Article

Local Structure Optimization Design of Floating Offshore Wind

Turbine Platform Based on Response Surface Analysis
Yajun Ren 1 , Mingxuan Huang 2 , Jungang Hao 1 , Jiazhi Wang 2 , Shuai Li 1 , Ling Zhu 1 , Haisheng Zhao 2, *
and Wei Shi 2

1 China Renewable Energy Engineering Institute, Beijing 100120, China; [email protected] (Y.R.);
[email protected] (L.Z.)
2 State Key Laboratory of Coastal and Offshore Engineering, Dalian University of Technology,
Dalian 116024, China; [email protected] (M.H.); [email protected] (J.W.);
[email protected] (W.S.)
* Correspondence: [email protected]

Abstract: The floating platform is a critical component of the floating offshore wind turbine (FOWT),
and its internal structure design plays a key role in ensuring the safe operation of the FOWT. In this
study, the local model of the floating platform was firstly parameterized, and a response surface model
was obtained by conducting an orthogonal test. The response surface model was then optimized
using a gradient descent algorithm. Finally, the internal structure arrangement was validated through
a safety calibration. The optimization results indicate that the maximum stress of the optimized
model is reduced by 22.12% compared to the original model, while maintaining the same mass,
centroid, and other mass-related parameters. The optimization significantly improves the safety of
the structure and provides valuable references for the design and construction of a FOWT platform.

Keywords: floating offshore wind turbine; structural strength; response surface optimization;
gradient descent method

Citation: Ren, Y.; Huang, M.; Hao, J.;

Wang, J.; Li, S.; Zhu, L.; Zhao, H.; Shi,
1. Introduction
W. Local Structure Optimization
Design of Floating Offshore Wind Developing offshore wind power is a crucial strategy to support the national “double
Turbine Platform Based on Response carbon” goals and drive the transition to sustainable energy [1]. As a green, renewable,
Surface Analysis. Energies 2024, 17, and sustainable energy source, offshore wind power has garnered increasing attention
6316. https://doi.org/10.3390/ worldwide. Over the past decade, the global offshore wind power market has experienced
en17246316 an average annual growth rate of 27%, with a cumulative installed capacity reaching
75.2 GW by the end of 2023 [2]. According to statistics, about 80% of the world’s wind
Academic Editor: Francesco
Castellani
energy resources are located in sea areas deeper than 50 m, where offshore wind energy
offers several advantages over traditional onshore wind power, including higher wind
Received: 5 November 2024 energy density, minimal impact on terrestrial ecosystems, and no land use conflicts. As
Revised: 8 December 2024 a result, the development of floating offshore wind turbines (FOWTs) has emerged as
Accepted: 11 December 2024 an effective approach to develop far-reaching sea wind energy, playing an increasingly
Published: 14 December 2024
important role in ensuring global energy security [3]. This is of great strategic importance
for the development of deep offshore resources [4].
The performance of the floating body must meet strict requirements to ensure safe
Copyright: © 2024 by the authors.
operation in the complex and changeable deep-sea environment. In the research on the
Licensee MDPI, Basel, Switzerland. normal operation of offshore wind turbines (OWTs), an exploration of the structural design
This article is an open access article of the floating offshore wind turbine platform helps to improve the safety redundancy
distributed under the terms and of the floating body. As a result, many scholars have recently focused on improving the
conditions of the Creative Commons motion response and safety performance of offshore platforms. Tian et al. [5] investigated
Attribution (CC BY) license (https:// the correlation between the main scale of the structure and the objective function, where
creativecommons.org/licenses/by/ the structural weight and motion response are regarded as the objective function and the
4.0/). NSGA-II algorithm is used to improve the computational efficiency. Yang et al. [6] studied

Energies 2024, 17, 6316. https://doi.org/10.3390/en17246316 https://www.mdpi.com/journal/energies

Energies 2024, 17, 6316 2 of 24

the reliability-based optimization of a tripod-supported OWT under dynamic constraints.

The meta-model technology was used to replace the finite element calculation method to
improve the efficiency of the calculation. Structural mass was used as the objective function,
while uncertain factors such as material properties, geometry, and load types were included
as random variables. Their approach, combining global optimization with Monte Carlo
simulation, demonstrated effective optimization of both structural weight and dynamics.
Saeed et al. [7] proposed a new design scheme for a multi-functional floating platform
based on a density-based topology optimization method, which can significantly improve
the stiffness and durability of the platform and reduce the total mass of the new optimized
structure to 40.82% of the main structure. Chen et al. used a fully coupled aerodynamic–
hydrodynamic net mooring time-domain model constructed by combining AQWA and
FAST with the open-source code F2A for a preliminary analysis of a turreted deep-sea
aquaculture vessel powered by wind energy [8]. Lu et al. [9] used topology optimization
to develop a lightweight design for the tripod structure of OWTs. Their method notably
reduced weight while improving fatigue resistance compared to the original structure.
Karimi et al. [10] applied a multi-objective genetic algorithm (GA) optimization method
to evaluate different platforms to find the minimum cost and maximum performance
wind turbine form under specified environment conditions and sea state spectrum. Lee
et al. [11] developed a neural response surface method (NRSM) for objective optimization
problems, which is used to optimize the generated response surface combined with the
second-generation non-dominated sorting genetic algorithm (NSGA-II).
Ferri et al. [12] designed a multi-objective optimization program to determine the
optimal substructure configuration for a 10 MW FOWT at a specific site. Choi et al. [13]
used a design of experiments (DOE) approach to minimize the number of calculation
models required, employing a GA with neural network approximation to optimize the
platform and substructure of the 3 MW spar wind turbine system, resulting in a reliable
structural form. Additionally, Benifla et al. [14] developed an optimization framework for
the floating foundation of OWTs based on a genetic algorithm. Dou et al. [15] explored
a conceptual optimization design framework for the floating foundation with a mooring
system, capable of accurately determining designs that meet optimal conditions. Leimeis-
ter et al. [16] developed a method to optimize the floating concept by using the global
limit state, where the NSGA-II optimization algorithm is selected to find the optimal de-
sign solution. Hall et al. [17] built an optimization framework for the support structure
of a FOWT based on a genetic algorithm. Furthermore, Hegseth et al. [18] proposed a
linearized aerodynamic-hydrodynamic-servo-elastic FOWT model combined with gradient
optimization and analytical derivative methods.
As mentioned above, current studies on optimizing the main dimensions of floating
bodies primarily focus on motion performance and total structural mass. The internal
stiffeners, bulkheads, and ribs of the platform contribute over 60% of the floating body’s
total mass, where their size and arrangement play a critical role in ensuring structural
safety, while the corresponding local structure optimization design is seldom carried
out previously.
In this study, an A-type floating body was selected as the research subject, with the aim
of optimizing its structural configuration to reduce maximum structural stress under ex-
treme sea conditions, based on an initially optimized model. First, the relationship between
internal structure variables, such as structural mass, centroid position, and maximum
stress, are determined through an orthogonal test. Based on the above results, the optimal
configuration relationship of reinforcement and other structures of floating foundation
structure are determined through the optimization process, employing the gradient descent
algorithm. Additionally, the optimized structure must be subjected to a strength verification
test to ensure its reliability under extreme sea conditions. Finally, the study outlines a
comprehensive optimal design methodology for the floating platform of FOWTs, based on
the steps and findings described below.
Energies 2024, 17, 6316 3 of 24

2. Theoretical Methods
FOWTs are located in a complex and changing marine environment and are constantly
subjected to the combined effects of wind and wave loads. These external forces have a
significant impact on the stability and performance of the wind turbine. In this study, the
theories of wind and wave are introduced as follows.

2.1. Wind Load Calculation Theory

Due to the different heights of each part of the fan, the average wind speed will also
be different, generally showing a gradient change rule [19]. The formula is as follows:
 z 1/n
v(z) = v(h) (1)
h
where z is the altitude; h is the reference height above sea level; v(h) is the average wind
speed at the reference height h; v(z) is the average wind speed at altitude z; and n is the
wind profile index of sea surface roughness, whose value is related to the time distance
and offshore distance when measuring wind speed is generally between 7~13 m/s.
In the strength analysis, wind load is an important factor to be considered. During
the normal operation of the FOWT, the area swept by the wind turbine blades should be
included when calculating the wind pressure area [20]. However, under extreme conditions,
when the wind turbine is in a downwind shutdown state, the wind load acting on the tower
becomes the dominant factor. In such cases, the wind load on the wind turbine tower can
be expressed as Equation (2) [21]:

F = 0.5CS CH ρα AV 2 (2)

where CS is the shape coefficient of the component; CH is the height coefficient, which
can refer to the specification DNVGL-OS-C301 [20]; ρα is the air density; A is the frontal
projection area of the windward surface of the component; and V is the wind speed.

