Ocean Engineering 262 (2022) 112157

Contents lists available at ScienceDirect

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

Prediction of dynamic responses for execution of marine operations using

partitioning of multimodal directional wave spectra and machine learning
regression models
Wilson Guachamin-Acero a ,∗, Jesús Portilla a,b
a
Departamento de Ingeniería Mecánica, Escuela Politécnica Nacional, Ecuador
b Research Center of Mathematical Modelling (MODEMAT), Quito, 17-01-2759 Ecuador

ARTICLE INFO ABSTRACT

Keywords: Safe execution of marine operations (MOs) depends on accurate prediction of vessel dynamic responses,
Machine learning which are necessary to help a superintendent make on-board decisions. However, it can be challenging to
Wave spectra partitioning conduct complex numerical simulations prior to execute a MO, especially at offshore sites characterized by
Marine operations
complex wave conditions, i.e. multimodal directional (2D) wave spectra. In this context, this article introduces
Regression model
a methodology for assessing vessel dynamic responses using characteristics of actual 2D wave spectra and
Boosted trees algorithm
regression models from machine learning. First, a state-of-the-art algorithm is used for 2D spectra partitioning
and spectral characterization (climate). This allows identifying wave systems that can occur at a study location,
and their integral parameters. For that, a data set for the North Atlantic Ocean, spanning 36 years is used.
Then, the main wave parameters of each wave system including significant wave height, peak period and peak
direction, are used as features to train and test regression models from machine learning tool kits. Accurate
predictions of dynamic responses were obtained with a ‘‘boosted trees’’ regression model considering all six
long-term wave systems (18 features) detected at the site. These findings are valuable as new tools to help
on-board personnel make safe and quick decisions.

1. Introduction distribution in the 2D frequency-direction domain, and it is at the

core of modern wave prediction models (Komen et al., 1994; Cavaleri
An accurate prediction of vessel dynamic responses is important et al., 2007), and observations methods (Holthuijsen, 2007; Nwogu,
when executing marine operations (MOs), which are normally carried 1989). Traditionally, waves have been described through main integral
out during calm weather conditions because operational limits are parameters, e.g., significant wave height (Hs) and peak period (Tp).
generally low and vessels are sensitive to wave action. Since wave
However, these parameters lose a significant amount of information of
spectra at sea are generally multimodal, it is advisable to consider the
the spectral structure. This is critical because in most cases, waves in
full spectral energy distribution for assessing operability and weather
windows (Guachamin-Acero and Li, 2018). any particular site are composed by those locally generated by wind
Planning critical marine operations requires complex numerical (wind-sea), and those generated far away and advected to the local site
simulations using directional (2D) wave spectra. However, these simu- (swell). By definition there can be only one wind-sea component, but
lations can be computationally expensive and lengthy, so they may be there can be several swell trains depending on the exposure of the site
prohibitive during the execution phase, because there is little time for to different oceanic conditions. For MOs, the swell signal is particularly
on-board decisions to be made. The motivation of the present study is to important because their long periods significantly affect vessel motion.
put together existing state of-the-art methodologies, both for 2D wave Therefore, it is fundamental to know the spectral energy distribution
spectra analysis, and for prediction of vessels dynamic response using in order to more precisely predict vessel’s motions. The description of
machine learning (ML) algorithms. This new method allows to shorten the wave spectrum can be facilitated and simplified using the concept
computational times of complex numerical simulations, making its use
of spectral partitions. A spectral partition is defined as a part of the
feasible during execution of marine operations.
wave spectrum that contains one and only one individual wave com-
At present, the wave spectrum is the basic variable to describe
ponent (wind-sea or swell) of a specific meteorological origin. Using
ocean surface waves. The wave spectrum represents the full energy

∗ Corresponding author.
E-mail address: [email protected] (W. Guachamin-Acero).

https://doi.org/10.1016/j.oceaneng.2022.112157
Received 25 January 2022; Received in revised form 23 June 2022; Accepted 28 July 2022
Available online 17 August 2022
0029-8018/© 2022 Elsevier Ltd. All rights reserved.
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

this concept, wave components can be described by typical integral Table 1

Main particulars of the construction vessel (Guachamin-Acero and Li, 2018).
parameters (e.g., Hs, Tp) without losing spectral information (Portilla
et al., 2009). Nowadays partitioning methods are implemented in many Parameter Notation Value Units

