You are on page 1of 13

Energy Conversion and Management 208 (2020) 112593

Contents lists available at ScienceDirect

Energy Conversion and Management


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

Comparative analysis and improvement of grid-based wind farm layout T


optimization
Giovanni Gualtieri
National Research Council, Institute of Bioecomony (CNR-IBE), Via Caproni 8, 50145 Firenze, Italy

ARTICLE INFO ABSTRACT

Keywords: Among the main grid-based wind farm layout optimization studies addressed in the literature, 14 layouts have
Wind farm layout optimization been recomputed by selecting the levelized cost of energy as a primary objective function. Relying on 120 wind
Gridded layout turbine combinations, a previously developed optimization method targeting best turbine selection has then
Literature case study been applied. All literature layouts were optimized, as capacity factors were (slightly) increased (78.89–80.90 to
Wind turbine database
83.02–83.07%), while levelized costs of energy were (significantly) reduced (130.37–370.42 to 54.01–142.64
Levelized cost of energy
Self-organizing map
$/MWh). This study concluded that neither the discrete nor the continuous optimization model can be re-
commended in all scenarios. In general, a capacity factor increase does not necessarily imply a decrease in
levelized cost of energy. The latter may be minimized by decreasing the overall wind farm capacity, the number
of turbines, or selecting turbines with lower rotor diameters or rated powers. By contrast, capacity factor may be
maximized by installing turbines with higher hub heights or lower rated speeds. Contradicting various findings,
using turbines with different rotor diameters, rated powers or hub heights is not recommended to minimize the
levelized cost of energy. Although addressed within several optimization studies, maximization of energy pro-
duction is a misleading target, as involving the highest costs of energy.

1. Introduction and crosswind spacing between the WTs; (ii) most suitable grid cells to
place the WTs. According to various Authors (e.g. [4]), since enabling
Wind farm layout optimization (WFLO) is commonly intended as to continuously vary the WTs placement, the continuous space search is
optimal positioning of wind turbines (WTs) in a wind farm (WF) more capable of achieving global optimal solutions. Conversely, unless
through minimization of power losses due to wake interferences be- when dealing with particularly dense grids (e.g. [9]), searching in a
tween the WTs to maximize the annual energy yield (AEY) [1]. In discrete space generally leads to a significant WFLO computational
reality, this is a limiting definition, as AEY is not the only factor to take simplification [10]. An improved approach has been addressed by
into account [2]. Since WF planning is an economic project, AEY various Authors (e.g. [11]) by combining both discrete and continuous
maximization should be combined with the minimization of the cost of models into a single framework in an attempt to incorporate the ad-
energy [3]. vantages offered by each model: less model-solving complexity the
former, and more flexible WT positioning the latter. To this aim, Long
1.1. Addressing wind farm layout optimization et al. [12] performed a thorough comparison of the advantages and
drawbacks in planning WF layout deriving from applying the discrete
When addressing the WFLO problem, two WF layout models are vs. the continuous model.
generally applied [4]: (i) the continuous model, allowing WTs to be The WFLO problem is non-convex (involving many optimal solu-
positioned anywhere in the WF, subject to constraints; (ii) the discrete tions) and cannot be solved using traditional optimization methods, but
model, only allowing WTs to be positioned at a finite number of loca- requires heuristic optimization techniques, which were implemented by
tions. When applying the continuous model (i), WTs may either be ir- most Authors [2]. Heuristic techniques include, e.g., Genetic Algo-
regularly placed (e.g. [5]), or positioned according to aligned (e.g. [6]) rithms (GA) [13], Particle Swarm Optimization (PSO) [3], Evolutive
or staggered (e.g. [7]) arrays. In the discrete approach (ii), the domain Algorithm (EA) [14], Viral System Algorithm (VSA) [8], or Extended
is divided into a discrete grid with WTs positioned in the centre of Pattern Search (EPS) [15]. GAs are used in more than 75% of WFLO
specific cells (e.g. [8]). Depending on the layout scheme, the best WTs studies [4]. All these algorithms are strongly affected by the size of the
positioning is pursued by seeking, respectively: (i) optimal downwind solution space. For example, if considering a gridded layout divided

E-mail address: giovanni.gualtieri@ibe.cnr.it.

https://doi.org/10.1016/j.enconman.2020.112593
Received 11 November 2019; Received in revised form 7 February 2020; Accepted 8 February 2020
0196-8904/ © 2020 Elsevier Ltd. All rights reserved.
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

Nomenclature LCoE levelized cost of energy [$/MWh], Eq. (25)


Lh, Lv horizontal and vertical WF full sizes [m]
Abbreviations NLAY overall number of generated WF layouts [–]
NLIT number of literature WF layouts [–]
AGL above ground level NP, ND, NH number of combinations in the WF with WTs having
AGL above ground level different Pr, D or Hhub [–]
AGL above ground level NT, NTH number of WTs, and number of WT combinations [–]
PL power law NWF number of WF layouts for each literature case study [–]
SOM self-organizing map P total wind power installed in the WF [kW]
WF wind farm Pe electric power output from a WT [kW], Eqs. (8), (13)
WFLO wind farm layout optimization Pr WT rated power [kW]
WT wind turbine r0 wake radius downstream a WT [m], Eq. (1)
rw radius of wake expansion caused by an upstream WT [m],
Variables Eq. (3)
v(z) wind speed at generic height z AGL [m/s], Eq. (20)
a WT axial induction factor [–], Eq. (2) v, vm wind speed, and mean wind speed [m/s]
A WT swept area [m2] v0 free stream wind speed [m/s]
AEY annual energy yield [MWh/y], Eq. (9) vact actual wind speed approaching the WT [m/s], Eq. (7)
CF capacity factor [%], Eq. (10) vhub actual wind speed approaching the WT at Hhub [m/s]
Ci annual cost at year i [$/kW/y] vi, vr, vo WT cut-in, rated and cut-off wind speeds [m/s]
Cini initial capital cost [$] z, z1, z2 height AGL, and height at z1 and z2 AGL [m]
CT WT thrust coefficient [–] z0 site’s aerodynamic surface roughness length [m]
D WT rotor diameter [m] zref wind reference height AGL [m]
Fsite power losses of a single WT depending on the site type zsite site elevation ASL [m]
[%], Eq. (17) α wake expansion coefficient of upstream WT [–], Eq. (4)
Ftot total power losses of a single WT experienced in the WF wind shear coefficient [–]
[%], Eq. (18) δ wind speed deficit [–], Eqs. (5), (6)
Fwake power losses of a single WT due to all wake interactions in Δh, Δv horizontal and vertical WF grid cell sizes [m]
the WF [%], Eqs. (11), (14) η WF efficiency [%], Eqs. (12), (15)
FWT power losses of a single WT depending on the WT system Φ WT design ratio [–], Eq. (21)
[%], Eq. (16) Ω mean-to-rated wind speed [%], Eq. (22)
Hhub WT hub height [m]

into Ncell cells where NWT WTs are to be installed, the number of pos- approach in [18], i.e. the limited number of installed WTs, and thus the
sible solutions is: Ncell!/[NWT!(Ncell–NWT)!] [2]. waste of land resource [4]. These studies targeted the same goals of
The WFLO problem is also complicated by the turbulence loads Mosetti et al. [18], i.e. maximization of AEY [21] (also via maximiza-
experienced by the WTs due to their mutual wake interactions in the tion of WF array efficiency [22]), and/or minimization of a fitness
farm. These increase the fatigue degradation of WTs, thus shortening value, given by the ratio of WF costs to AEY [23]. Regardless, the fol-
their expected lifetime and consequently increasing the overall project’s lowing deficiencies should be pinpointed in all such studies due to the
cost. Therefore, a proper balance between the maximization of energy use of: (i) the same WT model; (ii) a theoretical WT power curve; (iii) a
production and the minimization of structural loads should be targeted constant value for WT thrust coefficient (CT); (iv) a simplified objective
[16]. On the other hand, both the positive and negative environmental function (the fitness value), also relying upon (v) an overly-simplistic
impacts involved by the installation of a grid-connected WF should not cost model.
be ignored. For example, Tao et al. [17] demonstrated that, since in- Actually, as Chowdhury et al. [3] pointed out, the WFLO problem
tegration of the WF into the electrical network reduces air pollutant should address two synergistic issues: (i) optimal WT allocation, and (ii)
emissions thus saving remarkable expenditures on pollutant penalty optimal WT selection. While several literature studies exist on heuristic
costs, neglecting these costs undervalues the profit deriving from the algorithms, merely representing best WTs allocators (e.g. [24]), a
WF connection to the grid. By contrast, a relatively small amount of minority of works focused on seeking the best-suited characteristics of
economic benefit should be sacrificed for reducing the noise dis- commercial WTs [3], as in most cases only one single model was used,
turbance of the WF to neighbouring residents and animals. either when adopting the discrete [25] or the continuous [26] WFLO
model, or a combination of both [11]. In other words, rather than a
further WFLO assumption, WT characteristics should be treated as an
1.2. Literature survey array of variables to be carefully analysed for detecting the unique
combination required to solve the WFLO problem. To this aim, for ex-
Mosetti et al. [18] first addressed the optimal positioning of WTs in ample, Hayat et al. [27] analysed the advantage of alternating 2- and 3-
a WF, applying a GA algorithm and Jensen’s wake model [19]. They
blade WTs in a WF, while the benefits of vertically-staggered WFs were
dealt with: (i) a flat onshore site; (ii) three ideal wind conditions; (iii) explored placing small WTs between large WTs by Chatterjee and Peet
the same (theoretical) WT model installed over the whole WF; (iv) a
[28], as well as using different hub heights by Stanley et al. [29] or Wu
square (2 km × 2 km) gridded layout with 200-m square cells [18]. et al. [30]. A 3-variable analysis was addressed, e.g., by Mirghaed and
WFLO was pursued by maximizing AEY and minimizing the installation Roshandel [31], who targeted the minimization of WF costs through
costs of WTs [2]. Following and in the framework of this seminal work, variation of WT hub height, rotor diameter and rated power, while
several grid-based WFLO studies were addressed, either using the same parameters at WF-level such as the number of WTs and total installed
assumptions as in [18] (e.g. [20]), or slightly different ones that in-
power were also taken into account by Pookpunt and Ongsakul [32].
volved a denser grid up to 20 × 20 [1] or even 50 × 50 [9] cells. All of Following this survey, it is clear that all key factors – both at WT-
these studies attempted to address a recognized drawback of the