2.2. Wave Field Calculation Theory

In this study, the wave load is calculated through the three-dimensional potential
flow theory [22], which accounts for drag, inertial, and diffraction forces. This theory is
particularly suitable for calculating the interaction of large-scale structures and waves when
the structure is stationary. The velocity potential function of the wave field is shown in
Equation (3) [23]:

(1) (1) (1)

Φ(1) ( x, y, z) = Φ I ( x, y, z) + ΦD ( x, y, z) + ΦR ( x, y, z) (3)

(1)
where Φ(1) denotes the first-order velocity potential; ΦD is the first-order diffraction
(1)
potential; and ΦR denotes the first-order radiation potential.
Φ I is calculated as below:

igA cosh(Kz + d) −K ( xcosβ+ysinβ)

ΦI = · ·e (4)
ω cosh(Kd)

where g is the acceleration of gravity; A is the amplitude; K is the wave number; d is the
water depth; and β is the wave direction angle.
The boundary conditions for solving the velocity potential of diffracted and radiation
waves are as follows:
(1) Diffracted wave
Free surface condition:

∂Φ D
−ω 2 Φ D + g =0z=0 (5)
∂z
Energies 2024, 17, 6316 4 of 24

Object surface condition:

∂Φ D ∂Φ I
= − (6)
∂n S ∂n S
Seabed boundary condition:

∂Φ D
=0 (7)
∂z z=−h

Infinite condition:
∂ΦjD
!
√
lim R − ikΦjD =0 (8)
R→∞ ∂R

(2) Radiation wave

Free surface conditions:

∂2 ΦjR ω2 R
=− Φ z=0 (9)
∂n g j

Object surface conditions:

∂ΦjR
= nj (10)
∂n
S
Seabed boundary conditions:

∂ΦjR
=0 (11)
∂n
z=−h

Infinite condition:
√  ∂Φ 
lim R − ikΦ = 0 (12)
R→∞ ∂R

3. Development of Initial Model

In this study, a response surface optimization design method, incorporating the re-
sponse surface model and gradient descent algorithm, is employed to optimize an A-type
floating body. First, a parametric model was developed based on the initial model with
predefined boundary conditions, and the finite element analysis was conducted. Moreover,
the PB (Plackett–Burman) and CCD (Central Composite Design) tests were performed to
analyze the parameter sensitivity of the model variables and to establish the response sur-
face mode. The response surface equation was then optimized using the gradient descent
method. Finally, the optimized model was verified under extreme conditions. The specific
operation process is illustrated in Figure 1. The follow-up sections in this chapter describe
the development of the initial model and stress strength analysis.
 Energies2024,
Energies 2024,17,
17,6316
x FOR PEER REVIEW 55 of
of 24
25

Initial model
modeling
Finite element
solution
Set boundary
conditions

Response surface
Finite element Model verification
model
solution

Model
parameterization Meet the accuracy Meet the
requirements ? requirements ?

Parameter sensitivity
analysis Y Y
N
Gradient descent Get the optimal
CCD experimental method solution
design N

Figure1.1. Optimized
Figure Optimizeddesign
designprocess.
process.

3.1. Structural Parameters of Initial Model

3.1. Structural Parameters of Initial Model
The FOWT platform primarily consists of buoy columns, heave plates, pontoons, and
The FOWT platform primarily consists of buoy columns, heave plates, pontoons, and
diagonal and transverse braces, arranged in a triangular configuration. The wind turbine
diagonal and transverse braces, arranged in a triangular configuration. The wind turbine
tower is set at the top of one of the three columns, while the ballast tank and empty cabin
tower is set at the top of one of the three columns, while the ballast tank and empty cabin
are set inside the buoy column, heave plate, and pontoon to maintain the equilibrium
are set inside the buoy column, heave plate, and pontoon to maintain the equilibrium be-
between the platform’s overall gravity and buoyancy of the platform. The main parameters
tween the platform’s overall gravity and buoyancy of the platform. The main parameters
of the model and the schematic diagram of the model are presented in Table 1 and Figure 2,
of the model and the schematic diagram of the model are presented in Table 1 and Figure
respectively.
2, respectively.
Table 1. Main parameters of FOWT.
Table 1. Main parameters of FOWT.
Parameter of Model Value
Parameter of Model Value
Water Water depth
depth 100 100
m m
Center of gravity (0 m, 0 m, −7.2 m)
Center of gravity (0 m, 0 m, −7.2 m)
Platform mass 6035 t
Platform
Draft ofmass
platform 603525t m
Draft of platform
Ballast water density 25 m kg/m3
1025
Ballast water density 1025 kg/m3
The numerical simulation of the floating platform structure was performed in SESAM
2021 [24], a specialized software developed by Det Norske Veritas (DNV) (Oslo, Norway)
for the design, analysis, and evaluation of marine engineering and offshore structures.
Specifically, the structural model was established using GeniE 2021 and the strength analysis
of local structure was carried out in Sestra 2021, which are both modules incorporated in
the SESAM software 2021. Additionally, the pontoon is divided into six sections, numbered
1~6 from bottom to top. The thickness distribution of the bulkhead and the arrangement of
the internal stiffener are shown in Figures 3 and 4, respectively.
nergies 2024, 17, x FOR PEER REVIEW 6 of 25

Energies 2024, 17, 6316 6 of 24

Figure 2. Appearance of newly designed FOWT.

The numerical simulation of the floating platform structure was performed in

SESAM 2021 [24], a specialized software developed by Det Norske Veritas (DNV) (Oslo,
Norway) for the design, analysis, and evaluation of marine engineering and offshore
structures. Specifically, the structural model was established using GeniE 2021 and the
strength analysis of local structure was carried out in Sestra 2021, which are both modules
incorporated in the SESAM software 2021. Additionally, the pontoon is divided into six
sections, numbered 1~6 from bottom to top. The thickness distribution of the bulkhead
and the arrangement of the internal stiffener are shown in Figure 3 and Figure 4, respec-
Figure 2. Appearance
tively. of newly designed
Figure 2. Appearance of newlyFOWT.
designed FOWT.

The numerical simulation of the floating platform structure was performed in

SESAM 2021 [24], a specialized software developed by Det Norske Veritas (DNV) (Oslo,
Norway) for the design, analysis, and evaluation of marine engineering and offshore
structures. Specifically, the structural model was established using GeniE 2021 and the
strength analysis of local structure was carried out in Sestra 2021, which are both modules
incorporated in the SESAM software 2021. Additionally, the pontoon is divided into six
sections, numbered 1~6 from bottom to top. The thickness distribution of the bulkhead
and the arrangement of the internal stiffener are shown in Figure 3 and Figure 4, respec-
tively.

Energies 2024, 17, x FOR PEER REVIEW 7 of 25

Figure 3. Bulkhead thickness distribution map.

Figure 3. Bulkhead thickness distribution map.

Figure 3. Bulkhead thickness distribution map.

Figure
Figure 4. Layout
4. Layout diagram
diagram of stiffening
of stiffening rib. rib.

3.2. Marine Environmental Conditions

For the strength study of the platform structure, the hydrostatic pressure, structural
weight, wind load, and wave load are considered, which are applied to the floating plat-
form by linear superposition through multiplication with their respective sub-coefficients
Energies 2024, 17, 6316 7 of 24

3.2. Marine Environmental Conditions

For the strength study of the platform structure, the hydrostatic pressure, structural
weight, wind load, and wave load are considered, which are applied to the floating platform
by linear superposition through multiplication with their respective sub-coefficients [23],
and the load sub-coefficients are shown in Table 2.

Table 2. Partial safety factors for loads.

Hydrostatic Structural
Design Load Wave Load Wind Load
Pressure Weight
Coefficients 1.3 1.3 1.0 1.0

(1) Wind load distributed on the wind turbine tower

The wind load can be calculated by the horizontal wind force from Equation (2), which
takes into account wind speed, structural body type coefficient, air density, windward area,
and other factors to accurately determine the magnitude of the load exerted by the wind
on the structure. Herein, the limit state of wind load is considered with a wind speed of
51.5 m/s. The distributions of wind loads acting along the height of the tower, as well as
the total load on the blades, is provided in Table 3.

Table 3. Wind load distributed along the height of wind turbine tower.

