Modeling of Powering Requirements for a Pelagic Trawler Leifur Arnar Kristjánsson
Faculty of Engineering - University of Iceland
Modeling of Powering Requirements for a Pelagic Trawler Leifur Arnar Kristjánsson
A thesis submitted in partial ful llment of the requirements for the degree of Magister Scientiarum
Department of Mechanical and Industrial Engineering Faculty of Engineering University of Iceland June 2005
A thesis submitted in partial ful llment of the requirements for the degree of Magister Scientiarum in Mechanical and Industrial Engineering at the University of Iceland. Modeling of Powering Requirements for a Pelagic Trawler Leifur Arnar Kristjánsson Author email:
[email protected] c Leifur Arnar Kristjánsson 2005 ° Committee in charge: Prof. Guðmundur R. Jónsson, chair Jón Ágúst Þorsteinsson, Ph. D. Moderator: Assoc. Prof. Ólafur Pétur Pálsson
Abstract Pure theoretical calculations of a ship's resistance in still water is a hard if not impossible task. With the additions of the in uence of wind, ocean waves and dierent loading cases of the ship, the task becomes even harder. For determining the velocity of the vessel, the thrust delivered by the propeller must be in an equilibrium with the total resistance of the ship. This makes the calculation of the shaft power and velocity of the vessel for given conditions even a more challenging task, given all the estimations that must be made for the calculations of the propeller's thrust. Arti cial neural networks oer an alternative way to tackle complex and ill-de ned problems. They can learn from examples, are fault tolerant in the sense that they are able to handle noisy and incomplete data, are able to deal with non-linear problems and once trained can perform predictions and generalisations at high speed. They are particularly useful in system modeling, such as implementing complex mapping and system identi cation. In this thesis, a model describing the shaft power and velocity of a speci c pelagic trawler is created using arti cial neural networks. The simulation model is a mapping, able to nd the required shaft power to drive the ship to a given velocity, for dierent weather conditions and sea states. The model is created for two operational states of the trawler, the steaming and trawling states and was originally made to be used in a real-time energy management system onboard a speci c trawler. However, the method proposed by the thesis is also intended for other ships, such as cargo vessels. The usage of arti cial neural networks turned out to suit the problem well and the results from the proposed model look promising compared to previously used methods. It is therefore suggested that the proposed model should be implemented and used to replace the current model in future versions of the energy management system.
Keywords:
Ship's Resistance; Shaft Power Requirements; Pelagic Trawler; Energy
Management; Arti cial neural-networks; System Identi cation
i
Ágrip á íslensku Nákvæm líkanagerð af mótstöðu skipa í lygnum sjó hefur í gegnum tíðina reynst mönnum er tt og oft á tíðum ómögulegt verkefni. Auk þess ækist líkanagerðin þegar taka þarf tillit til umhver sins og mismunandi hleðslu skipa. Við ákvörðun á hraða þarf þrýstikraftur skrúfu skipsins að vera í jafnvægi við mótstöðu þess sem gerir verkefnið enn er ðara með tilliti til þeirra aðferða sem þróaðar hafa verið fyrir útreikning á mótstöðu skipskrokka og þrýstikrafts skrúfu. Tauganet bjóða upp á annars konar aðferðir við útfærslu verkefnisins. Þau henta vel til líkanagerðar á ólínulegum tengslum inn og útmerkja og geta lært út frá mælingum. Í þessu verkefni er tauganetslíkan fyrir ása og hraða uppsjávarveiðiskips kynnt. Líkanið er hermlíkan sem notast við innmerki eins og stýringar skipstjórnarmanna, ástand skipsins hverju sinni, áhrif umhver sins og hleðslu skipsins. Líkanið var í fyrstu gert til notkunar innan orkustjórnunarker s um borð í togaranum, en auk þess getur sú aðferð sem hér er kynnt einnig verið notuð við gerð líkana af öðrum tegundum skipa, allt frá togurum upp í utningaskip. Notkun tauganeta reyndist vel við vinnslu verkefnisins og niðurstöður líkansins reyndust vel samanborið við fyrri aðferðir sem notaðar hafa verið. Því er lagt til að aðferðin verði tekin upp og notuð í framtíðarútfærslum á orkustjórnunarker nu sem hún var þróuð fyrir.
Efnisorð:
mótstaða skipa; ása ; uppsjávarveiðiskip; togari; orkustjórnun; tauganet;
líkanagerð
iii
Preface This thesis for the degree of Magister Scientiarum has been developed in collaboration with Marorka from mid-year 2004. During this time I've been able to combine my interest in system identi cation with the energy considerations for a shing trawler under dierent and everchanging weather conditions. Emphasis has been put on making the thesis more or less self-contained and as approachable as possible for the reader.
Acknowledgements I want to express my sincere gratitude to Professor Guðmundur R. Jónsson for supervising the work and to Dr. Jón Águst Þorsteinsson, my assistant advisor and Director of Marorka, who gave me the opportunity to work on my studies. Eskja hf, the owner of the ship being modeled, deserve special thanks for permitting the use of their data, which is the basis of the project. I'm also greatly indebted to Gísli Viggósson, the Director of Research and Development at the Icelandic Maritime Administration for providing the ocean wave data which contributed greatly to the thesis. Additionally, my colleagues at Marorka deserve many thanks for the nice working atmosphere during the year I've been working on my thesis. Financial support for this work was provided by Reykjavik Energy (Orkuveita Reykjavíkur). Their support made my graduate studies possible and is gratefully appreciated.
Leifur Arnar Kristjánsson Faculty of Engineering, University of Iceland Reykjavík, June 2005.
v
Table of Contents 1 Introduction
1
1.1
Background and motivation . . . . . . . . . . . . . . . . . . . . . . . .
1
1.2
Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1.3
Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2 Ship Powering 2.1
5
Hull resistance in still water . . . . . . . . . . . . . . . . . . . . . . . .
6
2.1.1
Holtrop and Mennen's power prediction method . . . . . . . . .
6
2.1.2
Usage of tank test results
. . . . . . . . . . . . . . . . . . . . .
7
2.2
Trawl resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.3
External in uences on resistance . . . . . . . . . . . . . . . . . . . . . .
7
2.3.1
Wind loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.3.2
Ocean surface waves . . . . . . . . . . . . . . . . . . . . . . . .
8
Propeller propulsion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2.4.1
Controllable pitch propellers . . . . . . . . . . . . . . . . . . . .
11
2.4.2
Ducted propellers . . . . . . . . . . . . . . . . . . . . . . . . . .
11
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.4
2.5
3 System Identi cation 3.1
3.2
13
Tasks in system identi cation . . . . . . . . . . . . . . . . . . . . . . .
14
3.1.1
Selection of model inputs and outputs
. . . . . . . . . . . . . .
14
3.1.2
Model architecture . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.1.3
Other aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
Arti cial neural networks . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.2.1
The perceptron . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.2.2
Multilayer perceptron network (MLP)
. . . . . . . . . . . . . .
16
3.2.3
Activation function . . . . . . . . . . . . . . . . . . . . . . . . .
18
3.2.4
Network training . . . . . . . . . . . . . . . . . . . . . . . . . .
18
vii
4 Data and Preprocessing
23
4.1
Characteristics of the pelagic trawler . . . . . . . . . . . . . . . . . . .
23
4.2
Available operational data . . . . . . . . . . . . . . . . . . . . . . . . .
25
4.2.1
Measured variables . . . . . . . . . . . . . . . . . . . . . . . . .
25
4.2.2
Handling of corrupt data . . . . . . . . . . . . . . . . . . . . . .
25
4.2.3
Other preprocessing . . . . . . . . . . . . . . . . . . . . . . . . .
26
Derived data and estimations . . . . . . . . . . . . . . . . . . . . . . .
27
4.3.1
State separation . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
4.3.2
Loading conditions . . . . . . . . . . . . . . . . . . . . . . . . .
28
4.3.3
Ocean waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
4.3.4
Ocean current . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4.3.5
Aging and other time passing eects . . . . . . . . . . . . . . .
29
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.3
4.4
5 Model Implementation
31
5.1
General description . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
5.2
Current implementation . . . . . . . . . . . . . . . . . . . . . . . . . .
33
5.3
Model quality metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
5.4
Construction of the proposed model . . . . . . . . . . . . . . . . . . . .
34
5.4.1
Base model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
5.4.2
Training data . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.4.3
Feature selection and extraction . . . . . . . . . . . . . . . . . .
38
5.4.4
Network size . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.4.5
Training method . . . . . . . . . . . . . . . . . . . . . . . . . .
43
5.4.6
Reduced number of output variables . . . . . . . . . . . . . . .
43
5.4.7
Results for the proposed model . . . . . . . . . . . . . . . . . .
44
Composite models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
5.5.1
Input space decomposition . . . . . . . . . . . . . . . . . . . . .
45
5.5.2
Results for the composite models . . . . . . . . . . . . . . . . .
45
Hybrid modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.6.1
Feature selection . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.6.2
Objective function . . . . . . . . . . . . . . . . . . . . . . . . .
47
5.6.3
Results for the hybrid model . . . . . . . . . . . . . . . . . . . .
47
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
5.5
5.6
5.7
6 Model Analysis and Applicability
49
6.1
Results for the test tours . . . . . . . . . . . . . . . . . . . . . . . . . .
49
6.2
Factors in uencing the ship powering . . . . . . . . . . . . . . . . . . .
53
6.3
6.4
6.2.1
Operational . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
6.2.2
Environmental
. . . . . . . . . . . . . . . . . . . . . . . . . . .
56
Model applicability . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
6.3.1
Performance based on velocity . . . . . . . . . . . . . . . . . . .
61
6.3.2
Performance based on heading . . . . . . . . . . . . . . . . . . .
62
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
7 Summary and Conclusions
65
7.1
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
7.2
Further work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
A Data Information
69
A.1 Overview of the tours . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
A.2 State separation criteria . . . . . . . . . . . . . . . . . . . . . . . . . .
70
A.3 Coordinate conventions . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
Bibliography
73
engineering; The application of scienti c and mathematical principles to practical ends such as the design, manufacture, and operation of e cient and economical structures, machines, processes, and systems. The American Heritage Dictionary
Chapter 1 Introduction In recent years increasing emphasis has been put on the reduction of emission of greenhouse gases and other environmental considerations that are partly or mainly related to oil combustion. Because of these concerns along with higher oil prices, the necessity for improved e ciency of systems burning oil has become increasingly more important. The aim of most manufacturers and designers, building system components which produce power by the use of oil, is to make those components as e cient as possible. Despite that, it is not always given that these components and other energy users that rely on the power generation are always run at their optimal e ciency level when they have been installed. Especially where they become part of bigger energy systems and run under various conditions. It can be assumed that with increased number of system components in energy systems, the possibility for improved e ciency also increases, which applies to both the right combination of hardware as well as economical operation of the systems.
1.1
Background and motivation
Based on the discussion above it becomes clear that the greatest potential for energy e ciency improvements is at the planning phase of a project which then decreases through the following phases of detailed engineering, implementation and operation. A new method for increased e ciency at planning stages during the design of a shing vessel was introduced by Thorsteinsson (2004) where energy users on board are selected and connected based on simulation and optimization results following given preconditions such as operational pro les of the system in question. As a part of his study, models of system components used on board typical shing vessels were created and have since then been implemented as a computer software program as described 1
2
Introduction
Ch. 1
by Stefansson (2004). The development of the methods discussed above have been adopted by Marorka, an Icelandic engineering company specializing in energy management and energy system research for the shing industry. Their software consists of Energy-System Design Tool, (EDT) a design and simulation software package intended for consultants, designers and machine manufacturers and Maren Energy Management System, (Maren). Maren makes use of real-time data from the machinery on board a vessel in conjunction with the simulation models created in EDT to assist in control and decision making with the aim of minimizing the oil usage on board. Included in Maren is a data acquisition module where the measured data and resulting calculations are stored in a database available for later usage. Studies with similar objective were also reported by Journée et al. (1987), where mathematical models of the propulsion system of a ship were created and used to give the shipmasters insight into its performance.
1.2
Objective
The objective of this thesis is to use the measurements collected in the Maren database to create an improved model presenting the shaft power requirements and log velocity1 for a speci c pelagic trawler. The main goal is to get more accurate results than current simulation models mentioned in section 1.1 are giving. This is done using system identi cation methods based on empirical data where emphasis is put on detecting factors that aect the propulsion system's power usage and are not included in the models being used today. The main usage of the project can be considered twofold as described below:
• Onshore analysis of vessel's operation at dierent operating conditions; and • The proposed method can t into Maren and due to its simulation ability be applicable as a part of its real-time decision support system It is important to notice that the proposed model is not intended to replace the current models which are used at the design stages of a new vessel since they will be created from empirical data. The necessary measurements are not available until after the ship has been in operation and must therefore be adapted for every ship and can not be used as a general model. 1 The
velocity can be described both as the log velocity which is relative to the ocean and ground
speed which is relative to the ground.
Sec. 1.3.
1.3
Outline
3
Outline
In chapter 2 the methods currently applied in the EDT models are discussed and explained including the hull's resistance, trawl resistance and the thrust provided by the propeller. Classical methods for the modeling of ocean wave are also discussed. This chapter should give the reader an overview of the subject of ship powering in general. Chapter 3 discusses the theory of modeling with emphasis on modeling from empirical data. The discussion is kept as general as possible at the beginning and then focuses on the usage of arti cial neural networks for multivariate nonlinear regression as used during the implementation described in chapter 5. Before the implementation of the model is discussed, chapter 4 is devoted to the data availability and data handling in general. The results of the model derived are analyzed in chapter 6 where the applicability of the model is also discussed. Conclusions and ideas for further work are nally given in chapter 7.
