“Haste makes waste”

“Haste makes waste”

Empirical evidence linking shipping speed behaviour to the

energy efficiency gap in shipping

(Final version)

Student:

Bas de Boer (2810425)

Master Thesis:

MSc Technology & Operations Management

University of Groningen, Faculty of Economics and Business July 18th_{, 2016}

Supervisors:

Dr. ir. P. Buijs (first supervisor) Dr.ir. S. Fazi (co-assessor)

ABSTRACT

C

ONTENTS

Acknowledgements ...

1 Introduction ... 1

2 Background ... 3

2.1 Energy efficiency in shipping ... 3

2.2 Sailing speed ... 4 2.3 Time in ports ... 5 3 Methodology ... 6 3.1 Data collection ... 6 3.2 Pre-processing ... 6 3.2.1 Variable selection ... 6 3.2.2 Trips ... 7 3.2.3 Data normalization ... 8 3.3 Data analysis ... 9 3.3.1 Hierarchical clustering ... 9 3.3.2 Fuel consumption ... 11 3.3.3 Early arrivals ... 11 4 Findings... 13 4.1 Vessel type A ... 13 4.1.1 Descriptive statistics ... 13 4.1.2 Speed profiles ... 13 4.1.3 Fuel consumption ... 15 4.1.4 Early arrivals ... 15 4.2 Vessel type B ... 17 4.2.1 Descriptive statistics ... 17 4.2.2 Speed profiles ... 17 4.2.3 Fuel consumption ... 18 4.2.4 Early arrivals ... 18 4.3 Vessel type C... 20 4.3.1 Descriptive statistics ... 20 4.3.2 Speed profiles ... 20 4.3.3 Fuel consumption ... 21 4.3.4 Early arrivals ... 22

4.4 Overview vessel types A, B and C ... 24

6 Limitations and Future Research ... 28

7 Conclusion ... 29

References ... 30

Appendix A: Cluster analysis (vessel type A) ... 32

Appendix B: Cluster analysis (vessel type B) ... 34

A

CKNOWLEDGEMENTS

This thesis represents the last part in finalizing the master programme in Technology & Operations Management. Throughout this year, most of my time and effort was dedicated to finishing this Master. Despite my neglected social life, I can proudly look back at this successful year where my intellectual capabilities were challenged more than ever. Apart from my own motivation and hard work, I would not been able to achieve this alone. I would like to take this opportunity to express my graduate to everyone who supported me throughout this journey towards graduation.

First of all, I would like to thank Dr.ir. Paul Buijs for triggering my interest in the project on ship monitoring data at the master thesis fair. Next to the appreciation he deserves for his overall role as supervisor for his students in general, I would like to thank him for the personal and enthusiastic guidance throughout the project as well. He really supported me in the trajectory from the beginning to the end and provided me with very useful feedback. I would also like to thank Dr. Stefan Fazi in his role as co-assessor and the feedback he was willing to provide.

I would also like to thank René Ratering and Harald Rügebregt from RRISIS for their involvement in the project. They provided the data and offered me a place to sit at their office where I was always welcome work on my thesis or to discuss the research with them. Their field experience was very useful in deciding on certain assumptions and relating the findings to practice.

[1]

1 I

NTRODUCTION

Improving energy efficiency has become crucial nowadays for shipping companies due to increased fuel prices and environmental concerns. According to Christiansen et al. (2013), fuel cost can represent the majority of a vessel’s total operating expenses while the shipping industry as a whole constitutes for around 3% of global CO2 emissions and is expected to grow (Buhaug et al., 2009). Various technical

and operational energy efficiency improvements in shipping are identified in literature and available to be adopted with the current state of technology and knowledge (Buhaug et al., 2009; Eide et al., 2009; Eide et al., 2011; Faber et al., 2009). Even though most of the environmental improvements are cost-effective as well, they are not always realized in practice. This inconsistency between available efficiency improvements and actual implementation in practice is known as an “energy efficiency gap” (Johnson & Styhre, 2015; Johnson et al., 2014).

One of the most promising and most cost-effective operational energy efficiency improvements in shipping is speed reduction (i.e. slow steaming). An approximately proportional third-power relationship between sailing speed and a vessel’s daily fuel consumption is generally accepted in literature (Wang et al., 2013). Shipping speed can thus significantly impact operating cost and emissions, which resulted in an increased attention in recent research (Christiansen et al., 2013; Psaraftis & Kontovas, 2013; Psaraftis & Kontovas, 2014).

Eide et al. (2011) mention a great potential for lowering speed at sea by using unproductive time of ships in ports to reduce both costs and emissions. This specific potential is studied by Johnson and Styhre (2015), who looked at two ships in short sea shipping (mostly bulk shipping) in Europe. They studied voyage reports and found that half of the time in ports was unproductive due to factors such as early arrivals, port congestions, or ports that were closed for the night or the weekend. A common practice among shipper is that they sail fast at early stages of a trip because they face the risk of arriving too late (Poulsen & Sornn-Friese, 2015). It can be expected that this kind of behaviour more often lead to early arrivals while they consume more fuel than necessary.

The majority of research on shipping speed and energy efficiency often contain mathematical models that mainly focus on solving the model rather than formulating a model in a reasonable realistic way (Psaraftis & Kontovas, 2014). They probably lack extensive data regarding actual shipping practices in order to develop their models. Other studies on energy efficiency like Johnson et al. (2014) and Jafarzadeh and Utne (2014) only capture human perceptions by conducting interviews and surveys. These studies typically define barriers or drivers for possible improvements that still have to be explained (Johnson & Styhre, 2015). Empirical research on actual shipping practices seems to be missing in order to understand the interplay between sailing speed and port times in this part of the energy efficiency gap. This is consistent with Christiansen et al. (2013), who noticed that some of the recent research focusses more on theoretical contributions rather than real operations. They point to the need for future research on the uncertainties in sailing times and port times in maritime transportation. The extent to which this variability exist in these aspects may also complicate the improvements in energy efficiency.

[2]

operations is missing, appropriate improvements are hard to identify and implement, which could be a cause of the existence of the energy efficiency gap in shipping. In the search for a better understanding of actual practices and possibilities for improvement in this part of the energy efficiency gap, this thesis aims to answer the following main- and sub research questions:

“Is there a relationship between sailing speed behaviour at sea and unproductive time at ports due to early arrivals and how is it linked to the energy efficiency gap?”

 Can different sailing speed behaviours at sea be classified into typical speed profiles?

 What is the impact of different sailing speed behaviour at sea on a vessel’s fuel consumption?

 To what extent do early arrivals occur and how does it impact fuel consumption?

 Is there typical sailing speed behaviour that more often lead to early arrivals?

This research provides various theoretical contributions. Actual sailing speed behaviour at sea is addressed for the first time and focusses on actual shipping practices, which has been explicitly called for by several researchers (Johnson & Styhre, 2015; Christiansen et al., 2013). By using a different approach than Johnson and Styhr (2015), the unproductive time at ports are addressed by analysing five more ships in the same sector, enabling the comparison between results and generalization to a certain extent. Findings can increase knowledge on the presence of this part of the energy efficiency gap and can be used for improving the applicability and development of new and existing mathematical models (Psaraftis & Kontovas, 2013; Psaraftis & Kontovas, 2014). The variability in sailing speed and port times that are addressed reflect on the extent to which operational variability exist. This knowledge is increasingly important according to Fransoo & Lee (2013) and barely incorporated by researchers (Christiansen et al., 2013). The exploratory approach will demonstrate the possibilities of such shipping log data and stimulate ideas for future research as well.

