Identification of Basic Conditions and Disturbances for Offshore Wind Turbine Maintenance

(1)

Identification of Basic

Conditions and

Disturbances for Offshore

Wind Turbine

Maintenance

DD-MSc Technology & Operations Management

26th December 2016 Student: B.A. Vree Egberts

(2)

Abstract

(3)

Master’s Thesis:

MSc Technology and Operations Management, University of Groningen MSc Operations and Supply Chain Management, Newcastle University EBM028A30 & NBS8399

B.A. Vree Egberts

Hoofdstraat 54a, 7981AB Diever, the Netherlands

b.a.vree.egberts@student.rug.nl / b.vree.egberts@gmail.com Tel: +31(0)6 21943361

student number RUG: 2357852 student number NUBS: 150636102

(4)

Acknowledgement

The completion of the Master’s thesis would not have been possible without the assistance of some people, whom I would like to thank. Firstly, I wish to express a sincere thank you to the men and women who agreed to participate in my study on offshore wind turbine maintenance. With their help I have been able to collect the information required to conduct this research. Furthermore, I would like to thank my supervisors Dr. E. Ursavas from University of Groningen and Dr. J. Dong from Newcastle University for their helpful and valuable discussions and feedback throughout this thesis project. Also, I am grateful to Albert Schrotenboer, MSc for his support throughout the whole project. Finally, I would like to thank my family and friends for their support in completion of my studies and the project in particular.

(5)

List of Figures

1 Logistics network for offshore wind farms (GL Garrad Hassan, 2013b) 3

2 Example maintenance route and schedule . . . 4

3 Conceptual framework . . . 9

4 Empirical wind speed time series 2012/2013 . . . 23

5 Sampling monthly average . . . 25

6 Dutch offshore wind farms in the North Sea . . . 27

7 Cheapest feasible insertion per instance per planning horizon . . . . 33

8 Average cost per scenario . . . 36

9 Average cost per scenario . . . 38

B1 Boxplot of empirical wind speed data . . . 52

B2 Fitted deseasonalised wind speed data CDF and PDF . . . 52

B3 Boxplot of simulated wind speed data . . . 53

B4 Correlation between wind speed and wave height (R=0.69) . . . 53

(8)

List of Tables

1 Maintenance routing and scheduling for OWFs literature overview . 6

2 List of consulted experts . . . 10

3 Quality criteria . . . 12

4 Statistical characteristics of simulated and actual wind speed data . 24 5 Specification of the vessels used in the generated data . . . 27

6 Weather window for each vessel . . . 28

7 Data for dataset BC1, instance 1 . . . 29

8 Specification of the vessels used in scenario 3 . . . 30

9 BC1 outputs (planning horizon 3 days, 24 jobs) . . . 32

10 BC2 outputs (planning horizon 7 days, 45 jobs) . . . 32

11 Proportion distribution per cost component for each scenario . . . . 34

12 Cost savings different scenarios (3 day planning horizon) . . . 34

13 Cost savings different scenarios (7 day planning horizon) . . . 37

14 Detailed scheduling and routing in BC1, instance 1 . . . 42

(9)

List of Abbreviations

CVRP Capacity restricted Vehicle Routing Problem

CTV Crew Transfer Vessel

DVRP Dynamic Vehicle Routing Problem

FSV Field Support Vessel

MLE Maximum Likelihood Estimation

MMSE Minimum Mean Squared Error

MSE Mean Squared Error

O&M Operations and Maintenance

OWF Offshore Wind Farm

PDPTW Pick-up and Delivery Problem with Time Windows SCADA Supervisory Control and Data Acquisition

SOV Service Operation Vessel

VRP Vehicle Routing Problem

(10)

1 Introduction

Renewable energy is playing an increasing central role in the provision of energy services to people globally (REN21, 2015). According to UNEP (2015), the total investment in renewable energy increased six-fold during the last decade. Wind power is among the fastest growing energy generation systems in the world (Kaiser and Snyder, 2013) and has become the lowest-cost option for new power generating capacity in an increasing number of markets (REN21, 2015). The availability of large areas in which to locate major projects, the higher wind speeds and lower turbulence levels in the offshore environment encouraged operators to invest in offshore wind farms (OWFs) (Dalgic et al., 2015). However, the installation, oper-ation and maintenance costs are much higher for OWFs compared to onshore wind farms, since more resources are needed to install and maintain a wind turbine at sea (Irawan et al., 2017). Besides this, due to the harshness of the marine environ-ment and the rapid changes in weather conditions, offshore wind power capacity has reduced levels of system availability (Shafiee, 2015b). Availability of onshore wind farms is typically between 95-99%, as opposed to 60-70% for OWFs (Shafiee, 2015a).

