491
1 INTRODUCTION
Offshore wind energy has emerged as a key player in
the global transition towards renewable energy [4].
Over the past decade, offshore wind installations have
grown rapidly, with global capacity surpassing
67.4 GW by 2024 [16], and projections indicate
continued expansion, particularly in regions such as
the North Sea and the East China Sea. This growth is
driven by the advantages of offshore wind over
onshore alternatives, higher and more consistent wind
speeds, reduced land use conflicts, and proximity to
densely populated coastal demand centers.
Despite its advantages, such as higher wind speeds and
reduced land use constraints, offshore wind farms face
significant operational challenges, particularly harsh
marine conditions, such as high winds, waves,
corrosion, and limited accessibility, creating a complex
and costly environment for maintenance activities. In
addition, accessing offshore turbines often requires
specialized vessels, helicopters, or remotely operated
vehicles, with substantial logistical planning. This not
only inflates costs but also introduces high operational
uncertainty, as weather windows narrow and delays
become frequent. The harsh marine environment and
unpredictable wind conditions complicate preventive
and corrective maintenance operations. Unscheduled
maintenance interventions can lead to prolonged
downtimes and high logistical costs, making efficient
maintenance scheduling essential [1].
Current maintenance strategies typically combine
preventive maintenance, which follows scheduled
intervals but can lead to unnecessary interventions;
corrective maintenance, which addresses failures as
they occur, often resulting in prolonged downtime and
high reactive costs; and condition-based and predictive
maintenance, which utilize sensor data and machine
Maintenance Activities Coordination for Offshore Wind
Farms Integrating Multivariate Stochastic Models
Y. Salgado Duarte & J. Szpytko
AGH University of Kraków, Kraków, Poland
ABSTRACT: The efficiency and reliability of offshore wind farms are significantly influenced by their
maintenance strategies. Effective maintenance coordination is crucial to minimize downtime and consequently
maximize reliability. This paper proposes a methodology for integrating multivariate stochastic models into the
maintenance scheduling process of offshore wind farms, enabling a more data-driven approach. The proposed
framework accounts for the temporal and spatial dependencies of wind variations to optimize maintenance
activities, ensuring minimal disruptions to energy production using the Expected Energy Not Supplied (EENS)
indicator. We validate our approach using case studies that demonstrate its effectiveness through a sensitivity
analysis. The results highlight more than 6% deviations in the EENS estimate for scenarios with the same
conditions, where the only difference is the wind speed simulation approach. The change in approach impacts
the overall allocation of maintenance activities to offshore wind farms.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 19
Number 2
June 2025
DOI: 10.12716/1001.19.02.19
492
learning to forecast failures, reducing unexpected
outages and improving reliability.
However, even the most advanced strategies often
rely on deterministic models that do not fully
incorporate the stochastic, spatio-temporal nature of
offshore wind conditions. This omission can
significantly undermine planning accuracy. For
example, variability in wind speed and direction, and
their correlation between spatially distributed
turbines, has a positive impact on energy production
and accessibility, although these are often
oversimplified in existing models [2].
As offshore wind farms scale in size and number,
the complexity of coordinating maintenance schedules
between multiple geographically dispersed units
increases. There is a pressing need for integrated
probabilistic decision-making frameworks that
account for environmental uncertainty, asset
conditions, and system-level constraints.
This paper addresses this need by proposing a
methodology that incorporates multivariate stochastic
wind modeling into the offshore maintenance
scheduling process. The approach accounts for both
temporal and spatial wind dependencies and
introduces more realism into the simulation of wind
behavior. Using the Expected Energy Not Supplied
(EENS) as a reliability indicator, the framework is
validated through sensitivity analysis and case studies.
Ultimately, the goal is to minimize downtime, enhance
system reliability, and enable smarter data-driven
maintenance planning. This paper proposes a revised
methodology for coordinating offshore wind farms'
maintenance activities that leverages multivariate
stochastic wind modeling to account for a more real-
driven approach, introducing new considerations into
the modeling process. The methodology, subject to
review, is taken from [3], and the study presented in
this article aims to address the following key research
questions.
How can multivariate stochastic wind models be
effectively integrated into maintenance decision-
making?
What are the potential impacts of integrating
multivariate stochastic wind models on offshore wind
farm maintenance scheduling?
How does the proposed approach compare with
existing maintenance optimization strategies in terms
of results?
The remainder of this paper is organized as follows.
Section 2 reviews the relevant literature on
maintenance planning and stochastic wind modeling,
and at the end of the section, the proposed
methodology and the new considerations addressed in
this contribution are presented. Section 3 presents the
parameterization, results and discussion of the
optimization solution for the case study, and Section 4
concludes with recommendations for future research.
2 MATERIALS AND METHODS
The maintenance of offshore wind farms has been
extensively studied in the literature, with a strong
emphasis on preventive and predictive maintenance
strategies [5]. First, this section reviews existing work
on maintenance planning for offshore wind farms and
highlights the role of stochastic wind modeling in
improving modeling and, consequently, decision
making. Then, knowing that the scale of offshore wind
farms is increasing daily and, certainly, the complexity
of coordinating maintenance schedules across multiple
geographically dispersed units increases, we present a
revised methodology with new considerations that
proposes an integrated probabilistic decision-making
framework that accounts for environmental
uncertainty, asset conditions, and system-level
constraints.
2.1 Review of the literature
Maintenance strategies have been proposed to
optimize offshore wind farm operations, mainly
preventive maintenance [7], predictive maintenance
[5], and condition-based maintenance [6]. As we know,
preventive maintenance schedules interventions based
on fixed time intervals, reducing unexpected failures,
but potentially leading to unnecessary service.
Predictive maintenance predicts the remaining useful
life of the equipment to anticipate potential failures,
enabling early detection of problems and timely
maintenance actions. Condition-based maintenance
uses real-time sensor data and predictive analytics to
schedule interventions based on actual component
conditions. Certainly, the most recent state-of-the-art in
this domain is the integration of these strategies. For
example, studies have shown that condition-based
maintenance significantly improves operational
efficiency by reducing unnecessary maintenance
actions and increasing turbine availability [8].
