www.nature.

com/scientificreports

OPEN Novel methods for wind speeds

prediction across multiple locations
Oleg Gaidai1, Jingxiang Xu1*, Ping Yan1, Yihan Xing2, Yu Wu1 & Fuxi Zhang1

This article provides two unique methodologies that may be coupled to study the dependability of
multidimensional nonlinear dynamic systems. First, the structural reliability approach is well suited
for multidimensional environmental and structural reactions and is either measured or numerically
simulated over sufficient time, yielding lengthy ergodic time series. Second, a unique approach to
predicting extreme values has technical and environmental implications. In the event of measurable
environmental loads, it is also feasible to calculate the probability of system failure, as shown in this
research. In addition, traditional probability approaches for time series cannot cope effectively with
the system’s high dimensionality and cross-correlation across dimensions. It is common knowledge
that wind speeds represent a complex, nonlinear, multidimensional, and cross-correlated dynamic
environmental system that is always difficult to analyze. Additionally, global warming is a significant
element influencing ocean waves throughout time. This section aims to demonstrate the efficacy of
the previously mentioned technique by applying a novel method to the Norwegian offshore data set
for the greatest daily wind cast speeds in the vicinity of the Landvik wind station. This study aims
to evaluate the state-of-the-art approach for extracting essential information about the extreme
reaction from observed time histories. The approach provided in this research enables the simple and
efficient prediction of failure probability for the whole nonlinear multidimensional dynamic system.

Accurate estimations of wind resource and operational conditions are beneficial for the cost-effective growth of
the wind energy sector. Extreme value analysis is of relevance in a variety of domains since it deals with statistical
inference on extreme values. Environmental extremes (such as river flow, wind speeds, temperature, and rainfall)
and engineering are two of the key areas of concentration (e.g., structural reliability and strength of materials).
In general, utilizing traditional theoretical reliability methods to calculate the dependability of practical envi-
ronmental systems may be rather difficult. The latter is often brought on by a large number of system freedoms
and random variables that control the environment system. In theory, it is possible to estimate the dependability
of complex environmental systems in a straightforward manner by using either a sufficient number of measure-
ments or direct Monte Carlo simulations. For many complicated environmental systems, the experimental or
computational costs may be prohibitive. Authors have developed a novel reliability method for environmental
systems with the goal of lowering measurement or computational costs.

Methods
It is challenging to estimate structural system reliability using classic engineering ­methods1,2. The latter is usually
due to a large number of degrees of system freedom and random variables governing dynamic systems. The reli-
ability of complex structural systems can be straightforwardly estimated either by having enough measurements
or by direct numerical Monte Carlo s­ imulations3–11. However, experimental and computational often are unaf-
fordable for many complex engineering dynamic systems. Authors have introduced a novel reliability method
for structural systems aiming to reduce measurement or numerical computational costs.
Ocean waves are usually assumed to be a random ergodic process (stationary and homogenous). Consider
MDOF (multi-degree of freedom) offshore structure subjected to ergodic environmental loadings, for example,
originating from surrounding waves and wind. MDOF structural response vector process
R(t) = (X(t), Y (t), Z(t), . . .), that has been either measured or simulated over a long enough time interval (0, T).
Unidimensional response process maxima over the entire time span (0, T) denoted as XTmax = max X(t),
0≤t≤T
YTmax = max Y (t), ZTmax = max Z(t), . . ..
0≤t≤T 0≤t≤T
Let X1 , . . . , XNX represent local maxima of the time-dependent process X(t) at discrete monotonically rising
time moments t1X < · · · < tNXX in (0, T). Similar definitions may be provided for the additional MDOF response

1
Shanghai Engineering Research Center of Marine Renewable Energy, College of Engineering Science and
Technology, Shanghai Ocean University, Shanghai, China. 2University of Stavanger, Stavanger, Norway. *email:
jxxu@shou.edu.cn

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 1

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 1.  Illustration of how two exemplary processes, X and Y, are combined into a new synthetic vector R(t).