routine analysis, wave models, and can used with all kinds of data Displacement ▽ 2.55 × 104 tons
Length 𝐿 140 m
(model, satellite, and in-situ observations).
Breadth 𝐵 30 m
On the other hand, machine learning algorithms are becoming Draught 𝑇 6 m
increasingly popular and they are being applied to predict dynamic Metacentric height 𝐺𝑀 7.5 m
responses of marine structures and environmental actions. Machine Vertical position of COG above keel 𝑉 𝐶𝐺 8 m
learning aims to identify patterns through iterative procedures in-
cluding self training, which is necessary to optimize a model and
predict variables. To date, substantial research is carried out in this
in winter, and it is the reference for the development of engineering
topic for many engineering applications, but developments for pre-
standards such as DNV-OS-H101 (2011) and International Organization
dicting dynamic responses in the context of marine operations are
for Standardization (2015).
still limited. For instance, Bjørni et al. (2021) used artificial neural
The selected offshore site is located at 47 𝑁 336.58 E with a water
networks (ANNs) with JONSWAP (Joint North Sea WAve Project)
depth of 3000 m. Fig. 1 shows its main spectral characteristics. Fig. 1(a)
analytical spectra to predict mooring line forces of the OC3 Hywind
represents the long-term occurrence distribution of partitions, which
wind turbine. They found a 71% accuracy in predicted responses, with
points to the main wave systems present in the area, labeled from
substantial computational efficiency. Other machine learning methods
1 to 6 (0 deg indicates North, directions are ‘‘going to’’). Fig. 1(b)
based on regression schemes have been used by Bassam et al. (2022)
shows the average seasonal magnitude, and Fig. 1(c) shows the Hs
to predict ferry speeds, assessing various operating conditions, from
distributions (in boxplot form) characterizing each wave system. We
which Gaussian regression process were found to produce the best
observe that wave systems 1 and 2 are stronger and also persistent all
responses. Other applications employ ANNs for predicting occurrences.
year round with a considerable magnitude. Fig. 1(d) also shows the
For instance, Doong et al. (2018) developed a warning model to predict
joint probability of pairs of WSs occurring simultaneously (with Hs >
coastal freak wave occurrence. Seven parameters including misalign-
1 m each, indicating that WSs 1 and 2 may co-exist relatively frequently
ment between wind and waves allowed an accuracy rate over 90%. In
(13% of the cases), see Portilla et al. (2009), or Modemat (2020) for
addition, Li et al. (2019) developed a ANN algorithm to predict short-
further details.
term wave forces on an oscillating body using real time measurements
Integrated Hs, Tp and peak direction (𝜃𝑝 ) of the most energetic
of wave elevation. Results from the prediction model are fed into a
spectral peaks of each spectrum are obtained via partitioning. These
controller, which is affected for deviations in the predicted results. Gra-
parameters are then used to model equivalent analytical 2D JONSWAP
cia et al. (2021) used ANN and gradient boosted decision trees for
improving prediction of significant wave height using analytical mod- spectra, see e.g. Fig. 2.
els, achieving an error reduction of about 36%. Tapoglou et al. (2021) In these conditions Guachamin-Acero and Portilla-Yandún (2021)
predicted significant wave height using satellite wave images and wave have shown that heading a vessel into the largest wave component does
buoy data as input. They obtained rms errors of about 0.23 m for Hs not necessarily gives the best responses as one should expect, because
less than 3 m. such responses are affected by the presence of other wave systems. This
In this paper we develop a methodology that combines the use of illustrates some challenges of executing MOs in these sea environments.
spectral parameters of long-term wave systems (WSs) as features for
regression models in machine learning. The aim is to make predictions 2.2. Vessel’s characteristics
of vessel’s dynamic responses during execution of marine operations,
while optimizing computational time. The main characteristics of the heavy lift vessel considered in this
paper are listed in Table 1.
2. Assessment of dynamic responses using 2D hindcast and JON-
SWAP wave spectra 2.3. Vessel dynamic properties and responses

The dynamic responses of a floating vessel will be assessed using Dynamic responses of the vessel can be obtained by solving Eq. (1),
actual 2D wave spectra and their corresponding JONSWAP parameter- given in matrix form (Century Dynamics Limited, 2002), where 𝑴 and
izations. 𝑨 are the structural and the added mass matrices respectively, 𝑩 is
the damping matrix, 𝑿 is the position vector, 𝑲 is the stiffness matrix
2.1. Spectral wave information and 𝑭 𝑒𝑥𝑡 is the total external force vector, including first- and second-
order wave forces and mooring line forces. Note that added mass,
At present, third generation numerical wave models are the stan- damping and harmonic external forces are frequency and direction
dard for wave prediction and hindcasts. These models use the wave dependent. For linear and weakly non-linear dynamic systems, heading
spectrum as working variable. In addition, most in-situ wave measuring of the vessel and direction of wave actions, Eq. (1) can be solved in the
devices are based on this variable as well (Smit et al., 2021), and frequency-domain to obtain the responses 𝑿.
modern satellites are being equipped with sensors capable of observing
− 𝜔2 [𝑴 + 𝑨 (𝜔, 𝜃)] 𝑿 (𝜔) + 𝑖𝜔𝑩 (𝜔, 𝜃) 𝑿 (𝜔) + 𝑲𝑿 (𝜔) = 𝑭 𝑒𝑥𝑡 (𝜔, 𝜃) (1)
it from space (Hasselmann et al., 2013). In the present study we use
model data, mainly due to the advantage of having continuous series In addition, by defining a mass and inertia properties for the hull
over long spans, but the method is valid for any type of data. sections, vertical bending moment (VBM) response amplitude oper-
The ERA-Interim archive (henceforth ERAI) from the European ators (RAOs) and phase angles can be computed (for details refer
Center from Middle Range Weather Forecasts (ECMWF) (Dee et al., to Guachamin-Acero and Portilla-Yandún (2021)). Fig. 3 shows Roll
2011) contains global time histories of 2D hindcast wave spectra. Data and VBM RAOs computed using the state-of-the-art diffraction code
is available every 6 h ranging from January 1979 to December 2015 Ansys AQWA (Century Dynamics-ANSYS Inc., 2011). In turn, these
(37 years). RAOs can be used to compute response spectra, and thus, dynamic
The focus of the present study is the North Atlantic Ocean, which is responses using the 2D wave spectra defined in Section 2.1.
generally characterized by the presence of various wave systems com- Fig. 4 shows examples of time sequences of 4xrms roll and VBM
posed of wind seas and swells. This area is prone to harsh conditions responses of the vessel computed for every 2D wave spectrum of the