Loading Positions (m) Wind Thrust Force (N)

29.75 45,529.12
40.25 62,669.36
51.75 69,545.74
63.25 74,016.26
74.75 73,998.12
86.25 76,616.77
97.75 78,147.65
109.57 82,906.32
119 m (Height of rotor center) 1,450,000

(2) Wave conditions under the most unfavorable conditions

Long-term forecast design wave method is a commonly used method in the field of
marine and shipbuilding engineering. It is mainly based on spectral analysis to forecast
ship profile wave load, statistically characterizing its frequency response function and
long-term distribution. It can determine the design wave parameters, such as wave height,
period, direction, etc., providing a basis for the design of the floating platform structure
and the strength assessment in order to more realistically reflect the actual wave loading
condition [25].
The calculation of wave loads on marine structures can generally be judged according
to the ratio of the characteristic size D and wavelength λ. For large-size components
(D/λ > 0.15), the wave loads were mainly calculated using the potential flow theory.
Based on the wave direction, wave height, and period tabulated in Table 4, the velocity
potential was determined according to Laplace’s equation and the boundary conditions
of free surface, object surface, and the seabed. Then, Bernoulli’s equation was employed
to calculate the fluid pressure, and finally, the wave load was obtained by integrating the
pressure on the surface of the structure [26]. In addition, the hydrostatic pressure and
structural weight were easily obtained according to the principles of hydrostatics and
Newton’s second law, respectively.
 Energies 2024, 17, 6316 8 of 24

Table 4. Design wave for strength calculation.

Condition Number Direction (◦ ) Amplitude (m) Period (s)

LC1 0 8.17 6.8
LC2 0 6.46 9.6
LC3 0 6.96 10
LC4 0 8.17 19
LC5 75 8.05 7.2
LC6 90 7.87 25
LC7 90 8.05 27
LC8 105 8.05 7.2
LC9 180 6.46 9.6
LC10 180 6.96 10
LC11 180 8.17 19
LC12 270 7.87 25
LC13 270 8.05 27

3.3. Establishment of Finite Element Model

(1) Setting boundary conditions
In order to make up for the limitation that mooring system cannot be set directly on
the platform in GeniE software 2021, equivalent boundary conditions were adopted to
simulate the mooring effect [25], as shown in Figure 5. Regarding the translational degrees
Energies 2024, 17, x FOR PEER REVIEW of freedom at the center point of heave columns, the first point (SP1) is constrained in 25
9 of the X-,
Y- and Z-axis directions; the second point (SP2) is restricted along Y- and Z-axis directions;
and the third point (SP3) is limited only in the Z-axis direction [27].

Figure
Figure 5. Model
5. Model boundary
boundary condition.
condition.

(2) (2)
FiniteFinite element
element model model
Based Based
on theon structural
the structural parameters
parameters of theof floating
the floating platform,
platform, the initial
the initial modelmodel
was was
established
established in GeniE
in GeniE software,
software, consisting
consisting of components
of components suchsuch as buoy,
as buoy, pontoon,
pontoon, heaveheave
plate, diagonal brace, and transverse brace. To accurately represent the
plate, diagonal brace, and transverse brace. To accurately represent the real properties of real properties
of the structure,
the structure, thepontoon,
the buoy, buoy, pontoon, heavebulkhead,
heave plate, plate, bulkhead,
and riband ribwere
plate platemodeled
were modeled
as
as plate elements, while the diagonal brace, transverse brace, and internal
plate elements, while the diagonal brace, transverse brace, and internal skeleton were skeleton
con-were
constructed
structed using beamusingelements.
beam elements. Additionally,
Additionally, the interior
the interior was separated
was separated by ribs,
by ribs, and the
and the
overall
overall strength
strength of model
of the the model
was was ensured
ensured by theby steel
the steel structure
structure framework.
framework. ThisThis
finitefinite
element
element modelmodel
was was meshed
meshed withwith a size
a size of 2asm,shown
of 2 m, as shown in Figure
in Figure 6. 6.
 plate, diagonal brace, and transverse brace. To accurately represent the real properties of
the structure, the buoy, pontoon, heave plate, bulkhead, and rib plate were modeled as
plate elements, while the diagonal brace, transverse brace, and internal skeleton were con-
structed using beam elements. Additionally, the interior was separated by ribs, and the
Energies 2024, 17, 6316 9 of 24
overall strength of the model was ensured by the steel structure framework. This finite
element model was meshed with a size of 2 m, as shown in Figure 6.

Figure 6.
Figure Finiteelement
6. Finite elementmodel
modelof
offloating
floating offshore
offshore wind
wind turbine
turbine platform.
platform.

3.4. Stress Calculation of Initial Structure Model

3.4. Stress Calculation of Initial Structure Model
According to DNV GL-OS-C103 [27], the structural design should satisfy the strength
According
requirement to DNV
below GL-OS-C103
the material yield[27], the
stress structural
with a properdesign
safetyshould
factor.satisfy
In thisthe strength
study, Q355
requirement below the material yield stress with a proper safety factor. In this
steel, which exhibits a yield strength of 355 MPa, was employed. After taking into account study, Q355
steel, which
a safety exhibits
factor a yield
of 1.25, the strength
structuralof stresses
355 MPa,were
was required
employed. toAfter taking
remain into284
below account
MPa.
aThe
safety
results of the structural strength assessment under 13 conditions are listed, andThe
factor of 1.25, the structural stresses were required to remain below 284 MPa. all
results of the the
cases satisfy structural
requiredstrength assessment
specifications. under
Based 13 conditions
on this are listed,
validated initial andstructural
model, all cases
satisfy the required
optimization specifications.
was subsequently Based
carried on this validated initial model, structural opti-
out.
mization was subsequently carried out.
4. Response Surface Optimization
The response surface method employed in this study is a widely used optimization
technique that constructs a mathematical model to describe the relationship between
parameters and response variables. This is achieved by fitting regression equations based
on a set of sample points generated through orthogonal experimental design, and is
applied to predict the optimal solution of the model under the set constraints [28]. The
orthogonal design serves to reduce the number of tests required and improve optimization
efficiency [29,30]. In this process, an orthogonal design table is used to disperse the variable
values, ensuring a balanced exploration of the design space. Analysis of variance, regression
analysis, model fitting, and optimization analysis were performed using Design Expert
2021, an experimental design software developed by Stat-Ease Inc. (Minneapolis, MN,
USA) [31]. Design Expert mainly analyzes the data derived from SESAM 2021 where the
results are imported into Design Expert, and the analysis of variance function of Design
Expert is used to evaluate the effects of different parameters on the structural performance.
The mathematical relationship between the structural response and the parameters is
explored with the help of regression analysis in order to construct an accurate regression
model, and optimization analyses are carried out in order to find the optimal solution for
the structural design.

4.1. Mathematical Model for Response Surface Optimization

4.1.1. Design Variables
The floating platform is mainly divided into three parts, viz., pontoon, buoy, and
heave plate. In the optimization process, the bulkhead thickness, stiffener spacing, and
stiffener size of the above parts are parameterized, including a total of 10 variables. The
correspondence between variables and structural dimensions, as well as the designed
variable value, are shown in Table 5. Therefore, the new floating body structure can be
constructed by combining the changed variable. That is, the design variable is:

X = ( A, B, C, D, E, F, G, H, I, J ) T (13)
Energies 2024, 17, 6316 10 of 24

Table 5. Parametric modeling variable design table.

