Ocean Engineering 108 (2015) 636–647

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Uncertainty in forecasted environmental conditions for reliability

analyses of marine operations
Asle Natskår a,n, Torgeir Moan a, Per Ø. Alvær b
a
Centre for Ships and Ocean Structures (CeSOS), Norwegian University of Science and Technology (NTNU), Trondheim, Norway
b
DNV GL Marine Operations, VMO, Norway

art ic l e i nf o a b s t r a c t

Article history: Marine operations, e.g., the sea transport of heavy objects and the installation of offshore units and
Received 3 September 2014 equipment, need to be planned and executed with proper consideration for environmental conditions
Accepted 24 August 2015 and operational limits with respect to vessel motions and structural loads. Marine operations with a
Available online 19 September 2015
limited duration, usually less than 72 h, are typically designed as weather-restricted operations. The
Keywords: environmental design criteria are thus predefined, and the actual weather conditions are confirmed by
Marine operations weather forecasts issued immediately prior to the start of such an operation. Marine operations of longer
Weather forecast duration are typically designed as weather-unrestricted operations, and the environmental conditions
Weather criteria are calculated based on long-term statistics, possibly depending on the season. More detailed informa-
Forecast uncertainty
tion about uncertainties in weather forecasts could increase the feasible duration of weather-restricted
operations. The uncertainty inherent in weather forecasts, notably that in the significant wave height, is
studied. Further, a method to assess the reliability of weather forecasts is described. Data from the
Norwegian Sea are used to quantify the uncertainty in forecasted data. The probability of exceeding the
design criteria used in the planning of a weather-unrestricted marine operation can be estimated based
on forecast statistics. The corresponding uncertainty can be incorporated into structural reliability ana-
lyses.
& 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction of such a transported object are large, involving economic loss and
possibly pollution of the environment. Most likely, there will also
The work presented in this paper is part of a research project be delays in the project, and some loss of reputation may be suf-
regarding the level of reliability inherent in marine operations. The fered by the companies involved. It is therefore necessary to
uncertainty in the environmental conditions, and hence in the quantify the uncertainties inherent in such an operation. The
wave- and wind-induced load effects, that are considered in the environmental conditions are important input for the planning of
planning of an operation is important with respect to the overall marine operations, particularly with regard to the motion analysis
reliability level. The scope of this paper encompasses the study of of floating vessels. Hence, the uncertainty in the environmental
methods to account for the uncertainty inherent in weather conditions and how it is accounted for in the planning/design
forecasts for marine operations. This is of interest, e.g., with regard exert a considerable effect on the safety level of such an operation.
to structural design for sea fastening (i.e., the design of structures Marine operations can be designed in accordance with several
to secure a transported object to the transport vessel), but the Standards and Guidelines. These operations are generally defined
as either weather-restricted or weather-unrestricted operations,
approach is more general. The marine operations considered
depending on their duration. For weather-restricted operations
herein are specially planned, non-routine operations of limited
planned in accordance with DNV (2011) or GL Noble Denton
duration related to the load transfer, transport and installation of
(2013), the uncertainty inherent in the weather forecasts of the
objects, typically in the offshore oil and gas industry. The need for
significant wave heights and wind speeds is accounted for by a so-
special planning may arise because the transported object is large
called α-factor; see Eq. (2). For weather-unrestricted operations,
and/or heavy or has a high economic value or a long replacement
the weather criteria cannot be based on forecasts but instead must
time. Therefore, the consequences of severe damage to or total loss be based on long-term statistical data on the environmental
conditions.
n
Corresponding author. The uncertainty in the forecasting of significant wave heights is
E-mail address: [email protected] (A. Natskår). quantified by comparing the forecasted wave heights with the

http://dx.doi.org/10.1016/j.oceaneng.2015.08.034
0029-8018/& 2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 637

actual waves at the location. Instead of observed/measured sig- extreme weather conditions over a longer period, the operation
nificant wave heights, hindcast data are used. The uncertainty is reference period may be increased accordingly. By contrast, in
described by estimating the mean values and standard deviations areas and/or seasons in which the corresponding reliable weather
of the difference and ratio between the hindcasted and forecasted forecasts are not considered realistic, a shorter limit is to be
wave heights. Data from the Norwegian Sea are applied to quantify applied.
the uncertainty in the forecasted data. If an operation is weather restricted, then the design environ-
The objective is to incorporate the uncertainty resulting from mental criteria are defined in an early phase of the project.
weather forecasts into reliability analyses for marine operations. Weather-restricted operations are beneficial because the owner, or
For most marine operations, the environmental loads govern the his representative, may define the necessary environmental cri-
planning and structural design, and hence, the uncertainty in teria (with the understanding that more strict environmental cri-
forecasted environmental conditions is important input for these teria may lead to more wait time before the operation can com-
analyses. The uncertainty in the forecasted significant wave height mence). The operation may commence when the weather fore-
is studied in this paper. The intent is to address reliability analyses casts indicate acceptable environmental conditions. The uncer-
in a separate paper. tainty in the weather forecasts and how to include this uncertainty
in the planning of the operation thus become key issues.
To account for the uncertainty in weather forecasts, the
2. Planning of marine operations operational environmental limits must be less than those con-
sidered in the design. According to DNV (2011) and GL Noble
2.1. Design standards for marine operations Denton (2013), the operational limit on the significant wave height
can be expressed as
Marine operations must be designed in accordance with certain
Hs, oper = αHs, design (2)
standards/guidelines. We are aware of two such standards and two
such guidelines: where α is a parameter ( ≤ 1) that depends on both the duration of
the operation and the level of forecasting and/or monitoring. In

DNV-OS-H101, Marine Operations, General, DNV (2011). DNV (2011), α also depends on the significant wave height used in

GL Noble Denton (2013), General Guidelines for Marine Projects.
the design. The parameters for the base case, with one weather

