Avgouleas & Sclavounos 2014

PART C: Natural Sciences and Mathematics
Fuel-Efficient Ship Routing

Kyriakos Avgouleasa and Paul D. Sclavounosb
a
Lieutenant Commander HN, Submarine Shipbuilding Program, Hellenic Navy Detachment of
Skaramangkas, Greece
b
Professor, Department of Mechanical Engineering, Massachusetts Institute of Technology,
Cambridge MA, USA
Abstract. Optimal weather routing of ships seeks to determine the voyage route, in rough
seas, which minimizes a certain metric or criterion (e.g. transit time, consumed bunker
fuel, pollutant emissions etc.). In this paper the problem of minimum-fuel navigation in a
sea state is studied. Advanced methods of hydrodynamic simulation are employed to
compute the added resistance in waves and the responses of an example-ship in random
incident waves. A fictitious trans-Atlantic route is introduced. Given the wave forecast of
the geographical region of interest, the problem of determining the route that minimizes
fuel consumption is numerically solved using an improved Dynamic Programming
algorithm. It is concluded that significant savings in bunker fuel may result from the
implementation of a decision support system based on state-of-the-art hydrodynamics,
wave forecasting and optimization methods.
Keywords: Routing; optimization; hydrodynamics; resistance; fuel consumption.
PACS: 02.30.Yy, 47.35.Lf
INTRODUCTION
The shipping industry consumes about 5% of the world oil production in bunker fuel for
maritime transportation. That corresponds to approximately 4 million barrels of oil a day. This
fact warrants an effort to optimize navigation in a way that will bring the fuel cost down to the
extent possible. A new approach to the solution of the optimal routing problem is presented in
this paper. A combination of frequency domain methods for the computation of the ship
hydrodynamics, state-of-the-art weather forecasts and advanced dynamic programming
algorithms is utilized to generate trajectories (ship routes in the ocean) which minimize the
required fuel while meeting certain safety restrictions associated with the severity of the sea.
The optimization problem associated with the selection of the fuel-minimizing route of a ship
in a sea state is challenging. In calm weather the optimal route within the constraints imposed
by sea lanes is usually the loxodrome (or the orthodrome for long routes), and the fuel
consumption is known with a high degree of certainty given the sailing distance and the calm
water resistance and propulsion characteristics of the vessel. In rough weather a number of
complexities arise. The severity of the sea states to be encountered during the trip is not known
a priori with certainty and must be estimated from weather forecasts supplied by a routing
service. When sailing in rough seas, safety criteria must be met that protect the integrity of the
hull and the cargo. They affect the vessel speed and heading in a sea state. Reliable
ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

C-39
NAUSIVIOS CHORA, VOL. 5, 2014
seakeeping methods must be available in order to predict the vessel responses and the
increase of her resistance in waves – the added resistance – given the attributes of the wave
system. Finally, a robust and efficient dynamic optimization algorithm must be available that
may be executed in real time to determine the optimal speed and heading of the vessel in rough
weather in order to minimize the fuel consumption, subject to the safety constraints. These
challenges are addressed in the present paper and an optimal ship routing algorithm is
developed that leads to a notable reduction of the vessel fuel consumption in rough weather.
Literature Review
The problem of ship weather routing has been a subject of extensive research for many
decades now. A comprehensive, albeit not exhaustive, review of the research work in the field
starting from the 1950’s can be found in Avgouleas (2008). Some of the developments in the
last 10 years will be briefly presented here. Rathje and Beiersdorf (2005) developed a shipboard
routing assistance (SRA) software to prevent containerships from encountering dangerous
conditions with respect to seakeeping behavior (parametric rolling, slamming, exceedance of
bending moment and shear force threshold values). Montes (2005) proposed a method for the
automation of the Optimum Track Ship Routing (OTSR) system used by the US Navy. In
Abramowski et al. (2006) a formal solution of the minimal time ship routing problem is presented
on the basis of Pontryagin’s maximum principle. Tsujimoto and Tanizawa (2006) solve the
constrained optimization problem of minimum fuel routing, using the augmented Lagrange
multiplier method. Böttner (2007) describes a decision support system able to provide optimal
alternatives in case the ship finds herself in degraded condition (hull damage, rudder/propulsion
failure etc.). The work of Szlapczyńska (2007) and extensions thereof (Szlapczyńska and
Smierzchalski 2007-2009, Krata and Szlapczyńska 2011, Szlapczyńska 2013) focuses on
deterministic weather routing of a sail assisted ship using Multi-Objective Evolutionary Algorithm
(MOEA). Panigrahi et al. (2008) carry out a simulation of wave climate in the Indian Ocean and
use it to minimize voyage time for a cargo ship route in the region. The optimization utilizes the
Dijkstra algorithm and speed loss in waves is calculated from empirical curves. In Panigrahi et
al. (2012) the hydrodynamics of ship motions and added resistance in waves are treated
differently. Empirical speed loss curves are abandoned and standard frequency domain
seakeeping computation is adopted. Sen and Padhy (2009) conducted a similar study using a
Dijkstra algorithm as well, but with linear strip theory for hydrodynamic simulation. Wisniewski et
al. (2009) utilize evolutionary algorithm to determine the minimum time route which safely avoids
a tropical cyclone. Mezaoui et al. (2009) solve the unconstrained constant-speed minimum fuel
problem in coastal navigation using high resolution forecasts and a Dijkstra optimization
scheme. Dolinskaya et al. (2009) deduce shortest-time optimal paths inside convex regions of
areas that a ship can sail. Their analysis is restricted to short-range optimal routing. Marie and
Courteille (2009) deal with the deterministic dual-objective optimization of minimum
time/minimum fuel route using Multi-Objective Genetic Algorithm (MOGA) and identifying Pareto
optimal solutions. Hinnenthal and Clauss (2010) solve the same problem using similar principles
but they quantify the robustness of the final solution exploiting ensemble forecasts. Delitala et al.
(2010) explore the results of climatological simulations in combination with weather routing.
Bruns et al. (2011) present a software tool for fuel efficient speed profile selection using
potential flow CFD for calm water resistance prediction, strip theory for added resistance in
waves and RANS modeling for the propulsor (both in design and off-design conditions).
Gershanik (2011) addresses the challenges and benefits of weather routing optimization. The
author advocates the use of the classical discrete dynamic programming algorithm (backward
recurrence) for the minimum time or the minimum fuel optimal routing problem. Pipchenko
(2011) solves the deterministic minimum work ship routing using genetic algorithm or a Nelder-
http://nausivios.snd.edu.gr/nausivios
C-40
Mead heuristic method. Maki et al. (2011) propose a route optimization method that considers
not only fuel efficiency but also ship stability (the risk of parametric rolling in particular). A real-
coded genetic algorithm (GA) is used for the multi-objective optimization problem. Kobayashi et
al. (2011) optimize the route of a containership in a simulated trans-oceanic passage,
accounting for ship maneuvering dynamics. Dolinskaya (2012) introduces the notion of
restricted turning radius in optimal path finding of a vessel in an inhomogeneous, non-stationary
environment. Kosmas and Vlachos (2012) apply the simulated annealing method for the
minimization of a cost function defined by the weighted sum of voyage time and a comfort
parameter related to safety. Ilus and Heikkinen (2012) present a speed profile optimization
approach which relies on historical and operational data collected for a specific route or leg of a
voyage. The aim is to optimize energy usage while meeting prescribed constraints. Shao et al.
(2012) solve the deterministic minimum fuel routing problem using a discrete 3D dynamic
programming algorithm. Skoglund et al. (2012) present a method to obtain optimal routes in a
directed graph using the concept of Pareto optimality. Both deterministic and ensemble
forecasts are used. Tsou and Cheng (2013) combine an ant colony algorithm with a genetic
algorithm for the solution of the minimal fuel and pollutant emission weather routing problem. In
their recent work, Marie and Courteille (2014) construct a fuzzy logic model with data collected
from ship’s sensors to minimize fuel consumption of a sail-assisted motor vessel.
Model Overview
An example-ship is used from the Series 60 hullform, with the characteristics shown in table
1. A standard MARIN B-Series propeller is selected for this ship and a MAK 9M25C medium-
speed, four-stroke diesel engine is matched with the propeller via a mechanical gearbox.
Resistance in calm water is calculated from standard statistical methods (Holtrop 1984).
Added resistance in waves is computed using SWAN1, a frequency domain CFD code which
uses 3D Rankine Panel Methods to simulate steady and unsteady potential flows around
vessels with realistic hull shapes. Unlike most conventional strip and slender-body theories,
SWAN1 exploits robust computational hydrodynamics methods to produce reliable results
where approximate theories fail. The case of quartering and following waves can be mentioned
as an example. At certain speeds they correspond to the sub-critical reduced frequency regime
(τ<¼) and represent challenging hydrodynamic problems intractable by strip theory. The same is
true for the computation of added resistance in waves. Added resistance is a second order
effect of central importance to weather routing, as it influences both the calculation of fuel
consumption and the prediction of engine overload in a sea state. SWAN1 computes added
resistance using direct integration and its results have been calibrated against numerous
experiments carried out for America’s Cup yachts (Sclavounos and Nakos 1993). In addition to
resistance, all ship responses in waves are computed by SWAN1. In particular, pitch and heave
motions and velocities near the ends of the hull are required to calculate probabilities of
slamming and deck wetness. Limiting values imposed on these probabilities define safety
constraints in the formulation of the minimum fuel consumption problem. Spectral analysis is
used to estimate the mean added resistance and RMS values of the ship responses in a sea
state with a given direction, significant wave height and modal period. The vessel hydrodynamic
properties are computed a priori in the frequency domain across a wide range of encounter
frequencies, ship speeds and wave headings and stored in a database for use in the solution of
the fuel minimizing ship routing problem. Therefore, the time-consuming solution of the ship
hydrodynamic problem in the time domain is avoided.