Variable Number Variable Name Variable Value

Bulkhead thickness of buoy for 1~6 sections 0.018 m, 0.016 m, 0.016 m, 0.016 m, 0.016 m, 0.016
A
(Figure 3) m
B Bulkhead thickness of heave plate 0.019 m
C Bulkhead thickness of pontoon 0.010 m
D Spacing between pontoon longitudinal stiffener 12.8◦
E Number of transverse stiffeners in buoy 22
height = 1 m, web thickness = 0.015 m,
F Size of float stiffener
width = 0.3 m, top flange thickness = 0.015 m
G Spacing between heave plate stiffener 7◦
height = 0.7 m, web thickness = 0.01 m,
H Size of heave plate stiffener size
width = 0.2 m, top flange thickness = 0.01 m
I Number of transverse stiffeners in pontoon 54
height = 0.035 m, web thickness = 0.01 m,
J Size of pontoon transverse stiffener
width = 0.2 m, top flange thickness = 0.01 m

4.1.2. Objective Function

The objective of this optimization design is to optimize the local structure of the floating
platform by searching for the minimum solution of structural stress, i.e., the minimum
value of f (X), where f (X) is obtained by the experimental design software, Design Expert
software, based on the results of orthogonal tests. In summary, the optimized mathematical
model is:
X = A, B, C, D, E, F, G, H, I, J ) T
Min f ( X )
(14)
s.t.( Mass) L ≤ Mass ≤ ( Mass)U
s.t.(COMz) L ≤ COMz ≤ (COMz)U
where ( Mass) L and ( Mass)U are the lower and upper limit of platform mass; and (COMz) L
and (COMz)U are the lower and upper limit of platform center of mass, respectively.
The optimization of floating platform involves two key steps. The first step is con-
cerned with the hydrodynamic optimization of the overall structure. This optimization
requires the precise determination of key parameters such as mass, center of gravity, mo-
ment of inertia, and main scale. The goal of this optimization is to design a floating platform
with excellent hydrodynamic characteristics and lower steel consumption, the detailed
description of which referring to the reference [32] published by our group recently. These
optimized parameters will provide a solid foundation for subsequent work.
The second step, viz., the current research work, involves optimizing the local struc-
tural strength. In the local structural optimization process, the thickness of internal rein-
forcement, the size of steel bars, and the layout of compartments are used as optimization
variables. Due to practical constraints, it is difficult for the locally optimized model to
correspond precisely to the parameters of the initial model. Herein, the constraints for
the second step of local structural optimization are set to keep the variation of structural
parameters (mass, center of mass) within ±5%, based on an in-depth study of the hydro-
dynamic performance of the floating body. Within this tolerance (±5%), the impact on
the hydrodynamic performance of the floating body is negligible, and at the same time,
sufficient space is provided for the optimization of the local structural dimensions [32].

4.2. Plackett–Burman Test

The significant factor is determined by executing the Plackett–Burman test, viz., com-
paring the difference between the two levels of each factor and the overall difference and
conducting the parameter sensitivity analysis of the variables with the least number of
tests [33]. The orthogonal test with n = 12 (Number of tests) is designed by investigating the
10 variables and defining the maximum stress value as the response value. The factor level
and coded value of Plackett–Burman design are presented in Table 6, where variables A, B,
Energies 2024, 17, 6316 11 of 24

C, F, H, and J are all expressed in the form of variation coefficient to reflect the variation
visually, viz., the ratio of changed size to initial size.

Table 6. Plackett–Burman design factor level and coding value.

Variable Number
Test Number
A B C D E F G H I J
Lower limit 0.9 0.9 0.9 19 5 0.9 6 0.9 25 0.9
Upper limit 1.1 1.1 1.1 23 9 1.1 8 1.1 29 1.1

After orthogonal tests, the distribution of the maximum and minimum stress results is
Energies 2024, 17, x FOR PEER REVIEW in Figure 7. It is obvious that LC4 and LC11 are the most unfavorable among12
shown theof13
25

working conditions, thus the subsequent analysis should focus on these two conditions.

Figure7.7. Plackett–Burman
Figure Plackett–Burman test
test results
results distribution.
distribution.

The
The F-value is is aacrucial
crucialstatistical
statisticalmetric
metric in in regression
regression analysis,
analysis, which
which is usually
is usually em-
employed to evaluate
ployed to evaluate whether
whether the model
the model as a as a whole
whole is statistically
is statistically significant.
significant. Specifi-
Specifically, if
cally, if the computed
the computed F-valueF-value
is large is andlarge
theand the corresponding
corresponding p-value,p-value,
derived derived
from thefrom the
F-distri-
F-distribution,
bution, is less is lessthe
than than the pre-established
pre-established significance
significance levellevel
of α of α = 0.05,
= 0.05, it indicates
it indicates thatthat
the
the model is significant. In this context, by comparing the magnitudes
model is significant. In this context, by comparing the magnitudes of the F-values of dif- of the F-values of
different variables, the relative importance or significance of these
ferent variables, the relative importance or significance of these variables can be ranked. variables can be ranked.
The
The higher
higher thethe F-value
F-value of of aa variable,
variable, the the more
more significant
significant itsits contribution
contribution to to the
the model.
model.
The R 2 and R2 are important indicators of closeness of the linear relationship between
The 𝑅 and 𝑅 adj are important indicators of closeness of the linear relationship between
multiple
multipleindependent
independentvariables
variablesand andaa dependent
dependentvariable,
variable, where
wherean an absolute
absolute valuevalue closer
closer
to
to 1 indicates a stronger linear relationship between the independent variables and
1 indicates a stronger linear relationship between the independent variables and the
the
2
dependent
dependentone. one.In Inregression analysis,R𝑅 rises
regressionanalysis, riseswith
withthethe
increasing
increasing number
number of independent
of independ-
variables, eveneven
ent variables, if theif new
the newindependent
independent variables
variableshavehave
no substantial
no substantialexplanatory
explanatory signifi-
sig-
cance to the dependent variable, which may induce misjudgment. However, the R 2 can
nificance to the dependent variable, which may induce misjudgment. However, the adj 𝑅
make up for
can make upthis
for shortcoming
this shortcoming by taking into account
by taking the number
into account the number of independent
of independent variables
var-
and the sample size, making the assessment of goodness of
iables and the sample size, making the assessment of goodness of fit more objective fit more objective and precise,
and
which
precise, provides a reliable abasis
which provides for choosing
reliable basis for the reliablethe
choosing model andmodel
reliable judging andthejudging
validitytheof
independent variables. Below are the formulas for R2 and R2adj [34]:
validity of independent variables. Below are the formulas for 𝑅 and 𝑅 [34]:
) = β=0 +
f ( X𝑓(𝑋) 𝛽 β+1 X
𝛽1𝑋+ +
β 2𝛽
X2𝑋++
. .⋯
. ++β𝛽k X𝑋k ++ϵ𝜖 (15)
(15)

𝑆𝑆𝑇 = 𝑓(𝑋 ) − 𝑓(𝑋) (16)

𝑆𝑆𝑅 = 𝑓 (𝑋 ) − 𝑓(𝑋) (17)

Energies 2024, 17, 6316 12 of 24

n  2
SST = ∑ f (X i ) − f (X) (16)
i =1
n   2
SSR = ∑ fˆ( X i − f ( X ) (17)
i =1

R2 = SSR/SST (18)
n−1
R2adj = 1 − (1 − R2 ) (19)
n−j−1
where X1 , X2 , . . . , Xk are the independent variables, β 0 , β 1 , . . . , β k are the regression
coefficients, ε is the random error, SST is the sum of squared total deviations, SSR is the
computed sum of squared regressions where n is the number of samples, f ( Xi ) is the i−th
observation, f ( X ) is the sample mean of f ( X ), fˆ( Xi ) is the ith observation predicted by the
regression model, and j is the number of independent variables.
From Table 7, the order of magnitude of the F-values shows that the variable significance
of the condition 4 model meets the requirements (B > J > H > C > E > D > A > F > G > I). The
model correlation coefficient R2 is 0.9155 determining an excellent model correlation. The
correction coefficient R2adj is 0.8672, viz., the model can explain more than 86.72% of the
response value changes. Table 8 shows that the p-value of the model under LC11 is 0.0682
more than 0.05, illustrating the model does not fit well in the regression area in this study,
which is led by too many unrelated factors the model contains. Therefore, the variance
analysis of the model was performed again after removing the irrelevant factors, as shown
in Table 9. According to the F-value, the degree of influence is sorted: J > B > H > C > E > D
> F > A > I > G. In this case, it is found that the model correlation coefficient R2 is 0.9150,
indicating a high model correlation, and the correction coefficient R2adj is 0.8665, meaning
the model can explain more than 86.65% of the response value changes.
The results of the Plackett–Burman test are identified by the Lenth method [35]. The
semi-normal probability of the standardized effects for LC4 and LC11 are obtained as
shown in Figure 8. It is found that the standardized effect points of factors B, C, H, and J
are far from the fitting point, which means the thickness of the heave plate and pontoon
bulkhead as well as the size of the heave plate and pontoon stiffener are the four factors
affecting the maximum stress. According to the Pareto diagram of the standardized effects
of factors as shown in Figure 9, it can be seen that the t-value of the statistical test for factors
B, C, H, and J exceeds the limit for determining the significance of the factors; therefore,
these factors can be further proved to be significant factors.

Table 7. Significance analysis of condition 4 results.