ISO 19901-6 (2009) Petroleum and natural gas industries.
forecast available, are shown in Table 1. The α-factor can be
Specific requirements for offshore structures. Part 6: Marine
increased if the wave height at the site of the operation is mon-
Operations.
itored and if there is a meteorologist on site (because the presence

London Offshore Consultants Limited, Guidelines for Marine
of a meteorologist will increase the confidence in weather fore-
Operations, LOC (1997).
casts at that location). In the DNV method, α accounts for the
uncertainty in the weather forecast based on the planned duration
A key parameter for a marine operation is the duration. It is
defined as the best estimate plus an ample margin to account for ( TDur = TPOP ), but the forecasted wave height must be less than
inaccuracies in schedule and delays. This is the approach used in Hs, oper for the operation reference period, TR. In the GL Noble
all referenced standards. Using the notation from DNV (2011), the Denton method, α is based on the operation reference period
operation reference period, TR, is defined as follows: (TDur = TR ). It should be noted that the safety formats (load and
material factors) are somewhat different in the DNV and GL Noble
TR = TPOP + TC (1) Denton formulations and that the corresponding αs may not be
directly comparable.
where TPOP is the planned operation period and TC is the estimated
maximum contingency time. (The estimated maximum con-
tingency time is often between 50% and 100% of the planned 2.3. Weather-unrestricted operations
operation period, unless more accurate information is known.)
Operations with durations longer than three days are typically
2.2. Weather-restricted operations weather unrestricted. The separation between these two cate-
gories is important, as these two types of operations will be
If the duration of the operation is less than 72 h, then the designed differently with respect to environmental loads.
operation can be defined as a weather-restricted operation. An Weather-unrestricted marine operations are not planned based
operation can also be defined as weather restricted if it can be on weather forecasts, because the duration of such an operation is
halted and the handled object brought into safe conditions during longer than the duration over which weather forecasts are con-
the same period. For a sea transport operation, this means that the sidered reliable. Instead, the environmental conditions used for
route must be divided into several legs, and ports or areas of planning must be based on long-term statistics. The environmental
shelter along the transport route must be predefined. Updated
weather forecasts are received regularly throughout the entirety of Table 1
The parameter α as a function of operation duration from DNV (2011) and GL Noble
such an operation.
Denton (2013) for the case of one weather forecast and no wave monitoring. In
Traditionally, the maximum duration of a weather-restricted DNV (2011), the parameter definition is valid only for the North Sea and the
operation has been three days including contingency time, i.e., Norwegian Sea and is given as a function of the design wave height.
TR ≤ 72 h . This limit is stated in ISO 19901-6 (2009), GL Noble
TDur (h) DNV GL Noble Denton
Denton (2013), LOC (1997), NORSOK (2007). DNV (1996/2000) also
adhered to this limit until 2011. In DNV (2011), however, the Hs = 2 m 4m ≥ 6m
maximum operation period was increased to four days including
contingency provided that the planned operation period is less 12 0.76 0.79 0.8 0.69
than three days. 24 0.73 0.76 0.78 0.65
48 0.68 0.71 0.74 0.59
For marine operations in areas and seasons in which it can be
72 0.63 0.68 0.72 0.54
demonstrated that weather forecasts are capable of predicting any
638 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647

loads will then be based on a set of conditions with a given (low) Forecasts (ECMWF) and the US-based National Centers for Envir-
probability of being exceeded. onmental Prediction (NCEP).
In the planning of an unrestricted operation, the environmental The ECMWF system includes atmospheric variables, such as
criteria for the design must be based on long-term statistics wind, temperature and precipitation, in addition to waves for
accounting for offshore applications. The forecasts are based on the Ensemble
Prediction System (EPS) (see, e.g., Saetra and Bidlot, 2004). The

the geographical area, dynamical weather system is then simulated several times, each

the season of the year and time changing the initial conditions slightly. The forecasters

the duration of the operation. receive data from the ECMWF and perform their own evaluations
and interpretations of the results, on which the weather forecasts
The extreme values of the wave heights may be calculated are then based. The forecasts from different meteorologists may
based on scatter diagrams, e.g., those from BMT Ltd (1986) or DNV therefore differ for the same location and time.
(2014). It should be noted that the scatter diagrams from DNV are For projects involving the installation of structures at offshore
based on visual observations of the sea and may therefore, to some locations, weather forecasts are issued throughout the project
extent, include the effects of heavy weather avoidance (i.e., the period, which may be several years. In this paper, we will consider
largest waves are never observed). forecasted data for the Skarv oil and gas field, which is located
For commercial projects, more accurate data may be purchased, 210 km west of Sandnessjoen, Norway, at a water depth of 350–
e.g., from Fugro Oceanor. Data from Fugro Oceanor are derived 450 m. The weather forecasts are provided by BP.
from hindcast models and are calibrated against satellite data and, The forecasts include, amongst other information, the sig-
where available, in situ wave buoy data (FugroOceanor, 2012). nificant wave height Hs and the zero-crossing period Tz for wind-
A study by Shu and Moan (2008) of a VLCC (very large crude generated waves, swell and the total sea. The relation between
carrier) and a bulk carrier indicated that the use of data from Fugro these significant wave heights is Hs, total sea = (Hs2, wind waves + Hs2, swell )0.5.
Only the total sea, i.e., the significant wave height resulting from
Oceanor yielded amidships bending moments that were approxi-
both wind generated waves and swell, has been assessed in this
mately 15% larger than those deduced from the scatter diagrams of
paper. The lead time is defined as the numbers of hours from the
DNV (2014).
time when the forecast is issued until the time for which it applies.
Another alternative is the computer program Safetrans from
The first set of forecasted values is for a three-hour lead time.
Marin (2007), which contains a large environmental database and
Forecasted values are generally given every three hours for the
can provide wave statistics for certain transport routes and
first 72 h (for some forecasts, 69 h) and every six hours thereafter
seasons.
until the 168th hour. The data are for the year 2011 and include
1150 forecasts (generally three forecasts per day, with four fore-
casts for certain days).
3. Description of environmental conditions
3.2. Hindcast data
3.1. Weather forecasts
The formula given in Eq. (23) is a simple hindcast model, in
Several global systems are available for weather forecasters, which the significant wave height is estimated from the wind
e.g., those from the European Centre for Medium-Range Weather speed. This model does not, however, include the effects of wind