C-41
TABLE 1. Principal Dimensions of Example-Ship.

Description Symbol Unit Value
Length at DWL L m 101.7
Beam at DWL B m 14.28

Draft at DWL T m 5.7
Freeboard at DWL (bow) H m 2.85
Displacement  tonne 5700

Wetted Surface Area S m2 1950
Block Coefficient CB - 0.676
Prismatic Coefficient CP - 0.688
Midship Section Coefficient CM - 0.982
TABLE 2. Propeller Characteristics of Example-Ship.

Description Symbol Value
Diameter (m) D 3.2
Pitch Ratio P/ D 1.2
Expanded Area Ratio EAR 0.75

Number of Blades z 5
TABLE 3. Engine Characteristics of Example-Ship.

Description Value
Type MAK 9M25C
Rated Power (MCR) (kW) 3000
Rated Speed (rev/min) 750
Gear Ratio 4.45:1
A third generation model, WAM cycle 4, developed at the European Center for Medium-
Range Weather Forecasts (ECMWF) and adapted by the University of Athens (UoA), Greece,
provides the weather forecast for the region of interest. The forecast is uploaded daily and
covers a total of 168 hours. A fictitious cross-Atlantic route starting from Cape St.Vincent,
Portugal, and ending at Norfolk Virginia, USA, defines the nominal voyage track considered in
the present study. It is assumed that the nominal (calm water) speed for this trip is 13 knots. The
nominal sailing distance is 3138.6 nautical miles following the shortest route, which is the great
circle route (orthodrome). At that speed the trip would last 10.06 days. Mercator projection is
used throughout. For the part of the route not covered by the wave forecast, calm water is
assumed.
Keeping the start and end points fixed (“hard” constraints) and the “expiration time” of the
chartering contract also fixed (i.e. the time at destination), a special Dynamic Programming (DP)
algorithm is implemented in MATLAB ® which calculates the optimal route with backward
C-42
recurrence. To safeguard against the shortcomings of the standard DP algorithm, namely the
large requirements for memory storage and processing capacity, Iterative Dynamic
Programming (IDP) is employed. IDP is an advanced extension to the standard DP algorithm
(Luus 2000). It retains all the benefits of DP without suffering the infamous drawback known as
the “curse of dimensionality”.
HYDRODYNAMICS
Calm Water Resistance
Calm water resistance mainly consists of three components: friction, form and wave drag.
The friction drag coefficient is calculated in this paper by the ITTC 1957 model-ship correlation
line
0.075
Cf  (1)
(log Re 2)2
10
The form drag coefficient is calculated from the Holtrop regression model (Holtrop 1984). It is
given in the form of an enhancement factor k to the friction drag coefficient, based on the hull
geometry and accounting for the presence of bilge keels (if present) and appendages. In this
example, a bare Series 60 hull model is entered in the resistance calculations and the Holtrop
resistance curve is shown in figure 1.
Calm Water Resistance Prediction - Series 60 hull
300
Holtrop
250
bare hull resistnce (kN)
200
150
100
50
0
7 8 9 10 11 12 13 14 15 16
speed (knots)
FIGURE 1. Steady Resistance from Holtrop Method.
The total calm water resistance coefficient is expressed as:
Ctotal  (1  k )  C f  Cw  Ca (2)
where Ca is the model-ship correlation allowance and Cw the wave-making resistance

coefficient, both obtained from the Holtrop statistical method. For a ship with total wetted surface
S moving with constant forward speed U the total resistance is given by:
1
Rcalm  Ctotal    S U 2 (3)
2

C-43
Propulsion
The thrust T and torque Q of a propeller with diameter D rotating at a speed n in a fluid with
density ρ can be expressed in non-dimensional form as thrust and torque coefficients, KT and KQ
respectively:
T
KT  (4)
 n2 D4
Q
KQ  (5)
 n 2 D5
The speed at which the propeller is moving in the wake is the speed of advance UA:
U A  (1  w) U (6)
The parameter w is an average measure of the wake effect and is termed wake fraction. The
speed of advance can be non-dimensionalized to a parameter known as the advance ratio:
UA
J (7)
nD
The rate at which fuel mass is consumed in the engine is:
q  sfc( PB )  PB (8)
The engine break power PB is related to resistance through the following expression:
RtotalU
PB  (9)
 DTRM
The quasi-propulsive efficiency ηD can be expressed as the product:

D  RoH (10)
The relative rotative efficiency ηR and hull efficiency ηH are parameters dependent mainly on
the geometry of the hull. Although these parameters depend on speed too, the dependence on
hull geometry dominates and they are often assumed constant in naval architecture. For the
particular example-hull used here their values are 1.035 and 1.097 respectively, interpolated
from the propulsion factors of the Series 60 parent models (Todd 1963). The open water
propeller efficiency can be expressed in terms of the propeller parameters:
KT J
o  (11)
2 KQ
The term ηTRM in equation (9) stands for the transmission efficiency, which can be decomposed
further into a product of shaft and gearbox efficiency:
TRM  SGB (12)
C-44
In the example considered here the gearbox and shaft efficiencies are assigned typical values of
0.96 and 0.98 respectively.
Ship Dynamics in Waves

SWAN1 solves the linear forward speed seakeeping problem of the three-dimensional flow
around vessels of arbitrary geometry. Linearity allows for the solution of the problem using
frequency domain methods. The ambient waves are harmonic and so are the ship motions. Real
sea states are modeled by spectral analysis using the frequency domain results from SWAN1.
The computational domain for this study is shown in figure 2. The rectangular grid extends 50m
from the bow to the upstream free surface boundary, 100m from the stern to the downstream
boundary and 140m from the ship’s centerline to the lateral boundary.
Z
X
Y
FIGURE 2. SWAN1 Computational Grid.
A ship-fixed reference frame Oxyz with the z=0 plane coinciding with the calm water surface
and the positive x-axis coinciding with the positive ship axis on the waterline is related to the
earth-fixed OXYZ frame through the Galilean transformation:
X  x  Ut (13)
Yy (14)
Zz (15)
The total velocity potential Φ(x,t) can be decomposed into a sum of two terms:
(x, t )  (x, t )   (x, t ) (16)
The first term on the right hand side represents the steady flow potential, which can be further
decomposed into a basis flow and a disturbance flow:
(x, t )   (x, t )  (x, t ) (17)
The basis flow potential  (x, t ) accounts for the thickness effect of the ship’s hull as it

C-45
encounters the uniform ambient stream of velocity -U (the moving ship is equivalently treated as
fixed against an oncoming uniform flow). The disturbance potential  (x, t ) accounts for the
presence of the steady wave disturbance that forms the Kelvin wave pattern. The sum of these
quantities satisfies the zero normal flux condition on the mean position of the hull. In SWAN1 the
linearization of the steady flow is carried out around the double-body basis flow. The free
surface is replaced by a rigid wall at z=0, on which:

0 (18)
z
 (x, t ) represents the potential of the flow around the hull and its image above the plane z=0. By
this representation, the hydrodynamic end effects are modeled more accurately than in typical
strip and slender-body theories in which the basis flow is just the ambient free stream. The field
equation, the free surface and radiation conditions and the aforementioned boundary conditions
formulate a boundary value problem (BVP), which is solved numerically in SWAN1. The solution
 corresponds to a steady outgoing wave pattern known as the Kelvin wake.
The second term on the right hand side of (16) is the unsteady velocity potential associated
with the ambient (harmonic) wave, its interaction with the hull and the resulting ship motions. It
can be expressed as:

 (x, t )  e  (x)eit  (19)
In the above expression the harmonic term oscillates with the encounter frequency ω:
  o  kU cos  (20)
where ωo is the absolute ambient wave frequency defined relative to the earth fixed coordinate
system and k the corresponding wavenumber, given by the dispersion relation in deep water:
 2
k (21)
g
Following the standard convention for relative wave direction, the angle β is measured from the
stern (β=180o means head wave).
The complex potential  in (19) is a superposition of the incident wave potential  I , the
diffraction potential  D and the radiation potentials  j , j=1,…,6 for all modes of motion:
6
  I  D  
j 1
j (22)
SWAN1 solves the relevant BVP for the diffraction potential  D and calculates the (complex)
excitation forces from the linearized Bernoulli equation extended to account for the modeling of
end effects by the double-body basis flow:
  
P     i  U    I   D  (23)
 x 
C-46
Xi   Pn dS
SB
i (24)
Furthermore, upon solving the BVP for the radiation potentials  j SWAN1 provides the complex
amplitudes of motion for all six degrees of freedom in the form of Response Amplitude
Operators (RAO), normalized by the ambient wave amplitude:
 j ( )
RAO j ( )  , j=1,…,6 (25)
A
Examples of RAOs in heave and pitch in oblique (bow and stern quartering) waves for a Froude
number of 0.21 (13 knots) are shown in figures 3 to 6. The RAO’s are plotted against the
encounter period of the wave.
Heave RAO , Fn =0.21173

1.5
 =135o
1
|3/A| (m/m)
0.5
0
0 5 10 15 20 25 30
T (sec)
FIGURE 3. Heave RAO in Bow Waves at 13 Knots.
Pitch RAO , Fn =0.21173

4
 =135o
3.5
2.5
|5/A| (deg/m)
1.5
0.5
0
0 5 10 15 20 25 30
T (sec)
FIGURE 4. Pitch RAO in Bow Waves at 13 Knots.

