Ocean Engineering 109 (2015) 298–308

Contents lists available at ScienceDirect

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

Artificial neural networks applied to port operability assessment

I. López a,n, M. López b, G. Iglesias c
a
University of Santiago de Compostela, Hydraulic Engineering, EPS, Campus Universitario s/n, 27002 Lugo, Spain
b
Interdisciplinary Centre of Marine and Environmental Research (CIIMAR), Faculdade de Engenharia da Universidade do Porto (FEUP), Rua Dr. Roberto Frias,
s/n, 4200-465 Porto, Portugal
c
School of Marine Science and Engineering, Plymouth University, Plymouth PL4 8AA, UK

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

Article history: Waves are one of the main factors that can disturb port operations, from berthing to cargo loading and
Received 30 April 2015 unloading. Wave heights within port basins are typically assessed by means of numerical models based on the
Accepted 10 September 2015 outer (offshore) wave conditions, the bathymetry and the port layout. The aim of this work is to implement an
Available online 29 September 2015
artificial neural network (ANN) model which, upon training and validation, will be capable of determining
Keywords: wave agitation within a port basin based on deep-water wave buoy observations alone. In the training the ANN
Port of Ferrol model acquires knowledge on the problem from a series of examples, and thereafter applies this self-acquired
Wave agitation knowledge to other (new) cases. To select the ANN architecture most appropriate for this task a comparative
Harbour tranquillity study involving 65 options is carried out using the k-fold cross-validation technique. Upon validation, the ANN
Port operability
model is used to carry out a sensitivity analysis in which the influence of the different input variables on the
Artificial intelligence
wave parameters in the basin is quantified. Finally, the model is applied to a case study—the Exterior Port of
Artificial neural network
Ferrol—in order to evaluate wave agitation inside the basin and its influence on port operations.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction alternatives; nevertheless, they have numerous drawbacks: they are

expensive and time-consuming, and subject to laboratory and scale
One of the main functions of a harbour is to provide safe effects. Numerical models provide an approximate solution of the dif-
anchorage for vessels in order to facilitate the loading and ferential equations describing wave propagation. There are fundamen-
unloading operations of cargo and passengers (Goda, 2002). To tally two types of models: spectral models, which provide a phase-
perform these functions, the weather conditions inside the basin averaged description of wave propagation, e.g., SWAN (Rusu and
(wind, currents and, importantly, waves) must not exceed certain Guedes Soares, 2011, 2013); and phase-resolving models, including mild
thresholds, which depend on the type of operation (mooring, slope models (Chen et al., 2005; Hamidi et al., 2012) or Boussinesq
loading or unloading) and cargo involved (Elzinga et al., 1992; models (Bruno et al., 2009; Su et al., 2015). The main advantage of
PIANC, 1995; Puertos del Estado, 1999). Among the different spectral models is that they consider the generation and development
weather elements affecting port operations, wave height inside the of wind waves; on the flip side, they cannot resolve phenomena like
basin is one of the sources of most of the inoperative hours (Tsinker, diffraction or reflection for which the phase information is essential. On
2004). It follows that knowledge of wave agitation in the basin is the contrary, the mild slope and Boussinesq models do resolve the
essential for the planning, design, construction, security and phase, but can only be implemented in small areas given their high
operation of ports. This is not, however, an easy task: as waves computational cost. In all cases, wave observations from buoys or
approach the coast, propagating from deep to shallow water, they pressure gauges within the basin are recommended for validation.
interact with the seabed and shore through shoaling, refraction, This study proposes an alternative method based on artificial
diffraction, reflection and breaking. The ensemble effects of these intelligence (AI), in particular, artificial neural networks (ANNs). Over
phenomena change both the height and direction of waves, causing the last few years, ANNs have been successfully applied to different
attenuation in some areas and concentration of energy in others. ocean and coastal engineering studies, including the reliability of
In general, to assess the wave height inside a basin, the use of coastal structures (Iglesias et al., 2008; Kim et al., 2014), tidal prediction
experimental or numerical models is required. Physical (or experi- (Chang and Lin, 2006; Lee, 2004), long-wave prediction (López and
mental) models could in principle be seen as one of the best Iglesias, 2013), wind direction forecasting (Tagliaferri et al., 2015), wave
forecasting (Deo et al., 2001; Kashikar and Mane, 2014; Makarynskyy,
n
Corresponding author. Tel.: þ 34 982 285900. 2004; Vimala et al., 2014) or beach morphology (Hashemi et al., 2010;
E-mail address: [email protected] (I. López). Iglesias et al., 2009b). There are also several studies addressing the

http://dx.doi.org/10.1016/j.oceaneng.2015.09.016
0029-8018/& 2015 Elsevier Ltd. All rights reserved.
 I. López et al. / Ocean Engineering 109 (2015) 298–308 299

assessment of harbour tranquillity (Kankal and Yüksek, 2012; Londhe Nørgaard et al., 2003). ANNs, as a part of AI applications, emulate
and Deo, 2003). Nevertheless, these previous attempts rely on physical the behaviour of biological neural networks and, like them, their
experiments and numerical model tests to train and validate the ANN, main characteristic is the adaptive learning, i.e., tasks are not
so do not suppose an alternative to the conventional models. The ANN required to be programmed, an ANN learns from experience.
model implemented in this work estimates the short-wave height Furthermore, ANNs have also capacity to generalise from previous
inside a harbour from deep-water buoy observations, and, from the examples or to abstract the main characteristics of data series.
knowing of the limiting operational conditions, enables the assessment As aforementioned, the ANNs are composed of simple proces-
of port operability. The ANN is trained and validated based on field sing nodes called neurons. The artificial neurons receive a number
observations from wave buoys, making use of the same input data than of inputs and process them to produce the output: each input is
conventional models do. However, the ANN adds a significant advan- multiplied by a weight, reinforced by a bias and the result is com-
tage over these models: it is several orders of magnitude faster than a puted by a transfer function (also known as activation function) to
conventional model, allowing studying the whole set of wave condi- delimit the output. The process of adjusting the value of the weights
tions and not only the most representative cases. and biases to model the data is referred to as training or learning. In
The methodology is illustrated through a case study: the Exterior addition, using a learning algorithm, ANNs are able to adjust these
Port of Ferrol (Galicia, NW Spain). It is a new basin located in the parameters themselves. In order to provide an independent mea-
entrance of Ría de Ferrol (43.467°N, 8.334°W) (Fig. 1). It constitutes a sure of how well the ANN responds another set of examples that
very interesting case study because joins three main characteristics: have not been used to train is also needed, called validation set.
(i) as an outer port it is much more exposed to wave action than the For this work, multilayer feed-forward back-propagation neural
traditional ones; (ii) its large size, too much large to be modelled by a networks (hereafter, for the sake of brevity, back-propagation
phase-resolving numerical model (which are compulsory for a har- neural networks) were chosen on account of their generalisation
bour where diffraction is the dominating phenomenon), and (iii) a capabilities (Hagan et al., 1996), which has been shown in a
container terminal will soon be inaugurated (a traffic with a parti- number of applications (e.g., Hashemi et al., 2010; Iglesias et al.,
cularly high susceptibility to wave height disturbances). 2009a; Kankal and Yüksek, 2012). This ability makes possible to
train the ANN with only a representative number of cases and
achieve good results in other totally new to the network. As its
2. Materials and methods name suggests, back-propagation networks consist of several lay-
ers: an input layer, one or more hidden layers and an output layer.
2.1. Artificial neural networks The number of neural layers and neurons per layer is referred to as
ANN architecture. The training of these ANNs requires a set of
An ANN consists in a set of highly interconnected elements, so- examples—inputs and targets—, also known as training set, that
called neurons, working in parallel (Haykin, 1999; Jang et al., 1997; shows the network which result it must obtain for a given input.