16

14

12

10
H,m

0
0 50 100 150 200 250 300 350 400
Day number
Fig. 1. Forecasted and hindcasted significant wave heights for a lead time of seven days.
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 639

fluctuations in time and space or any effects of the sea bottom few days before and after this date are shown together with the
topology, etc. A surface ocean wave model that does include such forecasted wave heights issued 72 and 168 h before. The max-
effects is used by meteorologists to hindcast wave data. imum wave height was not captured in any of the forecasts, but
The hindcast data used in this study were provided by the the three-day forecast was, as expected, closer than the seven-day
Norwegian Meteorological Institute (DNMI, http://www.met.no). forecast.
The data are based on the WAM Model of the Wamdi Group The difference between the hindcasted and forecasted sig-
(1988). The quality of these hindcast data compared with the nificant wave heights is also illustrated for two different lead times
observed data is very good; see, e.g., Reistad et al. (2011). There- in Fig. 3 for one two-week period in June and one in December.
fore, in this paper, these hindcast data are used instead of
observed data.

3.3. Comparing forecast and hindcast data 4. Model uncertainty in weather forecasts

Forecasted and hindcasted wave heights are shown in Fig. 1. 4.1. Statistical models
The figure shows the significant wave height in the year 2011 as a
function of time in days (the first of January is day no. 1, and so on). The environmental conditions are an important input for
The hindcasted significant wave height is shown in black. The marine operations. The significant wave height and wind speed
forecasted significant wave height shown in red is given for a lead are key information obtained from weather forecasts. For certain
time of 168 h, i.e., these data represent the weather as forecasted marine operations, the wave periods and wave directions may also
seven days before (e.g., the wave height shown on day 30 is taken be important. The uncertainties inherent in the environmental
from the weather forecast issued on day 23, and so on). No details conditions predicted by forecasts, e.g., the significant wave height,
can be seen from the plot, but it is apparent that the trend is can be quantified using two different mathematical models, one
predicted quite well, even if the maximum significant wave additive and one multiplicative model. The statistical parameters
heights (the peak values) are not forecasted. The maximum sig- for these models are estimated based on hindcasted and fore-
nificant wave height at this location in 2011 was 15.8 m on day no. casted data. The additive model is formulated as follows:
329 (i.e., on 2011-11-25) at 18.00 h. In Fig. 2, the wave heights a
Δ = Ztrue − Z predicted (3)
16
Hindcasted values
where Δ is a stochastic variable. In our case, the predicted values
14 Forecast, TL = 72 h are obtained from weather forecasts, whereas the true values are
Forecast, TL = 168 h the hindcast data. In the multiplicative model, the stochastic
12
variable χ is defined as follows:
10
Ztrue
χ=
Hs, m

8 Z predicted (4)
6 Because the stochastic variable (Δ or χ) depends on the wave
4 height, lead time (or forecasting period) and season, the mean
value and the standard deviation of this variable are also functions
2
of these parameters.
0 The statistical parameters necessary for a realization of the
324 326 328 330 332 334
stochastic variable Δ (or χ) are calculated via the standard for-
Day number
mulas using the software package (Matlab, 2010).
Fig. 2. Forecasted Hs values for lead times of three days and seven days together
with hindcast Hs values for November 20 (day no. 324) to 30, 2011
4.2. Uncertainty as a function of lead time
4 TL = 72h
We now consider the significant wave height given in a
ΔH (TL), m

TL = 168h From June 9. to 23.

2 weather forecast and define a stochastic variable ΔHs (TL ) as fol-
lows:
s

0
ΔHs (TL ) = Hs, hindcast − Hs, forecast (TL ) (5)
−2
160 162 164 166 168 170 172 174 where TL is the lead time. (A similar definition may also be used in
Day number the multiplicative model.) In Fig. 6, ΔHs is shown as a function of
lead time.
4 The correlation between the forecasted and hindcasted sig-
ΔH (TL), m

From December 14. to 28.

2 nificant wave heights is shown for several lead times in Fig. 4. As
expected, the correlation is initially high and decreases with
s

0 increasing lead time. In Fig. 5, the wave periods (spectral peak

periods, Tp) are shown in a similar manner. It is apparent that the
−2
348 350 352 354 356 358 360 362 correlation between forecasted and hindcasted significant wave
Day number heights is higher than the correlation between forecasted and
hindcasted wave periods. In fact, it seems preferable to use a fitted
Fig. 3. The difference between hindcasted and forecasted significant wave heights
( ΔHs (TL ) in Eq. (5)) for one two-week period in June and one in December 2011 for conditional probability distribution for the wave period but to base
lead times of three and seven days the wave height solely on the weather forecast.
640 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647

Fig. 4. Forecasted versus hindcasted significant wave height; note that for the longest forecast period, the largest waves are not forecasted

4.3. Uncertainty within a forecast period 4.4. Quantification of model uncertainty

In Eq. (5), the forecasting uncertainty is considered as a func- The methodology for estimating the uncertainty is as follows:
tion of the lead time alone. For most marine operations, however,
the primary concern is not whether a certain weather condition
Establish the observed data and weather forecasts for a certain
occurs exactly when forecasted but rather whether it occurs at all period of time at a chosen location.
during the marine operation. Given the maximum wave height

Calculate ΔHs,max (TR ) from Eq. (6) and χHs,max from Eq. (9).