C-47
Heave RAO , Fn =0.21173

1.4
 =45o
1.2
|3/A| (m/m)
0.8
0.6
0.4
0.2
0
0 5 10 15 20 25
T (sec)
FIGURE 5. Heave RAO in Quartering Waves at 13 Knots.
Pitch RAO , Fn =0.21173

1.6
 =45o
1.4
1.2
1
|5/A| (deg/m)
0.8
0.6
0.4
0.2
0
0 5 10 15 20 25
T (sec)
FIGURE 6. Pitch RAO in Quartering Waves at 13 Knots.
Added Resistance in Waves

Added resistance is a hydrodynamic force of paramount significance to the ship routing
problem. The effect of waves manifests itself in the fuel consumption through the existence of
this additive resistance force. Standard methods exist for the calculation of added resistance,
but with relatively narrow range of applicability in terms of wave direction (Gerritsma and
Beukelman 1972) or Froude numbers (Faltinsen et al. 1980).
SWAN1 implements the direct pressure integration method and accurately models the
hydrodynamic effects near the ends of the ship which are essential for the robust prediction of
the added resistance. The total velocity potential is given by (16). The total resistance is the x-
component of the hydrodynamic force:
C-48
RT    Pn dS
SB
1 B (26)
The above force is obtained by integrating the total hydrodynamic pressure:
  1 
P        gz  (27)
 t 2 
around the instantaneous submerged hull surface SB. Replacing  by the steady flow potential
 in (27) and integrating the resulting P over the mean surface of the submerged hull, the calm
water resistance follows:
Rcalm    Pn dS
SB
1 B (28)
The difference between the total and calm water resistance is oscillatory, time dependent and
can be written as the sum of three components:
Rtotal  Rcalm  R1  R2  R3 (29)
R1     Pn dS
SB
1 B (30)
R2    P n dS
SB
1 B (31)
R3    Pn dS
 SB
1 B (32)
In the above relations,  S B represents the fluctuation of the mean wetted surface S B that
accounts for the ship motions,  P is the correction to the steady pressure P at some rigid point
on the instantaneous position of the hull in waves and  n1 is the difference between the x-
component of the unit normal vector at the instantaneous and mean positions of the hull. By
expanding (29) in a Taylor series around the mean position of the hull, the quantities  S B ,  P ,  n1
appear explicitly in terms of  ,  and their gradients evaluated at the mean positions of the hull
and the waterline. Since the latter quantities are available from the solution of the linear
seakeeping problem, the added resistance can be calculated. For a monochromatic wave of
amplitude A and frequency ω, at an angle of incidence β the added resistance is defined as the
mean of (29), namely:
Rw (,  )  R1  R2  R3 (33)
In SWAN1 the output is expressed as an Added Resistance Operator (ARO) normalized by the
square of the ambient wave amplitude:

C-49
Rw (,  )
ARO(,  )  (34)
A2
Figures 7, 8 plot ARO vs. encounter period for the same wave headings as in figures 3 to 6.
Added Resistance Operator, Fn =0.21173

80
 =135o
70
60
50
ARO (kN/m2)
40
30
20
10
0
0 5 10 15 20 25 30
T (sec)
FIGURE 7. Added Resistance Operator in Bow Waves at 13 Knots.
Added Resistance Operator, Fn =0.21173

8
 =45o
6
2
ARO (kN/m2)
-2
-4
-6
-8
-10
-12
0 5 10 15 20 25
T (sec)
FIGURE 8. Added Resistance Operator in Quartering Waves at 13 Knots.
For irregular seas characterized by a spectrum S(ω0,θ) the mean added resistance can be
found by integrating the ARO and the spectral density over all constituent absolute wave
frequencies and headings:

 2

Rw  2 d 0
0
 d S ( , )  ARO[ ( ), ]
0 0 (35)

2
C-50
As explained below, wave and swell are treated as locally unidirectional, propagating along
the mean directions computed by the forecasting model. In this setting integration over θ in (35)
need not be carried out. In the evaluation of the mean added resistance by (35) the dependence
of the ARO on the encounter frequency defined by (20) has been used to transform the
integration in (35) over the absolute wave frequency.
SEA STATE MODELING AND FORECASTING

By means of an advanced CFD tool like SWAN1, the ship hydrodynamics in calm water and
in waves is properly modeled, as outlined in the preceding section. In addition to resistance and
powering, predicting the variability of the ocean environment is a necessity in optimal ship
navigation. In this paper the representation of the sea state is derived from the results of the
ECMWF global version of the WAM numerical model. The model solves the wave transport
equation (i.e. the governing law of wind-wave energy balance) without any assumption on the
spectral shape. The solution is the spectral density given at 28 different frequencies and 24
wave directions for every grid point on the global map (Emmanouil et al. 2007). After processing
the spectral information, the output contains significant wave height, maximum expected wave
height, mean and peak frequencies and mean wave direction. For swell waves the output
provides swell height, mean direction and mean frequency. These results constitute the wave
and swell forecasts which are promulgated daily in 1-hour time intervals at each grid point. For
reasons of computational efficiency the study presented here makes use of the forecast data in
3-hour time steps. The spatial resolution of 0.5×0.5 degrees corresponds to an area of roughly
55×55 km.
Although WAM4 calculates the spectrum, the enormous size of the data files makes it
impossible to process the direct information of spectral density S ( , , , , t ) as a function of
position (i.e. longitude  -latitude  ), frequency ω, direction θ and time t. Practical considerations,
therefore, dictate the need to back-fit the output parameters of the model (namely mean
frequency and wave height) to a standard spectrum which is anticipated to represent the actual
sea state reasonably well. A bi-modal Bretschneider spectrum is selected for this purpose. By
virtue of linearity, two separate single-peaked Bretschneider spectra can be superimposed, one
describing the local storm (i.e. wind-generated waves) the other describing the swell. The
expression for each of them is:
 4 2 0.437( /  )4
S ( )  0.278 H e (36)
5
where  and H stand for the mean frequency and height of the wave (or swell) respectively. H
is related to the significant wave height H s via:
Hs
 1.6 (37)
H
The forecast time interval of 3 hours corresponds to a time window for which the
assumption of stationarity of the sea state holds well. Stationarity, in turn, permits the utilization
of frequency domain methods and spectral analysis in hydrodynamics which together with
linearity allow the calculation of the mean added resistance from expression (35) and the
estimation of the ship response statistics from analogous expressions. For example, the
variance of heave motion (mode 3 in SWAN1) can be easily calculated, given the wave

C-51
spectrum and the RAO in heave:



2
 32  S (o ) RAO3 (o ) do (38)
0
The variances of other derived responses may be evaluated in a similar manner and used in
the equations of the constraints in the optimal control problem in question.
All hydrodynamic attributes of a particular vessel of interest in the minimum fuel routing
problem are contained in the ship response RAOs and the added resistance ARO. These
quantities are computed once, in advance, as functions over all absolute wave frequencies, ship
speeds and headings in unidirectional waves and stored in a hydrodynamic database
customized for each vessel. They are subsequently used in expressions (35) and (38) together
with the wave spectra supplied by the weather forecasts. As will be seen below the mean
added resistance and RMS values of the ship responses are the quantities that enter the
minimum fuel routing problem and their evaluation is very efficient given the ARO and RAOs
which are retrieved from the hydrodynamic database.
Treating the spectrum as locally unidirectional in the vicinity of each position on the map (grid
point x), the variance in equation (38) and the added resistance in (35) are both functions of the
local prevailing direction of propagation θ0 of the sea state (i.e. the mean direction). Technically,
the local directionality of the sea state is modeled by multiplying the spectral density by a
“spreading function”, which in this setting is a delta function located at θ0:
S (0 , )   (  0 )  S (0 ) (39)
Integration of an expression like (35) over all angles with the spectrum cast in the form of (39)
eliminates all but one direction θ0. If the assumption of long-crested waves propagating in one
direction is lifted, some directional spreading should be included in the model at the higher
computational cost incurred by the double integration in (35). The use of different spreading
functions (i.e. delta, cos2(), cosh-2(), Hasselmann formula, Mitsuyasu formula etc.) and their
effect on the optimal solution and computational time will be addressed in future work.
MINIMUM FUEL ROUTING
Powering and Fuel Consumption in Rough Seas