However, these approaches (preventive, predictive,
and condition-based) often neglect the stochastic
nature of offshore wind conditions, which can affect
accessibility and the feasibility of scheduled
maintenance operations.
Stochastic modeling of wind speed and turbulence
plays a crucial role in offshore wind energy forecasting
and decision making. Traditional wind forecasting
models rely on deterministic methods, which do not
capture the spatial-temporal uncertainties of offshore
wind patterns. On the contrary, stochastic models,
such as Markov processes, autoregressive integrated
moving average (ARIMA) models and Gaussian
mixture models, have been employed to provide
probabilistic wind forecasts [9].
Multivariate stochastic models extend these
approaches by considering the correlations between
different environmental variables, such as wind speed,
wind direction, and wave height. This improves
predictive accuracy and allows for more robust
maintenance planning [10]. For example, recent
research has shown that integrating these models into
offshore wind maintenance scheduling can improve
accessibility predictions and optimize vessel dispatch
planning [11].
Studies have explicitly incorporated multivariate
stochastic wind modeling into maintenance
coordination frameworks. Existing research focuses
primarily on single-variable wind speed forecasts,
neglecting interactions between wind, wave
conditions, and turbine reliability metrics. The need for
493
an integrated approach that captures these
dependencies is evident, particularly for offshore wind
farms, where accessibility constraints are highly
dependent on weather [12].
Recent advances in probabilistic optimization have
shown promise in addressing these challenges.
Methods such as Monte Carlo simulations, Bayesian
inference, and reinforcement learning have been
applied to improve maintenance decision making
under uncertainty [13]. However, there is still a gap in
integrating these approaches with multivariate
stochastic wind models to develop a fully adaptive and
proactive maintenance framework.
The reviewed literature highlights the interest in
optimizing offshore wind farm maintenance through
predictive and data-driven approaches. However,
there remains a lack of comprehensive frameworks
that integrate multivariate stochastic wind models into
maintenance coordination.
2.2 Contribution of this paper.
This paper aims to contribute by presenting a revised
methodology that uses multivariate probabilistic
modeling techniques to improve the maintenance
scheduling modeling process. As previously
highlighted, the methodology, subject to review, is
taken from [3]. However, in this section, we present
how the new consideration, essentially multivariate
stochastic wind modeling, is incorporated into the
existing methodology.
Figure 1 shows the revised flow diagram of the
optimization problem. Now, the wind energy and
thermal units are explicitly separated, although all the
generation units are aggregated at the system level. As
we can see in Figure 1, the heuristic optimization
algorithms propose a set of xi = Mi,1 starting
maintenance activity time for each generator unit
considered in the system, given the constraints of the
simulation window. The remaining parameters needed
to simulate the stochastic capacity of the units are
directly sourced from a SQL (Structured Query
Language) database.
Specifically, in the case of wind energy, we
introduce in this contribution a modeling change,
which aims to consider correlations in the wind speed
simulation. Essentially, considering correlations in
wind speed simulation means that the wind speed
characteristic in a region of the offshore wind farm is
homogeneous and that all wind turbines should
deliver a similar energy path. As we know, in the case
of wind energy, the primary source of energy is the
wind, so it is necessary first to simulate the wind speed
and then, using the characteristic function of the wind
turbine, translate the wind speed into energy.
In this research, we simulate the most likely wind
speed v with a Weibull model, estimating the shape
and scale parameters of the probability distribution
function with the mean and standard deviation of the
historical wind speed characteristics of the region.
Once the Weibull distribution is parameterized, the
inverse cumulative distribution function is used to
simulate the wind speed values v by generating
uniform random numbers u with (1).
( )
1/
lnvu
=−


(1)
where v is the wind speed, β and δ are the shape and
scale parameters of the Weibull probability
distribution function, and u is a uniform random
number between [0, 1].
As we can see, the number of random numbers u to
be generated depends on t = 1, 2, …, T, which is the
hourly simulation window.
Regardless of how the correlations can be
measured, we intend to assess the impact of correlated
wind speed simulations on the risk indicators of the
system, specifically assessing the disruptions to energy
production using the Expected Energy Not Supplied
(EENS) indicator, and eventually in the maintenance
scheduling activities to be conducted in the offshore
wind farms.
In this paper, we use the Gaussian elliptical copula
to simulate uniformly correlated samples U. In the
Gaussian copula, the correlation parameter ρ controls
the dispersion of random values. Certainly, values of ρ
close to one mean more correlated marginals and more
homogeneous wind speed simulations. In Section 3, we
illustrate the impact of highly correlated wind speed
simulations on the wind farm delivery energy path.
As we may know, copula functions try to capture
the dependence between marginals through the copula
parameter. Certainly, each density may have a
different operating space. In the case of the
Archimedean copulas, the parameter manages the
dispersion of the random numbers. For instance,
higher parameter values result in less dispersion of the
random values. When the parameter is close to one, the
random values are not correlated.
On the other hand, the Gaussian and t-copula are
elliptical copulas. In these cases, the correlation
parameter ρ controls the dispersion of the random
values. Values of ρ close to one, more correlated
marginals. Within the elliptical copulas, the t-copula
has two parameters; therefore, the t-copula offers
another feature, the tail dependency. This additional
parameter allows the t-copula to simulate values at the
corners of the distribution space. This flexibility is
given by the degrees of freedom υ. For instance, fixing
the correlation parameter and changing the degrees of
freedom to higher values, in a t-copula density, results
in a Gaussian copula. That said, t-copula is more
flexible than Gaussian.
We selected the Gaussian copula in this
investigation for two main reasons. First, elliptical
copulas can easily model high dimensions, since the
diagonal of the correlation matrix has the same
dimension as the number of components considered.
Archimedean copulas are practically implemented for
only two components. Second, since we do not have a
quantitative perception of the right value for the
degree of freedom υ in the t-copula and is certainly
more costly from the computation point of view, we
decided to use Gaussian over t-copula because we aim
to assess if correlated wind values impact the delivery
energy path and maintenance scheduling activities
allocation, and for that, Gaussian is sufficient.
494
Figure 1. Flow diagram to solve the optimization problem of maintenance scheduling.