2
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

and WF-level – should be concurrently analysed to actually perform a studies and implemented in several WFLO software, including WAsP
WFLO comprehensive study. [37] and WindPro [38].
According to Jensen’s model, the wake expands linearly behind a
1.3. Goals and contributions of the paper WT and a constant wind speed deficit in the radial direction x occurs
[39]. If considering an upstream WTi having a rotor radius Ri, the radius
The goals of the present study are the following: of the wake spreading downstream WTi is [14]:

1 ai
(i) to comparatively analyse main grid-based WFLO studies addressed r0, i = Ri
in the literature after their recomputation using more real-world 1 2ai (1)
settings; where ai is WTi axial induction factor [39]:
(ii) to optimize them by applying a previously developed method [33]
which targets an optimal selection of commercial-scale WTs; ai = 0.5(1 1 CT , i ) (2)
(iii) to provide insights into optimization functions and parameters that
mostly influence WFLO. with CT,i the WTi thrust coefficient.
At radial distance xij of a downstream WTj, the radius of wake ex-
Literature studies where an optimal 10 × 10 (2 km × 2 km) grid- pansion due to WTi is [39]:
based layout was achieved against wind scenario ‘c’ defined in [18] rw, ij = r0, i + i x ij (3)
have been recomputed. According to goal (i), the first contribution of
the present paper is to overcome deficiencies (detailed in section 1.2) where αi is WTi-induced wake expansion coefficient [14]:
affecting such studies, for which the following were never applied: (i) 0.5
120 combinations of commercial WTs rather than using a single WT i =
ln(Hhub, i z 0 ) (4)
model; (ii) experimental (instead of theoretical) power curves for each
WT; (iii) a continuous function of wind speed rather than a constant with Hhub,i the hub height of WTi and z0 site’s roughness length.
value for CT; (iv) the levelized cost of energy (LCoE) as an objective A single wake between two WTs causes a velocity deficit given by
function instead of the simple fitness value; (v) the thorough NREL cost [39]:
model rather than the simplified cost model proposed in [18]. 2
Performing optimization of such studies by applying the method r0, i Aoverlap, ij
ij = 2ai
developed in [33] (goal ‘ii’) enables to deliver the paper's second con- rw, ij Aj (5)
tribution: since the WTs best selection rather than the WTs best posi-
tioning is pursued, layout optimization – contrary to those studies – has where Aj is WTj swept area, and Aoverlap,ij the overlapping area between
been performed by varying the type of WTs while retaining the number Aj and WTi-induced wake area at xij. Aoverlap,ij can be calculated by
of WTs and their placement as assumed. Since the WFLO method pre- applying the method described in [39].
sented in [33] was previously applied to the continuous WF layout Summarizing, for a WTj subject to multiple wakes induced by Ni
model, a paper's further contribution is not only to test its applicability upstream WTi, the total velocity deficit can be calculated by summing
to the discrete model, but also to make a straight comparison – since up the Ni contributions provided by Eq. (5) [39]:
application conditions are similar – between the discrete and the con- 2 2
Ni Ni
tinuous WFLO model: this offers the opportunity to possibly determine 2 r0, i Aoverlap, ij
j = ij = 2ai
whether one approach is superior to the other. i=1 i=1
rw, ij Aj
(6)
An additional contribution while targeting goal (iii), is a multi-
purpose and a multi-variable analysis of grid-based WFLO studies that Ultimately, actual wind speed approaching the downstream WTj is
have never been addressed before: to concurrently consider all possible given by:
optimization functions (power production, WF efficiency, capacity
vact , j = v0 (1 j) (7)
factor or cost of energy) and parameters that mostly influence WFLO –
both at WT- and WF-level – allows to genuinely achieve a compre-
hensive WFLO study.
2.2. Wind power output and wind farm efficiency
2. Methods
The electric power Pe(v) generated by a real WT is defined as [40]:
Wake losses between the WTs were calculated by using Jensen’s 1 1
Pe (v ) = m e Pm (v ) = m e Cp P ( v ) = Cp A v3 = A v3
model, while all other losses were assessed through an analytic method. 2 m e
2 T

A previously-developed curve as a function of wind speed was used for (8)


CT, while the NREL model was applied for WF cost analysis. SOMs were
where Pm(v) is the mechanical power extractable from wind power P(v)
used to provide insights into WF layout optimization functions and
available for a WT with swept area A affected by a wind speed v [41], ρ
parameters that influence WFLO the most.
is site’s air density, Cp is the WT power coefficient, ηm and ηe are me-
chanical transmission and electric conversion efficiencies of the WT,
2.1. Wake losses: Jensen’s model
andηT = Cpηmηe is total power efficiency [42].
Annual energy yield (AEY) of a WT over a 1-year period (t =8760 h)
When a uniform free wind speed v0 hits a WT, a cone of slower and
is [43]:
more turbulent air develops behind the WT, causing the so-called ‘wake
effect’ [26]. Wake interactions between WTs are accounted for by the 1
AEY = t Pe (v ) f (v ) dv = t m e Pm (v ) f (v ) dv = t A v 3f (v ) dv
wake models. Jensen's analytical wake model, although one of the 2 T
0 0 0
oldest, is still very effective in WFLO because of its simplicity [4].
(9)
Evidence that Jensen's model outperformed other wake models has
been provided at both onshore (e.g. [34]) and offshore (e.g. [35]) lo- Capacity factor CF is the ratio of AEY to the energy (Er) that the WT
cations. Originally proposed by Jensen [19] and later developed by could have produced if operated at its rated power over the same period
Katic et al. [36] for studying WFLO, the model is applied within most [44]:

3
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

CF =
AEY attempted to deal with this issue, proposing a CT continuous curve as a
Er (10) function of Cp following a best-fit procedure. Very few WFLO studies
considered CT as a value varying with wind speed [2], as for example
Overall power losses a generic WTj undergoes in the WF caused by
the study performed by Ali et al. [45].
all wake interactions are:
Based on CT experimental curves provided by manufacturers for 50
Pact , j onshore commercial WTs [46], a further improvement has been pro-
Fwake, j = 1
P0 (11) posed in [33] by applying a best-fit procedure, which returned the
following CT(v) curve:
where Pact,j and P0 represent the power productions corresponding to
vact,j and v0 in Eq. (7). CT = (5.47581845 10 6v 5.00641402 + 1.132584887) 1
(19)
WF efficiency, returning the amount of energy extracted from total
Eq. (19) has been compared against the CT values averaged from the
available energy, is the ratio of total actual power output to total ideal
considered 50 WTs, eventually returning the scores summarized in
power output. For a WF including N WTs, it yields [44]:
Table S1 of the Supplementary material. Eq. (19) is particularly accu-
N
rate when the WTs operate above their rated speed. Across the full wind
Pact , j
Pact j=1
speed range (3–25 m/s), a mean bias of 0.007 m/s, root mean square
= = error of 0.027 m/s, and R2 of 0.993 are achieved.
Pideal N P0 (12)
Mosetti et al. [18] introduced a WT theoretical model whose power 2.5. Wind speed vertical profile
output as a function of wind speed resulted from applying the laws of
Betz and the momentum in the airflow passing through the WT swept The PL is the wind speed vertical extrapolation model most widely
area [2]. Assuming ηT = 40%, ρ = 1.225 kg/m3, and D = 40 m, Eq. (8) used in wind energy applications (73.5%) because of its simplicity and
can thus be expressed through the following theoretical equation [14]: greatest accuracy [47]. Based on the PL, a known wind speed value at
Pe (v ) = 0.3 v 3 (13) height z1 is extrapolated to height z2 (i.e., WT hub height) by [47]:

with Pe(v) in kW. Eq. (13) was used in all grid-based WFLO literature z2
v (z2 ) = v (z1)
studies analysed. z1 (20)
In the special case of the WT theoretical model, a modified ex-
where v(z1) and v(z2) are wind speeds at heights z1 and z2, respectively,
pression for Eq. (11) can be derived by using Eq. (13):
and is the wind shear coefficient. Although is a function of various
Pact , j 3
vact ,j
parameters (e.g., atmospheric stability, terrain roughness, height range)
Fwake, j = 1 = 1 and experiences significant inter-daily and inter-annual variations [48],
P0 v03 (14)
when the site’s estimates are not available a rough value as a sole
Accordingly, combining Eq. (13) into Eq. (12) yields [14]: function of the landscape can be used (e.g. [49]).
N N
3
Pact , j 0.3 vact ,j 2.6. Specific wind turbine parameters
P j=1 j =1
= act = =
Pideal N P0 N (0.3 v03 ) (15) Specific WT parameters, derived from basic parameters, may be
calculated to better analyse WT characteristics, and thus assess how
2.3. Turbine-related, site-related and total losses their particular regulation may influence WT performances.
For example, the WT design ratio [42]:
For any WT, power losses other than wake losses can be quantified = vr vi (21)
by applying the method reported in [41]. Losses linked to gearbox,
quantifies the range of the region a WT operates between its vi and vr
generator, converter, and unavailability & repair are calculated as de-
values.
pending on the WT system; losses due to electric grid connection, icing/
A further WT parameter is the ratio of the site’s mean wind speed to
soiling, and other generic factors are calculated as a function of the type
the WT rated wind speed [33]:
of site.
WT-related losses FWT are calculated as: = vm vr (22)
FWT = 1 [(1 fgearbox ) (1 fgenerator ) (1 fconverter ) (1 funav )] This mean-to-rated wind speed Ω quantifies the percentage of the
(16) site’s wind speed the WT exploits on average when operating at rated
conditions.
while site-related losses Fsite are calculated as:
Fsite = 1 [(1 fgrid ) (1 fice ) (1 fother )] (17) 2.7. Wind farm cost estimation