The equations of motion of the ship sailing in the ocean can be written in spherical
coordinates as:
1
 (t )   U (t )  cos  p(t )  (40)
Rearth
1
 (t )   U (t )  sin  p(t )  (41)
Rearth  cos  (t ) 
The speed of the vessel is denoted by U and the course is denoted by p (measured relative to
the true north). U refers to the speed through water (STW), which in the absence of currents
coincides with the speed over ground (SOG). The latter is accurately measured by the GPS
receivers on board. Ocean currents are not considered in the present study, but their inclusion
in the model is straightforward. The spherical coordinates in the above equations are the latitude
C-52
  
  ,  , measured from the equator (positive north), and the longitude     ,  
 2 2
measured from the Greenwich meridian (increasing eastbound). Rearth is the mean radius of the
earth. The differential distance between positions  ,   and   d ,   d   , if measured along
the rhumb line (loxodrome) connecting the two locations, is given by:
dS  U (t )  dt  Rearth  d (t )2  cos2  (t )   d  (t )2 (42)
The loxodrome is the constant heading route connecting any two locations on the sphere. On
a Mercator map this route appears as a straight line. On the other hand, the shortest distance
between two points on the sphere is obtained if these two points are connected with a great
circle arc. This is the orthodrome route. The great circle distance between points 2 , 2  and
2 , 2  is:
S  Rearth  cos1  sin(1 )sin(2 )  cos(1 ) cos(2 ) cos(2  1 )  (43)
It is customary in navigation to approximate the great circle route with a sequence of rhumb line
segments. This is the approach adopted throughout this paper.
x(t )   (t )  (t )  is
T
The position vector the state vector of the system, while
u(t )  U (t ) p(t )  is the control vector. Assuming quasi-steady conditions, the thrust delivered by
T
the propulsor is balanced by the total resistance:
Rcalm (U )  Rw (U , p,  ,  , t )
T (U , p,  ,  , t )  (44)
1 t
T R
where t denotes time to distinguish it from the thrust deduction factor defined as t  .
T
Combining equations (4), (6), (7) and (44) yields the propeller load curve:
Rcalm (U )  Rw (u, x, t )
KT  J2 (45)
 (1  t )(1  w) U D 2 2 2
0.9
Thrust coefficient
Torque coefficient (x10)
0.8 Open water efficiency
0.7 Load curve
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.12 0.24 0.36 0.48 0.6 0.72 0.84 0.96 1.08 1.2
Advance ratio
FIGURE 9. Typical B-Series Chart. Intersection of load curve with thrust coefficient curve uniquely
defines all the operational parameters of the propeller.

C-53
Equation (45) together with the KT curve of the B-series chart define a nonlinear system of
equations for the unknowns KT and J. Upon solving this system, the parameters KQ and o are
readily calculated from (5) and (11) respectively. Figure 9 shows a graphical solution on the
chart. The point of intersection of the load curve with the thrust coefficient curve uniquely
determines the advance ratio J, from which the corresponding KQ and o can be read off the
chart. From equations (8)-(12), the fuel rate can be expressed formally as a function of position
in the ocean, control setting (i.e. speed and course) and time:
q(u, x, t )  sfc  U (46)
R o (u, x, t ) H S GB
The product of the last two terms in (46) is the engine load (break power). The dependence of
the term sfc on engine load has been suppressed for brevity. Figure 10 demonstrates this
dependence for the MAK 9M25C engine.
Specific Fuel Consumption Curve

205
Specific Fuel Oil Consumption (g/kWh)
200
195
190
185
180
20 30 40 50 60 70 80 90 100
Engine Break Power (% of MCR)
FIGURE 10. Specific Fuel Consumption as a Function of Engine Load (data source: ref [22]).
The integral of (46) over the entire duration of the trip defines a scalar performance index:
tf
I  sfc  
0 R o (u, x, t )  H  S GB
 Udt (47)
The above index quantifies the total amount of fuel which has been consumed upon reaching
the fixed final time tf . It is the cost function to be minimized in the optimization process.
Inequality Constraints
While seeking to minimize (47) there are certain limiting factors which in a typical optimal
control problem appear as inequality constraints. Violation of these constraints is prohibited at
all times and this restriction changes the structure of the solution.
C-54
State Constraints, namely limits in  and  , could represent regions of forbidden navigation,
such as shallow waters, land or ice. The numerical simulation in this study was carried out in the
Atlantic Ocean below the arctic circle, so the possible influence of shallow waters and ice is not
investigated.
Control Constraints are bounds imposed by the propulsion plant capabilities. For example,
the particular ship cannot exceed 15.45 knots in calm seas with the given engine at the
particular loading condition, assuming clean hull. In addition, the engine load (an implicit
function of the controls) is not permitted to exceed an envelope function g(PB,ne) which is
depicted in figure 11, otherwise the engine will be overloaded. PB and ne represent the engine
load and speed respectively.
Engine Load Diagram (MAK 9M25C)

100
95
90
Engine Load (% MCR)
85
80
75
70
65
60
80 85 90 95 100 105
Engine Speed (% rated speed)
FIGURE 11. Engine Operating Envelope (data source: ref. [22]).
Therefore, the control bounds can be described as follows:
PB U , p   g  PB , ne  (48)
U min  U  U max (49)
In principle, the lower bound for speed should be zero. However, in the optimization example
described below it was found that there is no gain in lowering the speed below 7 knots. As
verified by SWAN, the responses in speeds lower than 7 are practically those of the ship at zero
speed. Regardless of the initial conditions or the trial forecasts, it was observed that the
program would always choose to avoid severe weather by altering the ship’s course while
retaining speed above 7. Furthermore, there exists a lower bound for engine speed to avoid
excessive vibrations. In order to save computational effort the lower bound Umin was set to 7
knots. This limit corresponds (in calm water) to an engine speed of 311 revs/min, well above the
minimum allowable speed of 250 revs/min for this particular engine type (ref. [22]).

C-55
Safety Constraints are enforced to ensure the secure transportation of the cargo and the
physical integrity of the ship herself, while moving in rough seas. Quantifying safety is not a
trivial task. The approach proposed here is to enforce “hard” inequality constraints on the
probabilities of occurrence of certain extreme events associated with sea severity. Two such
events are green water on deck (or deck wetness) and slamming. Other motion-related
constraints could also be specified, at increased computational cost, depending on the type of
vessel and nature of risks that must be taken into account when sailing in severe weather. For
instance, parametric rolling is a dangerous situation that should be avoided especially in
container transportation. In what follows, only restrictions on deck wetness and slamming have
been considered, as they are deemed sufficient to demonstrate the concept and the benefits of
constrained optimal routing. It should be noted that these two events are related to pitching and
heaving, two longitudinal modes of ship motion which can be predicted by SWAN1 with
remarkable accuracy. Considering a point Π1 in the uppermost part of the bow (bulwark), deck
wetness occurs when the relative motion exceeds the freeboard f at point Π1. The probability of
this event follows the Rayleigh distribution (Ochi 1998) and is given by:
f2

2 r2
P(water on deck)  e (50)
The variance of the relative motion,  r 2 , is calculated by an expression similar to (38). Slamming
occurs when the relative motion at a point Π2 in the keel near the bow exceeds the draft H at
that point and when the relative velocity exceeds a critical value:
Vcr  0.093 g  L (51)
The critical velocity Vcr, gravitational acceleration g and waterline length L enter equation (51) in
SI units. Assuming that the two events are statistically independent, the probability of slamming
follows the Rayleigh distribution:
 H2 V2 
 2  cr2 
 2 2V 
P(slamming)  e  r  (52)
where  r 2 and  V 2 are the variances of the relative motion and velocity, respectively, at point Π2.
When the frequency of occurrence of slamming or deck wetness rises above certain limits,
the motions are so severe that the ship’s safety is compromised and voluntary speed reduction
is effected. The limiting probabilities assume the following values (Faltinsen 1990):
P(slamming)  0.03 (53)
P(water on deck)  0.07 (54)
Substituting (50) and (52) in (53) and (54), respectively, and taking the logarithms of both sides
the safety constraints take the form of inequalities involving the RMS values of the relative
motion and velocities at selected points on the hull. The RMS values of these and other
derivative seakeeping quantities may be evaluated easily from their definition and the use of
expressions analogous to (38).
C-56
Formulation of the Optimal Routing Problem

The complete formulation can be summarized in the standard format for constrained
optimization problems:
tf
minimize I  sfc  
0 R o (u, x, t )  H  S GB
 Udt
subject to:
Dynamic constraint:
1
 (t )   U (t )  cos  p(t ) 
Rearth
1
 (t )   U (t )  sin  p(t ) 
Rearth  cos  (t ) 
Control bounds:
PB U , p   g  PB , ne 
U min  U  U max
Safety constraints:
 H2 V2 
 2  cr2 
 2 2 V 
P(slamming)  e  r   0.03
f2