Source Sum of Squares Degree of Freedom Mean Square F p Significant

Model 6.905 × 1015 10 6.905 × 1014 289.45 0.0457 Yes
A 4.481 × 1013 1 4.481 × 1013 18.78 0.1444 No
B 2.084 × 1015 1 2.084 × 1015 873.46 0.0215 Yes
C 9.426 × 1014 1 9.426 × 1014 395.10 0.0320 Yes
D 1.684 × 1014 1 1.684 × 1014 70.58 0.0754 No
E 3.057 × 1014 1 3.057 × 1014 128.13 0.0561 No
F 4.470 × 1013 1 4.470 × 1013 18.74 0.1445 Yes
G 1.140 × 1013 1 1.140 × 1013 4.78 0.2731 No
H 1.323 × 1015 1 1.323 × 1015 554.57 0.0270 Yes
I 6.278 × 1012 1 6.278 × 1012 2.63 0.3517 No
J 1.975 × 1015 1 1.975 × 1015 827.71 0.0221 Yes
Residual 2.386 × 1012 1 2.386 × 1012 — — —
Cor Total 6.908 × 1015 11 — — — —
Energies 2024, 17, 6316 13 of 24
Energies 2024, 17, x FOR PEER REVIEW 14 of 25

Table 8. Significance analysis of condition 11 results.

Table 9. Significance analysis of condition 11 results (after correction).
Source Sum of Squares Degree of Freedom Mean Square F p Significant
Source Sum of Squares Degree of Freedom Mean14Square F p Significant
Model 6.994 × 1015 10 6.994 × 10 129.89 0.0682 No
Model
A 6.979
8.128 × 10×1310
15
1 9 7.754
8.128 × 1013 × 10
14 73.52
15.09 0.0135
0.1604 Yes
No
BA 8.128
1.988 × 1013
× 10 15 1 1 8.128
1.988 × 10 × 1013
15 7.71
369.12 0.1090
0.0331 NoYes
CB 1.988
1.037 ×151015
× 10 1 1 1.037 × 1015 × 1015
1.988 192.50
188.47 0.0458
0.0053 Yes
Yes
D 1.484 × 10 14 1 1.484 × 1014 × 1015 27.56 0.1198 No
C 1.037 × 1015 1 1.037 98.29 0.0100 Yes
E 2.251 × 1014 14 1 2.251 × 1014 41.79 0.0977 No
D 1.484 × 10 1 1.484 × 1014 14.07 0.0643 No
F 8.338 × 1013 1 8.338 × 1013 15.48 0.1584 No
GE 2.251 ×131014 1 2.251
1013 × 10 21.34 0.0438 Yes
14
1.571 × 10 1 1.571 × 2.92 0.3372 No
HF 8.338 ×151013 1 8.338
1015 × 10 7.91 0.1066 No
13
1.355 × 10 1 1.355 × 251.61 0.0401 Yes
H
I 1.355
3.555 ×131015
× 10 1 1 3.555 × 13
1.355
10 × 1015 128.46
6.60 0.0077
0.2363 Yes
No
JI 2.026 × 10 15 1 2.026 × 1015 × 1013 376.25 0.0328 Yes
3.555 × 1013 1 3.555 3.37 0.2078 No
Residual 5.385 × 10 12 1 5.385 × 1012 × 1015 — — —
J 2.026 × 1015 1 2.026 192.10 0.0052 Yes
Cor Total 7.000 × 1015 13 11 — — — —
Residual 2.109 × 10 2 1.055 × 1013 — — —
Cor Total 7.000 × 1015 11 — — — —
Table 9. Significance analysis of condition 11 results (after correction).
The results of the Plackett–Burman test are identified by the Lenth method [35]. The
Source Sum of Squares Degree of Freedom Mean Square F p Significant
semi-normal probability of the standardized effects for LC4 and LC11 are obtained as
15
6.979 × shown 14
Model 10 in Figure 8.9It is found that 7.754the×standardized
10 73.52 points of0.0135
effect factors B, C, Yes
H, and J
A 8.128 × 1013 1 8.128 × 1013 7.71 0.1090 No
are far from the fitting point, which means the thickness of the heave plate and pontoon
B 1.988 × 1015 1 1.988 × 1015 188.47 0.0053 Yes
bulkhead as well as1 the size of the heave 15 plate and pontoon stiffener are the fourYesfactors
C 1.037 × 1015 1.037 × 10 98.29 0.0100
D affecting
1.484 × 10 14 the maximum1 stress. According
1.484 × 10 to
14 the Pareto diagram
14.07 of the standardized
0.0643 Noeffects
E of
2.251 × 10 factors
14 as shown 1 in Figure 9, it can
2.251 × 10be 14seen that the
21.34 t-value of the
0.0438statistical test for
Yes
F 10 13
8.338 × factors B, C, H, and 1 J exceeds the 8.338 limit 13
× 10for determining 7.91 the significance
0.1066 of theNo factors;
H 1015
1.355 × therefore, these factors
1 1.355 ×
can be further 1015 to be significant
proved 128.46 factors.
0.0077 Yes
I 3.555 × 1013 Overall, too many 1 variables3.555
in the
13
× 10response 3.37 analysis0.2078
surface might lead toNo the re-
J 1015
2.026 × sponse surface 1
model being 2.026
inaccurate × 10 15
[36]. 192.10
Therefore, only the 0.0052
top four Yes with
variables
Residual 2.109 × 1013 2 × 1013 — — —
greater
15 influence were retained, 1.055including the thickness of the heave plate and pontoon
Cor Total 7.000 × 10 11 — — — —
bulkhead as well as the size of the heave plate and pontoon stiffener.

Figure 8. Half-normal plot of stress results under condition 4 (left) and condition 11 (right).
 Energies 2024,
Energies 2024, 17,
17, 6316
x FOR PEER REVIEW 1514of
of 25
24

Figure
Figure 9.
9. Normalized
Normalized effect
effect Pareto
Pareto diagram
diagram of
of stress
stress results
results under
under condition
condition 44 (left)
(left) and
and condition
condition
11 (right).
11 (right).

4.3. Response
Overall, Surface
too manyAnalysis (Central
variables in theComposite
response Design)
surface analysis might lead to the response
surface model
Based being
on the inaccurate
conclusion [36].previous
of the Therefore, only the four
subsection, top four variables
variables werewith greater
selected for
influence were retained, including the thickness of the heave plate and
the Central Composite Design test [37,38], including the thickness of the heave plate bulk-pontoon bulkhead
as wellthe
head, as thickness
the size ofofthe
theheave
pontoonplatebulkhead,
and pontoon stiffener.
the size of the heave plate stiffener, and the
size of the pontoon stiffener, which were relabeled as A, B, C, and D, respectively. By set-
4.3. Response Surface Analysis (Central Composite Design)
ting the parameter range and interval to 0.8~1.2 and 0.025, respectively, a total of 332 com-
Basedofonorthogonal
binations the conclusiontestofcombinations
the previous subsection,
were obtained.four variables
The stress were selected for
maximum the
value,
Central Composite Design test [37,38], including the thickness of the heave
mass, and centroid position were taken as the response values in the test, and the ultimate plate bulkhead,
the thickness
goal of the pontoon
was to obtain bulkhead,
the smallest maximum the size of the
stress heave
value plateeach
under stiffener, and condition.
working the size of
the pontoon stiffener, which were relabeled as A, B, C, and D, respectively.
Meanwhile, the parameter ranges of mass and centroid were kept within ±5% of the initial By setting the
parameter range and interval to 0.8~1.2 and 0.025, respectively, a
model parameters as shown in Figure 10. Table 10 shows that the experimental designtotal of 332 combinations
of orthogonal
factor level andtest combinations
coded value. were obtained. The stress maximum value, mass, and
centroid position were taken as the response values in the test, and the ultimate goal was
Energies 2024, 17, x FOR PEER REVIEW 16 of 25
to obtain
Table the smallest
10. Central maximum
Composite stress levels
Design factor valueand
under
codedeach working condition. Meanwhile,
values.
the parameter ranges of mass and centroid were kept within ±5% of the initial model
parameters as Number
shown in Figure 10. Table 10 shows that Variable Number design factor level
the experimental
Test
and coded value. A B C D
Lower limit 0.800 0.800 0.800 0.800
Upper limit 1.200 1.200 1.200 1.200

Figure 10. Central Composite Design test results distribution diagram.

Figure 10. Central Composite Design test results distribution diagram.

The results obtained by significance analysis of the model are shown in Table 11. It is
observed that the four response value models of the von Mises stress under LC4 and LC
11, as well as the mass and the center of mass, are extremely significant (p < 0.05). As
shown in Table 12, the model correlation coefficient 𝑅 and the correction coefficient 𝑅
Energies 2024, 17, 6316 15 of 24

Table 10. Central Composite Design factor levels and coded values.