4
Introduction
Ch. 1
Chapter 2 Ship Powering The modeling of hull's resistance and the thrust delivered by ship's propeller is a di cult task and is usually, as described in the following chapter, derived using models based on regression analysis. The purpose of this chapter is to demonstrate the classical approach to the problem of ship powering and to introduce the ground that some of the methods described in chapter 5 are in uenced by. The following sections outline the general terms used in classical approach of the treatment of ship powering. This includes the resistance and the thrust that must be generated by the propeller to overcome the resistance in order to move the ship at speci c velocity. To give the reader insight into some of the main terms and notation used for the power at various places in the propulsion system they are given in gure 2.1. The gure shows a simpli ed diagram of the ship's propulsion system and shows the transmission of power from the prime mover to the propeller. The total resistance,
Brake Power, PB Prime Mover (Main Engine)
Shaft Power, P S
Reduction Gear, ? GB
Shaft and Bearings, ?S
Propeller ?R ?O
Hull ?H
Figure 2.1: Main components of a typical propulsion system showing the transmission of engine power to the propeller along with e ciencies.
R, that must be overcome consists of hull resistance in still water (section 2.1), trawl resistance (section 2.2) and resistance due to external factors such as ocean waves and wind (section 2.3) and can be described with
R = Rhull + Rtrawl + Rexternal .
(2.1)
The thrust generated by the propeller to overcome the total resistance is discussed in section 2.4. 5
6
Ship Powering
2.1
Ch. 2
Hull resistance in still water
With the usage of Bernoulli's law the general approach, which is still being used today, the resistance force, Rhull , on a ship in still water is given by
1 Rhull = ρocean V 2 Chull S 2
(2.2)
where ρocean is the ocean density, V the velocity of the vessel, Chull hull resistance coe cient and S the wetted surface of the ship's hull. The wetted surface of the ship's hull is the area of the ship that is underwater and is therefore highly dependant on the movements and loading of the ship at any given time. The coe cient of hull resistance
Chull is a dimensionless factor used to describe the performance of a ship's hull. It is generally considered to consist of two factors Cf (1 + k) and Cw presenting the frictional resistance and wave making resistance respectively according to
Chull = Cf (1 + k) + Cw
(2.3)
where k is a form factor describing the viscous resistance of the hull form. The wave making resistance refers to the energy loss caused by waves created by the vessel as it moves in still water.
2.1.1
Holtrop and Mennen's power prediction method
There are many resistance prediction methods known in the world of naval architects although Holtrop and Mennen's method presented by Holtrop and Mennen (1982) and Holtrop (1984) is probably the most commonly used. The method is based on regression analysis of tests made on 334 models and as said in Watson (1998), many naval architects nd it to give acceptable results although the number of the formula is rather complicated and the physics behind them are not clear at all which often is the case in regression analysis. At the time of the release of the Holtrop method, one of its main attractions beside its accuracy, was how it could be implemented in a computer program since many of the previously known methods involved reading from diagrams. From that time other power prediction methods have surfaced although the Holtrop method is still the most commonly used. Among those new methods is a proprietary regression model based on arti cial neural networks using data from MARINTEK1 , (Koushan, 2003). As with most power prediction methods it is mainly used at early stages of ship design before any physical models are created. 1 Norwegian
Marine Technology Research Institute.
Sec. 2.3.
7
Trawl resistance
The Holtrop method is also extensively used to determine the wake fraction coe cient and the trust deduction connected to propeller e ciency as described in section 2.4.
2.1.2 Usage of tank test results The total hull resistance, Rhull , of a ship can also be determined through model experiment results from towing tank as commonly used in the ship industry. These methods are usually carried out at later stages of ships design and give more accurate results than results from power prediction methods. The power prediction methods are still the best tool at early design stages since they do not require any physical model creation. Discussion about the usage of tank test results is given by Watson (1998).
2.2
Trawl resistance
The resistance of midwater (pelagic) trawl is expressed in similar manner as the resistance of a hull and is given as
1 Rtrawl = ρocean V 2 CDN ST 2
(2.4)
where ST is the total area of the trawl twine, V the velocity of the trawl relative to the ocean and CDN is the drag coe cient which can be expressed as a function of the Reynold's number according to Hu et al. (2000). A method for determining the total area of the twine is presented by Ferro (1981).
2.3
External in uences on resistance
In addition to the resistance aecting ship in still water there are several things in uencing the total resistance that change with time and position of the ship such as resistance due to wind (section 2.3.1) and ocean surface waves (section 2.3.2). The exernal resistance can therefore be described as
Rexternal = Rw + Raw
(2.5)
where Rw is the resistance caused by wind and Raw is the average resistance addition caused by ocean waves.
8
2.3.1
Ship Powering
Ch. 2
Wind loads
Wind resistance varies with time and location and is a function of the ship's sail area, wind velocity and angle relative to the ship at given time. Method to estimate the resistance added due to the direct aect of wind on a vessel was presented in Isherwood (1972) given by
where Vrw
1 2 Rw = ρair Vrw CXw AT (2.6) 2 is the velocity of the wind relative to the ship, CXw is a polynomial with
coe cients as a function of the wind's direction and the ship's structure and AT is the transverse projected wind area. Complete polynomials are given by Isherwood (1972). The wind can increase the resistance of a cargo vessels with high superstructures and a lot of cargo on the deck considerably while the aect on pelagic trawlers such as the one discussed by this study is considered to be a lot lower although it follows the same principles. Another method which is used by the EDT resistance model was presented by Comstock (1967).
2.3.2
Ocean surface waves
Due to the complexity of ocean wave systems it is usually di cult or even impossible to predict for the response of the vessel when encountering a wave system although predominant wave direction of wave travel and wave period can usually be determined. The information from these predominant waves can be used to determine the frequency
of encounter, ωe , which is, as the name implies, the frequency at which a ship moving forward encounters the waves where velocity, V , and the heading of the ship, µ, is taken into consideration. The equation for frequency of encounter in deep water is given by
ωe = ω −
ω2 V cos(µ) g
(2.7)
where ω is the frequency of the predominant waves and g is the gravitational acceleration.
Added resistance in regular waves Methods for the calculation of average resistance addition Raw due to ocean waves are presented by Gerritsma and Beukelman (1972) for larger wave lengths in head to beam waves where the ship motions in uence the resistance and by Faltinsen et al. (1980) which applies for shorter wave lengths. The latter method is derived by a direct pressure integration approach while the primer method is based on radiated
Sec. 2.3.
9
External in uences on resistance
energy method. Another method based on integrated pressure method, suitable for all wave directions, was presented by Boese (1970) and is explained in Journée and Massie (2001). All the abovementioned methods show that the mean added resistance is proportional to the wave amplitude squared. As shown in gure 2.2 the direction of the ocean wave relative to the vessel is important where almost all combinations of height and direction of the ocean wave increase the resistance of the vessel, (Bowditch et al., 1995).
20 18
Following seas
Ship speed (knots)
16 Beam seas
14 12 10
Head seas
8 6 4 2 0
2
4
6
8
10
12
14
16
18
20
22
Wave height (ft.)
Figure 2.2: Performance curves showing the aect of wave height and direction on the ground speed of a cargo ship, (Bowditch et al., 1995).
Wave estimation Since few vessels are equipped with wave measurement devices the information about the wave must be fetched through services providing this information for the desired area or predicted based on other available data such as wind and statistical probabilities. The Beaufort scale introduced by Sir Francis Beaufort in 1805 measures wind by observing ocean waves at sea and can give indication about the average relationship between the wind data and the signi cant wave height 2 , H1/3 , and wave periods. Another scale has since been adopted by the 17th ITTC (1984) showing the relationship between the sustained wind speed and the signi cant wave height and the modal wave period. The scales are currently the best tool available to estimate ocean waves based 2 The
average wave height of the highest 1/3 of waves from a recorded set.
10
Ship Powering
Ch. 2
solely on local wind information (wind sea) but since the waves are also in uenced by what happens at distant areas at dierent times (swell) other methods used to forecast waves have been developed. The oldest and simplest one use empirical relationship between the wind speed, fetch and duration while the latest ones use buoy and satellite altimeter wave data in conjunction with improved simulation models, (Stewart, 2004).
2.4
Propeller propulsion
The resistance from the hull and external conditions discussed above must be overcome by the propeller propulsion to move the ship at desired velocity. The dierence between propulsion testing in open water and behind a hull can be described with the introduction of two factors. The wake fraction coe cient, w, describes the changes of the propeller plane from being uniform in open water tests due to ship's displacement of water and the boundary layer around the hull and the thrust deduction coe cient,
t, describes the addition of resistance caused by the propeller since it increases the velocity of the water along the hull. With the assistance of the wake fraction coe cient and the thrust deduction coe cient the hull e ciency given by
ηH =
1−t PE = , PT 1−w
(2.8)
can be introduced as the ratio between the eective power, PE , and the thrust power,
PT , which the propeller delivers to the water. Holtrop (1984) provides a method for determining both the thrust deduction coe cient and the thrust deduction coe cient. In order to nd the open water power, PO , that must be delivered to the propeller the thrust power must be divided by the open water e ciency, i.e.
ηO =
KT J PT = . PO KQ 2π
(2.9)
A widespread method for determining open water characteristics of Wageningen Bseries propellers, which is the probably the most widely used propeller (Carlton, 1994), was reported by Oosterveld and van Oossanen (1975) and consists of polynomials obtained through regression analysis of test data from 120 propeller models. The method gives the performance characteristics of the propellers through the thrust coe cient,
KT , and the torque coe cient, KQ , based on its structure. Using those coe cients the trust, T , and the torque, Q, produced by the propeller can be found according to
KT =
T ρn2 D4
(2.10)
Sec. 2.4.
11
Propeller propulsion
and
Q (2.11) ρn2 D5 respectively, where n is the rotational velocity of the propeller and D is its diameter. KQ =
The non-dimensional terms KT and KQ are generally presented as a function of the
advance coe cient, given by
Va (2.12) nD where Va is the speed of advance, i.e. the ow velocity through the propeller which is J=
lower than the ship's velocity due to the wake velocity. The relative rotative e ciency,
ηR =
PO PP
(2.13)
re ects the dierence in torque in wake and in open water and must therefore be used to determine the power that must be delivered to the propeller through the propulsion shaft. Holtrop (1984) provides a method for determining the relative rotative e ciency. Finally the required shaft power, PS , can be found by dividing the shaft e ciency
ηS to the propeller power, PP . Another method, beside the one presented by Oosterveld and van Oossanen (1975), for determining the e ciencies of a propeller was provided by Reich and Barai (2000), where arti cial neural networks are used on empirical data.
2.4.1 Controllable pitch propellers The term propeller pitch is similar to a pitch of a wood screw and describes how long the propeller would screw itself into soft wood. The majority of propellers are either
xed pitch propellers (FPP) i.e. propellers with constant pitch or controllable pitch propellers (CPP) where the pitch of the blades can be changed while underway. The term pitch for CP propellers is often given in percentages ranging from 0% when the blade tips are orthogonal to the rotation of the propeller and 100% at full possible pitch.
2.4.2 Ducted propellers Ducted propellers, as their name implies, are surrounded by a ring that ts closely around the propeller. Originally the ducts where added to the propellers in order to protect them but it was soon found out that under special conditions the e ciency of the propellers can be increased with the usage of ducts. This applies to vessels
12
Ship Powering
Ch. 2
that need relatively high thrust and sail at low velocity and is therefore particularly good for tug boats and shing trawlers. Special polynomials for nding KT and KQ were derived for Wageningen ducted propellers by the authors of (Oosterveld and van Oossanen, 1975) and can be found in (Carlton, 1994).
2.5
Summary
The powering of shing vessels is mainly aected by the form of the ship's hull and the loading of the ship which changes the wetted surface area and the interface of the hull to the ocean, the external weather conditions and external equipment such as shing trawls. Beside these major resistance components, additional resistance can be related to steering and fouling. The resistance of the hull must then be overcome by the propeller's thrust that can be changed by changing the pitch of the propeller or by changing its rotational velocity. The relationship from the rotation of the propeller shaft, through the functionality of the propeller to speci c velocity by overcoming the resistance of the hull is highly non-linear and hard if not impossible to model using models derived only from rst principles.
Chapter 3 System Identi cation The importance of high quality models is fundamental at most disciplines where modules are used for system identi cation for various purpose such as prediction and simulation. The previously mentioned EDT models are no exception where the biggest energy user on board a pelagic trawler is the propeller which drives the vessel. As with most systems the model quality determines the upper bound on the quality of the whole Maren system and will therefore in uence the accuracy of all results concluded from that particular model. The application of the model presented by this thesis can be utilized in the eld of decision support, where it is used to nd an optimal operating point during a real-time operation, or an optimal input pro le to construct instructions for the system operator. An important aspect regarding the development of the model is the absence of feedback values since its purpose is not to serve as a control model. In the world of modeling there are basically three approaches that are generally used during system identi cation; white box modeling where models are derived from rst principles; black box modeling which are mainly based on experimental data and grey box modeling which uses a bit of both of the the other two methods. Due to the good availability of empirical data, the complexity of rst principles related to the problem presented in chapter 2 and its high non-linearity, the decision was taken to rely on the data and focus on black box modeling although known relations will be used to some extent.
13
14
3.1
System Identi cation
Ch. 3
Tasks in system identi cation
In its simplest form, the problem of model creation, is to optimize the structure and parameters of the model based on minimizing the error between its outputs and the known outputs for known input parameters. As shown in gure 3.1 the process of creating a model with the usage of empirical input and output data (supervised learning ) can be considered to be quite xed, independent of the methods selected. Training Data Input
Output
Model
-
+
Objective Function
Model parameters
Optimization Method
Figure 3.1: Supervised learning. Desired output of the model is compared to the given output, the error is evaluated by the objective function and the optimisation method takes care of adapting the parameters of the model.
3.1.1
Selection of model inputs and outputs
The selection of model outputs is usually quite clear from the start since it must represent the purpose of the whole hassle of creating a model. The selection of the input variables can on the other hand be a little more tricky and often requires some investigation. This can be based on supervised methods which are usually complex optimisation problems in the case of nonlinear models or unsupervised methods such as principle
component analysis (PCA) which makes it possible to discard irrelevant inputs and by that means reduce the complexity of the model. Finally the selection can be based on prior knowledge using heuristics for correct combination of logical inputs. In some cases it might be reasonable to use some kind of input space decomposition technique to separate the input variables based on dierent output behavior and by that means create separate models for dierent states or conditions of the input variables.