N N

Estaca de
Bares buoy
Atlantic
Ocean Bay of
Biscay Vilán-Sisargas
buoy

Atlantic
Spain Ocean
Mediterranean Sea

0 200 400 600 km 0 25 50 100 km

N Interior Port
of Ferrol

Exterior Port
of Ferrol
Ría de Ferrol

Ferrol-1
tide gauge

0 1 2 3 km

Fig. 1. Location of the Exterior Port of Ferrol, the Vilán-Sisargas and Estaca de Bares deep-water wave buoys and the Ferrol-1 tide gauge.
300 I. López et al. / Ocean Engineering 109 (2015) 298–308

Hm0 (m) Uw (m/s)

>5 > 10
4−5 8 − 10
3−4 6−8
2−3 N 4−6 N
1−2 2−4
0−1 0−2
NW NE NW NE

W 0% E W 0% E
5% 2%
10% 4%
15% 6%
20% 8%
25% 10%
SW SE SW SE

S S

Fig. 2. Wave and wind roses at the Estaca de Bares buoy.

Thus, during the training, the weights and biases are adjusted generalisation of the ANN, the angular measurements of wave direc-
iteratively to minimise the ANN performance function between tion were converted into Cartesian coordinates as follows:
the targets and the actual network outputs, usually, the mean

θmx ¼ sin θm ; ð1Þ
square error (MSE).
There are two main issues that must be prevented when

θmy ¼ cos θm : ð2Þ
dealing with back-propagation networks. First, the random initi-
alisation of weight values: the same ANN, with the same archi- The same procedure was followed with the wind direction
tecture and the same input data will not always generate the same measurements
output. And second, the over-fitting: although the training error is

θwmx ¼ sin θwm ; ð3Þ
small, the validation error is large, i.e., the ANN does not generalise
when facing new cases.

θwmy ¼ cos θwm : ð4Þ

2.2. ANNs applied to the study of port operability As a statistical summary of these datasets, the wind and wave roses
for the Vilán-Sisargas and Estaca de Bares buoys are presented in
The parameters that determine the wave height inside the basin Figs. 2 and 3, respectively. Waves at Estaca de Bares buoy are pre-
can be classified into two main groups: (i) parameters that define the dominantly propagating from the northwest quadrant (Fig. 2 – left),
characteristics of the incident waves; and (ii), parameters that define being the west-northwest the prevailing direction. The most common
the configuration of the basin, i.e., bathymetry, coastal morphology, significant wave heights vary between 1 and 3 m (about 75% of the
position and characteristics of the breakwater, etc. When working with time). Winds show a more irregular distribution with the highest
ANNs, the information of the basin configuration is unnecessary since percentage of data within the northeast and southwest quadrants
they determine the values of wave height inside the basin just inter- (Fig. 2 – right); the prevailing direction is the east-northeast. The wave
preting a set of examples that show to the ANN towards which values distribution at Vilán-Sisargas buoy (Fig. 3) is very similar to that of the
of wave height a given deep-water observation leads. previous buoy, with most of waves coming from the northwest
quadrant, and the most common heights comprised between 1 and
2.2.1. Inputs and targets set-up 3 m. The prevailing direction is the northwest. The wind distribution
The inputs of the ANN were obtained from observations from (Fig. 3 – right) is again concentrated in the northeast and southwest
deep-water wave buoys operated by Puertos del Estado (Spanish quadrants, with northeast as the prevailing direction in this case.
Port Authority). Taking into account the location of the Port of Finally, an additional input was provided to the ANN to take
Ferrol, data from two different buoys—Vilán-Sisargas and Estaca into account the changes in sea level produced by the great tidal
de Bares—were used (Fig. 1). Estaca de Bares buoy dataset covers range existing in the area (4.50 m). The tidal time series was
from 1996/06/19 to 2012/02/10, whereas Vilán-Sisargas dataset computed on the basis of the tidal constituents (Pawlowicz et al.,
covers from 1998/05/12 to 2013/01/30. In both cases, data were 2002), following
recorded hourly. These records, however, present diverse blank X
n

periods—due to malfunctioning or maintenance of the buoy—dis- ξ¼ Ai cos σ i t  φi ð5Þ
i¼1
tributed throughout the year. Altogether, the datasets for the
Estaca de Bares and Vilán-Sisargas buoys comprise 72,308 and where ξ is the sea-level variation, Ai is the amplitude of each ith
84,839 data points, respectively, covering approximately 15 years. harmonic constituent, σi is its angular frequency, φi its phase and n
In order to correctly characterise the incident wave climate, the is the number of tidal constituents.
following spectral wave parameters were used: spectral significant As mentioned above, the ANN needs, besides the inputs, a set of
wave height (Hm0), peak period (Tp) and mean wave direction (θm). In targets that shows to the network which result it must obtain for a
addition, taking into account that wind and waves are closely related, given input. Thus, time series of wave height inside the basin are
the mean wind speed (Um) and mean wind direction (θwm) were used. required, which were obtained from measurements from the
Given the angular nature of the mean wave and wind directions, two Ferrol-1 tide gauge located at the front face of the dock and
similar measurements can have a very different representation value operated by Puertos del Estado (Fig. 1). The dataset extends from
(e.g., 359° and 1°). To address this issue, which can lead to a poor 2006/12/22 to 2013/03/30, with a twenty-minutely frequency. In
 I. López et al. / Ocean Engineering 109 (2015) 298–308 301

Hm0 (m) Uw (m/s)

>5 > 10
4−5 8 − 10
3−4 6−8
2−3 N 4−6 N
1−2 2−4
0−1 0−2
NW NE NW NE

W 0% E W 0% E
5% 2.4%
10% 4.8%
15% 7.2%
20% 9.6%
25% 12%
SW SE SW SE

S S

Fig. 3. Wave and wind roses at the Vilán-Sisargas buoy.

