many
problems the best-known heuristic algorithms have running times too large to be of practical use,
despite the fact that they might have polynomial time execution. To overcome this issue, more complex
and intelligent heuristics have been developed and are commonly referred as metaheuristics and
hyper-heuristics [327]. Metaheuristics are high level procedures designed to find near optimal
solutions more efficiently than common heuristics. Hyper-heuristics are higher level procedures
designed to find or generate a lower-level procedure that may provide a near-optimal solution to a
complex problem. Metaheuristics are commonly based on optimization processes observed in the
nature (e.g., Simulated Annealing (SA) [338–344], Particle Swarm Optimization (PSO) [327,345],
Genetic Algorithm (GA) [327,346–348], etc.). Some of the multiple remarkable aspects of
metaheuristics are that they make only a few assumptions about the function being extremized, they are
relatively easy to implement, and do not require the objective function to be continuous or differentiable.
Even though, metaheuristics are designed to be problem-independent, some problems that are successfully
solved by one metaheuristic algorithm may not be solved when using another metaheuristic algorithm
(i.e., metaheuristics are not universal and are subject to the NFLT). In addition, they may require more
elementary operations to perform the optimization tasks compared with simple heuristics or
calculus-based algorithms. In this regard, the decision of using a metaheuristic algorithm must be based on
the problem complexity rather than in its implementation easiness, since for some problems calculus-based
algorithms can outperform heuristics and metaheuristics.
Metaheuristics are classified in three different groups: constructive-based, local search-based, and
population-based metaheuristics. Constructive metaheuristics are based on a set of heuristics to build a
feasible solution. The Greedy Randomized Adaptive Search Procedure (GRASP) [327,349] is considered
one of the most powerful constructive metaheuristics and consists of two phases: (1) an iterative
procedure made up from a greedy randomized construction (i.e., a semi-greedy heuristic) and
(2) a local search iterative procedure, which may be a calculus-based local search procedure or a
heuristic/metaheuristic-based local search procedure, that improves the previously constructed solution.
Yin and Wang [265] implemented a GRASP algorithm complemented with a Variable Neighborhood
Search (VNS) metaheuristic algorithm and showed improved results when comparing them with the
results of Mosetti et al. [28] and Grady et al. [232], who employed a genetic algorithm to optimize
Energies 2014, 7 6982

wind farms under simplified conditions. However, the shortcomings of the proposed GRASP procedure,
and other procedures as well, are the need to discretize the search space to define the possible locations
of the WTs integrating the wind farm, and the fact that the inclusion of additional variables is not
straightforward, requiring a complete reformulation of the state/solution representation.
Local search metaheuristics are based on the fundamental idea in which feasible solutions of similar
quality are related. The hope is that if a solution has good quality, then a related solution will also be
good (and possible better). The notion of two solutions being related is handled differently in different
local search-based algorithms. Local search procedures iteratively modify the value of the design
variables, by performing deterministic or non-deterministic actions, resulting in neighbor solutions.
The algorithm selects the best neighbor solution at each iteration and repeats the process until no better
solutions are found in the neighborhood. The main advantages of local search algorithms are that they
are very easy to implement and have few driving parameters. At the same time, they can be very
powerful solution procedures for several problems [350,351] including the WFDO problem. The main
drawback of local search is that if the search space is too complex (i.e., involving many local extreme
points and/or many optimal solutions), the algorithm may not converge to the optimal solution.
To overcome this limitation, local search metaheuristics use additional heuristics to decide which is the
best neighbor solution (i.e., neighbor solutions with lower quality may be selected).
Different local search procedures have been proposed to solve the WFDO problem. Yin and
Wang [265] adopted VNS as a complementary local search procedure to their GRASP implementation,
as previously mentioned. VNS was first introduced in 1997 [352] and is quite similar to common local
search procedures such as the calculus-based multi-start procedure. VNS performs a local search
from multiple neighborhoods and accepts neighbor solutions only if an improvement is made.
The difference between the VNS and the multi-start algorithms is that the VNS systematically changes
the neighborhoods being explored in two phases: the first phase involves a gradient decent procedure
to identify local extreme points, and the second phase involves a random perturbation which allows the
algorithm to explore other neighborhoods.
Another successful local-search metaheuristic applied to the WFDO problem is the Simulated
Annealing (SA) algorithm [327,338]. SA is based on the process of annealing of solids, which is a
thermal process for obtaining low energy states in solids by exposing them to a heat bath. The annealing
process consists of the following two steps [343]: (1) increase the temperature of the heat bath until
the solid melts; and (2) slowly decrease the temperature of the heat bath until the particles arrange
themselves in the low energy ground state of the solid. In the melt phase, the constituent particles
arrange themselves randomly, and in the cooling stage the particles arrange themselves in an organized
lattice. The analogy between the annealing process and the optimization process is that potential
solutions of the studied problem are comparable to states of a physical system undergoing a process of
annealing, and the quality of a potential solution is equivalent to the energy of a state [338]. While local
search (and variable local search) generally will accept transitions between neighbor solutions only if
its quality is better than the current solution, SA has the capacity of allow transitions between solutions
whose qualities are worse than the current solution in the hope of avoiding getting stuck in a local
extreme point. The transition to a worst solution is modeled as a probabilistic process, which relies
on the Metropolis criterion [327], and depends on the actual state and the new state quality. The
probability of rejection of states with lower quality increase as the search continues (i.e., SA behaves
Energies 2014, 7 6983

as a greedy algorithm), thus simulating the arrangement of particles in an organized lattice during the
cooling process. The SA’s ease of implementation, convergence properties, and ability to escape local
extreme points have made it a popular technique during the past two decades. The SA-based
metaheuristic has been implemented to solve different versions of the WFDO problem showing better
or equal performance than other metaheuristics such as the genetic algorithm [154,156].
Population-based metaheuristics encompass all solution procedures that use multiple
agents or individuals (i.e., potential solutions), which are interacted between them, to trace
out for better individuals. Population-based algorithms rely on the idea in which the best
characteristics of the best individuals can be combined to produce a better individual. The
Genetic Algorithm (GA) is the most used population-based metaheuristic in the WFDO
problem [56,134,135,156,159–161,163,165,199,200,207,235,237,245,250,254,278,301,307,353–357].
The GA [327,346] refers to a class of adaptive search procedures based on the principles derived from
natural evolution and genetics. In the GA, genes represent the design variables, and potential solutions
(i.e., individuals) are represented by chromosomes. Like other methods, as the GA proceeds through
generations (i.e., iterations), it holds a set of possible solutions referred to as the population. Based on
the fitness value (i.e., the quality of a solution), which is determined through the evaluation of the
population’s individuals, two individuals (i.e., the parents) are selected and are combined through
genetic operators (crossover and mutation) resulting in two new individuals (e.g., the offspring). Many
different variants of genetic algorithms have been considered in the literature [327] differing in the
way genetic operators act over the population and the way individuals are represented. Research has
demonstrated that the performance of a genetic algorithm relies upon the proper choice of the selection
and crossover mechanisms and the proper design of the individual representation.
Another successful population-based metaheuristic algorithm applied to the WFDO problem is
the Particle Swarm Optimization (PSO) algorithm [184,192,193,201,203,212,243,264,269,353], which
was developed by Eberhart and Kennedy in 1995 [345]. The PSO algorithm is inspired by the social
behavior of fish schooling and bird flocking. The PSO algorithm relies on the idea in which different
paths having different length and complexity can lead to an optimal solution (i.e., different particles,
or individuals that represent a possible solutions, can take different paths that lead to the same
destination, which is assumed to be the place where the performance metric is the best). The PSO
algorithm tracks the best position and velocity of each particle integrating the swarm and updates both
quantities iteratively. The particles will move in the search space until the stopping criteria is met,
which is established when the optimal location has been determined or a given number of iterations
have been performed.
A very detailed study that contrasts the different SA, GA and PSO implementations used in the
WFDO problem is given in [154]. Additional solution procedures based on Ant Colony Optimization [195]
and Neural Networks [262] have been used to solve different versions of the WFDO problem.
All approaches have shown promising results. However, these solution procedures have been
implemented in different scenarios having variable complexity, unknown optimal solutions, substantial
differences in main objectives and model constraints, thus limiting any objective conclusion of their
effectiveness while solving the WFDO problem. In this regard, a systematic evaluation of solution
procedures in realistic wind farm scenarios will provide a basis for the judgment of the performance of
each solution procedure in the solution of the WFDO problem. In addition, the performance evaluation
Energies 2014, 7 6984

of additional heuristics and metaheuristics based on the TABU search, scatter search, memetic
algorithms, cuckoo search, etc. has not yet been performed.

3.3.4. Multi-Step, Nesting and Hybridization

New strategies have been developed to handle complex problems, which may contain embedded
optimization problems. The most basic strategy is the multi-step optimization, which consists of computing
a solution to the problem by first using one type of optimization procedure. The solution is later given
as an input to a second optimization procedure, which in turn aims to improve the quality of the given
solution. The process can be repeated systematically by using the same optimization procedures or by
including additional optimization procedures. Multi-step procedures are commonly used to handle
the limitations of certain algorithms (e.g., algorithms that need the discretization of the search space)
by systematically using another algorithm(s) that has different limitations. If the problem contains
embedded optimization problems, then the common procedure is to obtain the best solution to the
general problem to later give it as an input to the embedded optimization problems (e.g., in the WFDO
problem, the first step may be to determine the best wind farm layout to later optimize the design of the
inner electrical infrastructure layout; this procedure, whoever, may compromise the financial balance
of the project). The multi-step procedure has been adopted in the solution of the WFDO problem
several times. In the multi-fidelity approach (i.e., approaches that systematically increase the complexity
of the problem being treated) proposed by of Réthoréet al. [166], within the TOPFARM framework,
the first multi-fidelity level can be categorized as a multi-step procedure, in which the solution of a
Simple Genetic Algorithm (SGA) is given as an input to a Sequential Linear Programming (SLP)
algorithm. It must be noted that higher levels in the proposed multi-fidelity approach consider the inclusion
of advanced objective functions and sub-models and, therefore, cannot be categorized as multi-step
procedures. In addition, it should be noted that some metaheuristic algorithms, such as GRASP [265],
include multi-step procedures as a part of their optimization strategies. González et al. [277], within
the scope of the WFDO problem, demonstrated that sequential multi-step procedures do not lead to the
optimal solutions of problems having embedded optimization problems, instead of global procedures
(or nested procedures) that take into account all the design variables at once.
Solution procedures using the concept of nesting include different algorithms in sub-levels during
the overall optimization. The difference with the multi-step approach is that in the nesting approach a
combination of two or more algorithms, usually a population-based metaheuristic and a local search
algorithm, are used sequentially through the iterative process. The optimization of the design variables
is performed first by the leading algorithm (e.g., population-based metaheuristic) and then, in the same
iteration, the nested algorithm(s) improve the solution quality by performing additional modifications
to the design variables. The nesting approach, as suggested by González et al. [277], might lead to
improved solutions in contrast with muti-step approaches.
Hybridization [327,358–360] procedures are commonly associated with nesting procedures. However,
in a strict sense, hybridization refers to the development of a new method that combines the best
characteristics of other individual methods. Thus, the hybrid method is expected to have superior
performance when contrasting with the performance of the individual methods composing the hybrid
method. Hybrid methods are hardly found in the literature and are generally classified as new algorithms.
Energies 2014, 7 6985

4. Research Needs and Trends in Wind Farm Design and Optimization

The general research trend in WFDO is focused on balancing the trade-off between the development
of complex sub-models (by gradually including additional adverse factors and variables) and the
inherently complexity added to the problem. Although computational technologies are being improved
quickly, they are generally the limiting factor for the treatment of the WFDO problem. Thus, researchers
and wind farm designers rely on advanced solution procedures such as metaheuristics to tackle the
problem. In addition, researchers rely on additional concepts that help setting the optimization framework
such as the preprocessing and the multi-fidelity procedures. The concept of preprocessing, within the
WFDO problem scope, refers to a set of computations performed to build a wind farm scenario,
previous to the optimization procedure. Preprocessing includes the definition of the local wind field
(e.g., through the use of CFD techniques), pre-convolution of WT power curves, pre-calculation of
individual loading cases for WT components using aero-elastic simulations, construction of databases
having relevant information (e.g., WT characteristics such as maximum height, minimum height or
costs), etc. Preprocessed data can be interpolated during the optimization procedure, thus significantly
improving computational times. However, preprocessing techniques typically require additional
assumptions (e.g., conventional CFD techniques used to estimate the local wind resource assume
steady state conditions) and depend on the accuracy of the interpolation procedure. Preprocessing
techniques can be combined with other relevant techniques such as the multi-fidelity approaches in
which the complexity of a problem is gradually increased during the optimization procedure. As
proposed in the novel work of Réthoré et al. [166], the concept of multi-fidelity can significantly
improve computational times and solution quality. However, both the preprocessing and multi-fidelity
approaches have been scarcely applied to the WFDO problem, thus denoting an important area of research.
New issues regarding the potential conflict of interests (e.g., effects of the presence of nearby
projects that are not necessarily wind farms) and the design of multiple wind farms under uncertain
conditions have been treated recently [171,308] and are recognized as potential problems that will
be more frequent in the near future. However, due to the different scales of such problems, certain
assumptions and models (e.g., wind field modeling and wake modeling) may not hold and will require
additional research to determine its validity and/or its extension.
Another important area of research in WFDO focuses on the validation of the various sub-models
described in Section 3.2. In addition, the complete validation of the outputs of the WFDO procedures
remains a complex problem to solve. The specific research trends on the performance metrics, sub-models
and solution methods are discussed in the next sub-sections.

4.1. Performance Metrics

Previous research works [135,199] have demonstrated that the outputs of the WFDO process are
highly sensitive to the choice of the objective function (i.e., there is high dependency of the best wind
farm layout on chosen performance metric). However, a complete sensitivity analysis describing the
effects of using different performance metrics and including different adverse factors during the
optimization procedures has not yet been performed. In addition, a complete sensitivity analysis of
Energies 2014, 7 6986

the resultant optimized layouts considering different atmospheric conditions (e.g., for a wide range of
inflow conditions and different thermal stratification conditions) has not yet been performed.
The general trend of the optimized results obtained among the different approaches adopted in the
literature is to install few and large capacity WTs, in contrast to installing many small capacity WTs, in
irregular arrays in which wake effects are minimized. An interesting trend observed in all the reviewed
works, without regard to the chosen objective metric and optimization procedure, is that significant
improvements (in terms of the selected performance metric) can be obtained due to the application of
systematic optimization methods over micro-siting methods based on rules-of-thumb.
The energy yield alone has proven not to be the best objective function for practical purposes, since
the net economic benefits obtained from the wind farm may not always increase monotonically with
increasing converted energy. This is mainly due to the effects of wakes and turbulence on the power
conversion and loading and/or by the required costs for the acquisition of WTs and the construction of
civil engineering and electrical infrastructure. Therefore, the wind farm with the largest energy
conversion does not necessarily have to be the one that produces the best economic profits. The need
for all-encompassing design variables requires the costs to be computed besides the energy conversion
yield. However, significant improvements in the traditional way to calculate the annual energy
production of wind farms (that may include the inclusion robust multivariate probability density functions
and the inclusion of additional factors such as turbulence, wind shear or WT dynamic loading) are
required in order to improve the energy yield predictions and lower the impacts of risk and uncertainty
on the expected economic returns.
The most suitable performance metrics are the NPV and the FB, since the wind farm purpose is to
produce the largest possible economic profit. The CoE alone cannot provide conclusions about the
viability of a particular project, thus it should be incorporated as a complementary economic metric.
The inclusion of additional finance metrics such as benefit cost ratio (the present value of all benefits
divided by the present value of all costs), the internal rate of return (IRR, is the value of the discount
rate that gives a NPV of zero), and discounted value indicators such as the payback period (the capital
cost of the project divided by the annual average return) might improve the understanding of the
absolute financial prospections in the WFDO problem and will include the possibility of performing
additional analysis that might be necessary when considering different energy markets. In this regard,
the inclusion of dynamic models for energy price setting and the integration of regulation models for
energy marketing remain complex problems yet to be solved.