Calculate the statistical parameters.
that is predicted to occur during a certain period, the probability

Choose the statistical distribution for the variables. For each
that this wave height will be exceeded can be estimated.
data set, fit the distribution to the data.
A stochastic variable can be defined based on the additive

Extract percentile values for ΔHs and χHs as functions of lead time
model as follows:
and forecasted Hs, either based on the chosen statistical dis-
ΔHs, max (TR ) = Hs, hc, max − Hs, fc, max (6) tribution or directly from the data sets.

where The mean values and standard deviations for realizations of

these stochastic variables (i.e., for the data from 2011) are shown
Hs, fc, max = max {Hs, forecast (τ )} for t ≤ τ ≤ t + TR
τ (7) in Figs. 7 and 8 for the additive and multiplicative models,
respectively. (Note that whereas a perfect weather forecast would
is the maximum forecasted significant wave height during the
yield a mean value of 0 m for ΔHs, max , the corresponding mean value
operation reference period and
would be equal to 1.0 for χHs, max .)
Hs, hc, max = max {Hs, hindcast (τ )} for t ≤ τ ≤ t + TR Figs. 7 and 8 present all-year data. However, the majority of
τ (8)
marine operations are performed during the summer season, and
is the maximum hindcasted significant wave height during the therefore, seasonal data may be more relevant to use for analysing
same period. Similarly, a stochastic variable can be defined based the uncertainties in weather forecasts. The data are therefore
on the multiplicative model as follows: divided into two seasons: the summer season, from April to Sep-
tember, and the winter season, from October to March. (Data are
Hs, hc, max
χ Hs, max (TR ) = often divided into four seasons or into monthly data, but because
Hs, fc, max (9)
the data for only one year are considered here, only two seasons
are defined.)
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 641

Fig. 5. Forecasted versus hindcasted peak wave periods

2 1.5
Hs,fc=0 − 2
Mean of ΔHs
Mean of ΔHs

1 2−4
1
4−6
0
6−
0.5
−1

−2 0
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Lead time, hours Forecast period, hours
4 3
Stand. dev. of ΔHs
Stand. dev. of ΔHs

Hs,fc=0 − 2
3
2−4 2
2 4−6
≥6 1
1

0 0
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Lead time, hours Forecast period, hours

Fig. 6. Mean values and standard deviations of the stochastic variable ΔHs as a Fig. 7. Statistical parameters for the stochastic variable ΔHs, max , based on the
function of lead time for four groups of forecasted significant wave heights: 0–2 m, additive model, as functions of the forecast period for four groups of forecasted
2–4 m, 4–6 m and >6 m . significant wave heights: 0–2 m, 2–4 m, 4–6 m and >6 m

In Figs. 9 and 10, the means and standard deviations of ΔHs, max and is, to some extent, also observed for the seasonal data pre-
are given for the summer and winter seasons for two groups of sented in Figs. 9 and 10. In the groups corresponding to significant
forecasted wave heights. The corresponding parameters for the wave heights of 2–4 m and 4–6 m, the mean and the standard
variable χHs, max are shown in Figs. 11 and 12. deviation are both larger during the winter season than during the
It is evident that the means and standard deviations of ΔHs, max summer season, except for Hs for the 4–6 m group for forecast
depend on the size of the forecasted significant wave height; this periods of less than 50 h. (Notably, this variation could also, to
is particularly apparent from the all-year data presented in Fig. 7 some extent, be attributed to the fact that the forecasted wave
642 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647

1.5 1
Mean of χHs

Mean of ΔHs
1
Hs,fc=0 − 2
0.5
2−4
0.5
4−6
6−
0 0
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Forecast period, hours Forecast period, hours
0.4 2
Stand. dev. of χHs

Stand. dev. of ΔHs

Summer
0.3 1.5 Winter
All year
0.2 1
0.1 0.5

0 0
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Forecast period, hours Forecast period, hours
Fig. 8. Statistical parameters for the stochastic variable χ Hs, max , based on the Fig. 10. Statistical parameters for the additive model as functions of the forecast
multiplicative model, as functions of the forecast period for four groups of fore- period for the summer (April–September) and winter (October–March) seasons for
casted significant wave heights: 0–2 m, 2–4 m, 4–6 m and >6 m . forecasted wave heights between 4 and 6 m

1.5
1 Mean of χHs
1
Mean of ΔHs

0.5 Summer
0.5
Winter
All year
0
20 40 60 80 100 120 140 160
0
20 40 60 80 100 120 140 160 Forecast period, hours
Forecast period, hours
Stand. dev. of χHs

2 0.4
Stand. dev. of ΔHs

Summer
1.5 Winter 0.3
All year 0.2
1
0.1
0.5
0
20 40 60 80 100 120 140 160
0
20 40 60 80 100 120 140 160 Forecast period, hours
Forecast period, hours
Fig. 11. Statistical parameters for the multiplicative model for the summer (April–
Fig. 9. Statistical parameters for the additive model as functions of the forecast September) and winter (October–March) seasons for forecasted wave heights
period for the summer (April–September) and winter (October–March) seasons for between 2 and 4 m
forecasted significant wave heights between 2 and 4 m.
corresponding interval ranges from −5% to +10% of the estimates.
heights are larger during the winter season and could thus reflect For the skewness and kurtosis, the uncertainties are much larger.
behaviour similar to that observed in Fig. 7 rather than an actual The 95% confidence interval varies with the forecast period, but it
seasonal variation.) ranges from approximately −50% to +100% of the estimates
The effects of the forecasted wave height and season are (hence, single values in the confidence interval may be from
smaller for χHs, max . In Fig. 8, the mean values are essentially iden- 0.5 times to 2 times the corresponding estimates). It is well known
tical for all wave height groups, and the spread in the estimated that estimates of skewness and kurtosis suffer from a relatively
standard deviation also appears not to depend on the forecasted large uncertainty for data sets of limited size. However, these
wave height. parameters are used here only to identify a plausible statistical
Based on the above findings, we choose to use the multi- distribution to be fitted to the data set, and hence, the uncer-
plicative model and to include all wave heights in a single group tainties are considered acceptable.
when performing the analysis. The behaviours of the statistical
parameters in this analysis are shown in Fig. 13.
5. Statistical description of environmental conditions
4.5. Uncertainty in point estimates
5.1. Significant wave heights
The uncertainties in the model parameter estimates are cal-
culated via a bootstrap method using Matlab (2010). The 95% A short-term sea state is described by a significant wave height
confidence interval for the mean value in Fig. 13 is found to be and a wave period and possibly also by other wave spectral
within ±1% of the estimates. For the standard deviation, the parameters (bandwidth parameter, doubly peaked spectrum, etc.).
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 643