total, the Ferrol-1 measurements comprise 152,518 data points (as order to reduce the number of ANN architectures tested. In the first
for the deep-water wave buoys the tide gauge records present stage, 57 different architectures were trained. With one hidden layer,
diverse blank periods). architectures with a number of hidden neurons varying from five to
The last step involved standardising the datasets that come eighty in steps of five were compared (i.e., 5, 10, 15, 20, 25, 30, 35, 40,
from observations (Vilán-Sisargas, Estaca de Bares and Ferrol-1) by 45, 50, 55, 60, 65, 70, 75 and 80 hidden neurons), giving a total of 16
discarding those data points not present in all sets simultaneously architectures. With two hidden layers, the architectures compared
(tide is not included since it was reconstructed from tidal con- were those resulting from the combination of 5, 10, 15, 20, 25, 30 and
stituents as desired). In the case of the Ferrol-1 tide gauge data, a 35 neurons in the first hidden layer and 1, 2, 3, 4, 5 and 10 neurons in
prior step is required to transform twenty-minutely data into the second hidden layer; in total 41 ANN architectures (10 neurons in
hourly data, consisting in averaging the three values of wave the second hidden layer were only used in those cases where the
height in an hour. All things considered, the input and target number of neurons in the first hidden layer is equal or higher than 10).
dataset reached 25,848 data points. In the second stage, the selection of the ANN architecture was
To sum up, the matrix of inputs comprises 15 components: improved by comparing eight additional architectures covering all the
seven from each one of the deep-water buoys (significant wave options in the vicinity of the best architecture of the previous stage.
height, peak period, Cartesian components of the mean wave From this point onwards, an ANN with, for instance, fifteen input
direction, mean wind speed and Cartesian components of the neurons, five neurons in the hidden layer and one output neuron will
mean wind direction) and the astronomical tide. The vector of be referred to as a [15-5-1] network.
targets comprises only the wave height inside the dock. Thus, the The transfer function used for the hidden neurons was the tan-
ANN model can be mathematically expressed as sigmoid function
0 V 1
1
H m0 ; T Vp ; θmx ; θmy ; U Vm ; θwmx ; θwmy ;
V V V V
y ¼ 2 1 þ e  2x 1 ð7Þ
B C
H Fm0 ¼ f ANN B
@ H m0 ; T p ; θmx ; θmy ; U m ; θwmx ; θwmy ;
E E E E E E E C
A ð6Þ whose nonlinearity allows the model to account for nonlinear
ξ relations between the data. For the output neuron, a linear transfer
function was used
where the superscripts V and E represent the Vilán-Sisargas and
y¼x ð8Þ
Estaca de Bares buoys, respectively; and the superscript F repre-
sent the basin of Exterior Port of Ferrol. This combination yielded excellent results in previous appli-
cations of ANNs to engineering problems (Iglesias et al., 2009a;
2.2.2. ANN model set-up López and Iglesias, 2014, 2013).
One of the main difficulties when it comes to implement back- As regards the training algorithm, the Levenberg–Marquardt
propagation neural networks is to determine the ANN archi- algorithm (LMA) (Hagan and Menhaj, 1994) was selected for this
tecture; in fact, developing a method to automatically determine work. It constitutes an intermediate algorithm between the Gauss–
the number of hidden neurons, without conducting a thorough Newton method and gradient descent algorithm, and addresses the
experimental study of the various options, remains a current limitations of each of those techniques: it combines the rapid
research topic. Thus, the architecture of the ANN is usually convergence of the Gauss–Newton method with the independence
determined using a trial and error methodology (Haykin, 1999). of the initial value of the weights of the gradient descent algorithm.
The number of neurons in the input and output layers are, how- To prevent over-fitting, the Bayesian regularisation algorithm
ever, determined by the nature of the problem. In this case, the (MacKay, 1992a, 1992b) was implemented. The principle of the
input layer must consist of 15 source nodes, corresponding to the Bayesian regularisation technique is based on the addition of a
right side of Eq. (6), and the output layer of one neuron, corre- penalty term (the mean of the sum of squares of the ANN weights)
sponding to the one output variable in the same equation. to the performance function, so that for two ANNs with a similar
In order to find the optimum number of hidden neurons and error, the best one is that with the lowest values of the weights.
hidden layers, a comparative study of different architectures with one The Bayesian approach has the benefit that no dataset is needed to
and two hidden layers and different number of neurons in them, was help the ANN to prevent over-fitting (as in other methods such as
carried out. The comparative study was performed in two stages in early stopping). The last step in the ANN set-up was the
302 I. López et al. / Ocean Engineering 109 (2015) 298–308

normalisation of the inputs and targets in the range [  1, 1] to where o and t are the average values of the outputs and targets,
make the training more efficient. respectively.

2.2.3. ANN validation 2.3. Sensitivity analysis

Among the different methods of validation existent (e.g., hold-out
validation, leave-one-out or bootstrapping) in this study the k-fold Sensitivity analysis can be defined as the study of how uncer-
cross-validation was used. In this method, the data are divided into k tainty in the output of a model can be attributed to different sources
subsets of (approximately) equal size and the validation process is of uncertainty in the model input (Saltelli et al., 2008), in other
repeated for k iterations. Each time, one of the k subsets is used as words, how sensitive the output of a model is to changes in the
validation set and the remaining (k 1) as the training data. Finally, value of the inputs. This technique is usually applied to ANNs to
the average error across all k trials is computed. Thus, how the data is overcome their black-box nature (in the sense that there is a poor
divided lacks importance: every data point has the opportunity to be knowledge of their internal mathematical structure) and thus
in a test set once and to be in a training set k 1 times. identifying the role or contribution of the different variables in the
To accomplish the k-fold cross-validation whatever value of k outputs of the ANN. Furthermore, it can be used to check for the
smaller than the size of the dataset is valid (where k equals the presence of unnecessary inputs, i.e., redundant variables which do
sample size the validation method is called leave-one-out). A priori, not affect the performance of the ANN model (Bowden et al., 2005).
a large value of k may be desirable, since a higher k involves a Among the different methods existent in the literature, e.g., PaD,
greater number of error estimates, and also the size of the training perturb, weights, profile or stepwise (a review of these methods can be
set is closer to the size of the entire dataset. When k is large, found for instance in Gevrey et al., 2003), the profile method (Lek et al.,
however, the overlap between the training sets increases and there 1996) was used in this work, which has been applied successfully in
is a reduction in the size of the validation set; this leads to a worst other coastal engineering applications of ANNs, e.g., (Castro et al., 2014).
estimation accuracy of the generalisation of the ANN. A widely used In this method, one of the input variables—the explanatory input—is
solution that provides a good balance between pros and cons is the divided into a number of equal intervals—the scale—between its
cross-validation of 10 iterations, also called 10-fold cross-validation. minimum and maximum values. A scale value of 12 is usually adopted,
This was the value used in this work. Thus, the total dataset, which and is the value used in the present study. An output is obtained across
was composed of 25,848 entries (each one made up of 15 inputs the range of values of the explanatory input while the remaining inputs
and a corresponding target), was divided randomly into 10 subsets are held constant at set values (minimum, first quartile, median, third
of approximately 2585 entries. quartile and maximum). Thus, for each one of the 12 values of the
In addition, each of the k trials of the cross-validation proce- explanatory input, five output values are obtained, which are reduced
dure was run 10 different times in order to, irrespective of the to the median value. The procedure is repeated switching the expla-
initial values of weights, determine the architecture with the natory input. The final result is a set of response curves of the output
greatest generalisation capabilities. Thus, each one of the ANN across the range of values of each one of the inputs.
architectures was trained 100 times: 10 times for each of the 10
cross-validation iterations. In total, 5700 ANNs were trained in the
2.4. Port operability analysis
first training stage and 800 more in the second stage, to reach a
total of 6500 ANNs tested.
In order to ensure port operability, wave height inside the har-
The performance of the different architectures was compared
bour basin cannot exceed certain thresholds. These thresholds or
on the basis of the MSE, its square root, i.e., the root mean square
limiting operational conditions depend, as aforementioned, both on
error (RMSE), and the normalised root mean square error
the type of operation to perform (mooring, loading or unloading)
(NRMSE), defined by
and the type of vessel in question. The limiting operational condi-
" #1=2
1 1X N tions are usually defined based on the movements of vessels at
2
NRMSE ¼ ðt  oi Þ ð9Þ berth (Elzinga et al., 1992; PIANC, 1995), however, it is also possible
t max  t min N i ¼ 1 i
to establish a criterion based on the wave conditions that lead to
where oi and ti represent each i-th output and target, respectively; these movements (Puertos del Estado, 1999). In this work, the latter
N is the total number of cases of training or validation; and tmax procedure was followed, by using the values (Table 1) recom-
and tmin are the maximum and minimum values of the targets, mended by the Spanish Port Authority (Puertos del Estado, 1999),
respectively. The performance of an ANN architecture was repre- which have been used in other harbour management case studies
sented by the average NRMSE of the 10 iterations of the cross- (González-Marco et al., 2008). In addition, Table 2 gives the mean
validation procedure; and, similarly, the performance of the ANN acceptable area downtimes due to adverse environmental condi-
for each k-iteration was calculated as the average of the 10 trials tions, again, following the recommendations of the Spanish Port
carried out for that iteration. Authority (Puertos del Estado, 1999).
Additionally, the standard deviation of error between the dif- After the validation process, the resulting ANN model was
ferent architectures was also calculated applied to the study of the port operability of the Exterior Port of
" #1=2 Ferrol. Thus, the generalisation capability of the ANN was used to
1X N
estimate the wave agitation inside the basin from the observations
σ¼ ðE k  E Þ2 ð10Þ
Ni¼1 from the deep-water wave buoys (Fig. 4). In this way, the gen-
eralisation dataset covered approximately 15 years (1998–2012)
where Ek and E are the average NRMSE of each k-fold and the total
rather than the 5 years covered by the tide gauge measurements. It
average NRMSE of each architecture, respectively.
enabled, therefore, a more accurate characterisation of the average
Finally, the correlation coefficient was used, defined as
values of wave height inside the basin and, from the limiting
P
N
 operational conditions (Table 1), of the non-operational hours of
ðoi  oÞ t i  t
i¼1
the area. The monthly and annual average values of non-
R¼" #1=2 " #1=2 ð11Þ operational hours were calculated dividing the number of hours
P
N N