3.1.2
Model architecture
For the selection of the model architecture there are several aspects that must be considered, both with respect to the system being modeled and the proposed usage of the model. Due to the known nonlinearity of the proposed model it soon came
Sec. 3.2.
Arti cial neural networks
15
clear that the selected method had to be able to comply with these requirements which ruled out possibility such as linear regression and because of the fact that models with parametric architecture often make drastic assumptions about the behavior of the system they were soon ruled out, (Nelles, 2001). Interesting method that was taken into consideration as the method of choice was a neuro-fuzzy method called ANFIS. This is a neuro-fuzzy network and as such has a functionality equivalent to fuzzy inference systems but has a network architecture and trainable parameters which can be compared to conventional neural networks. The fuzzy models are therefore not solely designed by expert knowledge such as pure fuzzy logic but are also trained from data. ANFIS was originally presented by Jang (1993) and is discussed and explained in (Jang et al., 1997). The main reason for the selection of the arti cial neural network architecture mainly lies in those assumptions made above along with the requirement of the model to be able to have multiple outputs and the big availability of literature and tools.
3.1.3 Other aspects Some aspects of the tasks in system identi cation such as the structure and complexity of the model along with the optimisation of the model parameters are discussed in section 3.2 while others such as the validation and testing of the model are left for the implementation in chapter 5.
3.2
Arti cial neural networks
Arti cial neural networks were originally motivated by the principles observed in nervous systems and are loosely based on its functionality were neurons receives signals from other neurons, generates and distributes response to other neurons based on the input and some internal state. It should be noted that the discussion about the networks in this section only represents the networks used by this thesis and are far from being a complete discussion of the types and capabilities of arti cial neural networks in general.
3.2.1 The perceptron An example of an arti cial neuron is shown in gure 3.2, where g is the activation
function usually representing the nonlinearity and Σ is the summation of the input
16
System Identi cation
Ch. 3
values x and the corresponding weights w. The term perceptron is often used for this type of neuron although it was originally used by Rosenblatt (1962) for a singlelayer network with threshold activation function, (Bishop, 1999). The output of the
wj0 x. 1 . . xi . .
wj1 g wji
aj
zj
Figure 3.2: An arti cial neuron model often referred to as perceptron. The input to the neuron consists of weighted values from the input and a bias.
perceptron is given by
zj (x) = g
à d X i=1
! wji xi + wj0
=g
à d X
! wji xi
(3.1)
i=0
where g is the activation function, which is a function of the input xi and the weight
wji . The bias can also be included as weight by the introduction of x0 = 1. Although a single arti cial neuron is not very powerful for system identi cation, a collection of arti cial neurons can represent a whole arti cial neural network of arbitrary size and structure which has proven to be a powerful tool for both classi cation and
regression problems. The most widely used arti cial neural network type is the so called multilayer perceptron network (MLP) which is a set of single-layer perceptrons constructed in layers with separate layers for input and output variables and zero or more layers in between. The layers between the input and output layers are often referred to as hidden layers since they are not observable from the input nor the output variables. The networks used in this study are fully connected, feedforward networks which means that all neurons in layer i receive signals form every neuron in layer i − 1 and connect to every neuron in layer i + 1 without any feedbacks to previous layers.
3.2.2
Multilayer perceptron network (MLP)
A collection of several perceptrons as the one shown in gure 3.2 can make up the previously mentioned multilayer perceptron network in gure 3.3. The MLP network is the most widely used neural network architecture and is often used as a synonym for all neural networks, (Nelles, 2001).
Sec. 3.2.
17
Arti cial neural networks
Input layer d inputs
Hidden layer M hidden neurons (1)
w11
x1 x2
Output layer c = 2 outputs
(2) w11
(1) w12
. . . . .
y1 . . .
w1p(1)
xd
w2(2) 0
(1)
w10
y2
(2) w1M
(1)
wM0
z20
x10
Figure 3.3: Two layer MLP network with p input variables, sigmoid activation function in the hidden layer with M hidden neurons and linear activation function for the k (in this case two) outputs.
The naming conventions used in this section are presented in gures 3.2 and 3.3 where xj stands for one of the d input variables. The weights, w, for the input into the hidden layer have subscripts indicating the number of the input they come from and the number of the hidden neuron they go to respectively and their superscript is the number of layer they connect to. The total number of hidden neurons is M . The outputs of the j neuron is referred to as aj before it has been transformed using the activation function, g , and is then called zj after its transformation. The output of the network is then called yk where the total number of outputs is c. A complete function for the representation of the MLP network shown in gure 3.3 is given by
yk = ge
ÃM X j=0
(2) wkj g
à d X
!! (1) wji xi
(3.2)
i=0
where the bias is absorbed into the weights. The activation function of the output layer is presented as ge to emphasize that it must not be the same function as the one used for the hidden layer. In this study the activation function for the output layer is linear, i.e. ge(a) = a. In (3.2), which shows the k output of the network, the primer subscript of the weights represents the number of the receiving neuron while the latter is for the neuron sending the signal. The weight's subscript represents the number of the activation layer where M is the total number of hidden neurons and d is the number of outputs. Finally
g is the activation function of the hidden layer.
18
System Identi cation
3.2.3
Ch. 3
Activation function
The activation function which transforms the input of a neuron to its output is typically chosen to be of saturation type and the most commonly used are sigmoid functions such as the logistic function logistic(x) =
1 . 1 + exp(x)
(3.3)
and the hyperbolic tangent
tanh(x) =
exp(x) − exp(−x) 1 − exp(−2x) = . exp(x) + exp(−x) 1 + exp(−2x)
(3.4)
These functions map the interval (−∞, ∞) onto (0, 1) and (−1, 1) respectively and their derivate can be expressed as a simple function of their output which is an advantage in any gradient-based optimisation method used during the training of an MLP network. Another important property of the sigmoid functions, which get their name from their shape, is their ability to represent both linear tendencies in the case of low weight values and highly nonlinear behavior with the usage of higher weights.
3.2.4
Network training
The training of the network aims at minimizing the objective function which is calculated from the comparison of the network outputs and the desired output according to the training data. At each iteration the weights of the network are updated according to the selected optimisation method until the network has been trained to ful ll previously de ned stop criteria. Training usually aims at minimizing the error of the objective function with consciousness regarding the generalization of the model which is the ability of a model to represent other situations than those speci cally given in the training set. The ability to learn instead of just remembering known situations. The methods presented here only apply to the training of the weights of the network after its size and structure has been set. Other methods which are either constructive
methods where new nodes are added to undersized network or pruning algorithms where oversized network is trimmed down are known in the search for the optimal size of the network.
Training data normalization Training of the networks can be made more e cient by applying preprocessing of the input data. Among the most commonly used methods are standardizing where the
Sec. 3.2.
19
Arti cial neural networks
mean of the data set is set to zero and the standard deviation is set to unity given by
x eni
xni − xi = σi
(3.5)
where xi is the mean for variable i, xni is an instance of variable i in pattern n (or at time n) and σi is the standard deviation of variable i. Another commonly used method is called scaling where all the data fall within a speci ed range, (Demuth and Beale, 1998).
Feature extraction and selection The measured and estimated data presented in chapter 4 must in some cases be processed further on order to get better results from the training of the network model. The main objectives of feature extraction are to unite some of the input variables into a lower dimensional feature space while the objective of feature selection is to reduce the number of input variables to those who add contribution to the model. Commonly implemented feature extraction method is PCA which transforms input data so that all input vectors will be uncorrelated and removes input variables that do not contribute enough to the output variables, (Demuth and Beale, 1998). For feature selection, prior
knowledge is of great importance and heuristics such as forward selection where one new input variable is added at a time until some stop criteria for the changes in train error is reached are also common.
Weight initialization If the training data has been preprocessed as discussed above, it should be of order unity and the initial values of the weight can therefore be randomly selected in the order of unity as well. Otherwise a solution for the weight, where some weight had markedly dierent values from other, had to be found, (Bishop, 1999). Beside the most common and simplest ways to distribute the initial weights by random selection, other methods have been created. Among those is the Nguyen-Widrow initialization algorithm which chooses the values in order to distribute the active of each region of each neuron in the layer evenly across the layer's input space, (Demuth and Beale, 1998). This method has been shown to outperform most common random techniques, (Pavelka and Prochazka, 2004).
20
System Identi cation
Ch. 3
Objective function The most commonly used objective functions for regression problems using the MLP network are the mean square error or the sum of square error shown in N
E=
c
1 XX (yk (xn ; w) − tnk )2 2 n=1 k=1
(3.6)
where tnk is the target variable for the k output of the n input pattern. The objective function can also contain penalty term to limit the error to some interval or to prevent certain conditions to arise.
Backpropagation Backpropagation 1 is a method for calculating the derivatives of a MLP network nodes with respect to its weights which are then adjusted using the derivatives according to the selected weight update algorithm. This method gets its name from the way the derivatives are calculated from the result of the objective function backwards to the hidden layers. Having obtained the error E from an objective function its derivatives with respect to weights w can be expressed as
∂E n = δ j zi ∂wji
(3.7)
where δj ≡ ∂E n /∂aj and zi = ∂aj /∂wji which can be seen from gure 3.2 and (3.1). The δk , for the output nodes, can be obtained by
δk ≡
∂E n ∂E n ∂yk ∂E n = = g 0 (ak ) ∂ak ∂yk ∂ak ∂yk
(3.8)
while the δj for the hidden nodes can be found by the application of the chain rule
δj ≡
∂E n X ∂E n ∂ak = ∂aj ∂ak ∂aj k
(3.9)
where the rst factor is just the δk of the k node and the second factor is obtained by 0
noting that if node i connects directly to node k then ∂ak /∂ai = gi wki otherwise its zero, which gives us the backpropagation formula
δj = g 0 (aj )
X
wkj δk .
(3.10)
k 1 Due
to its popularity backpropagation has since then often been used as a synonym for the applied
training method or even for the MLP network architecture.
Sec. 3.2.
21
Arti cial neural networks
Weight update algorithm Having used backpropagation to nd the derivatives of the error E with respect to each weight w for each input pattern n the total derivate can be obtained using
∂E X ∂E n = . ∂w ∂w n
(3.11)
When all input patterns are presented at the same time as shown above the method is called batch learning while using on-line learning the weights would be updated after each input pattern. There are several optimization techniques available for updating the weights of the network where the simples is the gradient descent method which uses xed step size in the direction of the steepest descent in every iteration. Using Levenberg-Marquardt algorithm, a combination of the steepest gradient descent ∂E/∂w and the Newton direction H−1 ∂E/∂w as shown in
w(t + 1) = w(t) − (H + λI)−1
∂E , ∂w
(3.12)
where t represents each epoch. The weights are updated using the derivatives after each epoch (iteration) according to (3.12) where λ is dynamically adjusted parameter and H is the Hessian matrix. The updated weights are then forward propagated thought the network, the error calculated using the objective function, gradients calculated and nally the weights updated again. This procedure is continued until speci ed stop criteria is ful lled. According to Demuth and Beale (1998) the Levenberg-Marquardt algorithm appears to be the fastest method for training moderate-sized feedforward neural networks beside the fact that it has very e cient MATLAB implementation. Other sources such as Reed and MarksII (1999) state that the Levenberg-Marquardt algorithm seems to be good on moderately sized problems when compared to conjugate gradient descent and gradient
descent.
Stop criteria Due to the fact that an MLP network can approximate any smooth function to arbitrary degree of accuracy with increased number of hidden layer neurons2 , care must be taken during its training to avoid over tting. Among the methods used to determine when to stop training is a method called early stopping where the training set is divided into 2 This
is so-called universal approximator, (Reed and MarksII, 1999)
22
System Identi cation
Ch. 3
two subsets where part of the data is actually used for training and a part used for
validation. The training error normally decreases during training until it reaches zero (in case of smooth, noise free data) while the validation error will normally start to increase at some point which is a strong sign for possible over tting of the data. At that time the training is stopped and the weights from the iteration containing the lowest validation error are used. The training data, including validation data is used during the training phase of the network while test data which is never used for the validation of the model during training is used to evaluate its quality independent of the training phase.
Chapter 4 Data and Preprocessing The majority of the available data used by this study is fetched from the database underlying the Maren system onboard the pelagic trawler. This data is either direct measurement from various meters installed onboard or data derived from those measurements using well known rst principles. The available data range includes tours over a period of four months, total 15 tours that last from few days to over a week. The period for blue whiting tours includes tours in October, November and December 2004 while the capelin tours are all from January 2005. No new meters were speci cally installed as a part of this study, since its aim is to use the data already being collected onbard, using the Maren system. Because of the nature of the model, its input data consists of operational variables of the vessel along with measurements describing some environmental conditions derived from the measurements. It contains no information regarding the size or form of the vessel as would be the case for a more general model. Some basic characteristics of the vessel being used are given in section 4.1, followed by the available operational data which is discussed in section 4.2. The remaining of the chapter is dedicated to data derived from the measured variables and some additional information that were collected especially for this study.
4.1
Characteristics of the pelagic trawler
The data used in this study is taken from the pelagic trawler Jón Kjartansson SU-111. Its overall length is close to 70m where length between perpendiculars is 62m. The breadth on waterline is around 10m and draught is close to 7m. The form of the vessel can be seen in gure 4.1. 23
24
Data and Preprocessing
Ch. 4
Figure 4.1: Form of the hull and the location of the main engine for the pelagic trawler, Jón Kjartansson SU-111, used in this study.
The velocity of the ship is mainly controlled by changing the pitch of the propeller as described in section 2.4.1 although the velocity can also be controlled through the two available rotational velocities of the main engine. The higher velocity is 750 RPM producing maximum power of 4920 kW while the lower is 650 RPM giving 3120 kW at maximum load, (Wartsila). For later usage, the speci c fuel consumption curves are given in gure 4.2 where it can be seen how the minimum speci c fuel consumption is reached around 85% load.