2
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 1. Wave system components for North Atlantic Ocean 47 𝑁 336.58 E. (a) Occurrence distribution of wave systems. (b) Mean value of monthly averaged Hs for each wave
system. (c) Hs magnitude distribution in boxplot form. (d) Joint occurrence probability of pairs of WS. (Modemat (2020), https://modemat.epn.edu.ec/nereo/).

Fig. 2. Examples of directional (2D) wave spectra. (a) Multimodal hindcast spectrum. (b) Equivalent analytical directional JONSWAP spectrum with total spreading angle of 180
deg.

time series shown in the figure. It can be observed that actual dy- from the corresponding ones using hindcast 2D spectra. Some statistical
namic responses can be significantly different when using analytical 2D parameters of the error in the responses are shown in Fig. 5 for different
JONSWAP spectra. For the period January 1981–December 1982, the ranges of roll and VBM. It is apparent in this figure that the error bias
error is computed by subtracting responses for 2D JONSWAP spectra is not sensitive to these response levels of roll and VBM. In contrast,

3
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 3. Response amplitude operators. (a) Roll RAO. (b) Midships VBM RAO.

Fig. 4. Examples of dynamic response time histories computed using 2D hindcast wave spectra and 2D JONSWAP spectra. (a) Roll responses for 90 deg heading respect to the
true North. (b) Midships VBM responses for 90 deg heading respect to the true North.

the interquartile range increases for increasing magnitudes. Given the see e.g. Hasselmann et al. (1988), Komen et al. (1994), and the launch
fact that vessels used in marine operations are highly sensitive to of the first satellites equipped with instruments capable of observing
the wave energy distribution (Guachamin-Acero and Li, 2018), and the wave spectrum, namely the Synthetic Aperture Radar (SAR), the
considering the differences in dynamic responses shown in Figs. 4 possibility to incorporate data assimilation (DA) into wave models
and 5, in this paper we use actual 2D hindcast spectra for analysis became evident. However, wave data assimilation can only be effec-
of MOs. Moreover, these differences in responses are the reason for tive if the different wave components (i.e., wind sea and swell) are
the occurrence of unexpected responses during of execution of marine analyzed separately, because total integral parameters (e.g., combined
operations, especially heavy lifting. This fact was experienced during Hs) tend to compensate errors between these components. Therefore,
the execution of the Sheringham and Shoal wind farm project, which the distinction of waves between wind sea and swell is a pretty evident
caused delays and important economic losses (Li, 2016). one, mainly because these two regimes are subject to different physics,
In summary, by considering safety and challenges associated with and so they can be recognized even visually. Wind sea is more irregular,
the execution of marine operations, it is evident that more accurate broad banded in direction and characterized by higher frequencies, and
methods for predicting vessel dynamic responses need to be developed. constantly affected by breaking. In turn, swell is more linear, long,
For marine operations, these methods should make use the actual 2D and smooth (Holthuijsen, 2007). This distinction marks the onset of
wave spectra. wave partitioning. However, evident as these components are, finding a
simple mathematical scheme to perform this task automatically proved
3. Methodology less simple, very often a cutoff period (or frequency) at 10 s was used as
a rule of thumb, but this is too crude for most applications. In addition,
This section describes the numerical method used for 2D wave spec- strictly speaking, swell should be better described as swells, as in most
tra partitioning and its use for developing machine learning regression parts of the globe, there is more than one persistent swell characterizing
models to predict dynamic responses. the wave climate.
Most partitioning algorithms are based on the watershed algo-
3.1. Wave spectra partitioning rithm (Hasselmann et al., 1996), which is also called the mountaineer
scheme due its implementation. Starting at any bin of the spectral
Spectral partitioning was first introduced for data assimilation pur- matrix the algorithm searches the steepest ascent, and continues in that
poses, but soon after this concept gained relevance in nearly all fields of way until reaching a local peak. The collection of all points climbing
wave data analysis. With the advent of third generation wave models, to the same local peak constitutes (by definition) a partition.

4
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 5. Error statistical parameters for responses computed between January 1981 and December 1982. (a) Roll error statistics for 90 deg vessel heading. (b) VBM error statistics
for 0 deg vessel heading.