Previously, with (1), uncorrelated random numbers
ui were generated for each unit i considered in the
system, to obtain the wind speed values vi for each
wind turbine. Now, we simulate a matrix with
dimension [I, T] of correlated marginals U[0,1](ρ), where
ρ is the correlation parameter controlling the
dispersion of random values, I is the number of
offshore wind turbines within the wind farm, and T is
the simulation window in hourly resolution. Then,
using the UI,T, and the parametrized Weibull
distribution, we obtain a matrix of VI,T, which is the
correlated wind speed simulations as shown in (2), and
then, with the power characteristics function of the
wind turbine, we translate the wind speed values into
energy. For wind turbines, the variable Pi(t|v)
represents the wind power of the i-th wind turbine at
the t-th time, as illustrated in Figure 1.
(2)
At this point, we end the description of the
modeling changes. The following modeling steps
remain unchanged as presented in [3]. In Section 3, we
present the parameterized scenario used to assess the
impact of the modeling change introduced in this
paper.
3 RESULTS AND DISCUSSION
In this section, we describe scenario parameterization,
then we illustrate the experiment designed to measure
the deviations in the EENS estimations when
correlations in the wind speed are considered, and then
we provide the solution for the system maintenance
scheduling activities under highly correlated wind
speed conditions using two heuristic optimization
algorithms. The parameterization of each algorithm is
also presented, as well as a comparison between them.
The base scenario described in [3] is partially used
to validate the revised methodology. However, we
recap key information here to highlight what is
relevant for the analysis of this contribution. The IEEE
Reliability Test System (IEEE-RTS) suggested by the
Subcommittee on Applications of Probabilistic
Methods provides the basis for the scenario and is
particularly suitable for implementing maintenance
planning solutions [15]. Tables 1 and 2 show the
parameterization implemented in the scenario under
study. The original IEEE-RTS has 32 generators and an
overall capacity of 3405 MW. However, a modification
is introduced in the scenario to integrate offshore wind
farms. The wind turbine parameters assumed to make
up the offshore wind farms modeled in this
contribution are listed in Table 3. This modified
scenario replaced three 100 MW Oil/Stream generator
units with offshore wind farms. In particular, three
farms with 50 wind turbines of 2 MW, making a total
of 300 MW, which is the same capacity as the replaced
generator units.
In this modified scenario, even if the replacement
total nominal capacity is the same (300 MW), there is a
difference in the primary source of energy. Usually, in
the case of Oil/Stream generators, the primary source
of energy is always available. In the case of wind farms,
the primary source of energy is the wind and it is
subject to changes and variability. Therefore, the
contribution of wind energy will never be 100%; in
other words, even if we have 300 MW installed, in
practice, the potential energy to be used will certainly
be less. Selecting the same nominal capacity points out
this difference.
495
In this modified scenario, as we introduced three
offshore wind farms, the number of individual
generator units increased from 32 to 179. Figure 2
shows the connection schema of the wind farms with
the Power System, which is the connection to the grid.
Essentially, three wind farms are connected in parallel
and located in different locations, and the wind
characteristics are also related to the location.
Since we use a Monte Carlo method to estimate the
EENS, we assume that β = 5% (error criterion) and T =
8760 hours (simulation time window).
Table 1. Parameterization of Scenario (A)
No.
Units
Unit Capacity (MW) i
MTTF (hours)
MTTR (hours)
5
12
2940
60
4
20
450
50
6
50
1980
20
4
76
1960
40
4
155
960
40
3
197
950
50
1
350
1150
100
2
400
1100
150
150
2
3650
55
Table 2. Parameterization of Scenario (B)
No.
Units
Mi,k (hours/year)
Maintenance (hours/year) Di,k
NMTi
5
672
168, 168
2
4
672
168, 168
2
6
672
168, 168
2
4
672, 672
168, 168, 168
3
4
672, 672, 672
168, 168, 168, 168
4
3
672, 672, 672
168, 168, 168, 168
4
1
672, 672, 672, 672
168, 168, 168, 168, 168
5
2
672, 672, 672, 672, 672
168, 168, 168, 168, 168, 168
6
150
672, 672, 672, 672, 672
168, 168, 168, 168, 168, 168
6
Table 3. Wind speed and wind turbine parameters
Parameter
Values Considered
µsw
19.52 km/h
σsw
10.99 km/h
Pr
2 MW
vci
15 km/h
vr
36 km/h
vco
80 km/h
Once the parameterization is known, we assess the
deviations in EENS when correlations in the wind
speed are considered within each wind farm. The
differences are clearly illustrated in Figures 3-6 and
discussed in this section.
Figure 3 shows the wind farm 1 delivery energy in
the first five days when the wind simulation is
assumed to be uncorrelated. As we can see, even if the
total capacity of wind farm 1 is 100 MW, there is no
peak value above 35 MW in the simulation. On the
other hand, Figure 4 shows the same wind farm 1, but
under the assumption that wind speed simulations are
highly correlated. Certainly, the outcome of wind farm
delivery is between these two extreme simulations
(uncorrelated and highly correlated) in a real scenario,
and the shape of how the energy is delivered changes
drastically.
In the system analyzed, we consider three wind
farms. Figure 5 shows the energy delivered in the first
five days if uncorrelated wind simulations are
assumed. As we can see, 300 MW of wind capacity only
delivers a peak value slightly above 90 MW. On the
other hand, Figure 6 shows the delivery energy when a
highly correlated wind speed is assumed in each wind
farm, but also assuming that each wind farm is in a
different location, which means that there should be no
correlation between wind farms. Again, the shape of
the delivery energy changes, and we can see peak
values reaching almost 250 MW.
Table 4. Variations in the correlation of wind speed.
Dev.
Wind Farm 1
Wind Farm 2
Wind Farm 3
EENS
6.23%
0.99
0.99
0.99
20,385
5.60%
0.90
0.90
0.90
20,265
4.93%
0.80
0.80
0.80
20,137
4.40%
0.70
0.70
0.70
20,035
2.79%
0.60
0.60
0.60
19,725
2.03%
0.50
0.50
0.50
19,579
1.73%
0.40
0.40
0.40
19,522
1.56%
0.30
0.30
0.30
19,490
1.18%
0.20
0.20
0.20
19,416
0.33%
0.10
0.10
0.10
19,253
0.00%
0.00
0.00
0.00
19,190
Figure 2. Power system and offshore wind farms connection schema.