Variable Number
Test Number
A B C D
Lower limit 0.800 0.800 0.800 0.800
Upper limit 1.200 1.200 1.200 1.200

The results obtained by significance analysis of the model are shown in Table 11. It is
observed that the four response value models of the von Mises stress under LC4 and LC 11,
as well as the mass and the center of mass, are extremely significant (p < 0.05). As shown
in Table 12, the model correlation coefficient R2 and the correction coefficient R2adj are all
above 0.98 with an excellent model fitting degree, where R2adj means all models can explain
more than 98% of the response value changes. Therefore, each regression model can be
used to analyze and predict the above-mentioned four response value models (A, B, C, D).

Table 11. Results obtained by model significance analysis.

Model Degree of Freedom Sum of Squares Mean Square F p Significant

von Mises stress (LC 4) 3.329 × 1017 14 2.378 × 1016 1541.21 <0.001 Yes
von Mises stress (LC 11) 3.318 × 1017 14 2.370 × 1016 1331.08 <0.001 Yes
Mass 6.353 × 1012 14 4.538 × 1011 4.04 × 1010 <0.001 Yes
COMz 42.98 14 3.07 1.03 × 107 <0.001 Yes

Table 12. Model correlation coefficient and correction coefficient.

Model R2 R2adj
von Mises stress (LC 4) 0.9855 0.9894
von Mises stress (LC 11) 0.9833 0.9825
Mass 0.9999 0.9999
COMz 0.9999 0.9999

Based on the obtained regression model, the error distribution of the value predicted by
the model and the correlation between the predicted and experimental value are shown in
Figure 11. To better evaluate the model, an index, the prediction errors not exceeding ±2 SD
(standard deviation), is defined as a criterion for excellent prediction model. According to
the error distribution maps of the four models in Figure 11a,c,e,g, the sample prediction
error values are within ±2 SDs (standard deviations), while no anomalous sample points
indicate the high prediction accuracy of the model. From the correlation diagram of the
model predicted value and the experimental value in Figure 11b,d,f,h, it is obvious that the
sample points are all near the 45◦ diagonal with a small difference, which indicates a high
goodness fit between predicted and experimental values.
The change range of the limiting mass and centroid is selected to be ±4% for predicting
the minimum stress in the feasible region. Meanwhile, the A variable (heave plate bulkhead
thickness) and C variable (heave plate stiffener size) are adjusted to 0.8~1.4 and 0.7~1.2,
respectively, to make the variable prediction point fall within the variable range, as shown
in Table 13.
 ple points indicate the high prediction accuracy of the model. From the correlation dia-
gram of the model predicted value and the experimental value in Figure 11b,d,f,h, it is
Energies 2024, 17, 6316 16 of 24
obvious that the sample points are all near the 45° diagonal with a small difference, which
indicates a high goodness fit between predicted and experimental values.

(a) Condition 4 (b) Condition 4

(c) Condition 11 (d) Condition 11

(e) Mass (f) Mass

Figure 11. Cont.

Energies 2024, 17, 6316
x FOR PEER REVIEW 1817of
of 25
24

(g) COMz (h) COMz

Figure
Figure 11.
11. Prediction
Prediction error
error distribution
distribution diagram
diagram (a,c,e,g)
(a,c,e,g) and
and correlation
correlation diagram
diagram between
between predicted
predicted
value and experimental value (b,d,f,h).
value and experimental value (b,d,f,h).

The change range of the limiting mass and centroid is selected to be ±4% for predict-
ing the
Table 13.minimum
Estimation stress
range. in the feasible region. Meanwhile, the A variable (heave plate
bulkhead thickness) and C variable (heave plate stiffener size) are adjusted to 0.8~1.4 and
Lower Limit
0.7~1.2, respectively, for the variable
to make Upper Limit for
prediction Lowerfall
point Limit for the
within Upper Limitrange,
variable for
Name Goal
LC4 LC4 LC11 LC11
as shown in Table 13.
A In range 0.8 1.4 0.8 1.4
B In range 0.8
Table 13. Estimation range. 1.2 0.8 1.2
C In range 0.7 1.2 0.7 1.2
D In range Lower Limit for Upper
0.8 1.2 Limit for Lower
0.8 Limit for Upper1.2
Limit for
Name
von Mises stress
Goal
Minimize and in range 1.98 × 108 LC4 8
3.86 × 10LC4 108
1.98 ×LC11 × 108
3.86LC11
Mass In range 5.80 × 106 6.28 × 106 5.80 × 106 6.28 × 106
A In range 0.8 1.4 0.8 1.4
COMz In range −7.488 −6.912 −7.488 −6.912
B In range 0.8 1.2 0.8 1.2
C In range 0.7 1.2 0.7 1.2
D InThe response surface diagram
range 0.8 of the maximum 1.2 von Mises stress
0.8 value, mass, and
1.2centroid
von Mises stress position within the
Minimize and in range prediction range
1.98 × 10 8 under LC4 and
3.86 × 10 LC11
8 are presented
1.98 × 10 8 in Figures
3.8612 and
× 10 8 13,

Mass respectively.
In range The feasible region
5.80 × 106for minimum solution
6.28 × 106 of the maximum
5.80 × 106 stress
6.28 × 106 is
value
COMz determined
In range by the boundaries−7.488of the limit mass
−6.912and centroid position. In Figure
−7.488 −6.91212, the
predictions for mass and centroid position based on the response surface model under LC4
remain within the allowable limits. The minimum solution of the predicted von Mises
The response surface diagram of the maximum von Mises stress value, mass, and
stress maximum value is 2.21 × 108 Pa, with the credibility being as high as 0.88.
centroid position within the prediction range under LC4 and LC11 are presented in Figure
In Figure 13, it is found that the mass and COMz predicted by the response surface
12 and Figure 13, respectively. The feasible region for minimum solution of the maximum
model under LC11 remain within the allowable range, where the minimum solution of the
stress value is determined by the boundaries of the 8limit mass and centroid position. In
largest predicted von Mises stress value is 2.23 × 10 Pa, whose credibility value is 0.882
Figure 12, the predictions for mass and centroid position based on the response surface
with a fine predicted feasibility. Moreover, the variable values can be roughly estimated,
model under LC4 remain within the allowable limits. The minimum solution of the pre-
thanks to the optimal value prediction point in the response surface model falls within the
dicted von Mises stress maximum value is 2.21 × 108 Pa, with the credibility being as high
range of the four variables; subsequently, the response surface equation needs to be solved
as 0.88.
by using a numerical calculation method to determine the specific variable value.
In Figure 13, it is found that the mass and COMz predicted by the response surface
model under LC11 remain within the allowable range, where the minimum solution of
the largest predicted von Mises stress value is 2.23 × 108 Pa, whose credibility value is 0.882
with a fine predicted feasibility. Moreover, the variable values can be roughly estimated,
thanks to the optimal value prediction point in the response surface model falls within the
range of the four variables; subsequently, the response surface equation needs to be solved
by using a numerical calculation method to determine the specific variable value.
 Energies
Energies 2024, 2024, 17, x FOR PEER REVIEW
17, 6316 18 of 24 19 of 25

(a) Interaction between factor A and factor B.

(b) Interaction between factor A and factor D.

(c) Interaction between factor B and factor D.

Figure 12. Cont.

 Energies
Energies 2024, 2024, 17, x FOR PEER REVIEW
17, 6316 19 of 24 20 of 25

(d) Interaction between factor C and factor D.

Figure
Figure 12. 12. Response
Response surfaceunder
surface model model under LC4.
LC4.

(a) Interaction between factor A and factor B.

(b) Interaction between factor A and factor D.

Figure 13. Cont.

 Energies
Energies 2024, 2024, 17, x FOR PEER REVIEW
17, 6316 20 of 24 21 of 25

(c) Interaction between factor B and factor D.

(d) Interaction between factor C and factor D.

Figure
Figure 13. 13. Response
Response surfaceunder
surface model modelLC11.
under LC11.