components Y (t), Z(t), . . . using Y1 , . . . , YNY ; Z1 , . . . , ZNZ,…. All MDOF R(t) components and their unidimen-
sional maxima are considered to be bigger than zero in this i­ nvestigation12.
The objective is to determine the MDOF system failure probability, also known as the exceedance probability
1 − P = Prob(XTmax > ηX ∪ YTmax > ηY ∪ ZTmax > ηZ ∪ . . .) (1)
(ηX ,ηY ,ηZ ,...)
  
with P = pXTmax ,YTmax ,ZTmax ,... XTmax , YTmax , ZTmax , . . . dXTmax dYNmax
Y
dZNmax
z
. . . being the non-exceedance
(0,0,0,...)
probability for critical unidimensional response values ηX , ηY , ηZ,…; with ∪ denoting logical unity operation;
and pXTmax ,YTmax ,ZTmax ,... denoting the global maxima joint probability density over the time period (0, T). In reality,
it is difficult to directly estimate the joint probability distribution due to its large dimensionality and the data
collection limits.
The moment in time when either X(t) exceeds ηX , or Y (t) exceeds ηY , or Z(t) exceeds ηZ , and so
on, the dynamic system is regarded as efficiently failed. Pre-defined failure levels ηX , ηY , ηZ ,…are indi-
vidual for every unidimensional response component of R(t). XNmax = max{Xj ; j = 1, . . . , NX } = XTmax ,
YNmax = max{Y ; j = 1, . . . , N } = Y max,Z max = max{Z ; j = 1, . . . , N X} = Z max,…. , see Fig. 1.
Y j Y T Nz j Z T
 X The nextX step is to sort time instants  unidimensional response component maxima
t1 < · · · < tNX ; t1 < · · · < tNY ; t1 < · · · < tNZ in non-decreasing temporal order into one single merged
Y Y Z Z
time series vector t1 ≤ · · · ≤ tN , with tN = max{tNXX , tNYY , tNZZ , . . .}, N = NX + NY + NZ + · · ·. With tj repre-
senting unidimensional response component maxima of one of the MDOF structural response components.
So having R(t), one must concurrently and continuously screen for each unidimensional response component
temporal maxima, noting its exceedances of MDOF limit vector (ηX , ηY , ηZ , . . .) in any of its unidimensional
response components X, Y , Z, . . . Each unidimensional−
→ response component maxima are then combined into
a single temporal non-decreasing vector R = (R1 , R2 , . . . , RN ) based on the merged time vector t1 ≤ · · · ≤ tN .
Finally, the unified MDOF limit vector (η1 , . . . , ηN ) is now introduced with each component ηj being either ηX ,
ηY or ηZ …, depending on which of X(t) or Y (t), or Z(t) … corresponds to the current unidimensional response
component maxima with the running index j , ­see12–16.
Scaling parameter 0 <  ≤ 1 is now added toartificially lower  limit values for all unidimensional response
components, namely the new MDOF limit vector ηX , ηY , ηz , .. . with ηX ≡ ·ηX , ≡ ·ηY , ηz ≡ ·ηZ , … is now
introduced, ­see17. The unified MDOF limit vector η1 , . . . , ηN  has each component η as either η , η or η ….
j X Y z
The latter specifies the probability distribution function P() as a function of artificial parameter  , note that
P ≡ P(1) from Eq. (1). The non-exceedance probability function P() may be expressed as follows

P() = Prob{RN ≤ ηN , . . . , R1 ≤ η1 }
 
 
= Prob RN ≤ ηN RN−1 ≤ ηN−1 , . . . , R1 ≤ η1 } · Prob{RN−1 ≤ ηN−1

, . . . , R1 ≤ η1 }
(2)
N

= Prob{Rj ≤ ηj |Rj−1 ≤ η1j−

, . . . , R1 ≤ η1 } · Prob(R1 ≤ η1 )
j=2

The cascade of approximations based on conditioning is described. In practice, dependencies between adja-
cent local maxima Rj are not insignificant; hence the following one step (defined as conditioning level k = 1)
memory approximation may be implemented

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 2

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Prob{Rj ≤ ηj |Rj−1 ≤ ηj−1


, . . . , R1 ≤ η1 } ≈ Prob{Rj ≤ ηj |Rj−1 ≤ ηj−1

(3)

for 2 ≤ j ≤ N (defined as conditioning level k = 2). Then Eq. (3) can be further developed as