496
Figure 3. One wind farm simulated delivery energy with no-correlated wind speed.
Figure 4. One wind farm simulated delivery energy with fully correlated wind speed.
Knowing that the shape of the delivery energy
changes, we first quantify the holistic impact using the
EENS indicator (MWh / year) and a predetermined
maintenance scheduling activity solution for this case
study, taken from [3]. Then we gradually introduce the
correlations as described in Table 4.
As we can see in Table 4, deviations above 6% are
observed in the EENS indicator. In fact, when an
uncorrelated wind speed is assumed in the simulation,
the dimension of the capacity of the system can be
overestimated, as Figure 7 shows. In this experiment,
we kept the random seeds unchanged to isolate the
actual impact of the correlations. Additionally, Table 5
shows deviations slightly higher than 3% in the annual
energy delivered by wind farms when correlations are
introduced. This result confirms that assuming
correlations in wind speed impacts not only the
individual energy delivery of each wind farm, but also
the overall behavior of the system, since the deviations
in the EENS indicator are higher than the deviations of
energy in the wind farms.
Table 5. Energy delivered (MWh/year)
Wind Farm
Uncorrelated
Correlated
Deviation
1
186,607
183,381
1.76%
2
185,665
190,449
-2.51%
3
185,814
191,879
-3.16%
Once presented with the impacts of the new
assumption, we solve the maintenance scheduling
problem under the new conditions, specifically, under
a highly correlated assumption (ρ = 0.99). For this aim,
we tested two heuristic optimization algorithms. In this
paper, we use MATLAB [14] in the proposed
implementation, and we use well-implemented
MathWorks algorithms for the solution, specifically
Particle Swarm Optimization (PSO) and Surrogate
Optimization. Independent of the algorithm strategy,
we assume that the optimization process ends when
the difference between two consecutive evaluations in
the objective function is less than 1.0E-06. Tables 6 and
7 show the maintenance scheduling activities solution
for each algorithm, and Figures 8 and 9 show the
convergence of the algorithms.
In the case of Surrogate, the algorithm required
8,950 evaluations in the objective function to achieve an
EENS value of 20,798 MWh/year. The PSO required
26,000 evaluations (260 iterations and populations of
100 candidates) to achieve 17,306 MWh/year. Although
Surrogate requires less execution time, the PSO can
find a better solution. Since we face a nondeterministic
polynomial time (NP) problem addressed with
heuristics, algorithms are prone to finding a local
minimum. Therefore, for this problem, PSO provides
the best solution.
497
Figure 5. Three wind farms simulated delivery energy without correlated wind speed.
Figure 6. Three wind farms simulated the delivery energy with a fully correlated wind speed.
As we can see, the PSO adapted the solution to
the new modeling conditions, since the EENS
achieved = 17,306 MWh/year is 18% lower than the
previous scheduling solution with the new
modeling conditions, EENS = 20,385 MWh/year.
Figure 10 illustrates the system capacity versus
system load under the new conditions, and Figure
11 shows the spatial distribution of the solution for
maintenance scheduling activities.
Table 6. Maintenance scheduling solutions for offshore wind farms
Unit
No.
Capacity
Start Maintenance
PSO
Surrogate
1
12
1727
'13-Mar-1900 23:00:00'
7053
'21-Oct-1900 21:00:00'
2
12
4007
'16-Jun-1900 23:00:00'
1392
'28-Feb-1900 00:00:00'
3
12
4674
'14-Jul-1900 18:00:00'
6344
'22-Sep-1900 08:00:00'
4
12
1117
'16-Feb-1900 13:00:00'
6296
'20-Sep-1900 08:00:00'
5
12
589
'25-Jan-1900 13:00:00'
0
'01-Jan-1900 00:00:00'
6
20
5562
'20-Aug-1900 18:00:00'
6214
'16-Sep-1900 22:00:00'
7
20
4999
'28-Jul-1900 07:00:00'
5406
'14-Aug-1900 06:00:00'
8
20
4548
'09-Jul-1900 12:00:00'
3954
'14-Jun-1900 18:00:00'
9
20
1190
'19-Feb-1900 14:00:00'
4970
'27-Jul-1900 02:00:00'
10
50
6091
'11-Sep-1900 19:00:00'
1549
'06-Mar-1900 13:00:00'
11
50
5958
'06-Sep-1900 06:00:00'
1287
'23-Feb-1900 15:00:00'
12
50
6600
'03-Oct-1900 00:00:00'
2496
'15-Apr-1900 00:00:00'
13
50
4432
'04-Jul-1900 16:00:00'
1436