P 2
ðoi  oÞ2 ti  t in which the limiting operational conditions are exceeded by the
i¼1 i¼1 total number of hours of the period of time considered.
 I. López et al. / Ocean Engineering 109 (2015) 298–308 303

Table 1 (Tables 4 and 5). As the number of neurons in the first or second
Limiting operational conditions for loading and unloading operations. hidden layer increases, the training error decreases (Table 4). However,
the lowest validation errors are obtained in the [15-30-10-1] archi-
Type of vessel Hs
tecture (Table 5); for ANNs with more neurons (e.g., those ANNs with
Oil tankers o 30,000 DWTa 1.5 m 35 neurons in the first hidden layer) the validation error increases. In
30,000–200,000 DWT 2.0 m any case, the inclusion of a second hidden layer helps little, and
4 200,000 DWT 2.5 m
actually produces a worse approximation than the achieved by most of
Bulk-carriers Loading 1.5 m
Unloading 1.0 m the one-hidden-layer architectures; in particular, the performance of
Gas tankers o 60,000 m3 1.2 m the best two-hidden-layer architecture is improved by the ANNs of
4 60,000 m3 1.5 m one hidden layer with 40–75 hidden neurons.
General cargo 1.0 m In view of the results of the first stage, the second stage was
Containers, RO-ROs and Ferries 0.5 m
carried out with one-hidden-layer ANNs, varying the number of
a
DWT: Deadweight tonnage. hidden neurons one by one in the vicinity of the [15-60-1] archi-
tecture. The average NRMSE for the architectures tested in this
Table 2 second training stage are presented, together with the archi-
Mean acceptable area downtimes due to adverse environmental conditions. tectures [15-55-1], [15-60-1] and [15-65-1] of the previous stage,
in Table 6. As previously happened, the training error decreases
Area characteristics Downtimes
when the number of hidden neuron increases. As regards the
Terminals operating with regular lines 200 h/year 20 h/month validation error, it is concentrated in a narrow range of values
Terminals not operating regular lines 500 h/year 50 h/month between 4.4587% ([15-57-1] and [15-60-1] architectures) and
4.4673% ([15-59-1] architecture), being difficult to discern which is
the best architecture. Thus, the standard deviation of the NRMSE
for the 10 iterations of the cross-validation was also calculated
(Table 6), obtaining an estimation of how much the error varies
Estaca de Bares depending on the dataset used for validation. Paying attention to
buoy data
the standard deviation, the architecture which achieves the lowest
values (σ E 0.0933%) is the [15-57-1].
Vilán-Sisargas
buoy data
Tide gauge data To sum up, the ANN model that estimates the wave height inside
the basin with the lowest average NRMSE and with the best gen-
1995 2000 2005 2010 eralisation capabilities (it is the most independent of the dataset
year used for validation) is the based on a [15-57-1] architecture.

Subset for training and validating the ANN model 3.2. Performance of the ANN model
Subset for computing wave agitation inside the
basin In this section, the results of the best performing ANN architecture,
Fig. 4. Scheme of the different datasets used in the ANN implementation (training
[15-57-1], are presented in greater detail. To this end, of the 100
and validation) and for the computation of wave agitation inside the basin. training runs, the one corresponding to the minimum NRMSE value
was chosen (NRMSEtraining ¼ 4.0949%, NRMSEvalidation ¼ 4.4587%).
3. Results This is the ANN model used hereinafter.
The generalisation of the ANN model is checked by comparing
3.1. Comparative study of the ANN architectures the values of the wave height measured by the tide gauge with the
corresponding values estimated by the ANN. The linear regression
The first step is to compare the performance of the different graph for the training dataset is presented in Fig. 5. The agreement
ANN architectures in the estimation of the wave height inside the between the measured and estimated values (which add up to
basin of the Exterior Port of Ferrol. Thus, the average MSE, RMSE 23,263 data points for the training dataset), taking into account
and NRMSE of the 57 architectures studied in the first stage are that the problem is extremely difficult, is certainly good. The best
presented in Tables 3–5. linear fit is given by
In the case of one-hidden-layer architectures (Table 3), the ANNs ytrain ¼ 0:9176x þ 0:0301; ð12Þ
with fewer hidden neurons ([15-5-1], [15-10-1] and [15-15-1]) have
the greatest errors, both in training and validation. They are too and the correlation coefficient is R ¼0.9585. Although the linear
simple to adequately capture the information contained in the regression of the training set comprises data points known by the
training dataset—as evidenced by the large errors for this subset— ANN (the ANN does not have to generalise them), checking that
and this inability prevents a good generalisation in the validation the network has ‘learned’ adequately is the fundamental starting
point for a proper generalisation to new cases.
stage. As the number of neurons increases, the training error
The results for the validation dataset, which comprises 2585
decreases. Similarly, the validation error decreases until reach the
data points, are shown in Fig. 6. In this case, the correlation
[15-60-1] architecture, however above this number of neurons
coefficient is, again, very close to unity (R¼0.9506) and the best
there is a change in the trend: although the training error continues
linear fit line is close to the perfect fit. Its equation is given by
to decrease, the validation error starts to increase. In these cases,
the number of hidden neurons is so large that the ANN learns the yvalidation ¼ 0:9223x þ 0:0294: ð13Þ
problem very well (very small training error) but it is not able to
Finally, in order to give a quantitative estimation of the
improve its generalisation capabilities. In any case, the validation
advantages of using ANNs, the accuracy of the ANN model is
error increases slightly, a fact that shows the good performance of compared with a multiple linear regression (MLR) model of the
the Bayesian regularisation algorithm for preventing over-fitting. form
A similar situation occurs in two-hidden-layer ANNs. The archi-
tectures with fewer hidden neurons have, again, the greatest errors Y ¼ β 0 þ β 1 X 1 þ … þ β n X n þ ε; ð14Þ
304 I. López et al. / Ocean Engineering 109 (2015) 298–308

Table 3
Average of the MSE, RMSE, NRMSE and correlation coefficient (R) for the 16 ANN architectures with one hidden layer studied in the first stage of the comparative study
(training and validation datasets).