230
240 750 RPM 650 RPM
220 210 200 190 180 1000
2000 3000 4000 Brake power [kW]
Specific fuel consumption [g/kWh]
Specific fuel consumption [g/kWh]
240
750 RPM 650 RPM
230 220 210 200 190 180
40
60 80 Engine load [%]
100
Figure 4.2: Speci c fuel consumption of the vessel as a function of the brake power and engine load for the two possible engine rotational velocities.
The data used by this study includes the usage of two dierent types of pelagic trawls, one for catching capelin and the other for blue whiting where a single trawl is used in both cases. Beside the two types of trawling periods, each tour includes steaming states where the vessel sails to, from and between shing areas. The term
state will be used throughout this thesis to describe the operational state of the vessel which are mainly the two previously mentioned states, trawling and steaming. Since the proposed model is only intended for this speci c trawler the gures given above, along with the type of the trawler are descriptive enough. If the intension was
Sec. 4.2.
Available operational data
25
to create a general model for all, or some speci c size ranges of pelagic trawlers, the parameters describing the trawler would of course play role in the design of the model and the demand for data would be a lot greater.
4.2
Available operational data
The available operational data is collected from meters that are installed as a part of the Maren system or from previously existing PLCs or systems, using the NMEA and OPC protocols and stored in an underlying database. The resolution of the measurements in the Maren database is 15s although the measurements are not taken nor logged simultaneously. Although the measurements in the database only have resolution of 15s, the measurements are available to the Maren system at shorter interval and in current installation the EDT model is calculating the status of the vessel every 2s. During this study it is assumed that all meters are correctly calibrated and no special study is conducted with respect to their accuracy or noise. Incorrect measurements used as an input data will not aect the quality of the model, if it is only shifting or some kind of ampli cation error, but any later corrections made to those measurements require the model to be trained again using the new data or corrections of the old data.
4.2.1 Measured variables Currently there are over 750 dierent measured entities stored in the database along with 39 variables that are derived from the measurements. Of these measurements there are around 20 that can in some way, directly or indirectly, be connected to the modeling of the required shaft power of the vessel and are therefore used during the study. In general, these variables can be classi ed as direct operational variables such as rotational velocity of the main engine and the pitch of the propeller; environmental
variables such as wind and ocean current ; and the indirect variables that are used to estimate the operational condition at each time. In addition to the variables mentioned above, the variables used to train the model, the shaft power and the log velocity, are taken from the Maren database.
4.2.2 Handling of corrupt data During the implementation, each variable is collected in a vector where data is synchronized every 15s resulting in a complete set. These sets can contain various kinds of corruption where the most common ones are either frozen data where the meter keeps
26
Data and Preprocessing
Ch. 4
sending the same value although the conditions are changing and in missing data where nothing is logged. The handling of those corruptions can be done in various ways depending on the intended usage of the data. In the case when the quantity of the available data is su ciently large and the data values is independent of the data itself, i.e. modeling of the system is not depending on preceding values as is common in time series analysis, the simplest method is to discard the whole data pattern for that particular time. For cases where missing data must be lled in with values, one approach is to model these intervals based on the known data although it is prone to problems, especially if the series contain a lot of corrupted data. The available data contained series without any missing data which was used to show that the data values were more or less independent of the data it self. The method of discarding data patterns where some of the data was missing was therefore chosen. For the cases where the output data had remained unchanged over ten sampling patterns, all the ten patterns were removed from the data set. For some input variables the occurrence of consecutive recurrent values can often be normal. In those cases the variables were left unchanged (such as the loading of the vessel) while other input variables which are known to change a lot or contain a lot of noise where removed in a similar manner as the output variables.
4.2.3
Other preprocessing
The content of this section focuses on the preprocessing of the data in terms of its availability, correctness and accuracy and does not involve the real content of the data such as feature extraction or special preprocessing for the neural network such as
normalization. Outliers may occur in the data set as a result from wrongly observed or recorded measurement. They may in many cases result in a gross error and should in those cases be excluded from model tting. In the cases were outlier removal was tested the
Grubbs test was used, (NIST). Among other cases of simple measurement handling include the changes of relative data such as wind direction and ship direction to absolute data since none of the model input data can contain relative data due to the fact that the vessel direction must be controllable as an input into the simulation model.
Sec. 4.3.
4.3
27
Derived data and estimations
Derived data and estimations
Beside the directly measured data, other factors that are known to aect the vessel's shaft power and velocity should be included in the model. The three most important factors that will be estimated, either from the data measured on board or from other available external services are the state of the ship, the loading of catch onboard and the state of the ocean waves. The possible aect of fouling is also discussed.
4.3.1 State separation The overall operation of a trawler can be broken into operational states within each shing tour. The example used here represents a typical tour for a vessel catching pelagic species such as blue whiting and is shown in gure 4.3. The vessel leaves
Fishing area A. Harbor
B. Steaming
1. Pay out
E. Harbor
Steamingotother fishingarea
2. Trawling
3. Hauling
D. Steaming
4. Pumping
Figure 4.3: An example of a typical shing tour and its state separation for purse seiners catching pelagic species
harbor, steaming to the sh area. The states from 1 to 4 are then repeated until the vessel is full or the maximum time for the shing tour has passed. Then the vessel returns to harbor where the catch is landed. The state de nition is an important part of the model and is used to decide which input parameters or even which model is to be used at each state. Beside that, the pumping condition is used to estimate the loading of the vessel as discussed in section 4.3.2. The criteria for the state separation for the steaming and trawling states mainly relies on the shaft power, the velocity of the vessel and the values of the trawl wire tension and length. In addition to that, the current for the sh pumps is monitored for the detection of the pumping state. The complete criteria is given in appendix A.2.
28
Data and Preprocessing
4.3.2
Ch. 4
Loading conditions
An important factor for the vessel's resistance is its loading, since it changes the interface between the hull and the ocean. The loading consists of various factors where the most important and changeable is the loading of catch in the shtanks. Factors such as oil usage and other consumption on board might aect the loading inconsiderable compared to the sh loading. These factors are left out due to their insigni cant in uence on the total loading and their unavailability. Since there are no measurements available regarding the status of the mixed seawater1 nor amount of catch in the shtanks the loading must be estimated by other means. This has been done be combining the knowledge of the amount of sh landed after each tour and the total time of the pumping state in the corresponding tour. Based on this information and the capacity of the pumps used to pump the sh on board, the additional loading due to the catch can be estimated from the time of the pumping state. It is also assumed that the loading into the sh tanks is always performed in the same way since the loading of each tank is currently not being measured. With the addition of those measurements, the loading information would be a lot more accurate and could provide better input to the model. In addition to that, the way the sh is loaded into the tanks could also be a part of the model. A more di cult thing to estimate is the aect of loading of sh in the trawl during each trawling period. Correlation between the trawl wire tension and the shaft power or the log speed was not detected which indicates that the loading of the sh into the trawl is a minor factor in the total trawl resistance. Beside that no measurements for the sh loading are available in the Maren database although these measurements are being made onboard the vessel.
4.3.3
Ocean waves
The Maren database contains no information about the ocean waves or the movements of the vessel in the waves. This information is on the other hand available from various weather services where the waves are usually identi ed by three parameters, the
signi cant wave height, wave period and wave direction as explained in section 2.3.2. The historical ocean wave information used in this study was obtained from the Icelandic Maritime Administration were the processed data is available for every six hours (at 0, 6, 12, 18 hours) on a grid with resolution of 0.5◦ . This data was then 1 For
erations
use with pelagic species the sea must be mixed with water (30/70) due to sh quality consid-
Sec. 4.4.
Derived data and estimations
29
interpolated using cubic splines to get values for the maximum resolution of the EDT database data, 15s. In the absence of the data from the Icelandic Maritime Administration an investigation was made by using the wind information to generate wave data using a table for the North Sea Areas which gives an approximation of the main properties of the ocean waves. The table is a result from the Joint North Sea Wave Project (JONSWAP) and is just a rough estimation of the state at each time and merely takes into consideration that the waves grow with the time in constant wind, (Journée and Massie, 2001). The usage of this method did not improve the proposed model and was soon dropped.
4.3.4 Ocean current The ocean current is more or less left out of this study. In the presence of the required information about ocean currents the ground speed can always be calculated from the log velocity. The ground speed is therefore left out of the scope of the proposed neural network model but might be a part of the postprocessing when the information of the current is available. For most Icelandic shing vessels the log velocity is of interest during the trawling states while the captain is more concerned about the ground speed while steaming to shing area at certain location or back to harbor with the catch.
4.3.5 Aging and other time passing eects There are certain relations to time that are known to aect the total resistance of the vessel such as fray of the trawl twins and paint erosion of the hull (fouling). In order to implement these factors into the model, two counters were created, one for the paint erosion which is always ticking as long as the vessel is sailing and the other is for each of the two trawls that only tick during the usage of that particular trawl. These counters must be reset when new trawls are taken into usage or when the painting of the hull is renewed in a dockyard. Since neither the usage of the trawls or the age of the painting is known with respect to passed sailing and trawling time, the counters were set to zero at the beginning of the study. This should serve as a proof of concept while they must be correctly set again when new trawl or painting is taken into usage.
30
Data and Preprocessing
4.4
Ch. 4
Summary
The preprocessing of the data presented in this chapter is only related to the data in general and does not include any of the additional preprocessing made for feature selection or other actions to assist the neural network. The estimated measurement error is not addressed at this time and remains an open issue for future work. The data is partly taken directly from the Maren database although important variables must also be estimated from the available information. Since information about the ocean wave is not measured onboard the vessel, that data was provided by the Icelandic Maritime Administration. An example of the input and output variables used is given in gure 4.4 showing
°
the data for the test tours presented in chapter 5. Input: WindAngle
200 0 50 40 30 20 10
200
300
400
500
600 700 800 Input: WindStrength
80 60 40 20 0 150 140 130 1500 1000 500
100
200
300
400
500
600
700 800 Input: Pitch
100
200
300
400
500
600
700 800 Input: RPM
100
200
300
400
500
600 700 800 Input: FishLoading
6 4 2
100
200
300
400
500
600 700 800 Input: WaveHeight
300 200 100
100
200
300
400
500
600 700 800 Input: WaveDirection
3 2.5 2 14 12 10 8 6 4 2
100
200
300
400
500
600
100
200
300
400
500
600 700 800 Output: VelocityLog
100
200
300
400
500
600 700 800 Output: ShaftPower
100
200
300
400 Samples
500
600
kW
knots
°
m
tonnes
rpm
%
knots
100
3000 2000 1000
700 800 Input: StateNum
700
800
Figure 4.4: Example of the input and output variables used. The data ranges over the four test tours presented in section 5.4.2 where the number of data points have been reduced to around 800.
Chapter 5 Model Implementation This chapter unites the content of previous chapters. The methodology described in chapter 3 is used on the data presented in chapter 4 to create a model for the propulsion system described in chapter 2. After general description of the problem, current implementations are addressed and nally the construction and evolution of the proposed model is explained.
5.1
General description
It is extremely di cult to derive shaft power model from rst principles where it would be practically impossible to estimate or measure some of the required parameters at all times. The model is assumed to be highly connected to the ocean surface waves and weather conditions at given times while the best known models for the resistance of the hull and the thrust given by the propeller are based on regression analysis, often with little connection to known rst principles, (Journée and Massie, 2001; Watson, 1998). Due to those reasons along with the number of desired outputs of the model and the fact that model will be created based on the measurements and estimations discussed in chapter 4, the neural network architecture is a promising method for the identi cation of the system. Especially since there are no intermediate measurements available to identify the characteristics of the hull's resistance or the propeller separately. Other methods such as the Holtrop method discussed in section 2.1.1 require inputs that are hard to measure oshore and with the addition of known methods for oshore eects such as ocean waves the model becomes complicated, including the polynomials related to the calculation of the propellers thrust. As stated in chapter 1, the objective is to create a simulation model based on
known external conditions and operational conditions of the vessel in the absence of 31
32
Model Implementation
Ch. 5
direct measurements of the vessel's behaviour. The proposed model can therefore be used to simulate the behavior of the vessel at other conditions than those it is exposed to at current time. It should therefore be emphasized that the model is not a control model aiming at moving the vessel to a certain condition, it is rather a simulation model to be incorporated in a decision support system where decisions are made from optimal condition derived from the model. The following variables are of interest for the simulation of the power requirements for the pelagic trawler:
Required shaft power, PS : The power delivered from the main engine after the e ciency reduction in the gear. It aects the load of the engine along with other interesting variables such as fuel consumption.
Log velocity, vlog : Velocity of the vessel relative to the ocean. Ground speed can be calculated from the log velocity if information about the ocean currents is given. The model should be able to predict the quantities given above form the relevant inputs, which in the case of the pelagic trawler are used to describe:
Propeller pitch: This is the main control mechanism for controlling the velocity of the vessel.
Propeller revolution: The main engine of the vessel is con gured to have two possible settings for the rotational velocities.
Vessel heading: Used to nd out the relative angles of both waves and wind direction.
Loading of the vessel: Contributes to the changes of the wetted surface of the hull.
State of the vessel: Determines whether the vessel is in trawling or steaming state and contains information about the trawl type being used.
Wind: Both the direction and velocity of the wind are assumed to aect the resistance of the vessel.
Ocean waves: The signi cant wave height, its period and direction. Provided that representative measurements or estimations of those variables is available the model will rely on neural networks to nd the relationship between the inputs and the outputs.
Sec. 5.3.
5.2
33
Current implementation
Current implementation
In the discussion of current implementation the model is compared to the implementation of the model for hull and propeller from the EDT solver (hereafter referred to as the EDT model). The comparison is mainly made on accuracy level while other characteristics of the implementations such as performance will be left out of the study. The EDT model uses the Holtrop method discussed in section 2.1.1 for the still water resistance calculation of the hull with the addition of wind calculation. The thrust and other factors of the propeller are calculated as discussed in section 2.4. Since the model is made from known methods, loosely based on rst principles, it is not black box as the proposed model and number of intermediate results can therefore be read from its results. The EDT model is based on separate models for hull and propeller and its solution therefore has to be based on an iterative process due to the fact that the hull's resistance is dependant on the ship's velocity which is based on the thrust given by the propeller. The EDT collects all available equations and solves the problem using Newton-Raphson method.