From a practical point of view, findings can be used to create awareness among shipping companies about their energy efficiency in terms of sailing speed behaviour and unproductive time at ports. This is important for the whole shipping industry and in particular for bulk shipping activities in Europe since energy costs and regulations are on the rise in this sector (Johnson et al., 2014). The potential for reduced costs and emissions are quantified based on real operations that can encourage shipping companies and ports to implement energy efficiency measures and justify investments for them. Unique findings and insights are enabled by the availability of such shipping log data and stimulates the adoption of monitoring services.

This research is based on shipping log data that is typically measured real time and collected in a database. The data contains variables such as position (GPS), speed, trim, but also a lot of other variables related to the state of the engine or weather conditions. Besides the shipping log data, coordinates and information of the planned routes are also collected in separate data tables. The data is collected by monitoring systems that are installed on five different ships. The ships are active in short sea shipping/bulk shipping and sail mostly around Europe but frequently cross the ocean as well when freight opportunities come by.

[3]

2 B

ACKGROUND

The vessels investigated in this research generally fit within definitions of shipping activities in short sea shipping and bulk shipping. The vessels and types of cargoes are conform to the typical characteristics in short sea shipping mentioned by Paixão and Marlow (2002). Namely that vessels are often relatively small compared to typical ocean going vessels, vary from traditional bulk carriers to container vessels, and can carry various types of bulk cargo and project cargo.

Most transportation is between ports around Europe but ocean crossings frequently occur as well. This is contradicting to the general definitions often used for short sea shipping that imply that ships do not cross an ocean (Paixão and Marlow, 2002). According to the company involved, this definition rarely holds in practice as most vessels in this sector sail under charter contracts but are active on the spot-market as well, meaning that ocean crossings occur when possibilities for earnings come by.

2.1 E

Increasing the energy efficiency may be of particular importance at shipping companies in Europe’s short sea shipping due to competition from rail and truck transportation (Paixão & Marlow, 2002) and the introduction of more strict environmental policies and regulations like emission taxes (Johnson & Styhre, 2015).

Energy efficiency in transportation can be defined as the amount of the energy (i.e. fuel) used per transported good over a certain distance. Increasing energy efficiency thus involve the reduction of energy used (e.g. fuel consumption) while maintaining the same output (Croucher, 2011). Shipping is a rather energy intensive sector compared to other sectors as energy costs can rise up to 50% of the total costs for a shipping company (Jafarzadeh & Utne, 2014).

There are several ways to increase energy efficiency in shipping where a distinction is often made between technical and operational improvements (Buhaug et al., 2009; Eide et al., 2009; Eide et al., 2010; Faber et al., 2009). Technical improvements include a more efficient hull design or the use of low-carbon fuels and renewable energy. At the operational level, routing and scheduling can be improved but also trim optimization and speed reduction belong to the energy efficiency improvements at the operational level and are often cost-efficient as well. One of the most promising operational improvements is speed reduction at sea, especially when enabled by using reduced unproductive time otherwise spend in port.

Energy efficiency improvements are not always implemented in practice while they are often readily available and cost effective as well. This inconsistency is known as the “energy efficiency gap” and is often tried to be explained by various types of barriers. For example Johnson et al. (2014) and Jafarzadeh and Utne (2014) mention the inaccuracy of information, incompatibility between technologies and operations, not using information and split incentives among stakeholders. They also argue that the inability to accurately measure energy consumption or not utilizing the installed measurement equipment makes it hard to justify energy efficient investments.

[4]

2.2 S

Increasing fuel prices, depressed market conditions and environmental concerns have brought a new perspective to shipping speed in recent years (Psaraftis & Kontovas, 2014). Research in this area is rapidly growing as shown by several recent reviews on this specific topic (Psaraftis & Kontovas, 2013; Psaraftis & Kontovas, 2014). Among many factors, sailing speed is the main determinant of a ship’s fuel consumption (Meng et al., 2016), whereas fuel consumption on its turn represents a great share of operational expenses. A third-power relationship, or cubic law, between speed and fuel consumption is widely adopted in these models, meaning that the fuel consumption of a ship in one time unit is proportional to the sailing speed to the power of three (Meng et al., 2016). According to Psaraftis and Kontovas (2014), the cubic approximation is reasonable for bulk carriers, or ships of small size. Wang and Meng (2012) empirically confirmed that this is a good approximation for container ships under 20 knots but argue that when historical data is available, a more accurate function should be determined and used because the exact relation is vessel-specific. Although the aim of this research is not focussed on the precise fuel consumption – sailing speed relation, the literature does indicate that speed reduction can significantly reduce costs and mitigate emissions (Corbett et al., 2009). Reducing speed may be involuntary due to safety reasons in extreme weather conditions (Prpić-Oršić et al., 2014) but also voluntarily to reduce fuel consumption, which is known as slow steaming. Slow steaming has its technical and economic drawbacks. From a technical perspective, there is a limit to what extend a ship can be slowed down. Reducing speed below a certain threshold might not be efficient anymore but can still be achieved by deactivating one or more cylinders or install a so called ‘slow steaming kit”, provided by engine manufacturers (Psaraftis & Kontovas, 2014). From an economical perspective, Psaraftis and Kontovas (2010) argue that speed reduction can lead to revenue losses because of charter rates, inventory costs and increased transit times. Under these circumstances, it is necessary to increase fleet size in order to maintain a certain service level. In this research, the economic drawback is not taken into account as the focus will be on speed reduction by using unproductive time that was otherwise spend in port. This comes down to a more efficient use of the same time while maintaining the same output instead of taking more time for a certain output that leads to possible revenue losses.

Due to the non-linear relationship, sailing at a constant speed is more energy efficient than using varying speeds. It means that a ship retains a stable speed throughout the trip and avoid costly, high speeds at early or later stages of a trip. Using a fixed speed is a common assumption in literature when assessing shipping speed (Psaraftis & Kontovas, 2014). In practice, ships may show different sailing speed behaviour during a shipping leg (i.e., a trip between two ports). Based on interviews, Poulsen and Sornn-Friese (2015) mention imprecise voyage instructions and fast sailing at early stages of the trip as the main causes for large speed variations. Sailing fast at early stages of the trip is a common practice to buffer against delays at later stages of the trip. Different scenarios can result in the same average speed, but may vary in speed during the trip.

[5]

2.3 T

There is much uncertainty involved when it comes to port times in maritime transportation (Christiansen et al., 2013; Fransoo & Lee, 2013). According to Johnson and Styhre (2015), the two ships they analysed spent 40% their time in ports, and at half of this time being unproductive. Part of the time in port is unavoidable and involves productive activities such as cargo handing while the other part is often unproductive due to various reasons. Johnson and Styhre (2015) indicate early arrivals, closed ports, port congestion or waiting for pilot as the main reasons for longer waiting times in ports. Other reasons mentioned by Du et al. (2015) are time-varying draft restrictions due to tides but also the widely used hurry up and wait (HUW) policy by ship operators, resulting in high fuel waste, increased emissions, and long waiting times at ports.