Summarizing, total losses Ftot affecting any WT in the farm are given An optimization function (the fitness value) was introduced to ad-
by: dress the WFLO problem [50]:
Ftot = 1 [(1 Fwake) (1 FWT ) (1 Fsite )] (18) Objective = Cost Pe (23)
where Pe is the total yearly power output extracted from the WF and
2.4. Turbine thrust coefficient Cost is the overall yearly cost of the WF, defined as [18]:
2 1 0.00174 N2
One of the major WT operational parameters, thrust coefficient CT is Cost = N + e
3 3 (24)
also crucial for achieving an accurate wake analysis [7]. In the large
majority of WFLO studies (e.g. [32]), a constant value of 0.88 is as- with N the number of WTs installed in the WF.
sumed for CT. However, as remarked by Serrano et al. [2], a CT fixed As noted in section 1.2, both a rough objective function (Eq. 23) and
value may lead to significant errors in evaluating the wake effect. Re- a simplified cost model (Eq. (24)) were used within most grid-based
lying on CT curves from 13 commercial WTs, Abdulrahman et al. [7] WFLO studies (e.g. [23]). A more economically accurate approach was

4
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

proposed, e.g., by Castro et al. [13], who suggested considering the similar feature values are arranged close to each other on the map,
project’s net present value (NPV) as an optimization function since while dissimilar data result in neurons allocated on different map edges.
taking into consideration crucial parameters such as the initial capital Therefore, the grid of neurons allows to easily detect the relationships
cost, its discount rate and the WF full lifetime. In addition to NPV, between the variables and, possibly, the cluster structure in the original
Shamshirband et al. [51] proposed that the project’s interest rate of data [55].
return (IRR) should also be considered. According to this perspective, The SOM algorithm performs an iterative training process where
LCoE is probably the most reliable metric to be used for assessing the any input neuron is connected to all output neurons. Once the single
economic performances of a WF project [3]. The LCoE is the ratio be- winner neuron is identified, neurons are connected to adjacent neurons
tween lifetime costs and lifetime electricity production, both discounted by a neighbourhood function. Gaussian, Cut Gaussian, bubble, and
back to a common year through a discount rate reflecting the average Mexican hat are the most commonly used neighbourhood functions.
capital cost [52]. For a WF project, LCoE may be calculated as [41]: Another two parameters that control this learning procedure are the
n learning rate and the neighbourhood radius, both decreasing mono-
Cini + i=1
Ci (1 + d )i tonically with the training steps [55].
LCoE = n
i=1
AEY (1 + d )i (25)
3. Layouts from literature studies
with Cini the initial capital cost, Ci the annual cost at year i from in-
stallation, d [%] the investment’s interest rate, and n the operational
A total of 14 grid-based layouts developed in the literature have
lifetime [years]. More accurate analytical models than Eq. (24) since
been comparatively analysed (Fig. 1). Their characteristics and scores
also taking Pr into account were proposed (e.g. [53]). The most reliable
are summarized in Table 1. All layouts have been built-up based on
cost model is likely the one developed at NREL by Fingersh et al. [54],
Mosetti et al. [18], who proposed a square gridded domain divided into
where the cost for each single WT component and subsystem are cal-
10x10 possible WT locations. Assuming the installation of a theoretical
culated. This model is fully detailed in Table 5 of [32], where: (i) Cini is
WT with D = 40 m, Hhub = 60 m, yielding a power output as in Eq.
expressed as a function of R, Pr and Hhub by means of 23 equations; (ii)
(13), they assumed square grid cells with a size of 5D = 200 m, with an
Ci is expressed as a function of Pr and AEY based on 3 equations.
overall layout extent therefore equal to 50D × 50D (2 km × 2 km).
Only layouts based on wind scenario ‘c’ defined in [18] have been
2.8. Self-organizing map considered herein.
Mosetti et al. [18] achieved an optimal layout comprising 15 WTs
The SOM is an unsupervised learning artificial neural network (Fig. 1a). Only Ulku & Alabas-Uslu [10], proposing a layout with WTs
model that operates a nonlinear projection from a large data space to a positioned along the first and last rows (Fig. 1k), used the same number
small grid of neurons [55]. The original data space (input layer) is of WTs, while all studies following [18] largely increased this number.
projected to a grid of neurons (output layer) while maintaining the All Authors employed WTs with the same Hhub value (60 m), except
topological and metric relationships in the original data. The SOM can MirHassani & Yarahmadi [26], who also used WTs with Hhub = 78 m
recognize groups of similar input variables so that neurons having (Fig. 1h), and a combination of 8 WTs with Hhub = 50 and 17 WTs with

Fig. 1. Analysed literature gridded layouts (Δh = Δv = 200 m; Lh = Lv = 2 km): (a) Mosetti et al. [18]; (b) Grady et al. [50]; (c) Emami & Noghreh [56]; (d) Turner
et al. [57]; (e) Patel et al. [25]; (f) Parada et al. [1]; (g) MirHassani & Yarahmadi [26]; (h) MirHassani & Yarahmadi [26]; (i) MirHassani & Yarahmadi [26]; (j)
Abdelsalam & El-Shorbagy [20]; (k) Ulku & Alabas-Uslu, ‘a’ [10]; (l) Ulku & Alabas-Uslu, ‘b’ [10]; (m) Ulku & Alabas-Uslu, ‘c’ [10]; (n) Ulku & Alabas-Uslu, ‘d’ [10].
Hhub is 60 m for all layouts except 78 m (h) and 50 & 78 m (i). Wind conditions as in Fig. S1 apply for all layouts.

5
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

Table 1
WFLO literature studies analysed and compared.a,b,c,d
Layout Study Ref. Design Hhub (m) No WTs η (%) Pe (kW/y) Fitness (×10−4)

Mosetti et al. [18] Fig. 1a 60 15 84 3,695 36.1


(b) Grady et al. [50] Fig. 1b 60 39 86.62 32,038 8.031
(c) Emami & Noghreh [56] Fig. 1c 60 28 91 32,262
(d) Turner et al. [57] Fig. 1d 60 39 32,453
(e) Patel et al. [25] Fig. 1e 60 39 91.40 33,810 7.96
(f) Parada et al. [1] Fig. 1f 60 40 93.62 34,173 7.875
(g) MirHassani & Yarahmadi [26] Fig. 1g 60 39 50,639
(h) MirHassani &Yarahmadi [26] Fig. 1h 78 25 36,832
(i) MirHassani &Yarahmadi [26] Fig. 1i 50 & 78 25 34,132
(j) Abdelsalam & El-Shorbagy [20] Fig. 1j 60 41 34,461 8.155
(k) Ulku & Alabas-Uslu, ‘a’ [10] Fig. 1k 60 15 12,103 10.5
(l) Ulku & Alabas-Uslu, ,’b’ [10] Fig. 1l 60 28 20,776 9.5
(m) Ulku & Alabas-Uslu, ‘c’ [10] Fig. 1m 60 39 26,932 10.0
(n) Ulku & Alabas-Uslu, ‘d’ [10] Fig. 1n 60 46 30,424 10.4

a
For each layout, original characteristics and scores are reported.
b
Values are missing where not reported in the works.
c
Eq. (15) used to calculate η.
d
Wind conditions as in Fig. S1 apply for all layouts; for normalization purposes, they also apply for layouts (g), (h) and (i), although in the original work
zref = 78 m was set [26].