'01-Mar-1900 20:00:00'
14
50
885
'06-Feb-1900 21:00:00'
5142
'03-Aug-1900 06:00:00'
15
50
5794
'30-Aug-1900 10:00:00'
4219
'25-Jun-1900 19:00:00'
16
76
614
'26-Jan-1900 14:00:00'
2980
'05-May-1900 04:00:00'
17
76
6216
'17-Sep-1900 00:00:00'
5656
'24-Aug-1900 16:00:00'
18
76
5531
'19-Aug-1900 11:00:00'
4924
'25-Jul-1900 04:00:00'
19
76
4648
'13-Jul-1900 16:00:00'
1667
'11-Mar-1900 11:00:00'
20
155
4295
'28-Jun-1900 23:00:00'
5039
'29-Jul-1900 23:00:00'
21
155
2340
'08-Apr-1900 12:00:00'
4697
'15-Jul-1900 17:00:00'
22
155
130
'06-Jan-1900 10:00:00'
63
'03-Jan-1900 15:00:00'
23
155
5575
'21-Aug-1900 07:00:00'
4574
'10-Jul-1900 14:00:00'
24
197
275
'12-Jan-1900 11:00:00'
5105
'01-Aug-1900 17:00:00'
25
197
3
'01-Jan-1900 03:00:00'
972
'10-Feb-1900 12:00:00'
26
197
4546
'09-Jul-1900 10:00:00'
4850
'22-Jul-1900 02:00:00'
27
350
1317
'24-Feb-1900 21:00:00'
530
'23-Jan-1900 02:00:00'
28
400
1840
'18-Mar-1900 16:00:00'
2749
'25-Apr-1900 13:00:00'
29
400
1642
'10-Mar-1900 10:00:00'
1689
'12-Mar-1900 09:00:00'
30
2
3880
'11-Jun-1900 16:00:00'
1294
'23-Feb-1900 22:00:00'
31
2
3589
'30-May-1900 13:00:00'
2115
'30-Mar-1900 03:00:00'
32
2
3085
'09-May-1900 13:00:00'
1531
'05-Mar-1900 19:00:00'
33
2
599
'25-Jan-1900 23:00:00'
736
'31-Jan-1900 16:00:00'
34
2
4388
'02-Jul-1900 20:00:00'
1519
'05-Mar-1900 07:00:00'
35
2
2122
'30-Mar-1900 10:00:00'
2420
'11-Apr-1900 20:00:00'
36
2
4389
'02-Jul-1900 21:00:00'
1800
'17-Mar-1900 00:00:00'
37
2
1902
'21-Mar-1900 06:00:00'
4141
'22-Jun-1900 13:00:00'
38
2
242
'11-Jan-1900 02:00:00'
1566
'07-Mar-1900 06:00:00'
39
2
3961
'15-Jun-1900 01:00:00'
1423
'01-Mar-1900 07:00:00'
40
2
1721
'13-Mar-1900 17:00:00'
3541
'28-May-1900 13:00:00'
41
2
1364
'26-Feb-1900 20:00:00'
2290
'06-Apr-1900 10:00:00'
42
2
3247
'16-May-1900 07:00:00'
61
'03-Jan-1900 13:00:00'
43
2
340
'15-Jan-1900 04:00:00'
2219
'03-Apr-1900 11:00:00'
44
2
7
'01-Jan-1900 07:00:00'
4118
'21-Jun-1900 14:00:00'
45
2
4378
'02-Jul-1900 10:00:00'
3994
'16-Jun-1900 10:00:00'
46
2
4182
'24-Jun-1900 06:00:00'
1593
'08-Mar-1900 09:00:00'
47
2
4278
'28-Jun-1900 06:00:00'
331
'14-Jan-1900 19:00:00'
48
2
3
'01-Jan-1900 03:00:00'
3441
'24-May-1900 09:00:00'
49
2
726
'31-Jan-1900 06:00:00'
1763
'15-Mar-1900 11:00:00'
50
2
228
'10-Jan-1900 12:00:00'
3167
'12-May-1900 23:00:00'
51
2
2241
'04-Apr-1900 09:00:00'
2153
'31-Mar-1900 17:00:00'
52
2
4064
'19-Jun-1900 08:00:00'
2353
'09-Apr-1900 01:00:00'
498
53
2
2583
'18-Apr-1900 15:00:00'
3018
'06-May-1900 18:00:00'
54
2
3732
'05-Jun-1900 12:00:00'
4
'01-Jan-1900 04:00:00'
55
2
2658
'21-Apr-1900 18:00:00'
3652
'02-Jun-1900 04:00:00'
56
2
4351
'01-Jul-1900 07:00:00'
3770
'07-Jun-1900 02:00:00'
57
2
2955
'04-May-1900 03:00:00'
3072
'09-May-1900 00:00:00'
58
2
83
'04-Jan-1900 11:00:00'
3164
'12-May-1900 20:00:00'
59
2
391
'17-Jan-1900 07:00:00'
3559
'29-May-1900 07:00:00'
60
2
930
'08-Feb-1900 18:00:00'
1816
'17-Mar-1900 16:00:00'
61
2
4364
'01-Jul-1900 20:00:00'
948
'09-Feb-1900 12:00:00'
62
2
3390
'22-May-1900 06:00:00'
574
'24-Jan-1900 22:00:00'
63
2
1600
'08-Mar-1900 16:00:00'
4341
'30-Jun-1900 21:00:00'
64
2
2213
'03-Apr-1900 05:00:00'
2561
'17-Apr-1900 17:00:00'
65
2
2253
'04-Apr-1900 21:00:00'
1010
'12-Feb-1900 02:00:00'
66
2
4
'01-Jan-1900 04:00:00'
678
'29-Jan-1900 06:00:00'
67
2
752
'01-Feb-1900 08:00:00'
579
'25-Jan-1900 03:00:00'
68
2
3008
'06-May-1900 08:00:00'
3060
'08-May-1900 12:00:00'
69
2
2232
'04-Apr-1900 00:00:00'
1287
'23-Feb-1900 15:00:00'
70
2
3665
'02-Jun-1900 17:00:00'
2890
'01-May-1900 10:00:00'
71
2
353
'15-Jan-1900 17:00:00'
322
'14-Jan-1900 10:00:00'
72
2
349
'15-Jan-1900 13:00:00'
291
'13-Jan-1900 03:00:00'
73
2
754
'01-Feb-1900 10:00:00'
1455
'02-Mar-1900 15:00:00'
74
2
3403
'22-May-1900 19:00:00'
3835
'09-Jun-1900 19:00:00'
75
2
4043
'18-Jun-1900 11:00:00'
2888
'01-May-1900 08:00:00'
76
2
1951
'23-Mar-1900 07:00:00'
997
'11-Feb-1900 13:00:00'
77
2
2325
'07-Apr-1900 21:00:00'
1056
'14-Feb-1900 00:00:00'
78
2
4085
'20-Jun-1900 05:00:00'
2241
'04-Apr-1900 09:00:00'
79
2
1663
'11-Mar-1900 07:00:00'
36
'02-Jan-1900 12:00:00'
80
2
2357
'09-Apr-1900 05:00:00'
1900
'21-Mar-1900 04:00:00'
81
2
601
'26-Jan-1900 01:00:00'
4392
'03-Jul-1900 00:00:00'
82
2
2627
'20-Apr-1900 11:00:00'
4327
'30-Jun-1900 07:00:00'