1.5 defined as a stochastic variable. In the additive model, this variable

is defined as follows:
Mean of χHs

1 Hs, max = hs, fc + ΔHs, max (10)

Summer where ΔHs, max is defined in Eq. (6). The mean value and the stan-
0.5
Winter
dard deviation are equal to
All year
0 μ Hs, max = hs, fc + μ ΔHs, max
20 40 60 80 100 120 140 160 (11a)
Forecast period, hours
Stand. dev. of χHs

σ Hs, max = σ ΔHs, max (11b)

0.4
0.3 In the multiplicative model, the maximum wave height is
0.2 expressed as follows:

0.1 Hs, max = χ Hs, max ·hs, fc (12)

0 where χHs, max is defined in Eq. (9). The mean value and the standard
20 40 60 80 100 120 140 160
Forecast period, hours deviation are equal to

Fig. 12. Statistical parameters for the multiplicative model for the summer (April– μ Hs, max = hs, fc ·μ χ Hs, max (13a)
September) and winter (October–March) seasons for forecasted wave heights
between 4 and 6 m.
σ Hs, max = hs, fc ·σ χ Hs, max (13b)
1.5 It is evident from Fig. 13 that the skewness is positive and that
the kurtosis is larger than three. Thus, the log-normal distribution
Mean of χHs

1 may be suitable (see, e.g., Hahn and Shapiro, 1967, Fig. 6-1).
Under the assumption of a log-normal distribution, the prob-
0.5 ability density function for the maximum significant wave height
during the operation reference period for a given weather forecast
0 may be expressed as follows:
20 40 60 80 100 120 140 160
Forecast period, hours 1 2
fHs, max (hs, max ) = e−(1/2)((ln hs, max − μln Hs ) / σ ln Hs)
0.4 σ ln Hs hs, max 2π (14)
Stand. dev. of χHs

where the mean value and the standard deviation of the logarithm
0.3
of Hs are calculated as follows:

0.2 μln Hs = ln (hs, fc ) + μln χ (15a)

0.1
20 40 60 80 100 120 140 160 σ ln Hs = σ ln χ (15b)
Forecast period, hours
2 where hs, fc is given in meters and
Skewness of χHs

1.5 ⎛ ⎛σ ⎞2 ⎞
⎜ ⎟ ⎟
ln ⎜ 1 + ⎜⎜
χ Hs , max
σ ln χ = ⎟ ⎟
1 ⎝ μ χ Hs, max ⎠ ⎠
⎝ (16a)
0.5

0
20 40 60 80 100 120 140 160 1 2
μln χ = ln (μ χ Hs, max ) − σ ln χ
Forecast period, hours 2 (16b)
15 The cumulative distribution function is
Kurtosis of χHs

⎛ ln (hs, max ) − μln H ⎞

10 F Hs, max (hs, max ) = Φ ⎜ s
⎟
⎝ σ ln Hs ⎠ (17)
5
where Φ ( ) is the cumulative distribution function for the standard
normal distribution.
0
20 40 60 80 100 120 140 160
Forecast period, hours 5.2. Wave period
Fig. 13. Mean value, standard deviation, skewness and kurtosis of the stochastic
variable χ Hs, max , based on the multiplicative model, as functions of the forecast The wave period is given in the weather forecast and may be
period. included in the analysis in a similar manner as is the wave height.
However, because of the relatively low correlation between fore-
Given a forecasted significant wave height (i.e., the maximum casted and hindcasted periods (see Fig. 5), this is not done here.
predicted value during the forecast period) hs, fc , the maximum The uncertainty in the wave periods can instead be accounted for
significant wave height to be used in an operational design may be using a statistical distribution that is conditional upon the
644 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647

Table 3
Mean wind speeds and standard deviations based on the conditional Weibull dis-
tribution (see Eqs. (19)–(21)) with coefficients from Table 2.

Hs (m) μV | Hs (m/s) σ V | Hs (m/s)

1 4.6 2.7
2 7.0 3.0
4 10.9 3.0
6 14.2 2.9
8 17.2 2.7
10 19.9 2.5
12 22.4 2.4
14 24.7 2.2

Fig. 14. Wind speed at a height of 10 m averaged over one hour versus significant
wave height (hindcast data) at the Skarv field for 2010 and 2011, together with the
wave-wind relation obtained from Holthuijsen's formula, Eq. (24), and the fitted
Weibull distribution, i.e., the mean value from Eq. (20) with error bars of ±2σ V | Hs
(see Eq. (21))

Table 2
Coefficients obtained by fitting Eq. (19) to the data
from the Skarv field for 2010 and 2011.

