1 Methodology for prediction of wave spectral parameters to support

2 execution of marine operations using multiple-output regression

3 analysis

4 Jonathan Prócel1
1∗
Wilson Guachamin-Acero
*1
Departamento de Ingenierı́a Mecánica, Escuela Politécnica Nacional
Quito, Ecuador 17-01-2759

5 March 2, 2023

6 Abstract

7 Execution of a marine operation (MO) requires coordinated actions of several vessels conducting
8 simultaneous and sequential offshore activities. From all environmental actions, waves are generally
9 the most important because of their high energetic level. Forecast wave spectral parameters are used
10 for planning and execution of MOs. For critical operations, the use of forecast wave spectral param-
11 eters can be complemented with buoy measurements. This paper proposed to use synthetic statistics
12 of vessel dynamic responses to predict wave spectral parameters using multi-output machine learning
13 (ML) regression algorithms, observations and long term spectral characteristics. It is observed that
14 random forest regression has good performance for practical purposes. Moreover, the performance
15 of the random forest models outperforms other deep learning models reported in scientific literature.
16 Findings from this research can be valuable for ”real time” assessment of wave spectral parameters,
17 which are necessary to support decision making during execution of marine operations.
18 Keywords: Machine learning, wave spectra partitioning, marine operations, regression model,
19 boosted trees algorithm.

*
Corresponding author. Tel: +593 981286528 ; Email: [email protected]

1
20 1 Introduction

21 The execution phase of marine a marine operation (MO) (e.g. topside installation) is a complex set

22 of sequential and simultaneous activities from one or various vessels (positioning of transport barge,

23 topside lift-off, moving topside to location, lowering topside, mating process). All these activities have

24 their own operational limits, which can be given in terms of motion responses or environmental param-

25 eters such as wave spectral parameters, and wind and current speed. Operational limits are compared

26 with motion responses (e.g., computed using forecasted wave parameters) to identify workable weather

27 windows prior to start an activity (Guachamin-Acero and Li, 2018). Thus, accurate forecast of wave

28 parameters is necessary to make on-board decisions and decide whether to start or not an operation.

29 To support decision making during execution of critical MOs, apart from wave forecasts, it may

30 be necessary to conduct wave measurements using buoys. This is typically the case for float-over and

31 heavy lifting operations. Considering the vessel as a wave buoy, several researchers have proposed

32 various methodologies to predict real-time wave spectral parameters i.e, significant wave height (Hs),

33 peak period (Tp) and wave direction. The term ”real-time” considers the time lag (e.g. less than 8

34 minutes) between the actual estimate of sea state parameters and measurements of the vessel responses

35 (Pascoal and Soares, 2009).

36 To predict sea state parameters, there are basically two methods: model-based and data-driven.

37 In general, model-based methods rely on the response amplitude operator (RAO). The wave spectral

38 ordinates can be computed from cross response spectra and the RAO of a vessel. For instance, Nielsen

39 (2006) predicted the directional (2D) wave spectra from vessel responses using parametric (e.g. JON-

40 SWAP) and Bayesian models, which performed well when compared with the wave spectra from radar

41 system. To reduce the time lag and the non-stationarity of sea states, several authors preferred to

42 analyze time histories of vessel responses directly in the time-domain. As the output is the actual

43 wave surface elevation, the method computes responses of the vessel using the RAOs and irregular

44 wave amplitude parameters in the state-space. These parameters are updated based on gains set in

45 algorithms such as the Kalman filter until the responses match the measured ones. Thus, the resulting

46 surface elevation can be converted into a wave spectrum. Nielsen et al. (2015) applied the Arnovsky

47 filter (Aranovskiy and Bobtsov, 2012) to asses the frequency of the vessel responses, whose amplitudes

48 and phases were determined using a non-linear least squares fitting, and by applying the inverse of

2
49 the RAO the wave elevation was reconstructed. In a similar fashion, Pascoal et al. (2017) assessed the

50 directional wave spectrum of an oceanographic vessel using data from seaborn sensors and the Kalman

51 filter.