4.2. Sub-Modeling of Adverse Factors

A wide range of varying conclusions about the different effects of the adverse factors, described in
Section 2.2, produce in the resultant optimized wind farms were provided in the literature. However,
most of the conclusions are based on sub-models in which the effects of such factors were limited or
oversimplified. In addition, the conclusions are commonly based on contrasts between pairs of factors
that are usually in conflict. Only a few works were able to address the overall effects of including
different factors [54–56,199,234,278,310]. Two general kinds of effects in the resultant wind farm
layouts were observed; some factors (e.g., infrastructure costs) induce pressure to concentrate WTs in
high dense irregular clusters within the wind farm area while other factors (e.g., wake effects,
Energies 2014, 7 6987

turbulence and noise) induce pressure to separate WTs. Both trends affect the optimal number of WTs
and the wind farm size. These observed trends, however, are highly dependent on the chosen performance
metric and further analyses are required to understand which factors, and under which conditions,
dominate during the WFDO process. In addition, the inaccuracies and validity ranges of the common
used sub-models may be influencing the shape of the obtained near-optimal solutions, thus the validation
and extension of sub-models is considered the most important aspect to be improved in the development
of the WFDO problem. Some of the most important modeling limitations are described next:
Sub-models describing the impacts of WT wakes are limited, since crude assumptions about the
wake expansion, wake merging, wake rotation, partial wake operation, wake meandering and 2-way
interactions with large scale turbulence structures are commonly considered.
State-of-the-art sub-models describing the structural degradation costs of the different WT members
and the costs associated to O&M are still not validated, and rely on aero-elastic codes that simulate the
loads that each component will undergo while operating, as described in Sections 3.2.2 and 3.2.6.
Although the quantification of WT structural degradation is a complex task that depends on the local
climatology, the particular WT design, and its operation, it is expected that some WT members (e.g., gear
box, main shaft, and WT blades among others) will be more affected than others while operating.
Thus, significant research is required to identify which components are the most affected, and under
what conditions, and to improve the cost prediction of their degradation.
The civil infrastructure cost models are significantly different comparing onshore and offshore wind
farms as described in Sections 3.2.3 and 3.2.6. State-of-the-art infrastructure cost models for offshore
wind farms consider all relevant factors during the design of the WT substructures and foundations
(e.g., seabed depth, tower mass as a function of wind and wave loads among others). However, current
existing models for onshore projects consider that foundation costs are proportional to the land bearing
capacity. The inclusion of additional variables such as the expected WT loading and the type and
capacity of the WT foundation might improve the civil infrastructure modeling in onshore sites.
The most comprehensive civil and electrical infrastructure cost models are highly sensitive to time
variations of the price of their constituent components. In addition, an important lack of modeled
physics in power transmission was observed, which in turn affect the predictions of the effective
expected electric energy conversion. An adequate combination of improved models describing electrical
infrastructure costs, electrical losses, electrical components capacity and reliability is required.
An important aspect that has been largely neglected is the topography of the wind farm area.
The topography does not only affect the computation of the local wind field but also affects the
development of other adverse factors such as noise, wakes, turbulence and electromagnetic fields.
Literature works commonly consider a flat square area and assume, with the exception of a few
works [55,219], that the wind field characteristics are identical throughout the entire site. A flat square
area assumption certainly holds for offshore sites, but it is very unrealistic for onshore sites, where
terrain is rarely flat, has non-uniform roughness and irregular boundaries. The reason behind ignoring
the topography complexity is due to the fact that it is hard to obtain reliable estimations of its effects
though simple models and cheap computational routines. Thus, additional assumptions are commonly
considered when optimizing onshore wind farms (e.g., it is commonly assumed that WT wakes follow
the terrain shape and are not affected by the presence of the ground). However, as previously mentioned,
the use of commercial software (see Section 4.4) and the concept of preprocessing might help to overcome
Energies 2014, 7 6988

some of these limitations. Therefore, the description of how wakes, turbulence, noise and electromagnetic
fields evolve in the domain while considering the local topography remains an important problem
to address. In addition, common approaches in the literature neglect the effects of topography in the
design of the civil engineering and electrical infrastructure. In complex terrain, the total length (and
therefore the total cost) of the auxiliary roads or the MVAC collection lines can be significantly
different from that for flat sites. This fact limits the applicability of the traditional Euclidean minimum
spanning tree and the Steiner tree approaches and induces the necessity for the development of new
robust approaches.
The uncertainty and risk modeling of wind farms has been scarcely treated in the scope of the
WFDO problem as pointed out in Section 3.2.5. Two important unaddressed problems in the topic have
been identified: (1) only a small set of factors that contribute to the overall uncertainty were included
in the optimization frameworks of the literature works; (2) The treatment of uncertainty is based on the
assumption in which the stochastic variation of certain variables can be described using predefined
statistical distributions. The assumptions, however, are not validated and require additional research.
Environmental and/or social impacts within the WFDO procedures have been neglected in most
works as pointed out in Section 3.2.4. Although their inclusion into an optimization framework demands
significant efforts in both the modeling and optimization stages, they should not be discarded as they
usually influence the viability of wind projects.

4.3. Solution Methods

The general research trend in the solution methods for the generalized WFDO problem is focused
on balancing the trade-off between the obtained solution quality and the computational requirements.
Hence, the adoption of metaheuristic algorithms in integrated schemes, while employing the concepts
of multistep, nesting and hybrid optimization is necessary to tackle the problem. Investigation related
to the adequate comparison of solution procedures in realistic wind farm scenarios, and while
employing the same state-of-the-art performance metrics and sub-models is required. This is due to the
fact that certain comparisons between different solution procedures, that are reported in the literature
are meaningless [241,242], since the adopted performance metrics were different or the amount and
type of contrasted algorithms were limited. In addition, the evaluation of other optimization algorithms
that have not been used to attempt solve the WFDO problem may provide a wider and more objective
perspective about which type of algorithms perform best.
Another important aspect to consider is that some solution procedures rely on the relaxation of
the complexity of the problem (i.e., relaxation of the model or variables complexity). While this may
provide interesting conclusions about the mathematical behavior of the problem, relaxations over
already simplified models can lead to important inaccuracies, which in turn impact not only the expected
wind farm performance but also the solution shape of the problem. Therefore, the proposed solution
procedures must tackle the problem without affecting both the mathematical form of the model and the
design variables. Many solution procedures, such as the ones based on the Genetic Algorithm, require
the discretization of the wind farm site, thus limiting the available locations where WTs can be erected,
which in turn impacts the solution quality of the problem as demonstrated in [242,259]. Although
it has been suggested to use irregular, finer and non-uniform discretization grids to improve these
Energies 2014, 7 6989

limitations [242], the selection of the type, size and refinement of the grid is equally hard as selecting
the WT individual locations over a continuous space, thus denoting an important area of improvement
for such solution procedures. In addition, most of the optimization procedures reported in the literature
have assumed a constant number of WTs or a maximum number of WTs, justified by the assumed
limited investment capacity of the wind farm designer. The assumption, however, is very unrealistic,
since the financial resources on the worldwide market are very large. In practice, what limits the size
of a wind farm (in terms of installed capacity) are practical considerations, such as the maximum
amount of energy that can be sold off in an offtake agreement, the limited site space, and very often the
transmission capacity. Other approaches, in which the discretization of the wind farm site is considered,
have linked the fixed available WT locations with the number of WTs, thus reducing the number of
design variables with the implication of restricting the possible wind farm layouts and thus the solution
quality of the problem, as previously discussed. The determination of the optimal number of WTs,
assuming no predefined investment capacity or any linkage to the process of domain discretization, is
complex and the best procedures perform a sensitivity analysis to determine its optimum value.
However, sensitivity analyses may require long computation times and will depend on the effectiveness
of the solution procedure to find optimal or near-optimal solutions for scenarios having different
predefined number of WTs. It should be noted that optimized wind farm layouts having different
amount of WTs can be significantly different, thus hindering the determination of the optimal number
of WTs based on solutions to scenarios having different number of WTs. In this regard, the development
of adaptive algorithms that include the number of WTs as a design variable is required.
Most of the studied wind farm scenarios assumed a square boundary. The reason behind this trend
is related to the ability of the proposed solution procedures to handle restrictions. As the boundary
shape complexity is increased, the complexity of the wind farm model constraints increase, thus affecting
the performance of the proposed solution procedures. In order to deal with this issue, a novel
multi-boundary constraint model, which is based on the optimal design of polygonal geometries used
to approximate the shape of complex boundaries, and the incorporation of a ray intersection method,
used to determine if WTs are located on feasible locations, was recently developed by Gu and
Wang [361]. The robust model of Gu and Wang may be successfully implemented in different
optimization frameworks, thus enhancing the constraints management.
The solution procedures proposed in the literature have been able to find better solutions than the
ones proposed by Mosetti et al. [28], traditional rules-of-thumb (e.g., equally spaced or grid-like siting
schemes) and initial random solutions, but none of them have been able to assess the quality of the
solution found (i.e., no upper bound of the objective value, with the exception of the objective value if
no wake effects are taken into account, has been provided). Therefore, it is inadequate to assume that
the proposed algorithms reached optimal solutions without knowing the shape and the objective value
of optimal solution. As discussed in Section 2.5, optimal solutions of the WFDO problem can be verified
only for small instances, and further comparisons between solution procedures, assuming larger scenarios,
can be performed only in a relative basis, if no other information is available or can be inferred about
the optimal solution.
None of the reviewed works discussed how the algorithm-driving parameters were tuned. A poor
tuning process can significantly affect the performance of the tested algorithms, as discussed in
Section 3.3.2. In addition, none of the authors of the reviewed works reported the average efficiency of
Energies 2014, 7 6990

the tested algorithm. It is well known that heuristic and metaheuristic algorithms require several
executions to reach near-optimal solutions. In this regard, statistical design of experiments could
provide useful information to determine the average efficiency of the tested solution procedures and to
determine the best value of the algorithm-driving parameters.
The complexity of the WFDO problem requires a multi-objective framework, in which the evaluation
of the performance metric(s) and all adverse factors can be performed systematically. The obvious
limitation of such framework is the natural non-parallelizable characteristics of the problem, which in
turn imply that a significant amount of time will be required to evaluate potential solutions throughout
the optimization procedure. However, detailed research on the performance improvements of incorporating
parallelism on the embedded optimization problems, such as the civil engineering and infrastructure
design, are yet to be performed.

4.4. Commercial Software

Software tools are now being used to optimize wind farms. The most widely used commercial
programs, including its strategies, are presented next.
The Wind Atlas Analysis and Application Program (WAsP) v11 [48], developed in 1987 by the
RISØ National Laboratory, is considered the industry standard software for wind resource assessment
and wind farm micro-siting. The software includes two different approaches for the wind resource
quantification and wind resource mapping. The first approach is based on the so-called Wind Atlas
methodology, which in turn is based on a set of linearized flow equations and is suitable for sites showing
modest terrain complexity and near uniform roughness conditions. The second approach is based on
the Reynolds-Averaged Navier-Stokes equations complemented with a standard k–ε turbulence model,
which is suitable to be used in complex terrain. Both approaches perform the procedures described in
Section 2.1 to obtain accurate estimations of the local wind field and rely heavily on the quality
and quantity of the measured data at the site. The WAsP software includes the possibility to
perform different types of analysis regarding the wind farm production, the wind farm efficiency
(incorporating the Park model, which is based on the Jensen-Katic wake model), wind data analysis,
wind climate estimation, and wind atlas generation. However, it lacks an optimization framework for
automated micro-siting. Nonetheless, upcoming versions of WAsP will incorporate the TOPFARM
framework [164–166,306,354] as a complementary module that will allow to perform systematic
optimization of wind farms.
WindPRO v2.9 [49] is a robust tool developed by EMD International A/S that incorporates the
WAsP engine for wind resource quantification. WindPro includes several different modules for the
simulation and quantification of the wind farm energy production (incorporating the WAsP Park model
to determine wake-based and turbulence-based energy losses) and environmental impacts (including
noise impact, shadow flicker, and visual impacts). In addition, it includes modules for the electrical
layout design and its optimization (including the quantification of power losses) and a robust financial
balance model. WindPro optimizes the wind farm layout using the annual energy production as a
performance metric while accounting for noise thresholds. However, shadow flicker, visual impacts
and financial prospections are not accounted during the optimization procedure. The optimization is
based on a pre-calculated wind resource map made up from the WAsP engine, which accounts for both
Energies 2014, 7 6991

the topography- and the roughness-based induced effects. Different types of optimization processes,
based on greedy heuristics, can be performed: (1) a stochastic and gradual addition of WTs into the
wind farm (without changing the preexisting WT positions) until no more WTs can be placed or a
predefined maximum number of WTs is reached; (2) an iterative modification of the wind farm layout
(by changing WT separation distances assuming a staggered or grid-like siting scheme); (3) an iterative
filling up of the available land with WTs, resulting in a high dense WT cluster; or (4) optimization of
noise operation modes for fixed wind farm layouts. The optimization processes depend on a set of
predefined parameters such as the minimum distance between WTs, a predefined installed capacity per
unit area and predefined restricted zones.
The WindFarmer v5.2 [50] tool, developed by DNV-GL Garrad Hassan, is similar to WindPro, as it
incorporates the WAsP engine for wind resource quantification and offers a set of different modules
for the design and optimization of wind farms. It includes the possibility of perform analysis regarding
the energy production, noise impacts, visual impacts, shadow flicker, turbulence effects, electric system
design, uncertainty modeling and financial balance prospections. Wake losses are calculated with the
use of the DNV GL’s CFD model, which is based on the Reynolds-Averaged Navier-Stokes equations
complemented with an eddy viscosity turbulence closure. The optimization procedures are based on
greedy heuristics, which can either maximize the annual energy production or the financial balance of
the project. The software includes the possibility of adding a customized financial model and actively
updates the selected performance metric during the optimization process. The optimization process
includes environmental impacts (noise, visual impact, shadow flicker and turbulence intensity) as
constraints with predefined thresholds.
Meteodyn WT v4 [51] is a CFD-based tool for wind resource assessment and wind farm design.
The wind flow modeling approach is based on the Reynolds-Averaged Navier-Stokes equations, which
are complemented with a one-equation turbulence model (i.e., the turbulent kinetic energy equation).
The software is able to assess the annual energy production of the wind farm while accounting for
wake effects (which are modeled using the Jensen-Katic wake model). However, it currently lacks an
optimization framework for automated micro-siting as well as of additional modules to perform
environmental impact and uncertainty assessments.
Wind Farm v4.2 [52] is a general tool, made by ReSoft, for the analysis, design and optimization of
wind farms. The software is capable of performing wind resource assessments based on the linearized
steady-state flow model MS-Micro/3 (having the same technical background as WAsP), developed by
the Atmospheric Environment Service of Environment Canada and Zephyr, and predict long term
variations in wind speed using the Measure-Correlate-Predict (MCP) method. The wind farm energy
yield is estimated based on pre-constructed wind maps while accounting for wake losses, which are
modeled using different wake models including the Ainslie and the WAsP Park models, and includes
additional losses caused by turbulence. Similarly to WindFarmer and WindPro, Wind Farm is capable
of performing noise estimations (based on the ISO 9613 standard [315]), as well as visual and shadow
flicker assessments. The optimization module of WindFarm determines the optimum layout of WTs
based on a pre-defined set of constraints such as the WT minimum separation, separation from houses,
noise level and ground slope. The performance metric can be either the AEPFarm or the CoE
(considering a simple form of the capital costs as the value to be minimized) of the wind farm. The
optimization strategy is based on a greedy local search heuristic. The optimization process is
Energies 2014, 7 6992