Neurons in the hidden layer Training Validation

MSE (m2) RMSE (m) NRMSE (%) R MSE (m2) RMSE (m) NRMSE (%) R

5 0.0029 0.0539 4.8155 0.9421 0.0030 0.0543 4.8514 0.9412

10 0.0027 0.0515 4.5988 0.9473 0.0027 0.0523 4.6691 0.9456
15 0.0025 0.0503 4.4913 0.9498 0.0026 0.0514 4.5912 0.9475
20 0.0024 0.0495 4.4163 0.9515 0.0026 0.0509 4.5450 0.9485
25 0.0024 0.0488 4.3544 0.9529 0.0026 0.0506 4.5143 0.9493
30 0.0023 0.0482 4.3002 0.9541 0.0025 0.0504 4.4956 0.9497
35 0.0023 0.0477 4.2555 0.9551 0.0025 0.0502 4.4834 0.9500
40 0.0022 0.0472 4.2166 0.9559 0.0025 0.0501 4.4767 0.9501
45 0.0022 0.0468 4.1769 0.9567 0.0025 0.0500 4.4654 0.9504
50 0.0021 0.0464 4.1396 0.9575 0.0025 0.0500 4.4651 0.9504
55 0.0021 0.0460 4.1093 0.9582 0.0025 0.0500 4.4623 0.9505
60 0.0021 0.0456 4.0742 0.9589 0.0025 0.0499 4.4590 0.9506
65 0.0021 0.0453 4.0453 0.9595 0.0025 0.0500 4.4631 0.9505
70 0.0020 0.0450 4.0154 0.9601 0.0025 0.0501 4.4690 0.9504
75 0.0020 0.0446 3.9859 0.9607 0.0025 0.0501 4.4733 0.9503
80 0.0020 0.0444 3.9609 0.9612 0.0025 0.0503 4.4882 0.9499

Table 4 Table 5
Average of the MSE, RMSE, NRMSE and correlation coefficient (R) for the 41 ANN Average of the MSE, RMSE, NRMSE and correlation coefficient (R) for the 41 ANN
architectures with two hidden layers studied in the first stage of the comparative architectures with two hidden layers studied in the first stage of the comparative
study (training dataset). study (validation dataset).

Neurons in Neurons in second hidden layer Error Neurons in Neurons in second hidden layer Error
first hidden estimator first hidden estimator
layer 1 2 3 4 5 10 layer 1 2 3 4 5 10

5 0.0031 0.0028 0.0028 0.0028 0.0028 MSE (m2) 5 0.0032 0.0029 0.0029 0.0029 0.0028 MSE (m2)
0.0550 0.0534 0.0530 0.0527 0.0524 RMSE (m) 0.0553 0.0539 0.0539 0.0535 0.0532 RMSE (m)
4.9069 4.7650 4.7352 4.7079 4.6819 NRMSE (%) 4.9407 4.8097 4.7891 4.7748 4.7484 NRMSE (%)
0.9331 0.9433 0.9440 0.9447 0.9453 R 0.9328 0.9422 0.9422 0.9431 0.9431 R
10 0.0029 0.0026 0.0025 0.0025 0.0025 0.0024 MSE (m2) 10 0.0030 0.0027 0.0027 0.0026 0.0026 0.0026 MSE (m2)
0.0524 0.0507 0.0503 0.0501 0.0499 0.0492 RMSE (m) 0.0532 0.0517 0.0516 0.0514 0.0514 0.0510 RMSE (m)
4.6824 4.5234 4.4930 4.4709 4.4568 4.3946 NRMSE (%) 4.7530 4.6179 4.6032 4.5879 4.5856 4.5526 NRMSE (%)
0.9385 0.9491 0.9498 0.9503 0.9506 0.9520 R 0.9368 0.9468 0.9472 0.9475 0.9476 0.9484 R
15 0.0025 0.0024 0.0024 0.0024 0.0024 0.0023 MSE (m2) 15 0.0026 0.0026 0.0026 0.0026 0.0026 0.0026 MSE (m2)
0.0501 0.0493 0.0489 0.0487 0.0485 0.0478 RMSE (m) 0.0514 0.0510 0.0510 0.0509 0.0510 0.0506 RMSE (m)
4.4727 4.3984 4.3668 4.3491 4.3333 4.2648 NRMSE (%) 4.5895 4.5492 4.5558 4.5466 4.5507 4.5157 NRMSE (%)
0.9502 0.9519 0.9526 0.9530 0.9534 0.9549 R 0.9475 0.9484 0.9483 0.9485 0.9484 0.9294 R
20 0.0024 0.0023 0.0023 0.0023 0.0023 0.0022 MSE (m2) 20 0.0026 0.0026 0.0026 0.0026 0.0026 0.0025 MSE (m2)
0.0492 0.0483 0.0479 0.0476 0.0475 0.0467 RMSE (m) 0.0510 0.0509 0.0509 0.0509 0.0507 0.0503 RMSE (m)
4.3957 4.3122 4.2778 4.2527 4.2409 4.1663 NRMSE (%) 4.5492 4.5459 4.5456 4.5483 4.5252 4.4931 NRMSE (%)
0.9520 0.9538 0.9546 0.9551 0.9554 0.9570 R 0.9484 0.9485 0.9486 0.9485 0.9490 0.9498 R
25 0.0024 0.0022 0.0022 0.0022 0.0022 0.0021 MSE (m2) 25 0.0026 0.0026 0.0026 0.0026 0.0026 0.0025 MSE (m2)
0.0486 0.0474 0.0470 0.0468 0.0466 0.0458 RMSE (m) 0.0508 0.0511 0.0508 0.0508 0.0507 0.0503 RMSE (m)
4.3354 4.2321 4.2002 4.1803 4.1586 4.0867 NRMSE (%) 4.5324 4.5633 4.5361 4.5380 4.5260 4.4899 NRMSE (%)
0.9533 0.9556 0.9562 0.9567 0.9571 0.9586 R 0.9488 0.9482 0.9488 0.9488 0.9490 0.9499 R
30 0.0023 0.0022 0.0021 0.0021 0.0021 0.0020 MSE (m2) 30 0.0026 0.0026 0.0026 0.0026 0.0026 0.0025 MSE (m2)
0.0479 0.0467 0.0463 0.0461 0.0459 0.0450 RMSE (m) 0.0505 0.0510 0.0510 0.0509 0.0506 0.0503 RMSE (m)
4.2792 4.1679 4.1308 4.1138 4.0975 4.0173 NRMSE (%) 4.5082 4.5511 4.5504 4.5437 4.5212 4.4893 NRMSE (%)
0.9545 0.9569 0.9577 0.9581 0.9584 0.9601 R 0.9494 0.9484 0.9485 0.9486 0.9492 0.9499 R
35 0.0022 0.0021 0.0021 0.0021 0.0020 0.0020 MSE (m2) 35 0.0025 0.0026 0.0026 0.0026 0.0026 0.0026 MSE (m2)
0.0474 0.0462 0.0457 0.0454 0.0452 0.0443 RMSE (m) 0.0504 0.0508 0.0512 0.0509 0.0508 0.0506 RMSE (m)
4.2335 4.1267 4.0803 4.0539 4.0386 3.9528 NRMSE (%) 4.4962 4.5325 4.5745 4.5481 4.5372 4.5170 NRMSE (%)
0.9555 0.9578 0.9588 0.9593 0.9596 0.9614 R 0.9497 0.9489 0.9479 0.9486 0.9488 0.9493 R