Johnson and Styhre (2015) and Du et al. (2015) mention a large potential for just-in-time arrivals by reducing speed at sea when delays in ports are foreseen. This requires good communication between ship and shore besides a rational sailing behaviour by shippers. Increased port productivity is also mentioned as an improvement to prevent unproductive waiting times. Although port efficiency is not within the control of the shipping company, the impact of port congestion and port time randomness is an increasingly important factor (Lee et al., 2015). A better understanding of waiting times in ports could be used in ship scheduling to predict or even avoid port congestions.

[6]

3 M

ETHODOLOGY

In this research, an empirical data analysis on sailing speed behaviour and unproductive port times is conducted to gain a better understanding of the energy efficiency gap in shipping. By analysing detailed and real operational shipping log data, actual sailing speed behaviour at sea is captured instead of the commonly used speed averages over long time periods. Empirical evidence on sailing speed behaviour and unproductive port times provides guidance for increasing energy efficiency and challenges the validity of basic assumption made in quantitative theoretical models. Findings don’t show what shipping companies claim they do, or believe they do, but rather provide evidence on how they actually perform their operations, which is currently lacking in literature.

3.1 D

A mixture of large data tables containing (1) shipping log data, (2) route info and (3) route coordinates (i.e. waypoints) is used and provided by a third-party company that offers monitoring services to shipping companies. The data is the main resource for this research and contains data of five ships, representing three different types that among other things vary in size, deadweight and engine power (Table I). The data covers a timespan between 1 to 3 years, depending on the ship, and is collected real time by monitor systems installed on the vessels. Within this timespan, hundreds of variables are continuously measured where the mean values are collected at rate of 10 minutes at all times (i.e. either when sailing or not).

Table I. Characteristics per type of vessel

Type Size Main engine (kW) Deadweight (tons) Dataset

A B C Small Medium Large 1520 2970 4500 3850 12000 14600 1 ship 2 ships 2 ships

3.2 P

-

Prior to the actual data analysis, relevant variables need to be selected, different trips need to be distinguished and the data needs to be prepared in order to be used as input for the hierarchical cluster analysis. Due to the unstructured nature of the logger data, routes and waypoints, most of the time and effort in this research consists of pre-processing the data to eventually capture the underlying value.

3.2.1 Variable selection

Table II provides an overview of the variables selected for the research. Measured values are labelled with timestamps, and can therefore be classified as time-series data. Speed is expressed in nautical miles per hour (knots) and is used for determining the sailing speed behaviour and to see whether a vessel is sailing or not. Fuel consumption in liter/min is used to calculate the average fuel consumption during the trip and the GPS coordinates (latitude and longitude) in decimal degrees define the geographical position on earth of the corresponding ship and are required to match with the route waypoints to couple the shipping log data to their corresponding trips.

Table II. Selected variables from shipping log data

Datetime Speed Fuel Flow Latitude Longitude

[7] 3.2.2 Trips

The three data sources (1) shipping log data, (2) route info and (3) waypoints are used to determine when the vessels where actually sailing on a planned route (i.e. trip) in order to distinguish and compare the different trips. The measured shipping log data is coupled to their corresponding routes by matching the timestamps between data tables. The route IDs in the route info data table are subsequently matched with the route coordinates of the planned route. Finally, the actually measured GPS coordinates of the vessels (shipping log data table) are matched with the start- and end coordinates of the planned routes (waypoint data table) at all timestamps to determine when vessels actually started and ended their trips.

A clear definition of a trip is formulated in order to treat them equally and make them comparable. Hence, a single trip is defined as to start at the moment the vessel leaves its port of departure and ends when it leaves its port of destination, after visiting this port. This means that the each trip starts with a sailing part at sea followed by a part within a port. The sailing part covers the sailing speed behaviour at sea whereas the port part represents the total port time (i.e. productive and unproductive). In all cases, it is assumed that the vessel aims at leaving the port as fast as possible because both the vessel and the port authorities benefit from shorter turnaround times in the port. The process to identify the different trips is automated due to the vast amount of data (+- 1 to 2 million observations per vessel). To validate the results of the automated procedure, the individual trips are manually inspected and corrected when necessary.

Various assumptions and decisions are made throughout the process in order to automate the process and obtain usable outcomes. Based on observations in the data, validation of the trips, and discussions with a field expert the following is assumed:

 Trips under a length of 100 nautical miles (NM) are excluded from the analysis. These trips are assumed to be less prone to planning and speed decisions made by shippers that can affect the arrival time and are not interesting for this research.

 Vessels often wait outside a port due to various reasons whereas early arrivals are just one of them. A range of 15 NM from the end-coordinate of a route (i.e. port of destination) is used to define a port area such that waiting time outside a port is captured within the defined port part of a trip. Hence, whenever a vessel is within a range of 15 NM from its port of destination, it is considered to be in or around port. This enables the possibility to distinguish between the sailing time and port time of a defined trip to ultimately assess the sailing speed behaviour during the sailing part and the total port time in the port part.

 In some cases, vessels were sailing at sea (i.e. not within port area) without available route info in the route data table. These cases could not be defined as a trip from port to port and are left out of the analysis.

[8] 3.2.3 Data normalization

Different trips may show a relative similar shape in sailing speed behaviour, but at a different absolute

speed. These cases are perceived similar in this research and adjusted such the cluster algorithm

assigns these trips to the same cluster. To achieve this, all i-th measured actual speeds v on the j-th trip is normalized to a relative speed factor zij as difference from the average speed over the entire trip

with formula (1):

𝑧

=

𝑣

Ṽ

Where is:

zij - i-th normalized value as factor of the mean speed at trip j, vij - i-th actual measured speed at trip j,

Ṽj - Average of i values in vector V, representing the sailing part of trip j

Similar sailing speed behaviour can occur relatively at the same part of the trip but vary in absolute

distance because of the differences in trip lengths. The covered distance at each observation i is

calculated by cumulating the great-circle distance (shortest distance over the earth’s surface) between the GPS coordinates at the previous observation i-1 and the GPS coordinates of the current observation i. To assess the sailing speed behaviour between trips over the same relative distance segments of the trip, the cumulative distance at every observation i is normalized to a [0-1] scale with formula (2):

𝑓

=

𝑑

𝑚𝑎𝑥 𝐷𝑗

Where is:

fij - i-th normalized covered distance as factor of the total distance at trip j, dij - i-th covered distance at trip j,

Dj - Vector of i values, representing cumulative distances at trip j (i max = total distance) The different trips contain unequal data points due to differences in speed, trip length and is a common issue in clustering of time-series data (Liao, 2005). Handling unequal data point can basically be executed in two ways Liao (2005): Either the cluster algorithm or the input data can be modified such that existing cluster algorithms appropriate for static data can be applied. The latter is applied in this research by converting the raw time-series data of the trips into a vector with a reduced and equal dimension (Liao, 2005). Although information is lost in this process, this approach decreases computation time, simplifies visualization and the analysis and results might even be more accurate in case all observations i are used (Laureshyn et al., 2009).

The dimension of the unequal data points at all trips j is normalized by aggregating the normalized values Zij at every i-th observation where dij reached a 0.05 increase, resulting in 20 data points per trip

which each represents the average normalized speed as factor of the mean speed over approximately each 5% segment k of the total distance of trip j with formula (3):

𝑧𝑠

=

∑

𝑧

|𝑠|

Where is:

zssj - Average normalized value i of normalized values Zij over segment s at trip j, zij - i-th normalized speed at trip j,

[9]

3.3 D

After modifying the data during the pre-processing stages, the actual data analysis is conducted to obtain the final results. The normalized trips serve as input for the hierarchical clustering method, which is discussed first. Secondly, the method for calculating the impact of certain sailing speed behaviour on fuel consumption is discussed. The method for assessing whether certain sailing speed behaviour more often leads to early arrivals is discussed lastly.