repeatedly restated in order to avoid local minima or maxima. At each restart, each turbine might be
moved randomly within a specified movement limit, which is reduced by a specified step size on each
restart until the movement limit is reduced to zero.
WindSim v6.2 [47] is a CFD-based tool for wind resource assessment and wind farm design and
optimization. The wind flow modeling approach is based on the Reynolds-Averaged Navier-Stokes
equations, which are complemented with different kind of turbulence models (e.g., standard k–ε,
RNG k–ε and the standard k–ω). The wind farm energy yield is estimated based on pre-constructed
wind maps while accounting for wake losses (which can be quantified using different wake models
such as the Jensen-Katic [23], Larsen [25] or the Ishihara [288] models). The objective performance
metric of the proposed optimization framework is the Net Present Value (NPV) of the project.
The optimization procedure aims to define the best location of a predefined number of WTs while
accounting for multiple restrictions such as the minimum distance between WTs, terrain inclination,
flow inclination, effective turbulence intensity (which is estimated from both the measured turbulence
intensity at the wind farm site and the simulated turbulent kinetic energy), wind shear, and extreme
winds (which are estimated by the method of Independent Storms) [362]. The optimization process is
repeated while varying the number of WTs (N) in the range 1 ≤ 𝑁 ≤ 𝑁𝑚𝑎𝑥 , where Nmax is a predefined
maximum number of WTs. The output of the overall optimization process is a concave curve
describing the NPV variation of the project as a function of the number of installed WTs, which in turn
define the best number of WTs and the best layout design. Two different algorithms may be selected to
perform the optimization task; the local search Simulated Annealing metaheuristic algorithm, which
may or may not take into account wake effects, modeled by the Jensen–Katic model, during the
optimization process; or a local search calculus-based algorithm that optimizes the wind farm layout
while considering a relaxed mixed integer nonlinear model. The optimization process ignores the
civil engineering and electrical layout design and any other environmental impacts during the
optimization process.
AWS Truepower OpenWind® Enterprise [46,363] software is a general GIS tool for wind resource
assessment (based on a mass-conservation wind flow model) and wind farm micro-siting optimization
(which include both wake and turbulence effects in energy conversion, which are modeled using
different engineering wake and turbulence models). OpenWind® implements either the AEPFarm or the
CoE as objective performance metric during the wind farm micro-siting optimization, in which the
decision variables are the number and position of WTs and the civil engineering (auxiliary roads) and
the electrical infrastructure layout design, while considering existing infrastructure and multiple
physical constraints. The software includes the possibility of performing shadow flicker, visual impact,
noise impact (based on the ISO 9613 [315]), uncertainty and control (e.g., WT scheduling)
assessments. Such assessments, however, are not considered during the optimization process.
OpenWind’s solution procedure is based on an iterative non-deterministic greedy heuristic algorithm,
in which the WT coordinates are modified by random Gaussian-distributed perturbations. The number of
WTs is updated iteratively depending on predefined WT capacity factor thresholds. A set of different
heuristic rules are adopted to obtain feasible solutions during the optimization process (i.e., a set of
restrictions are given to the program, which in turn define feasible zones in which WTs can
be located).
Energies 2014, 7 6993

4.5. Misconceptions

Several misconceptions that have been propagating in the literature were identified and are
summarized next:
 Different denominations have been arbitrarily given to the engineering wake models
(e.g., wind farm models, flow models and linear flow models) and have been used indistinctly
through the literature. Generally, wind farm models refers to holistic models that take into
account not only wake effects but other factors such as the ones described in Section 2.2 as
well. Flow models are commonly referred to as models based on the fundamental laws of fluid
mechanics, which quantify the fluid-fluid or fluid-structure interactions. Although wake models
may quantify the interactions between different layers of a fluid flow, they are less fundamental
than the laws of fluid mechanics. Thus, clear distinctions must be provided to avoid generating
confusion. The Jensen-Katic wake model is commonly referred to as a linear wake model.
This denomination is given since the wake expansion is assumed to be a linear function of
downwind distance from the WT. This, however, introduces a misleading concept about the
mathematical complexity of the model. The Jensen-Katic wake model is a non-linear model,
Equation (39), that quantifies the wind speed evolution of the wake of WT.
 Different performance metrics were adopted in the literature, including “the cost of
power” [28,232]. As pointed out by Tesauro et al. [169], the denomination “cost per KW of
power produced” contains two logical errors. (1) The total cost of a wind farm, composed of
fixed and variable costs, is commonly calculated on a 20-year baseline and is therefore based on
integral quantity, while the wind farm power [kW], defined as the energy converted per unit of
time, is a differential quantity by nature. The ratio between the two is meaningless. The Annual
Energy Production (AEP) on the contrary, is defined as the integral of power over a period of
one year and is measured in [kWh/y]. Thus, the ratio of the wind farm total costs and the AEP
(i.e., the cost per kWh of energy) and has an adequate physical meaning and is referred as the
Cost of Energy; (2) The denomination “power production of the wind farm” is against the first
law of thermodynamics, since energy cannot be produced or created, thus and adequate
denomination would be “the power conversion of the wind farm”.
 A linear combination of objective functions, involving different type of variables, is numerically
possible but is inadequate due to misleading dimensioning. Examples of linear combination of
performance metrics are given in [28,239].
 The computational complexity theory aims to classify problems according to their type and
complexity. Although optimization problems are the most treated problems in the literature,
other different kind of problems exist such as functional problems or decision problems.
Functional problems are problems that produce an output for every given input. Both the type
of inputs and outputs can be significantly different, including real or integer numbers, strings,
or even abstract objects. Decision problems are problems that return a Boolean answer for a
given set of input parameters. Depending on the type of problem treated and its time or space
complexity, different complexity denominations are given. Decision problems that can be
evaluated by a non-deterministic Turing machine in polynomial time are referred as problems
with NP complexity. Optimization problems that can be evaluated by a non-deterministic
Energies 2014, 7 6994

Turing machine in polynomial time are referred as problems with NPO complexity. Functional
problems that can be evaluated by a non-deterministic Turing machine in polynomial time are
referred as problems with FNP complexity. In the literature, it is common to find inadequate
complexity denominations (e.g., the WFDO problem has been referred as an NP-hard optimization
problem [56,152–154]), which in turn are commonly based on subjective reasoning rather than
in formal demonstrations based on suitable reduction methods, as described in Section 2.5.

5. Conclusions

This work presented an exhaustive survey of the state of the art in the wind farm design and
optimization problem. After almost four decades of research, a significant amount of contributions of
variable complexity has been reported in the literature, which have enhanced both the understanding of
the physical process of wind energy conversion and the techno-economical modeling of wind farms.
This topic is currently gaining popularity in the industry, because of the potential economic improvements
it offers. However, the applicability of the proposed optimization frameworks in the design of real
wind farm projects is currently limited by the modeling complexity of the different physical processes
involved during the wind energy conversion, by the current optimization strategies, and by the available
computational resources. In addition, the slow inclusion of physics and the use of idealistic scenarios
to perform comparisons between different implementations is limiting the applicability of the developed
approaches in the design of real wind farms.
The most important research trend in the WFDO problem is focused on improving the formulations
describing the expected energy conversion of the wind farm, the WT wakes and turbulence impacts,
the WT structural fatigue and degradation, the various types of environmental impacts, the overall
electric system losses and reliability, and the uncertainty and risk management. New holistic models
are required to improve the wind farm performance modeling and its optimization. In addition,
optimization frameworks must encompass all the design variables during the micro-siting process,
since current existing approaches have limited the number of design variables and their degrees of
freedom, thus significantly affecting the potential benefits. The specific research agenda in WFDO is
summarized next:
- Inclusion of all state-of-the-art sub-models, describing all the adverse factors, and design
variables in an all-encompassing procedure.
- Sensitivity analysis of the effects the selection of different objective performance metrics
produce in the resultant optimized wind farms.
- Sensitivity analysis of the effects that each adverse factor produces in the optimized wind
farms, while considering realistic scenarios.
- Systematic evaluation of different solution procedures, each solution procedure being fine-tuned.
- Investigation of the effects the different uncertainties produce in the optimized wind farms.
- Improvement of sub-models, following the observations given in Section 4.
- Inclusion of control aspects, systematic optimization of multiple wind farms and modeling of
the effects of potential conflict of interests.
Energies 2014, 7 6995

Acknowledgments

Support from (1) the EUDP FarmOpt Project J.nr. 64013-0405, (2) the Research Chair for Wind
Energy at Tecnológico de Monterrey, under contract CAT158, and (3) the Mexican Council for
Science and Technology (CONACyT) is gratefully acknowledged.

Author Contributions

All authors were involved during the development of the manuscript. José F. Herbert and
Oliver Probst performed the literature review and writing of Sections 1, 2.1–2.5. José F. Herbert,
Pierre-Elouan Réthoré, and Gunner Chr. Larsen performed the literature review and writing of
Sections 3.1 and 3.2. JoséF. Herbert and Krystel K. Castillo performed the literature review and writing
of Section 3.3. Finally, all authors participated in the discussion and writing of Sections 4 and 5.

Conflicts of Interest

The authors declare no conflict of interest.

References

1. Global Wind Statistics 2012; Global Wind Energy Council (GWEC): Brussels, Belgium, 2013;
pp. 1–4.
2. Global Wind Report 2012; Global Wind Energy Council (GWEC): Brussels, Belgium, 2013;
pp. 1–72.
3. AWEA Wind Industry Market Reports for 2013; American Wind Energy Association (EWEA):
Washington, DC, USA, 2013.
4. Templin, R.J. An Estimate of the Interaction of Windmills in Widespread Arrays; Laboratory
Report LTR-LA-171; National Aeronautical Establishment: Ottawa, ON, Canada, 1974.
5. Crafoord, C. An Estimate of the Interaction of a Limited Array of Windmills; Report DM-16;
Department of Meteorology, University of Stockholm: Stockholm, Sweden, 1975; Volume 77.
6. Newman, B.G. The spacing of wind turbines in large arrays. Energy Convers. 1977, 16, 169–171.
7. Moore, D.J. Depletion of available wind power by a large network of wind generators.
In International Conference on Future Energy Concepts; Institution of Electrical Engineers:
London, UK, 1979; pp. 302–305.
8. Milborrow, D.J. The performance of arrays of wind turbines. J. Wind Eng. Ind. Aerodyn. 1980, 5,
403–430.
9. Lipman, N.H.; Bossanyi, E.A.; Dunn, P.D.; Musgrove, P.J.; Whittle, G.E.; Maclean, C.
Fluctuations in output from wind turbine clusters. Wind Eng. 1980, 4, 1–7.
10. Bossanyi, E.A.; Maclean, C.; Whittle, G.E.; Dunn, P.D.; Lipman, N.H.; Musgrove, P.J.
The efficiency of wind turbine clusters. In Proceedings of the Third International Symposium on
Wind Energy Systems, Lyngby, Denmark, 26–29 August 1980; pp. 1–16.
11. Voutsinas, S.G.; Rados, K.G. A method for the aerodynamically optimal design of wind parks.
In Proceedings of the 15th Wind Energy Conference, Wind Energy Conversion, London, UK,
6–9 October 1993; pp. 231–238.
Energies 2014, 7 6996

12. Patel, M.R. Wind and Solar Power Systems, 1st ed.; CRC Press LLC: Boca Raton, FL, USA,
1999; p. 1–350.
13. Bansal, R.; Bhatti, T.; Kothari, D. On some of the design aspects of wind energy conversion
systems. Energy Convers. Manag. 2002, 43, 2175–2187.
14. Ammara, I.; Leclerc, C.; Masson, C. A viscous three-dimensional differential/actuator-disk
method for the aerodynamic analysis of wind farms. J. Sol. Energy Eng. 2002, 124, 345–356.
15. Jangamshetti, S.H.; Ran, V.G. Optimum siting of wind turbine generators. IEEE Trans.
Energy Convers. 2001, 16, 8–13.
16. Xie, J.; Qui, Y. Siting of large wind farm and optimum wind turbine-site matching.
Acta Energ. Sol. Sin. 2001, 4, 466–472.
17. Kiranoudis, C.T.; Voros, N.G.; Maroulis, Z.B. Short-cut design of wind farms. Energy Policy
2001, 29, 567–578.
18. Vermeulen, P. A Wind Tunnel Study of the Wake of a Horizontal Axis Wind Turbine; Netherlands
Organisation for Applied Scientific Research (TNO): Apeldoorn, The Netherlands, 1978; pp. 1–19.
19. Lissaman, P.B.S. Energy effectiveness of arbitrary arrays of wind turbines. J. Energy 1979, 3,
323–328.
20. Vermeulen, P. An experimental analysis of wind turbine wakes. In Proceedings of the 3rd
International Symposium on Wind Energy Systems, Lyngby, Denmark, 26–29 August 1980;
pp. 431–450.
21. Jensen, N.O. A Note on Wind Generator Interaction; RisøNational Laboratory: Roskilde, Denmark,
1983; pp. 1–17.
22. Crespo, A.; Manuel, F.; Moreno, D.; Fraga, E.; Hernández, J. Numerical analysis of wind turbine
wakes. In Delphi Workshop on Wind Energy Applications; Bergeles, G., Chadjivassiliadis, J., Eds.;
European Wind Energy Association: Delphi, Greece, 1985; pp. 15–25.
23. Katic, I.; Højstrup, J.; Jensen, N.O. A simple model for cluster efficiency. In Proceedings of
the European Wind Energy Conference and Exhibition 1986, Rome, Italy, 7–9 October 1986;
pp. 407–410.
24. Lissaman, P.B.S.; Gyatt, G.W.; Zalay, A.D. Numeric Modeling Sensitivity Analysis of the
Performance of Wind Turbine Arrays; Pacific Northwest Laboratory: Richland, WA, USA, 1982;
pp. 1–122.
25. Larsen, G.C. A Simple Wake Calculation Procedure; Report Risø-M-2760; Risø National
Laboratory: Roskilde, Denmark, 1988; pp. 1–53.
26. Ainslie, J.F. Calculating the flowfield in the wake of wind turbines. J. Wind Eng. Ind. Aerodyn.
1988, 27, 213–224.
27. Kaminsky, F.C.; Kirchhoff, R.H.; Sheu, L.-J. Optimal spacing of wind turbines in a wind energy
power plant. Sol. Energy 1987, 39, 467–471.
28. Mosetti, G.; Poloni, C.; Diviacco, B. Optimization of wind turbine positioning in large windfarms
by means of a genetic algorithm. J. Wind Eng. Ind. Aerodyn. 1994, 51, 105–116.
29. International Electrotechnical Commission (IEC) 61400 Standard, Subclasses 1–27.
Available online: http://www.iec.ch/ (accessed on 18 March 2014).
30. European Wind Energy Association (EWEA). European Best Practice Guidelines for Wind
Energy Development; European Wind Energy Association: Paris, France, 2002; pp. 1–26.
Energies 2014, 7 6997