52 Regarding data-driven models, most of them are based on time histories of vessel responses, which

53 are preprocessed and segmented in shorter duration samples. Algorithms for feature extraction e.g.,

54 pattern recognition and labelling are applied. Training and testing are conducted by applying su-

55 pervised and unsupervised machine learning algorithms, which can be used to assess wave spectral

56 parameters. Using amplitudes and phase differences of two-years records of 6 degrees of freedom

57 (DOF) motions of a frigate type vessel, Mak and Düz (2019) applied convolutional neural networks

58 (CNN) and recurrent neural networks (RNN) for multivariate regression of time series of measured

59 responses to assess relative direction of waves. Cheng et al. (2020a) used a deep neural network al-

60 gorithm (SEENET) to predict wave height and directions. SEENET uses CNN blocks to assess and

61 fusion hierarchical frequency features of segments of response time series of a vessel with forward speed.

62 Later ,Cheng et al. (2020b) developed a deep learning model (SpectralSeaNet), which uses 2D CNNs

63 as a classifier of spectrograms of segmented time series of shipborne sensors. High classification accu-

64 racy is achieved increasing the number of shipborn sensors, and including correlation and combination

65 of spectral images of the segmented time series. Then, the sea state parameters are predicted using

66 CNN. In general, SpectralSeaNet performs better than other algorithms based on CNN and Long Term

67 Short Memory (LTSM). Moreover, Mittendorf et al. (2022) compared the performance of frequency-

68 and time-domain models using 1.5 years time histories of the 6 DOF of a container vessel transiting in

69 the Northern Atlantic Ocean. Comparing with radar measurements, the former model predicted more

70 accurate spectral parameters (form spectral ordinates) than the latter.

71 Based on the above, several algorithms exist for prediction of wave spectral parameters based on

72 frequency- and time-domain analyses. In the frequency domain, the spectral ordinates are assessed

73 from the RAOs and the cross response spectra. The analyses in the time-domain allows for shortening

74 the time lag between prediction of sea parameters and measurement of vessel responses, but require

75 more computational effort, especially for deep neural network algorithms. In this paper, we propose

76 the use of large 2D hindcast wave spectra data sets to compute vessel response statistics, e.g. root

77 mean squared (rms). These responses can be used to train a multi-output regression models. Since

78 most of marine construction vessels are equipped with seaborn sensors, it is possible to use rms value

3
79 of these responses to asses sea state parameters (i.e. Hs, Tp and wave direction) of dominant wave

80 systems. Information on estimated parameters can be supported by wave climate data and observations

81 to support decision making on-board offshore construction vessels.

82 2 Methodology

83 This section describes the methodology for prediction of sea state parameters (i.e. Hs, Tp, wave

84 direction) using a machine learning multi-output regression algorithm, which is trained using directional

85 (2D) hindcast actual wave spectra and vessel dynamic parameters.

3. Features:
2. Training 4. Machine
vessel
data learning
reponse
1. Data set: regression model
statistics
wave spectral
parameters
and vessel
response
statistics 5. Test data

Assessment of
Regression model
performance
10. Final
6. Select a assessment of Hs,
machine learning Tp, wave direction
algorithm

8. Predicted wave 9. Long-term wave

Figure 1: Machine learning methodology

86 2.1 Wave data set

87 Vessel response statistics, e.g., standard deviation (rms) or significant responses (2 x rms) can be cal-

88 culated from actual 2D hindcast wave spectra and vessel transfer functions. These spectra is obtained

4
 89 from reanalysis data, nowadays available worldwide e.g., at the National Oceanic and Atmospheric

90 Administration (NOAA) or the European Center for Medium Weather Forecast (ECMWF), which

91 use third generation numerical models to asses 2D wave spectra. The ERA-Interim database from

92 ECMWF 2D hindcast wave spectra for most offshore sites. These records span long periods, see e.g.,

93 (Dee et al., 2011). In this paper, we use a dataset for the North Atlantic Ocean (47 N 336.58 E) that

94 spans over 37 years from January 1979 to December 2015. Figure 2 shows long term wave statistics of

95 wave six wave systems (wind seas and swells), including mean average Hs, peak frequencies and direc-

96 tion, being systems 1 and 2 the most energetics ones, see figure 2(b). Figure 2(a) shows for instance

97 that system 3 is more likely a swell because its energy is highly concentrated at low frequencies a a

98 narrow directional band. This information will be valuable when assessing wave spectral parameters

99 en Sec..