Hhub = 78 m (Fig. 1i) being the only case with multiple Hhub values. As
a result of the optimization algorithms that aimed at maximizing the Table 2
downwind distance between the WTs, in layouts (e) and particularly (j) Assumptions and constraints for WF layouts.
the majority of WTs are positioned towards the outer boundaries of the Section Description Parameter Value(s)
WF [25]. In layout (d) WTs are positioned according to a quite regularly
staggered design very similar to that of an array layout. Conversely, Site
Location onshore
layout (f) is scattered and irregular, with WTs mostly located in the
Topography flat
outermost zone of the WF, particularly on the prevailing wind direc- Elevation zsite 0 m ASL
tions (270–350°) [1]. The optimized layouts proposed by Ulku & Roughness length z0 0.30 m
Alabas-Uslu [10], particularly (m) and (n), are structured in arrays with Wind shear coefficient 0.15
WTs aligned along the grid columns. Apart from layout (a), the fitness Air density ρ 1.225 kg/m3
value – where available – exhibits a narrow range (7.875–10.5x10-4), Wind conditions
with layout (f) returning the lowest value. Wind case study multi-directional wind with
variable wind speed
(vm = 14 m/s, Fig. S1)
4. Assumptions and constraints Reference height zref 60 m AGL
Vertical profile v(z) varying based on PL (Eq. (20))
Assumptions and constraints, derived from the literature, apply for
Wind turbines
both (i) recomputation and (ii) optimization of WFLO literature layouts No. models NT
(Table 2). (i) layouts 1
recomputation
(ii) layouts 39
4.1. Site characteristics and wind conditions
optimization
No. combinations NTH
Site characteristics and wind conditions are basically the same as (i) layouts 1 (2)a
those assumed in [33]. Following [49], =0.15 was considered herein. recomputation
Wind scenario ‘c’ defined in [18] – plotted in Fig. S1 of the supple- (ii) layouts 120
optimization
mentary material – has been assumed, which refers to a height (zref) of Thrust coefficient CT varying with wind speed (Eq.
60 m AGL. To normalize all compared studies, zref = 60 m was also (19))
assumed for layouts (g), (h) and (i), although in the original work Power curve experimental
zref = 78 m was set [26]. Wind values at zref are adjusted to the proper WF layout
Hhub of any WT by applying the PL (Eq. (20)) with = 0.15. Although a Design gridded (row/column)
theoretical one, this wind scenario is an accepted reference for most Size (Fig. 1)
horizontal full size Lh 2000 m
grid-based WFLO studies (e.g. [58]). It was assumed herein to nor-
vertical full size Lv 2000 m
malize comparison among all literature studies and to retain the scores horizontal grid cell size Δh 200 m
originally achieved therein. In any case, this wind scenario is the most vertical grid cell size Δv 200 m
realistic among the three defined in [18]. On the other hand, refined WT placement study-dependent (Fig. 1)
wind condition settings such as the one proposed by Haces-Fernandez WT minimum spacing 4 rotor diameters
WT selection criteria
et al. [59], where wind speed and direction were segmented into bins to
(i) layouts recomputation
create a 2-D matrix with all possible combinations, were not considered Manufacturer and the WT as in Table 4 (case ‘a’)
since they are beyond the scope of this work. model
(ii) layouts optimization
Manufacturer(s) the same for the whole WF
4.2. Wind turbines
Model(s) the same WT along the row,
and varied along the column
The WT database developed in the computation tool detailed in [60]
was used. This database has since been upgraded to include over 350 a
Values in brackets refer to layout (i) in Table 1.

6
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

onshore commercial models. WT characteristics were retrieved from 2002–2020 inflation rate. For WT-related losses, the fx percentages
manufacturers’ websites as well as from online WT databases such as depending on WT system, derived from [41] and given in Table S2,
Wind turbines models [46] and WindPower [61]. were used in Eq. (16); for site-specific losses, fgrid = 2%, fice = 2% and
It is accepted (e.g. [32]) that minimum crosswind and particularly fother = 0% were set in Eq. (17). To calculate LCoE, d = 5% and
downwind WT spacing should be observed in order to mitigate wake n = 20 years were set in Eq. (25). This procedure generates two out-
interactions between WTs in the WF. This constraint should also be puts: (i) a summary spreadsheet (with NWF records), listing character-
considered for any wake model’s application. For example, Jeon et al. istics and scores of all generated layouts; (ii) NWF spreadsheets, one for
[34] demonstrated that the errors in predicting velocity deficit by each layout, reporting characteristics and scores of each WT in the
Jensen’s wake model can be accepted for a downstream WT spacing farm. Summarizing, NWF = 1,758 layouts are generated and analysed
above 3.75D. Therefore, consistently with Pookpunt and Ongsakul by the WFLO procedure for each of NLIT literature layouts. Since
[32], a minimum downwind distance of 4D was assumed between ad- NLIT = 14, the overall number of layouts is:
jacent cells in the gridded layout: since Δh = Δv = 200 m, this con- NLAY = NWF*NLIT = 1,758*14 = 24,612.
straint yielded a maximum D = 50 m (Table 2). Furthermore, a A far more simplified WFLO procedure is activated when re-
minimum Pr = 200 kW was set for withdrawing the very small WT computing the original literature studies as, for each case, only one
models. Eventually, a total of NT = 39 WTs have been used, extracted single WT (at most with two Hhub values) is imported from the WT
from the WT database after setting D ≤ 50 m and Pr ≥ 200 kW. Their database, and only one layout (NWF = 1) is generated and then ana-
characteristics are reported in Table 3. Notably, for any WT, experi- lysed: in this case, the number of processed layouts (NLAY = NWF*NLIT)
mental power curves have been used rather than approximated (e.g. is 14.
[32]), best-fitted (e.g. [3]), or theoretically-constructed (e.g. [18] and
all following grid-based WFLO studies) power curves. Since WTs with 6. Results and discussion
multiple hub heights were handled as distinct WTs, the overall number
of WT combinations was actually NTH = 120. The considered literature layouts have been recomputed and then
optimized. In doing so, real WT models (exhibiting experimental power
4.3. Wind farm layouts curves supplied by manufacturers) have been considered instead of the
theoretical WT model (featuring the theoretical power curve defined in
WF layouts are presented in Fig. 1. When literature studies are re- Eq. (13)) originally used in each of those studies. For the same reason,
computed, a single WT model is used. When layouts are optimized, the the more general Eq. (12) rather than Eq. (15) was used to calculate WF
same WT selection criterion proposed in [33] is followed, which is efficiency. Key factors that influence layout optimization the most have
detailed in the following. Once each WT is retrieved from the database, also been investigated.
all possible WT combinations obtained by varying model, D and Hhub
are implemented. For any layout instance, a common manufacturer is 6.1. Recomputation of literature layouts
chosen for all WTs to be installed in the farm. WTs are the same for each
row, while they are varied from one row to the other. The particular The optimal layouts achieved in the original studies (Table 1) have
case where a unique WT model is installed throughout the farm is been recomputed by considering the same application conditions
considered, too. (Table 2). In doing so, the Nordex N43-600 wt has been selected,
proving to be the model within the WT database that best-fits the
5. Wind farm layout optimization procedure characteristics of the theoretical WT defined in [18]. WT characteristics
are detailed as (a) in Table 4, while the WT power curve is shown in Fig.
The applied WFLO procedure (Fig. S2) consists of two stepwise S3a. Although not the same model as in [18], this WT offers the ad-
Fortran modules: a layout generator, and a layout analyser. The same vantage of having two Hhub values in addition to 60 m (i.e. 50 and 78 m)
procedure is used to perform both layout recomputation and optimi- which enable to also run the layouts (h) and (i) proposed in [26].
zation. This procedure is a variation of the one developed in [33]. Table 5 summarizes the scores achieved after recomputing all lit-
In the most general case (i.e., the optimization approach), based on erature layouts by employing the Nordex N43-600 wt It is apparent
the WT selection criteria, the layout generator ingests from the database that, irrespective of the number of installed WTs, the overall WF ca-
all WTs with D ≤ 50 m and Pr ≥ 200 kW. Each layout in Fig. 1 is then pacity and corresponding energy production, all layouts return very
loaded to enable allocation of each WT accordingly. After processing all similar scores of η and CF, as shown by their very narrow ranges:
available WT combinations, the module generates NWF layouts, each 98.81–99.61% (η), and 78.89–80.90% (CF). By contrast, a very wide
including sequential code, database identifier and coordinates of all range is returned for Pe (7,101–21,773 kW/y) and LCoE
allocated WTs. During the second stage, all previously generated lay- (130.37–370.42 $/MWh). Overall, it should be noted that – except for
outs are ingested by the layout analyser, with characteristics of any WT layout (a) – significantly higher power output was achieved in layouts
retrieved from the database. The layout analyser performs the compu- optimized in the original studies (Table 1). These findings support the
tation of CT(v), vertical extrapolation of wind speed, Jensen’s model hypothesis that the heuristic algorithms implemented within each study
and corresponding losses/efficiency, losses linked to the site and WT, likely optimized all such layouts while targeting both Pe and CF max-
energy output scores, and cost model. The NREL cost model [54] was imization rather than LCoE minimization.
used to calculate LCoE after setting [$] as the currency and the average Layout (a) returns the lowest LCoE (130.37 $/MWh), while layout

Table 3
Summary of characteristics of all processed onshore WTs.a
Score Rated power Rotor diameter Hub height Wind speeds Design ratio

Cut-in Rated Cut-off


Pr (kW) D (m) Hhub (m) vi (m/s) vr (m/s) vo (m/s) Φ

Median value 600 43 50 3.5 14.0 25.0 4.3


Range 220–900 26–50 30–78 2.5–5.0 11.0–18.0 20.0–28.0 3.1–5.3

a
Total WT models: NT = 39. Total WT combinations (sorted by Hhub): NTH = 120.

7
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

Table 4
Characteristics of WTs used or achieved while running all WFLO literature layouts.a
Approach Application condition Model Layout(s) Pr (kW) D (m) Hhub (m) vi (m/s) vr (m/s) vo (m/s) Systemb
(case)

(i) layouts recomputation


(a) Recomputing Nordex All 600 43 50 3 13.5 25 A
N43-600 60
78

(ii) layouts optimization


(b) Minimizing LCoE Vergnet (a) 220 26 60 3.5 17 20 A
GEV 26/220
(c) Maximizing CF Windtec 650 (a) 600 50 71.5 3.5 11 20 C

a
Power curves for each WT are plotted in Fig. S3.
b
Coding for WT systems is reported in Table S2 [41].