Prob{Rj ≤ ηj |Rj−1 ≤ ηj−1


, . . . , R1 ≤ η1 } ≈ Prob{Rj ≤ ηj |Rj−1 ≤ ηj−1
 
, Rj−2 ≤ ηj−2 } (4)

with 3 ≤ j ≤ N (defined as level k = 3). Equation (4) presents subsequent refinements of local consecutive
maxima independence assumption. The latter reflects the statistical dependence between neighbouring local
maxima with ever-increasing
  accuracy. As the MDOF response process R(t) was assumed ergodic, the probability

pk () := Prob Rj > ηj Rj−1 ≤ ηj−1 j−k+1 ≤ ηj−k+1 } for j ≥ k is independent of j and entirely on the condi-
 ,R 

tioning level k
Pk () ≈ exp(−N · pk ()) , k ≥ 1. (5)
that Eq. (5) is similar to the mean up-crossing rate formula for the probability of ­exceedance4. There is gradual
convergence with respect to the conditioning parameter k
P = lim Pk (1); p() = lim pk () (6)
k→∞ k→∞

Note that Eq. (5) k = 1 case corresponds to the non-exceedance probability relationship with the mean up-
crossing rate function
 ∞
P() ≈ exp (−ν + () T); ν + () = ζ pRṘ (, ζ )dζ (7)
0

with ν + () being the mean up-crossing rate of the response level   for the above
 assembled non-dimensional
vector R(t) assembled from the scaled MDOF system response ηXX , ηYY , ηZZ , . . . . The mean up-crossing rate from
 
Rice’s formula given in Eq. (7) with pRṘ being the joint probability density for R, Ṙ with Ṙ being time derivative
18
R (t), ­see . Equation (7) relies on the Poisson assumption, that is, up-crossing events of high  levels (in this
′
paper it is  ≥ 1) can be assumed to be independent.
In the preceding, the assumption of ergodicity was used. For non-stationary cases, the illustration will be
done for offshore engineering and naval architecture areas of engineering applications. Given, for example, the
scattered diagram
 of m = 1, . . . , M environmental states, each short-term environmental state has the probability
qm, so that M m=1 qm = 1. Next, the long-term equation can be introduced
M

pk () ≡ pk (, m)qm (8)
m=1

with pk (, m) being the similar function as in Eq. (6), but corresponding to a short-term environmental state
with the number m.
Next, the Naess-Gaidai (NG) a­ pproach11 is presented briefly. The above-described pk () functions are often
tail-regular, i.e. for extreme
 values of approaching extreme level 1. More specifically, for  ≥ 0, the distribution
tail behaves like exp −(a + b)c + d with a, b, c, d being fitted constants for appropriate tail cut-on 0 value.
NG method assumes
 
pk () ≈ exp −(ak  + bk )ck + dk ,  ≥ 0 (9)
   
When ln ln pk () − dk is plotted against ln(ak  + bk ), almost linear tail behaviour is often seen. The loga-
rithmic level was optimized by minimizing the error function F with regard to its four parameters ak , bk , ck , pk , qk
 1
     2
F ak , bk , ck , pk , qk = ω() ln pk () − dk + (ak  + bk )ck d,  ≥ 0 (10)
0

with 1 being an appropriately selected distribution tail cut-off number for which the breadth of the confidence
interval (CI) is acceptable. The optimal parameters ak , bk , ck , pk , qk may be calculated using the sequential quad-
ratic programming (SQP) approach implemented in NAG Numerical L ­ ibrary19.

Novel extrapolation method. Accurate prediction of extreme values is a typical and demanding engi-
neering dependability challenge, particularly when data is insufficient. Consequently, the development of
unique, efficient, and precise extrapolation methods is of enormous practical value.
Consider a stationary stochastic process X(t), either simulated or measured over a given time interval
0 ≤ t ≤ T , and which is represented as a sum of two independent stationary processes X1 (t) and X2 (t), namely
X(t) = X1 (t) + X2 (t) (11)
Note that this study seeks to provide a generic technique applicable to predicting extreme values for a broad
range of loads and reactions for various vessels and offshore constructions. For the process of interest X(t),
marginal PDF (probability density function) pX may be obtained in two ways:

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 3

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

A. By directly extracting pXA from the provided data set, i.e. time series X(t),
B. By separately extracting PDFs from
 the process components X1 (t) and X2 (t), namely pX1 and pX2, then using
convolution pXB = conv pX1 , pX2 .