5.3
Model quality metrics
The criteria used to estimate the quality of the model is based on two error estimates for each of the two output variables. The primer is the standard deviation given by s PN Pc n 2 n=1 k=1 (yk − tk ) (5.1) σ= Nc where yk is the result from the model while tnk is the target output. Both are given for patterns from n=1 to N where the number of variables is taken from k =1 to c. The latter estimate is the mean absolute error de ned as N
EM AE
c
1 XX = |yk − tnk | . N c n=1 k=1
(5.2)
The overall quality is then graded from the standard deviation and the absolute mean for both the shaft power and the log velocity according to
grade = 100(σvlog + EM AE,vlog ) + (σPS + EM AE,PS ),
(5.3)
which was derived heuristically where higher emphasis is put on the shaft power. This was done since experiments showed that the factor between the error estimates for the two variables was little over 100, which was therefore reduced to 100 to put a
34
Model Implementation
Ch. 5
little more emphasis on the shaft power. These quality estimates were used for all types of testing, including dierent models (section 5.4 to section 5.6) while other factors such as quality estimations regarding ease of maintenance and complexity were only estimated from rationality.
5.4
Construction of the proposed model
The construction of the model can be considered to be twofold; the decision regarding size and topology of the network and the training of the network with respect to selection and preprocessing of input parameters. A schematic gure of the proposed model is shown in gure 5.1 where the model includes pre- and postprocessing along with the neural network. Unless otherwise noted the methods used are the ones listed in section 3.2. x(t) Proposed Model x(t-k)
y(t)
Figure 5.1: Structure of the proposed neural network model where single model is used to map the transformation between all input variables to the desired outputs; shaft power and log velocity.
5.4.1
Base model
A benchmark for further development of the model was obtained using base model as shown in gure 5.2. Only the presumably most important input variables were used, either raw from the Maren database or as estimated without any changes. To connect
Wind strength Wind angle Propeller pitch Propeller RPM Wave height Wave angle Loading Operation state Ship’s heading
Preprocessing
Neural network
Postprocessing
- Incomplete sets removed - Standardization - Angles made relative to the ship. - Angles transformed to [1;-1]
- Single hidden layer - Seven hidden neurons - tanh activation func. - Linear transformation for output - Available train tours - 800 train points - 70% train / 30% valid. - SSE performance func.
- Standardization
Log velocity Shaft power
Figure 5.2: The properties of the base model used as a benchmark for the testing of variations made to the model.
Sec. 5.4.
35
Construction of the proposed model
the wave and wind to its direction all angles were made relative to the heading of the ship. The structure of the neural network was obtained from crude experiments and based on what was expected to give good results according to literature. The results for the base model are used as a benchmark for further studies and can be seen in table 5.1. These results are from a model created using all available train tours without any preprocessing, except for the removal of incomplete sets due to logging errors or frozen meters. Shaft power Base training (benchmark)
Log velocity
EM AE
σ
EM AE
σ
127
157
0.63
0.93
Grade 441
Table 5.1: Benchmark results from the base model. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
Note about the training of the models During each test cycle the models were trained 100 times and the best model was selected from these training sessions. Attention was paid to the distribution of the grades where the optimum according to the test tours didn't turn out to be the most frequently calculated result. An example of this, for three dierent train sessions of the base model, can be seen in gure 5.3. The reason why the three gures do not 15
10 8
10 10
6 4
5
5
2 0 450
500
550 Grade
600
0 450
500 550 Grade
600
0
450
500 550 Grade
600
Figure 5.3: Distribution of three dierent train sessions for the base model where the x-axis gives the grade.
give exactly the same results lies in the random seeding during the initialization of the network being trained (section 3.2.4). Beside that the distribution of the grade indicates rather at landscape being modeled.
36
Model Implementation
5.4.2
Ch. 5
Training data
The high amount of data available for this study will not always be the case since the proposed method is to be used for creating model which should be available as soon as possible after measurements have started. Dierent model training is also required for every new type of shing gear and some estimations regarding the acceptable number of training tours should therefore be made. As mentioned earlier the available data can in general be split into to two operating states where the trawling state can again be split into dierent states depending on the shing gear being used. Two dierent trawl types were investigated as a part of this study, the capelin trawl and the blue whiting
trawl. The division of available data is shown in gure 5.4 were part of the data was reserved as a test set which was in no way used during the training phase of the model, while the rest was used using dierent number of tours and dierent selection of the data for training and validation. The reason why the last tours of each trawl type period were used for testing lies in the future usage of the model where there are some time passing eects that may in uence the model and its training should be generalized for future behavior. Before the train data was subjected to the training, it was reduced by taking samples with equal intervals from the tours. The original data had resolution of 15s, but after being reduced the interval between the points was often more than one hour. This was done to lower the training time of the model. For each test, the train tours used are randomly selected and ordered, using each tour only once for every training.
Blue whiting
1
2
Chapelin
3
5
6
16
17
2
3
5
6
8 9
10
11
9
10
11
12
13
14
15
Time
Training set (including validation)
1
8
Testing set
12
13
16
17
14
15
Figure 5.4: Division of the tours into training (including validation) and testing sets for the two available trawl types.
Sec. 5.4.
37
Construction of the proposed model
Results for dierent combination of training data Result for dierent combination of train data, dierent number of train tours used and dierent combination of train/validation data are shown in table 5.2. Shaft power
Log velocity
Grade
EM AE
σ
EM AE
σ
1. Tnr. 11
228
268
1.00
1.40
736
5. Tnr. 10 1 11 8 13
114
156
0.59
0.87
416
9. Tnr. 8 2 12 3 11 10 1 5 13
112
149
0.57
0.82
401
106
137
0.61
0.85
389
= 5; n = 1000. Ratio(tr/val) 80/20 119
140
0.64
0.87
409
M (1) = 7; n = 1500. Ratio(tr/val) 70/30 125
154
0.56
0.83
418
Dierent number of tours and combination
Dierent number of hidden neurons, train points and ratio btw. train and validation points
M (1) = 3; n = 500. Ratio(tr/val) 90/10 M
(1)
Table 5.2: Results for dierent numbers of hidden neurons M , training points n and the ratio between training and validation points. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
From gure 5.5, representing the results for dierent number of train tours, it can be noted that increased number of train tours1 improves the accuracy of the model to some extend. For the given data the limitation for clear improvements is reached after about ve tours although minor improvements can be reached using more tours. The usage of more than ve tours does on the other hand lower the mean grade from the 100
Grade, logarithmic base 10
train sessions conducted. Due to the result dierence between training of the models Mean Minimum Most common 3
10
1
2
3
4 5 6 Number of train tours used
7
8
9
10
Figure 5.5: The gure shows the general behaviour of the minimum, mean and most frequent values of the grade with increased number of train tours. 1 Increased
number of train tours does not mean increased number of train points.
38
Model Implementation
Ch. 5
(section 5.4.1), the train sessions were also evaluated by plotting the most frequently calculated optimum and the mean grade for all the 100 training sessions. In general this did not change the main results since all three metrics showed similar behaviour. Various combinations of ratio between train and validation data were also tested where the results indicate that the ratio of training data should be close to 90% using only 10% of the data for validation. The accuracy of the model then decreases as the portion of the training data is lowered and due to that, care must be taken for the possibility that the results can be tailor made for the test set and do not represent the circumstances in general. The number of train data does not seem to aect the results as long as it is kept over 500 points.
5.4.3
Feature selection and extraction
The feature selection and extraction performed consists mainly of two things; the selection of the input variables and their preprocessing in terms of adding prior knowledge to the interpretation of the variables. In addition to that preprocessing using PCA is tested.
Wind The measured wind information consists of two variables. One representing the wind angle on the range αwind ∈ [0; 360[ and the other for the wind velocity measured in m/s. The rst step in the preprocessing of the wind angle was to change it from being described as the whole circle to being described on the interval from 1 to −1 representing the wind to the aft and front of the ship respectively. It is therefore assumed that the aect of the wind is symmetrical around the heading of the ship, which raises considerations regarding the direction of the ocean wave with respect to the vessel, which must be in the same half as the wind if no extra measures are to be taken. One possible method is to introduce an indicator which is set to 1 as long as the direction of the wind and wave are to the same side of the ship and 0 otherwise. For the calculation of the relative wind, the angle of the wind and heading of the ship are required along with the velocity of the ship. Under normal circumstances, during simulation, the velocity of the vessel is not available and must therefore be fed back from the simulation model as shown in gure 5.6. The angle of the wind relative to the vessel was tested both as an individual input variable, with no connection to the wind's velocity, and by the usage of compression functions where the strength and angle are combined into one input variable, (Li et al.,
Sec. 5.4.
39
Construction of the proposed model
Encounter frequency Relative wind velocity
Simulation
vt
v t-1
Figure 5.6: The velocity of the vessel is required to calculate the relative wind velocity. Same methods are used to calculate the encounter frequency for the ocean wave.
2001). Two types of compression functions were used, a simple one given as (5.4)
f (vw , αw ) = vw cos(αw )
and shown in gure 5.7(a) and an advanced one which was originally derived for the wave data and is explained in the following section. Similar to the expected aect from wind, the simple compression function assumes that head wind increases the resistance of the vessel while following wind reduces the resistance.
Affect on velocity
Affect on velocity
0 20 10 0 −10 −20 0
0 10 20 Velocity [knots]
100 Angle [°]
(a) Simple compression
−5
−10 0
0 5 Height [m]
100 Angle [°]
(b) Advanced compression
Figure 5.7: Compression functions applied to the wind and wave data in order to incorporate the angle information. (a) gives a simple compression according to (5.4) while (b) has more advanced compression heuristically created based on literature and given in (5.5).
Ocean wave Similar measures were taken with respect to the direction of the wave as were made in the wind section. Beside the two compression functions, tests were made using all wave variables separately. The advanced compression function was created in search for the behaviour discussed in section 2.3.2 and shown in gure 2.2 where the input variables were combined and weighted with the function shown in gure 5.7(b) and given by
f (H1/3 , µ) = H1/3 (−0.14H1/3 + 0.6 cos(µπ/180)).
(5.5)
40
Model Implementation
Ch. 5
That way the representation of the wave is reduced to two parameters. Since the calculation of the encounter frequency (2.7) requires the velocity of the vessel the measured velocity is used during the training of the network while the simulated velocity is fed back as an input to the model during simulation.
Note: The wind velocity and the ocean wave height are supposed to play similar roles for the compression functions above. Equations and gures should therefore be read with that in mind.
Trawl types During this study, two dierent types of trawls were tested, where the trawl for the blue whiting is known to cause greater resistance than the capelin trawl. The dierence of means for the required shaft power during trawling of blue whiting and the capelin trawl was also con rmed for two samples taken at similar weather and velocity conditions and shown in gure 5.8. One input variable was therefore used to represent the three
30
Blue whiting
Steaming 20
Capelin
10 0
500
1000
1500
2000 2500 Shaft power [kW]
3000
3500
Figure 5.8: Selected sets shows the shaft power for the vessel at velocities between 4 knots and 6 knots in wind below 20 knots. The three distributions represent the steaming state and the two types of trawls tested.
dierent states of the ship simulated. For the steaming state the variable was set to 2, while the two trawling states were represented by 3.1 and 3.2. The selection of the values is mainly related to conventions used during the programming where 3 was used to represent all trawling stages and 1 was used for the pumping stage. Another possibility was to use indicators set to 0 and 1 where two variables would be needed for the three states presented in this model.
Sec. 5.4.
Construction of the proposed model
41
Dierence series Since changes at certain time are often likely to in uence the system with some delay, dierence series were created for some of the input variables. These series represent the changes in that particular variable and are created as follows; ∆x(t) = x(t) − x(t − 1). The closest example of this might be the pitch of the propeller where it is rather likely that changes require extra power in addition to the power needed to keep the propeller at steady state. To avoid loosing information about the characteristics of the original series, the dierence series were always used along with the original series they were derived from. This is the closest this study gets to implementing dynamic behaviour into the system.
Other variables Other input variables such as the operational parameters (propeller pitch and rotation) and the estimated variables such as the loading of the vessel were not pre-processed any dierently than all the variables introduced to the network by standardization.
PCA selection Selections of input data were presented to PCA as implemented by the Matlab function
PREPCA. In the primer test, all available input data was used and then reduced by the PCA while only selected set, still bigger than in the benchmark, was used in the latter case.
Results for the feature selection and extraction In order to evaluate the aect of dierent approaches regarding the input data preprocessing the options presented in the sections above were tested while other aspects of the network remained xed. The main results are shown in table 5.3. The simple compression gave the best result for the wind preprocessing while the advance compression function returned the best results for the preprocessing of the wave data. The relatively good results for the tests where either the wave or wind data is omitted are interesting. The reason for this probably lies in the high correlation between these input variables and the fact that neural network often perform better on fewer input variables. In similar manner, the period of the wave did not improve the model, which may be related to its correlation to the wave height ( gure 6.10(c)). The dierent combinations of dierence series did not seem to aect the results in
42
Model Implementation
Shaft power
Log velocity
Ch. 5
Grade
EM AE
σ
EM AE
σ
Wind: omitted
125
142
0.63
0.90
419
Wind: simple compression
125
152
0.56
0.84
417
Wave: omitted
121
141
0.58
0.85
405
Wave: advanced compression
112
126
0.55
0.84
376
Wind
Ocean Wave
Table 5.3: The aect of dierent combinations of input data preprocessing. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
the same way as the wind and wave data. In similar manner, the inclusion of aging did not improve the model which is probably due to the short period included in the train set. Preprocessing using PCA is left out since it did not improve the results of the base model.
5.4.4
Network size
The size of the network mainly consists of two factors, the number of hidden layers and the number of hidden neurons in each layer. Due to the universal approximation capabilities of a MLP network with one hidden layer, emphasis was put on its usage although some instances of MLP networks with two hidden layers and networks without hidden layer (linear regression) were also tested. Since the optimal number of hidden neurons depends in a complex way on many things such as numbers of input and output variables, number of training cases, amount of noise in the targets, complexity of the function to be learned, architecture of the network, type of hidden neuron activation function, training algorithm and the regularization method there are no simple general rules that can be applied to decide how many hidden neurons should be used based on the number of inputs and outputs, (Sarle, 2004).