83
2
4379
'02-Jul-1900 11:00:00'
1035
'13-Feb-1900 03:00:00'
84
2
4364
'01-Jul-1900 20:00:00'
3722
'05-Jun-1900 02:00:00'
85
2
2004
'25-Mar-1900 12:00:00'
1002
'11-Feb-1900 18:00:00'
86
2
796
'03-Feb-1900 04:00:00'
4115
'21-Jun-1900 11:00:00'
87
2
1170
'18-Feb-1900 18:00:00'
1126
'16-Feb-1900 22:00:00'
88
2
3736
'05-Jun-1900 16:00:00'
3215
'14-May-1900 23:00:00'
89
2
1061
'14-Feb-1900 05:00:00'
2344
'08-Apr-1900 16:00:00'
90
2
4348
'01-Jul-1900 04:00:00'
1617
'09-Mar-1900 09:00:00'
91
2
4109
'21-Jun-1900 05:00:00'
159
'07-Jan-1900 15:00:00'
92
2
1545
'06-Mar-1900 09:00:00'
984
'11-Feb-1900 00:00:00'
93
2
60
'03-Jan-1900 12:00:00'
1933
'22-Mar-1900 13:00:00'
94
2
112
'05-Jan-1900 16:00:00'
1773
'15-Mar-1900 21:00:00'
95
2
4249
'27-Jun-1900 01:00:00'
2342
'08-Apr-1900 14:00:00'
96
2
1423
'01-Mar-1900 07:00:00'
3441
'24-May-1900 09:00:00'
97
2
126
'06-Jan-1900 06:00:00'
3616
'31-May-1900 16:00:00'
98
2
816
'04-Feb-1900 00:00:00'
2982
'05-May-1900 06:00:00'
99
2
233
'10-Jan-1900 17:00:00'
1694
'12-Mar-1900 14:00:00'
100
2
1229
'21-Feb-1900 05:00:00'
1659
'11-Mar-1900 03:00:00'
101
2
961
'10-Feb-1900 01:00:00'
632
'27-Jan-1900 08:00:00'
102
2
2851
'29-Apr-1900 19:00:00'
49
'03-Jan-1900 01:00:00'
103
2
4145
'22-Jun-1900 17:00:00'
3170
'13-May-1900 02:00:00'
104
2
207
'09-Jan-1900 15:00:00'
1227
'21-Feb-1900 03:00:00'
105
2
244
'11-Jan-1900 04:00:00'
1699
'12-Mar-1900 19:00:00'
106
2
1323
'25-Feb-1900 03:00:00'
20
'01-Jan-1900 20:00:00'
107
2
4174
'23-Jun-1900 22:00:00'
1779
'16-Mar-1900 03:00:00'
108
2
1683
'12-Mar-1900 03:00:00'
513
'22-Jan-1900 09:00:00'
109
2
4392
'03-Jul-1900 00:00:00'
3286
'17-May-1900 22:00:00'
110
2
565
'24-Jan-1900 13:00:00'
3653
'02-Jun-1900 05:00:00'
111
2
1783
'16-Mar-1900 07:00:00'
3676
'03-Jun-1900 04:00:00'
112
2
3555
'29-May-1900 03:00:00'
4040
'18-Jun-1900 08:00:00'
113
2
4309
'29-Jun-1900 13:00:00'
972
'10-Feb-1900 12:00:00'
114
2
2212
'03-Apr-1900 04:00:00'
2447
'12-Apr-1900 23:00:00'
115
2
4368
'02-Jul-1900 00:00:00'
2922
'02-May-1900 18:00:00'
116
2
2827
'28-Apr-1900 19:00:00'
0
'01-Jan-1900 00:00:00'
117
2
518
'22-Jan-1900 14:00:00'
3094
'09-May-1900 22:00:00'
118
2
4391
'02-Jul-1900 23:00:00'
2896
'01-May-1900 16:00:00'
119
2
3059
'08-May-1900 11:00:00'
1579
'07-Mar-1900 19:00:00'
120
2
2347
'08-Apr-1900 19:00:00'
4178
'24-Jun-1900 02:00:00'
121
2
873
'06-Feb-1900 09:00:00'
1476
'03-Mar-1900 12:00:00'
122
2
441
'19-Jan-1900 09:00:00'
25
'02-Jan-1900 01:00:00'
123
2
1876
'20-Mar-1900 04:00:00'
4356
'01-Jul-1900 12:00:00'
124
2
2066
'28-Mar-1900 02:00:00'
3851
'10-Jun-1900 11:00:00'
125
2
2572
'18-Apr-1900 04:00:00'
364
'16-Jan-1900 04:00:00'
126
2
2472
'14-Apr-1900 00:00:00'
2006
'25-Mar-1900 14:00:00'
127
2
4392
'03-Jul-1900 00:00:00'
3041
'07-May-1900 17:00:00'
128
2
1953
'23-Mar-1900 09:00:00'
3566
'29-May-1900 14:00:00'
129
2
967
'10-Feb-1900 07:00:00'
1514
'05-Mar-1900 02:00:00'
130
2
2021
'26-Mar-1900 05:00:00'
4216
'25-Jun-1900 16:00:00'
131
2
45
'02-Jan-1900 21:00:00'
2987
'05-May-1900 11:00:00'
132
2
1525
'05-Mar-1900 13:00:00'
2502
'15-Apr-1900 06:00:00'
133
2
1909
'21-Mar-1900 13:00:00'
757
'01-Feb-1900 13:00:00'
134
2
41
'02-Jan-1900 17:00:00'
3321
'19-May-1900 09:00:00'
135
2
4392
'03-Jul-1900 00:00:00'
1333
'25-Feb-1900 13:00:00'
136
2
3084
'09-May-1900 12:00:00'
4344
'01-Jul-1900 00:00:00'
137
2
570
'24-Jan-1900 18:00:00'
2328
'08-Apr-1900 00:00:00'
138
2
3065
'08-May-1900 17:00:00'
1103
'15-Feb-1900 23:00:00'
139
2
226
'10-Jan-1900 10:00:00'
0
'01-Jan-1900 00:00:00'
140
2
4388
'02-Jul-1900 20:00:00'
573
'24-Jan-1900 21:00:00'
141
2
1241
'21-Feb-1900 17:00:00'
4392
'03-Jul-1900 00:00:00'
142
2
4346
'01-Jul-1900 02:00:00'
1731
'14-Mar-1900 03:00:00'
143
2
345
'15-Jan-1900 09:00:00'
516
'22-Jan-1900 12:00:00'
144
2
3239
'15-May-1900 23:00:00'
3204
'14-May-1900 12:00:00'
145
2
2356
'09-Apr-1900 04:00:00'
4235
'26-Jun-1900 11:00:00'
146
2
418
'18-Jan-1900 10:00:00'
1320
'25-Feb-1900 00:00:00'
147
2
2376
'10-Apr-1900 00:00:00'
3596
'30-May-1900 20:00:00'
148
2
1
'01-Jan-1900 01:00:00'
4327
'30-Jun-1900 07:00:00'
149
2
3861
'10-Jun-1900 21:00:00'
29
'02-Jan-1900 05:00:00'
150
2
0
'01-Jan-1900 00:00:00'
3613
'31-May-1900 13:00:00'
151
2
2
'01-Jan-1900 02:00:00'
4354