Both pXA and pXB are approximations of the PDF goal, pXA is simpler to implement, however B) would yield a
more precise estimate of the desired PDF, pX . Convolution permits extrapolation of the directly derived empiri-
cal PDF, pXA , without supposing any particular extrapolation functional class, such as generalized extreme value
distributions (GEV), required to extrapolate distribution tail towards the design low probability level of interest.
The majority of current extrapolation techniques, which are frequently used in engineering practice, depending
on the assumption of certain extrapolation functional classes, e.g.3–9.
The two independent component representation provided by Eq. (11) is seldom accessible; consequently,
one may seek artificial methods to estimate, pX1 and pX2, or in the most straightforward situation, discover two
identically distributed process components, X1 (t) and X2 (t) with pX1 = pX2 . This work focuses on the latter
possibility, in which processes, X1 (t) and X2 (t) are dispersed equally. Given the directly estimated distribution,
pX , as in option A, the present study’s objective would be to identify the component distribution, pX1, such that
 
pX = conv pX1 , pX1 (12)
therefore, limiting the scope of this research to a deconvolution situation. To illustrate the latter concept of how
to estimate the unknown distribution, pX1, reliably and then enhance (or extrapolate) the supplied empirical
distribution, pX .

Discrete convolution. This section provides a concise overview of commonly held knowledge surrounding
the discrete convolution of two vectors. The convolution of two vectors, u and v , illustrates the vector component
overlap as v glides over u. Convolution corresponds algebraically to multiplying polynomials whose coefficients
are the elements of u and v.
Let m = length(u) and n = length(v). Then w is the vector of length m + n − 1, whose k-th element is
m
   
w(k) = u j v k−j+1 (13)
j=1
  
The sum is over all the values of j that lead to legal subscripts for u j and v k − j + 1 , specifically
j = max(1, k + 1 − n) : 1 : min(k, m). When m = n, as will be the case in this paper, the latter gives
w(1) = u(1) · v(1)
w(2) = u(1) · v(2) + u(2) · v(1)
w(3) = u(1) · v(3) + u(2) · v(2) + u(3) · v(1)
··· (14)
w(n) = u(1) · v(n) + u(2) · v(n − 1) + · · · + u(n) · v(1)
···
w(2n − 1) = u(n) · v(n)
Having obtained u = v = (u(1), . . . , u(n)) , one may progressively acquire w -components
w(n + 1), . . . , w(2n − 1), as the index ranges from n + 1 to 2n − 1. Clearly, the latter would extend vector w
into a support domain that is twice as long as the original distribution support domain, i.e. doubling the pX .
distribution support length, (2n − 1) · �x ≈ 2n · �x = 2XL, compared to the original distribution support length
n · x = XL , where x is the constant length of each discrete distribution bin. In other words, convolution may
convect distribution tail attributes farther "downstream" or deeper inside the tail.
In Eq. (14), values of u and v outside the length of the vector are constrained to be zero; however, this is not
the case in practice. Consequently, later in this research, vectors u and v will be extrapolated linearly on the
logarithmic scale; see the next section, in which u the probability distribution function will represent u and v
by, fX1, and w by fX .
Note that w = (w(1), . . . , w(n)) is a discrete representation of the empirical target distribution, pX from
“Results” section, with n denoting the length of distribution support [0, XL ]. For the sake of simplicity, this work
is confined to the situation of one-sided positive valued random variables, i.e. X ≥ 0.
In the remainder of this study, only deconvolution scenarios, i.e. u = v in Eq. (14), will be examined. Accord-
ing to Eq. (12), the distributions, pX and, pX1 will correspond to the vectors w and u,, respectively.
From Eq. (15) is evident that given w = (w(1), . . . , w(n)) one can successively
√ locate unknown components
u = v = (u(1), . . . , u(n)), beginning with the first component u(1) = w(1) , then the second u(2) = 2u(1) w(2)
,
and so on until u(n).
As will be addressed further in this work, the authors propose simple linear extrapolation of self-deconvoluted
vector (u(1), . . . , u(n)) towards (u(n + 1), . . . , u(2n − 1)). i\In other words pX1 will have its tail linearly extrapo-
lated in the range (XL , 2XL ). Note that pX1 can be called the deconvoluted distribution, that in discrete form is
represented by the estimated vector u. Using Eq. (13) the original vector w will be clearly extended and extrapo-
lated into a support domain that is twice longer than the original distribution support domain, i.e. doubling
the pX distribution support length (2n − 1) · �x ≈ 2n · �x = 2XL , relative to the original distribution support
length n · x = XL . To smooth out the tail of the original distribution, pX , the authors conducted, pX tail inter-
polation, since the tail of the CDF distribution, is often rather regular for large tail values x. In particular, the