(h) the highest CF (80.90%). The number of installed WTs significantly 6.2. Optimization of literature layouts
affects the results, as shown by the near matching scores from 28- wt
layouts (c) and (l), as well as those from 39- wt layouts (b), (d), (e), (g) For each literature layout, a total of NWF = 1758 layout instances
and (m). Scores from 15- wt layout (k) are very similar to those from have been generated, ultimately allowing the detection of only one
15- wt layout (a), too, particularly as minimum Pe and LCoE are basi- solution that satisfies the selected optimization function. Table 6 shows
cally the same, and values of all other indicators are only slightly worse. the comparative results of layouts that minimize LCoE, while Table 7
Pe scores are also a strict function of the number of installed WTs. shows those of layouts that maximize CF.
Contrary to suggestions by MirHassani and Yarahmadi [26] that using All recomputed literature layouts (Table 5) have been optimized by
WTs with different Hhub leads to an increase in the power output, with the present optimization method, as both LCoE values have been re-
Pe = 11,962 kW/y layout (i) does not improve performances of the duced and CF values increased. While LCoE reduction is substantial
layout (h), which has the same design but WTs with identical Hhub. (130.37–370.42 to 54.01–142.64 $/MWh, Table 6), CF increase is
From the scores in Table 5, it is apparent that LCoE is a much more marginal (78.89–80.90 to 83.02–83.07%, Table 7). It is worth noting
straightforward economic indicator than the fitness value, because not that for all literature case studies and both optimization conditions, all
only does it exhibit a wider variation range than the latter (Table 1), but optimal layouts have been achieved by using identical WTs over the
it also completely reverses the ranking of (a) from the worst (Table 1) to whole WF: this means that a unique combination for Pr, D and Hhub
the best (Table 5) layout. (NP = ND = NH = 1) applies. Furthermore, for all case studies, the
Significantly improved η values are achieved with respect to the same WT model proved to minimize LCoE, and the same WT model to
original applications (Table 1, where reported), where, however, Jen- maximize CF. As shown in Table 4, LCoE minimization is accomplished
sen’s wake model was applied using CT = 0.88. To assess the differ- by using the Vergnet GEV 26/220 wt (whose power curve is plotted in
ences against the present study where the CT(v) approach is im- Fig. S3b), while CF maximization is achieved by using the Windtec
plemented, the CT = 0.88 option has been also analysed herein (see 650 wt (power curve in Fig. S3c). With respect to layouts recomputa-
Table S3). It is apparent that η values are lower (overall ranging tion, LCoE minimization is accomplished by reducing Pr and D, and
94.69–98.18%), thus indicating that using CT = 0.88 is more con- concurrently increasing vr in the selected WT (Table 4), while CF
servative than using CT(v). The impacts are however quite marginal on maximization is accomplished – for the same Pr – by increasing D and
Pe (7,025–21,512 kW/y), CF (77.94–80.50%), and LCoE decreasing vr . These outcomes agree with suggestions, e.g., by Abdul-
(130.50–371.29 $/MWh). rahman and Wood [7], that small-medium sized WTs with low hub
height should be installed to minimize LCoE.

Table 5
Comparative results of WF literature layouts recomputed using a single WT model.a,b,c,d
Layout Ref., and Hhub Cumulated Mean Cumulated Mean Cumulated

No WTs P (MW) η (%) Pe (kW/y) AEY (MWh/y) CF (%) LCoE $/MWh)

(a) Mosetti et al. [18], 60 m 15 9 99.61 7116 62,335 79.06 130.37


(b) Grady et al. [50], 60 m 39 23.4 99.21 18,480 161,884 78.97 316.16
(c) Emami & Noghreh [56], 60 m 28 16.8 99.35 13,273 116,271 79.01 231.00
(d) Turner et al. [57], 60 m 39 23.4 99.23 18,481 161,897 78.98 316.16
(e) Patel et al. [25], 60 m 39 23.4 99.08 18,473 161,826 78.95 316.18
(f) Parada et al. [1], 60 m 40 24 99.13 18,950 165,998 78.96 323.92
(g) MirHassani & Yarahmadi [26], 60 m 39 23.4 99.04 18,471 161,806 78.94 316.19
(h) MirHassani & Yarahmadi [26], 78 m 25 15 99.41 12,135 106,301 80.90 207.38
(i) MirHassani & Yarahmadi [26], 50 & 78 m 25 15 99.53 11,962 104,784 79.74 207.70
(j) Abdelsalam & El-Shorbagy [20], 60 m 41 24.6 99.05 19,419 170,112 78.94 331.67
(k) Ulku & Alabas-Uslu [10], ‘a’, 60 m 15 9 98.90 7,101 62,209 78.91 130.43
(l) Ulku & Alabas-Uslu [10], ‘b’, 60 m 28 16.8 98.81 13,253 116,094 78.89 231.08
(m) Ulku & Alabas-Uslu [10], ‘c’, 60 m 39 23.4 98.95 18,466 161,766 78.92 316.21
(n) Ulku & Alabas-Uslu [10], ‘d’, 60 m 46 27.6 98.82 21,773 190,732 78.89 370.42

a
Characteristics of each WF layout are depicted in Fig. 1.
b
Eq. (12) used to calculate η.
c
Wind conditions as in Fig. S1 apply for all layouts; for normalization purposes, they also apply for layouts (g), (h) and (i), although in the original work
zref = 78 m was set [26].
d
The same WT (Table 4, case ‘a’) is used for all WFLO studies.

8
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

Table 6
Comparative results of WF literature layouts optimized by LCoE minimization.a,b,c,d
Layout Ref., and Hhub Cumulated Mean Cumul. Mean Cumul.

No WTs P (MW) η (%) AEY (MWh/y) CF (%) LCoE ($/MWh)

(a) Mosetti et al. [18], 60 m 15 3.30 99.63 21,928 75.86 54.01


(b) Grady et al. [50], 60 m 39 8.58 99.22 56,945 75.76 122.61
(c) Emami & Noghreh [56], 60 m 28 6.16 99.37 40,902 75.80 91.17
(d) Turner et al. [57], 60 m 39 8.58 99.25 56,950 75.77 122.60
(e) Patel et al. [25], 60 m 39 8.58 99.10 56,924 75.74 122.62
(f) Parada et al. [1], 60 m 40 8.80 99.14 58,392 75.75 125.47
(g) MirHassani & Yarahmadi [26], 60 m 39 8.58 99.05 56,917 75.73 122.62
(h) MirHassani & Yarahmadi [26], 78 m 25 5.50 99.36 36,519 75.80 82.59
(i) MirHassani & Yarahmadi [26], 50 & 78 m 25 5.50 99.39 36,522 75.80 82.59
(j) Abdelsalam & El-Shorbagy [20], 60 m 41 9.02 99.07 59,839 75.73 128.33
(k) Ulku & Alabas-Uslu [10], ‘a’, 60 m 15 3.30 98.90 21,882 75.69 54.05
(l) Ulku & Alabas-Uslu [10], ‘b’, 60 m 28 6.16 98.82 40,835 75.67 91.20
(m) Ulku & Alabas-Uslu [10], ‘c’, 60 m 39 8.58 98.96 56,902 75.71 122.63
(n) Ulku & Alabas-Uslu [10], ‘d’, 60 m 46 10.12 98.83 67,090 75.68 142.64

a
Characteristics of each WF layout are depicted in Fig. 1.
b
Eq. (12) used to calculate η.
c
Wind conditions as in Fig. S1 apply for all layouts.
d
The same WT (Table 4, case ‘b’), installed over the whole WF, achieved to optimize all layouts.

Table 7
Comparative results of WF literature layouts optimized by CF maximization.a,b,c,d
Layout Ref., and Hhub Cumulated Mean Cumul. Mean Cumul.

No WTs P (MW) η (%) AEY (MWh/y) CF (%) LCoE ($/MWh)

(a) Mosetti et al. [18], 60 m 15 9.00 99.63 65,494 83.07 125.03


(b) Grady et al. [50], 60 m 39 23.40 99.25 170,232 83.05 309.33
(c) Emami & Noghreh [56], 60 m 28 16.80 99.38 122,231 83.06 224.86
(d) Turner et al. [57], 60 m 39 23.40 99.28 170,236 83.05 309.33
(e) Patel et al. [25], 60 m 39 23.40 99.14 170,216 83.04 309.33
(f) Parada et al. [1], 60 m 40 24.00 99.18 174,587 83.04 317.01
(g) MirHassani & Yarahmadi [26], 60 m 39 23.40 99.10 170,211 83.04 309.34
(h) MirHassani & Yarahmadi [26], 78 m 25 15.00 99.38 109,134 83.06 201.82
(i) MirHassani & Yarahmadi [26], 50 & 78 m 25 15.00 99.41 109,137 83.06 201.82
(j) Abdelsalam & El-Shorbagy [20], 60 m 41 24.60 99.11 178,941 83.04 324.69
(k) Ulku & Alabas-Uslu [10], ‘a’, 60 m 15 9.00 98.98 65,459 83.03 125.04
(l) Ulku & Alabas-Uslu [10], ‘b’, 60 m 28 16.80 98.90 122,182 83.02 224.87
(m) Ulku & Alabas-Uslu [10], ‘c’, 60 m 39 23.40 99.02 170,200 83.03 309.34
(n) Ulku & Alabas-Uslu [10], ‘d’, 60 m 46 27.60 98.90 200,729 83.02 363.10

a
Characteristics of each WF layout are depicted in Fig. 1.
b
Eq. (12) used to calculate η.
c
Wind conditions as in Fig. S1 apply for all layouts.
d
The same WT (Table 4, case ‘c’), installed over the whole WF, achieved to optimize all layouts.