31. Wind Energy Siting Handbook, 1st ed.; American Wind Energy Association (EWEA):
Washington, DC, USA, 2008; pp. 1–183.
32. American Wind Energy Association. Available online: http://www.awea.org/ (accessed on
18 March 2014).
33. Best Practice Guidelines for Wind Energy Development; The British Wind Energy Association:
London, UK, 1994; pp. 1–27.
34. Global Wind Energy Council (GWEC). Global Wind Reports 2006–2012. Available online:
http://www.gwec.net/publications/global-wind-report-2/ (accessed on 18 March 2014).
35. Sustainable Energy Authority of Ireland (SEAI). How to Develop a Wind Farm. Available
online: http://www.seai.ie/Renewables/Wind_Energy/Wind_Farms/Wind_Farm_Development/
How_to_develop_a_wind_farm/ (accessed on 18 March 2014).
36. Burton, T.; Sharpe, D.; Jenkins, N.; Bossanyi, E. Wind Energy Handbook; John Wiley & Sons, Ltd.:
New York, NY, USA, 2001.
37. Rodman, L.C.; Meentemeyer, R.K. A geographic analysis of wind turbine placement in Northern
California. Energy Policy 2006, 34, 2137–2149.
38. Punt, M.J.; Groeneveld, R.A.; van Ierland, E.C.; Stel, J.H. Spatial planning of offshore wind farms:
A windfall to marine environmental protection? Ecol. Econ. 2009, 69, 93–103.
39. Lee, K.-H.; Jun, S.-O.; Pak, K.-H.; Lee, D.-H.; Lee, K.-W.; Park, J.-P. Numerical optimization of
site selection for offshore wind turbine installation using genetic algorithm. Curr. Appl. Phys.
2010, 10, S302–S306.
40. Van Haaren, R.; Fthenakis, V. GIS-based wind farm site selection using spatial multi-criteria
analysis (SMCA): Evaluating the case for New York State. Renew. Sustain. Energy Rev. 2011,
15, 3332–3340.
41. Ghaderi, S.F.; Zakerinia, M.S.; Vaezi, E. Optimal placement of wind turbine farms in Iran using
new ranking method. In Proceedings of the 2012 Second Iranian Conference on Renewable
Energy and Distributed Generation (ICREDG), Tehran, Iran, 6–8 March 2012; pp. 132–136.
42. Probst, O.; Cárdenas, D. State of the art and trends in wind resource assessment. Energies 2010,
3, 1087–1141.
43. Belu, R.; Koracin, D. Statistical and spectral analysis of wind characteristics relevant to wind
energy assessment using tower measurements in complex terrain. J. Wind Energy 2013, 2013, 1–12.
44. Bailey, B.H.; McDonald, S.L.; Bernadett, D.W.; Markus, M.J.; Elsholz, K.V. Wind Resource
Assessment Handbook: Fundamentals for Conducting a Successful Monitoring Program;
TAT-5–15283–01; National Renewable Energy Laboratory: Golden, CO, USA, 1997; pp. 1–79.
45. Brower, M.C. Wind Resource Assessment: A Practical Guide to Developing a Wind Project,
1st ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2012; p. 296.
46. AWS Truepower OpenWind. Available online: http://www.awsopenwind.org/ (accessed on
25 February 2014).
47. WindSim. Available online: http://www.windsim.com/ (accessed on 25 February 2014).
48. Wind Atlas Analysis and Application Program (WAsP). Available online: http://www.wasp.dk/
(accessed on 25 February 2014).
49. EMD WindPro. Available online: http://www.emd.dk/WindPRO/ (accessed on 25 February 2014).
Energies 2014, 7 6998

50. GL Garrad Hassan WindFarmer. Available online: http://www.gl-garradhassan.com/en/software/

GHWindFarmer.php (accessed on 25 February 2014).
51. Meteodyn Meteorology & Dynamics. Available online: http://meteodyn.com/en/ (accessed on
25 February 2014).
52. ReSoft WindFarm. Available online: http://www.resoft.co.uk/English/ (accessed on 25 February 2014).
53. Bechmann, A.; Sørensen, N.N.; Berg, J.; Mann, J.; Réthoré, P.-E. The bolund experiment,
Part II: Blind comparison of microscale flow models. Boundary-Layer Meteorol. 2011, 141,
245–271.
54. Chen, L.; MacDonald, E. Considering landowner participation in wind farm layout optimization.
J. Mech. Des. 2012, 134, 1–6.
55. Elkinton, C.N. Offshore Wind Farm Layout Optimization; University of Massachusetts Amherst:
Amherst, MA, USA, 2007; pp. 1–218.
56. Castro Mora, J. Optimización Global de Parques Eólicos Mediante Algoritmos Evolutivos;
Universidad de Sevilla: Sevilla, Spain, 2008; pp. 1–278. (In Spanish)
57. Larsen, G.C.; Madsen, H.A.; Larsen, T.J.; Troldborg, N. Wake Modeling and Simulation;
Risø National Laboratory for Sustainable Energy: Roskilde, Denmark, 2008; Volume 1653,
pp. 1–28.
58. Sanderse, B. Aerodynamics of Wind Turbine Wakes: Literature Review; Report ECN-E-09-016;
Energy Research Centre of the Netherlands: Petten, The Netherlands, 2009; pp. 1–46.
59. Vermeer, L.J.; Sørensen, J.N.; Crespo, A. Wind turbine wake aerodynamics. Prog. Aerosp. Sci.
2003, 39, 467–510.
60. Barthelmie, R.J.; Hansen, K.; Frandsen, S.T.; Rathmann, O.; Schepers, J.G.; Schlez, W.; Phillips, J.;
Rados, K.; Zervos, A.; Politis, E.S.; et al. Modelling and measuring flow and wind turbine wakes
in large wind farms offshore. Wind Energy 2009, 12, 431–444.
61. Barber, S.; Chokani, N.; Abhari, R.S. Wind turbine performance and aerodynamics in wakes
within wind farms. In Proceedings of the European Wind Energy Conference and Exhibition
2011, Brussels, Belgium, 14–17 March 2011.
62. Lange, B.; Waldl, H.-P.; Guerrero, A.G.; Heinemann, D.; Barthelmie, R.J. Modelling of offshore
wind turbine wakes with the wind farm program flap. Wind Energy 2003, 6, 87–104.
63. Dahlberg, J.-Å.; Thor, S.-E. Power performance and wake effects in the closely spaced lillgrund
wind farm. In Proceedings of the European Offshore Wind 2009 Conference and Exhibition,
Stockholm, Sweden, 14–16 September 2009; pp. 1–10.
64. Adaramola, M.S.; Krogstad, P.-Å. Wind tunnel simulation of wake effects on wind turbine
performance. In Proceedings of the European Wind Energy Conference and Exhibition 2010,
Warsaw, Poland, 20–23 April 2010; pp. 1–8.
65. Barthelmie, R.; Frandsen, S.; Jensen, L.; Mechali, M.; Réthoré, P.-E. Verification of an efficiency
model for very large wind turbine clusters. In Proceedings of the Copenhagen Offshore Wind
2005, Copenhagen, Denmark, 26–28 October 2005; pp. 1–9.
66. Barthelmie, R.J.; Rathmann, O.; Frandsen, S.T.; Hansen, K.S.; Politis, E.; Prospathopoulos, J.;
Rados, K.; Cabezón, D.; Schlez, W.; Phillips, J.; et al. Modelling and measurements of wakes in
large wind farms. J. Phys. Conf. Ser. 2007, 75, doi:10.1088/1742-6596/75/1/012049.
Energies 2014, 7 6999

67. Barthelmie, R.J.; Pryor, S.C.; Frandsen, S.T.; Hansen, K.S.; Schepers, J.G.; Rados, K.; Schlez, W.;
Neubert, A.; Jensen, L.E.; Neckelmann, S. Quantifying the impact of wind turbine wakes on
power output at offshore wind farms. J. Atmos. Ocean. Technol. 2010, 27, 1302–1317.
68. Barthelmie, R.J.; Jensen, L.E. Evaluation of wind farm efficiency and wind turbine wakes at the
Nysted offshore wind farm. Wind Energy 2010, 13, 573–586.
69. Seifert, H.; Kroning, J. Recommendations for spacing in wind farms. In Proceedings of the
European Wind Energy Conference and Exhibition 2003, Madrid, Spain, 17 June 2003; pp. 2–7.
70. Meyers, J.; Meneveau, C. Optimal turbine spacing in fully developed wind farm boundary layers.
Wind Energy 2012, 15, 305–317.
71. Højstrup, J. Spectral coherence in wind turbine wakes. J. Wind Eng. Ind. Aerodyn. 1999, 80,
137–146.
72. Chamorro, L.P.; Porté-Agel, F. A Wind-tunnel investigation of wind-turbine wakes: Boundary-layer
turbulence effects. Boundary-Layer Meteorol. 2009, 132, 129–149.
73. Frandsen, S.; Barthelmie, R.; Pryor, S.; Rathmann, O.; Larsen, S.; Højstrup, J. The Necessary
Distance between Large Wind Farms Offshore—Study; Report Risø-R-1518(EN); Risø National
Laboratory: Roskilde, Denmark, 2004; pp. 1–29.
74. Gelman, R. 2012 Renewable Energy Data Book; Meshek, M., Ed.; National Renewable Energy
Laboratory: Golden, CO, USA, 2012; pp. 1–128.
75. Morthorst, P.E.; Auer, H.; Garrad, A.; Blanco, I. Part III: The economics of wind power. In Wind
Energy—The Facts: A Guide to the Technology, Economics and Future of Wind Power;
Earthscan, European Wind Energy Association: London, UK, 2009; pp. 197–258.
76. Morthorst, P.E.; Jacobsen, H. Costs & Prices. In Wind Energy—The Facts; European Wind
Energy Association: Brussels, Belgium, 2004; Volume 2, pp. 93–110.
77. Hopewell, P.D.; Castro, F.; Bailey, D.I. Optimising the design of offshore wind farm collection
networks. In Proceedings of the 41st International Universities Power Engineering Conference,
Newcastle Upon Tyne, UK, 6–8 September 2006; pp. 84–88.
78. Lumbreras, S.; Ramos, A. Optimal design of the electrical layout of an offshore wind farm
applying decomposition strategies. IEEE Trans. Power Syst. 2013, 28, 1434–1441.
79. Pattanariyankool, S.; Lave, L.B. Optimizing transmission from distant wind farms. Energy Policy
2010, 38, 2806–2815.
80. Zhao, M.; Chen, Z.; Blaabjerg, F. Optimisation of electrical system for offshore wind farms via
genetic algorithm. IET Renew. Power Gener. 2009, 3, 1–12.
81. Zhao, M.; Chen, Z.; Hjerrild, J. Analysis of the behaviour of genetic algorithm applied in
optimization of electrical system design for offshore wind farms. In Proceedings of the 32nd Annual
Conference on IEEE Industrial Electronics (IECON 2006), Paris, France, 7–10 November 2006;
pp. 2335–2340.
82. Cheng, H.-P.; Yu, M.-T. Effect of the transmission configuration of wind farms on their capacity
factors. Energy Convers. Manag. 2013, 66, 326–335.
83. Gardner, P.; Garrad, A.; Jamieson, P.; Snodin, H.; Nicholls, G.; Tindal, A. Volume 1: Technology.
In Wind Energy—The Facts; European Wind Energy Association: Brussels, Belgium, 2004;
Volume 1, pp. 1–92.
Energies 2014, 7 7000

84. Nandigam, M.; Dhali, S.K. Optimal design of an offshore wind farm layout. In Proceedings of
International Symposium on Power Electronics, Electrical Drives, Automation and Motion 2008,
Naples, Italy, 11–13 June 2008; pp. 1470–1474.
85. Korytowski, M.J. Comparative Analysis of Medium Voltage DC and AC Network Infrastructure
Models; University of Pittsburgh: Pittsburgh, PA, USA, 2009; pp. 1–135.
86. Pappala, V.S.; Wilch, M.; Singh, S.N.; Erlich, I. Reactive power management in offshore wind
farms by adaptive PSO. In Proceedings of the 2007 International Conference on Intelligent
Systems Applications to Power Systems, Kaohsiung, Niigata, Japan, 5–8 November 2007; pp. 1–8.
87. Li, L.; Zeng, X.; Zhang, P. Wind farms reactive power optimization using genetic/tabu hybrid
algorithm. In Proceedings of the 2008 International Conference on Intelligent Computation
Technology and Automation (ICICTA), Hunan, China, 20–22 October 2008; pp. 1272–1276.
88. Chen, P.; Siano, P.; Chen, Z.; Bak-jensen, B. Optimal allocation of power-electronic interfaced
wind turbines using a genetic algorithm—Monte carlo hybrid optimization method. In Wind
Power Systems; Wang, L., Singh, C., Kusiak, A., Eds.; Springer: Berlin, Germany, 2010; pp. 1–23.
89. Kansal, S.; Sai, B.B.R.; Tyagi, B.; Kumar, V. Optimal placement of wind-based generation in
distribution networks. In Proceedings of the IET Conference on Renewable Power Generation
(RPG 2011), Edinburgh, UK, 6–8 September 2011; pp. 1–6.
90. Shafiullah, G.M.; Oo, A.M.T.; Ali, A.B.M.S.; Stojcevski, A. Influences of wind energy integration
into the distribution network. J. Wind Energy 2013, 2013, 1–22.
91. Zhao, M.; Chen, Z.; Blaabjerg, F. Probabilistic capacity of a grid connected wind farm based on
optimization method. Renew. Energy 2006, 31, 2171–2187.
92. World Health Organization (WHO). Energy, sustainable development and health. In Proceedings
of the Fourth Ministerial Conference on Environment and Health, Budapest, Hungary,
23–25 June 2004; pp. 1–122.
93. Colby, W.D.; Dobie, R.; Leventhall, G.; Lipscomb, D.M.; McCunney, R.J.; Seilo, M.T.;
Søndergaard, B. Wind Turbine Sound and Health Effects an Expert Panel Review;
American Wind Energy Association: Chicago, IL, USA, 2009; pp. 1–85.
94. Minnesota Department of Health. Public Health Impacts of Wind Turbines; Minnesota
Department of Health Environmental Health Division: Minneapolis, MI, USA, 2009; pp. 1–32.
95. The Potential Health Impact of Wind Turbines; Chief Medical Officer of Health (CMOH)
Report; Queen’s Printer for Ontario: Ontario, ON, Canada, 2010; pp. 1–14.
96. Mann, J.; Teilmann, J. Environmental impact of wind energy. Environ. Res. Lett. 2013, 8, 1–3.
97. United Nations Environment Programme. Environmental Impact Assestment. Available online:
http://www.unep.org/ (accessed on 18 March 2014).
98. Torrance, M. IFC Performance Standards on Environmental & Social Sustainability: A Guidebook;
LexisNexis Canada: Ontario, ON, Canada, 2012; p. 439.
99. Baidya Roy, S.; Traiteur, J.J. Impacts of wind farms on surface air temperatures. Proc. Natl.
Acad. Sci. USA 2010, 107, 1–6.
100. Zhang, W.; Markfort, C.D.; Porté-Agel, F. Experimental study of the impact of large-scale wind
farms on land–atmosphere exchanges. Environ. Res. Lett. 2013, 8, 1–9.
Energies 2014, 7 7001