(a) (b)

Figure 2: Wave system components for North Atlantic Ocean 47 N 336.58 E. (a) Occurrence distri-
bution of wave systems. (b) Mean value of monthly averaged Hs for each wave system. (Modemat
(2020), https://modemat.epn.edu.ec/nereo/)
.

100 2.2 Vessel response statistics data set

101 In the literature review it was observed that actual time series of vessel responses were recorded during

102 trips. The benefit or reading actual responses is to reduce uncertainties, especially for the transfer

103 functions (Majidian et al., 2022). In this paper, we consider that actual RAOs of heavy lift vessels

104 are available or else, they can be obtained from model tests, and used for tuning numerical models.

5
105 Table 1 shows the main particular of the vessel studied in this paper. Figure 3(a) shows an example

106 of the RAO, which though spectral analysis is converted into time histories of response statistics, see

107 e.g. Fig. 3(b). These time histories can be easily obtained at office, months ahead of executing a MO.

108 If the MO involves a nonlinear numerical model or a transient simulation (nonstationary process), the

109 time-domain method should be used. However, if the process is stationary process or the system is

110 linear, frequency-domain methods can be applied. In this paper, we select the response statistics from

111 6-DOF, and the motion of a point with coordinates (X= -25, Y= -11.5, Z= -22) m, typical for a jack-up

112 leg, measured from the center of gravity.

Table 1: Main particulars of the construction vessel (Guachamin-Acero and Li, 2018)
Parameter Notation Value Units
Displacement 5 2.55 × 104 tons
Length L 140 m
Breadth B 30 m
Draught T 6 m
Metacentric height GM 7.5 m
Vertical position of COG above keel V CG 8 m

7 Roll
Pich
6 Heave

5
4xrms [deg]

4
3
2
1
0
3 -05 7 9 1 1 3 5
9-0 9 9-0 979-0 979-1 980-0 980-0 980-0
197 197 197 1 1 1 1 1
Date

(a) (b)

Figure 3: Examples of vessel RAO and response time histories. (a) Roll RAO. (b) Example of motion
response statistics time histories.

6
113 2.3 Training data and feature selection

114 The sea-born sensors installed on board a vessel generally include the Motion reference unit (MRU)

115 that measure the motions of the 6-DOF of the COG, and other accelerometers used to monitor e.g.,

116 lifting points, sea-fastening components, etc. At first, all these parameters can be used to generate

117 synthetic response statistics and then, a correlation matrix can be used to select those with higher

118 correlation coefficients. Figure 4 shows a matriz including the following parameters: Surge (X), sway

119 (Y), heave (Z), roll (Rx), pitch (Ry), yaw (Rz), Hs, Tp, cos(wave dir) DirC and sin(wave dir) DirS. The

120 last two parameters are selected instead of the wave direction to avoid circular ambivalence (Mittendorf

121 et al., 2022). Except for DirC and DirS, all response and wave parameters have important correlation

122 coefficients. Parameters with significant correlation coefficients are consdiered for feature selection.

123 All data can be split in training and test data, typically, 80 and 20 % respectively. The above can be

124 applied to speciofic set of data, e.g. summer or winter season, in order to improve the performance of

125 the ML regression algorithms.

126 2.4 Machine learning regression algorithm

127 In this paper, we asses the performance of the multi-output and random forest regression models.

128 2.4.1 Multi-output linear regression model

129 The prediction of multiple variables can be conducted from equation (2).

Y =X ×W +b+e (1)

130 Where Y is a m x n output matrix which contains n rows for (Hs, Tp and wave direction) and m

131 columns for for the number of records. X is a matrix m x p that contains the response statistics of the

132 features. W is a matrix p x n with linear regression coefficients. Moreover, b is a vector 1 x n that

133 includes slope coefficients, and e is an m x n error or noise matrix. The coefficients from W and b are

134 computed by minimizing the residual squared error. To reduce the error and to avoid overfiting the

135 squared of the Frobenius (F) norm is applied. A regularization type L2 (Ridge) is applied via the α

136 parameter.