The algorithm so defined will detect any 3D feature of the energy point to a different and specific meteorological origin. Once these WSs
distribution, however, not all of them are relevant wave systems be- have been identified, their partial spectra (partition) can be extracted
cause the spectral distribution may contain secondary features (noise). and characterized individually. For instance Fig. 1(c) shows their Hs
Therefore, a post-processing step is in general necessary to identify distributions, and any other of their parameters can be obtained in
(only) the most relevant partitions. Here we use the scheme described the same fashion. This scheme has been applied at global scale to
in Portilla et al. (2009), which uses image processing tools to assist the ERAI data to produce a global atlas of spectral wave conditions
the algorithm assess the most relevant partitions, without critically de- (GLOSWAC), see Portilla-Yandún (2018). These results are available at
pending on tuning parameters. Moreover, in this method, the labeling https://modemat.epn.edu.ec/nereo/.
of partitions as wind-sea or swell is relieved, but it is considered as an
additional attribute through the wave age parameter. 3.2. Machine learning regression model
Post-processing is carried out in the form of smoothing through the
convolution operation between the spectral matrix S, and the filter 𝜅, To predict responses, linear and boosted trees regression models are
as indicated in Eq. (2). used, from the machine learning package turicreate. These algorithms
are open source and developed in Python language. They can handle
∑
1 ∑
1
̂ 𝑗) = 𝜅(𝑚, 𝑛) × 𝑆(𝑖, 𝑗) =
𝑆(𝑖, 𝜅(𝑚, 𝑛)𝑆(𝑖 − 𝑚, 𝑗 − 𝑛) (2) large data sets and features used to train and test the models. The
𝑚=−1 𝑛=−1 process is schematized in Fig. 6, which shows that wave spectral
This step is used in combination with thresholding to avoid exces- parameters and actual vessel dynamic responses are split into training
sive smoothing. This means that all partitions with marginal energy (80%) and testing (20%) data. From the training data, we select specific
content (e.g., below 2% of the total), are combined to their closest features (i.e. Hs, Tp and 𝜃𝑝 ) of specific wave systems. These features
partition. More details of this algorithm are provided by Portilla et al. together with actual dynamic responses are used to optimize ML re-
(2009). gression models (to obtain regression coefficients). Once an optimum
model is found, its performance is evaluated using the test data. If the
3.1.1. Long-term spectral characterization model is satisfactory, it can be used to predict dynamic responses.
The characterization of long series of wave spectra informs us about Linear regression model
its long-term structure. In particular, which are the most common A dependent variable such as Roll (𝜙) can be estimated by a combi-
frequencies, directions, magnitudes, among others. Standard statis- nation of various independent variables 𝑥𝑖 (features) such as 𝐻𝑠1 , 𝐻𝑠2 ,
tical parameters such as the mean spectrum, can provide to some 𝑇 𝑝1 , 𝑇 𝑝2 , 𝜃𝑝1 , 𝜃𝑝2 , see Eq. (3). Where 𝛽0 is the intercept, 𝛽1 to 𝛽𝑛 are
extent that information. However, most of these parameters tend to the slope coefficients of the independent variables and 𝜖 is the model’s
smooth out important details of the spectral structure. In turn, it was error component.
found by Portilla-Yandún et al. (2015) that a more skillful parameter
𝛷 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ 𝛽𝑛 𝑥𝑖𝑛 + 𝜖 (3)
for elucidating spectral components is the ‘‘occurrence distribution of
partitions’’, see Fig. 1(a). To derive this distribution, each partition The intercept and slope coefficients are selected after minimizing
is represented by a pair of spectral coordinates, which in practice a residual sum of squares or loss function 𝑔(𝜷), see Eq. (4). Where
correspond to its peak frequency or corresponding Tp, and 𝜃𝑝 . Each 𝜙 are the actual values of the dependent variable. To this end, the
of these pair of coordinates becomes a point in the spectral domain, method of the gradient descent is applied to find the optimum values
and the overall counting of these points is the sought distribution. of the estimated parameters as shown in Eq. (5). Where 𝜂 is a stepsize
There are two main advantages of this parameter, the first is that it parameter (between 0 and 1) and ∇𝑔 is the gradient of the loss function.
has exactly the same (Tp, 𝜃𝑝 ) structure of the individual spectrum, so Convergence is achieved when ∇𝑔 approaches zero.
the correspondence is direct. The second is that the different clusters
∑
𝑁
( )2
formed are very distinct, for visibility, but also for identification of 𝑔(𝜷) = arg min 𝜙 𝑖 − 𝛷𝑖 (4)
the main wave systems at work. Fig. 1(a) shows this distribution 𝛽 𝑖=1
for the selected site in the Atlantic Ocean. The color scale in this
𝜷 (𝑖+1) = 𝜷 (𝑖) ± 𝜂∇𝑔(𝛽 (𝑖) ) (5)
case indicates the number of occurrences in each spectral grid point.
Furthermore, each of the clusters observed can be determined using Gradient boosted trees model
(again) the partitioning approach. These spectral clusters identify long- Decision trees can allow linear and non-linear interaction between
term spectral wave systems or wave families, and physically they input and output. As a simplified example, Fig. 7(a) shows that roll

5
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 6. Machine learning methodology.