The performance of the ANN is better than that of the MLR

where Y indicates the dependent variable (in this case the wave height
model for the whole set of error estimators. For instance, the
inside the basin of the Exterior Port of Ferrol), Xi represent each of the
values of the RMSE and NRMSE of the MLR model are 30% higher
independent variables (the 15 inputs of the ANN model), βi represents
than those of the ANN, and in the case of the MSE 68% higher. In
each of the predicted parameters and ε is the error term. First, the
fact, even the ANNs with the simplest architectures (Tables 3 and
values of the predicted parameters were obtained from the training
5) overcome the performance of the MLR.
dataset. Second, the MLR model was applied to the validation set to
According to all these results, the ANN model can be con-
obtain the values of wave height. Finally, the results of both models,
sidered as a suitable model for the study of the port operability
ANN and MLR, were compared with the real measured values by means
through the estimation of wave agitation from offshore
of the MSE, the RMSE, the NRMSE and the correlation coefficient
observations.
(Table 7).
 I. López et al. / Ocean Engineering 109 (2015) 298–308 305

Table 6
Average of the MSE, RMSE, NRMSE, correlation coefficient (R) and standard deviation (σ) for the 11 ANN architectures studied in the second stage of the comparative study
(training and validation datasets).

Neurons in the hidden layer Training Validation

MSE (m2) RMSE (m) NRMSE (%) R MSE (m2) RMSE (m) NRMSE (%) R σ (%)

55 0.0021 0.0460 4.1093 0.9582 0.0025 0.0500 4.4623 0.9505 0.0998

56 0.0021 0.0459 4.1002 0.9584 0.0025 0.0500 4.4636 0.9505 0.0985
57 0.0021 0.0459 4.0949 0.9585 0.0025 0.0499 4.4587 0.9506 0.0933
58 0.0021 0.0458 4.0872 0.9586 0.0025 0.0500 4.4644 0.9504 0.0964
59 0.0021 0.0457 4.0815 0.9587 0.0025 0.0500 4.4673 0.9504 0.0993
60 0.0021 0.0456 4.0742 0.9589 0.0025 0.0499 4.4590 0.9506 0.0939
61 0.0021 0.0456 4.0690 0.9590 0.0025 0.0500 4.4651 0.9504 0.0953
62 0.0021 0.0455 4.0625 0.9591 0.0025 0.0499 4.4596 0.9506 0.0963
63 0.0021 0.0454 4.0573 0.9592 0.0025 0.0499 4.4597 0.9506 0.0937
64 0.0021 0.0454 4.0528 0.9593 0.0025 0.0500 4.4615 0.9505 0.0953
65 0.0021 0.0453 4.0453 0.9595 0.0025 0.0500 4.4631 0.9505 0.0969

1.4 Table 7
Data Points Wave height prediction performance of the ANN and MLR models (validation
F F dataset).
[Hm0]estimated = [Hm0]measured
1.2
Best Linear Fit Model MSE (m2) RMSE (m) NRMSE (%) R

1.0 ANN 0.0025 0.0499 4.4587 0.9506

MLR 0.0042 0.0649 5.7938 0.9147
]estimated

0.8
3.3. Relative influence of the input variables
[Hm0

0.6
F

The results of the sensitivity analysis carried out through the

profile method are presented in Fig. 7. The response curves
0.4 R = 0.9585 represent the relative contribution of each input variable to the
value of the output (the wave height into the basin), which can be
0.2 quantified by the range of variation of the output (from maximum
to minimum) for the range of values of the input considered.
The spectral significant wave height is the variable that, by far,
0
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 has the greatest effect on the output value (Fig. 7a), being that
F corresponding to the Vilán-Sisargas buoy, with a range of variation
[Hm0]measured of the output value of approximately 0.90 m, the most important
input. The next variable in order of importance is the peak period
Fig. 5. Linear regression analysis for the training dataset.
(Fig. 7b); again, the input from the Vilán-Sisargas buoy is the more
significant (variation of approximately 0.25 m). In fact, the influence
of the inputs from the Vilán-Sisargas buoy is, in general, higher than
those from the Estaca de Bares buoy. This result was to be expected
1.4
taking into account the location of the buoy with respect to the Port
Data Points
F
[Hm0 F
]estimated = [Hm0]measured
of Ferrol (Fig. 1) and the prevalence of the sea states coming from
1.2 the fourth quadrant (Figs. 2 and 3). In any case, the contribution of
Best Linear Fit
the Estaca de Bares dataset is far from negligible, with the wave
1.0 height, period and wing velocity of this buoy among the most
important inputs (second, seventh and fourth positions in order of
]estimated

importance, respectively). Interestingly, with the highest values of

0.8
these inputs, the minimum values of the wave height inside the
basin of the Port of Ferrol are achieved. Finally, the importance of
[Hm0

0.6 the variables which express magnitudes (wave height, period, wind
F

velocity) over the variables which express directions (wave and

0.4 R = 0.9506 wind direction components) must be pointed out. The exception is
the tidal level, whose influence on the output value is almost
negligible, and which constitute the least relevant input.
0.2 According to these results, the selection of the input variables
can be considered good. Thus, the seven variables selected from
0 deep-water wave buoys (significant wave height, peak period,
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 Cartesian components of the mean wave direction, mean wind
F speed and Cartesian components of the mean wind direction) play
[Hm0]measured
a significant role in the output of the ANN. In those cases where
Fig. 6. Linear regression analysis for the validation dataset. the influence of an input from a given buoy is low, e.g., wind
306 I. López et al. / Ocean Engineering 109 (2015) 298–308

1.2 1.2
Estaca de Bares buoy Estaca de Bares buoy
1 Vilán−Sisargas buoy 1 Vilán−Sisargas buoy

(m)

(m)
0.8 0.8
Hm0

Hm0
F 0.6 0.6

F
0.4 0.4
0.2 0.2
0 5 10 5 10 15 20 25
Hm0 (m) Tp (s)

1.2 1.2
Estaca de Bares buoy Estaca de Bares buoy
1 Vilán−Sisargas buoy 1 Vilán−Sisargas buoy
(m)

(m)
0.8 0.8
Hm0

Hm0
0.6 0.6
F

F
0.4 0.4
0.2 0.2
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
θmx θmy

1.2 1.2
Estaca de Bares buoy Estaca de Bares buoy
1 Vilán−Sisargas buoy 1 Vilán−Sisargas buoy
(m)

(m)

0.8 0.8
Hm0

Hm0

0.6 0.6
F

0.4 0.4
0.2 0.2
0 5 10 15 20 −1 −0.5 0 0.5 1
Um (ms ) −1 θwmx

1.2 1.2
Estaca de Bares buoy Ferrol−1 tide gauge
1 Vilán−Sisargas buoy 1
(m)