'01-Jul-1900 10:00:00'
152
2
2288
'06-Apr-1900 08:00:00'
1621
'09-Mar-1900 13:00:00'
153
2
4183
'24-Jun-1900 07:00:00'
127
'06-Jan-1900 07:00:00'
154
2
304
'13-Jan-1900 16:00:00'
4097
'20-Jun-1900 17:00:00'
155
2
3052
'08-May-1900 04:00:00'
2257
'05-Apr-1900 01:00:00'
156
2
626
'27-Jan-1900 02:00:00'
0
'01-Jan-1900 00:00:00'
157
2
116
'05-Jan-1900 20:00:00'
1095
'15-Feb-1900 15:00:00'
158
2
524
'22-Jan-1900 20:00:00'
1380
'27-Feb-1900 12:00:00'
159
2
150
'07-Jan-1900 06:00:00'
2507
'15-Apr-1900 11:00:00'
160
2
1861
'19-Mar-1900 13:00:00'
3505
'27-May-1900 01:00:00'
161
2
1955
'23-Mar-1900 11:00:00'
416
'18-Jan-1900 08:00:00'
162
2
1868
'19-Mar-1900 20:00:00'
1881
'20-Mar-1900 09:00:00'
163
2
99
'05-Jan-1900 03:00:00'
1864
'19-Mar-1900 16:00:00'
164
2
4348
'01-Jul-1900 04:00:00'
2508
'15-Apr-1900 12:00:00'
165
2
2887
'01-May-1900 07:00:00'
3075
'09-May-1900 03:00:00'
166
2
4005
'16-Jun-1900 21:00:00'
72
'04-Jan-1900 00:00:00'
167
2
21
'01-Jan-1900 21:00:00'
2780
'26-Apr-1900 20:00:00'
168
2
495
'21-Jan-1900 15:00:00'
2387
'10-Apr-1900 11:00:00'
169
2
2936
'03-May-1900 08:00:00'
1972
'24-Mar-1900 04:00:00'
170
2
1171
'18-Feb-1900 19:00:00'
3668
'02-Jun-1900 20:00:00'
171
2
2984
'05-May-1900 08:00:00'
3336
'20-May-1900 00:00:00'
172
2
4377
'02-Jul-1900 09:00:00'
2398
'10-Apr-1900 22:00:00'
173
2
4386
'02-Jul-1900 18:00:00'
3395
'22-May-1900 11:00:00'
174
2
46
'02-Jan-1900 22:00:00'
1600
'08-Mar-1900 16:00:00'
175
2
3508
'27-May-1900 04:00:00'
4040
'18-Jun-1900 08:00:00'
176
2
120
'06-Jan-1900 00:00:00'
2137
'31-Mar-1900 01:00:00'
177
2
29
'02-Jan-1900 05:00:00'
2397
'10-Apr-1900 21:00:00'
178
2
3650
'02-Jun-1900 02:00:00'
3151
'12-May-1900 07:00:00'
179
2
746
'01-Feb-1900 02:00:00'
3975
'15-Jun-1900 15:00:00'
Table 7. Probabilistic impact assessment of maintenance
scheduling solutions considering wind farms.
Risk Indicator EENS (MWh/year)
PSO
Surrogate
17,306 [16,441; 18,171]
20,798 [19,758; 21,838]
499
Figure 7. Variations due to correlated wind speed assumptions.
Figure 8. Convergence of the surrogate algorithm.
Figure 9. Convergence of the PSO algorithm.
500
Figure 10. Generation capacity considering wind farms.
Figure 11. Maintenance scheduling schema.
4 CONCLUSIONS
This study introduced a revised methodology to
optimize the coordination of maintenance activities in
offshore wind farms within power systems. By
integrating multivariate stochastic wind simulations
into a maintenance scheduling framework, the
proposed model captures the spatial and temporal
dependencies inherent in offshore wind behavior, an
aspect oversimplified or overlooked in previous
results. Through extensive simulations based on IEEE-
RTS and a revised methodology drawn from previous
work, the results reveal that incorporating correlated
wind speed significantly affects the estimation of
Expected Energy Not Supplied (EENS). Specifically,
deviations greater than 6% in EENS were observed
under the same operational conditions when
comparing uncorrelated and highly correlated wind
scenarios. In addition, the new solution for
maintenance scheduling activities is 18% lower than
the previous solution under the new modeling
conditions. These findings confirm that the wind speed
correlation plays a critical role in system reliability and
should be taken into account in realistic maintenance
planning.
The use of copula-based modeling to simulate
correlated wind speeds, combined with heuristic
optimization algorithms (such as PSO and Surrogate
models), allowed for a nuanced and effective
scheduling of maintenance activities across dispersed
units. Among the heuristics tested, PSO achieved the
lowest EENS values, although further validation
through sensitivity analysis is needed to establish its
consistency across various scenarios. Certainly, this
modeling change introduced in this paper exposes
deviations (up to 6%) in the Expected Energy Not
Supplied (EENS) indicator under the same operational
conditions, purely due to changes in how wind speed
correlations are handled. Such deviations have a direct
impact on the allocation and timing of maintenance
activities, indicating that overlooking wind correlation
can lead to suboptimal or even misleading
maintenance strategies.