To make offshore wind power cost-competitive with other sources of renewable energy, the reliability, availability and maintainability of OWFs need to be signi-ficantly improved (Shafiee, 2015b). One way to increase availability and preserve profitability of offshore wind energy is to reduce the costs of operations and main-tenance (O&M). The O&M costs of a typical offshore wind turbine with a twenty-year lifetime account for 20-35% of the lifetime power generation cost (Ortegon et al., 2013). One way to reduce O&M costs is to make maintenance activities more efficient by optimising maintenance schedules and the routing of vessels (Ir-awan et al., 2017). This becomes increasingly important as OWFs become larger and are constructed farther from shore, increasing the vessel travel times to and within OWFs.

(11)

optimised to exploit these weather windows, especially as good weather periods are limited.

Research that combines routing of maintenance vessels with scheduling of mainten-ance tasks at OWFs is limited. There are a few models that have been published in the last few years. However, one of the major limitations of the several pro-posed models is that they do not take stochastic weather conditions into account. According to Irawan et al. (2017), a major challenge for maintenance routing and scheduling models for application as an operational decision support tool for a real OWF is the uncertainty and variability associated with weather conditions. However, there is little to no knowledge on this subject. Furthermore, basic con-ditions for offshore wind turbine maintenance have not been sufficiently identified. Identifying the true parameters of these basic conditions and the disturbances of supply chains for offshore wind turbine maintenance will help optimise mainten-ance schedule modelling.

Thus, the main research question is: What are the basic conditions and disturb-ances of supply chains for offshore wind turbine maintenance? These phenomena can be explored by interviewing industry professionals (Karlsson, 2009). Based on the results of the interviews, input data can be generated that tries to repres-ent the O&M of typical, currrepres-ent OWFs as accurately as possible. Idrepres-entified basic conditions and disturbances can be put into practice via a newly developed tech-nician allocation and routing model (Schrotenboer et al., 2017). This multi-period service planning and routing model aims to find the routes for a fleet of vehicles. One addition of this thesis to the current literature is the use of stochastic weather conditions. In this thesis, weather windows have been simulated based on a large dataset containing weather information from over ten years.

(12)

2 Problem description

Figure 1 provides a rough overview of the logistics network for OWFs. Several key players can be identified which play an important role in the logistics network: the OWF owner, the maintenance service provider(s) and the external logistics service provider. Several vessels are used to perform maintenance. First, jack-up vessels are required to replace big turbine components like gearboxes or blades. Second, hotel vessels and service operations vessels (SOV) can be chartered by the maintenance service provider. Both hotel vessels and SOVs stay in the OWF and are used for performing maintenance tasks. Hotel vessels are only used as accom-modation vessels but still need additional vessels to transfer technicians from the vessel to the turbines. However, as the hotel vessel is located in the OWF, offshore transfer times are significantly shortened. The same counts for SOVs which can directly transfer technicians onto turbines via a gangway. Third, field support ves-sels (FSVs) (Dinwoodie et al., 2015) can, for instance, be used to perform seismic, hydro-graphic and oceanographic surveys. Furthermore, helicopters can be used to transfer technicians onto the turbines, which is a costly option. Finally, crew transfer vessels (CTVs), also called work boats, are used to transfer technicians from the O&M base to the turbines. These vessels are used daily and can serve multiple turbines per day. This thesis deals with the routing for CTVs.

(13)

The maintenance routing and scheduling problem consists of several factors that need to be considered. Each OWF has its own set of turbines that need to be main-tained and the maintenance that needs to be performed can be either preventive or corrective. O&M bases are located at the port nearby the OWF. However, offshore accommodation platforms serving one or more wind farms can also be considered as an O&M base. Each onshore O&M base contains resources such as vessels, technicians and warehouses. After the maintenance schedules have been fixed by the maintenance service provider, the following aspects must be addressed (Irawan et al., 2017): the type of vessels that will be used for each period, the op-timal route for each vessel to visit the turbines in the OWFs, and the number of each skill type of technicians required by each vessel for the given maintenance schedule and vessel route.