Fig. 7. Illustration of the regression trees method using two independent variables. (a) Independent variables and actual responses. (b) Example of a decision tree.

responses can be characterized by two wave parameters, e.g. Hs and Cases of analysis
Tp. Colored points refer to intensity in roll responses, which can be Fig. 1 shows the 6 long-term wave systems found at the study
separated based on conditions applied to Hs and Tp, using a decision location in the North Atlantic. To assess the influence of the number
tree, see Fig. 7(b). Likewise, several splitting alternatives can be eval- of wave parameters and the type of regression model, various cases are
uated, and the one giving the less variance in the results is selected. analyzed, and they are summarized in Table 2.
This is the first weak learner 𝛷0 (𝒙) in Eq. (6), which basically predicts
4. Results
the mean value of the corresponding leaf node. To predict residuals,
every predicted response is subtracted from the corresponding actual
4.1. Spectral partitioning and feature engineering
response. These residuals are the leaf nodes of the next decision tree,
similar to the one shown in Fig. 7(b). The next weak learner 𝛷1 (𝒙) Using the spectral partitioning described in Section 3.1, wave spec-
follows Eq. (7), where the partial derivative can be computed from the trum parameters for each wave system were computed. Wave systems
loss function, see e.g. (4) and 𝜂 is again a stepsize. The second weak do not always occur simultaneously, see e.g. Figs. 1(d) and 2, which
learner will predict the residuals from the previous weak learner. To means that some records for some wave systems will be empty. To deal
𝜕𝑔
predict a response, the product 𝜂 × − 𝜕𝛷 will be added to Eq. (7). This with this issue the Hs values for those wave systems is set equal to zero,
1
procedure is repeated for the desired number of iterations (number of while the peak directions and frequencies are set to their mean values
trees or weak learners), which can be increased together with the depth estimated from historical data used to plot Fig. 1(a). This artificial
of a tree for better accuracy. For a defined number of trees, Eq. (6) can manipulation has no effect on the predicted responses but is a necessary
predict a dynamic response of a structure. artifice for the machine learning algorithms to make use of all available
data. Figs. 8(a)–8(c) show examples of time series for Hs, Tp and 𝜃𝑝 ,
𝛷𝐵 (𝒙) = 𝛷0 (𝒙) + 𝛷1 (𝒙) + 𝛷2 (𝒙) + ⋯ (6) for each of the six wave systems. Based on the peak periods observed
in 8(b) it is evident that each system may classify as wind sea or swell
𝜕𝑔 due to their natural evolution. In other words, each system has a strong
𝛷1 (𝒙) = 𝛷0 (𝒙) + 𝜂 × − (7)
𝜕𝛷0 local nature. Moreover, Fig. 8(c) illustrates marked specific directions
of each wave system.

6
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Table 2
Analysis cases.
Analysis Hs feature Tp feature Peak dir feature No. features Regression model
case No.
1 Hs1 Tp1 Dir1 3 linear
2 Hs1, Hs2 Tp1, Tp2 Dir1, Dir2 6 linear
3 Hs1, Hs2, Hs3, Hs4 Tp1, Tp2, Tp3, Tp4 Dir1, Dir2, Dir3, Dir 4 12 linear
4 All Hs All Tp All Dir 18 linear
5 Hs1 Tp1 Dir1 3 boosted trees
6 Hs1, Hs2 Tp1, Tp2 Dir1, Dir2 6 boosted trees
7 Hs1, Hs2, Hs3, Hs4 Tp1, Tp2, Tp3, Tp4 Dir1, Dir2, Dir3, Dir 4 12 boosted trees
8 All Hs All Tp All Dir 18 boosted trees

Fig. 8. Examples of time histories for features, and roll responses for various headings. (a) Hs features. (b) Tp features. (c) Peak direction features. (d) Actual roll responses for
some vessel headings.

Using RAOs from Fig. 3, and 2D hindcast spectra, for some heading when using 6 and 12 features (e.g. spectral parameters from 2 and 4
of the vessel, roll and VBMs were computed using spectral analysis, see wave systems), and are very similar to the actual values for 18 features,
e.g. Fig. 8(d). see Fig. 9(d). Similar findings are observed for VBMs, see Fig. 10.
Errors are obtained by subtracting predicted responses from actual
4.2. Assessment of machine learning regression models values. For the case with 18 features, from actual ones, Figs. 11(a) and
11(b) shows an example of error time histories for roll and VBM re-
For each analysis case listed in Table 2, the performance of each spectively, while their corresponding statistics are shown in Figs. 11(c)
regression model for roll responses is summarized in Table 3. For the 11(d). It can be observed that the bias and the interquartile range
test data, the performance of case 8, using a ‘‘boosted trees regres- increases when the magnitude ranges increase. This means that for low
sion model’’ and 18 wave features produces the smallest errors, and response values (typical for marine operations) the bias and interquar-
therefore is the preferred model in this paper. tile range will be much smaller, see Section 5.

4.3. Predicted responses 5. Discussion

To assess the performance of the regression models for the features During the planning phase of a given marine operation, thousands of
listed in Table 2, Fig. 9 compares actual and predicted roll responses for complex numerical simulations can be conducted for various operation
90 deg heading. As expected, when using features from one wave sys- scenarios, but this can be prohibitive during the execution phase, espe-
tem only (3 features), the predicted responses are significantly different cially for wave conditions different from those used during the planning
from the actual ones, see Fig. 9(a). Predictions substantially improve phase. For weather-restricted MOs (those with durations of less than

7
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Table 3
Assessment of regression models using roll test data.
Analysis No. features Regression model max error rms error
case [deg] [deg]
1 3 linear 31.3 4.3
2 6 linear 29.7 3.8
3 12 linear 32.5 3.0
4 18 linear 16.5 2.4
5 3 boosted trees (iterations=20, depth=20) 29.0 4.0
6 6 boosted trees (iterations=20, depth=20) 25.4 3.1
7 12 boosted trees (iterations=20, depth=20) 22 1.8
8 18 boosted trees (iterations=20, depth=20) 12.5 1.2

Fig. 9. Example of predicted roll responses for 90 deg vessel heading respect to the true North. (a) Using 3 features: Analysis cases 1 and 4 (b) Using 6 features: Analysis cases
2 and 5. (c) Using 12 features: Analysis cases 3 and 6. (d) Using 18 features: Analysis cases 4 and 8.