The types of vessels are analysed separately in order to effectively compare results and to prevent mixed results that might be caused by the vessel’s characteristics (e.g., more impact of weather conditions on speed profile at small sized vessel). Port times are expected to be different because more cargo on a larger vessel logically takes more time to unload than less cargo on a smaller vessel. 3.3.1 Hierarchical clustering

Clustering is a common activity in data mining research and generally aims for the identification of homogeneous groups within a dataset (Liao, 2005). An extensive overview of different clustering techniques is given by Jain and Dubes (1988). In general, a distinction can be made between partitional clustering and hierarchical clustering. Hierarchical clustering techniques have the advantage that a predefined number of clusters is not required before the analysis (Weijermans & van Berkum, 2009). Due to the exploratory nature in this first attempt to identify different speed profiles in shipping, the number of expected clusters is unknown in advance. Therefore, the hierarchical clustering procedure is applied in this research.

More specifically, the Ward’s method (Murtagh & Legendre, 2014) is applied in this research and is observed at other studies as well were similar types of datasets were used (Weijermars & van Berkum, 2009; Gerbec et al., 2002). The Ward’s method is an agglomerative, bottom-up approach which minimizes the within group-dispersion at every iteration by using the minimum sum-of-squares criterion (Murtagh & Legendre, 2014). The technique makes use of the Euclidean distances between the trips, based on the normalized values at all 20 segments of the trip vectors. The procedure starts with a separate “cluster” for every trip included in the analysis. Clusters are paired at every iteration through the minimization of the sum-of-squares variance until all trips are linked in one cluster. An important issue in cluster analysis is the evaluation of the cluster results. Visualization is a crucial verification technique that is often used (Halkidi et al., 2001). A first indication on the number of appropriate clusters within the dataset is determined by a “scree plot” (Figure 1a) and the KL-statistic (Figure 1b) as proposed by Krzanowski and Lai (1988) and also used by Larson et al. (2005). A scree plot shows the within cluster sum of squares (WCSS) over a range of cluster solutions (i.e. number of clusters to choose). An observed “elbow” (circled red) suggests the number of clusters at the scree plot. The KL-statistic determines the error improvement between a cluster solution with k clusters and k + 1 clusters and suggests the cluster solution with the highest score computing for each solution with k clusters by using formulas (4) and (5):

𝐷𝐼𝐹𝐹(𝑘) = (𝑘 − 1)2/𝑝𝑊𝑘−1− 𝑘 2 𝑝_𝑊 𝑘 (4) 𝐾𝐿(𝑘) = 𝐷𝐼𝐹𝐹(𝑘) 𝐷𝐼𝐹𝐹(𝑘 + 1) (5) Where is: K - Number of clusters

[10]

Figure 1 (a+b). Example of a scree plot (a) and a KL-statistic graph (b)

The cluster procedure is visualized by a dendrogram (Figure 2) which is a hierarchical tree that shows the linkages between clusters made at each iteration until one cluster is formed. The suggested number of clusters by the scree pot and KL-statistic is assessed by cutting the dendrogram at a height such that the number of “branches” is equal to the suggested number. “Cutting the tree” at a fixed height is a rather static approach and one of the drawbacks in this method. In case of unsatisfying results, large clusters are divided in sub-clusters through a more dynamic selection of clusters at different heights in de dendrogram to achieve better results. These improvements are realised through manual inspection and makes it less applicable for a larger amount of trips. Improvements through this approach comply with the suggested improvement techniques by Laureshyn et al. (2009), mentioning that a larger number of clusters can be chosen or cluster analysis is performed within clusters that appeared to have a large variation. The dendrogram structure provides visual guidance in the above mentioned cluster improvement techniques.

Figure 2. Example of a dendrogram

5 10 15 20 0 5 10 15 Scree plot # clusters W C S S 0 .8 1 .2 1 .6 KL-statistic # clusters K L-S ta ti s ti c 2 3 4 5 6 7 8 9 5 10 15 20 0 5 10 15 Scree plot # clusters W C S S 0 .8 1 .2 1 .6 KL-statistic # clusters K L-S ta ti s ti c 2 3 4 5 6 7 8 9

8 4

27 33 34 25 29 36 40 28 41 24 35 42

1 5

32 18 38

9 2

21 43 15 26 20

3

37 10 14 17 22 30 12 19 11 39

6

31

7

16 13 23

0

.0

0

.5

1

.0

1

.5

2

.0

Vessel type 1 - Cluster dendrogram

[11] 3.3.2 Fuel consumption

Due to the non-linear relationship between speed and fuel consumption (Section 2.3), sailing at a constant speed during the entire trip is theoretically more efficient than varying speeds at different segments of the trip, even when the average speed is equal.

The impact on the observed sailing speed behaviours on fuel consumption is assessed by a comparison with the case where vessels would have sailed constant at the same average speed over all 20 segments. In this case, the vessel would have arrived at its destination at the exact same time, but by sailing at a constant speed instead of varying speeds. The extent to which sailing speed behaviour impacts fuel consumption thus depends on the shape of the different speed profile and the degree of deviations from the average speed.

A ship’s actual fuel consumption is influenced by many determinants and in an extremely complicated way (Meng et al., 2016; Psaraftis & Kontovas, 2014), which makes it difficult to isolate the impact of speed fluctuations on the empirically observed fuel consumption. Therefore, the generally accepted cubic law between speed and fuel consumption is used to calculate the theoretical difference in fuel consumption.

The cubic function assumes that the difference in fuel consumption per time unit is the difference in speed to the power of three. Because the speed behaviour is expressed in 20 segments of 5% of the total distance of the trip, the fuel consumption is calculated per distance unit (L/NM) and is commonly used to express energy efficiency. Based on the normalized speeds per segment, the total fuel consumption at trip j is expressed as factor in difference from the case where constant speed would have been retained and is calculated by setting up formula (6):

𝐹𝐹𝐶𝑗= ∑ (𝑍𝑠𝑠𝑗3∗ (_𝑍𝑠1 𝑠𝑗)) 20 𝑖=1 ∑20𝑖=1𝑍𝑠𝑠𝑗 = ∑ 𝑍𝑠𝑠𝑗 2 20 𝑖=1 20 = 0,05 ∑ 𝑍𝑠𝑠𝑗 2 20 𝑖=1 (6) Where is:

FFCj - Fuel consumption as factor from fuel consumption at constant average speed at trip j, 𝑍𝑠𝑠𝑗3 - Difference factor in fuel consumption based on normalized speed over segment s at trip j,

1/𝑍𝑠𝑠𝑗 - Factor of the time needed to cover the distance over segment s with speed Vnij at trip j with

respect to time needed at average speed.

∑ Zssj - Sum of speed factors over 20 segments with an average of 1.00, making the sum 20 at all

times.

By using this approach, the impact of a certain sailing speed behaviour on fuel consumption is determined. It is assumed that all other conditions during the trip would have remained equal in case of sailing a constant speed. Furthermore, it is assumed that all speed changes were within control of the shipper (e.g., no increases or decreases due to weather conditions) and made around the design speed of the ship. Contrary to slow or near-zero speed, the speed range around the design speed is where the cubic function is most applicable (Psaraftis & Kontovas, 2014).