Sec. 5.4.
43
Construction of the proposed model
Results based on network size The main results from dierent executions of the proposed model using dierent number of hidden neurons and layers are given in table 5.4. Shaft power
Log velocity
Grade
EM AE
σ
EM AE
σ
272
337
0.92
1.07
808
261
235
2.35
3.26
1157
Without hidden layer (linear regression)
M (1) = N aN ; M (2) = N aN Single hidden layer
M (1) = 1 M
(1)
=2
133
169
0.65
0.89
456
M
(1)
=3
124
153
0.57
0.85
420
Table 5.4: Main results for dierent sizes of the network including networks with none and one hidden layer. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
Based on the results in table 5.4, linear regression is not good enough for the description of the two variables being modeled and a network with single hidden layer gives even worse results. The number of neurons in the hidden layer need not to be more than 3 and the range from 3 to 7 provides the best results. Using more neurons or networks with two hidden layers, including over 5 neurons in each layer, seem to create to complicated model where the optimum is harder to nd during training.
5.4.5 Training method Two types of stop criteria, Early Stopping and Bayesian Regularization were tested on the base model setup for dierent optimisation algorithms, Levenberg-Marquardt and Gradient Descent. The combination of the Levenberg-Marquardt optimisation algorithm and Early Stopping returned the best results using fewest numbers of epochs although the Bayesian Regularization also gave good results.
5.4.6 Reduced number of output variables A part of this study is a comparison to the hybrid model presented in section 5.6 where only one output variable is being calculated. Some tests were therefore performed using only one output variable at a time, rst the shaft power and then the log velocity. The
44
Model Implementation
Ch. 5
results from those tests are similar to the tests where two output variables are simulated at the same time and are therefore not presented in more detail here.
5.4.7
Results for the proposed model
Based on the study performed and described in the preceding sections a new model was created as speci ed in the following list. Its performance is given in table 5.5.
Input variables: RPM and pitch of the propeller were used along with the loading of the vessel and its state. Wind and wave data were used and processed as given below.
Feature extraction: Simple compression was used for the wind data while the advanced compression returned the best results for the wave data.
Training data: The ratio between training and validation data was set to 90/10 although it raised some questions and considerations addressed in section 5.4.2. Number of training points was set to 500.
Network structure: A network with single hidden layer including as low as three hidden neurons seems to provide good results for the given input data. The proposed model uses three hidden neurons in a single hidden layer.
Training method: The Levenberg-Marquardt optimization with Early Stopping returned the best results. Shaft power Proposed Model
Log velocity
EM AE
σ
EM AE
σ
99
125
0.59
0.79
Grade 362
Table 5.5: Results for the proposed model. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
5.5
Composite models
Although implementation of a single model is by far the most convenient method for the training, execution and maintenance of the model, other investigations were also performed. Construction of several models based in input space decomposition, where each of the models were created for speci c range of speci c input variable, was also tested. The main objective of those experiments was to get more accurate and even simpler models. This is shown schematically in gure 5.9.
Sec. 5.6.
45
Composite models
x(t) x(t-k)
Network selection based on input variables
Network #1
Network #2 y(t)
. . .
Figure 5.9: Structure of the model based on input space decomposition where new network model is created for some known subset of the input space.
5.5.1 Input space decomposition The rst input space decomposition to be implemented was made with respect to the state of the trawler where one model was created from the steaming data and two models from the trawling data, depending on the trawl type being used. The benchmark setup, described in section 5.4, was used resulting in the gures given in table 5.6. The table also shows results from experiments using decomposition based on rotational velocity.
5.5.2 Results for the composite models Beside the fact that composite models are more complex with respect to usage and maintenance, this method did not outperform the method used for the proposed model shown in section 5.4. The main results for composite models are given in table 5.6. Shaft Power
Log Velocity
Ratio
EM AE
σ
EM AE
σ
Steaming
144
140
0.54
0.74
412
0.38
Trawling
129
160
0.88
1.22
498
0.62
Grade
Grade
Grade
465 RPM ≈130
125
163
0.69
0.91
447
0.05
RPM ≈150
132
168
0.65
0.96
461
0.95
460
Table 5.6: Results for the contribution of each sub-model along with the total quality of the combined model compared to the proposed model. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
46
Model Implementation
5.6
Ch. 5
Hybrid modeling
A hybrid model was also studied where the objective was to model the error between the EDT model and the measured outputs. The idea was to identify the variables that did not contribute to the EDT model as they should and those that are not taken into account in the EDT model such as the loading of the vessel and the aect of ocean wave. This new model is called supplementary model and shown in gure 5.10. xprior
Supplementary network model
xadditional
ysupplement y
xprior
Prior model (EDT)
yprior
Figure 5.10: Structure of the hybrid model where the prior model is the EDT model while the supplementary model is supposed to model the error of the prior model using neural network.
The training data for the hybrid model consist of the dierence between the measured variables and the results calculated from the EDT model as given by (desired) yˆ = y − yˆprior . supplement
(5.6)
It is therefore highly dependant on any changes to the EDT model which makes it a bad choice with respect to maintenance. On the contrary from the complete model presented in section 5.4, some of the input parameters that are essential to the general modeling of the shaft power and velocity of a vessel may not be required by the supplementary model if they are already counted for in the prior model. Another diverse from the proposed neural network model is the number of outputs which in this case is reduced to one variable, namely the required shaft power. This distinct is made due to the fact that the EDT model uses the velocity of the vessel as an input parameter and calculates the required pitch to obtain the desired velocity. Because of this the velocity replaces the pitch in the possible input set.
5.6.1
Feature selection
Beside the pitch being replaced by the velocity of the ship, the same subset as used for the base model was used for basic testing of the hybrid model. Tests were made where the selection of input variables was based on prior knowledge about the EDT
Sec. 5.7.
47
Hybrid modeling
model and visual comparisons of its results to measured values where it is known to lack two presumably important input variables, loading and ocean waves. In that case the velocity and rotation of the propeller were excluded from the test.
5.6.2 Objective function The objective function for the hybrid structure was made a little dierent from the previously shown SSE error with the introduction of a penalty. The main purpose of the penalty was to keep the modeled error between the value of the error generated by the EDT model and zero. The results for this modi ed performance function were then compared to the results given by the SSE.
5.6.3 Results for the hybrid model The results for the hybrid model do not look promising as can be seen in table 5.7 where the velocity factors from the benchmark test of the proposed model have been added to the grade for easier comparison.
Shaft power
Grade
EM AE
σ
Base training (benchmark, SSE)
275
503
926
Base training using PCA and pitch+rpm
220
289
657
Table 5.7: Results for the dierent versions of the hybrid model discussed in this section. Table shows the values for the dierence series between the model results and measured results. Shaft power is given in kW and velocity in knots.
It is quite clear that the dierence between the EDT model and the measured data cannot be corrected only with the addition of the factors originally missing from the EDT model. The usage of the velocity and the RPM increased the accuracy of the model although it is far from being comparative with the results from the proposed model presented in section 5.4. Beside that, the results from the usage of the custom made performance function did not turn out to give as good results as rst tests indicated.
48
5.7
Model Implementation
Ch. 5
Summary
Since the rather tedious modeling based on composition of many models (section 5.5) did not outperform the simple usage of one model for all operating states (section 5.4) and the hybrid model (section 5.6) did not show the expected results, these models will not be discussed any further. The results for the proposed model do on the other hand look promising and are further discussed and analyzed in chapter 6.
Chapter 6 Model Analysis and Applicability This chapter presents the results from the implementation of the proposed model and demonstrates the in uence of the selected input variables on its outputs, the shaft power and the log velocity. Examples of the model's applicability are also given.
6.1
Results for the test tours
Figures 6.1 and 6.2 show the results for the shaft power and log velocity respectively. These results are derived using the proposed model, including the four test tours as a function of time. It should be noted that the results are based on the same input parameters as used for the testing in chapter 5, where corrupt data due to frozen meters have been removed. The original test set contains almost 90,000 points but has been reduced to around 1000 for increased clarity for the reader. Another thing that should be noted is that the time is not given for continuous time interval since some operating states such as pumping, preparing and pay out are not simulated and the time series contain four tours occurring at dierent times of the year. Scattering results of gure 6.3 were produced to show how the neural network estimation output matches the desired patterns. The shaft power estimations seems to give good results over the whole range while the velocity is less accurate for intermediate and low velocities. The decreased accuracy of the velocity for the trawling states can also be seen in gure 6.2 where the variance in the model is not enough to follow the actual velocity. The reason for this is probably related to the behaviour of the trawl and the in uence of the trawl winches on the total resistance. The histogram shown in gure 6.4 gives the dierence between the actual output and the estimated output for the shaft power and the log velocity using the test tours. 49
50
Model Analysis and Applicability
Measured values Proposed model
3000
Shaft power [kW]
Ch. 6
2000 1000 100
200
300
400
500
600
700
800
900
Shaft power [kW]
500
1000
Error
0
−500 100
200
300
400
500 Samples
600
700
800
900
1000
Figure 6.1: Results for the proposed model as a function of time and comparison with the measured
Log velocity [knots]
values for the test data.
14 12 10 8 6 4 2
Measured values Proposed model
Log velocity [knots]
100
200
300
400
500
600
700
800
900
4
1000
Error
2 0 −2 −4 −6 100
200
300
400
500 Samples
600
700
800
900
1000
Figure 6.2: Results for the proposed model as a function of time and comparison with the measured values for the test data.
51
Results for the test tours
3000 2000 1000
Actual log velocity [knots]
Actual shaft power [kW]
Sec. 6.1.
500
1000
1500 2000 2500 Estimated shaft power [kW]
2
4
6 8 10 Estimated log velocity [knots]
3000
3500
14 12 10 8 6 4 2 12
14
Figure 6.3: Comparison between the proposed model and the measured values for the test data.
250
140 120
200
100 150 80 60
100
40 50 20 0
−600 −400 −200 0 200 Shaft power error [kW]
400
0
−6
−4 −2 0 2 Log velocity error [knots]
4
Figure 6.4: Histogram showing the dierence between the actual output and the estimated output for the shaft power and the log velocity.
52
Model Analysis and Applicability
Ch. 6
Autocorrelation1 function of the prediction error for the test data is shown in gure
Sample autocorrelation
Sample autocorrelation
6.5. The fact that the autocorrelation function for the velocity error lies within or
1 Auto correlation for the shaft power error 0.5 0 −0.5
0
5
10
15 Lag
20
25
30
1 Auto correlation for the log velocity error 0.5 0 −0.5
0
5
10
15 Lag
20
25
30
Figure 6.5: Auto correlation function for the dierence between the predicted and simulated values of the shaft power and log velocity. Function is calculated for the test tours.
very close to its con dence bounds indicate that the error does not contain repeated patterns that should be modeled with some missing input variable. The prediction error for the required shaft power does not give as promising results, which indicate that the model can still be improved with introduction of new input variables or with improved training. A method for estimating the error bars for the prediction interval of the model is given by Penny and Roberts (1998) but is left as a further work. The data range for the input and output parameters still provides an important knowledge about the possible operating ranges of the model and may be used as to give idea about the accuracy of the model on those regimes. The histograms for the wind and wave shown in gure 6.11 along with the histogram of the pitch given in gure 6.6 show the feasible range of input values for the model.
1 The
autocorrelation of a white noise signal is close to 0 for all time > 0.
Sec. 6.2.
53
Factors in uencing the ship powering
Frequency [%]
25 20 15 10 5 0
0
10
20
30
40 50 60 70 Propeller pitch [%]
80
90
100
Figure 6.6: Histogram for the data showing the pitch of the propeller. Gives an idea about the feasible operational range for the model.
Comparison with EDT implementation Although the implementation using the EDT was known to lack some important parameters such as the loading of the vessel (major) and the ocean wave (minor) the modeling of the dierence between it and the measured data did not prove to be successful (section 5.6). Using the grade introduced in section 5.3 the EDT model is graded at 913 using the test tours while the proposed model is graded at 362. The accuracy of the EDT model will not be discussed in more detail here and the reason for the dierence will not be investigated to more extend than described above.
6.2
Factors in uencing the ship powering
Since the models presented by this study are completely based on black-box methods it can be hard to determine the individual aect of each of the input variables due to internal con icts between them. The clearest example of this is probably the correlation between the ocean waves and wind where the waves are mainly generated by the wind as discussed in section 6.2.2. In order to evaluate which factors aect the ship the most, the operational and environmental factors are investigated.
6.2.1 Operational The operational factors are the ones controllable by the shipmasters, including the captain and the engineers onboard. The two main factors are the propeller's pitch and the rotation of the main engine although the loading of the vessel is also categorized with the operational parameters.
54
Model Analysis and Applicability
Ch. 6
Pitch and RPM of the propeller The pitch of the propeller and the rotational velocity of the main engine play the main role for the speed control of the vessel. Figure 6.7(a) shows the aect of the pitch and rpm on the ship's velocity along with the required shaft power. The data in the gure assumes still weather and no wave during the steaming state. In gure 6.7(b) the
3000 2500
10
2000
8
1500 6
1000
4 2
500 0
50 Propeller pitch [%]
Difference 150 RPM 3000 130 RPM 2500
500 400 300
2000 1500
200
1000
100 0
500
0 100
(a) Shaft Power v Pitch
2
4 6 8 10 12 Log Velocity [knots]
Shaft Power Difference [kW]
12
3500 Shaft Power [kW]
Log velocity [knots]
14
Shaft Power [kW]
Velocity, 150 Velocity, 130 Shaft power, 150 Shaft power, 130
14
(b) Shaft Power v Velocity
Figure 6.7: Figure (a) shows the log velocity and shaft power for the two available rotational velocities where the thick lines represent 150 RPM and the dotted lines stand for the log velocity. Figure (b) shows the shaft power dierence between the two rotational velocities.
dierence in required shaft power between running the engine at higher speed and lower speed is shown. In order to evaluate the e ciency with respect to fuel consumption the load of the engine must also be taken into account (see section 6.3.1).