4.4. Gradient DescentDescent

4.4. Gradient Method Method
SolutionSolution
The gradient descentdescent
The gradient methodmethod is a widely used numerical
is a widely algorithm
used numerical for solving
algorithm the
for solving the
minimum value of the objective function. The optimization process
minimum value of the objective function. The optimization process begins at a random begins at a random
point and progresses
point toward the
and progresses minimum
toward value along
the minimum the direction
value along theofdirection
the steepest descent
of the steepest de-
definedscent
by thedefined
objectiveby function.
the objective To enhance
function. robustness,
To enhance therobustness,
search process shouldprocess
the search initiate should
from multiple
initiatestarting points within
from multiple starting thepoints
design space.the
within After a certain
design space.number
After aofcertain
iterations,
number of
the optimal solution can be identified [39].
iterations, the optimal solution can be identified [39].
For the For
optimal targeting
the optimal targeting f ( Xi ), the
functionfunction 𝑓(𝑋 update formulaformula
), the update of the gradient descentdescent
of the gradient
methodmethod
is as follows [40]:
is as follows [40]:
Xi + 1 = Xi − α ∇ f ( Xi ) (20)
𝑋 = 𝑋 − 𝛼∇𝑓(𝑋 ) (20)
where Xi+1 is the updated parameter vector, Xi is the current parameter vector, α is the
learningwhere 𝑋 is
rate, which is the
used updated
to controlparameter
the step sizevector,
of the𝑋 update,
is the current
and ∇ fparameter vector, 𝛼 is the
( Xi ) is the gradient
learningfunction
of the objective rate, whichf ( Xis
i ) used
at X i .to control the step size of the update, and ∇𝑓(𝑋 ) is the gra-
dient of the objective function 𝑓(𝑋 ) at 𝑋 .
Based on the response surface equations established in Section 4.3, the constraint
conditions andBased on the response
optimization surface
objectives equations
are set establishedthe
up; subsequently, in Section
optimal4.3, the constraint
solution of the con-
responseditions
surfaceand optimization
model is obtained objectives
after 10,000are set up; subsequently,
iterations. the optimal
The optimization solution
process and of the
response
results are shownsurface
in Figuremodel14 andis obtained
Table 14, after 10,000 iterations. The optimization process and
respectively.
results are shown in Figure 14 and Table 14, respectively.
Energies 2024, 17, x FOR PEER REVIEW 22 of 25
Energies 2024, 17, 6316 21 of 24

(a) Maximum stress under LC4. (b) Maximum stress under LC11.

(c) Mass under LC4. (d) Mass under LC11.

(e) Center of mass under LC4. (f) Center of mass under LC11.
Figure 14. Gradient descent method solving process.
Figure 14. Gradient descent method solving process.
Table 14. Optimal results.
Table 14. Optimal results.
Variable Number Target Value
Loading Condition Variable Number Target Value
Loading Condition von Mises Stress
A A
B BC C D D von Mises Stress (Pa) Mass(kg)
Mass (kg) COMz(m)
COMz (m)
(Pa)
LC4 LC4 1.297 1.116
1.297 0.788 0.788
1.116 1.049 1.049 2.263 × 10×8 108
2.263 6.246 66
6.246××1010 −−7.488
7.488
8
× 10× 108 6 −−7.488
LC11 LC11 1.365 1.163
1.365 0.742 0.742
1.163 0.995 0.995 2.260
2.260 6.245××10106
6.245 7.488

4.5. Results
4.5. Results Verification
Verification
Toverify
To verifythe
themodel
modelobtained
obtainedby bythe
theoptimization
optimizationprocess,
process,aafinite
finiteelement
elementmodel
modelwaswas
establishedininGeniE
established GeniEsoftware
software 2021
2021 and
and imported
imported intointo HydroD
HydroD software
software 20212021 for struc-
for structural
strength calculation.
tural strength calculation.
The
The von Misesstress
von Mises stressof
ofthe
theoptimal
optimalmodel
modelisisshown
shownin inTable
Table14,14,where
wherethe
thepredicted
predicted
maximum
maximumstructural
structuralvon
vonMises
Misesstress
stressunder
underLC4
LC4isis226.3
226.3MPa,
MPa,larger
largerthan
than226
226MPa
MPaof ofthe
the
case under LC11. Therefore, the predicted model for the case under LC11 is determined to
Energies 2024, 17, 6316 22 of 24

be the optimal solution. As shown in Table 15, the mass and centroid position are increased
by 3.36% and 3.85%, respectively, which satisfies the requirement that the gap between the
initial and optimized model is within ±5%. Moreover, the maximum von Mises stress value
of the optimized model is reduced by 22.12% after optimization, achieving the purpose of
improving the structural safety in this study.

Table 15. Comparison before and after optimization.

Variable Name Optimized Model Variable Values Comparison

A Bulkhead thickness of heave plate 0.0259 m —
B Bulkhead thickness of pontoon 0.01163 m —
height = 0.5894 m,
C Size of heave plate stiffener web thickness = 0.00842 m, width = 0.1684 m, —
top flange thickness = 0.00842 m
height = 0.0348 m,
D Size of pontoon transverse stiffener web thickness = 0.00995 m, width = 0.199 m, —
top flange thickness = 0.00995 m
— Mass 6.245 × 106 kg 3.36%
— COMz −7.48826 m 3.85%
— von Mises stress 2.260 × 108 Pa −22.12%

5. Conclusions
In this paper, the floating wind turbine platform is taken as the research object, the local
structure is parametrically analyzed using the orthogonal test method to obtain the optimal
arrangement, and the optimal results are verified in terms of strength. The following
conclusions are obtained.
(1) The initial model of the structure is established in GeniE software 2021, which is
mainly composed of pontoon, buoy, and heave plate. The mass of the initial model
steel structure is 6035 t, and the coordinates of the center of mass position are (0, 0,
−7.2 m). The maximum stress value of the initial model was calculated in Sestra
software 2021 as 276 MPa.
(2) A total of 10 variables, namely bulkhead thickness, spacing of stiffener, and size of
stiffener, were parameterized for the three parts of pontoon, buoy, and heave plate. A
Plackett–Burman test was used to analyze the significance of the 10 variables, which
were ranked according to the p-value. It was determined that the thickness of the
heave plate and pontoon bulkhead, and the section size of heave plate and pontoon
stiffener are the most important factors influencing the magnitude of the structural
stress in the floating platform.
(3) The Central Composite Design method was used to further analyze the above four
variables. A total of 332 orthogonal tests were established, the factor levels of the
selected variables were in the range of 0.8~1.2, the stress maxima, mass, and center
of mass were used as the objective functions, all response surface models had p-
values less than 0.0001, and the model correlation coefficients and model correction
coefficients were above 0.98. The response surface models can well predict the changes
of the response values of the stress maximum, mass, and center of mass under various
working conditions.
(4) Based on the gradient descent algorithm, the minimum solution of the stress maximum
value was solved by controlling the function range of mass and center of mass. After
optimization, the maximum von Mises stress of the model was reduced by 22.12%
under extreme conditions, and the mass and center of mass were increased by 3.36%
and 3.85%, respectively, which satisfies the requirement of not exceeding ±5% between
the initial model and the optimized one, achieving the purpose of improving structural
safety in this study.
Energies 2024, 17, 6316 23 of 24