72 h), the state of-the-art marine operation standards such as DNV-OS- or stresses, or carrying out complex numerical simulations, in real
H101 (2011), recommend the use of forecasted Hs as design parameter time can be difficult or impractical. Therefore, it is generally preferred
(modified by alpha factors), which is compared with its corresponding to simply compare the operational limits in terms of environmental
operational limit for assessing weather windows. However, offshore parameters (e.g. Hs and Tp) with the corresponding forecasted values.
vessels are sensitive to the wave period, but this parameter is not taken Dynamic responses can differ significantly if computed with total in-
into consideration. Recent investigations for assessing operational lim- tegral wave parameters (i.e Hs, Tp, 𝜃𝑝 ) or with specific wave systems
its included Tp using directional JONSWAP spectra, see Guachamin parameters as in the present study. To date, either these parameters or
Acero et al. (2016). These limits in terms of Hs and Tp were shown the full wave spectrum can be provided by meteorological institutes.
to be significantly different when directional hindcast spectra were In such cases, efficient algorithms can be used to predict dynamic
applied (Guachamin-Acero and Li, 2018). This fact is also evident in responses for each sea state during the execution phase. In this way,
Fig. 4, where dynamic responses computed from 2D JONSWAP and weather windows can be established more precisely to assist decision
hindcast spectra are sometimes significantly different. Nowadays, most making.
meteorological institutes are capable to provide hindcast and forecasted Wave parameters (i.e Hs, Tp, 𝜃𝑝 ) from all wave systems have shown
2D wave spectra. Thus, it is possible to compute accurate responses by to improve prediction of machine learning algorithms. When using 18
using the already available directional wave spectra. features (for 6 wave systems), predicted responses are close the actual
As it is recognized both by standards and by the offshore industry, ones, see Figs. 9(d) and 10(d). In fact, errors are smaller for lower
a vessel’s captain or superintendent needs practical and simple infor- responses, and they increase as the level of actual response increase,
mation to make quick and safe decisions. Monitoring vessel responses see Fig. 11.

8
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 10. Example of predicted VBM responses for 0 deg vessel heading respect to the true North. (a) Using 3 features: Analysis cases 1 and 4 (b) Using 6 features: Analysis cases
2 and 5. (c) Using 12 features: Analysis cases 3 and 6. (d) Using 18 features: Analysis cases 4 and 8.

To improve the accuracy of machine learning algorithms, the train- 5.1. Practical implementation of the methodology for marine operations
ing data can be selected based on the season. For example, if a marine
operation is planned for the month of June, the prediction models The results of this study can be incorporated into an on-board
can be trained with computed responses corresponding to that month. system to help field engineers and superintendents making decisions.
This is because for a given month not all wave systems necessarily The procedure can be summarized as follows:
occur at the same time and some of them will dominate, and thus, During the planning phase. These activities can be conducted at office
the coefficients for linear regression will allow a better fit, and the (on-land).
decision trees will predict more accurate responses for a given depth
• Step 1: Obtain hindcast 2D wave spectra for the offshore site where
of the tree and number of iterations. Fig. 12(a) shows time histories
the marine operations are to be executed.
of Roll responses and their corresponding errors for the months of
• Step 2: Conduct wave partition analysis to identify wave parameters,
June during the period 1979–2015. Using the boosted trees regression i.e. Hs, Tp, 𝜃𝑝 of every wave system.
model, the maximum error is 4.07 deg and the rms error is 0.79 deg,
• Step 3: Compute statistics of response parameters (e.g. significant
which are smaller than the corresponding values shown in Table 3, values of roll, heave, etc.) that limit an operation for planned head-
which means that including seasonality can improve the prediction. ings of the vessel using wave data from step 1. Time-domain and
For marine operations, the operational roll limits are generally less frequency-domain methods can be conducted.
than 2 deg (Clauss et al., 1990), and thus, error statistics are assessed • Step 4: Create prediction models using boosted trees regression. Se-
for lower roll interval limits. Fig. 12(b) shows that the error bias and lect between 70 to 80% of wave data for training a prediction model.
the interquartile range are small especially for low roll limits. These According to this paper, the boosted trees regression model with all
statistical parameters can be used for reliability analysis of MOs. wave parameters of each wave system can be used as features. Assess
Regarding the computational costs of the models used in this paper, the prediction model using the remaining 30 or 20% of the data.
the prediction using hundreds of wave records, such as those shown in
Fig. 11(a) can be done within a few seconds in an ordinary laptop. In During the execution phase. These activities are conducted offshore
turn, complex time-domain simulations, involving transients and non- (on-board a vessel).
linearities, which are common in marine operations, can take several • Step 5: Obtain forecasted 2D wave spectra. These are generally given
hours. In those cases, machine learning regression models offer an every one, three or six hours for the next 7 days. Alternatively, wave
effective and accurate alternative to predict dynamic responses. radar or directional buoys can be used to obtain wave spectra.