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 4

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

NG extrapolation approach has been used; see “Methods” section. In addition, various nonlinear extrapolation
techniques may be readily included in the suggested method; however, this would involve certain assumptions
and biases.

Results
This section provides numerical findings based on the deconvolution extrapolation approach presented. As stated
in the Introduction, the deconvolution extrapolation method does not presuppose a particular functional class
to extrapolate the distribution tail.

Numerical results for exceedance probability distribution tail. Since in the majority of reliability
analysis engineering applications, it is more important to estimate the probability of exceedance, i.e. 1CDF where
CDF stands for cumulative density function than the marginal PDF, fX will represent the probability of exceed-
ance 1—CDF in this paper, analogous to the marginal probability density function PDF, pX in “Results” section.
However, the suggested strategy may be applicable to any monotonically declining concave or convex function
tail that is sufficiently regular.
The "shorter" version of the original data set was extrapolated and compared to predictions based on the
"longer" version to verify the recommended extrapolation strategy. Therefore, this study aims to demonstrate
the efficacy of the recommended extrapolation approach by at least a few orders of magnitude.
Based on the preceding reasoning, one may implement an iterative approach, whereas in the marginal PDF,
one can utilize 1-CDF and then construct a new artificially smoother CDF by integration. The latter may substan-
tially improve the extrapolation process if distribution tail anomalies are present owing to a lack of underlying
data.
Next, the process of discrete convolution, or rather de-convolution (since the objective was to determine the
deconvoluted 1—CDF distribution, fX1, given the empirical distribution, fX is based on the sequential solution of
Eq. (4). Since the deconvoluted values u = (u(1), . . . , u(n)) normally follow a monotonically diminishing trend
(the same was expected for the empirical parent distribution fX , it seems that certain final values of the result-
ant vector u , say, (u(n − L), . . . , u(n)) for some L < n. The latter is an unacceptable numerical inaccuracy since
only positive numbers may represent distributions. To address this numerical difficulty, the following scaling
method has been i­ mplemented12,16. As the pivot value, the lowest positive value fL of the specified distribution
tail fX is used. The scaling is thus only a linear transition along the vertical y-axis of the logarithmic decimal scale.
      
gX = µ log10 fX − log10 fL + log10 fL (15)
with gX (x) being scaled log10 version of the empirical base distribution fX , with the reference level fL being
intact. The scaling coefficient µ is conveniently chosen to avoid the occurrence of negative components in the
resulting fX1. For both numerical examples studied in this paper, µ = 1/3 served that purpose well. Then when
fX1 was found, and back convolution  fX = conv fX1 , fX1 as in Eq. (13) was done, the inverse scaling with µ−1
was performed to restore the original scale, with fX being extrapolated version of fX .