(m)

0.8 0.8
Hm0

Hm0

0.6 0.6
F

0.4 0.4
0.2 0.2
−1 −0.5 0 0.5 1 −2 −1 0 1 2
θwmy ξ (m)

Fig. 7. Response curves of each input variable to the prediction of the ANN model, obtained by the profile method.

velocity from the Vilán-Sisargas buoy (Fig. 7e), the input of the presented in Table 8. The non-operational hours indicated in the
second buoy does affect the ANN output. A similar situation hap- table are, in every case, annual and monthly average values.
pens with the wind and wave direction components from the The loading and unloading operations of oil tankers (either of
Estaca de Bares buoy whose low influence on the ANN output is less than 30,000 DWT, between 30,000 DWT and 200,000 DWT, or
compensated by the relevance of their equivalents from the Vilán- of more than 200,000 DWT) are guaranteed as their non-
Sisargas buoy. The exception is, as aforementioned, the tidal level, operational times do not reach one hour per year. An identical
which could be avoided in future applications of the ANN model. situation, without any downtime throughout the year, is that of the
gas tankers of whichever capacity (o60,000 m3 and 460,000 m3)
3.4. Assessment of the port operability and the bulk-carriers carrying out loading operations. By contrast,
bulk carriers carrying out unloading operations present a total of
Finally, the results of the port operability assessment for the 29.5 non-operational hours per year (0.3% of the year), dispersed in
Exterior Port of Ferrol, carried out from the values of wave height autumn and winter. In any case, this value is far below the accep-
generalised by the ANN model for the 15 years of offshore data, are table annual closure time due to non-operational conditions
 I. López et al. / Ocean Engineering 109 (2015) 298–308 307

Table 8
Non-operational percentage of time and non-operational hours for loading and unloading operations for the different types of vessels.

Type of vessel Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Annual

Oil tankers o30,000 DWT 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%

0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h
30,000–200,000 DWT 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%
0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h
4200,000 DWT 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%
0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h
Bulk-carriers Loading 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%
0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h
Unloading 2.2% 0% 0.1% 0% 0% 0% 0% 0% 0% 0.5% 0.2% 1.0% 0.3%
16.5 h 0h 0.5 h 0h 0h 0h 0h 0h 0h 3.5 h 1.7 h 7.2 h 29.5 h
Gas tankers o60,000 m3 0.1% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%
0.5 h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0.5 h
460,000 m3 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%
0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h 0h
General cargo 2.2% 0% 0.1% 0% 0% 0% 0% 0% 0% 0.5% 0.2% 1.0% 0.3%
16.5 h 0h 0.5 h 0h 0h 0h 0h 0h 0h 3.5 h 1.7 h 7.2 h 29.5 h
Containers, RO-ROs and ferries 51.3% 15.9% 28.5% 18.8% 7.0% 2.5% 4.2% 5.1% 7.2% 27.4% 35.6% 45.6% 20.7%
381.3 h 106.9 h 211.9 h 135.3 h 52.3 h 18.1 h 31.2 h 37.7 h 51.5 h 203.9 h 256.5 h 339.0 h 1825.7 h

recommended by the Spanish Port Authority to tramp shipping The architecture with 15 neurons in the input layer, 57 neurons in
(irregular shipping, mainly over nonstandard routes, with no defi- the hidden layer and one neuron in the output layer ([15-57-1])
nite schedule) that is set to 500 h (Table 2). Similarly, general cargo proved to be the best of the ANN architectures analysed and was
traffic also has a mean downtime of 29.5 h/year. In this case, chosen for the ANN model. A good agreement was obtained between
however, since general cargo is usually a liner traffic (regular the values of wave height predicted by this ANN model and those
shipping with a fixed route and schedule), the non-operational measured by the tide gauge used for validation, with the values of the
hours in January are very close to the acceptable threshold of 20 h correlation coefficient very close to unity (R40.95) and the lowest
per month (Table 2). validation error among the different ANN architectures compared
As regards the loading and unloading operations of containers, (NRMSE¼4.4587%). Once the ANN model was validated, a sensitivity
the non-operational hours amount to 20.7% of the time, a total of analysis was carried out to analyse the influence of the different input
1825.7 h per year. Containers are the ultimate vessel liners, with variables on the network output. We can conclude that wave height
acceptable downtime values of 200 h/year or 20 h/month (Table 2) and period are the two variables which most influence wave agitation
—well below the values obtained. In fact, the acceptable monthly inside the basin, with the information coming from tidal level and
downtime is not exceeded only in June (18.1 h). wind direction components being of lesser importance.
Finally, the ANN model was used to analyse the port operability
in the Exterior Port of Ferrol, using its generalisation capabilities to
4. Conclusions estimate the values of wave height during the extensive period
during which there are available records in the deep-water buoys
In this work, a method based on artificial intelligence, in par- (approximately 15 years data against the 5 available in the tide
ticular artificial neural networks (ANN), was developed in order to gauge). Based on this analysis and the limiting operational con-
estimate, from deep-water buoy observations, the values of wave ditions for loading and unloading operations, bulk-carriers and oil
height inside a harbour basin and, from those, the corresponding and gas tankers are the most appropriate traffics for this terminal;
levels of port operability. In particular, multilayer feed-forward container traffic is more problematic.
back-propagation neural networks were selected due to its In conclusion, the ANN model developed in this work is shown
excellent generalisation capabilities. To train the ANN, the Leven- as an excellent alternative to conventional models (physical and
berg–Marquardt algorithm was used, combined with Bayesian numerical) with the advantage that it does not require as many
regularisation to prevent over-fitting. The optimum ANN model elaborate data as conventional models do (only deep-water buoy
was selected from 65 different one- and two-hidden-layer archi- observations), it is relatively less complex, less rigid and has a lower
tectures by carrying out a k-fold cross-validation of 10 iterations. computational cost. Thus, the ANN model can be successfully used
In addition, each of the 10 trials of the cross-validation procedure to analyse very long datasets and thereby achieve more accurate
was run 10 times to prevent the effects of weight initialisation. results. The ANN model is specially intended for sites like the
Thus, a total of 6500 ANNs were run. The work was illustrated by Exterior Port of Ferrol, where the large dimensions of the study area
applying the ANN to a case study: the Exterior Port of Ferrol. complicate the implementation of accurate numerical models.
The first conclusion is the excellent capability of ANNs to deal
with this problem; in fact, the NRMSE values are, independently of
the ANN architecture, below 5%. There are, however, differences Acknowledgements
between architectures. It was found that the ANN architectures
with one hidden layer perform better than two-hidden-layer During this work I. López was supported by FPU Grant AP2010–
architectures, being the ANNs with a number of hidden neurons 4690 of the Spanish Ministry of Education, Culture and Sports
between 55 and 65 the architectures with the lowest validation (Ministerio de Educación, Cultura y Deporte). M. López wish to
errors. The ANNs with the simplest architectures provide the acknowledge the research Grant from Fundação para a Ciência e
worst results (both for training and for validation). On the con- Tecnologia (FCT), Portugal (Ref. SFRH/BPD/102112/2014). The
trary, ANNs with the highest number of hidden neurons although authors are indebted to the Spanish Port Authority (Puertos del
have very low training errors, indicative of a good learning, are not Estado), in particular to Dr. I. Rodríguez-Arévalo, Dr. E. Fanjul,
able to generalise when facing new cases. Ms. P. Gil and Ms. S. Pérez, for kindly providing the wave data.
308 I. López et al. / Ocean Engineering 109 (2015) 298–308