2 r2
P(water on deck)  e  0.07
initial and terminal conditions:
  (0)   0 
x(0)    
  (0)   0 
  (t f )    f 
x(t f )    
  (t f )    f 
   
and final time:
tf prescribed
In order to solve the problem numerically the discrete counterparts of the cost function and

C-57
dynamic constraint are needed. The sailing time is divided into N stages of length:
tf
t  (55)
N
The equations of motion are approximated as:
1
 (k  1)   (k )  tU (k ) cos p(k ) (56)
Rearth
1
 (k  1)   (k )  tU (k )sin p(k ) (57)
Rearth  cos  (k ) 
with  (1)  0 ,  (1)  0 ,  ( N  1)   f ,  ( N  1)   f and k  1, 2,..., N .
The integrand in (47) is the Lagrangian. In discrete form the integral is approximated by a sum
and the performance index becomes:
N
I  q U (k ), p(k ), (k ),  (k ), k t
k 1
(58)
NUMERICAL SOLUTION
Dynamic Programming
Dynamic Programming (DP) was introduced by Richard Bellman in the 50’s and has proven
to be an invaluable tool in optimal control, especially for problems that do not admit an analytical
solution. Many such problems do not satisfy optimality or existence and uniqueness conditions.
In these cases the classical calculus of variations formulation falls short. Yet, these problems
may possess an optimal solution which can be found by DP. Mathematical principles of the
method can be found in Bellman (1957). A standard DP algorithm is presented in Kirk (1970).
Iterative Dynamic Programming (IDP) is a recent evolution of the same idea. It applies the
recurrence relation known as the Bellman equation in an iterative fashion, converging to the
solution in each iteration. The IDP algorithm can be implemented with surprisingly few grid
points, thus eliminating the most prominent weakness of the standard DP algorithm: as the
number of dimensions grows the vast size of memory storage becomes prohibitive. In fact, the
IDP version adopted in this paper uses only a single grid-point. The algorithm outlined below in
10 steps, is adapted by Luus (2000):
1. Discretize the time interval [0, t f ] into N time stages of equal length t . t should be chosen to
coincide with the forecast interval. If t f  N  t , an additional time stage is required of length
t   t f  N  t .
2. Make an initial “guess” for the entire control sequence u (k ) (k  1,..., N ) . In the iteration
following the first, this will be the optimal control obtained in the previous iteration. Using u (k )
C-58
integrate (40)-(41) to generate an initial nominal trajectory x (k ) , where k  1,..., N  1 .

Alternatively, one can “guess” the nominal trajectory x (k ) and compute the corresponding initial
control history u (k ) .
3. Choose the number of controls cr (r  1, 2,..., m) for each element of the mx1 control vector u .
The total number of controls is C  c1  c2  ...  cm . Choose the control region contraction factor  .
4. Choose an increment u that determines the span of control region around the central value
u , i.e.
j 
u  u  j R (59)
The left superscript j denotes iteration number. The quantity j R is a scaling factor applied to
u . It is responsible for narrowing the control region around j u in every iteration. For the first
iteration 1R  1 .
cr
5. Discretize each element r of u into quantized values. These values are selected
2
randomly in the range u .
6. Start the iterations by setting j  1, j R  1 .
7. Move two steps back from the terminal state to the beginning of stage N  1 . This corresponds
to time t f  2t . Integrate forward along the next stage, that is from time t f  2t to time t f  t ,
C times using C different values of u :
u( N  1)  j u ( N  1)  u  j R (60)
If any of the elements of u( N  1) falls outside the allowable values for control, clip them at the
upper or lower bound as appropriate. The integration starts from the known state j x ( N  1) of
the current optimal path. Calculate the reached state xr ( N ) (r  1, 2,..., C) and integrate along the
final stage to close in on the (fixed) destination point (i.e. the terminal state). Calculate the
associated performance index I for each of the C different controls for this two-leg route. Find
the r for which I is minimized and store the corresponding ur ( N  1) as j 1u ( N  1) .
8. Move one step further back, to the beginning of stage N  2 (corresponding to time t f  3t ).
Starting from state j
x ( N  2) integrate one step forward, up to time t f  2t , using C different
values of u :
u( N  2)  j u ( N  2)  u  j R (61)
Calculate the reached state xr ( N  1) . For every r , carry out the integration of the remaining
trajectory up to the terminal state. Use the optimal control j 1u ( N  1) derived in the previous
step to integrate. Calculate the performance index for this segment of the path (i.e. from time
t f  3t to t f ). Among the C different values, store the minimum I and the corresponding
control as j 1u ( N  2) .
9. Repeat steps 7 and 8 until the initial state is reached (initial condition). Integrate forward
along the complete path using all available combinations of control as before and determine the

C-59
best one for the initial stage, i.e. j 1u (1) . This concludes the first iteration. A complete control
sequence j 1u (k ) , (k  1, 2,..., N) and an optimal trajectory j 1 x (l ) , l  1, 2,..., N 1 are now
available for the next iteration.
10. Reduce the size of the control region by setting j 1R   j R . Increase the iteration index j by
1 and repeat the algorithm starting from step 6.
t
The above algorithm assumes that the whole voyage duration f is covered by a wave/swell
forecast. If the duration of the voyage exceeds the maximum forecast time, as is the case in the
example below, the algorithm needs to be adapted to assume calm water conditions for the
segment of the route beyond forecast coverage. In this case, the remaining part of the route will
be the great circle arc to the destination point. Figure 12 shows the extent of forecast coverage
along the nominal great circle route.
Great Circle Route

60 W 30 W 0
forecast coverage
45 N
30 N
15 N
FIGURE 12. Forecast Grid Points Closest to the Great Circle Route.
Assumptions
The length of time stages, as defined in the algorithm above, is chosen to coincide with the
forecast interval of 3 hours. Such a short-term description of a sea state allows much confidence
in the stationarity assumption discussed earlier. The control policy (i.e. speed and course) is
piecewise constant over this time step. No accelerations are taken into account. The transients
associated with changing the control input (e.g. engine rpm) are very short relative to the
problem’s time scale and are neglected. Other components of resistance (such as appendage
drag, air resistance, trim or steering resistance) are not included in the analysis. These are
either steady components which can be easily incorporated in the model or relatively small in
magnitude. The effect of currents is also omitted, as discussed above.
At each stage (time) and for every state (spatial location) the program calculates fuel
consumption and checks for constraint violation. For that, it draws information from stored
forecast data. In the particular example presented here, the global forecast uploaded on the
FTP server of UoA on October 9, 2014 contains the relevant information. For the fraction of the
C-60
trip not covered by the forecast calm water is assumed. Ship response and added resistance
data are obtained from SWAN1, via the constructed hydrodynamic database, for 10 different
speeds (from 7 to 16 knots) and 5 different headings (from 0 to 180 degrees on either side). The
actual data are stored in memory in the form of large matrices. Intermediate values are obtained
by interpolation. The number of iterations is set to 20 and γ is chosen to be 0.7. The
discretization of controls yields a total of C=50 combinations of speed and course (5 speed
settings and 10 course settings are used in the optimization code).
A longer trip, on the order of several weeks, can be handled by the IDP algorithm as well. In
this case, more than one 168-hour forecasts are needed to cover the whole voyage duration.
The algorithm is executed at the starting time using the available forecast at that time. The
following day, as soon as the next forecast becomes available, the algorithm is executed again
with different initial conditions. These are naturally the coordinates of the ship’s new position.
The final time is also adjusted accordingly. Now the part of the total route not covered by a sea
state prediction is shorter. The program is run daily using updated forecasts in a similar fashion
until the destination is reached. This process may be executed very efficiently onboard the
vessel. It can, in fact, be implemented in short duration voyages as well.
RESULTS AND DISCUSSION

A trans-oceanic passage would normally be planned along the great circle route. In calm
seas this is the minimum fuel route, because it is the shortest one. Regardless, the rhumb line
route is often considered by the navigator as a reasonable alternative, especially when the great
circle routes pass closer to the poles and through more severe seas and/or navigational hazards
(e.g. icebergs). Table 4 below presents a comparison of calculation results for the two routes in
the numerical example worked out in this text. It can be seen that the loxodrome (rhumb line) is
2.3% longer than the orthodrome (great circle) and about 2.5% more fuel-expensive in calm
seas. In rough seas, on the other hand, computation of fuel consumption when sailing the great
circle route shows a 45.5% increase in fuel cost compared to the respective calm seas quantity.
For the rhumb line route the increase of fuel cost in waves is 14%.
TABLE 4. Great Circle vs. Rhumb Line Routes.