Synthetic example. This subsection studies the 3.65-day maximum wind speed process X(t) simulated
during the given time interval [0, T]. The underlying normalized non-dimensional stochastic processes U(t) has
been modelled as a stationary Gaussian process with zero mean value and a standard deviation equal to one.
Therefore it was assumed that the U(t) mean zero up-crossing rates satisfies the equality νU+ (0) = 103 /T , with
T = 1 year, the latter assumption is common in offshore wind ­engineering20.
For this numerical example, the «shorter» data record had been chosen to contain 104 data points, which is
equivalent to 100 years record, since wind speed maxima process X(t) has 365/3.65=102 data points per year;
the full (the latter in this study referred to as «longer») data set had been chosen to contain 106 data points (thus
100 times longer than the «shorter» data record) are equivalent to 10,000 years.
The underlying
 wind
 2speed
 process U(t) results in 3.65 days maximum analytical CDF distribution
FX3d (x) = exp −qexp − x2 for the wind speed maxima process X(t), with q = 10. Note that under the above
assumptions, the non-dimensional X -reference value x 100yr = 4.80, i.e. the wind speed with a return period of
100 years.
Note that artificial horizontal axis shift x → x − xshift with xshift = 1.5 has been done in order to operate only
with the distribution tail; in other words xshift = x0 from Eq. (7) has been chosen as the cut-on tail marker. The
latter makes an analytical estimate for the x 100yr = 4.80 − xshift = 3.3.
Figure 2 left presents synthetic wind speed data scaled results, Fig. 2 right presents the final unscaled results
of the proposed in this paper technique, namely the «shorter» decimal log scale fX tail, extrapolated by decon-
volution, along with «longer» data distribution tail and analytic distribution.
Finally, 100 independent numerical Monte Carlo (MC) runs identical to the one described above have been
performed, i.e. both «shorter» and «longer» wind speed maxima data records were generated, with a subsequent
comparison between the analytical 10,000 return period and the one estimated by the proposed deconvolution
extrapolation. In addition, The Naess-Gaidai (NG) extrapolation technique has been performed for validation
purposes. Validation using the above two methods suggested deconvolution extrapolation and taken as the
benchmarking reference NG extrapolation technique has shown similar accuracy (within acceptable 5% relative
deviation) in terms of 10,000 return period prediction.
Cross-validation or MC simulations are typically used to evaluate the performance of models in a robust
­manner1.

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 5

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 2.  Synthetic wind speed data. Left: Scaled f X 1 tail on the decimal log scale for the «shorter» data (cyan
squares) «longer» data (blue squares). Right: Unscaled «shorter» decimal log scale f X tail, raw (red squares) and
fitted (solid blue line, along with «longer» data (green line) and analytic (solid red line).

Figure 3.  Wind speed measurements locations according to Norwegian Meteorological I­ nstitute21, a blue circle
indicates the area of interest, https://​sekli​ma.​met.​no/.

Environmental example. It is generally known that the dynamics of wind speed is a highly nonlinear,
multidimensional, and cross-correlated dynamic system that is always difficult to analyze. Moreover, the system
dependability approach is crucial for buildings working in any particular region of interest and encountering
genuine and sometimes harsh conditions. This section aims to demonstrate the efficacy of the previously men-
tioned technique by applying a novel method to the Norwegian maximum daily wind cast speeds data set in the
vicinity of the Landvik wind station in southern Norway.
As five wind measurement sites, namely Landvik, Kjevik, Lindesnes Fyr, Lista Fyr, and Oksy Fyr, were selected
for this research, their respective measured daily peak daily wind cast speeds were established as five environmen-
tal system components (dimensions) X, Y , Z, . . . therefore illustrating a five-dimensional (5D) environmental
system.
During the 2010–2020 observation period, the extreme unidimensional response values for each of the five
wind measurement sites were determined by doubling the maximum daily highest daily wind cast speed.
Figure 3 presents wind measurement locations according to the Norwegian Meteorological I­ nstitute21; the
blue circle indicates the area of interest. Daily largest highest daily wind cast speed at the Landvik location during
the years 2010 – 2020 is shown in Fig. 4.

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 6

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 4.  Daily largest highest daily wind cast speed at Landvik location during the years 2010 – 2020,
Norwegian Meteorological ­Institute21.

Figure 5.  Left: unscaled raw daily−

largest highest daily wind cast speed data [m/sec], Right: scaled non-
→
dimensional assembled 5D vector R .