3.3.3 Early arrivals

[12]

Throughout this research, it is observed that many trips showed an expected zero-speed period near the end-coordinates of the route, i.e. the port of destination. Also observed is a slightly different pattern like displayed in Figure 3. The latter is a strong indication of a vessel that waits outside the port for a certain amount of time before actually entering the port (speed peak, circled red). Especially because of the very tiny speed fluctuations around zero before the speed peak, probably caused by waves and currents while anchored outside the port. This could mean that the vessel arrived too early and is defined as an early arrival indication.

Figure 3. Example of early arrival indication (EAI)

In case the early arrival indications are relatively more frequently observed at trips within a given cluster (speed profile), then it might implicate that this behaviour more often leads to early arrivals and unproductive waiting time. To explore if there is a relationship between a certain sailing speed behaviour and early arrivals, the number of early arrival indications are counted within all speed profiles through manual inspection.

An early arrival indication shows that the shipper could have reduced speed during the trip in order to reduce the fuel consumption and arrive just-in-time. The extent to which the speed could have been reduced depends on the duration of unproductive waiting time. Therefore, the amount of waiting time before the peaks at the early arrival indications are measured as well. The potential speed reduction and fuel savings at trip j with an observed early arrival indication is calculated by constructing formulas (7) and (8): 𝑣𝑝𝑗= 𝑑𝑗 𝑡𝑠𝑗+ 𝑡𝑤𝑗 (7) 𝐹𝑠𝑗= 1 − ( 𝑣𝑝𝑗 𝑣𝑎𝑗 )2 ₍₈₎ Where is:

tsj - The currently observed sailing time in hours to cover the trip distance dj at trip j, twj - The observed waiting time in hours outside the port at trip j,

vpj - the average potential speed that could have been used on trip j, vaj - the currently observed average speed in knots at trip j,

Fsj - Percentage of fuel that could have been saved by sailing at a lower speed vpj at trip j

By using this approach, the cubic law is applied to calculate the potential fuel savings. It is assumed that the speed could not have been reduced below the most energy efficient speed of the vessel. The most energy efficient speed is assumed to be similar at the vessels observed by Johnson and Styhre (2015). For vessel A, B, and C this threshold is assumed to be respectively 8 knots, 9.5 knots, and 10.5 knots. The values 8 knots and 9.5 knots mentioned by Johnson and Styrhe (2015) while 10.5 knots are estimated at Type C, based on the different characteristics of the vessels.

Trip (Sailing + Port time)

[13]

4 F

INDINGS

The findings are presented for each type of vessel, followed by an overview of similarities and differences in findings between the vessels. Besides the results from the analysis in speed profiles, fuel consumption, and early arrivals, the main statistics of each type are also presented to get a sense of the absolute values in the underlying data.

4.1 V

A

From the available dataset, 43 different trips were identified and normalized to be used as input in the cluster analysis. To get an idea about the underlying data, an overview of overall statistics of all trips made by vessel type A is presented in Figure 4. This figure shows how the trip lengths, the total port times, averages speed, and average fuel consumption of the trips are distributed. The mean trip length was 550 NM (SD = 549). An average speed of 9.35 Knots (SD = 1.05) with an average fuel consumption 24.13 L/NM (SD = 4.81) is observed. The resulting port times after the trips are on average 61.28 hours (SD = 47.72), disregarding an outlier of 1092 hours where it is assumed that the vessel was out of operation.

Figure 4. Histograms of relevant variables at vessel type A

4.1.2 Speed profiles

In the cluster analysis, both the scree plot and KL-statistic (Appendix A) suggested a 5-cluster solution, which was used as an initial indication in determining the clusters. A closer look revealed three different sub-clusters within the cluster 1+2+3 (Figure 5). The initially suggested cluster solution treated these trips in a single cluster because they are relatively similar in comparison to the other clusters (linked at a lower height). Treating these sub-clusters separately is desirable in this case as they show clearly different behaviour. Trips on the left and the right of Cluster 1 (Figure 5) are merged into Cluster 2. These trips show speed fluctuations of around 10% of the average speed during the entire trip. They are numerically different because the fluctuations don’t always occur at the same distance segments but basically represent the same speed profile in this research. The final proposed cluster solution contains seven clusters (Figure 6), representing different speed profiles of vessel type A where black lines are the individual trips and red lines the group averages.

[14]

Figure 5. Dendrogram of cluster analysis vessel type A

Figure 6. Cluster plots vessel type A

Different and recurring speed patterns are clearly present within the analysed trips. From all 43 trips, 11 trips (25.6%) were sailed at a rather constant speed (Cluster 1). 9 trips (21.9%) showed larger fluctuations around the average speed (Cluster 2). Another 10 trips (23.3%) kept a constant speed during the major length of the trip but clearly decreased in speed at the last 20% of the trip distance (Cluster 3) and looks like a conscious decision. 6 trips (14.0%) in Cluster 4 slowed down halfway the distance before speeding up again. At 3 trips (7.0%, Cluster 5), the vessel increased its speed during the whole trip whereas another 3 trips (7.0%, Cluster 6) showed the opposite. The one and only trip in Cluster 7 (2.3%) shows very deviating behaviour at a low average speed of 5.66 knots but a high average fuel consumption of 43.50 L/NM.

8 4 27 33 34 25 29 36 40 28 41 24 35 42 1 5 32 18 38 9 2 21 43 15 26 20 3 37 10 14 17 22 30 12 19 11 39 6 31 7 16 13 23 0 .0 0 .5 1 .0 1 .5 2 .0

Vessel type A - Cluster dendrogram

[15] 4.1.3 Fuel consumption

The impact of the different sailing speed behaviours on fuel consumption at vessel type A is presented in Table III that provides the average savings (in %) per trip per cluster.

Table III. Average fuel savings per trip per cluster in case of constant speed at vessel type A Clusters N Average potential fuel savings per trip

C1 C2 C3 C4 C5 C6 C7 33 (25.58%) 9 (20.93%) 10 (23.26%) 6 (13.95%) 3 (6.98%) 3 (6.98%) 1 (2.33%) 0.19% 1.20% 0.97% 2.37% 3.02% 4.32% 7.95% All trips 43 (100%)

Trips in Cluster 1 show rather constant sailing speed behaviour and show a near optimal fuel consumption at the corresponding average speed. On average, savings of 0.19% (SD = 0.21%) per trip could be achieved if the average speed could have been constantly retained. Trips in Cluster 2 show larger variations in speed, resulting in 1.20% (SD = 0.60%) average fuel savings per trip. 0.97% (SD = 0.40%) could have been saved on average per trip in Cluster 3 where a presumably conscious decrease in speed is observed at the last 20% of the trip. Trips at cluster 4, cluster 5, and cluster 6 have the most fuel saving potentials of respectively 2.37% (SD = 0.95%), 3.02% (SD= 1.58%), and 4.32% (SD = 1.98%). Cluster 7 is assumed to be an exceptional trip under assumed extreme conditions but does show that such extreme variations in speed could theoretically lead to about 8% increased fuel consumption compared to constant speed.

4.1.4 Early arrivals

In total, 8 out of 43 trips (18.60%) showed an early arrival indication at vessel type A. Detailed results on trips with early arrival indications within clusters are displayed in Table IV.