Table 6 shows that layout (a) (termed “L1”) returns the lowest LCoE continuous model and that, consistent with reviews, e.g., by Serrano
value (54.01 $/MWh), yet only slightly lower than layout (k) which et al. [2] or Shakoor et al. [4], a superior model cannot be clearly as-
achieves a score very close to this. This value is consistent with the certained.
global average LCoE value (60 $/MWh) reported for 2017-updated The CF values maximized in Table 7 match among all the considered
onshore wind projects by IRENA [52]. Using observed wind data and studies (83.02–83.07%), thus reinforcing the hypothesis raised in sec-
applying the NREL cost model, at the real site of Khaf in Iran (40-m tion 6.1 that optimization of such literature layouts tends to a common
vm = 10.5 m/s) Mirghaed and Roshandel [31] designed a square value, corresponding to maximum CF. Regardless, layout (a) (termed
(1.73 km × 1.73 km) layout with crosswind/downwind WT distances “L2”) also proves to be the best solution for maximizing CF. Comparing
of 3D × 6D: notably, they obtained an LCoE value of 55 $/MWh very Tables 6 and 7, it is apparent that the LCoE minimization target com-
close to the minimum values (54.01–54.05 $/MWh) obtained in the bines with satisfactory scores of CF (75.67–75.86%, Table 6), while this
current study (40-m vm = 13.2 m/s). Conversely, current LCoE minima is not the case for CF maximization, which overall is not associated to
are larger than those (21–31 $/MWh) calculated by Rahbari et al. [5] at particularly profitable scores of LCoE (125.03–363.10 $/MWh,
a real site in Teheran based on observed wind data (10-m vm = 6.38 m/ Table 7). Current CF values are lower than the maximum CF value
s), where they designed four different rectangular layouts (sized (92.17%) achieved in the previous WFLO study based on the continuous
1.5 km × 2.5 km) with WTs irregularly allocated. Current LCoE scores model [33]: although indirectly, this outcome is consistent with find-
are either lower or higher than minimum LCoE value (103.46 $/MWh) ings by Long et al. [12] that the continuous model is better than the
achieved in the previous WFLO study, based on the continuous model discrete model when maximizing the power output is set as WFLO
and run under quite similar application conditions (i.e. same site fea- target.
tures, layout shape and size, and wind conditions) [33]. This indicates Characteristics and performances of WTs installed in layouts opti-
that the discrete model can perform either better or worse than the mized by LCoE minimization and CF maximization are detailed,

9
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

respectively, in Table S4 (L1) and Table S5 (L2), while their placement Cini/Pr (Fig. S4n), and CF (Fig. S4p), meaning that – for any WT com-
within the WFs is depicted in Fig. S5. It is apparent that L1 and L2 bination in the WF – all these parameters are insensitive to the selected
correspond to the same layout (a), except that they employ two dif- layout design. This outcome is particularly meaningful for CF, as it
ferent WT models. As reported in Table 4, such WT models belong to confirms that whatever the chosen literature layout, this score does not
systems A and C as defined in Table S2, i.e. those accounting for the appreciably change. Conversely, patterns are layout-dependent for P
lowest overall WT-specific losses. If considering characteristics of all (Fig. S4d), total losses (Fig. S4m), AEY (Fig. S4o), and LCoE (Fig. S4q),
available WTs (Table 3), LCoE minimization (Table S4) is achieved meaning that choosing a specific layout dramatically affects WF power
using a WT exhibiting the lowest Pr and D, and almost maximum vr. production and particularly the project’s economic viability. 15- wt
These outcomes agree with suggestions by Chen et al. [40] that higher layouts (a) and (k) exhibit a similar pattern for P, AEY, and LCoE, with
vr values are required by sites with higher vm and result in lower WT- main discrepancies apparent in η. For layouts (k) and (l) there is a re-
related LCoE, and that larger WTs do not necessarily reduce LCoE. The levant frequency of layout instances built-up with a twofold combina-
values of Ω (84.29–84.47%) indicate that on average the WT operates tion of Pr (Fig. S4a) and D (Fig. S4b), while for layout (k) this also
below its rated speed. By contrast, CF maximization (Table S5) is involves Hhub values (Fig. S4c). Layout (a) proves to be the best solution
achieved through a WT of medium Pr, with minimum vr and Φ, and Hhub overall for maximizing η, minimizing total losses, and – most im-
close to the maximum (Table 3). Corresponding Ω values portantly – minimizing LCoE. By contrast, layout (a) returns the lowest
(130.73–131.00%) indicate that to a remarkable extent (over 30%) the overall AEY values. It is worth noting that the plots of AEY and LCoE
WT operates above its rated speed, thus largely exceeding the threshold can almost perfectly be overlapped. In terms of total power losses (Fig.
over which the WF overall cost increase is adequately compensated for S4m), wake losses are almost negligible (Fig. S4k), particularly if
by the revenues deriving from energy production. compared to losses linked to the site and WT (Fig. S4l). Furthermore, in
the majority of layout instances, WTs operate in wind regimes below
6.3. Insights into wind farm layout optimization key factors their rated speed (Fig. S4j).
After lumping all literature layout instances together, a correlation
Optimization functions and key factors that mostly affect WFLO analysis between all parameters has been performed (Table S6). Since
have been investigated after applying kernel density plots, correlation no correlation exists between CF and LCoE (r = −0.03), increasing CF
analysis and self-organizing maps to main WFLO parameters. can either reduce or increase LCoE. As shown by Figs. S2o and S2q, AEY
and LCoE are strongly correlated (r = 0.99): therefore, for a high wind
6.3.1. Kernel density and correlation analysis potential site, WFLO is surprisingly achieved by reducing the total en-
All layout instances generated per literature study have been ana- ergy production. LCoE is also strongly correlated to P (r = 1.00), thus
lysed in terms of the distribution of main WFLO parameters. The result indicating that the overall installed power rather than the mere number
of this analysis is summarized through the kernel density plots shown in of WTs (r = 0.84) should be reduced for reducing LCoE. Pr affects LCoE
Fig. S4, drawn by using the “sm.density.compare” tool [62] im- more than vr (r of 0.51 vs. –0.02), which contrasts with suggestions by
plemented in the “sm” R package [63]. For all literature layouts, quite Chen et al. [40] that vr is more influential on LCoE than Pr. Distributions
similar distributions may be observed for Pr (Fig. S4e), D (Fig. S4f), Hhub of Ci and AEY are strongly correlated as well (r = 1.00), meaning that
(Fig. S4g), vr (Fig. S4h), Φ (Fig. S4i), Ω (Fig. S4j), Fsite & FWT (Fig. S4l), annual expenditures increase with WF energy production. Total losses

Fig. 2. SOM component planes obtained while optimizing literature layouts for the following parameters: No. WTs, NP, ND, NH, P, Pr, D, Hhub, vr, Φ, Ω, η, Fsite & FWT,
Ftot, Cini/Pr, AEY, CF, and LCoE. Cumulated values are plotted for No. WTs, P and LCoE, while WF-averaged values for all other parameters. A total of 1,758 layouts by
14 literature case studies (NLAY = NWF*NLIT = 24,612) were processed.