In order to unify all five measured time series X, Y , Z ,... the following scaling was performed:
X Y Z
X→ ,Y → ,Z → ,... (16)
ηX ηY ηZ
making all five responses non-dimensional and having the same failure limit equal to 1. Next, all local maxima
from five → measured time series were merged into one single time series by keeping them
− → in time non-decreasing
−
order: R = (max{X1 , Y1 , Z1 , . . .}, . . . , max{XN , YN , ZN , . . .}) with the whole vector R being sorted according
to non-decreasing times of occurrence of these local maxima. −
→
Figure 5 presents an example of a non-dimensional assembled vector R , which is comprised of assembled
local maxima of raw daily greatest highest daily wind cast speed data. → The failure probability distribution tail was
−
extrapolated towards a return time of 100 years. Note that vector, R has no physical significance on its own, since
it is composed of wholly distinct response components. Index j is only an index of local maxima encountered
in a non-decreasing time sequence. Index j is only an index of local maxima encountered in a non-decreasing
time sequence. The «shorter» data record was constructed by extracting every tenth data point from the «longer»
data record of wind speed.
Figure 6 on the left shows the «shorter» data record, fX1 tail, derived by deconvolution as in Eq. (13), and then
linearly extrapolated in the terminal tail section to encompass the X1 range corresponding to the «longer» data
record. Figure 6 on the right displays the unscaled final outcomes of the approach suggested in this research,
namely the «shorter» decimal log scale, fX tail, extrapolated by deconvolution, together with the «longer» data
distribution tail and NG extrapolation.
It is seen that NG prediction matches the deconvolution method pretty well. Figure 6 on the right demon-
strates that the suggested technique works pretty well, is based on the "shorter" data set and produces distributions
that are relatively similar to those derived from the "longer" data set. Figure 6 on the right illustrates extrapolation
according to Eq. (9) as well as a unique deconvolution approach from Sect. 0 towards a failure state with a 100-
year return time, which is 1 and a bit beyond;  = 0.05 was chosen as the cut-on value. The dashed lines represent
the extended 95% confidence interval according to Eq. (10). According to Eq. (5), p() is exactly proportional to
the goal failure probability 1—P. (1). In accordance with Eq. (5), the system failure probability 1 − P ≈− →1 − P k (1)
may be computed. Note that in Eq. (5), N is the total number of local maxima in the response vector R . Due to
convergence with respect to k (see Eq. 6), it was determined that k = 5 was an adequate conditioning value (6).
Note that, although innovative, the above-described technique has the evident benefit of using existing meas-
ured data sets rather effectively, owing to its capacity to handle system multidimensionality and conduct correct
extrapolation using relatively small data sets.

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 7

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 6.  Wind speed data. Left: scaled f X 1 tail on the decimal log scale for the «shorter» data (cyan),
linearly extrapolated (dark blue). Right: unscaled raw «shorter» data (green) f X tail on the decimal log scale,
extrapolated by the deconvolution method (dark blue), along with «longer» raw data (red) and NG extrapolation
(cyan). Star indicates 100 years return period prediction.

Conclusions
Classical dependability approaches for time series cannot properly handle systems with high dimensionality and
cross-correlation between various system responses. The presented approach’s primary advantage is its ability to
examine the dependability of high-dimensional nonlinear dynamic systems.
This article examined the Norwegian greatest daily wind cast speeds obtained at two distinct wind speed
measurement locations near the Landvik wind station from 2012 to 2022.
The techniques presented in this study have been previously confirmed by application to a variety of simu-
lation models, albeit only for one-dimensional system responses, and highly accurate predictions have been
achieved in general. This research attempted to establish a multidimensional dependability approach that was
simple, robust, and applicable to a variety of applications.
It is possible to analyze both observed and numerically generated time series responses. It is shown that the
suggested strategy generated an acceptable confidence interval. Consequently, the proposed technique may
become useful for various nonlinear dynamic system dependability investigations.
In the case of Monte Carlo-type numerical simulation, unlike existing dependability approaches, the novel
method does not need to restart numerical simulation each time the system fails. It is possible to analyze both
observed and numerically generated time series responses.
Finally, the recommended technique applies to a vast array of technical applications. The offered example of
naval architecture does not restrict the scope of application for the proposed technique.

Data availability
Meteorological data used in this paper is publicly available at the Norwegian Meteorological Institute, https://​
sekli​ma.​met.​no/.