Figure 2 shows an example of a solution for performing maintenance tasks at an OWF. In this example, four maintenance tasks have to be performed within a specific weather window using one CTV. The CTV delivers the first crew to task A, after which it sails to maintenance task B. After safely transferring the technicians, the CTV travels to maintenance task C. The final maintenance crew is dropped off at maintenance task D and the CTV waits in the field until the maintenance crew at D is done with their task. With the maintenance crew (D) on board, the CTV sails back to turbine C and picks up the next maintenance crew. Next, the CTV picks up the crew at turbine A. The second to last step is to sail to maintenance task B and pick-up the final maintenance crew. Now that all maintenance crews are safely back on the CTV, it sails back to the O&M base, completing the route.

(14)

3 Literature review

This section discusses the maintenance routing and scheduling problem for OWFs. It discusses the available literature on the routing and scheduling of vessels and maintenance tasks. Related vehicle routing problems (VRP) are discussed next to better understand the maintenance and routing scheduling problem. As the routing and scheduling problem for OWFs incorporates the scheduling of techni-cians, workforce scheduling literature is discussed. The section concludes with the research framework used in the remainder of this thesis.

3.1 Maintenance routing and scheduling for offshore wind

farms

(15)

their developed model as an operational decision support tool is limited as it does not take stochastic weather conditions into account. In contrast, Besnard et al. (2013) use historical data regarding environmental conditions at OWFs to present a model for optimising the maintenance support organization of an OWF, but it does not deal with the routing and scheduling of maintenance. Table 1 provides an overview of the current literature on maintenance routing and scheduling for OWFs and their limitations.

Table 1: Maintenance routing and scheduling for OWFs literature overview

Article Study purpose Limitations

Dai et al. (2015)

This article introduces the routing and scheduling problem of a maintenance fleet for OWFs, which is to determine the optimal assignment of turbines and routes to the vessels in terms of costs.

Only small instances which have 4, 6, 8 offshore wind turbines in need for maintenance are tested in the

computational study.

Stålhane et al. (2015)

This paper studies the problem of finding the optimal routes and schedules for a fleet of vessels that are to perform maintenance tasks at an OWF.

Only one period for performing maintenance is considered in this article.

Irawan et al. (2017)

An optimisation model and a solution method for

maintenance routing and scheduling at OWFs are proposed in this study. The routes have to consider several constraints such as

weather windows, the availability of vessels, and the number of technicians available at the O&M base.

One major challenge for application as an operational decision support tool for a real OWF is the uncertainty and variability associated with weather conditions.

3.2

3.3 Workforce scheduling

As offshore wind turbine maintenance routing and scheduling incorporates the scheduling of technicians, workforce scheduling literature is discussed. Workforce scheduling aims to match employee shift arrangements to a time-fluctuating cus-tomer demand for service while satisfying all applicable regulations (Castillo et al., 2009). Previous studies have identified operational problems of scheduling work jobs that are to be processed by a set of workers (Goel and Meisel, 2013). The objective of these studies is usually to minimise the number of required workers and their travel effort or to maximise the number of tasks that can be processed within given time windows (Dohn et al., 2009; Li et al., 2005). According to Goel and Meisel (2013), these approaches often support different worker skills that must be considered.

(17)

Techawiboon-wong et al., 2006).

In the current OWF maintenance literature, technicians are classified on a categor-ical basis. For instance, electrcategor-ical, mechancategor-ical and electromechancategor-ical technicians are used in current literature to perform maintenance tasks. In case of categorical skills, there is no difference in skill level and the skills of a worker determine which tasks that worker can perform (De Bruecker et al., 2015). The skills of one person are not better or worse than the skills of another person. As each maintenance task in offshore wind turbine maintenance requires a different set of technicians, the workforce scheduling becomes more complex.

3.4 Conceptual framework

Figure 3 provides the conceptual framework of the OWF maintenance and rout-ing schedulrout-ing problem. It is based on the work of Irawan et al. (2017) and the framework is suggested to organise and approach the problem by providing an infrastructure for the research project. It is also used to identify basic conditions and disturbance in the offshore wind turbine maintenance industry. It consists of main inputs, constraints and outputs for the routing and scheduling process of maintenance in OWFs.

The main inputs required include the set of turbines that need to be maintained during the planning horizon (Irawan et al., 2017). Each maintenance task has sev-eral parameters: the maintenance/repair time, the number of technicians required (for each type of technician e.g. electrician, mechanical, and electromechanical), the availability and weight of spare parts needed, the recommended (last) period of maintenance and a penalty cost. Information is needed regarding the presence of the vessels during the maintenance operation on a turbine. If that is not the case, the vessel delivers the technicians and picks them up within the same time window after the maintenance activity has been completed. Moreover, information regarding vessel travelling (based on distance, fuel cost and speed of the vessel) and technician costs are required.