101. Longridge, M.W. The Impact of Transmission Lines on Bird Flight Behaviour with reference to
Collision Mortality and System Reliability; Bird Research Committee, ESCOM: Johannesburg,
South Africa, 1986.
102. Clausager, N. Impact of Wind Turbines on Birds; National Environmental Research Institute:
Copenhagen, Denmark, 1995. (In Danish)
103. Montes, R.M.; Jague, L.B. Effects of Wind Turbine Power Plants on the Avifauna in the CAMPO
de Gibraltar Region; Sociedad Espanola de Ornitologia (SEO/BirdLife): Gibraltar, Spain, 1995.
104. Hunt, W.G.; Jackman, R.E.; Hunt, T.L.; Driscoll, D.E.; Culp, L. A Population Study of Golden
Eagle in the Altamont Pass Wind Resource Area: Population Trend Analyses 1997; University of
California: Santa Cruz, CA, USA, 1998.
105. Thelander, C.; Rugge, L. Avian Risk Behaviour and Fatalities at the Altamont Wind Resource
Area—March 1998 to February 1999; National Renewable Energy Laboratory: Golden, CO,
USA, 2000.
106. Clausen, N.-E.; Mortensen, N.G.; Hansen, J.C.; Clausager, I.; Jensen, F.P.; Georgy, L.; Said, U.S.
Wind Farm Planning at the Gulf of Suez; Report Risø-R-1387(EN); Risø National Laboratory:
Roskilde, Denmark, 2004; pp. 1–97.
107. Petersen, I.K.; Christensen, T.K.; Kahlert, J.; Desholm, M.; Fox, A.D. Final Results of Bird
Studies at the Offshore Wind Farms at Nysted and Horns Rev, Denmark; National Environmental
Research Institute: Roskilde, Denmark, 2006; pp. 1–166.
108. Global Consumer Wind Study 2012; Vestas: Copenhagen, Denmark, 2012; pp. 1–40.
109. Manchado, C.; Otero, C.; Gómez-Jáuregui, V.; Arias, R.; Bruschi, V.; Cendrero, A. Visibility
analysis and visibility software for the optimisation of wind farm design. Renew. Energy 2013,
60, 388–401.
110. Vissering, J.; Sinclair, M.; Margolis, A. A Visual Impact Assessment Process for Wind Energy
Projects; Clean Energy States Alliance: Montpelier, VT, USA, 2011; pp. 1–60.
111. Buchan, N. Visual Assessment of Windfarms: Best Practice; University of Newcastle: Edinburgh,
Scotland, 2002; pp. 1–79.
112. Sullivan, R.G.; Kirchler, L.B.; Lahti, T.; Roché, S.; Beckman, K.; Cantwell, B.; Richmond, P.
Wind Turbine Visibility and Visual Impact Threshold Distances in Western Landscapes; Argonne
National Laboratory: Chicago, IL, USA, 2012; pp. 1–47.
113. Van den Berg, F.; Pedersen, E.; Bouma, J.; Bakker, R. Visual and Acoustic Impact of Wind
Turbine Farms on Residents; European Platform Against Windfarm: Groningen, The Netherlands,
2008; pp. 1–99.
114. Harding, G.; Harding, P.; Wilkins, A. Wind turbines, flicker, and photosensitive epilepsy:
Characterizing the flashing that may precipitate seizures and optimizing guidelines to prevent
them. Epilepsia 2008, 49, 1095–1098.
115. Pinder, J.N. Mechanical noise from wind turbines. Wind Eng. 1992, 16, 158–168.
116. The Assessment and Rating of Noise from Wind Farms; Energy Technology Support Unit (ETSU):
Harwell, UK, 1997.
117. Moriarty, P.; Migliore, P. Semi-Empirical Aeroacoustic Noise Prediction Code for Wind Turbines;
National Renewable Energy Laboratory (NREL): Golden, CO, USA, 2003; pp. 1–39.
Energies 2014, 7 7002

118. A Critical Appraisal of Wind-Farm Noise Propagation; Energy Technology Support Unit (ETSU):
Harwell, UK, 2000.
119. Noise Guidelines for Wind Farms; Ministry of the Environment: Ontario, ON, Canada, 2008;
pp. 1–20.
120. Southampton, M. Noise from Wind Turbines; British Wind Energy Association: London, UK,
2000; pp. 1–4.
121. Rogers, A.L.; Manwell, J.F. Wind Turbine Noise Issues; University of Massachusetts at Amherst:
Boston, MA, USA, 2004; pp. 1–19.
122. Stiesdal, H.; Kristensen, E. Noise control on the BONUS 300 kW wind turbine. In Proceedings
of the 15th British Wind Energy Association Conference; Mechanical Engineering Publications:
Bury St Edmunds, UK, 1993; pp. 335–340.
123. Occupational Noise Exposure Standards; 1910.95; Occupational Safety & Health Administration
(OSHA): Washington, DC, USA, 1991.
124. Recommended Practice for Wind Turbine Testing 4. Acoustics—Measurements of Noise
Emission From Wind Turbines; International Energy Agency (IEA): Paris, France, 1994.
125. Recommended Practice for Wind Turbine Testing 10. Measurements of Noise Emission from Wind
Turbines at Noise Receptor Locations; International Energy Agency (IEA): Paris, France, 1997.
126. Tennant, A.; Chambers, B. Radar signature control of wind turbine generators. In Proceedings of
the 2005 IEEE Antennas and Propagation Society International Symposium, Washington, DC,
USA, 3–8 July 2005; pp. 489–492.
127. Krug, F.; Lewke, B. Electromagnetic Interference on Large Wind Turbines. Energies 2009, 2,
1118–1129.
128. Theil, A.; Schouten, M.W.; de Jong, A. Radar and wind turbines: A guide to acceptance criteria.
In Proceedings of the 2005 IEEE Antennas and Propagation Society International Symposium,
Washington, DC, USA, 3–8 July 2005; pp. 1355–1361.
129. Sengupta, D.L. Electromagnetic interference effects of wind turbines. In Annual Meeting of the
Working Committee on EMI; International Energy Association (IEA): Copenhagen, Denmark,
1984; pp. 1–19.
130. Sengupta, D.L. Electromagnetic interference from wind turbines. In IEEE Antennas and
Propagation Society International Symposium; Institute of Electrical and Electronics Engineers
(IEEE): Orlando, FL, USA, 1999; Volume 3, pp. 1984–1986.
131. Derrick, L.J. Cherry Tree Wind Farm Electromagnetic Interference Assessment—Investigation of
Possible Impacts on Broadcasting and Radiocommunication Services; Infigen Energy: Sydney,
Australia, 2012; pp. 1–36.
132. French, T. Ararat Wind Farm Electromagnetic Interference Assessment Report; Renewable
Energy Systems Ltd: Sydney, Australia, 2009; pp. 1–26.
133. Lackner, M.; Rogers, A.; Manwell, J. Uncertainty analysis in wind resource assessment and wind
energy production estimation. In Proceedings of the 45th AIAA Aerospace Sciences Meeting and
Exhibit, Reno, NV, USA, 8–11 January 2007.
134. Afanasyeva, S.; Saari, J.; Kukkonen, S.; Partanen, J.; Pyrhönen, O. Optimization of wind farm
design taking into account uncertainty in input parameters. In Proceedings of the European Wind
Energy Conference and Exhibition 2013, Vienna, Austria, 4–7 February 2013; pp. 1–10.
Energies 2014, 7 7003

135. González, J.S.; Payán, M.B.; Riquelme-Santos, J.M. Optimization of wind farm turbine layout
including decision making under risk. IEEE Syst. J. 2012, 6, 94–102.
136. Gaumond, M.; Réthoré, P.-E.; Ott, S.; Peña, A.; Bechmann, A.; Hansen, K.S. Evaluation of the
wind direction uncertainty and its impact on wake modeling at the Horns Rev offshore wind
farm. Wind Energy 2013, 17, 1169–1178.
137. Fingersh, L.; Hand, M.; Laxson, A. Wind Turbine Design Cost and Scaling Model; National
Renewable Energy Laboratory (NREL): Golden, CO, USA, 2006; pp. 1–43.
138. Chen, L.; Macdonald, E. Effects of uncertain land availability, wind shear, and cost on wind farm
layout. In Proceedings of the International Design Engineering Technical Conferences & Computers
and Information in Engineering Conference IDETC/CIE, Portland, OR, USA, 4–7 August 2013;
pp. 1–13.
139. Arora, S.; Barak, B. Computational Complexity: A Modern Approach, 1st ed.;
Cambridge University Press: New York, NY, USA, 2009; p. 594.
140. Rich, E.A. Automata, Computability and Complexity: Theory and Applications, 1st ed.;
Prentice Hall: Upper Saddle River, NJ, USA, 2007; p. 1120.
141. Allender, E.; Loui, M.C.; Regan, K.W. Complexity classes. In Algorithms and Theory of
Computation Handbook; Atallah, M.J., Ed.; CRC Press: Boca Raton, FL, USA, 1998; pp. 1–39.
142. Blum, M. A machine-independent theory of the complexity of recursive functions. J. ACM 1967,
14, 322–336.
143. Miller, R.E.; Thatcher, J.W.; Bohlinger, J.D. Complexity of Computer Computations;
Miller, R.E., Thatcher, J.W., Bohlinger, J.D., Eds.; Springer US: Boston, MA, USA, 1972.
144. Turing, A.M. Systems of Logic Based on Ordinals; Princeton University: Princeton, NJ, USA,
1939, pp. 1–110.
145. Savage, J.E. Models of Computation, 1st ed.; Addison-Wesley Longman Publishing Co., Inc.:
Boston, MA, USA, 1998; p. 672.
146. Ausiello, G.; Paschos, V.T. Reductions, completeness and the hardness of approximability.
Eur. J. Oper. Res. 2006, 172, 719–739.
147. Ausiello, G.; Crescenzi, P.; Protasi, M. Approximate solution of NP optimization problems.
Theor. Comput. Sci. 1995, 150, 1–55.
148. Orponen, P.; Mannila, H. On Approximation Preserving Reductions: Complete Problems and
Robust Measures; Report C-1987-28; University of Helsinki: Helsinki, Finland, 1987; pp. 1–14.
149. Bürgisser, P. Completeness and Reduction in Algebraic Complexity Theory; Algorithms and
Computation in Mathematics; Springer: Berlin, Germany, 2000; Volume 7, p. 168.
150. Cook, S.A. The complexity of theorem-proving procedures. In Proceedings of the Third Annual
ACM Symposium on Theory of Computing—STOC ’71, New York, NY, USA, 3–5 May 1971;
pp. 151–158.
151. Karp, R.M. Reducibility among Combinatorial Problems. In Complexity of Computer
Computations; Miller, R.E., Thatcher, J.W., Bohlinger, J.D., Eds.; Springer: New York, NY,
USA, 1972; pp. 85–103.
152. Ozturk, U.A.; Norman, B.A. Heuristic methods for wind energy conversion system positioning.
Electr. Power Syst. Res. 2004, 70, 179–185.
Energies 2014, 7 7004

153. Khan, S.A.; Rehman, S. Computational intelligence techniques for placement of wind turbines:
A brief plan of research in Saudi Arabian perspective. In Proceedings of the 2010 IEEE
International Energy Conference, Manama, Bahrain, 18–22 December 2010; pp. 519–523.
154. Khan, S.A.; Rehman, S. Iterative non-deterministic algorithms in on-shore wind farm design:
A brief survey. Renew. Sustain. Energy Rev. 2013, 19, 370–384.
155. Paz, A.; Moran, S. Non deterministic polynomial optimization problems and their approximations.
Theor. Comput. Sci. 1981, 15, 251–277.
156. Herbert-Acero, J.-F.; Franco-Acevedo, J.-R.; Valenzuela-Rendón, M.; Probst-Oleszewski, O.
Linear wind farm layout optimization through computational intelligence. In Proceedings of
the 8th Mexican International Conference on Artificial Intelligence 2009, Guanajuato, México,
9–13 November 2009; pp. 692–703.
157. Beyer, H.G.; Pahlke, T.; Schmidt, W.; Waldl, H.-P.; de Witt, U. Wake effects in a linear wind
farm. J. Wind Eng. Ind. Aerodyn. 1994, 51, 303–318.
158. Cuerva, A.; Sanz-Andrés, A. The extended Betz–Lanchester limit. Renew. Energy 2005, 30,
783–794.
159. Kusiak, A.; Song, Z. Design of wind farm layout for maximum wind energy capture. Renew. Energy
2010, 35, 685–694.
160. Chen, Y.; Li, H.; Jin, K.; Song, Q. Wind farm layout optimization using genetic algorithm with
different hub height wind turbines. Energy Convers. Manag. 2013, 70, 56–65.
161. Chen, K.; Song, M.X.; Zhang, X. The investigation of tower height matching optimization for
wind turbine positioning in the wind farm. J. Wind Eng. Ind. Aerodyn. 2013, 114, 83–95.
162. Schmidt, J.; Stoevesandt, B. Wind farm layout optimization with wakes from fluid dynamics
simulations. In Proceedings of the European Wind Energy Conference and Exhibition 2014,
Barcelona, Spain, 10–13 March 2014; pp. 28–32.
163. Tran, R.; Wu, J.; Denison, C.; Ackling, T.; Wagner, M.; Neumann, F. Fast and effective
multi-objective optimisation of wind turbine placement. In Proceedings of the Fifteenth Annual
Conference on Genetic and Evolutionary Computation Conference, Amsterdam, The Netherlands,
6–10 July 2013; pp. 1381–1388.
164. Réthoré, P.-E.; Fuglsang, P.; Larsen, T.J.; Buhl, T.; Larsen, G.C. Topfarm Wind Farm
Optimization Tool; Report Risø-R-1234(EN); Risø National Laboratory: Roskilde, Denmark,
2011; pp. 1–91.
165. Réthoré, P.-E.; Fuglsang, P.; Larsen, G.C.; Buhl, T.; Larsen, T.J.; Madsen, H.A. TopFarm:
Multi-fidelity optimization of offshore wind farm. In Proceedings of the 21th International
Offshore and Polar Engineering Conference, Maui, HI, USA, 19–24 June 2011; pp. 516–524.
166. Réthoré, P.-E.; Fuglsang, P.; Larsen, G.C.; Buhl, T.; Larsen, T.J.; Madsen, H.A. TOPFARM:
Multi-fidelity optimization of wind farms. Wind Energy 2013, 1–20, doi:10.1002/we.1667.
167. Fagerfjäll, P. Optimizing Wind Farm Layout—More Bang for the Buck Using Mixed Integer
Linear Programming; Chalmers University of Technology and Gothenburg University: Gothenburg,
Sweden, 2010; p. 111.
168. Bauer, J.; Lysgaard, J. The offshore wind farm array cable layout problem: A planar open vehicle
routing problem. J. Oper. Res. Soc. 2014, 2014, 1–15.
Energies 2014, 7 7005