Coefficient Fitted value

c1 1.23
c2 0.55
c3 1.17
c4 0.00
c5 5.19 Fig. 15. Mean value (μln χ ) and standard deviation (σ ln χ ) of the logarithm of χ in the
c6 0.61 multiplicative model, based on the previously mentioned Skarv data for all fore-
casted wave heights, as a function of the operation period.

significant wave height. The period may be described by a log- The joint probability distribution of wave height and wind
normal distribution, as shown by Bitner-Gregersen and Haver speed is fV , Hs (v, h) = fHs (h) fV | Hs (v|h), where the conditional dis-
(1991). For design purposes, the distribution of the zero-crossing tribution of the wind speed can be described by a two-parameter
period, Tz, follows a log-normal distribution that is conditional on Weibull distribution; see Bitner-Gregersen (2005):
Hs: v k − 1 −(v / VC )k
fV | Hs (v|h) = k e
1 2 VCk (19)
fTz | Hs (t|h) = e−(1/2)((ln t − μln Tz ) / σ ln Tz)
σ ln Tz t 2π (18)
where the shape parameter is k = c1 + c2 hc3
and the scale para-
with parameters μln Tz = a1 + a2 ha3 and σ ln Tz = b1 + b2 eb3 h (see, e.g., meter is VC = c4 + c5 hc6 . The coefficients ci, i = 1, 2, … , 6, are esti-
DNV, 2014). The coefficients ai and bi, with i = 1, 2, 3, are esti- mated from real data. The mean value of the conditional wind
mated from real data. speed is
For engineering purposes, the upper and lower bounds on the ⎛ 1⎞
wave periods are given as functions of the significant wave height μV | Hs = E [V |Hs ] = VC Γ ⎜ 1 + ⎟
⎝ k⎠ (20)
by, e.g., DNV (2011), GL Noble Denton (2013), and LOC (1997).
and the variance is
5.3. Wind speed ⎛ ⎛ 2⎞ ⎛ 1 ⎞⎞
σV2| Hs = VC2 ⎜ Γ ⎜ 1 + ⎟ − Γ2 ⎜ 1 + ⎟ ⎟
⎝ ⎝ k⎠ ⎝ k ⎠⎠ (21)
The wind speed may be treated in a similar manner as the wave
where Γ is the gamma function (Γ (t ) = ∫ xt − 1e−x dx ).
∞
heights, namely, by comparing forecasted values with observed
0
values, resulting in a statistical description of the deviation. In Fig. 14, the hindcasted significant wave height and the
Alternatively, the wind speed can be inferred based on the joint hindcasted wind speed at a height of 10 m averaged over one hour
probability density function of wind speed and wave height. are plotted. Values are given for every three hours at the Skarv
The significant wave height for wind-driven waves is typically field in 2010 and 2011 (Table 3). In the same figure is also plotted
conditional upon the wind speed, as the wind creates such waves; the wind speed indicated by the conditional Weibull distribution
see, e.g., Johannessen et al. (2002). from Eq. (19), which is plotted with error bars equal to ±2σ V | Hs (i.e.,
However, in our scenario, the wind speed must be determined approximately the 95% confidence interval). The parameters used
for a given significant wave height. in Eq. (19) are given in Table 2.
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 645

Table 4
Select numerical values of μln χ and σ ln χ , from Fig. 15.

No. of days (–) TR (h) μln χ (–) σ ln χ (–)

1 24 0.055 0.112
2 48 0.066 0.119
3 72 0.079 0.134
4 96 0.084 0.153
5 120 0.095 0.176
6 144 0.111 0.204
7 168 0.127 0.224

6. Environmental conditions for weather-unrestricted

operations

6.1. Design environmental conditions

For weather-unrestricted operations, the environmental design

conditions are based on long-term statistics, possibly accounting
for seasonal variations. According to ISO 19901-6, weather- Fig. 16. Probability of exceeding Hs = 6 m as a function of the forecasted significant
unrestricted operations may be planned using environmental cri- wave height for three operation reference periods: 24 h, three days and seven days

teria with return periods estimated as a multiple of the operation

duration. A minimum of 10 times the duration of the operation are also given by other authors, e.g., Gran (1992, Eq. 3.4.63), with a
may be used (ISO 19901-6, 2009). (However, for operations with coefficient of 0.18 instead of 0.24 (for a wind speed V observed 10–
durations of up to seven days, environmental criteria based on 20 m above the sea surface). Note that the spread (e.g., the stan-
seasonal data with a return period of one year are recommended.) dard deviation) of the Hs data is not described by this formula.
An alternative is a method proposed by Lindemann (1986) for Inverting Eq. (23) yields the wind speed as a function of the sig-
calculating the design significant wave height as a function of the nificant wave height:
duration with a defined exceedance probability of 10%. This
Hs g
method is used by DNV (2011). V=
0.24 (24)