Description Unit Value
Great Circle (GC) Distance naut.miles 3138.6
Rhumb Line (RL) Distance naut.miles 3210.5
Nominal Voyage Time days 10.06
Calm Water Speed (GC) knots 13.0
Calm Water Speed (RL) knots 13.3
Calm Water Fuel Cost (GC) tonne 66.89
Calm Water Fuel Cost (RL) tonne 68.59
Predicted Fuel Cost (GC) tonne 97.31
Predicted Fuel Cost (RL) tonne 78.24
Figures 13 to 16 show the fraction of each trajectory in which some safety constraint is
violated. In the middle part of the great circle route the WAM forecast predicts wind waves and
swell with significant height locally in excess of 9m and 4m respectively. For the rhumb line the
predicted figures are about 3.5m and 3m respectively. In both routes the ship encounters these

C-61
sea states from head and bow directions. The rhumb line route violates only the deck wetness
limiting probability of occurrence.
Great Circle Route: Safety Hazards

60 W 30 W 0
deck wetness
45 N
30 N
15 N
FIGURE 13. Violation of Deck Wetness Constraint along the Orthodrome.

60 W 30 W 0
bow slamming
45 N
30 N
15 N
FIGURE 14. Violation of Slamming Constraint along the Orthodrome.
C-62

60 W 30 W 0
engine overload
45 N
30 N
15 N
FIGURE 15. Violation of Engine Overload Constraint along the Orthodrome.
Rhumb Line Route: Safety Hazards

60 W 30 W 0
deck wetness
45 N
30 N
15 N
FIGURE 16. Violation of Deck Wetness Constraint along the Loxodrome.

C-63
FIGURE 17. Evolution of Wind and Swell Wave System as Forecasted between 105h and 120h of
Voyage Time.
The algorithm described in the previous section is coded in MATLAB ® to generate the fuel
minimizing route. First, the code is executed without any constraint considerations. The
orthodrome enters the computations as the nominal route (initial guess). The output is depicted
in figure 18. The computed optimal route, after 25 iterations, saves 19.8% more bunker fuel than
the nominal great circle but only marginally less than the rhumb line (around 0.3%). However,
C-64
this route exceeds the specified threshold for deck wetness probability in the portion designated
by ‘x’ marks. This would normally be an infeasible route in the constrained counterpart of the
optimization problem, hence it would be ruled out during the code execution. To verify this, the
optimization code is run again checking for constraint violation this time. The result is graphically
reproduced in figure 19. The optimal control histories for ship’s course and speed settings are
shown in figures 20 and 21 respectively. Evidently, the optimal route turns sharply towards the
north while slowing down at the same time, in order to avoid the high head/bow seas around the
middle segment of the trajectory. Then it turns back towards the previous track. After the
expiration of the available forecast the gap to the destination point is bridged with a great circle
arc.
Unconstrained Min-Fuel Route


60 W 30 W 0
45 N
Rhumb Line
30 N
Great Circle
Optimal (unconstrained)
deck wetness
15 N
FIGURE 18. Optimal Route of the Unconstrained Minimum Fuel Problem.
Ship Routes on Mercator Projection

60 W 30 W 0
Rhumb Line
Great Circle
Optimal
45 N
30 N
15 N
FIGURE 19. Optimal Route of the Constrained Minimum Fuel Problem.

C-65
Optimal Course Profile

340
320
course (degrees) 300
280
260
240
220
200
0 50 100 150 200 250
time (hours)
FIGURE 20. Optimal Control History for Ship’s Heading.
Optimal Speed Profile

15
14
speed of advance (knots)
13
12
11
10
9
0 50 100 150 200 250
time (hours)
FIGURE 21. Optimal Control History for Ship’s Speed.
The structure of the computed optimal solution is intuitive, if one observes the sea state
evolution in the mid Atlantic during the voyage of interest (figure 17), bearing in mind that the
ship’s freeboard is only 2.85m. In this numerical experiment the behavior of the optimal solution
is explored under the influence of longitudinal motion-related constraints only. As a result,
turning the ship’s bow away from the waves both reduces added resistance and alleviates
heaving and pitching motions.
Convergence, Accuracy and Sensitivity

Speed of convergence and accuracy are both dominated by the control space C and the
contraction factor γ, as defined above in the IDP algorithm. Convergence is illustrated in figure
C-66
22 for the constrained optimization problem. To expedite convergence, the optimal solution
computed before for the unconstrained problem was used as the initial route in the algorithm. A
discussion about tuning the IDP parameters (C, γ, number of iterations) can be found in
Avgouleas (2008). The particular optimization scheme with the IDP settings mentioned above
was found to be the best compromise between speed of execution and accuracy of the solution.
Convergence of IDP Algorithm

86
85.5
85
84.5
Optimal Cost Function
84
83.5
83
82.5
82
81.5
81
0 2 4 6 8 10 12 14 16 18 20
Number of Iterations
FIGURE 22. Convergence to the Optimal Solution with γ=0.7.
To assess the sensitivity of the solution to the imposed safety constraints several trials were
conducted relaxing or lifting one constraint at a time and then both simultaneously. It is
concluded that relaxing the slamming constraint has no effect on the optimal solution. Deck
wetness, in contrast, is dominant. In all cases, violation of the slamming constraint occurred
together with violation of the deck wetness constraint, but the converse was not true. Altering
the limiting value for the probability of occurrence of this event (i.e. green water on deck)
changes the structure of the solution significantly. This suggests that the type and limiting values
of the constraints must be selected carefully and consistently with the risks that particular
vessels and their cargoes encounter in severe sea states. The observed redundancy of the
slamming constraint could justify enforcement of a different restriction, such as rolling angle or
acceleration. Some caution should be exercised regarding the accuracy of computation of ship
motions, if they are to be incorporated as constraints in the program. For example, prescribing a
rolling angle (RMS) restriction implies computation of rolling RAOs. If these computations are
performed on the basis of ideal flow principles the results may not be realistic, as roll is primarily
a viscous phenomenon.
Fuel Savings
There is no simple way to unambiguously quantify the savings in fuel cost in this constrained
optimization simulation. The lack of knowledge, in advance, of the navigation track the captain
would follow to avoid rough seas makes it difficult to directly compare it to the optimal solution
and evaluate its beneficial effect. An estimate may however be obtained by examining feasible
routes that satisfy the constraints and comparing them to the optimal.
Any route which brings the ship from the point of departure to the destination point without
violating the imposed constraints as well as the voyage final time belongs to the set of feasible
routes. The one yielding the lowest value of the assigned performance index (fuel cost, in

C-67
particular) is the optimal route. Given the difference in the forecasted severity of the sea
between loxodrome and orthodrome, as discussed earlier in this section, existence of a
plausible track was investigated to the south of the rhumb line route. This baseline-feasible, sub-
optimal route (figure 23) satisfies the prescribed constraints. It was determined by trial and error.
Calculation yields a fuel cost of 85.93 tonne for this route. The optimal route of figure 19
requires 5.7% less bunker fuel than the baseline feasible one. Table 5 summarizes the
calculated savings. Although the calculated saving rates are heavily dependent on the predicted
sea states of the particular example, the benefit of utilizing a decision support system for route
selection over sheer intuition or mariner’s experience is indisputable.
A Non-Optimal Feasible Route


60 W 30 W 0
Loxodrome
Feasible
45 N
30 N
15 N
FIGURE. 23. A Baseline Feasible Route South οf the Loxodrome.
TABLE 5. Summary of Fuel Savings.