7
Figure 4: Example of correlation matrix used for feature selection

8
 minW,b = ||X × W + b + e||F 2 + α||W ||F 2 (2)

137 2.4.2 Random forest regression model

138 A technique that combines the result from various decision trees. In fact, the prediction is the average

139 from the independent trees.

1 (b)
ŷij = Σb = 1B fj (xi ) (3)
B

140 Where, ŷij is the predicted variable j for an input i. B is the number of decision trees, fj is the

141 prediction algorithm for tree b, and xi is the input variable. Each tree is constructed using a bagging

142 technique, which trains each tree with the output variables and their characteristics.

143 2.5 Algorithm evaluation metrics

144 To assess the performance of the ML algorithms, we use two metrics. The mean squared error (MSE) is

145 defined by equation (4). Similarly, the mean absolute error MAE is defined by equation (5). While both

146 MSE and MAE approach zero, the model prediction is accurate. If MAE and MSE differ significantly,

147 over-fitting may occur. Additionally, the squared root of the mean squared error(RMSE) can be used

148 as a regression metric. The algorithm that gives the smallest error should be selected for prediction.

N
1 X 2
M SE = Yi − Ŷi (4)
N
i=1

N
1 X
M AE = |Yi − Ŷi | (5)
N
i=1

149 2.6 Cases of analysis

150 Figure 2 shows the 6 long-term wave systems found at the study location in the North Atlantic. To

151 assess the influence of the number of wave parameters and the type of regression model, various cases

152 are analyzed, and they are summarized in Table 2.

9
 Table 2: Analysis cases with feature input parameters
COG responses Node responses Spectral parameters No. fea-
tures
Analysis X Y Z Rx Ry Rz Xp Yp Zp Hs Tp cos(θp ) sin(θp ) Regression model
case
1 x x x x x x x 3 linear/random
forest
2 x x x x x x x x 4 linear/random
forest
3 x x x x x x x x x 5 linear/random
forest
4 x x x x x x x x x x 6 linear/random
forest
5 x x x x x x x x x x x x x 9 linear/random
forest

153 3 Results

154 3.1 Assessment of machine learning regression models

155 Table 3 summarizes the performance of the regression models listed in Table 2. Using the test data,

156 the best performance corresponds to the random forest regression model and 6 features (case No.4).

157 It is observed that the error reduces while increasing the number of features.

158 3.2 Predicted sea state parameters

159 Figure 5 shows examples of time histories for Hs, Tp and θp using the random forest regression model.

160 From a qualitative perception, the prediction model performs well for Hs and Tp and has a regular

161 performance for θp .

162 For a quantitative assessment of the random forest regression model, we compute RMSE for various

163 ranges of Hs, Tp and θp . Figure 6 shows for a given range or Hs, e.g., Hs < 6 m (see Fig. 6(a)), the

164 corresponding values of Tp and θp can be identified, see figures 6(b), 6(c) and 6(d). Figure 7 shows

165 boxplot diagrams for the errors of predicted variables. Figure 7 shows error statistics for various ranges

166 of Hs. For Hs < 3 m, it can be observed that 99 % of the errors lie within a range of approximately

167 0.24 m. For those Hs, 99 % of the corresponding errors for Tp lie within ± 1 s. Correspondingly the

168 directions are within ± 50 deg. From a practical point of view, the prediction of spectral parameters

169 are reasonable, especially when considering that θp can be verified or corrected with observations on-

170 board. Moreover, the performance of the random forest model is assessed by comparing its metrics