169. Tesauro, A.; Réthoré, P.-E. State of the art of wind farm optimization. In Proceedings of the
European Wind Energy Conference and Exhibition 2012, Copenhagen, Denmark, 16–19 April 2012;
pp. 1–11.
170. Samorani, M. The wind farm layout optimization problem. In PowerLeeds School of Business;
University of Colorado at Boulder: Boulder, CO, USA, 2010; pp. 1–21.
171. Serrano González, J.; Burgos Payán, M.; Santos, J.M.R.; González-Longatt, F. A review and
recent developments in the optimal wind-turbine micro-siting problem. Renew. Sustain. Energy Rev.
2014, 30, 133–144.
172. Donovan, S. Wind farm optimization. In Proceedings of the 40th Annual Conference of the
Operational Research Society of New Zealand, Wellington, New Zealand, 2–3 December 2005;
pp. 196–205.
173. Donovan, S. An improved mixed integer programming model for wind farm layout optimisation.
In Proceedings of the 41st Annual Conference of the Operational Research Society of New
Zealand, Christchurch, New Zealand, 30 November–1 December 2006; pp. 143–152.
174. Donovan, S.; Nates, G.; Waterer, H.; Archer, R. Mixed integer programming models for wind
farm design. In Proceedings of MIP 2008 Workshop on Mixed Integer Programming, New York,
NY, USA, 4–7 August 2008; pp. 1–21.
175. Donovan, S.B. Whither the Wind Blows: Wind Flow Modelling and Wind Farm Layout
Optimisation; The University of Auckland: Auckland, New Zealand, 2008; p. 92.
176. Rivas, R.A.; Clausen, J.; Hansen, K.S.; Jensen, L.E. Solving the turbine positioning problem for
large offshore wind farms by simulated annealing. Wind Eng. 2009, 33, 287–297.
177. Wan, C.; Wang, J.; Yang, G.; Zhang, X. Optimal siting of wind turbines using real-coded genetic
algorithms. In Proceedings of the European Wind Energy Conference and Exhibition 2009,
Marseille, France, 16–19 March 2009; pp. 1–6.
178. Chowdhury, S.; Messac, A.; Zhang, J.; Castillo, L.; Lebron, J. Optimizing the unrestricted
placement of turbines of differing rotor diameters in a wind farm for maximum power
generation. In Proceedings of the ASME 2010 International Design Engineering Technical
Conferences & Computers and Information in Engineering Conference, Montreal, QC, Canada,
15–18 August 2010; pp. 775–790.
179. Kusiak, A.; Zheng, H. Optimization of wind turbine energy and power factor with an evolutionary
computation algorithm. Energy 2010, 35, 1324–1332.
180. Neubert, A.; Shah, A.; Schlez, W. Maximum yield from symmetrical wind farm layouts.
In Proceedings of the 10th German Wind Energy Conference, Bremen, Germany,
17–18 November 2010; pp. 1–4.
181. Rašuo, B.; Bengin, A. Optimization of wind farm layout. FME Trans. 2010, 38, 107–114.
182. Rašuo, B.; Bengin, A.; Veg, A. On aerodynamic optimization of wind farm layout.
In Proceedings of the 81st Annual Meeting of the International Association of Applied
Mathematics and Mechanics, Karlsruhe, Germany, 22–26 March 2010; pp. 539–540.
183. Sood, P.; Winstead, V.; Steevens, P. Optimal placement of wind turbines: A Monte Carlo
approach with large historical data set. In Proceedings of the IEEE International Conference on
Electro/Information Technology 2010, Normal, IL, USA, 20–22 May 2010; pp. 1–5.
Energies 2014, 7 7006

184. Wan, C.; Wang, J.; Yang, G.; Zhang, X. Optimal micro-siting of wind farms by particle swarm
optimization. In Advances in Swarm Intelligence; Tan, Y., Shi, Y., Tan, K.C., Eds.; Springer:
Berlin, Germany, 2010; Volume 6145, pp. 198–205.
185. Wan, C.; Wang, J.; Yang, G.; Zhang, X. Particle swarm optimization based on Gaussian mutation
and its application to wind farm micro-siting. In Proceedings of the 49th IEEE Conference on
Decision and Control, Atlanta, GA, USA, 15–17 December 2010; pp. 2227–2232.
186. Archer, R.; Nates, G.; Donovan, S.; Waterer, H. Wind turbine interference in a wind farm layout
optimization mixed integer linear programming model. Wind Eng. 2011, 35, 165–178.
187. Han, X.; Guo, J.; Wang, P. Adequacy study of a wind farm considering terrain and wake effect.
IET Gener. Transm. Distrib. 2012, 6, 1001–1008.
188. Han, X.; Guo, J.; Wang, P.; Jia, Y. Adequacy study of wind farms considering reliability and
wake effect of WTGs. In Proceedings of the 2011 IEEE Power and Energy Society General
Meeting, San Diego, CA, USA, 24–29 July 2011; pp. 1–7.
189. Tzanos, J.; Margellos, K.; Lygeros, J. Optimal wind turbine placement via randomized
optimization techniques. In Proceedings of the 17th Power Systems Computation Conference,
Stockholm, Sweden, 22–26 August 2011; pp. 1–8.
190. Wagner, M.; Neumann, F.; Kalyan, V.; O’Reilly, U.-M. Optimizing the layout of 1000 wind
turbines. In Proceedings of the European Wind Energy Conference and Exhibition 2011,
Brussels, Belgium, 14–17 March 2011.
191. Xu, C.; Yan, Y.; Liu, D.Y.; Zheng, Y.; Li, C.Q. Optimization of wind farm micro sitting based
on genetic algorithm. Adv. Mater. Res. 2011, 347–353, 3545–3550.
192. Chowdhury, S.; Zhang, J.; Messac, A.; Castillo, L. Unrestricted wind farm layout optimization
(UWFLO): Investigating key factors influencing the maximum power generation. Renew. Energy
2012, 38, 16–30.
193. Chowdhury, S.; Tong, W.; Messac, A.; Zhang, J. A mixed-discrete Particle Swarm Optimization
algorithm with explicit diversity-preservation. Struct. Multidiscip. Optim. 2012, 47, 367–388.
194. Christie, D.; Bradley, M. Optimising land use for wind farms. Energy Sustain. Dev. 2012, 16,
471–475.
195. Eroğlu, Y.; Seçkiner, S.U. Design of wind farm layout using ant colony algorithm. Renew. Energy
2012, 44, 53–62.
196. Grilli, A.; Spaulding, M.; O’Reilly C.; Potty, G. Offshore wind farm macro and micro siting
protocol application to Rhode Island. In Proceedings of the 33rd International Coastal Engineering
Conference, Santander, Spain, 1–6 July 2012; pp. 1–9.
197. Haugland, J.K.; Haugland, D. Computing the optimal layout of a wind farm. In Proceedings of
the NIK-2012 Conference, Bodø, Norway, 19–21 November 2012; pp. 93–104.
198. Haugland, J.K. Optimization Models for Turbine Location in Wind Farms; The University of
Bergen: Bergen, Norway, 2012.
199. Kwong, W.Y.; Zhang, P.Y.; Morgenroth, M.; Romero, D.; Amon, C.; Moran, J. Wind farm
layout optimization considering energy generation and noise propagation. In Proceedings of the
International Design Engineering Technical Conferences & Computers and Information in
Engineering Conference IDETC/CIE, Chicago, IL, USA, 12–15 August 2012; pp. 1–10.
Energies 2014, 7 7007

200. O’Reilly, C.M.; Grilli, A.R.; Potty, G.R. Offshore wind farm siting using a genetic algorithm.
In Proceedings of the International Conference on Green Technologies (ICGT), Trivandrum,
Kerala, India, 18–20 December 2012; pp. 208–214.
201. Wan, C.; Wang, J.; Yang, G.; Gu, H.; Zhang, X. Wind farm micro-siting by Gaussian particle
swarm optimization with local search strategy. Renew. Energy 2012, 48, 276–286.
202. Kwong, W.Y.; Zhang, Y.P.; Romero, D.; Moran, J.; Morgenroth, M.; Amon, C. Multi-objective
wind farm layout optimization considering energy generation and noise propagation with NSGA-II.
J. Mech. Des. 2014, 136, 1–11.
203. Chowdhury, S.; Zhang, J.; Messac, A.; Castillo, L. Optimizing the arrangement and the selection
of turbines for wind farms subject to varying wind conditions. Renew. Energy 2013, 52, 273–282.
204. Eroğlu, Y.; Seçkiner, S.U. Wind farm layout optimization using particle filtering approach.
Renew. Energy 2013, 58, 95–107.
205. Pérez, B.; Mínguez, R.; Guanche, R. Offshore wind farm layout optimization using mathematical
programming techniques. Renew. Energy 2013, 53, 389–399.
206. Rodrigues, S.M. F.; Bauer, P.; Pierik, J. Modular approach for the optimal wind turbine micro siting
problem through CMA-ES algorithm. In Proceedings of 25th Belgium–Netherlands Conference on
Artificial Intelligence, Delft, The Netherlands, 7–8 November 2013; pp. 1561–1568.
207. Salcedo-Sanz, S.; Gallo-Marazuela, D.; Pastor-Sánchez, A.; Carro-Calvo, L.; Portilla-Figueras, A.;
Prieto, L. Evolutionary computation approaches for real offshore wind farm layout: A case study
in Northern Europe. Expert Syst. Appl. 2013, 40, 6292–6297.
208. Song, M.X.; Chen, K.; He, Z.Y.; Zhang, X. Bionic optimization for micro-siting of wind farm on
complex terrain. Renew. Energy 2013, 50, 551–557.
209. Tong, W.; Chowdhury, S.; Mehmani, A.; Messac, A. Multi-objective wind farm design: Exploring
the trade-off between capacity factor and land use. In Proceedings of the 10th World Congress on
Structural and Multidisciplinary Optimization, Orlando, FL, USA, 19–24 May 2013; pp. 1–8.
210. Wagner, M.; Day, J.; Neumann, F. A fast and effective local search algorithm for optimizing the
placement of wind turbines. Renew. Energy 2013, 51, 64–70.
211. Wilson, D.; Veeramachaneni, K.; O’Reilly, U.M. Cloud scale distributed evolutionary strategies
for high dimensional problems. In Proceedings of the 16th European Conference EvoApplications,
Vienna, Austria, 3–5 April 2013; pp. 519–528.
212. Xu, C.; Yang, J.; Li, C.; Shen, W.; Zheng, Y.; Liu, D. A research on wind farm micro-sitting
optimization in complex terrain. In Proceedings of the International Conference on aerodynamics
of Offshore Wind Energy Systems and Wakes, Lyngby, Denmark, 17–19 June 2013; pp. 669–679.
213. Zhang, P.Y.; Romero, D.A.; Beck, J.C.; Amon, C.H. Solving wind farm layout optimization with
mixed integer programming and constraint programming. In Proceedings of the 10th International
Conference CPAIOR, Yorktown Heights, NY, USA, 18–22 May 2013; pp. 284–299.
214. Tindal, A.; Johnson, C.; LeBlanc, M.; Harman, K.; Rareshide, E.; Graves, A.M. Site-specific
adjustments to wind turbine power curves. In Proceedings of American Wind Energy Association
Conference 2008, Houston, TX, USA, 1–4 June 2008; pp. 1–11.
215. Sumner, J.; Masson, C. Influence of atmospheric stability on wind turbine power performance
curves. J. Sol. Energy Eng. 2006, 128, 531–538.
Energies 2014, 7 7008

216. Rinne, H. The Weibull Distribution: A Handbook, 1st ed.; CRC Press: Boca Raton, FL, USA,
2008; p. 808.
217. Casella, G.; Berger, R.L. Statistical Inference, 2nd ed.; Duxbury/Thomson Learning: Pacific
Grove, CA, USA, 2002; p. 660.
218. Zhang, J.; Chowdhury, S.; Messac, A.; Castillo, L. A multivariate and multimodal wind
distribution model. Renew. Energy 2013, 51, 436–447.
219. Lackner, M.A.; Elkinton, C.N. An analytical framework for offshore wind farm layout
optimization. Wind Eng. 2007, 31, 17–31.
220. Zhang, J.; Chowdhury, S.; Messac, A.; Castillo, L. Multivariate and multimodal wind distribution
model based on kernel density estimation. In Proceedings of the ASME 2011 5th International
Conference on Energy Sustainability, Washington, DC, USA, 7–10 August 2011.
221. Carlos, S.; Sánchez, A.; Martorell, S.; Marton, I. Onshore wind farms maintenance optimization
using a stochastic model. Math. Comput. Model. 2013, 57, 1884–1890.
222. Osadciw, L.A.; Yan, Y.; Ye, X.; Benson, G.; White, E. Wind Turbine diagnostics based on
power curve using particle swarm optimization. In Wind Power Systems; Wang, L., Singh, C.,
Kusiak, A., Eds.; Springer-Verlag: Berlin, Germany, 2010; pp. 151–165.
223. Tian, Z.; Ding, Y.; Ding, F. Maintenance optimization of wind turbine systems based on
intelligent prediction tools. In Innovative Computing Methods and Their Applications to Engineering
Problems; Nedjah, N., dos Santos Coelho, L., Cocco Mariani, V., de Macedo Mourelle, L., Eds.;
Springer-Verlag: Berlin, Germany, 2011; pp. 53–71.
224. Walford, C.A. Wind Turbine Reliability: Understanding and Minimizing Wind Turbine Operation
and Maintenance Costs; Report SAND2006-1100; Sandia National Laboratories: Livermore,
CA, USA, 2006; pp. 1–27.
225. Wu, Y.; Zhao, H. Optimization maintenance of wind turbines using Markov decision processes.
In Proceedings of the 2010 International Conference on Power System Technology, Hangzhou,
China, 24–28 October 2010; pp. 1–6.
226. Rogers, J.; Fink, S.; Porter, K. Examples of Wind Energy Curtailment Practices; National Renewable
Energy Laboratory: Golden, CO, USA, 2010.
227. Fink, S.; Mudd, C.; Porter, K.; Morgenstern, B. Wind Energy Curtailment Case Studies; Report
NREL/SR-550-46716; National Renewable Energy Laboratory: Golden, CO, USA, 2009.
228. Chiodo, E.; Lauria, D.; Pisani, C.; Villacci, D. Reliability aspects in wind farms design.
In Proceedings of 2012 International Symposium on Power Electronics, Electrical Drives,
Automation and Motion (SPEEDAM), Sorrento, Italy, 20–22 June 2012; pp. 577–581.
229. Tavner, P.J.; Xiang, J.; Spinato, F. Reliability analysis for wind turbines. Wind Energy 2007, 10,
1–18.
230. Zhao, M.; Chen, Z.; Blaabjerg, F. Optimization of electrical system for a large DC offshore wind
farm by genetic algorithm. In Proceedings of Nordic Workshop on Power and Industrial
Electronics, Trondheim, Norway, 14–16 June 2004; pp. 1–8.
231. Elkinton, C.N.; Manwell, J.F.; McGowan, J.G. Offshore wind farm layout optimization
(OWFLO) project: An introduction. In Proceedings of Copenhagen Offshore Wind Conference,
Copenhagen, Denmark, 26–28 October 2005.
Energies 2014, 7 7009