9
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Fig. 11. Roll and VBM responses for 90 deg and 0 deg vessel heading respectively. (a)Actual roll, predicted roll (using boosted trees regression with 18 features) and roll error
time history examples. (b) Actual VBM, predicted VBM (using boosted trees regression with 18 features) and VBM error time history examples. (c) Statistics for roll error. (d)
Statistics for VBM error.

Fig. 12. Roll responses for 90 deg vessel heading during June (1979–2015). (a)Time histories of actual and predicted (using boosted trees regression with 18 features) responses,
and error. (b) Error statistics for low limits of roll.

10
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

• Step 6: Conduct wave partitioning and obtain wave parameters for Acknowledgments
each wave system.
• Step 7: Predict dynamic responses for every forecasted spectrum. Wave spectra were downloaded from the ECMWF website (https://
Using wave parameters from step 6 as an input to the prediction www.ecmwf.int/). Long term spectral wave parameters were obtained
model obtained in step 4. through the GLOSWAC web site (https://modemat.epn.edu.ec/nereo/).
• Step 8: Compare dynamic responses with operational limits, iden- The authors acknowledge funding from the project EPN-PIGR-1908.
tify weather windows, and decide whether to start/stop or not an
operation. References
The methodology proposed in this paper is simple and easy to
Bassam, A.M., Phillips, A.B., Turnock, S.R., Wilson, P.A., 2022. Ship speed prediction
implement on-board. Such a system can change the actual use of
based on machine learning for efficient shipping operation. Ocean Eng. 245,
forecasts to optimize marine operations and increase safety. 110449. http://dx.doi.org/10.1016/j.oceaneng.2021.110449.
Bjørni, F., Lien, S., Midtgarden, T.A., Kulia, G., Verma, A., Jiang, Z., 2021. Prediction
of dynamic mooring responses of a floating wind turbine using an artificial neural
6. Conclusions and recommendations
network. IOP Conf. Ser. Mater. Sci. Eng. 1201 (1), 012023. http://dx.doi.org/10.
1088/1757-899X/1201/1/012023.
6.1. Summary conclusions Cavaleri, L., Alves, J.-H., Ardhuin, F., Babanin, A., Banner, M., Belibassakis, K.,
Benoit, M., Donelan, M., Groeneweg, J., Herbers, T., et al., 2007. Wave modelling–
the state of the art. Prog. Oceanogr. 75 (4), 603–674. http://dx.doi.org/10.1016/
This paper introduces a methodology to improve predictions of j.pocean.2007.05.005.
vessels dynamic responses during marine operations. The method uses Century Dynamics-ANSYS Inc., 2011. AQWA reference manual version 14.0.
wave parameters of actual directional wave spectra and machine learn- Century Dynamics Limited, 2002. Aqwa FER Manual. Dynamic House, 86 Hurst Road,
ing regression models. Horsham, West Sussex.
Clauss, G., Riekert, T., et al., 1990. Operational limitations of offshore crane vessels.
A state of-the-art wave spectral partitioning algorithm was applied
In: Offshore Technology Conference. http://dx.doi.org/10.4043/6217-MS.
to assess wave parameters (i.e Hs, Tp and peak direction) of each wave Dee, D.P., Uppala, S.M., Simmons, A., Berrisford, P., Poli, P., Kobayashi, S., Andrae, U.,
system (multimodal 2D wave spectra) present at the study location. Balmaseda, M., Balsamo, G., Bauer, d.P., et al., 2011. The ERA-interim reanalysis:
This scheme was applied to a record spanning 37 years of data, of Configuration and performance of the data assimilation system. Q. J. R. Meteorol.
a North Atlantic offshore site, where 6 wave systems are detected. Soc. 137 (656), 553–597. http://dx.doi.org/10.1002/qj.828.
DNV-OS-H101, 2011. Offshore Standard DNV-OS-H101 Marine Operations, General.
Their corresponding wave parameters were used as input for regression
Norway, Det Norske Veritas.
models in machine learning algorithms. Doong, D.-J., Peng, J.-P., Chen, Y.-C., 2018. Development of a warning model for
Regression models from machine learning were trained and opti- coastal freak wave occurrences using an artificial neural network. Ocean Eng. 169,
mized to predict responses. A boosted trees model with 18 features (for 270–280. http://dx.doi.org/10.1016/j.oceaneng.2018.09.029.
6 wave systems) was able to predict accurate roll and vertical bending Gracia, S., Olivito, J., Resano, J., Martin-del Brio, B., de Alfonso, M., Álvarez, E.,
2021. Improving accuracy on wave height estimation through machine learning
moment responses of a vessel. The performance of the models improved techniques. Ocean Eng. 236, 108699. http://dx.doi.org/10.1016/j.oceaneng.2021.
by increasing the number of features and seasonality. 108699.
The present results are promising for the development of on-board Guachamin-Acero, W., Li, L., 2018. Methodology for assessment of operational limits
systems, especially for demanding marine operations where accurate including uncertainties in wave spectral energy distribution for safe execution
of marine operations. Ocean Eng. 165, 184–193. http://dx.doi.org/10.1016/j.
estimate of dynamic responses is required for on-board decision mak- oceaneng.2018.07.032.
ing. Guachamin Acero, W., Li, L., Gao, Z., Moan, T., 2016. Methodology for assessment
of the operational limits and operability of marine operations. Ocean Eng. 125,
308–327. http://dx.doi.org/10.1016/j.oceaneng.2016.08.015.
6.2. Recommendations and future work
Guachamin-Acero, W., Portilla-Yandún, J., 2021. A study on vessel fatigue damage
as a criterion for heading selection by application of 2D actual bimodal and
The performance of regression models can be assessed including JONSWAP wave spectra. Ocean Eng. 226, 108822. http://dx.doi.org/10.1016/j.
other environmental variables such as wind speed and current. oceaneng.2021.108822.
Hasselmann, S., Bruning, C., Hasselmann, K., Heimbach, P., 1996. An improved
Assess reliability of marine operations. Uncertainties can include
algorithm for the retrieval of ocean wave spectra from synthetic aperture radar
errors form regression models and forecasted wave parameters for var- image spectra. J. Geophys. Res. Oceans 101 (C7), 16615–16629, http://hdl.handle.
ious lead times. This is necessary to develop more robust and reliable net/21.11116/0000-0008-89EF-1.
prediction models. Hasselmann, K.F., Chapron, B., Aouf, L., Ardhuin, F., Collard, F., Engen, G., Hassel-
mann, S., Heimbach, P., Janssen, P., Johnsen, H., et al., 2013. The ERS SAR Wave
Mode: a Breakthrough in Global Ocean Wave Observations. Vol. 1326. ESA/ESTEC,
CRediT authorship contribution statement Special Publication, pp. 167–197.
Hasselmann, S., Hasselmann, K., Janssen, P., 1988. The WAM model - a third generation
ocean wave prediction model. J. Phys. Oceanogr. 18 (12), 1775–1810. http://dx.
Wilson Guachamin-Acero: Writing, Paper structure, Methodology,
doi.org/10.1175/1520-0485(1988)018<1775:TWMTGO>2.0.CO;2.
Numerical models, Machine learning prediction models. Jesús Por- Holthuijsen, L.H., 2007. Waves in Oceanic and Coastal Waters. Cambridge University
tilla: Generating input wave data, Wave partitioning, Reviewing, Proof Press.
reading. International Organization for Standardization, 2015. Ships and Marine Technology-
Offshore Wind Energy-Port and Marine Operations ISO 29400.
Komen, G.J., Cavaleri, L., Donelan, M., Hasselmann, K., Hasselmann, S., Janssen, P.,
Declaration of competing interest 1994. Dynamics and Modelling of Ocean Waves. p. 532, ISBN 0-521-47047-1.
Li, L., 2016. Dynamic Analysis of the Installation of Monopiles for Offshore Wind
The authors declare that they have no known competing finan- Turbines (Ph.D. Thesis). NTNU, Trondheim, Norway.
Li, L., Gao, Z., Yuan, Z.-M., 2019. On the sensitivity and uncertainty of wave energy
cial interests or personal relationships that could have appeared to
conversion with an artificial neural-network-based controller. Ocean Eng. 183,
influence the work reported in this paper. 282–293. http://dx.doi.org/10.1016/j.oceaneng.2019.05.003.
Modemat, 2020. The global signature of ocean wave spectra. http://https://modemat.
epn.edu.ec/nereo/. (Accessed 09 December 2020).
Data availability
Nwogu, O., 1989. Maximum entropy estimation of directional wave spectra from an
array of wave probes. Appl. Ocean Res. 11 (4), 176–182. http://dx.doi.org/10.
Data will be made available on request. 1016/0141-1187(89)90016-3.