10
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

are strongly correlated to site- and WT-related losses (r = 0.96), and 6.3.1), the SOMs of Ftot basically match with those of Fsite & FWT. Since
poorly correlated to η, and thus to array losses (r = 0.33): since site- the SOMs of η confirm that wake losses are almost negligible, it is also
specific losses do not change, WT-related losses are the component that confirmed that WF total losses are mainly driven by losses depending on
primarily affects total losses, meaning that the system the selected WT WT characteristics: therefore, the selection of WTs returning the lowest
belongs to (see Table S2) is pivotal in affecting WF overall losses. On the losses (see Table S2) should be prioritised. In any case, reducing total
other hand, a significant correlation (r = 0.86) was ascertained be- losses positively affects CF but not LCoE, as CF increases do not ne-
tween Ω and CF, yet it was not as strong (e.g., r > 0.95) as to imply cessarily result in LCoE decreases.
that Ω gives the same level of information as CF. The highest energy production may be obtained by the most ex-
pensive layouts, i.e. those that maximize the overall WF installed ca-
6.3.2. The self-organizing maps pacity and employ WTs with the highest values of Pr, D and Hhub.
To more thoroughly investigate WFLO key factors, an advanced However, AEY maximization is economically disadvantageous as they
analysis than both kernel density and correlation analysis (Section correspond to the highest values of LCoE. This means that maximizing
6.3.1) was accomplished through the use of SOMs. A dataset including the overall WF power output, as pursued within several studies (e.g.
NWF = 1758 layout instances multiplied by NLIT = 14 case studies, for a [57]), is a misleading WFLO goal.
total of NLAY = NWF*NLIT = 24,612 layouts, was used. The following 18 Summarizing, the SOMs proved useful in improving the multi-
variables were selected as component planes of the SOMs: No. WTs, NP, variable investigations addressed in the past literature either dealing
ND, NH, P, Pr, D, Hhub, vr, Φ, Ω, η, Fsite & FWT, Ftot, Cini/Pr, AEY, CF, and with WFLO (e.g. [31]), or WT optimal site matching (e.g. [40]). They
LCoE. Ci was not considered since its distribution is the same as AEY are also particularly helpful when WFLO problems involving multiple
(r = 1.00, Table S6). The following final SOM structure was achieved objective functions (as performed, e.g., by Abdulrahman and Wood [7])
for the above 18 variables: (i) an input layer, including 24,612 neurons are addressed. Since they are capable of representing the relationships
corresponding to all layout instances; (ii) an output layer, including 100 among main WF layout parameters through a single, compact plot, the
neurons, represented through 10x10 gridded hexagons. Before SOM SOMs allow to achieve a straightforward WFLO pattern, thus efficiently
training, all variables’ values were normalized such that their mean was coping with the non-convex, highly complex, multi-purpose and multi-
0 and their variance was 1 [64]. The SOMs were trained following the variable nature of the WFLO problem.
requirements of the Kohonen’s algorithm [55]: thus, 60,000 iterations
(over 500 times the number of gridded neurons) were used, along with 7. Conclusions
initial values of 0.9 (close to 1) for learning rate and 7 units (higher
than grid’s radius) for neighbourhood radius. The Mexican hat was A total of 14 grid-based layouts addressed in the literature and run
selected as neighbourhood function since it results in a larger variables’ based on wind scenario ‘c’ defined in Mosetti et al. [18] have been
variation range than the most widely used Gaussian function, while the recomputed. Several application improvements have been performed to
goodness of SOM was assessed using the Kaski and Lagus error metrics the original settings, including the use of: (i) commercial-scale WTs
[65], that combine quantization and topographic errors. The graphical instead of a theoretical WT; (ii) experimental instead of theoretical WT
output of SOM application, performed by using the Living For SOM tool power curves; (iii) CT varying with wind speed rather than the 0.88
(http://livingforsom.com), is shown in Fig. 2. constant value; (iv) LCoE as an objective function rather than the simple
In SOM topology of LCoE, lower-to-higher neurons are arranged fitness value; (v) the detailed NREL cost model in place of the overly-
bottom to top, with minimum and maximum values located centrally. A simplistic cost model defined in [18]. All layouts returned very similar
similar topology is exhibited by SOMs of P and AEY, and – to a lesser scores of farm efficiency (98.81–99.61%) and CF (78.89–80.90%), but
extent – Pr, D, and No. WTs: this means that LCoE may be minimized by widely varying scores of LCoE (130.37–370.42 $/MWh): this indicates
decreasing the overall WF installed capacity, the energy production, that CF maximization rather than LCoE minimization was the actual
rated powers, rotor diameters or the number of WTs. These scores target pursued by the heuristic algorithms implemented within each
confirm findings achieved in section 6.2 that lie behind the optimiza- WFLO study. Results indicated the 15- wt layout by Mosetti et al. [18]
tion of layout L1 (Table S4). The association of LCoE with NP, ND and NH as the optimal in minimizing LCoE (130.37 $/MWh), and the 25- wt
is weaker, as shown by their different SOM topologies: LCoE minima are layout by MirHassani and Yarahmadi [26] as the optimal in maximizing
associated to lower but not minimum values of NP, ND and NH, which CF (80.90%). As such, results achieved from this layout recomputation
are located in the centre of their respective maps. In any case, this proved to be a novelty in themselves, as they were quite different from
means that LCoE could be reduced by avoiding to install WTs with those returned from the original studies, which adopted energy yield
multiple Pr, D or Hhub (as accomplished for layout L1). This finding maximization and/or fitness value minimization as an objective func-
seems to contradict those from various Authors, suggesting that it is tion.
more profitable to suitably combine different WTs rather than installing In the present study, relying on 39 commercial WTs, a previously
identical WTs, particularly having different D [66] or different Hhub developed optimization method [33] was applied which targets optimal
[58]. Since the WT combinations in the farm reduce to just one for each WT selection while retaining the original number of WTs and their
WT per manufacturer, WT selection criteria set in section 4.3 are dra- placement as an assumption. Previously applied to the continuous WF
matically simplified. layout model, the method has also been successfully applied to the
Notably, CF neurons do not arrange in a mirrored fashion with re- discrete layout model. For each literature case study, a total of 1,758
spect to LCoE neurons, suggesting that maximizing CF does not imply layout instances were generated, overall resulting in 24,612 layouts.
minimizing LCoE: confirming the null correlation between CF and LCoE Method’s application proved to optimize all literature layouts, as CF
(section 6.3.1), a CF increase may either result in an LCoE increase or values were (slightly) increased (78.89–80.90 to 83.02–83.07%), while
decrease. Rather, SOM clustering of CF shows a neurons arrangement LCoE values were (significantly) reduced (130.37–370.42 to
quite similar to the ones of Hhub, Φ, Ω, and (mirrored) vr, suggesting that 54.01–142.64 $/MWh). It is worth noting that the layout by Mosetti
an increase (decrease for vr) in such variables leads to an increase in CF. et al. [18] proved to be the best both in minimizing LCoE (54.01
These results were anticipated in section 6.2 when analysing the $/MWh) and in maximizing CF (83.07%). Furthermore, for all litera-
characteristics of optimal layout L2 (Table S5). As for the SOM of Ω, it is ture case studies and both optimization conditions, all optimal layouts
confirmed that values close to 100% do increase CF, but also involve an were achieved by using identical WTs: this means that only one com-
overall energy waste, since resulting in an energy over-production (see bination of rotor diameter, rated power and hub height applies over the
AEY) not adequately compensated by a corresponding revenue (see whole WF. For all case studies, the same WT model (Vergnet GEV 26/
LCoE). Consistent with findings from the correlation analysis (section 220) minimized LCoE, and the same WT model (Windtec 650)

11
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

maximized CF. LCoE scores were either better or worse than those Acknowledgments
achieved in the previous WFLO study, implementing the continuous
model and run under similar application conditions [33]: therefore, a No funds were received to support the research conducted in this
model – be it discrete or continuous –cannot be recommended since paper.
neither clearly outperforms the other.
A thorough analysis has been performed to investigate optimization Appendix A. Supplementary data
functions and the parameters that mostly influence WFLO. Although
specifically applying to high wind potential sites where mid-sized WTs Supplementary data to this article can be found online at https://
are supposed to be installed, various conclusions can be made: doi.org/10.1016/j.enconman.2020.112593.

• the two optimization conditions are not associated, as a CF increase References


does not necessarily imply an LCoE decrease;
• LCoE can be minimized by decreasing farm overall capacity or the [1] Parada L, Herrera C, Flores P, Parada V. Wind farm layout optimization using a
Gaussian-based wake model. Renew Energy 2017;107:531–41.
number of WTs, as well as selecting WTs with lower rotor diameters
[2] Serrano González J, Burgos Payán M, Riquelme Santos JM, González-Longatt F. A
or rated powers;

review and recent developments in the optimal wind-turbine micro-siting problem.
conversely, CF maximization can be achieved by installing WTs with Renew Sustain Energy Rev 2014;30:133–44.
higher hub heights or lower rated speeds; [3] Chowdhury S, Zhang J, Messac A, Castillo L. Optimizing the arrangement and the

• LCoE can be minimized even using WTs operating on average


selection of turbines for wind farms subject to varying wind conditions. Renew
Energy 2013;52:273–82.
(slightly) below their rated speed; by contrast, matching or ex- [4] Shakoor R, Hassan MY, Raheem A, Wu YK. Wake effect modeling: a review of wind
ceeding their rated speed increases CF, but also results in overall farm layout optimization using Jensen’s model. Renew Sustain Energy Rev
2016;58:1048–59.
energy waste, since it leads to an over-production of energy that is [5] Rahbari O, Vafaeipour M, Fazelpour F, Feidt M, Rosen MA. Towards realistic de-
not adequately compensated by the corresponding revenue; signs of wind farm layouts: application of a novel placement selector approach.
• contradicting various findings, using WTs with different rotor dia- Energy Convers Manage 2014;81:242–54.
[6] Mustakerov I, Borissova D. Wind turbines type and number choice using combi-
meters, rated powers or hub heights is not the best approach for
natorial optimization. Renew Energy 2010;35(9):1887–94.
minimizing LCoE: this dramatically impacts on WT selection criteria, [7] Abdulrahman M, Wood D. Investigating the power-COE trade-off for wind farm
as only one WT combination per manufacturer (i.e. the same model) layout optimization considering commercial turbine selection and hub height var-
is suggested as being installed in the farm instead of other possible iation. Renew Energy 2017;102:267–78.
[8] Ituarte-Villarreal CM, Espiritu JF. Optimization of wind turbine placement using a
combinations; viral based optimization algorithm. Procedia Comput Sci 2011;6:469–74.
• WF total losses are mainly driven by losses depending on WT [9] Yang Q, Hu J, Law SS. Optimization of wind farm layout with modified genetic
algorithm based on boolean code. Wind Eng Ind Aerodyn 2018;181:61–8.
characteristics, while wake losses are almost negligible: however,
[10] Ulku I, Alabas-Uslu C. A new mathematical programming approach to wind farm
reducing total losses results in CF increases, but not necessarily in layout problem under multiple wake effects. Renew Energy 2019;136:1190–201.
LCoE decreases; [11] Long H, Zhang Z. A two-echelon wind farm layout planning model. IEEE Trans
• maximization of WF power generation is a misleading WFLO target, Sustainable Energy 2015;6(3):863–71.
[12] Long H, Zhang Z, Song Z, Kusiak A. Formulation and analysis of grid and coordinate
as it is achieved at the expense of a dramatic LCoE increase: there- models for planning wind farm layouts. IEEE Access 2017;5:1810–9.
fore, when AEY maximization and LCoE minimization became two [13] Castro Mora J, Calero Barón JM, Riquelme Santos JM, Burgos Payán M. An evo-
competing (and conflicting) goals, the latter should be prioritised. lutive algorithm for wind farm optimal design. Neurocomputing
2007;70(16–18):2651–8.
[14] Rajper S, Amin IJ. Optimization of wind turbine micrositing: a comparative study.
The following limitations affect the study, based on considering: (i) Renew Sustain Energy Rev 2012;16(8):5485–92.
an ideal onshore flat and sea-level site; (ii) a theoretical wind scenario; [15] DuPont B, Cagan J, Moriarty P. An advanced modeling system for optimization of
wind farm layout and wind turbine sizing using a multi-level extended pattern
(iii) grid-based layouts with fixed (square) shape, grid density (10x10),
search algorithm. Energy 2016;106:802–14.
and overall size (2 km × 2 km); (iv) the simplified Jensen’s wake [16] Réthoré PE, Fuglsang P, Larsen GC, Buhl T, Larsen TJ, Madsen HA. TOPFARM:
model; (v) a set of mid-sized WTs (D = 26–50 m and Multi-fidelity optimization of wind farms. Wind Energy 2014;17(12):1797–816.
Pr = 220–900 kW). Also, wake-induced fatigue degradation of WTs, [17] Tao S, Xu Q, Feijoo A, Hou P, Zheng G. Bi-hierarchy optimization of a wind farm
considering environmental impact. IEEE Trans Sustainable Energy 2020. https://
and thus more complex cost functions (as addressed, e.g., by Réthoré doi.org/10.1109/TSTE.2020.2964793.
et al. [16]), have been disregarded since they are beyond the scope of [18] Mosetti G, Poloni C, Diviacco B. Optimization of wind turbine positioning in large
this work. In the future, this method could be applied to onshore sites wind farms by means of a genetic algorithm. J Wind Eng Ind Aerodyn
1994;51:105–16.
with complex topography or offshore sites, hopefully being real and not [19] Jensen NO. A note on wind generator interaction. Risø-M-2411, Risø National
ideal sites. Furthermore, to generalize the results achieved thus far, it Laboratory, 1983.
could be helpful to deal with more real-world wind scenarios (e.g. [20] Abdelsalam AM, El-Shorbagy MA. Optimization of wind turbines siting in a wind
farm using genetic algorithm based local search. Renew Energy 2018;123:748–55.
[59]), representative of a wider range of locations or, following DuPont [21] Zergane S, Smaili A, Masson C. Optimization of wind turbine placement in a wind
et al. [15], to analyse the effects resulting from considering different farm using a new pseudo-random number generation method. Renew Energy
atmospheric stability conditions. 2018;125:166–71.
[22] De-Prada-Gil M, Alías CG, Gomis-Bellmunt O, Sumper A. Maximum wind power
plant generation by reducing the wake effect. Energy Convers Manage
2015;101:73–84.
CRediT authorship contribution statement [23] Marmidis G, Lazarou S, Pyrgioti E. Optimal placement of wind turbines in a wind
park using Monte Carlo simulation. Renew Energy 2008;33(7):1455–60.
[24] Wu Y, Zhang S, Wang R, Wang Y, Feng X. A design methodology for wind farm
Giovanni Gualtieri: Conceptualization, Data curation, layout considering cable routing and economic benefit based on genetic algorithm
Investigation, Methodology, Supervision, Validation, Visualization, and GeoSteiner. Renew Energy 2020;146:687–98.
Writing - original draft, Writing - review & editing. [25] Patel J, Savsani V, Patel V, Patel R. Layout optimization of a wind farm to maximize
the power output using enhanced teaching learning based optimization technique. J
Cleaner Prod 2017;158:81–94.
[26] MirHassani SA, Yarahmadi A. Wind farm layout optimization under uncertainty.
Declaration of Competing Interest Renew Energy 2017;107:288–97.
[27] Hayat I, Chatterjee T, Liu H, Peet YT, Chamorro LP. Exploring wind farms with
alternating two-and three-bladed wind turbines. Renew Energy 2019;138:764–74.
The authors declare that they have no known competing financial [28] Chatterjee T, Peet Y. Exploring the benefits of vertically staggered wind farms:
interests or personal relationships that could have appeared to influ- Understanding the power generation mechanisms of turbines operating at different
scales. Wind Energy 2019;22(2):283–301.
ence the work reported in this paper.