Received: 27 July 2022; Accepted: 9 November 2022

References
1. Bokde, N., Yaseen, Z. & Andersen, G. ForecastTB—An R package as a test-bench for time series forecasting—application of wind
speed and solar radiation modeling. Energies 13(10), 2578. https://​doi.​org/​10.​3390/​en131​02578 (2020).
2. Ditlevsen, O. & Madsen, H. O. Structural Reliability Methods (Wiley, 1996).
3. Gaidai, O. et al. Offshore renewable energy site correlated wind-wave statistics. Probab. Eng. Mech. https://​doi.​org/​10.​1016/j.​probe​
ngmech.​2022.​103207 (2022).
4. Gaidai, O., Xing, Y. & Xu, X. COVID-19 epidemic forecast in USA East coast by novel reliability approach. Res. Sq. https://​doi.​
org/​10.​21203/​rs.3.​rs-​15738​62/​v1 (2022).
5. Gaidai, O., Storhaug, G. & Naess, A. Extreme large cargo ship panel stresses by bivariate ACER2D method. Ocean Eng. 123, 432–439
(2016).
6. Gaidai, O. et al. Bivariate statistics of wind farm support vessel motions while docking. Ships Offshore Struct. 16(2), 135–143 (2020).
7. Gaidai, O., Storhaug, G. & Naess, A. Statistics of extreme hydro elastic response for large ships. Mar. Struct. 61, 142–154 (2018).
8. Gaidai, O., Storhaug, G. & Naess, A. Extreme value statistics of whipping response for large ships. In PRADS Proceedings. Practical
Design of Ships and Other Floating Structures, vol 2, 1210–1221 (2010).
9. Gaidai, O., Storhaug, G. & Naess, A. Extreme value statistics of ship rolling. In PRADS Proceedings. Practical Design of Ships and
Other Floating Structures, vol 2, 457–466 (2010).
10. Gao, H., Gaidai, O., Naess, A., Storhaug, G. & Xu, X. Improving container ship panel stress prediction, based on another highly
correlated panel stress measurement. Mar. Struct. 64, 138–145 (2018).
11. Naess, A. & Gaidai, O. Estimation of extreme values from sampled time series. Struct. Saf. 31(4), 325–334 (2009).

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 8

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol:.(1234567890)
www.nature.com/scientificreports/

12. Gaidai, O., Yan, P. & Xing, Y. A novel method for prediction of extreme wind speeds across parts of Southern Norway. Front.
Environ. Sci. https://​doi.​org/​10.​3389/​fenvs.​2022.​997216 (2022).
13. Sun, J. et al. Extreme riser experimental loads caused by sea currents in the Gulf of Eilat. Probab. Eng. Mech. https://​doi.​org/​10.​
1016/j.​probe​ngmech.​2022.​103243 (2022).
14. Xing, Y., Gaidai, O., Ma, Y., Naess, A. & Wang, F. A novel design approach for estimation of extreme responses of a subsea shuttle
tanker hovering in ocean current considering aft thruster failure. Appl. Ocean Res. https://​doi.​org/​10.​1016/j.​apor.​2022.​103179
(2022).
15. Xu, X. et al. Bivariate statistics of floating offshore wind turbine dynamic response under operational conditions. Ocean Eng. https://​
doi.​org/​10.​1016/j.​ocean​eng.​2022.​111657 (2022).
16. Gaidai, O., Yan, P. & Xing, Y. Prediction of extreme cargo ship panel stresses by using deconvolution. Front. Mech. Eng. https://​
doi.​org/​10.​3389/​fmech.​2022.​992177 (2022).
17. Naess, A., Leira, B. J. & Batsevych, O. Reliability analysis of large structural systems. Probab. Eng. Mech. 28, 164–168 (2012).
18. Rice, S. O. Mathematical analysis of random noise. Bell Syst. Tech. J. 23, 282–332 (1944).
19. Numerical Algorithms Group. NAG Toolbox for Matlab (NAG Ltd, 2010).
20. Naess, A. & Moan, T. Stochastic Dynamics of Marine Structures (Cambridge University Press, 2013).
21. Norwegian Meteorological Institute https://​sekli​ma.​met.​no/.

Author contributions
All authors contributed equally.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to J.X.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.

© The Author(s) 2022

Scientific Reports | (2022) 12:19614 | https://doi.org/10.1038/s41598-022-24061-4 9

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Vol.:(0123456789)
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com