Fish loading The aect of the vessel's loading turned out to be lower than expected where fully loaded ship, steaming at 14 knots only requires 150kW more than a light ship at same conditions. This behaviour can be seen on gure 6.8. Onboard the vessel being modeled, it is a custom to load the sh into the tanks in speci c order. The loading of the ship, as presented by this thesis, does therefore rely on the same loading order as traditionally used, although dierent loading should not aect the performance curve drastically. In order to evaluate dierent load cases the status of the sh tanks must be measured and used as an input for the model.
Shaft Power [kW]
2500
250 200
2000 1500
150
1000
100
500
50 2
4 6 8 10 12 Log Velocity [knots]
300
3000 Shaft Power [kW]
Light ship Fully loaded Difference
Shaft Power Difference [kW]
300
3000
0
55
Factors in uencing the ship powering
0 14
Light ship Fully loaded Difference
2500
250 200
2000 1500
150
1000
100
500
50
0
2
(a) 130 RPM
4 6 8 10 12 Log Velocity [knots]
Shaft Power Difference [kW]
Sec. 6.2.
0 14
(b) 150 RPM
Figure 6.8: The shaft power vs. velocity for fully loaded ship and light ship along with its dierence. Figures (a) and (b) show the values for 130 RPM and 150 RPM respectively.
Dierent trawl types As mentioned earlier two dierent trawl types were included as a part of this study, the capelin trawl and the blue whiting trawl. The two trawl types are usually used at similar velocities and do therefore require dierent amount of shaft power. This is mainly due to their size dierence and the depth they are used on. Figure 6.9 shows the shaft power requirements for the two trawl types for the commonly used trawling velocities from 3 knots to 6 knots. 3500
4000 150 RPM 130 RPM
2500 2000 1500 1000 500
150 RPM 130 RPM
3500 Shaft Power [kW]
Shaft Power [kW]
3000
3000 2500 2000 1500 1000 500
2
3
4 5 6 Log Velocity [knots] (a) Capelin trawl
7
0
2
2.5 3 3.5 4 Log Velocity [knots]
4.5
(b) Blue whiting trawl
Figure 6.9: Shaft power as a function of log velocity for the two trawl types tested. Figure (a) stands for the capelin trawl while (b) shows the values for the blue whiting trawl.
The shaft power requirements during the trawling, using the capelin trawl raise
56
Model Analysis and Applicability
Ch. 6
some considerations regarding the required rotational velocity. Since the required pitch is only around 60% the lower rotational velocity could be used which should therefore lie closer to the engines e ciency optimum as discussed in section 4.1. It should also be noted that the shape of the two curves, especially the one for the capelin trawl, are not as expected and not consistent with the shape of the steaming data. The reason for this probably lies in limited range of the available training data.
6.2.2
Environmental
Environmental factors are the ones related to the weather and can in no way be changed by the vessel's crew. In some cases the aect of the weather can be reduced by changing the operational factors discussed in section 6.2.1 along with the heading of the vessel. Figure 6.10 shows the relation between the wave and the wind data where it can be seen how the ocean wave is related to the wind. The direction of the wave and
180
0
180 Wave direction [°]
(a) Direction v direction
6 5 4 3 2 1
7
ITTC JONSWAP Wave height [m]
Wave height [m]
Wind direction [°]
7
10 20 30 40 Wind velocity [knots]
(b) Wind velocity v wave height
6 5 4 3 2 1 50
100 150 200 250 Wave length [m]
(c) Wave length v wave height
Figure 6.10: Relation between the wind and the wave data. (a) shows the connection between its directions while (b) gives the relation between the height of the wave against the velocity of the wind at given times and (c) shows the relationship between the wave height and length.
wind, shown in gure 6.10(a), show some tendencies toward relation although there are also a lot of points that indicate other. The connection between the ocean wave height and the wind velocity is shown in gure 6.10(b). It shows that higher ocean waves are usually related to higher wind velocity although the data used by this study does not t closely to the information given from the Joint North Sea Wave Project (JONSWAP) mentioned in section 4.3.3. The solid line on the gure shows the relation between the wind velocity and wave height according to 2/3
Vw = 10H1/3
(6.1)
Sec. 6.2.
57
Factors in uencing the ship powering
in which Vw is the velocity of the wind and H1/3 is the signi cant wave height. This relationship is given by ITTC2 recommendations for quick analysis. The relation between the ocean wave height and length is nally given in 6.10(c). As an assistance for the interpretation of the results given in the following sections and to give the reader some idea about the weather conditions used by this study, histograms for the ocean wave height and the wind velocity are given in gure 6.11.
Frequency [%]
Frequency [%]
15 10
5
0
20 40 60 Wind velocity [knots] (a) Wind
10 5 0
2
4 6 Wave heigth [m]
8
(b) Wave
Figure 6.11: Histogram for the wind and wave data used by this study. Frequency is given in percentage since the availability of the wind data was a lot greater than for the wave data.
In the following sections the in uence of wave and wind will rst be investigated individually as given by the proposed model and then combined where the reaction of the ship is studied for more realistic environmental conditions.
Ocean wave - height and direction The expected behaviour of the vessel under dierent wave height and direction was described in section 2.3.2. The only wave condition that led to reduction of the required shaft power or to increased log velocity was at following wave with rather low wave height. Figure 6.12 shows the results from the proposed model for steaming and trawling state at velocities of 11 knots and 4 knots respectively. The behaviour can be considered to be as expected where increased wave height cause the vessel to require increased shaft power. Head waves increase the resistance the most although most wave directions tend 2 The
International Towing Tank Conference is a voluntary association of worldwide organizations
that have responsibility for the prediction of hydrodynamic performance of ships and marine installations based on the results of physical and numerical modeling. http://ittc.sname.org/
58
Model Analysis and Applicability
Ch. 6
to increase the resistance. The exception to this is the following wave which lowers the
1750
Required Shaft Power [kW]
Required Shaft Power [kW]
total resistance for the velocities given in gure 6.12.
Following Wave Beam Wave Head Wave
1700 1650 1600 1550 1500
2
3 4 5 6 Significant Wave Height [m] (a) Steaming at 11 knots
7
2750 Following Wave Beam Wave Head Wave
2700 2650 2600 2550 2500 2450
2
3 4 5 6 Significant Wave Height [m]
7
(b) Trawling at 4 knots
Figure 6.12: Increased requirements for shaft power assuming xed velocity is shown in the gure where (a) stands for ship steaming at 11 knots and (b) shows trawling state using capelin trawl at 4 knots.
Although the period of the ocean wave is known to aect the resistance of the ship its application to the model did not give any improvements. This may be related to the high correlation between the period of the wave and its height as shown in gure 6.10(c). The length of the wave is given by λ ≈ 1.56T 2 , where λ is the wave length given in meters and T is the wave period given in seconds, (Journée and Massie, 2001).
Wind - velocity and direction The analysis of the wind eect is conducted in a similar manner as for the wave, where increased wind velocity is investigated for three dierent directions. The results for the wind, shown in gure 6.13, indicate linear relation between the wind and changed shaft power requirements. Opposite to the aect of the ocean wave, following wind reduces the resistance while the aect of strong head wind is similar to the aect of high ocean wave. Beam wind did not seem to aect the resistance of the vessel at all.
Wind and wave combined Due to the correlation between the wind and wave, it might have become hard for the neural network to distinguish between the in uence of those two variables to the output variables. It is therefore important to investigate the aect of wind and wave for circumstances that are likely to surface instead of investigating them separately as was done in the preceding sections.
59
Factors in uencing the ship powering
2000
Required Shaft Power [kW]
Required Shaft Power [kW]
Sec. 6.2.
Following Wind Beam Wind Head Wind
1800 1600 1400 1200
10
20 30 Wind velocity [knots]
40
2700 Following Wind Beam Wind Head Wind
2600
2500
2400
(a) Steaming
10
20 30 Wind velocity [knots]
40
(b) Trawling
Figure 6.13: Increased requirements for shaft power assuming xed velocity is shown in the gure where (a) stands for ship steaming at 11 knots and (b) shows trawling state at 4 knots.
Based on gure 6.10 the angles for the wind and wave are always kept equal in the following analysis while the relation between the wind strength and the wave height are assumed to be according to (6.1). It should be noted that these assumptions are only used for the visual analysis of the in uence of waves and wind and could be selected arbitrarily based on any likely conditions. Another way to select this would be to used the results from the JONSWAP project or by deriving relationship from the given data.
Shaft Power [kW]
2000 1800 1600 1400
150 100 50 Direction [°]
1,10
2,16
3,21
4,25
5,29
6,33
7,36
Height [m], Velocity [knots]
Figure 6.14: Shaft power requirements for steaming light ship at 11 knots depending on wind and wave. Directions of wind and wave are kept equal while the wind velocity is given as function of wave height according to (6.1).
As can be seen from gures 6.14 and 6.15 the form of the results for the aect of
60
Model Analysis and Applicability
Ch. 6
Ship’s velocity [knots]
12 11.5 11 10.5 10
50 100 150
7,36
6,33
5,29
4,25
3,21
2,16
1,10
Height [m], Velocity [knots] Direction [°]
Figure 6.15: Velocity pro le for a light ship depending on wind and wave with shaft power kept at 1800 kW. Directions of wind and wave are kept equal while the wind velocity is given as function of wave height according to (6.1).
wind and ocean wave on the velocity and the shaft power are similar to the preprocessing of those variables and just as expected. Since the models using those forms for preprocessing gave the best results it is rational to conclude that those are the forms best suited for the description of these environmental phenomena. Despite that, care must be taken in splitting up the aect of the wind and wave and try to interpret them individually due to its correlation mentioned earlier.
6.3
Model applicability
The following sections demonstrate some of the applicabilities provided by the model. A part of the discussion is repeated from last section with emphasis on energy e ciency and focus on the fuel consumption. For demonstration of the applicability, the fuel consumption per sailed mile is used as a performance index; l/nm (liter per nautical mile). The fuel consumption is calculated from the speci c fuel consumption (SFC) values introduced in section 4.1 which was evaluated using the EDT model of the ships's main engine. In average, the propulsion system uses 85% of the power produced by the main engine while the shaft generator uses the other 15% to produce electrical load required onboard. During the
Sec. 6.3.
61
Model applicability
demonstration, it is assumed that the shaft generator requires 400kW at all times, which is close to its average usage. This is assumed for both the steaming and trawling state of the vessel and are based on the data collected by the Maren system. Beside giving the shipmasters a good insight into the behaviour and the performance of the ship, the model can also be used as a part of operational decision system and routing system.
6.3.1 Performance based on velocity The velocity is the single most important factor on fuel consumption of a ship. The fuel consumption as a function of the vessel's velocity is shown in gure 6.16 for various conditions. For each condition there is an optimal velocity based on fuel consumption 90 130 rpm in calm sea 130 rpm in rough sea 150 rpm in calm sea 150 rpm in rough sea
Fuel consumption [l/nm]
80 70 60 50 40 30
2
4
6
8 10 Velocity [knots]
12
14
Figure 6.16: Fuel consumption in liters per nautical mile as a function of the vessel's velocity. The optimal velocity for each condition is marked with an 'x' on the gure.
which usually lies on the interval from 8 knots to 11 knots for the steaming state of the vessel. The gure also shows that for constant environmental conditions, constant speed will give better performance than steaming partly at low speed and partly at high speed. For the trawling state there is usually little allowance for selection of velocity and heading since the main objective is to follow the sh and not to maximise the distance traveled for every liter of fuel. The performance index introduced above may therefore not apply for the trawling state in the same manner as it does for the steaming state. A possible performance index for the trawling state might on the other hand be liter per tonne of sh which is hard to control since catch varies drastically between shing
62
Model Analysis and Applicability
Ch. 6
tours, shing areas and within each trawling period. The optimal trawling velocity with respect to fuel consumption for the blue whiting trawl is shown in gure 6.17. 210 150 rpm in calm sea 150 rpm in rough sea
Fuel consumption [l/nm]
205 200 195 190 185 180 175 170
2
2.5
3 3.5 Velocity [knots]
4
4.5
Figure 6.17: Fuel consumption in liters per nautical mile as a function of the vessel's velocity with blue whiting trawl. The optimal velocity for each condition is marked with an 'x' on the gure.
The optimums discussed above only describe the optimums with respect to fuel consumption, but would require calculations related to maximising the catch and additions of constraints related to time if they were to be used to maximize the total margin contribution. The ship may for example have to be at speci c harbour for landing at speci c time.
6.3.2
Performance based on heading
The performance changes based on heading, are solely due to changes of wave and wind directions relative to the ship. These changes are relatively low compared to the velocity changes but may in some cases, when applicable, lower the fuel consumption. Figure 6.18 shows the relation for the heading of the vessel relative to the wind and wave direction where 0◦ is a following wave. From the gure it can be seen how the direction of the ship becomes less important with respect to the fuel consumption as the vessel is loaded with more sh.
6.4
Summary
The factor that in uences the required shaft power the most is the desired velocity of the vessel which is obtained by increased rotational velocity and pitch of the propeller.
Sec. 6.4.
63
Summary
Fuel consumption [l/nm]
150
Light ship Half full Fully loaded
149 148 147 146 145 144 −150
−100
−50
0 50 Heading [°]
100
150
Figure 6.18: Fuel consumption in liters per nautical mile as a function of the vessel's heading. The velocity is kept constant at 4.5 knots at trawling state using the blue whiting trawl.
Other important factors are the catch loaded onboard the vessel, which reduces the velocity at xed shaft power and the external factors related to the weather. The eect of wind turned out to be highly related to its direction where head wind reduces the velocity at xed shaft power, beam wind shows no aect and following wind reduces the total resistance and therefore increases the velocity of the vessel at xed shaft power. The relation between the wind velocity and the shaft power was close to being linear for xed velocity. The relation between the wave height and the shaft power, to maintain required velocity, was close to being of second order where the direction of the wave had dierent aect than the direction of the wind. In general the wave increased the shaft power requirements at xed velocity, except when the direction became closer to being a following wave. For those conditions the wave slightly decreases the shaft power for medium size waves and then increases again for bigger waves. Finally the trawl being used changes the shaft power usage drastically where the blue whiting trawl requires more power since it is bigger and used at more depth than the capelin trawl. Other input variables such as the turning of the vessel, the ocean wave period and various dierence series where also tested but did not improve the model. Beside that there were no signs of increased resistance of the hull nor the trawls related to aging of those components.