Author Contributions: Conceptualization, M.H.; Methodology, Y.R., M.H. and J.W.; Software, Y.R., M.H.,
J.H. and J.W.; Validation, Y.R.; Formal analysis, Y.R., J.H. and L.Z.; Investigation, Y.R., M.H., J.H., S.L., L.Z.,
H.Z. and W.S.; Resources, J.H., J.W., S.L. and L.Z.; Data curation, J.H., J.W. and S.L.; Writing—original draft,
Y.R. and M.H.; Writing—review & editing, H.Z. and W.S.; Supervision, H.Z.; Funding acquisition,
W.S. All authors have read and agreed to the published version of the manuscript.
Funding: This work is financially supported by the China Renewable Energy Engineering Institute
Research Program (ZS-KJSD-20230005). Their financial supports are gratefully acknowledged.
Data Availability Statement: The original contributions presented in this study are included in the
article. Further inquiries can be directed to the corresponding author.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wang, F.; Hao, J.; Li, S.; Ren, J.; Xie, Y.; Zhang, B. Key technology and development trend of floating offshore wind power. Water
Power 2022, 48, 9–12+117.
2. Xia, Y. In 2023, global offshore wind power will add 10.8 GW of new installed capacity. WIND Energy Mag. 2024, 27, 42–44.
3. Liu, X.; Han, X.; Qin, M. Status of floating wind power technology and prospect of offshore wind power development in China.
China Offshore Oil Gas 2024, 36, 233–242.
4. Wan, L.; Moan, T.; Gao, Z.; Shi, W. A review on the technical development of combined wind and wave energy conversion
systems. Energy 2024, 294, 130885. [CrossRef]
5. Tian, X.; Sun, X.; Liu, G.; Xie, Y.; Chen, Y.; Wang, H. Multi-objective optimization of the hull form for the semi-submersible
medical platform. Ocean Eng. 2021, 230, 109038. [CrossRef]
6. Yang, H.; Zhu, Y.; Lu, Q.; Zhang, J. Dynamic reliability based design optimization of the tripod sub-structure of offshore wind
turbines. Renew. Energy 2015, 78, 16–25. [CrossRef]
7. Saeed, N.; Gong, J.; Wan, Y.; Long, K.; Saeed, A.; Mei, L.; Xiong, C.; Long, W.; Zhou, H.; Li, L. A novel design of multifunctional
offshore floating platform structure based on topology optimization. Eng. Struct. 2024, 306, 117782. [CrossRef]
8. Chen, M.; Huang, W.; Liu, H.; Hallak, T.S.; Liu, S.; Yang, Y.; Tao, T.; Jiang, Y. A novel SPM wind-wave-aquaculture system:
Concept design and fully coupled dynamic analysis. Ocean Eng. 2025, 315, 119798. [CrossRef]
9. Lu, F.; Long, K.; Zhang, C.; Zhang, J.; Tao, T. A novel design of the offshore wind turbine tripod structure using topology
optimization methodology. Ocean Eng 2023, 280, 114607. [CrossRef]
10. Karimi, M.; Hall, M.; Buckham, B.; Crawford, C. A multi-objective design optimization approach for floating offshore wind
turbine support structures. J. Ocean Eng. Mar. Energy 2017, 3, 69–87. [CrossRef]
11. Lee, J.C.; Shin, S.C.; Kim, S.Y. An optimal design of wind turbine and ship structure based on neuro-response surface method. Int.
J. Nav. Archit. Ocean. Eng. 2015, 7, 750–769. [CrossRef]
12. Ferri, G.; Borri, C.; Marino, E. Substructure optimization of a 10MW floating wind turbine for installations in the Mediterranean
Sea. J. Phys. Conf. Ser. 2024, 2647, 112011. [CrossRef]
13. Choi, E.; Han, C.; Kim, H.; Park, S. Optimal design of floating substructures for spar-type wind turbine systems. Wind Struct.
2014, 18, 253–265. [CrossRef]
14. Benifla, V.; Adam, F. Development of a genetic algorithm code for the design of cylindrical buoyancy bodies for floating offshore
wind turbine substructures. Energies 2022, 15, 1181. [CrossRef]
15. Dou, S.; Pegalajar-Jurado, A.; Wang, S.; Bredmose, H.; Stolpe, M. Optimization of floating wind turbine support structures using
frequency-domain analysis and analytical gradients. J. Phys. Conf. Ser. 2020, 1618, 042028. [CrossRef]
16. Leimeister, M.; Kolios, A.; Collu, M.; Thomas, P. Design optimization of the OC3 phase IV floating spar-buoy, based on global
limit states. Ocean Eng. 2020, 202, 107186. [CrossRef]
17. Hall, M.; Buckham, B.; Crawford, C. Evolving Offshore Wind: A Genetic Algorithm-Based Support Structure Optimization
Framework for Floating Wind Turbines. In Proceedings of the 2013 MTS/IEEE OCEANS-Bergen, Bergen, Norway, 10–14 June
2013.
18. Hegseth, J.M.; Bachynski, E.E.; Martins, J.R. Integrated design optimization of spar floating wind turbines. Mar. Struct. 2020, 72,
102771. [CrossRef]
19. Tang, Y.; Shen, G.; Liu, L. Structural Dynamics of Ocean Engineering; Tianjin University Press: Tianjin, China, 2008.
20. DNV GL. Stability and Watertight Integrity: DNVGL-OS-C301; DNV GL: Høvik, Norway, 2020.
21. Islam, M.T. Design, Numerical Modelling and Analysis of a Semi-Submersible Floater Supporting the DTU 10MW Wind Turbine.
Master’s Thesis, Norwegian University of Science and Technology, Trondheim, Norway, 2016.
22. Zhu, Y. Wave Mechanics in Ocean Engineering; Tianjin University Press: Tianjin, China, 1991.
23. Liu, Y. Research on Structural Strength Analysis Method of Single-Column Floating Wind Turbine; Shanghai Jiao tong University:
Shanghai, China, 2014.
24. SESAM User Manual, HydroD v4.10. In Wave Load & Stability Analysis of Fixed and Floating Structures. 2020. Available online:
https://manualzz.com/doc/7321164/det-norske-veritas-hydrod-user-manual (accessed on 1 November 2024).
Energies 2024, 17, 6316 24 of 24

25. Sang, S.; Yu, M.; Shi, X.; Yi, S.; Yang, S.; Fei, S. Structural Strength Calibration of Semi-submersible Floating Wind Turbine Based
on the Design Wave Method. Acta Energiae Solaris Sin. 2019, 40, 185–191.
26. Song, Z. Structural Design and Dynamic Characterization of Large-Scale Floating Offshore Wind Turbine Platform; Dalian University of
Technology: Dalian, China, 2022.
27. DNV GL-OS-C103; Structural Design of Columnstabilised Units-LRFD Method. DNV GL: Høvik, Norway, 2020.
28. Wang, Y.; Xiu, S. Multi-objective optimization design of motor mounts applying response surface methodology. Mech. Des. Manuf.
2021, 10, 42–44.
29. Suresh, R.; Basavarajappa, S.; Samuel, G.L. Predictive modeling of cutting forces and tool wear in hard turning using response
surface methodology. Procedia Eng. 2012, 38, 73–81. [CrossRef]
30. Ma, K.; Ren, F.; Wang, H.; Li, L.; Wu, D. Dynamic mechanical responses and freezing strengthening mechanism of frozen
sandstone with single flaw: Insights from drop weight tests and numerical simulation. Rock Mech. Rock Eng. 2024, 57, 1263–1285.
[CrossRef]
31. Liu, J.; Zhang, T.; Jin, P.; Zhu, B. Optimization of structural parameters of electric chain saw teeth based on Design-Expert.
For. Eng. 2024, 2, 142–150.
32. Wang, J.; Ren, Y.; Shi, W.; Collu, M.; Venugopal, V.; Li, X. Multi-objective optimization design for a 15 MW semisubmersible
floating offshore wind turbine using evolutionary algorithm. Appl. Energy 2025, 377, 124533. [CrossRef]
33. Quinlan, K.R.; Lin, D.K. Run order considerations for Plackett and Burman designs. J. Stat. Plan. Inference 2015, 165, 56–62.
[CrossRef]
34. Yang, X.; Huang, S.; Shi, J.; Wang, B.; Wang, K.; Dong, S.; Li, H. Research on day-ahead electricity price prediction method based
on dynamic data and multi-factor similarity day. Electr. Power Big Data 2024, 27, 35–44.
35. Lenth, R.V. Quick and easy analysis of unreplicated factorials. Technometrics 1989, 31, 469–473. [CrossRef]
36. Zhang, Z.; Li, Y.; Shi, L.; Wang, H. Improved response surface method and its application in structural reliability analysis. Eng.
Mech. 2007, 8, 111–115+187.
37. Benalia, A.; Baatache, O.; Derbal, K.; Khalfaoui, A.; Amrouci, Z.; Pizzi, A.; Panico, A. The use of central composite design (CCD)
to optimize and model the coagulation-flocculation process using a natural coagulant: Application in jar test and semi-industrial
scale. J. Water Process Eng. 2024, 57, 104704. [CrossRef]
38. Verma, A.S.; Yan, J.; Hu, W.; Jiang, Z.; Shi, W.; Teuwen, J.J. A review of impact loads on composite wind turbine blades: Impact
threats and classification. Renew. Sustain. Energy Rev. 2023, 178, 113261. [CrossRef]
39. Mustapha, A.; Mohamed, L.; Ali, K. An overview of gradient descent algorithm optimization in machine learning: Application in
the ophthalmology field. In Smart Applications and Data Analysis: Third International Conference, SADASC 2020, Marrakesh, Morocco,
25–26 June 2020; Proceedings 3; Springer International Publishing: Cham, Switzerland, 2020; pp. 349–359.
40. Rezk, H.; Ferahtia, S.; Djeroui, A.; Chouder, A.; Houari, A.; Machmoum, M.; Abdelkareem, M.A. Optimal parameter estimation
strategy of PEM fuel cell using gradient-based optimizer. Energy 2022, 239, 122096. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.