232. Grady, S.; Hussaini, M.Y.; Abdullah, M.M. Placement of wind turbines using genetic algorithms.
Renew. Energy 2005, 30, 259–270.
233. Elkinton, C.N.; Manwell, J.F.; Mcgowan, J.G. Offshore wind farm layout optimization (OWFLO)
project: Preliminary results. In Proceedings of the 44th AIAA Aerospace Sciences Meeting and
Exhibit, Reno, NV, USA, 11 January 2006; pp. 1–9.
234. Elkinton, C.N.; Manwell, J.; McGowan, J. Modeling the trade-offs in offshore wind energy
micrositing. In Proceedings of the WINDPOWER 2006 Conference and Exhibition, Pittsburgh,
PA, USA, 4–7 June 2006; pp. 1–12.
235. Elkinton, C.N.; Manwell, J.F.; Mcgowan, J.G. Optimization algorithms for offshore wind farm
micrositing. In Proceedings of the WINDPOWER 2007 Conference and Exhibition, Los Angeles,
CA, USA, 3–6 June 2007; pp. 1–23.
236. Elkinton, C.N.; Manwell, J.F.; McGowan, J.G. Optimizing the layout of offshore wind energy
systems. Mar. Technol. Soc. J. 2008, 42, 19–27.
237. Elkinton, C.N.; Manwell, J.F.; McGowan, J.G. Algorithms for offshore wind farm layout
optimization. Wind Eng. 2008, 32, 67–84.
238. Marmidis, G.; Lazarou, S.; Pyrgioti, E. Optimal placement of wind turbines in a wind park using
Monte Carlo simulation. Renew. Energy 2008, 33, 1455–1460.
239. Wu Yichun; Ding Ming Optimal choice of wind turbine generator based on Monte-Carlo method.
In Proceedings of the Third International Conference on Electric Utility Deregulation and
Restructuring and Power Technologies, Nanjing, China, 6–9 April 2008; pp. 2487–2491.
240. Wan, C.; Wang, J.; Yang, G.; Li, X.; Zhang, X. Optimal micro-siting of wind turbines by genetic
algorithms based on improved wind and turbine models. In Proceedings of the 48h IEEE
Conference on Decision and Control & 28th Chinese Control Conference, Shanghai, China,
16–18 December 2009; pp. 5092–5096.
241. Wang, F.; Liu, D.; Zeng, L. Modeling and simulation of optimal wind turbine configurations in
wind farms. In Proceedings of the 2009 World Non-Grid-Connected Wind Power and Energy
Conference, Nanjing, China, 24–26 September 2009; pp. 1–5.
242. Wang, F.; Liu, D.; Zeng, L. Study on computational grids in placement of wind turbines using
genetic algorithm. In Proceedings of the 2009 World Non-Grid-Connected Wind Power and
Energy Conference, Nanjing, China, 24–26 September 2009; Institute of Electrical and
Electronics Engineers (IEEE): Nanjing, China, 2009; Volume 2, pp. 1–4.
243. Aristidis, V.; Maria, P.; Christos, L. Particle Swarm Optimization (PSO) algorithm for wind farm
optimal design. Int. J. Manag. Sci. Eng. Manag. 2010, 5, 53–58.
244. Du Pont, B.L.; Cagan, J. An extended pattern search approach to wind farm layout optimization.
In Proceedings of the ASME International Design Engineering Technical Conferences &
Computers and Information in Engineering Conference IDETC/CIE 2010, Montreal, QC, Canada,
15–18 August 2010; pp. 1–10.
245. Emami, A.; Noghreh, P. New approach on optimization in placement of wind turbines within
wind farm by genetic algorithms. Renew. Energy 2010, 35, 1559–1564.
246. Li, X.; Wang, J.; Zhang, X. Equilateral-triangle mesh for optimal micrositing of wind farms.
In Proceedings of the 14th WSEAS international conference on Computers, Corfu Island,
Greece, 23–25 July 2010; pp. 187–195.
Energies 2014, 7 7010

247. Mittal, A. Optimization of the Layout of Large Wind Farms Using a Genetic Algorithm;
Case Western Reserve University: Cleveland, OH, USA, 2010, pp. 1–90.
248. Mustakerov, I.; Borissova, D. Wind turbines type and number choice using combinatorial
optimization. Renew. Energy 2010, 35, 1887–1894.
249. Rahmani, R.; Khairuddin, A.; Cherati, S.M.; Mahmoud Pesaran, H.A. A novel method for
optimal placing wind turbines in a wind farm using particle swarm optimization (PSO).
In Proceedings of the 9th International Power & Energy Conference, Singapore, Singapore,
27–29 October 2010; pp. 134–139.
250. Şişbot, S.; Turgut, Ö.; Tunç, M.; Çamdalı, Ü. Optimal positioning of wind turbines on Gökçeada
using multi-objective genetic algorithm. Wind Energy 2010, 13, 297–306.
251. Szafron, C. Offshore windfarm layout optimization. In Proceedings of the 9th International
Conference on Environment and Electrical Engineering (EEEIC), Prague, Czech Republic,
16–19 May 2010; pp. 542–545.
252. Amer, E.R.; Hamed, F.A. Optimal study design for Wind farm in Arwad Island. Energy Procedia
2011, 6, 721–735.
253. Chowdhury, S.; Zhang, J.; Messac, A.; Castillo, L. Characterizing the influence of land
configuration on the optimal wind farm performance. In Proceedings of the International Design
Engineering Technical Conferences (IDETC) and Computers and Information in Engineering
Conference (CIE), Washington, DC, USA, 28–31 Agust 2011; pp. 1–12.
254. Changshui, Z.; Guangdong, H.; Jun, W. A fast algorithm based on the submodular property for
optimization of wind turbine positioning. Renew. Energy 2011, 36, 2951–2958.
255. Ituarte-Villarreal, C.M.; Espiritu, J.F. Optimization of wind turbine placement using a viral based
optimization algorithm. Procedia Comput. Sci. 2011, 6, 469–474.
256. Ituarte-villarreal, C.M.; Espiritu, J.F. Wind turbine placement in a wind farm using a viral based
optimization. In Proceedings of the 41st International Conference on Computers & Industrial
Engineering, Los Angeles, CA, USA, 23–26 October 2011; pp. 672–677.
257. Lazarou, S.; Vita, V.; Ekonomou, L. Application of Powell’s optimisation method for the optimal
number of wind turbines in a wind farm. IET Sci. Meas. Technol. 2011, 5, 77–80.
258. Mustakerov, I.; Borissova, D. Wind Park Layout Design Using Combinatorial Optimization;
Al-Bahadly, I., Ed.; InTech: Rijeka, Croatia, 2011; pp. 1–22.
259. Wang, J.; Li, X.; Zhang, X. Genetic optimal micrositing of wind farms by equilateral-triangle
mesh. In Wind Turbines; Al-Bahadly, I., Ed.; InTech: Rijeka, Croatia, 2011; pp. 425–436.
260. Bonanni, A.; Conan, B. Wind farm optimization based on CFD model of single wind turbine
wake. In Proceedings of the European Wind Energy Conference and Exhibition 2012,
Copenhagen, Denmark, 16–19 April 2012; pp. 1–10.
261. Chen, L.; Macdonald, E. A new model for wind farm layout optimization with landowner decisions.
In Proceedings of 2012 Graduate Research Symposium, Ames, IA, USA, 29 March 2012; pp. 2–3.
262. Ekonomou, L.; Lazarou, S.; Chatzarakis, G.E.; Vita, V. Estimation of wind turbines optimal
number and produced power in a wind farm using an artificial neural network model.
Simul. Model. Pract. Theory 2012, 21, 21–25.
263. Rajper, S.; Amin, I.J. Optimization of wind turbine micrositing: A comparative study.
Renew. Sustain. Energy Rev. 2012, 16, 5485–5492.
Energies 2014, 7 7011

264. Veeramachaneni, K.; Wagner, M.; O’Reilly, U.-M.; Neumann, F. Optimizing energy output and
layout costs for large wind farms using particle swarm optimization. In Proceedings of the IEEE
Congress on Evolutionary Computation, Brisbane, Australia, 10–15 June 2012; pp. 1–7.
265. Yin, P.-Y.; Wang, T.-Y. A GRASP-VNS algorithm for optimal wind-turbine placement in wind
farms. Renew. Energy 2012, 48, 489–498.
266. Chen, K.; Song, M.X.; He, Z.Y.; Zhang, X. Wind turbine positioning optimization of wind farm
using greedy algorithm. J. Renew. Sustain. Energy 2013, 5, 1–15.
267. Feng, J.; Shen, W.Z. Wind Farm Layout: A Refinement Method by Random Search; Technical
University of Denmark: Lyngby, Denmark, 2013.
268. Geem, Z.W.; Hong, J. Improved formulation for the optimization of wind turbine placement in a
wind farm. Math. Probl. Eng. 2013, 2013, 1–5.
269. Pookpunt, S.; Ongsakul, W. Optimal placement of wind turbines within wind farm using binary
particle swarm optimization with time-varying acceleration coefficients. Renew. Energy 2013,
55, 266–276.
270. Rahmani, R.; Yusof, R.; Seyedmahmoudian, M.; Mekhilef, S. Implementing particle swarm
optimization in wind farm to place Wind Turbines. Aust. J. Basic Appl. Sci. 2013, 7, 77–84.
271. Rezaei Mirghaed, M.; Roshandel, R. Site specific optimization of wind turbines energy cost:
Iterative approach. Energy Convers. Manag. 2013, 73, 167–175.
272. Natural Resources Canada. RETScreen 4 International. Available online: http://www.
retscreen.net/ang/home.php (accessed on 18 March 2014).
273. Gifford, J.S.; Grace, R.C. Cost of Renewable Energy Spreadsheet Tool: A Model for Developing
Cost-Based Incentives in the United States; Report NREL/SR-6A20-50374; National Renewable
Energy Laboratory: Golden, CO, USA, 2013; pp. 1–43.
274. Krohn, S.; Morthorst, P.-E.; Awerbuch, S. The Economics of Wind Energy; European Wind
Energy Association: Roskilde, Denmark, 2009; pp. 1–156.
275. Blanco, M.I. The economics of wind energy. Renew. Sustain. Energy Rev. 2009, 13, 1372–1382.
276. Cory, K.; Schwabe, P. Wind Levelized Cost of Energy: A Comparison of Technical and
Financing Input Variables; National Renewable Energy Laboratory: Golden, CO, USA, 2009;
pp. 1–34.
277. González, J.S.; Rodriguez, A.G. G.; Mora, J.C.; Santos, J.R.; Payan, M.B. A new tool for wind
farm optimal design. In Proceedings of the 2009 IEEE PowerTech Conference, Bucharest,
Romania, 28 June–2 July 2009; pp. 1–7.
278. González, J.S.; Rodríguez, Á.G.G.; Mora, J.C.; Burgos Payán, M.; Santos, J.R. Overall design
optimization of wind farms. Renew. Energy 2011, 36, 1973–1982.
279. Larsen, G.C. A Simple Generic Wind Farm Cost Model Tailored for Wind Farm Optimization;
Report Risø-R-1710(EN); RisøNational Laboratory: Roskilde, Denmark, 2009; pp. 1–25.
280. González, J.S.; Gonzalez Rodriguez, A.G.; Mora, J.C.; Santos, J.R.; Payan, M.B. Optimization of
wind farm turbines layout using an evolutive algorithm. Renew. Energy 2010, 35, 1671–1681.
281. Crespo, A.; Hernández, J.; Frandsen, S. Survey of modelling methods for wind turbine wakes and
wind farms. Wind Energy 1999, 2, 1–24.
282. Choi, J.; Shan, M. Advancement of Jensen (Park) wake model. In Proceedings of the European
Wind Energy Conference and Exhibition 2013, Vienna, Austria, 4–7 February 2013; pp. 1–8.
Energies 2014, 7 7012

283. Frandsen, S.; Barthelmie, R.; Pryor, S.; Rathmann, O.; Larsen, S.; Højstrup, J.; Thøgersen, M.
Analytical modelling of wind speed deficit in large offshore wind farms. Wind Energy 2006, 9,
39–53.
284. Magnusson, M. Near-wake behaviour of wind turbines. J. Wind Eng. Ind. Aerodyn. 1999, 80,
147–167.
285. Ott, S.; Berg, J.; Nielsen, M. Linearised CFD Models for Wakes; Danmarks Tekniske
Universitet, RisøNationallaboratoriet for Bæredygtig Energi: Roskilde, Denmark, 2011.
286. Troldborg, N. Actuator Line Modeling of Wind Turbine Wakes; Technical University of
Denmark: Lyngby, Denmark, 2009; p. 142.
287. Larsen, G.C.; Réthoré, P.-E. A Simple Stationary Semi-Analytical Wake Model; Report
Risø-R-1713(EN); RisøNational Laboratory: Roskilde, Denmark, 2009.
288. Ishihara, T.; Yamaguchi, A.; Fujino, Y. Development of a new wake model based on a wind tunnel
experiment. In Proceedings of Global Wind Power 2004, Chicago, IL, USA, 28–31 March 2004.
289. Rados, K.; Larsen, G.; Barthelmie, R.; Schlez, W.; Lange, B.; Schepers, G.; Hegberg, T.;
Magnisson, M. Comparison of wake models with data for offshore windfarms. Wind Eng. 2001,
25, 271–280.
290. Crespo, A.; Herna´ndez, J. Turbulence characteristics in wind-turbine wakes. J. Wind Eng.
Ind. Aerodyn. 1996, 61, 71–85.
291. Larsen, G.C.; Hojstrup, J.; Madsen, H. Wind fields in wakes. In European Wind Energy
Conference and Exhibition 1996; European Wind Energy Association: Goteborg, Sweden, 1996.
292. Quarton, D.; Ainslie, J. Turbulence in wind turbine wakes. In Proceedins of the European Wind
Energy Conference and Exhibition 1989, London, UK, 10–13 July 1989.
293. Code of Practice for Loads and Safety of Wind Turbine Constructions; The Danish Society of
Engineers and the Federation of Engineers: Copenhagen, Denmark, 1992. (In Danish)
294. Characteristics of Atmospheric Turbulence Near the Ground. Part II: Single Point Data for Strong
Winds (Neutral Atmosphere); Engineering Sciences Data Unit (ESDU): London, UK, 1985.
295. Hassan, U. A Wind Tunnel Investigation of the Wake Structure Within Small Wind Turbine
Farms; UK Department of Energy: London, UK, 1992.
296. Lissaman, P.; Bate, E. Energy Effectiveness of Arrays of Wind Energy Conversion Systems;
Report AV-FR-7058; Aerovironment Inc.: Pasadena, CA, USA, 1977.
297. Lissaman, P. Energy Effectiveness of Arrays of Wind Energy Collection Systems; Report AV-R-6110;
Aerovironment Inc.: Pasadena, CA, USA, 1977.
298. Frandsen, S. On the wind speed reduction in the center of large clusters of wind turbines. J. Wind
Eng. Ind. Aerodyn. 1992, 39, 251–265.
299. Van Leuven, J. The Energetic Effectiveness of a Cluster of Wind Turbines; Universite Catholique
de Louvain: Louvain-la-Neuve, Belgium, 1992.
300. Rankema, D.J. Validation of Wind Turbine Wake Models: Using Wind Farm Data and Wind
Tunnel Measurements; Delft University of Technology: Delft, The Netherlands, 2007.
301. Conzález-Rodriguez, A.G.; Serrano-Conzález, J.; Riquelme-Santos, J.M.; Burgos-Payán, M.;
Castro-Mora, J.; Persan, S.A. Global Optimization of Wind Farms Using Evolutive Algorithms;
Wang, L., Singh, C., Kusiak, A., Eds.; Springer: Berlin, Germany, 2010; pp. 53–104.
Energies 2014, 7 7013