12
G. Gualtieri Energy Conversion and Management 208 (2020) 112593

[29] Stanley AP, Ning A, Dykes K. Optimization of turbine design in wind farms with [48] Gualtieri G. Atmospheric stability varying wind shear coefficients to improve wind
multiple hub heights, using exact analytic gradients and structural constraints. resource extrapolation: a temporal analysis. Renew Energy 2016;87:376–90.
Wind Energy 2019;22(5):605–19. [49] Sun H, Yang H, Gao X. Investigation into spacing restriction and layout optimiza-
[30] Wu YT, Liao TL, Chen CK, Lin CY, Chen PW. Power output efficiency in large wind tion of wind farm with multiple types of wind turbines. Energy 2019;168:637–50.
farms with different hub heights and configurations. Renew Energy [50] Grady SA, Hussaini MY, Abdullah MM. Placement of wind turbines using genetic
2019;132:941–9. algorithms. Renew Energy 2005;30:259–70.
[31] Mirghaed MR, Roshandel R. Site specific optimization of wind turbines energy cost: [51] Shamshirband S, Petković D, Ćojbašić Ž, Nikolić V, Anuar NB, Shuib NLM, et al.
iterative approach. Energy Convers Manage 2013;73:167–75. Adaptive neuro-fuzzy optimization of wind farm project net profit. Energy Convers
[32] Pookpunt S, Ongsakul W. Design of optimal wind farm configuration using a binary Manage 2014;80:229–37.
particle swarm optimization at Huasai district, Southern Thailand. Energy Convers [52] International Renewable Energy Agency (IRENA). Renewable Power Generation
Manage 2016;108:160–80. Costs in 2017. IRENA, ISBN 978-92-9260-040-2. Abu Dhabi, Jan. 2018. Available
[33] Gualtieri G. A novel method for wind farm layout optimization based on wind at: https://www.irena.org/publications/2018/Jan/Renewable-power-generation-
turbine selection. Energy Convers Manage 2019;193:106–23. costs-in-2017 [accessed 03/11/19].
[34] Jeon S, Kim B, Huh J. Comparison and verification of wake models in an onshore [53] Zhang J, Chowdhury S, Messac A, Castillo L. A response surface-based cost model
wind farm considering single wake condition of the 2 MW wind turbine. Energy for wind farm design. Energy Policy 2012;42:538–50.
2015;93:1769–77. [54] Fingersh LJ, Hand MM, Laxson AS. Wind turbine design cost and scaling model.
[35] VanLuvanee DR. Investigation of observed and modelled wake effects at Horns Rev Golden, CO: NREL Report No. TP-500-40566, Dec. 2006.
using WindPRO. Technical University of Denmark, Department of Mechanical [55] Kohonen T. Self-organizing maps. 2nd ed. Heidelberg, Germany: Springer; 2001.
Engineering; 2006. [56] Emami A, Noghreh P. New approach on optimization in placement of wind turbines
[36] Katic I, Hojstrup J, Jensen NO. A simple model for cluster efficiency. In: within wind farm by genetic algorithms. Renew Energy 2010;35(7):1559–64.
Proceedings of the European Wind Energy Association Conference and Exhibition, [57] Turner SDO, Romero DA, Zhang PY, Amon CH, Chan TCY. A new mathematical
1986, 407–10. programming approach to optimize wind farm layouts. Renew Energy
[37] Mortensen NG, Heathfield DN, Myllerup L, Landberg L, Rathman O. Wind atlas 2014;63:674–80.
analysis and application program: WAsP 8 help facility. Roskilde, Denmark: Risø [58] Chen Y, Li H, Jin K, Song Q. Wind farm layout optimization using genetic algorithm
National Laboratory; 2005. with different hub height wind turbines. Energy Convers Manage 2013;70:56–65.
[38] Nielsen P. The WindPRO Manual Edition 2.5. EMD International A/S, 2006. [59] Haces-Fernandez F, Li H, Ramirez D. Improving wind farm power output through
[39] Wang L, Tan AC, Cholette M, Gu Y. Comparison of the effectiveness of analytical deactivating selected wind turbines. Energy Convers Manage 2019;187:407–22.
wake models for wind farm with constant and variable hub heights. Energy Convers [60] Gualtieri G. Development and application of an integrated wind resource assess-
Manage 2016;124:189–202. ment tool for wind farm planning. Int J Renew Energy Res 2012;2(4):674–85.
[40] Chen J, Wang F, Stelson KA. A mathematical approach to minimizing the cost of [61] WindPower. The wind turbines and wind farms database, http://www.thewind-
energy for large utility wind turbines. Appl Energy 2018;228:1413–22. power.net; 2019 [accessed 03/11/19].
[41] Gualtieri G. Improving investigation of wind turbine optimal site matching through [62] Bowman AW, Azzalini A. Applied smoothing techniques for data analysis: the
the self-organizing maps. Energy Convers Manage 2017;143:295–311. kernel approach with S-plus illustrations. Oxford: Oxford University Press; 1997.
[42] Pallabazzer R. Parametric analysis of wind siting efficiency. J Wind Eng Ind [63] Bowman AW. ‘sm’ package: Smoothing Methods for Nonparametric Regression and
Aerodyn 2003;91(11):1329–52. Density Estimation. Version 2.2-5.6, 2018/09/27. Available at: https://cran.r-pro-
[43] Chang TJ, Wu YT, Hsu HY, Chu CR, Liao CM. Assessment of wind characteristics ject.org/web/packages/sm/index.html [accessed 03/11/19].
and wind turbine characteristics in Taiwan. Renew Energy 2003;28:851–71. [64] Luyssaert S, Janssens IA, Sulkava M, Papale D, Dolman AJ, Reichstein M, et al.
[44] Manwell JF, McGowan JG, Rogers AL. Wind energy explained: theory, design and Photosynthesis drives anomalies in net carbon-exchange of pine forests at different
application. 2nd ed. Chippenham, UK: John Wiley & Sons; 2010. latitudes. Glob Change Biol 2007;13:2110–27.
[45] Ali M, Matevosyan J, Milanović JV. Probabilistic assessment of wind farm annual [65] Kaski S, Lagus K. Comparing self-organizing maps. In: Proceedings of the
energy production. Electr Power Syst Res 2012;89:70–9. International Conference on Artificial Neural Networks (ICANN '96), Springer,
[46] Wind turbines models. https://en.wind-turbine-models.com; 2019 [accessed 03/ Berlin, Germany, 1996:809–14.
11/19]. [66] Chowdhury S, Zhang J, Messac A, Castillo L. Unrestricted wind farm layout opti-
[47] Gualtieri G. A comprehensive review on wind resource extrapolation models ap- mization (UWFLO): Investigating key factors influencing the maximum power
plied in wind energy. Renew Sustain Energy Rev 2019;102:215–33. generation. Renew Energy 2012;38(1):16–30.

13

You might also like