Table IV. Overview number of trips with early arrival indications per cluster at vessel type A Clusters # Trips Trips with early

arrival indications

% within cluster % of total early arrival indications C1 C2 C3 C4 C5 C6 C7 11 9 10 6 3 3 1 2 1 3 0 2 0 0 18.18% 11.11% 30.00% 00.00% 66.67% 00.00% 00.00% 25.00% 12.50% 37.50% 00.00% 25.00% 00.00% 00.00% Total 43 8 (18.60%)

Proportionally, trips within Cluster 3 (30.00%) and 5 (66.67%) showed most of the early arrival indications. At Cluster 3, this could mean that shippers realized at 80% of the distance that they were ahead of schedule and reacted by decreasing their speed at the last 20% of the distance but still arrived too early. The opposite could have happened at Cluster 5, namely a realization of being behind on schedule at 50% of the trip distance, trying to compensate by speeding up at the second half of the trip but overestimated the delay and arrived too early.

[16]

Figure 7. Histogram of unproductive waiting times at Type A

Table V. Speed reduction and fuel savings by using unproductive waiting time

Waiting time Speed reduction Fuel savings Average per trip with early arrival indication

Std. Dev. 9.5 hrs 17.5 hrs 0.84 knots 0.57 knots 15.95% 9.16%

On average, the unproductive waiting time outside the port was 9.5 hours at the trips that showed an early arrival indication. As can be observed in the histogram, a relatively large waiting time of around 52 hours was measured at one trip. While the other trips all showed waiting times of around 2-3 hours, this value strongly affects the average value.

Over the 8 trips that showed an early arrival indication, the average speed during voyage could have been reduced by an average of 0.84 knots (SD = 0.57) while satisfying the lower bound in absolute speed as described in Section 3.3.3. Theoretically, this could have saved an average of 15.95% (SD = 9.16%) of fuel at the trips that contained an early arrival indication.

Type A - Waiting time

[17]

4.2 V

B

4.2.1 Descriptive statistics

115 trips where successfully identified at this vessel. The average trip distance is 1142 NM (SD = 1047), with 10.32 Knots on average (SD = 0.89) and a fuel consumption of 47.39 L/NM (SD = 12.82). The total port time was on average 80.57 hours. Figure 8 gives an overview of how the trip lengths, resulting port times, and the averages of the speed and fuel consumption during the trips are distributed.

Figure 8. Histograms of relevant variables at vessel type B

4.2.2 Speed profiles

In the cluster analysis of vessel type B (Appendix B), the scree plot didn’t suggest a clear number of clusters. The KL-statistic indicated a 2-cluster solution but also showed a relatively high score for a 5-cluster solution. The 2-5-cluster solution resulted in high within-5-cluster variation. The 5-5-cluster solutions was an improvement but still not satisfying. The final cluster solution contains six clusters and are displayed in Figure 9.

[18]

A vast amount of trips (N=33, 28.69%) showed little variation in speed and represent Cluster 1. Another 20 trips (17.39%) showed similar behaviour but on average started to slightly decrease speed after around 80% of the trip, followed by a steep decrease at 95% and is defined as Cluster 2. Cluster 3 shows very fluctuating speeds and sails on average at a higher speed in the second half of the trip and contains 16 trips (13.91%). The 20 trips (17.39%) in Cluster 4 have similarities with Cluster 2 but vary more in speed and have on average a decreasing speed during the entire trip, followed by a steeper decrease in speed over the final 5% of the distance. A clear halfway speed decrease seems also present at this type of vessel and is carried out by the 8 trips (6.96%) in Cluster 5. The 18 trips (15.65%) in Cluster 6 showed the opposite behaviour as trips within Cluster 3 by decreasing their speed at the second half of the trip but with less speed fluctuations.

4.2.3 Fuel consumption

To see the impact of different sailing speed behaviour on the fuel consumption at this type of vessel, potential savings are theoretically calculated with the cubic law as well and listed in table VI.

Table VI. Average fuel savings per trip per cluster in case of constant speed at vessel type B Clusters N Average potential fuel savings per trip

C1 C2 C3 C4 C5 C6 33 (28.69%) 20 (17.39%) 16 (13.91%) 20 (17.39%) 8 (6.96%) 18 (15.65%) 0.20% 0.30% 1.63% 1.08% 1.46% 1.55% All trips 115 (100%) sdsddsdsds ds

On average, 0.20% (SD = 0.18%) per trip could theoretically be saved at Cluster 1 if the average speed was sailed constantly during the entire trip. For trips in Cluster 2 this is on average 0.30% (SD = 0.10) per trip, showing that the speed decrease in the final part of the trip does not have large consequences on fuel efficiency. Larger savings of 1.63% (SD = 0.69%), 1.08% (SD = 0.53%), 1.46% (SD = 1.79%), and 1.55% (SD = 1.00%) could have been achieved at the trips in cluster 3, 4, 5, and 6 respectively. Most interesting is cluster 4, where underlying conscious decisions are more likely to be the cause of the difference in speed between the first part and the second part of the trip.

4.2.4 Early arrivals

The number of early arrival indications is also counted for this type of vessel and reported in Table VII.

Table VII. Overview number of trip with EAIs per cluster at vessel type B Clusters Trips Trips with early

arrival indications

% within cluster % of total early arrival indications C1 C2 C3 C4 C5 C6 33 20 16 20 8 18 9 6 4 9 3 4 27.27% 30.00% 25.00% 45.00% 37.50% 22.22% 25.71% 17.14% 11.42% 25.71% 8.57% 11.42% Total 115 35 (30.43%)

[19]

or foreseen delays at the port. In that case, energy efficiency increased by lowering their speed such that unproductive waiting time at port was avoided.

The absolute durations in waiting time at the trips that showed an early arrival indication are presented in the histogram in Figure 10 while the potential speed reduction and fuel savings over these trips are listed in Table VIII.

Figure 10. Histogram of unproductive waiting times at Type B

Table VIII. Speed reduction and fuel savings by using unproductive waiting time

Waiting time Speed reduction Fuel savings Average per trip with early arrival indication

Std. Dev. 53.8 hrs 63.2 hrs 0.85 knots 0.74 knots 14.41% 11.65%

The unproductive waiting time at trips with an early arrival indication was on average 53.8 hours at vessel type B. The histogram shows how the durations are distributed. Most durations stay under 50 hours while some measured waiting times even took than a week.

Over the 35 trips that showed an early arrival indication, the average speed during voyage could have been reduced by an average of 0.85 knots (SD = 0.74). At various some trips the speed reduction was limited by the constraint described in Section 3.3.3 while at other trips no speed reduction was possible because the average speed was already below this bound. Overall, the speed reductions could theoretically have saved an average of 16.81% (SD = 10.83%) of fuel per trips that contained an early arrival indication.

Type B - Waiting time

[20]

4.3 V

C

4.3.1 Descriptive statistics

154 trips where successfully identified at vessel type C. The average trip distance at this type was 1478 NM (SD = 1347), at an average of 11.34 Knots (SD = 1.26) and a fuel consumption of 53.29 L/NM (SD = 13.11). The fuel consumption data contains values from 128 out of the 154 trips due to a period of fuel gauge malfunctioning. Figure 11 gives an overview of how the trip lengths, port times, averages speed, and average fuel consumption of the trips are distributed. The time spend in ports after the trips was on average 109.29 hours (SD = 81.84).