10
 Table 3: Assessment of regression models
Regression model
Linear multi-output Random Forest
Analysis Number of Output param- Max. RMSE MAE Max. RMSE MAE
case features eters error error
1 3 Hs (m) 0.65 0.02 0.11 0.89 0.02 0.10
Tp (s) 9.69 3.27 1.40 10.16 2.74 1.78
sin(θp ) 1.93 0.27 0.40 1.96 0.25 0.34
cos(θp ) 1.61 0.38 0.52 1.92 0.39 0.50
2 4 Hs (m) 0.65 0.02 0.10 0.88 0.01 0.09
Tp (s) 11.12 2.62 1.20 8.50 1.26 0.74
sin(θp ) 2.13 0.26 0.38 1.97 0.23 0.31
cos(θp ) 2.05 0.37 0.50 1.92 0.34 0.45
3 5 Hs (m) 0.61 0.02 0.10 0.73 0.01 0.09
Tp (s) 11.86 2.58 1.19 7.40 1.20 0.72
sin(θp ) 1.78 0.24 0.36 1.97 0.22 0.30
cos(θp ) 1.78 0.36 0.49 1.94 0.32 0.42
4 6 Hs (m) 0.81 0.01 0.08 0.78 0.009 0.07
Tp (s) 10.11 2.45 1.15 7.36 1.15 0.71
sin(θp ) 2.13 0.25 0.36 1.98 0.21 0.29
cos(θp ) 1.80 0.35 0.48 1.85 0.30 0.41
5 9 Hs (m) 0.81 0.01 0.08 0.78 0.009 0.07
Tp (s) 10.11 2.45 1.15 7.36 1.15 0.71
sin(θp ) 2.13 0.25 0.36 1.98 0.21 0.29
cos(θp ) 1.80 0.35 0.48 1.85 0.30 0.41

11
 (a)

(b)

(c)

(d)

Figure 5: Examples of time series for sea state parameters for case analysis No. 4. (a) Hs (b) Tp (c)
sin(θp ) (d) cos(θp )
.

12
171 with those from deep learning models available in scientific literature, see Sec. 4.

(a)

(b)

(c)

(d)

Figure 6: Examples of time series for sea state parameters for Hs < 6 m. (a) Hs (b) Tp (c) sin(θp ) (d)
cos(θp )
.

13
 (a) (b)

(c) (d)

Figure 7: Error statistics for sea state parameters. (a) Hs (b) Tp (c) θp (rad) using asin(θp )(d) θp
(rad) using acos(θp )
.

14
172 4 Discussion

173 During the execution of a MO, several simultaneous or sequential activities have to be conducted using

174 one or various offshore vessels. In general, offshore vessels use forecasted environmental parameters to

175 compare with operational limits and decide whether or not to start an operation. For critical MOs,

176 e.g., float over operations, it is common the use of buoys to monitor the wave elevation or directly

177 to asses wave spectral parameters. Using the vessel as a buoy to predict wave spectral parameters in

178 almost real time is beneficial to support on-board decision making.

179 Using large data sets of hindcast wave spectra to assess statistics of vessel dynamic responses

180 (e.g., rms values) using a tuned RAO is straightforward, and can be done for any marine operation

181 or activity. Accurate RAOs for various loading conditions can be obtained from model tests and sea

182 trials well in advance during the planning phase of a marine operation. Thus, it is not necessary to

183 have long records of actual measured vessel responses unless the RAOs are not available. However, it is

184 necessary to have ship-born sensors to assess response statistics for parameters used to train machine

185 learning algorithms. RMS values for vessel responses are always available in almost real time using

186 records from the last 10-15 min, and can be used to predict actual sea state parameters.

187 To show the performance of machine learning regression models, Table 4 compares RMSE and

188 MAE with those corresponding from deep learning methods available in literature. Mittendorf et al.

189 (2022) used deep learning algorithms to predict wave spectral parameters using 1.5 years of in-service

190 measurements for the six DOF of a container vessel transiting in the North Atlantic Ocean. The

191 INCEPTION (a deep learning algorithm than uses CNNs) and Multi Task Learning (MTL) algorithms

192 were applied and the maximum error is not reported. Regarding Hs and wave direction, the random

193 forest regression model predicts far better than the deep learning models. In contrast, the deep learning

194 models predict more accurate values for Tp. In overall, the performance of the random forest regression

195 model is better than those from deep learning.