64
Model Analysis and Applicability
Ch. 6
Chapter 7 Summary and Conclusions Formulation of a simulation model as the one proposed by this study presenting the shaft power and log velocity of a vessel has proven to be a challenging task. The assignment of creating a simulation program expected to be used both as the underlying model for onboard decision system as well as a tool for analysis and routing conducted onshore, can only use limited set of input information. Information that must be able to present the nonlinearity of the hull's resistance and the propeller's behaviour. Three dierent ways of modeling the shaft power and velocity of a pelagic trawler, all based on MLP network using empirical data, were studied. The data describing the operation of the vessel was mainly taken from the Maren system onboard the vessel or derived from the Maren data. This excludes the wave information, which was given by the Icelandic Maritime Administration. A complete model for the shaft power and the velocity of the vessel was made from scratch using the MLP network. This model soon became the proposed model and the focus was set on nding the best combination of input variable selection, its preprocessing and the topology of the network. Instead of using a single model as the one proposed above, it was also studied how the usage of composite models, each for a special condition, would perform compared to the single model. Two types of input decomposition were tested, one based on the operating state of the vessel and the other based on the rotational velocity of the propeller. These models only gave similar results as the usage of single model and did not outperform them as expected. It was therefore decided to put more emphasis on the single model due to it simplicity. The last model type tested was the hybrid model, where the focus was set on modeling the dierence between the EDT model and the measurements. The idea was to compensate for the parameters that are known to be lacking in the EDT model. The 65
66
Summary and Conclusions
Ch. 7
hybrid model did not give satisfactory results although some improvements from the current model were made. The calculations used throughout this study, along with the implementation of the model, were conducted using Matlab. The resulting model can easily be implemented in a standalone module such as a dll while the integration into Maren remains an open issue.
7.1
Conclusions
The experiments made on the model proposed by this thesis gave good results compared to the classical approach given by the EDT model and should suit its simulation purpose well. Although the model was not able to follow high frequency changes in the shaft power requirement, it followed the major changes very well which is an important feature for the energy oriented simulation it is meant to serve. It is therefore concluded that the model is very well suited for onshore simulation where the main purpose is to evaluate dierent options regarding the vessels operation and also as a part of an oshore decision support system conducting total operational optimisation. One of the advantages of the model over the EDT model, is being created from empirical data, which gives a lot more accurate model. The method being used has on the other hand disadvantages related to the range of the data being used during training, which limits its operational range and may give unreliable results at intervals that are not represented by the data being used. This mainly applies to the upper and lower bounds of speci c variables while intervals in between data can in most cases be considered reliable.
7.2
Further work
The initial idea behind this thesis was to model the resistance of the vessel's hull and its trawls including the aect of dierent weather conditions. Since the resistance by itself is hard to measure the idea was extended to the next available measurement, namely the shaft power, which introduced the aect of the vessel's propeller. Due to its importance to the shipmasters, the velocity was also taken into the model. For future work, this idea could be extended further, for example with the simulation of the trawl wire tension which is also important with respect to the total energy management onboard the vessel. Similar extensions could be made in the direction of the main engine with the simulation of the fuel consumption or some of its other interesting factors.
Sec. 7.2.
Further work
67
The theory behind simulation models assume that any given state can be simulated without the presence of any feedback measurements. Dierent from the simulation models are the prediction models used for control, where feedback values are available to correct next prediction. In a similar manner as introduced in this study, it would be interesting to create a prediction model for the trawl wire tension which would provide very valuable information for the control of the trawl winches. This kind of prediction model for the shaft power was presented by Xirios and Kyratatos (2000). Growing number of ships with Maren installations open up the possibility to create a general power prediction model with the introduction of some hull characteristics. The initial steps may be to take few similar vessels which only dier slightly and then extend the model as the availability of data grows. The introductions of new and more detailed variables might also assist for the creation of better model. The assumptions regarding the loading of sh into the ship, do for example generate some uncertainty which could be eliminated with the introduction of new measurement devices. Similarly it is always assumed that the sh is pumped onboard by lling the sh tanks in the same order all the time, which might not be the case for all vessels. Inaccuracy is also related to the aect of the ocean wave since the data used only represents the signi cant wave height with resolution of only 0.5◦ and 6 hours. Future work might include the search for the availability of more accurate data. This might even by measured onboard the ship being modeled. The data available was not created speci cally for this study and may therefore include tendencies of the captain to always react in the same manner to speci c conditions. The data may therefore be contaminated with this behaviour and could in some cases not be t to represent arbitrary situations that may arise. The changed behaviour of the captain due to the Maren decision support might therefore change the accuracy of the model. To make sure that the model is performing well it should therefore be monitored and reevaluated after speci c number of tours, or when some criteria for the reevaluation of the model is met. This might also be needed since the aect of fouling was not incorporated into the model although it is known to cause increased resistance of both the hull and the trawls.
68
Summary and Conclusions
Appendix
Appendix A Data Information A.1
Overview of the tours
The available tours are mainly from October 2004 to the end of January 2005 (see table A.1). The resolution of the data is 15s. Steaming and trawling columns represent the number of complete sets for the given tours. In the thesis, the tours are referred to using number given in the Id column. Id
Started
Ended
Steaming
Trawling
Fish / trawl type
1
10 Oct
15 Oct
8,433
11,219
Blue whiting
2
20 Oct
25 Oct
4,583
7,684
Blue whiting
3
25 Oct
27 Oct
5,173
2,615
Blue whiting
5
10 Nov
17 Nov
14,472
8,324
Blue whiting
6
18 Nov
26 Nov
8,726
31,061
Blue whiting
16
27 Nov
05 Dec
13,373
21,399
Blue whiting
17
08 Dec
13 Dec
6,592
14,882
Blue whiting
8
03 Jan
07 Jan
14,664
2,161
Capelin
9
07 Jan
11 Jan
2,964
5,187
Capelin
10
11 Jan
13 Jan
5,204
4,183
Capelin
11
13 Jan
17 Jan
6,014
6,303
Capelin
12
17 Jan
19 Jan
3,549
3,702
Capelin
13
19 Jan
20 Jan
3,458
2,422
Capelin
14
21 Jan
24 Jan
6,063
7,429
Capelin
15
24 Jan
30 Jan
7,880
10,003
Capelin
Table A.1: Dates and number of data points available for the tours used for training and testing.
69
70
A.2
Data Information
Appendix A
State separation criteria
Complete set for the state separation consists of wind, velocity, pitch, rpm, shaft power,
trawl wire information and electric current to pumps. The state detection is made according to the criteria given in table A.2. Unit
Steeming
Shaft power
[kW]
> 200
Log velocity
[knots]
>2
Trawling
Pumping
<1
Trawl wire tension
[kN]
Trawl wire length
[m]
< 20
> 200
Current to pumps
[amper]
< 20
> 200
>6 > 120
Table A.2: Criteria for the state separation.
A.3
Coordinate conventions
Absolute directions are represented using α as shown in gure A.1(a) while the angles representing directions relative to the vessel are given by µ according to gure A.1(b) where following wind and waves is given by zero and λ stands for the ocean wave length. Cosine transformations of the angles relative to the ship are frequently used where the
14 ,8 5
N
138,5˚
E
Vlog
121,6˚
(a) Absolute coordinates
(b) Relative to the ship
Figure A.1: The coordinate conventions used throughout this study. (a) demonstrates the absolute direction of the ship while (b) shows the angle and direction of the wave, wind and current where λ is the wave length.
angle of the wave, wind and current becomes -1 for head, 0 for beam and 1 for following directions.
Bibliography The American Heritage Dictionary. The american heritage dictionary of the english language, 2000. http://www.dictionary.com. Jon Agust Thorsteinsson. Modelling of Fishing Vessel Operation for Energy System
Optimisation. PhD thesis, Aalborg University, Aalborg, Denmark, 2004. G. Stefansson. Design and implementation of an equation solver for an energy management system. Master's thesis, University of Iceland, 2004. EDT. Maren EDT; User Manual. Marorka ehf., Reykjavik, Iceland, rst edition, January 2005. http://www.marorka.com/. Maren. Maren EMT; Notendahandbók biðlara. Marorka ehf., Reykjavik, Iceland, rst edition, January 2005. http://www.marorka.com/. J.M.J. Journée, R.J. Rijke, and G.J.H. Verleg. Marine performance surveillance with a personal computer. Technical Report 0753-P, Delft University of Technology, Ship Hydromechanics Laboratory, Mekelweg 2, 2628 CD Delft, The Netherlands, May 1987. http://dutw189.wbmt.tudelft.nl/ johan. J. Holtrop and G.G.J. Mennen. An Approximate Power Prediction Method. Interna-
tional Shipbuilding Progress, 29(335):166170, July 1982. J. Holtrop. A Statistical Re-analysis of Resistance and Propulsion Data. International
Shipbuilding Progress, 31(363):272276, November 1984. David G.M. Watson. Practical Ship Design. Elsevier Science Ltd., Kidlington, Oxford OX5 1GB, UK, 1998. Kourosh Koushan. Automatic hull form optimisation towards lower resistance and wash using arti cial intelligence. Ischia, Italy, October 2003. Fuxiang Hu, Ko Matuda, and Tadashi Tokai. Eects of drag coe cient for dynamic similarity on model testing og trawl nets. Fisheries science, 67:8489, 2000. 71
72
Bibliography
R.S.T. Ferro. The calculation of the twine area of a trawl net. Scottish Fisheries
Information Pamphlet, (5), January 1981. R. M. Isherwood. Wind resistance of merchant ships. Trans. of the Royal Institution
of Naval Architects, 115:327338, 1972. http://www.rina.org.uk. C. P. Comstock. Principles of Naval Architecture. The Society of Naval Architects and Marine Engineers, New York, 1967. J. Gerritsma and W. Beukelman. Analysis of the resistance increase in waves of a fast cargo-ship. International Shipbuilding Progress, 18(217), 1972. Odd M. Faltinsen, K. J. Minsaas, and N. Liapis S.O. Skjördal. Prediction of resistance and propulsion of a ship in a seaway. In Proc. Thirteenth Symp. On Naval
Hydrodynamics, pages 503530, Tokyo, Japan, 1980. P. Boese.
Eine einfache methode zur berechnung der wiederstandserhöhung eines
schi¤es in seegang. (258), 1970. J.M.J. Journée and W.W. Massie. Oshore hydromechanics. Netherlands, January 2001. http://www.shipmotions.nl. Nathaniel Bowditch, National Imagery, and Mapping Agency Sta. The American
Practical Navigator: Bowditch. National Imagery and Mapping Agency, Bethesda, Maryland, 1995 edition, 1995. Robert H. Stewart. Introduction To Physical Oceanography. Department of Oceanography, Texas A and M University, Texas, USA, september 2004 edition, 2004. J S Carlton. Marine Propellers and Propulsion. Butterworth-Heinemann Ltd, Linacre House, Jordan Hill, Oxford OX2 8DP, UK, 1994. M.W.C. Oosterveld and P. van Oossanen. Further Computer-analyzed Data of the Wageningen B-screw Series. International Shipbuilding Progress, 22(252):251262, 1975. Yoram Reich and S.V. Barai. A methodology for building neural networks from empirical engineering data. Engineering Application of Arti cial Ingelligence, 13(6): 685694, December 2000. Oliver Nelles. Nonlinear System Identi cation: From Classical Approaches to Neural
Network and Fuzzy Models. Springer-Verlag, Berlin, Germany, 2001.
73 Jyh-Sing Roger Jang.
ANFIS: Adaptive-Network-Based Fuzzy Inference System.
IEEE Transactions on Systems, Man, and Cybernetics, 23:665684, 1993. citeseer.lcs.mit.edu/jang93an s.html. J.-S. R. Jang, C.-T. Sun, and E. Mizutani. Neuro-Fuzzy and Soft Computing: A
Computational Approach to Learning and Machine Intelligence. Pentice-Hall, Inc., Upper Saddle River, NJ 07458, USA, 1997. Christopher M. Bishop. Neural Networks for Pattern Recognition. Oxford University Press Inc., Oxford OX2 6DP, UK, reprint edition, 1999. ISBN 0198538642. Howard Demuth and Mark Beale. Neural Network Toolbox: For Use with Matlab. 3 Apple Hill Drive Natick, MA 01760-2098, USA, fth edition, January 1998. A. Pavelka and A. Prochazka. Algorithms for initialization of neural network, 2004. http://dsp.vscht.cz/konferencematlab/matlab04/pavelka.pdf. Russel D. Reed and Robert J. MarksII. Neural Smithing: Supervised Learnin in Feedfor-
ward Arti cial Neural Networks. The MIT Press, Cambridge, Massachusetts, USA, 1999. Wartsila. Wartsila Vasa 32, Project Guide for Marine Applications. Wartsila NSD Finland Oy, Marine, Vaasa, Finland, May 1997. NIST.
Nist/sematech e-handbook of statistical methods.
Online, April 2005.
http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm. Shuhui Li, Donald C. Wunsch, Engar A. O'Hair, and Michael G. Giesselmann. Using neural networks to estimate wind turbine power generation. IEEE Transactions on
Energy Conversion, 16(3):276282, Septermber 2001. Warren S. Sarle. comp.ai.neural-nets faq, part 3 of 7: Generalization. Online, June 2004. http://www.faqs.org/faqs/ai-faq/neural-nets/part3/index.html. William D. Penny and Stephen J. Roberts. Error bars for linear and nonlinear neural network regression models. Technical report, Imperial College of Science, Technology and Medicine, London SW7 2BT., U.K., November 1998. Nikolaos I. Xirios and Nikolaos P. Kyratatos. A neural predictor of propeller load demand for improved control of diesel ship propulsion. In Proceedings of the 15th
IEEE International Symposium of Intelligent Control (ISIC 2000), 2000.