Figure 11. Histograms of relevant variables at vessel type C

4.3.2 Speed profiles

[21]

Figure 12. Cluster plots vessel type C

The cluster containing the most members (42, 27.27%) was again formed by trips that sailed at a reasonably constant speed, in this case cluster 6. Another similar behaviour is a speed decrease at the last part of the trip like the 14 trips (9.09%) in cluster 5, or the 14 (9.09%) similar but more extreme cases in cluster 8. The 27 trips (17.53%) in cluster 1 show on average a decreasing speed during over entire trip. Cluster 2 (4.55%) and cluster 3 (2.60%) are overlapping and show a temporary speed decrease around 20-40% of the average speed somewhere during the trip. Cluster 4 (12.99%) and cluster 7 (16.23%) are close to cluster 6 but do show some differences mainly at first part of the trip. Just as vessel type 1, one outlier that formed a cluster on its own was detected. This trip is not taken into account as it shows a relatively very low speed of 6.82 Knots and is assumed to be an irregular trip.

4.3.3 Fuel consumption

Also for this type of vessel the theoretical savings in fuel consumption are calculated by the cubic law and compared with the case where the average speed of the trips would be kept constant over all segments in the entire trip. The results are displayed in table IX.

[22]

Table IX. Average fuel savings per trip per cluster in case of constant speed at vessel type C Clusters N Average potential fuel savings per trip

C1 C2 C3 C4 C5 C6 C7 C8 C9 27 (17.53%) 7 (4.55%) 4 (2.60%) 20 (12.99%) 14 (9.09%) 42 (27.27%) 25 (16.23%) 14 (9.09%) 1 (0.65%) 1.26% 1.53% 4.70% 0.81% 0.36% 0.21% 0.62% 1.98% 9.90% All trips 154 0.95%

The trips in Cluster 6 sail quite constant and show a 0.21% difference to the fuel consumption at constant speed. A more constant speed at trips in Cluster 1 seems achievable but depends on the actual reasons behind their speed decrease over the entire trip. If this speed profile resulted from conscious decisions, an average of 1.26% per trip could theoretically have been saved. Trips following the speed profile in Cluster 5 or Cluster 8 are more likely to be based on rational behaviour. The 1.98% theoretical fuel savings on average per trip makes Cluster 8 particularly interesting. Although it remains unclear what the underlying reasons of the halfway speed-decreases at Cluster 3 are, trips could have saved as much as 4.70% of fuel savings per trip when sailing constantly at average speed. Over all 154 trips, an average of 0.95% of fuel per trip could theoretically be saved.

4.3.4 Early arrivals

Counting the number of early arrival indications at this type of vessel, showed that more than one-third (31.82%) of the trips resulted in waiting time outside the port of destination. Within clusters (Table X), trips in Cluster 3 performed relatively worse than other clusters with 75% of the trips with an early arrival indication but also contained only a few trips and might be misleading. More interesting is a 55.00% in Cluster 4 and a 38.10% in Cluster 6 that consist of respectively 20 and 42 trips that all sailed reasonably constant. Remarkable is that only 1 out of 14 trips (7.14%) showed an early arrival indication at Cluster 5, indicating that an expected delay might have been avoided by decreasing speed at the last part of the trip but was not observed at the other vessels. Cluster 8 on the other hand showed an early arrival indication at 5 out of 14 trips (35.71%), which is contradicting since a similar but even stronger speed decrease is observed when compared with Cluster 5.

Table X. Overview number of trip with EAIs per cluster at vessel type C Clusters # Trips Trips with early arrival

indications

% within cluster % of total early arrival indications C1 C2 C3 C4 C5 C6 C7 C8 C9 27 7 4 20 14 42 25 14 1 5 2 3 11 1 16 6 5 0 18.51% 28.57% 75.00% 55.00% 7.14% 38.10% 24.00% 35.71% 00.00% 10.20% 4.08% 6.12% 22.45% 2.04% 32.65% 12.25% 10.20% 00.00% Total 154 49 (31.82%)

[23]

Figure 13. Histogram of unproductive waiting times at Type C

Table XI. Speed reduction and fuel savings by using unproductive waiting time at Type C

Waiting time Speed reduction Fuel savings Average per trip with early arrival indication

Std. Dev. 42.6 hrs 52.7 hrs 0.74 knots 0.76 knots 12.74% 10.72%

The unproductive waiting time at trips with an early arrival indication was on average 42.6 hours at vessel type C. The histogram shows how the durations are distributed. Most durations are under 50 hours while few measured waiting times varied between larger values of around 100-250 hours.

Type C - Waiting time

[24]

4.4 O

A,

B

C

Various differences are observed in the underlying data between the three types of vessels. Type A (smallest vessel) sailed the shortest distances at the lowest average speed and fuel consumption and also showed the shortest port times. Type B (medium-sized vessel) shows higher values than Type A but lower values than type C (largest vessel). An overview is found in Table XII.

Table XII. Overview of average values in underlying data per type of vessel

Vessel type Trip length (NM) Speed (knots) Fuel consumption (L/NM) Total port time (hrs)

Type A Type B Type C 550 1142 1478 9.35 10.32 11.34 24.13 47.39 53.29 61.28 80.57 109.29

Several similarities in sailing speed behaviour are found between the types of vessels. A reasonably constant speed was observed at all types of vessels. Another frequent observed behaviour is a speed drop at the last part of the trip. Although there are differences in the moment (e.g. final part or earlier stage), size, and increment (e.g. abrupt or gradually) of the speed drop, this pattern is observed at all types of vessels. Also notable is the presence of a temporary speed decrease of around 20-40% during the trip. This behaviour is also observed at all types of vessels and occurred most of the times half-way the trip or around 70%. An overview of the roughly sketched trends and the corresponding clusters at all vessels are found in Table XIII.

Table XIII. Overview of main trends in speed profiles at types A, B and C

Reasonable constant Final speed-drop Gradual speed-drop Temporary speed-drop

Speed profile Sketch Type A – Cluster: 1 (25.6%) 3 (23.3%) 6 (7%) 4 (14%) Type B – Cluster: 1 (28.7%) 2(17.4%), 4 (17.4%) 6 (15.7%) 5 (7%) Type C – Cluster: 6 (27.3%) 5 (9.1%) 8 (9.1%) 2 (4.6%), 3 (2.6%) Impact on fuel consumption +- 0.20% +- 0.30 – 1.0% +- 1.5 – 4.3% +- 1.4 – 4.7%

The impact of sailing speed behaviour on fuel consumption is similar at all types of vessels (Table XIII) because the shapes and degree of deviation from the average speed does not differ much between vessels. Overall, sailing at a reasonable constant speed results in a near optimal fuel consumption of around 0.20% difference to optimal. More fluctuating speeds up to 20% from the average have a larger impact and lies around an 0.80-1.20% increase in fuel consumption at for example Cluster 2 at type A. Higher fluctuations of around 40% of the average speed could lead up to 5% in fuel consumption as observed at for example Cluster 3 at type C. Although Cluster 9 at type C represents only trip and is assumed to be an outlier, it does show that 60% speed fluctuations could theoretically lead to an almost 10% increase in fuel consumption.