By incorporating realistic wind behavior and
spatial-temporal dependencies, the proposed
methodology enables more robust, data-driven, and
risk-aware planning of maintenance operations in
offshore environments.
In the case of directions for further research, we
plan to explore real operational data from actual
offshore wind farms to validate the evidence of
correlations in wind speed and introduce a dynamic
and real-time scheduling approach by using real-time
data assimilation and online adaptive optimization
algorithms to dynamically adjust maintenance
schedules.
501
ACKNOWLEDGMENT
The Polish Ministry of Science and Higher Education has
financially supported the work.
REFERENCES
[1] Namkyoung Lee, Joohyun Woo, Sungryul Kim. A deep
reinforcement learning ensemble for maintenance
scheduling in offshore wind farms, Applied Energy,
Volume 377, Part A, 2025, 124431, ISSN 0306-2619.
[2] Xiangyong Feng, Shunjiang Lin, Yutao Liang, Xin Lai,
Mingbo Liu. Coordinated risk-averse distributionally
robust optimization for maintenance and generation
schedules of offshore wind farm cluster, International
Journal of Electrical Power & Energy Systems, Volume
159, 2024, 109993, ISSN 0142-0615,
https://doi.org/10.1016/j.ijepes.2024.109993.
[3] Salgado Duarte Y, Szpytko J, Reliability-oriented twin
model for integrating offshore wind farm maintenance
activities, Eksploatacja i Niezawodnosc Maintenance
and Reliability 2025: 27(3)
http://doi.org/10.17531/ein/199355.
[4] Markus Steen, Tuukka Mäkitie, Jens Hanson, Håkon
Endresen Normann, Developing the industrial capacity
for energy transitions: Resource formation for offshore
wind in Europe, Environmental Innovation and Societal
Transitions, Volume 53, 2024, 100925, ISSN 2210-4224,
https://doi.org/10.1016/j.eist.2024.100925.
[5] Rui He, Zhigang Tian, Yifei Wang, Yinuo Chen, Ming J.
Zuo, Predictive maintenance for offshore wind farms
with incomplete and biased prognostic information,
Ocean Engineering, Volume 322, 2025, 120541, ISSN 0029-
8018, https://doi.org/10.1016/j.oceaneng.2025.120541.
[6] Jiaxuan Luo, Xiaofang Luo, Xiandong Ma, Yingfei Zan, Xu
Bai, An integrated condition-based opportunistic
maintenance framework for offshore wind farms,
Reliability Engineering & System Safety, Volume 256,
2025, 110701, ISSN 0951-8320,
https://doi.org/10.1016/j.ress.2024.110701.
[7] Shuya Zhong, Athanasios A. Pantelous, Mark Goh, Jian
Zhou, A reliability-and-cost-based fuzzy approach to
optimize preventive maintenance scheduling for offshore
wind farms, Mechanical Systems and Signal Processing,
Volume 124, 2019, Pages 643-663, ISSN 0888-3270,
https://doi.org/10.1016/j.ymssp.2019.02.012.
[8] Papadopoulos, P., Coit, D. W., & Aziz Ezzat, A. (2023).
STOCHOS: Stochastic opportunistic maintenance
scheduling for offshore wind farms. IISE Transactions,
56(1), 115.
https://doi.org/10.1080/24725854.2022.2152913.
[9] Qian Sun, Jinxing Che, Kun Hu, Wen Qin, Deterministic
and probabilistic wind speed forecasting using
decomposition methods: Accuracy and uncertainty,
Renewable Energy, Volume 243, 2025, 122515, ISSN 0960-
1481, https://doi.org/10.1016/j.renene.2025.122515.
[10] Xiaolin Ge, Quan Chen, Yang Fu, C.Y. Chung, Yang Mi,
Optimization of maintenance scheduling for offshore
wind turbines considering the wake effect of arbitrary
wind direction, Electric Power Systems Research, Volume
184, 2020, 106298, ISSN 0378-7796,
https://doi.org/10.1016/j.epsr.2020.106298.
[11] Christian Gundegjerde, Ina B. Halvorsen, Elin E.
Halvorsen-Weare, Lars Magnus Hvattum, Lars Magne
Nonås, A stochastic fleet size and mix model for
maintenance operations at offshore wind farms,
Transportation Research Part C: Emerging Technologies,
Volume 52, 2015, Pages 74-92, ISSN 0968-090X,
https://doi.org/10.1016/j.trc.2015.01.005.
[12] S. Astariz, J. Abanades, C. Perez-Collazo, G. Iglesias,
Improving wind farm accessibility for operation &
maintenance through a co-located wave farm: Influence
of layout and wave climate, Energy Conversion and
Management, Volume 95, 2015, Pages 229-241, ISSN 0196-
8904, https://doi.org/10.1016/j.enconman.2015.02.040.
[13] Jack C Kennedy, Daniel A Henderson, Kevin J Wilson,
Multilevel emulation for stochastic computer models
with application to large offshore wind farms, Journal of
the Royal Statistical Society Series C: Applied Statistics,
Volume 72, Issue 3, June 2023, Pages 608627,
https://doi.org/10.1093/jrsssc/qlad023.
[14] The MathWorks Inc. MATLAB Version: 24.1.0.2603908
(R2024a) Update 3, Natick, Massachusetts 2024. The
MathWorks Inc. https://www.mathworks.com.
[15] Grigg C. et al. The IEEE Reliability Test System-1996. A
report prepared by the Reliability Test System Task Force
of the Application of Probability Methods Subcommittee.
IEEE Transactions on Power Systems 1999; 14(3): 1010-
1020, https://doi: 10.1109/59.780914.
[16] Global Offshore Wind Report. (2024) https://wfo-
global.org/wp-content/uploads/2024/04/WFO-Report-
2024Q1.pdf