11
W. Guachamin-Acero and J. Portilla Ocean Engineering 262 (2022) 112157

Portilla, J., Ocampo-Torres, F.J., Monbaliu, J., 2009. Spectral partitioning and iden- Smit, P., Houghton, I., Jordanova, K., Portwood, T., Shapiro, E., Clark, D., Sosa, M.,
tification of wind sea and swell. J. Atmos. Ocean. Technol. 26 (1), 107–122. Janssen, T., 2021. Assimilation of significant wave height from distributed ocean
http://dx.doi.org/10.1175/2008JTECHO609.1. wave sensors. Ocean Model. 159, 101738. http://dx.doi.org/10.1016/j.ocemod.
Portilla-Yandún, J., 2018. The global signature of ocean wave spectra. Geophys. Res. 2020.101738.
Lett. 45 (1), 267–276. http://dx.doi.org/10.1002/2017GL076431. Tapoglou, E., Forster, R.M., Dorrell, R.M., Parsons, D., 2021. Machine learning for
Portilla-Yandún, J., Cavaleri, L., Van Vledder, G.P., 2015. Wave spectra partitioning satellite-based sea-state prediction in an offshore windfarm. Ocean Eng. 235,
and long term statistical distribution. Ocean Model. 96, 148–160. http://dx.doi. 109280. http://dx.doi.org/10.1016/j.oceaneng.2021.109280.
org/10.1016/j.ocemod.2015.06.008.

12