196 4.1 Use of wave climate and observations

197 Figures 5(c), 5(d), 7(c) and 7(d) show that prediction of wave direction is not always accurate. This

198 is because wave systems 1 and 2 are the most energetic all year round, see Fig. 2. This can be a

199 characteristic for offshore sites with various dominant wave systems. The predicted wave direction can

15
 Table 4: Assessment of the random forest regression model using 6 features
Sea state Random forest regression INCEPTION MTL MTL+
parameter
Max. RMSE MAE RMSE MAE RMSE MAE RMSE MAE
error
Hs (m) 0.78 0.009 0.07 0.425 0.268 0.341 0.179 0.324 0.171
Tp (s) 7.36 1.15 0.71 0.381 0.216 0.356 0.170 0.371 0.192
sin(θp ) 1.98 0.21 0.29 4.910 2.372 5.031 2.548 5.082 2.569
cos(θp ) 1.85 0.30 0.41 4.910 2.372 5.031 2.548 5.082 2.569

200 be verified or easily corrected with on-board observations.

201 4.2 Unimodal vs multimodal wave spectral parameters

202 Guachamin-Acero and Portilla (2022) compared statistics of vessel responses (e.g., 4 x RMS) using

203 actual 2D hindcast wave spectra and those corresponding using 2D unimodal JONSWAP spectra. They

204 showed that the errors increase for increasing ranges of dynamic responses. Obviously, it would be ideal

205 to have available on-board and in real time the actual 2D wave spectra, but it seems to be still difficult.

206 Several efforts have been made to assess the actual 2D spectra, see e.g., (Nielsen, 2006; Pascoal et al.,

207 2017), but the resolution seems to be not sufficient to compute accurate dynamic responses. This

208 is mainly because the short length of the time series of vessel responses used to reconstruct a wave

209 spectrum (e.g., less than 10 min), which is needed to assess the spectrum in real time. Based on the

210 above, research efforts focus on prediction of wave spectra parameters (i.e. Hs, Tp, θp ) to construct

211 an analytical 2D JONSWAP spectrum, which can be used for analysis of MOs. As stated earlier, this

212 model will add uncertainties in the dynamic responses, however, in absence of better alternatives, it is

213 highly valuable to support decision making on-board vessels.

214 Guachamin-Acero and Portilla (2022) showed that ML regression models can predict accurate

215 statistics of dynamic responses using all spectral parameters from all wave systems, instead of using a

216 2D actual spectra. However, prediction of all these spectral parameters is even more challenging, and

217 no related research has been reported in scientific literature.

16
218 5 Summary conclusions and Recommendations

219 5.1 Summary conclusions

220 This paper deals with an assessment of machine learning regression models to predict wave spectra

221 parameters, i.e., significant wave height (Hs), peak period (Tp) and wave direction using algorithms

222 trained with synthetic data of vessels responses statistics, which are computed using vessel transfer

223 functions and directional hindcast wave spectra from a location in the North Atlantic Ocean. The

224 models have a good performance, especially the multi-output random forest regressor.

225 The random forest regression model predicts better results than the multi-output linear regression

226 model. The precision increases by increasing the number of features. The best performance was

227 achieved using statistics of motion responses for the six degrees of freedom of the vessel and that

228 corresponding to the motions of a point node.

229 A benchmarking study showed that the random forest regressor can predict more accurate spectral

230 parameters than other deep learning models, except for Tp. However, the prediction is accurate enough

231 for practical use.

232 Results from this study are useful for developing simple but robust on-board systems to help

233 superintendents make a quick and safe decision.

234 5.2 Recommendations and future work

235 Prediction models for wave spectral parameters of 2D multimodal spectra need to be developed. This

236 is necessary for accurate assessment of dynamic responses.

237 Uncertainty in predicted wave spectra parameters should be included in operational limits by

238 including safety factors.

239 ACKNOWLEDGMENTS

240 Wave spectra were dowloaded from the ECMWF website (https://www.ecmwf.int/). Long term spec-

241 tral wave parameters were obtained trough the GLOSWAC web site (https://modemat.epn.edu.ec/nereo/).

242 The authors acknowledge funding from the project EPN-PIS-2105.

17
243 References

244 Aranovskiy, S., Bobtsov, A., 2012. Output harmonic disturbance compensation for nonlinear plant,

245 in: 2012 20th Mediterranean Conference on Control & Automation (MED), IEEE. pp. 386–391.