Description Value
Fuel Cost of Baseline Feasible Route (tonne) 85.93
Fuel Cost of Nominal Route (tonne) 97.31
Fuel Cost of Optimal Route (tonne) 81.06
Fuel Savings Relative to Nominal (%) 16.7
Fuel Savings Relative to Baseline Feasible (%) 5.7
CONCLUSIONS
Fuel-efficient navigation is explored and a solution to the minimum cost routing problem is
proposed. An understanding of the governing physics of the problem in question has been
emphasized. The dynamics of ship motions and sea state evolution is best captured by state-of-
the-art tools like SWAN1 and WAM4. Particular attention should be paid to the added resistance
in waves, as it is a driving factor in the vessel routing problem. A comprehensive model for the
C-68
optimal control of a ship in waves is developed and a solution has been shown to exist and has
been generated using Dynamic Programming. This technique proves to be particularly
appealing as it circumvents problematic questions of existence and uniqueness of solution,
optimality conditions, existence of derivatives and gradients, all basic requirements of the
classical calculus of variations. Furthermore, it is ideally suitable for fast, real time
implementation onboard the vessel. Iterative Dynamic Programming, an intelligent alternative to
the standard DP algorithm, is found to produce fast results with reasonable accuracy. The main
advantage of this alternative, pointed out by Luus (2000) who introduced it, lies in its robustness
in obtaining the global optimum. Although there is always a possibility of obtaining a local
optimum, careful selection of the IDP parameters usually eliminates this risk and it is proven to
give remarkable results in challenging problems, compared to existing well-established methods
such as sequential quadratic programming and others. The structure of the optimal solution is
strongly affected by the enforced constraints. The findings reveal a great potential in the
application of optimal routing methods in shipping. The example-ship used in this paper is a
small cargo liner. The magnitude of the derived savings projects to a notable reduction of
bunker fuel expenses, especially for tankers, bulk carriers and containerships.
Accurate weather forecasts need to be the subject of a continuous research effort for the
results of the present fuel efficient routing algorithm to be significant and useful. The current
formulation carries out the optimization of the expected value of the cost function conditioned on
a given weather forecast known deterministically. Current weather forecasting technologies
allow for the accurate prediction of the expected sea states days in advance. This enables the
direct implementation of the optimal ship routing method developed in the present article for the
reduction of fuel consumption for trips of the order of weeks by breaking the sailing time into
sub-intervals over which the weather forecast is known with a high degree of certainty.
The shipping industry consumes approximately 4 million barrels of oil daily. Assuming that
cargo vessels sail into severe sea states approximately 25% of their sailing time, the
implementation of the fuel efficient routing algorithm developed in the present study, even using
a modest 5% fuel saving rate, would result approximately into a daily reduction of fuel
consumption of approximately 50,000 barrels of oil (=0.25x0.05x4,000,000). Other than the cost
savings by the shipping industry this reduction of crude oil consumption results in a reduction of
emissions of CO2 and other greenhouse gases into the atmosphere. The combined liquid fuels
obtained from the refining of an average barrel of oil will produce a minimum of 317kg of CO2
when consumed. Therefore a reduction of fuel consumption by the shipping industry of 50,000
barrels of oil daily would result in a reduction of 15,850 tonnes of CO2 emissions. These figures
may climb higher under conditions that result in greater fuel savings rates such as the 26%
estimated by Tsujimoto and Tanizawa (2006), or the 22.1% estimated by Mezaoui et al. (2009)
etc.
The optimal routing methods developed in the present study are readily applicable to the
optimal sailing of yachts at the America’s Cup level and in open ocean racing. The cost function
that must be minimized when sailing a yacht is the time to destination. Weather uncertainty
arises from the wind and wave forecasts and accurate forecasts are currently available a day or
more in advance. The resistance of sailing yachts in calm water and in a sea state is analogous
to that of ships, augmented by the induced drag caused by the wake of the keel, winglets,
rudder and sails. The “propulsion” of a sailing yacht arises from the lift force exerted by the wind
on the sails. The controls to be optimally selected by the dynamic programming algorithm are
the settings of the rudder and trim tabs and the trimming of the sails. Given the advanced state
of development of the hydrodynamic and aerodynamic performance attributes of hull forms,
appendages and sails the introduction of optimal navigation strategies based on real time
dynamic programming methods stands to offer a significant edge in competitive yacht sailing.

C-69
A Look Ahead
The model presented in this paper belongs to the broader category of white-box models,
namely models fully transparent in the incorporation of the governing physics to the best
possible extent. On the other extreme, black-box models exist that are oblivious of the
governing physics but rely on real time on-board measurements of relevant parameters to
derive the quantities of interest using system identification techniques. The abundance of
existing literature makes use of either approach, or combination of both (the so-called “grey-box”
models) and claims fuel savings ranging from 2% to over 25%. The savings anticipated by
deterministic optimal routing models are unlikely to materialize with the degree of confidence
expected by ship owners and operators. The cornerstone of weather routing is wave
forecasting. As already mentioned above, the state of the art in wave modeling and forecasting
is deterministic, yet the nature of predicting the future sea state is by definition stochastic.
Furthermore, variation of a range of parameters would lead to poor prediction of fuel
consumption even under the hypothesis of perfect knowledge of future weather. Just to name a
few, such parameters could be hull/propeller fouling, engine performance degradation, propeller
performance in waves, loading condition, water density, presence of unpredicted currents,
bunker fuel heating value and so on. A stochastic optimal weather routing formulation would
seek to determine the optimal control strategy which remains optimal under all state variables
and uncertainties associated with the problem. This task is challenging in multi-dimensional and
stochastic settings and is being addressed as outlined below.
Recent advances in control theory allow the explicit solution of the multi-dimensional
stochastic optimal control problem by casting the ship evolution equations in matrix state-
space form. In the context of fuel-efficient ship routing the states include: the ship fuel
consumption, the propeller RPM, the vessel coordinates relative to an earth-fixed frame, the
vessel horizontal velocities, the vessel yaw angle, the vessel yaw rate, the rudder angle, the
rudder angular velocity and the parameters discussed in the preceding paragraph. Casting the
coupled resistance, maneuvering and seakeeping problems in a multi-dimensional state-space
form and allowing for the weather forecasts to be known only stochastically leads to an
analytical and efficient determination of the ship route with the minimum fuel consumption
using the methods described in Yong and Zhou (1999). The generality and efficiency of this
optimal control algorithm would allow its real-time implementation on a ship using a desktop
PC. Its implementation to the fuel-efficient ship routing problem is very promising and is the
subject of ongoing research.
ACKNOWLEDGMENTS
The contribution of Professor George Kallos from the University of Athens and Professor
George Galanis from the Hellenic Naval Academy, Greece, is deeply acknowledged. They
provided the output of the WAM4 numerical model for wave and swell forecasting.
REFERENCES
1. Abramowski, Przemyslaw, Tomasz Abramowski, and Zenon Zwierzewicz. 2006. “Formal Solution of
Ship Weather Routing Problem via Pontryagin’s Maximum Principle”. Síntesis Tecnológica. 3 (1): 27-
31.
2. Avgouleas, K. 2008. “Optimal Ship Routing”. SM Thesis, Massachusetts Institute of Technology.
3. Bellman, R.E. 1957. Dynamic Programming. Princeton, N.J.: Princeton University Press.
4. Böttner, C. U. 2007. “Weather Routing for Ships in Degraded Condition”. International Symposium of
Maritime Safety, Security and Environmental Protection, Athens, Greece.
C-70
5. Bruns, A., K. Christiansen, and D. Rossow. 2011. “FSG.EcoPilot – an Onboard Tool for Fuel Efficient
Speed Profile Selection”. In 10th International Conference on Computer and IT Applications in the
Maritime Industries (COMPIT'11), Berlin, Germany.
6. Delitala, A. M. S., Gallino, S., Villa, L., Lagouvardos, K.,Drago, A. 2010. “Weather Routing in Long-
Distance Mediterranean Routes”. Theoretical and Applied Climatology 102 (1-2): 1-2.
7. Dolinskaya, I. S., M. Kotinis, M. G. Parsons, and R. L. Smith. 2009. “Optimal Short-Range Routing of
Vessels in a Seaway”. Journal of Ship Research, Society of Naval Architects and Marine Engineers 53
(3): 121-9.
8. Dolinskaya, Irina S. 2012. “Optimal Path Finding in Direction, Location, and Time Dependent
Environments”. Naval Research Logistics (NRL) 59 (5): 325-39.
9. Emmanouil, G., G. Galanis, and G. Kallos. 2007. “Assimilation of Radar Altimeter Data in Numerical
Wave Models: an Impact Study in Two Different Wave Climate Regions”. Annales Geophysicae, 25(3):
p. 581-596.
10.Faltinsen, O.M., K. J. Minsaas, N. Liapis, and S. O. Skørdal. 1980. “Prediction of Resistance and
Propulsion of a Ship in a Seaway”. In 13th Symposioum on Naval Hydrodynamics. T. Inui. Editor. The
Shipbuilding Research Association of Japan: Tokyo. p. 505-530.
11.Faltinsen, O.M. 1990. Sea Loads on Ships and Offshore Structures. Cambridge Ocean Technology
Series, 1, Cambridge; New York: Cambridge University Press.
12.Gerritsma, J., and W. Beukelman. 1972. “Analysis of the Resistance in Waves of a Fast Cargo Ship”.
International Shipbuilding Progress, 19(217): p. 285-293.
13.Gershanik V.I. 2011. “Weather Routing Optimisation - Challenges and Rewards. Proceedings of the
Institute of Marine Engineering, Science and Technology Part A: Journal of Marine Engineering and
Technology. 10 (3): 29-40.
14.Hinnenthal, J., and G. Clauss,. 2010. “Robust Pareto-Optimum Routing of Ships Utilising Deterministic
and Ensemble Weather Forecasts”. Ships and Offshore Structures 5 (2): 105-14.
15.Hotrop, J. 1984. “A Statistical Re-Analysis of Resistance and Propulsion Data”. International
Shipbuilding Progress, 31(363): p. 272-276.
16.Ilus, T., and A. Heikkinen. 2012. “Challenges in Vessel Speed Optimization”. In 11th International
Conference on Computer and IT Applications in the Maritime Industries (COMPIT'12), Liege, Belgium.
17.Kirk, D.E. 1970. Optimal Control Theory: an Introduction. Englewood Cliffs, N.J.: Prentice-Hall.
18.Kobayashi, E., T. Asajima, and N. Sueyoshi. 2011. “Advanced Navigation Route Optimization for an
Oceangoing Vessel”. In Methods and Algorithms in Navigation – Marine Navigation and Safety of Sea
Transportation, eds. Adam Weintrit, Tomasz Neumann, 149-155CRC Press.
19.Kosmas, O. T., and D. S. Vlachos. 2012. “Simulated Annealing for Optimal Ship Routing”. Computers
& Operations Research 39 (3) (3): 576-81.
20.Krata, P., and J. Szlapczynska. 2011. “Weather Hazard Avoidance in Modeling Safety of Motor-Driven
Ship for Multicriteria Weather Routing”. In Methods and Algorithms in Navigation – Marine Navigation
and Safety of Sea Transportation, eds. Adam Weintrit, Tomasz Neumann, 165-172CRC Press.
21.Luus, R. 2000. Iterative Dynamic Programming. Boca Raton: Chapman & Hall/CRC.
22.MAK 9M25C Project Guide –Propulsion, Issue July 2008, Caterpillar Motoren GmbH & Co. KG,
(downloaded from www.stet.pt).
23.Maki, Atsuo, Youhei Akimoto, Yuichi Nagata, Shigenobu Kobayashi, Eiichi Kobayashi, Shigeaki
Shiotani, Teruo Ohsawa, and Naoya Umeda. 2011. “A New Weather-Routing System that Accounts for
Ship Stability Based on a Real-Coded Genetic Algorithm”. Journal of Marine Science and Technology
16, (3) (09): 311-322.
24.Marie, S., and E. Courteille. 2009. “Multi-Objective Optimization of Motor Vessel Route”. In Marine
Navigation and Safety of Sea Transportation, ed. Adam Weintrit, 411-418CRC Press.
25.Marie, S., and E. Courteille. 2014. “Sail-Assisted Motor Vessels Weather Routing Using a Fuzzy Logic
Model”. Journal of Marine Science and Technology 19: 265-279.
26.Mezaoui, B., K. Takashima, and R. Shoji. 2009. “On the Fuel Saving Operation for Coastal Merchant
Ships Using Weather Routing”. In Marine Navigation and Safety of Sea Transportation, ed. Adam
Weintrit, 431-436CRC Press.
27.Montes, A. 2005. “Network Shortest Path Application for Optimum Track Ship Routing”. MSc thesis,
Naval Postgraduate School.
28.Ochi, M.K. 1998. Ocean Waves: the Stochastic Approach. Cambridge Ocean Technology Series, 6,
Cambridge, U.K.; New York: Cambridge University Press.