Table XIV. Overview of average values in underlying data per type of vessel Vessel

type

% of trips with early arrival indications

Average waiting time at port

Average potential speed reduction

Average potential fuel savings

Type A Type B Type C 18.60% 30.43% 31.82% 9.5 hrs 53.8 hrs 42.6 hrs 0.84 knots 0.85 knots 0.74 knots 15.95% 14.41% 12.74%

[25]

5 D

ISCUSSION

Speed profiles

In prior research on energy efficiency in shipping, speed is often treated as a constant parameter during their voyages (Psaraftis & Kontovas, 2014). This assumption is refuted by the results of the hierarchical clustering analysis in this research that show the presence of a wide variety in sailing speed behaviour at sea. Despite this variety, there are interesting and recurring patterns in the main trends that can be classified into typical speed profiles.

One major speed profile at all types of vessels does represent a reasonable constant sailing speed behaviour during the entire trip. 25.6%, 21.4%, and 27.3% of the trips at respectively vessel type A, B, and C followed this behaviour. Other speed profiles like Cluster 4 and 7 at type C also showed similar behaviour but the threshold whether these profiles are considered as constant speed can be argued. As studied by Prpić-Oršić & Faltinsen (2012), speed losses can be caused by waves and wind at open sea and is the most likely cause for the minor, but continuous fluctuations in speed that are observed at most of the speed profiles.

Besides the minor fluctuations, the majority of trips showed considerable speed fluctuations where the main trends are more likely to be caused by conscious decisions made by the shippers. For example, a frequently observed sailing speed behaviour at all types of vessels showed an abrupt speed drop or gradual speed drop in the last part of the trip, or already halfway the trip. This kind of behaviour is observed at 30.3% of the trips at Type A in Cluster 3 and 6. At Type B, 34.8% of the trips shows this effect slightly in Cluster 2 and 4, but more extreme in Cluster 6 (15.6%). The same holds for Type C, where this effect is slightly present at 9.1% the trips in Cluster 5 and more extreme at another 9.1% in Cluster 8. The speed reductions seem like rational decisions and may indicate a response to realizing that one is sailing ahead on schedule or to an expected delay at the port of arrival. Being ahead on schedule could be caused by shippers that sail unnecessary fast at early stages of the trip to buffer time against delays at later stages of the trip. This seems to comply with Poulsen and Sornn-Friese (2015), who mention that this behaviour is a common practice among shippers can be the reason behind this observed speed profile. Reducing speed at a later stage of the trip obviously saves fuel compared to remaining the same high speed but moreover indicates that the shipper could have anticipated at an earlier stage. Sailing at a lower and constant speed onwards the beginning could have prevented to be ahead on schedule while consuming less fuel.

Another remarkable and recurring sailing speed behaviour is observed at 13% of the trips at Type A in Cluster 4, at 5.2% the trips at vessel B in Cluster 5, and at 7.1% of the trips at Type C in Cluster 2 and 3. At these trips, a temporary speed decrease of around 20-40% is observed somewhere during the trip. This may be decided due to safety concerns in extreme weather conditions (Prpić-Oršić et al., 2014) or perhaps due to instructions from the shore to slow down because of port congestion but got instructed to sail faster at a later stage because problems were resolved at the port of destination. In this research, the behaviour at this speed profile and the undiscussed speed profiles only allow for speculation on their rationality but does show that vessels sail far from constant at most trips.

Fuel consumption

[26]

sailing speed behaviour at sea and their impact on fuel consumption. Hence, there were no expectations on possible numerical results.

The speed profiles with constant speed resulted in a theoretically near optimal fuel consumption with only 0.20% fuel saving potential. Average fuel savings of 0.30% to 2.00% per trip are observed at speed profiles that showed a clear speed drop at the last part of the trips. These speed profiles are likely caused by the common practice to sail fast at earlier stages of the trip, meaning that actual improvements should be realistic. The speed profiles that contain a temporary speed decrease of around 20-40% somewhere during the trip can impact fuel consumption between 1.5 - 4.7%. Higher fluctuations in speed of around 60% above average can even result in a 10% higher fuel consumption. Whether savings are realistic depends on the reason behind the behaviour. In case of safety reasons in bad weather conditions, this could not be easily avoided. Overall, it is worth striving for sailing at a constant speed while large speed fluctuations (+- 40%) should be avoided if possible.

Despite the limited savings in terms of percentage, absolute fuel savings and emission reductions are rather impressive when taking into account that shipping companies own entire fleets of ships that operate 24/7, that fuel costs can represent up to 50% of total operation costs at a shipping company (Jafarzadeh & Utne, 2014) and that the worldwide shipping industry constitutes for around 3% of global CO2 emissions (Buhaug et al., 2009). To illustrate, saving 1% of fuel consumption on 25 vessels of Type

C that consume around 14700 L/day (15 tons), comes down to saving 3675 Litres (14700*25*0.01) of fuel per day. Even more interesting is that basically all it requires is a change in behaviour rather than a monetary investment.

Early arrival indications

By examining the indications of early arrivals at the different speed profiles, there is no typical sailing speed behaviour found that tends to lead relatively more often to early arrivals and unproductive waiting time.

[27]

The durations in waiting times further expose the inefficiencies. At trips that contain early arrival indications, an average waiting time of 9.5 hours was found at Type A. When unproductive waiting times were avoided and used at sea, an average speed reduction of 0.84 knots could have theoretically resulted in an average fuel saving of 15.95% per trip. This was 53.8 hours, 0.85 knots and 14.41% at Type B. Type C showed 42.6 hours, 0.74 knots and 12.74%. Despite different waiting times, similar speed reductions and fuel savings are observed and can be explained by differences in trip lengths. Using a few hours waiting time at sea to reduce speed has a relatively large effect at shorter trips (Type A) than longer trips (Type B and C). Some early arrival indications showed very long waiting times between 100 – 300 hours. It is questionable whether these observations can be related to the various reasons for unproductive waiting time as Johnson and Styhre (2015) found that unproductive waiting times mostly accounted for 0 to 36 hours. Another explanation could be that the vessels were out of operation at these moment and just waited for freight opportunities. In this case there is still no need for fast sailing and speed could have been reduced as a fuel saving solution in depressed freight markets (Poulsen and Sornn-Friese, 2015).

Theoretical implications

This empirical research carries various theoretical implications regarding the sailing speed behaviour of vessels at sea, unproductive waiting times at ports and the port times in general. Through the empirical data analysis, lacking knowledge on actual shipping operations is addressed and provides input that can be used for the development of new and existing models that currently focus more on theoretical contributions rather than real operations (Christiansen et al., 2013). As reviewed by Psaraftis & Kontovas (2014), sailing speed is often treated as a fixed variable during voyage. The findings on sailing speed behaviour at sea provide new insights on how to treat sailing speed in future research as it can affect fuel consumption and estimated sailing times. This also addresses the significant uncertainty in sailing times and port times in maritime transportation, which is barely considered by researchers (Christiansen et al., 2013). Both findings on unproductive port times as well as total port times relate to the operational variability that is increasingly important to incorporate in research and was explicitly called for by Fransoo and Lee (2013). Assumptions on normally or uniformly distributed port times (Qi & Song, 2012) are challenged by showing that port time distributions over multiple ports does not tend take the shape of a normal or uniform distribution. Overall, the research show researchers the possibilities of studying such shipping log data and stimulate new ideas for future research. Besides, the used clustering method could be extended to other shipping sectors or even to other transportation modes to for example assess truck fuel efficiency or driving times between certain routes at certain times of the day to see the influence of road congestions or infrastructure on speed.

Practical implications