302. Kaiser, K.; Langreder, W.; Hohlen, H. Turbulence correction for power curves. In Proceedings of
the European Wind Energy Conference and Exhibition, Madrid, Spain, 16–19 June 2004.
303. Larsen, T.J.; Hansen, A.M. How 2 HAWC2, the User’s Manual; Risø National Laboratory:
Roskilde, Denmark, 2007.
304. Prim, R.C. Shortest connection networks and some generalizations. Bell Syst. Tech. J. 1957, 36,
1389–1401.
305. Kruskal, J.B. On The Shortest Spanning Subtree of A Graph and the Traveling Salesman
Problem; American Mathematical Society: Providence, RI, USA, 1956; pp. 48–50.
306. Larsen, G.C.; Buhl, T.; Madsen, H.A.; Troldborg, N.; Larsen, T.J.; Ott, S.; Mann, J.; Nielsen, M.;
Markou, H.; Réthoré, P.-E.; et al. Topfarm—Next Generation Design Tool for Optimisation of
Wind Farm Topology and Operation; Report Risø-R-1805(EN); Risø National Laboratory:
Roskilde, Denmark, 2011; pp. 1–95.
307. González, J.S.; Payán, M.B.; Santos, J.R. A new and efficient method for optimal design of large
offshore wind power plants. IEEE Trans. Power Syst. 2013, 28, 3075–3084.
308. Serrano-González, J.; Burgos-Payán, M.; Riquelme-Santos, J. Design of neighboring large
offshore wind farms: A game theory approach. In Proceedings of the European Wind Energy
Conference and Exhibition 2013, Vienna, Austria, 4–7 February 2013.
309. Serrano-González, J.; Burgos-Payán, M.; Riquelme-Santos, J.M. An improved evolutive
algorithm for large offshore wind farm optimum turbines layout. In Proceedings of 2011 IEEE
Trondheim PowerTech, Trondheim, Norway, 19–23 June 2011; pp. 1–6.
310. Serrano-González, J.; Burgos-Payán, M.; Riquelme-Santos, J. Optimum design of transmissions
systems for offshore wind farms including decision making under risk. Renew. Energy 2013, 59,
115–127.
311. Investigation of the Interactions between Wind Turbines and Radio Systems Aimed at
Establishing Co-Siting Guidelines, Phase 1: Introduction and Modelling of Wind Turbine
Scatter; Energy Technology Support Unit (ETSU): Harwell, UK, 1997.
312. Moglia, A.; Trusszi, G.; Orsenigo, L. Evaluation methods for the electromagnetic interferences
due to wind farms. In Proceedings of the Conference on Integration of Wind Power Plants in the
Environment and Electric Systems, Rome, Italy, 7–9 October 1996.
313. Van Kats, P.J.; van Rees, J. Large wind turbines: A source of interference for TV broadcast
reception. In Wind Energy and the Environment; Swift-Hook, D.T., Ed.; IEE Energy Series:
London, UK, 1989; pp. 95–104.
314. Hall, H.H. The assessment and avoidance of electromagnetic interference due to windfarms.
Wind Eng. 1992, 16, 326–337.
315. Acoustics—Attenuation of Sound during Propagation Outdoors; ISO 9613–2; International
Organization for Standardization: Geneva, Switzerland, 1996; pp. 1–24.
316. Wagner, S.; Bareis, R.; Guidati, G. Wind Turbine Noise; Springer-Verlag: Berlin, Germany, 1996.
317. Bekendtgørelse om Stoj fra Windmoller; Danish Statutory Order on Noise from Windmills:
Copenhagen, Denmark, 1991. (In Danish)
318. González, J.S.; Payán, M.B.; Riquelme-Santos, J.M. Wind farm optimal design including risk.
In Proceedings of the International Symposium Modern Electric Power Systems 2010, Wroclaw,
Poland, 20–22 September 2010; pp. 1–6.
Energies 2014, 7 7014

319. Kühn, M.; Harland, L.A.; Bierbooms, W.A.A.M.; Cockerill, T.T.; Ferguson, M.C.; Göransson, B.;
van Bussel, G.J.W.; Vugts, J.H. Opti-OWECS Final Report Volume 1: Integrated Design
Methodology for Offshore Wind Energy Conversion Systems; Report IW-98140R; Delft
University of Technology: Delft, The Netherlands, 1998.
320. Kooijman, H.J.T.; de Noord, M.; Volkers, C.H.; Machielse1, L.A.H.; Hagg, F.; Eecen, P.J.;
Pierik, J.T.G.; Herman, S.A. Cost and Potential of Offshore Wind Energy on the Dutch Part of
the North Sea; The Energy Research Centre of The Netherlands: Petten, The Netherlands, 2001.
321. Veldkamp, D. Chances in Wind Energy, a Probabilistic Approach to Wind Turbine Fatigue
Design; Delft University: Delft, The Netherlands, 2006.
322. Zhang, P.Y. Topics in Wind Farm Layout Optimization: Analytical Wake Models, Noise
Propagation, and Energy Production; University of Toronto: Toronto, ON, Canada, 2013; p. 67.
323. Land, A.H.; Doig, A.G. An automatic method of solving discrete programming problems.
Econometrica 1960, 28, 497–520.
324. Leyffer, S.; Linderoth, J.; Luedtke, J.; Miller, A.; Munson, T. Applications and algorithms for
mixed integer nonlinear programming. J. Phys. Conf. Ser. 2009, 180, doi:10.1088/1742-6596/
180/1/012014.
325. Schneider, J.J.; Kirkpatrick, S. Stochastic Optimization; Schneider, J.J., Kirkpatrick, S., Eds.;
Springer: Berlin, Germany, 2006; p. 563.
326. Trevisan, L. Combinatorial Optimization: Exact and Approximate Algorithms; Stanford University:
Stanford, CA, USA, 2011; p. 139.
327. Gendreau, M.; Potvin, J.-Y. Handbook of Metaheuristics; Gendreau, M., Potvin, J.-Y., Eds.;
International Series in Operations Research & Management Science; 2nd ed.; Springer US:
Boston, MA, USA, 2010; p. 648.
328. Horst, R.; Pardalos, P.; van Thoai, N. Introduction to Global Optimization, 2nd ed.; Springer:
Berlin, Germany, 2000; p. 357.
329. Pardalos, P.; Romejin, E.H. Handbook of Global Optimization, 2nd ed.; Kluwer Academic
Publishers: Dordrecht, The Netherlands, 2002; Volume 2, p. 572.
330. Bertsimas, D.; Tsitsiklis, J.N.; Tsitsiklis, J. Introduction to Linear Optimization; Athena Scientific:
Belmont, MA, USA, 1997; Volume 6, p. 608.
331. Wolsey, L.A.; Nemhauser, G.L. Integer and Combinatorial Optimization; Wiley-Interscience:
Hoboken, NJ, USA, 1999; p. 784.
332. Martí , R.; Gallego, M.; Duarte, A. A branch and bound algorithm for the maximum diversity
problem. Eur. J. Oper. Res. 2010, 200, 36–44.
333. Hahsler, M.; Hornik, K. TSP—Infrastructure for the traveling salesperson problem. J. Stat. Softw.
2007, 23, 1–21.
334. Wolpert, D.H.; Macready, W.G. No free lunch theorems for optimization. IEEE Trans.
Evol. Comput. 1997, 1, 67–82.
335. Goldberg, D.E. Computer-aided pipeline operation using genetic algorithms and rule learning.
PART I: Genetic algorithms in pipeline optimization. Eng. Comput. 1987, 3, 35–45.
336. Johnson, D.S.; Aragon, C.R.; McGeoch, L.A.; Schevon, C. Optimization by simulated annealing:
An experimental evaluation; Part II, graph coloring and number partitioning. Oper. Res. 1991,
39, 378–406.
Energies 2014, 7 7015

337. Johnson, D.S.; Aragon, C.R.; McGeoch, L.A.; Schevon, C. Optimization by simulated annealing:
An experimental evaluation; Part I, graph partitioning. Oper. Res. 1989, 37, 865–892.
338. Aarts, E.; Korst, J. Simulated Annealing and Bultzman Machine A Stochatic Approche to
combinatorial Optimization and Neutral Computing; Wiley: Hoboken, NJ, USA, 1989.
339. Abramson, D.; Krishnamoorthy, M.; Dang, H. Simulated annealing cooling schedules for the
school timetabling problem. Asia-Pac. J. Oper. Res. 1999, 16, 1–22.
340. Chu, K.W.; Deng, Y.F.; Reinitz, J. Parallel simulated annealing by mixing of states. J. Comput. Phys.
1999, 148, 646–662.
341. Eglese, R. Simulated annealing a tool for operational research. Eur. J. Oper. Res. 1990, 46,
271–281.
342. Fleischer, M. Simulated annealing past, present, and future. In Proceeding of the 1995 Winter
Simulation Confrence, Los Angeles, CA, USA, 3–6 December 1995; pp. 155–161.
343. Kirkpatrick, S.; Gelatt, C.D.; Vecchi, M.P. Optimization by simulated annealing. Science 1983,
220, 671–680.
344. Van Laarhoven, P. Theoretical and Computational Aspects of Simulated Annealing; Society for
Industrial and Applied Mathematics: Philadelphia, PA, USA,1988.
345. Kennedy, J.; Eberhart, R. Particle swarm optimization. In Proceedings of ICNN’95—International
Conference on Neural Networks, Perth, Australia, 27 November–1 December 1995; pp. 1942–1948.
346. Goldberg, D.E. Genetic Algorithm in Search of Optimization and Machine Learning;
Addison-Wesley Longman Publishing Co., Inc.: Boston, MA, USA, 1989.
347. Papadimitriou, C.H. Computational Complexity, 1st ed.; Addison-Wesley Professional: Boston,
MA, USA, 1993; p. 523.
348. Gen, M.; Cheng, R. Genetic Algorithms and Engineering Optimization; John Wiley & Sons, Inc.:
Hoboken, NJ, USA, 2000.
349. Feo, T.A.; Resende, M.G.C. A probabilistic heuristic for a computationally difficult set covering
problem. Oper. Res. Lett. 1989, 8, 67–71.
350. Chan, T.M.; Har-Peled, S. Approximation algorithms for maximum independent set of
pseudo-disks. In Proceedings of the 25th Annual Symposium on Computational Geometry—SCG ’09,
New York, NY, USA, 8–10 June 2009; p. 333.
351. Mustafa, N.H.; Ray, S. PTAS for geometric hitting set problems via local search. In Proceedings
of the 25th Annual Symposium on Computational Geometry—SCG ’09, New York, NY, USA,
8–10 June 2009; pp. 17–22.
352. Mladenovic, N.; Hansen, P. Variable neighborhood search. Comput. Oper. Res. 1997, 24,
1097–1100.
353. Bilbao, M.; Alba, E. GA and PSO applied to wind energy optimization. In XV Congreso
Argentino de Ciencias de la Computación; Universidad Nacional de La Plata: Jujuy, Argentina,
5–9 October 2009; pp. 50–59.
354. Larsen, G.C. TOPFARM-next generation design tool for optimization of wind farm topology and
operation…background, vision and challenges. In The Science of Making Torque From Wind
2010; European Wind Energy Association: Heraklion, Greece, 2010; pp. 437–448.
Energies 2014, 7 7016

355. Bilbao, M.; Alba, E. CHC and SA applied to wind energy optimization using real data.
In Proceedings of 2010 IEEE Congress on Evolutionary Computation, Barcelona, Spain,
18–23 July 2010; pp. 1–8.
356. Saavedra-Moreno, B.; Salcedo-Sanz, S.; Paniagua-Tineo, A.; Prieto, L.; Portilla-Figueras, A.
Seeding evolutionary algorithms with heuristics for optimal wind turbines positioning in wind
farms. Renew. Energy 2011, 36, 2838–2844.
357. Mokryani, G.; Siano, P. Optimal wind turbines placement within a distribution market environment.
Appl. Soft Comput. 2013, 13, 4038–4046.
358. Di Gaspero, L.; Schaerf, A.; Stützle, T. Advances in Metaheuristics; di Gaspero, L., Schaerf, A.,
Stützle, T., Eds.; Springer: Berlin, Germany, 2013; p. 183.
359. Talbi, E. Hybrid Metaheuristics; Talbi, E., Ed.; Springer: Berlin, Germany, 2013; p. 458.
360. Blum, C.; Puchinger, J.; Raidl, G.R.; Roli, A. Hybrid metaheuristics in combinatorial optimization:
A survey. Appl. Soft Comput. 2011, 11, 4135–4151.
361. Gu, H.; Wang, J. Irregular-shape wind farm micro-siting optimization. Energy 2013, 57, 535–544.
362. Meissner, C.; Vogstad, K.; Horn, U.W.-S. Park optimization using IEC constraints for wind
quality. In Proceedings of the European Wind Energy Conference and Exhibition 2011, Brussels,
Belgium, 14–17 March 2011.
363. Salcedo-Sanz, S.; Saavedra-Moreno, B.; Paniagua-Tineo, A.; Prieto, L.; Portilla-Figueras, A.
A review of recent evolutionary computation-based techniques in wind turbines layout optimization
problems. Cent. Eur. J. Comput. Sci. 2011, 1, 101–107.

© 2014 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article
distributed under the terms and conditions of the Creative Commons Attribution license
(http://creativecommons.org/licenses/by/4.0/).