C-71
29.Sen D. and Padhy C.P. 2009. “Development of a Ship Weather Routing Algorithm and its Application
for Indian Coastal Sea Routes”. International Journal of Ecology and Development 12 (W09): 59-74.
30.Panigrahi, J. K., Padhy, C. P., Sen, D., Swain, J.,Larsen, O. 2012. “Optimal Ship Tracking on a
Navigation Route between Two Ports: A Hydrodynamics Approach”. Journal of Marine Science and
Technology 17 (1): 59-67.
31.Panigrahi, J. K., Tripathy, J. K., Umesh,P.A.,. 2008. “Optimum Tracking of Ship Routes in 3g-WAM
Simulated Rough Weather Using IRS-P4 (MSMR) Analysed Wind Fields”. Journal of the Indian Society
of Remote Sensing 36 (2): 149-58.
32.Pipchenko, O. 2011. “On the Method of Ship's Transoceanic Route Planning”. In Methods and
Algorithms in Navigation – Marine Navigation and Safety of Sea Transportation. eds. Adam Weintrit,
Tomasz Neumann, 157-163CRC Press.
33.Rathe, H., and C. Beiersdorf. 2005. “Decision Support for Container Ship Operation in Heavy Seas –
Shipboard Routing Assistance” In 4th International Conference on Computer and IT Applications in the
Maritime Industries (COMPIT'05), Hamburg, Germany.
34.Sclavounos, P.D., and D.E.Nakos. 1993. “Seakeeping and Added Resistance of I.A.C.C. Yachts by
Three-Dimensional Rankine Panel Methods”. In 11th Chesapeake Sailing Yacht Symposium.
Annapolis MD.
35.Shao W., Zhou P.,Thong S.K. 2012. “Development of a Novel Forward Dynamic Programming Method
for Weather Routing”. Journal of Marine Science and Technology (Japan) 17 (2): 239-51.
36.Skoglund, L., J. Kuttenkeuler, and A. Rosén. 2012. “A New Method for Robust Route Optimization in
Ensemble Weather Forecasts”. Trita-AVE / KTH Aeronautical and Vehicle Engineering 2012:67 .
37.Szłapczyńska, J. 2007. “Multiobjective Approach to Weather Routing”. Advances in Marine Navigation
and Safety of Sea Transportation. Proceedings of TransNav 2007, Gdynia 2007. Gdynia Maritime
University.
38.Szłapczyńska, J. 2013. “Multicriteria Evolutionary Weather Routing Algorithm in Practice”. International
Journal on Marine Navigation and Safety of Sea Transportation, TransNav 7 (1): 61-5.
39.Szlapczyńska, J., and R. Smierzchalski. 2007. “Adapted Isochrone Method Improving Ship Safety in
Weather Routing with Evolutionary Approach”. International Journal of Reliability, Quality and Safety
Engineering. 14 (06): 635-645.
40.Szlapczyńska, J., and R. Smierzchalski. 2009. “Multicriteria Optimisation in Weather Routing”. In
Marine Navigation and Safety of Sea Transportation, ed. Adam Weintrit, 423-429CRC Press.
41.Todd, F.H. 1963. “Series 60 Methodical Experiments with Models of Single-Screw Merchant Ships”.
Washington: For sale by the Supt. of Docs., U.S. Govt. Print. Off.
42.Tsou, Ming-Cheng and Hung-Chih Cheng. 2013. “An Ant Colony Algorithm for Efficient Ship Routing”.
Polish Maritime Research. 20(3): 3-50
43.Tsujimoto, M., and K. Tanizawa. 2006. “Development of a Weather Adaptive Navigation System
Considering Ship Performance”. In Actual Seas, 25th International Conference on Offshore Mechanics
and Arctic Engineering, Proceedings of OMAE 2006, Germany.
44.Wisniewski, B., J. Chomski, and P. Medyna. 2009. “Application of the 1-2-3 Rule for Calculations of a
Vessel’s Route Using Evolutionary Algorithms”. In Marine Navigation and Safety of Sea
Transportation, ed. Adam Weintrit, 419-422CRC Press.
45.Yong, J. and X. Z. Zhou. 1999. Stochastic Controls: Hamiltonian Systems and HJB Equations.
Applications of Mathematics Series, 43, Springer-Verlag, New York.
C-72
View publication stats

Avgouleas & Sclavounos 2014

Uploaded by

Copyright:

Available Formats

Avgouleas & Sclavounos 2014

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Avgouleas & Sclavounos 2014

Uploaded by

Copyright:

Available Formats

PART C: Natural Sciences and Mathematics

Fuel-Efficient Ship Routing

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

TABLE 1. Principal Dimensions of Example-Ship.

Beam at DWL B m 14.28

Displacement  tonne 5700

Block Coefficient CB - 0.676

Prismatic Coefficient CP - 0.688

Midship Section Coefficient CM - 0.982

TABLE 2. Propeller Characteristics of Example-Ship.

Expanded Area Ratio EAR 0.75

TABLE 3. Engine Characteristics of Example-Ship.

FIGURE 1. Steady Resistance from Holtrop Method.

The total calm water resistance coefficient is expressed as:

where Ca is the model-ship correlation allowance and Cw the wave-making resistance

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

The rate at which fuel mass is consumed in the engine is:

The quasi-propulsive efficiency ηD can be expressed as the product:

TRM  SGB (12)

Ship Dynamics in Waves

FIGURE 2. SWAN1 Computational Grid.

(x, t )  (x, t )   (x, t ) (16)

(x, t )   (x, t )  (x, t ) (17)

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

Heave RAO , Fn =0.21173

FIGURE 3. Heave RAO in Bow Waves at 13 Knots.

Pitch RAO , Fn =0.21173

FIGURE 4. Pitch RAO in Bow Waves at 13 Knots.

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

Heave RAO , Fn =0.21173

FIGURE 5. Heave RAO in Quartering Waves at 13 Knots.

Pitch RAO , Fn =0.21173

FIGURE 6. Pitch RAO in Quartering Waves at 13 Knots.

Added Resistance in Waves

The above force is obtained by integrating the total hydrodynamic pressure:

Rtotal  Rcalm  R1  R2  R3 (29)

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

Added Resistance Operator, Fn =0.21173

FIGURE 7. Added Resistance Operator in Bow Waves at 13 Knots.

Added Resistance Operator, Fn =0.21173

FIGURE 8. Added Resistance Operator in Quartering Waves at 13 Knots.

SEA STATE MODELING AND FORECASTING

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

spectrum and the RAO in heave:

S (0 , )   (  0 )  S (0 ) (39)

MINIMUM FUEL ROUTING

Powering and Fuel Consumption in Rough Seas

dS  U (t )  dt  Rearth  d (t )2  cos2  (t )   d  (t )2 (42)

the propulsor is balanced by the total resistance:

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

Specific Fuel Consumption Curve

Engine Load Diagram (MAK 9M25C)

FIGURE 11. Engine Operating Envelope (data source: ref. [22]).

Therefore, the control bounds can be described as follows:

U min  U  U max (49)

ISSN:1791-4469 Copyright © 2014, Hellenic Naval Academy

Vcr  0.093 g  L (51)

P(slamming)  0.03 (53)

P(water on deck)  0.07 (54)