6.2. Exceedance probabilities for the wave height The wind speed according to Eq. (24) is also shown in Fig. 14.
(The wind speed in Eq. (24) is averaged over 10 min, whereas the
Consider a marine operation with a given duration and in a hindcasted wind speed, which is approximately 10% lower, is
given season for which the long-term statistics for the geo- averaged over one hour.) The significant wave height in Eq. (23)
graphical area yield a significant wave height for the design that is represents wind-generated waves. By contrast, the significant
equal to hs, d . Based on the log-normal distribution from Eq. (17), wave height assumed in long-term wave distributions describes
the probability that the actual (observed) significant wave height the total sea, i.e., it also includes swells. This means that when the
will be larger than the design value can be expressed as follows: wind speed is calculated via Eq. (24) using a significant wave
⎛ ln (hs, d ) − μln H ⎞ height from a long-term distribution representing the total sea,
Pe = P (Hs, max ≥ hs, d ) = 1 − Φ ⎜ s
⎟ the wind speed will be overestimated.
⎝ σ ln Hs ⎠ (22) In engineering, a simple method of calculating the wind speed
corresponding to a specified significant wave height is useful for
μln Hs and σ ln Hs are calculated from Eq. (15) for the maximum
operation design. As a simplification, the wind speed may be taken
forecasted wave height, hs, fc , during the operation period, TR.
to be a deterministic function of the wave height if a formula
similar to Eq. (24) is fitted to the upper limit in Fig. 14 (i.e., to
6.3. Simplified numerical values for the weather forecasting
μV | Hs + 2σ V | Hs ; see Eqs. (20) and (21)). The factor 0.24 is then
uncertainty
replaced with 0.15; hence, Vone hour = Hs g /0.15 ≈ 8 Hs . (A more
Using the expressions given in Eqs. (22) and (15), the prob- sophisticated curve could also be used, but Eq. (24) is both simple
ability of exceeding the design criteria over a period of several and convenient and has the traditional form.) This formula could
days can be estimated based on a given weather forecast. The be used for planning operations in the Norwegian Sea, bearing in
mean and the standard deviation of χ are given in Fig. 13 for all- mind that it is based on two years of data. Because the formula is
year data. In Fig. 15, the mean and the standard deviation of the valid for wind speeds averaged over one hour, it should be
logarithm of χ as calculated using Eq. (16) are shown. Select values transformed to correspond to the actual averaging time used in
of μln χ and σ ln χ are given in Table 4. the case under consideration. A one-minute averaging time is
often used, in which case the wind speed is approximately 20%
6.4. Simplified design wind speed higher than the one-hour wind speed (see, e.g., DNV, 2011). The
one-minute design wind speed can thus be approximated as fol-
A simple relationship between the significant wave height and lows:
the wind speed is given in Holthuijsen (2007, Section 6.3.2):
Vone minute ≈ 10 Hs (25)
V2
Hs = 0.24 where Hs is given in m and the wind speed is given in m/s. Note
g (23)
that this is merely an approximate formula for wind speed and is a
where V is the sustained wind speed (10-min average) 10 m above function of the significant wave height only. The relation between
the sea surface. Similar expressions with slightly varying constants wave height and wind speed at a certain location may depend on
646 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647

level. A probability of 10% that the design wave will be exceeded is

used in DNV (2011) for weather-unrestricted operations.
As an example, suppose that the maximum allowed forecasted
significant wave height for a certain marine operation is 5 m.
(Hence, the operation cannot begin before the forecasts indicate
Hs ≤ 5 m for the entire duration of the operation.) Using the values
from Table 4, the maximum significant wave heights obtained
from Eq. (26) are 6.4 m for a three-day operation and 7.6 m for a
seven-day operation, with Pe ¼0.1. (These would then be the
characteristic values to be used in design.)
The ratio between the forecasted wave height and the design
wave height for a three-day operation is 5/6.4 = 0.78. This value
can then be compared with the values from the design standards.
The ratio between the forecasted wave height and the design wave
height for a 72-h operation is 0.72 according to DNV (2011) (for
Hs, design ≥ 6 m ) and 0.54 according to GL Noble Denton (2013); see
Table 1. Hence, these standards yield conservative results com-
pared with our data set in this case.
For a seven-day operation period, no ratio is given by the
Fig. 17. Probability of exceeding Hs = 9 m as a function of the forecasted significant design standards, but the ratio between the forecasted and design
wave height for three operation reference periods: 24 h, three days and seven days.
wave heights according to the results from Table 4 is 5/7.6 = 0.66.
In general, standards for marine operations do not allow the
the wind and wave directions, storm duration, fetch, water depth planning and design to be based on weather forecasts when the
and possibly other covariates. However, Eq. (25) can be used for operation duration is more than three days unless it can be
engineering or feasibility studies and is valid for deep water and demonstrated that the relevant weather forecasts can predict any
unlimited fetch. extreme weather conditions over a longer period. The method
For structural reliability analyses, the conditional probability described in this paper may be used to assess the reliability of such
distribution given in Eqs. (19)–(21) may be used. weather forecasts, preferably based on a more extensive data set.
Because only weather forecasts for a single location are included in
the data set considered in this paper, the results cannot be directly
7. Reliability of a weather forecast applied elsewhere. However, they are considered to be repre-
sentative of extratropical conditions.
The probability of exceeding a certain significant wave height Because only one year of data is included, there will be a rather
given a certain forecasted wave height and operation period can large uncertainty in the calculated values; see, e.g., Moan et al.
be estimated from Eq. (22). (2005).
In Fig. 16, the probability of exceeding a significant wave height
of 6 m is plotted as a function of the forecasted wave height for
three different operation periods. Similarly, the probability of 8. Conclusions
exceeding a significant wave height of 9 m is plotted in Fig. 17.
It is observed that if the forecasted wave height is, e.g., 4 m, The uncertainty in environmental conditions based on weather
then the probability that the actual wave height will be greater forecasts has been studied. Forecasted significant wave heights
than 6 m after 24 h is approximately 10  3; after three days, it is have been compared with hindcasted values using an additive and
approximately 10  2; and after seven days, the probability of a multiplicative statistical model. In the additive model, the mean
exceedance is 10  1. The probability that the actual wave height value and the standard deviation are more strongly dependent on
will be greater than 9 m is negligible after either 24 h or three the forecasted wave height than in the multiplicative model,
days. After seven days, the probability of exceeding a wave height making the latter the preferred model in this study.
of 9 m is approximately 10  3. It is also apparent from Fig. 17 that if The uncertainty in the forecasts increases with increasing lead
the forecasted wave height is 6 m, then the probability of time, reducing the correlation between the forecasted and hind-
exceeding a 9 m wave height after seven days is approximately casted data. The correlation between forecasted and hindcasted
10  1. Hence, the weather forecasts still provide interesting infor- data is lower for wave periods than for significant wave heights for
mation even after one week. Because the results are sensitive to the same lead time. Therefore, it is preferable to model the wave
the tail of the distribution when extreme values are considered, period conditionally upon the wave height.
the values here should be taken as examples only. The wind speed may be modelled using a Weibull distribution
The primary concern in engineering is not the probability of that is conditional upon the wave height. For marine operations in
exceedance but rather how to calculate a design wave height. which the governing environmental load is caused by waves and
The design wave height can be estimated for a given probability of the corresponding wind speed must be estimated, a simple engi-
exceedance and a forecasted wave height. By substituting Eq. (15) neering method is proposed. The relationship between wind speed
into Eq. (22) and solving for Hs, d , the maximum significant wave and wave height was developed for wind-driven waves. Hence, if
height can be calculated as follows: the waves in fact contain swells in addition to wind-driven waves,
Hs, d = hs, fc exp (μln χ + σ ln χ Φ−1 (1 − Pe )) then the wind speed may be overestimated. Moreover, in sheltered
(26)
water or other cases in which there may be high wind speeds but
The probability of exceedance, Pe, should correspond to the safety small or no waves, this method is not applicable. In such cases, it
format used in the design of the marine operation. The safety will be necessary to determine the wind speed using other
factors used in the structural design (typically load and material methods, e.g., the return period approach based on the duration of
factors) will depend on the probability of exceeding a certain load the operation.
 A. Natskår et al. / Ocean Engineering 108 (2015) 636–647 647