246 Https://doi.org/10.1109/MED.2012.6265668.

247 Cheng, X., Li, G., Ellefsen, A.L., Chen, S., Hildre, H.P., Zhang, H., 2020a. A novel densely connected

248 convolutional neural network for sea-state estimation using ship motion data. IEEE Transactions

249 on Instrumentation and Measurement 69, 5984–5993. Https://doi.org/10.1109/TIM.2020.2967115.

250 Cheng, X., Li, G., Skulstad, R., Zhang, H., Chen, S., 2020b. Spectralseanet: Spec-

251 trogram and convolutional network-based sea state estimation, in: IECON 2020 The

252 46th Annual Conference of the IEEE Industrial Electronics Society, IEEE. pp. 5069–5074.

253 Https://doi.org/10.1109/IECON43393.2020.9254890.

254 Dee, D.P., Uppala, S.M., Simmons, A., Berrisford, P., Poli, P., Kobayashi, S., Andrae, U., Balmaseda,

255 M., Balsamo, G., Bauer, d.P., et al., 2011. The ERA-Interim reanalysis: Configuration and perfor-

256 mance of the data assimilation system. Quarterly Journal of the Royal Meteorological Society 137,

257 553–597. Http://dx.doi.org/10.1002/qj.828.

258 Guachamin-Acero, W., Li, L., 2018. Methodology for assessment of operational limits including un-

259 certainties in wave spectral energy distribution for safe execution of marine operations. Ocean

260 Engineering 165, 184–193. Https://doi.org/10.1016/j.oceaneng.2018.07.032.

261 Guachamin-Acero, W., Portilla, J., 2022. Prediction of dynamic responses for execution of marine

262 operations using partitioning of multimodal directional wave spectra and machine learning regression

263 models. Ocean Engineering 262, 112157. Https://doi.org/10.1016/j.oceaneng.2022.112157.

264 Majidian, H., Wang, L., Enshaei, H., 2022. Part. a: A review of the real-time sea-state estimation, using

265 wave buoy analogy. Ocean Engineering , 111684Https://doi.org/10.1016/j.oceaneng.2022.111684.

266 Mak, B., Düz, B., 2019. Ship as a wave buoy: Estimating relative wave direction from in-service

267 ship motion measurements using machine learning, in: International conference on offshore me-

268 chanics and arctic engineering, American Society of Mechanical Engineers. p. V009T13A043.

269 Https://doi.org/10.1115/OMAE2019-96201.

18
270 Mittendorf, M., Nielsen, U.D., Bingham, H.B., Storhaug, G., 2022. Sea state identification using

271 machine learning—a comparative study based on in-service data from a container vessel. Marine

272 Structures 85, 103274. Https://doi.org/10.1016/j.marstruc.2022.103274.

273 Modemat, 2020. The Global Signature of Ocean Wave Spectra. http://https://modemat.epn.edu.

274 ec/nereo/. Accessed: 2020-12-09.

275 Nielsen, U.D., 2006. Estimations of on-site directional wave spectra from measured ship responses.

276 Marine Structures 19, 33–69. Https://doi.org/10.1016/j.marstruc.2006.06.001.

277 Nielsen, U.D., Bjerregård, M., Galeazzi, R., Fossen, T.I., 2015. New concepts for ship-

278 board sea state estimation, in: OCEANS 2015-MTS/IEEE Washington, IEEE. pp. 1–10.

279 Https://doi.org/10.23919/OCEANS.2015.7404386.

280 Pascoal, R., Perera, L.P., Soares, C.G., 2017. Estimation of directional sea spectra from ship motions

281 in sea trials. Ocean Engineering 132, 126–137. Https://doi.org/10.1016/j.oceaneng.2017.01.020.

282 Pascoal, R., Soares, C.G., 2009. Kalman filtering of vessel motions for ocean wave directional spectrum

283 estimation. Ocean Engineering 36, 477–488. Https://doi.org/10.1016/j.oceaneng.2009.01.013.

19