References Kankal, M., Yüksek, Ö., 2012. Artificial neural network approach for assessing har-
bor tranquility: the case of Trabzon Yacht Harbor, Turkey. Appl. Ocean Res. 38,
23–31.
Bowden, G.J., Dandy, G.C., Maier, H.R., 2005. Input determination for neural net- Kashikar, V.R., Mane, S.J., 2014. Artificial neural network: an effective tool for
work models in water resources applications. Part 1—background and metho- forecasting wave height. Int. J. Eng. Res. Technol. 3 (7), 744–747.
dology. J. Hydrol. 301 (1–4), 75–92. Kim, D.H., Kim, Y.J., Hur, D.S., 2014. Artificial neural network based breakwater
Bruno, D., De Serio, F., Mossa, M., 2009. The FUNWAVE model application and its damage estimation considering tidal level variation. Ocean Eng. 87, 185–190.
validation using laboratory data. Coast. Eng. 56 (7), 773–787. Lee, T., 2004. Back-propagation neural network for long-term tidal predictions.
Castro, A., Carballo, R., Iglesias, G., Rabuñal, J.R., 2014. Performance of artificial Ocean Eng. 31 (2), 225–238.
neural networks in nearshore wave power prediction. Appl. Soft Comput. 23, Lek, S., Belaud, A., Baran, P., Dimopoulos, I., Delacoste, M., 1996. Role of some
194–201. environmental variables in trout abundance models using neural networks.
Chang, H., Lin, L., 2006. Multi-point tidal prediction using artificial neural network Aquat. Living Resour. 9 (1), 23–29.
with tide-generating forces. Coast. Eng. 53 (10), 857–864. Londhe, S.N., Deo, M.C., 2003. Wave tranquility studies using neural networks. Mar.
Chen, W., Panchang, V., Demirbilek, Z., 2005. On the modeling of wave–current Struct. 16 (6), 419–436.
interaction using the elliptic mild-slope wave equation. Ocean Eng. 32 (17–18), López, I., Iglesias, G., 2014. Efficiency of OWC wave energy converters: a virtual
2135–2164. laboratory. Appl. Ocean Res. 44, 63–70.
Deo, M.C., Jha, A., Chaphekar, A.S., Ravikant, K., 2001. Neural networks for wave López, M., Iglesias, G., 2013. Artificial Intelligence for estimating infragravity energy
forecasting. Ocean Eng. 28 (7), 889–898. in a harbour. Ocean Eng. 57, 56–63.
Elzinga, T., Iribarren, J.R., Jensen, O.J., 1992. Movements of moored ships in har- MacKay, D., 1992a. Bayesian interpolation. Neural Comput. 4 (3), 415–447.
bours. In: Proceedings of the 23rd Internatioanl Conference on Coastal Engi- MacKay, D., 1992b. A practical Bayesian framework for backpropagation networks.
neering (ICCE), Venice, Italy, pp. 3216–3229. Neural Comput. 4 (3), 448–472.
Gevrey, M., Dimopoulos, I., Lek, S., 2003. Review and comparison of methods to Makarynskyy, O., 2004. Improving wave predictions with artificial neural networks.
study the contribution of variables in artificial neural network models. Ecol. Ocean Eng. 31 (5–6), 709–724.
Nørgaard, M., Ravn, O.E., Poulsen, N.K., Hansen, L.K., 2003. Neural Networks for
Model. 160 (3), 249–264.
Modelling and Control of Dynamic Systems: A Practitioner's Handbook.
Goda, Y., 2002. Random Seas and Design of Maritime Structures. World Scientific,
Springer-Verlag, London.
Singapore.
Pawlowicz, R., Beardsley, B., Lentz, S., 2002. Classical tidal harmonic analysis
González-Marco, D., Sierra, J.P., Fernández de Ybarra, O., Sánchez-Arcilla, A., 2008.
including error estimates in MATLAB using T_TIDE. Comput. Geosci. 28 (8),
Implications of long waves in harbor management: the Gijón port case study.
929–937.
Ocean Coast. Manag. 51 (2), 180–201. PIANC, 1995. Criteria for Movements of Moored Ships in Harbours: A Practical
Hagan, M.T., Menhaj, M.B., 1994. Training feedforward networks with the Mar- Guide. Permanent International Association of Navigation Congresses. Perma-
quardt algorithm. IEEE Trans. Neural Netw. 5 (6), 989–993. nent Technical Committee II. Working Group 24, Brussels, Belgium.
Hagan, M.T., Demuth, H.B., Beale, M., 1996. Neural Network Design. PWS Publishing Puertos del Estado, 1999. ROM 3.1-99 Recommendation for Design the Maritime
Co., Boston, MA. Configuration of Ports, Approach Channels & Harbour Basins. Spanish Ministry
Hamidi, M.E., Hashemi, M.R., Talebbeydokhti, N., Neill, S.P., 2012. Numerical mod- of Public Works, Madrid, Spain.
elling of the mild slope equation using localised differential quadrature Rusu, E., Guedes Soares, C., 2011. Wave modelling at the entrance of ports. Ocean
method. Ocean Eng. 47, 88–103. Eng. 38 (17–18), 2089–2109.
Hashemi, M.R., Ghadampour, Z., Neill, S.P., 2010. Using an artificial neural network Rusu, L., Guedes Soares, C., 2013. Evaluation of a high-resolution wave forecasting
to model seasonal changes in beach profiles. Ocean Eng. 37 (14–15), 1345–1356. system for the approaches to ports. Ocean Eng. 58, 224–238.
Haykin, S., 1999. Neural Networks. A Comprehensive Introduction. Prentice Hall, Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., Gatelli, D., et al., 2008.
New Jersey. Global Sensitivity Analysis: The Primer. John Wiley & Sons.
Iglesias, G., López, I., Carballo, R., Castro, A., 2009a. Headland-bay beach planform Su, S., Ma, G., Hsu, T., 2015. Boussinesq modeling of spatial variability of infragravity
and tidal range: a neural network model. Geomorphology 112 (1–2), 135–143. waves on fringing reefs. Ocean Eng. 101, 78–92.
Iglesias, G., López, I., Castro, A., Carballo, R., 2009b. Neural network modelling of Tagliaferri, F., Viola, I.M., Flay, R.G.J., 2015. Wind direction forecasting with artificial
planform geometry of headland-bay beaches. Geomorphology 103 (4), neural networks and support vector machines. Ocean Eng. 97, 65–73.
577–587. Tsinker, G.P., 2004. Port Engineering: Planning, Construction, Maintenance, and
Iglesias, G., Rabuñal, J., Losada, M.A., Pachón, H., Castro, A., Carballo, R., 2008. A Security. John Wiley & Sons, Hoboken, New Jersey.
virtual laboratory for stability tests of rubble-mound breakwaters. Ocean Eng. Vimala, J., Latha, G., Venkatesan, R., 2014. Real Time wave forecasting using artificial
35 (11–12), 1113–1120. neural network with varying input parameter. Indian J. Geo-Mar. Sci. 43 (1),
Jang, J.S.R., Sun, C.T., Mizutani, E., 1997. Neuro-fuzzy and Soft Computing. Prentice 82–87.
Hall, Upper Saddle River.