According to the design standards for marine operations, only References

operations with planned durations of less than three days are
generally planned as weather-restricted marine operations, unless Bitner-Gregersen, E., Haver, S., 1991. Joint environmental model for reliability cal-
it can be demonstrated that the relevant weather forecasts are culations. In: ISOPE. Proceedings of the First International Offshore and Polar
Engineering Conference, vol. 1, Edinburgh, UK, pp. 246–253.
able to predict any extreme weather conditions over a longer Bitner-Gregersen, E.M., 2005. Joint probabilistic description for combined seas. In:
period. In that case, the operation reference period may be OMAE2005, p. 67382.
BMT Ltd, 1986. Global Wave Statistics.
increased. The data from the Norwegian Sea used in this paper do
DNV, 1996/2000. Rules for Planning and Execution of Marine Operations.
not reveal any specific limitations of the weather forecasts that DNV, 2011. Marine Operations, General. DNV-OS-H101.
would require a 72-h limit for weather-restricted operations. The DNV, 2014. Environmental Conditions and Environmental Loads. DNV-RP-C205.
FugroOceanor, 2012. 〈http://www.oceanor.no/services/〉.
method described in this paper may be used to assess the quality GL Noble Denton, 2013. General Guidelines for Marine Projects.
of weather forecasts for a period longer than three days based on Gran, S., 1992. Developments in Marine Technology. A Course in Ocean Engineering,
forecasts for the area of interest. vol. 8. Elsevier, Amsterdam.
Hahn, G., Shapiro, S., 1967. Statistical Models in Engineering. 1967. John Wiley &
The results may be used as input for structural reliability ana- Sons, Inc, New York (reprinted 1994).
lyses of marine operations. Holthuijsen, L.H., 2007. Waves in Oceanic and Coastal Waters. Cambridge University
Although the statistical model described in this paper considers Press, Cambridge, United Kingdom.
ISO 19901-6, 2009. Petroleum and Natural Gas Industries. Specific Requirements for
the significant wave heights, several other effects might have been Offshore Structures. Part 6: Marine Operations.
included. Seasonal variations and the dependences of the fore- Johannessen, K., Meling, T.S., Haver, S., 2002. Joint distribution for wind and waves
casted wave height have been discussed to some extent, but other in the northern North Sea. Int. J. Offshore Polar Eng. 12 (March), 1.
Lindemann, K., 1986. Extreme value statistics and exposure time: a simple method
covariates may also influence the results. The wave and wind to establish well defined criteria for marine operations. In: Offshore Technology
directions, water depth, geographical area, forecast provider or Conference, No. 5142 in OTC. OTC, Houston, pp. 519–525, May.
LOC, 1997. Guidelines for Marine Operations. London Offshore Consultants Limited,
other covariate effects could be included in these analyses. To
Ledbury, United Kingdom.
obtain the parameters for the engineering of an operation at a Marin, 2007. Safetrans. 〈http://www.marin.nl/web/JIPsNetworks/Public/Safetrans.
certain location, data reflecting the area of interest should be htm〉.
Matlab, 2010. Ver. 7.11. The MathWorks Inc., Natick, MA.
analysed. Moan, T., Gao, Z., Ayala-Uraga, E., 2005. Uncertainty of wave-induced response of
marine structures due to long-term variation of extratropical wave conditions.
Mar. Struct. 18 (May (4)), 359–382.
NORSOK, 2007. N-003, Actions and Action Effects.
Acknowledgements
Reistad, M., Breivik, Ø., Haakenstad, H., Aarnes, O.J., Furevik, B.R., Bidlot, J.-R., 2011.
A high-resolution hindcast of wind and waves for the North Sea, the Norwegian
The first author's work was financed by the Research Council of Sea, and the Barents Sea. J. Geophys. Res. 116, 18.
Saetra, Ø., Bidlot, J.-R., Potential benefits of using probabilistic forecasts for waves
Norway (RCN) through the Centre for Ships and Ocean Structures and marine winds based on the ECMWF ensemble prediction system. Weather
(CeSOS) at the Norwegian University of Science and Technology Forecast. 19 (4), 673–689, August.
(NTNU) as well as by support from the Education Fund of Det Shu, Z., Moan, T., 2008. Effects of avoidance of heavy weather on the wave-induced
load on ships. J. Offshore Mech. Arct. Eng. 130, 2.
Norske Veritas. This support is greatly appreciated. Wamdi Group, 1988. The WAM model—a third generation ocean wav prediction
Hindcast data were received from Magnar Reistad at the Nor- model. J. Phys. Oceanogr. 18, 1778–1810.
wegian Meteorological Institute and are also much appreciated.