(18)

Figure 3: Conceptual framework

(19)

4 Methodology

The purpose of this thesis is to discover the basic conditions and disturbances in the offshore wind turbine maintenance supply chain. The key focus is the routing and scheduling of the CTVs and the maintenance activities that need to be performed. The conceptual framework (Figure 3) is a starting point to explore the basic conditions and disturbances. This thesis identifies the basic conditions and disturbances of offshore wind turbine maintenance and puts them into practice via a multi-period service planning and routing model. Therefore, a part of this thesis is exploratory and uses a qualitative approach. Another part is quantitative, through which knowledge derived from interviews will be put into practice.

4.1 Interviews

According to Karlsson (2009), interviews are particularly effective for exploratory research where the phenomena are not understood and there are unknown vari-ables. The interviewing method used is semi-structured interviewing. According to Fylan (2005), semi-structured interviews can be used when the topics to be covered are known but the conversation has to be flexible since the structure will likely be different between subjects. To gather as much information as possible, a variety of experts in offshore wind turbine maintenance were interviewed. The experts selected had a minimum of two years of experience in the wind industry. Table 2 lists the consulted experts.

The interviews were conducted in person at the interviewee’s office to provide a nat-ural setting for them and to make in-depth questioning possible. Due to research constraints , only one researcher could be present during the interviews. To ensure that no relevant information was left undocumented, the interviews were recorded

Table 2: List of consulted experts

Interviewee Organisation Organisation type Position

A Wind farm owner/

management

B Logistics service provider

C Independent (BoP) maintenance

provider

D Turbine OEM/maintenance

service provider

E Wind farm owner/

(20)

and transcribed. This secured high quality interview transcripts (Karlsson, 2009). However, this was only done with the permission of the interviewee. Transcription was performed within seventy-two hours and transcripts were provided to the in-terviewees to review and revise the transcripts if necessary. The interviews lasted between 1 and 1,5 hours and were conducted between August and mid-November 2016. According to Karlsson (2009), an underlying principle of data collection in case research is triangulation, that is, combining different methods to study the same phenomena. Therefore, other data sources, such as company reports, a con-ference visit and academic research were used to increase data triangulation. This enhances the construct validity of this thesis (Voss et al., 2002).

The reviewed and, if necessary, revised transcripts were analysed. The transcripts were coded by means of a coding tree. To reduce the data retrieved, deductive and inductive coding was applied. Statements from interviews and excerpts from re-ports and academic literature were summarised using first order descriptive codes. The full list of descriptive codes can be found in Appendix A. Throughout this data analysis, the supportive software ATLAS.ti was used for coding and creating data overviews. The gathered data was validated by assessing the quality cri-teria for a qualitative research (Karlsson, 2009) in order to ensure the quality of the conclusions drawn from it. Table 3 shows the relevant quality criteria of the validity aspects. First, concerning reliability, an interview protocol was used in order to ensure that each interview was conducted in the same way. Next, state-ments made by the interviewees were matched in order to obtain internal validity. Furthermore, empirical and theoretical findings were compared to each other to indicate how variables impact the maintenance routing and scheduling. The ex-ternal validity is ensured by selecting experts on their dominant knowledge and experience area which allows to compare statements from practitioners.

4.2 Model-based research

(21)

Table 3: Quality criteria

Quality aspect Measures

Reliability Application of an interview protocol;

Interviews were recorded and transcribed if permission was obtained; Review and revision possibility for interviewees.

Internal Pattern matching of statements on the same offshore wind turbine maintenance elements;

Comparing empirical and theoretical findings.

External Experts selected on their dominant knowledge and expertise area; Comparing of statements from practitioners on the same offshore wind turbine maintenance elements across different settings (e.g. different OWF’s). Construct Data triangulation by interviewing multiple experts on offshore wind

turbine maintenace elements;

Method triangulation by using multiple sources of information (literature, reports, interviews).

windows based on wind speeds and significant wave heights.

(22)

5 Results

Based on the conceptual framework (Figure 3) and following the collection of qualitative data, this chapter presents results on uncertainties in the offshore wind turbine maintenance supply chain. Basic conditions and disturbances of the off-shore wind turbine maintenance supply chain are introduced and implications for the to be developed models are discussed.

5.1 Basic conditions of offshore wind turbine maintenance

Several basic conditions for the supply chain of offshore wind turbine maintenance have been identified. The composition of maintenance teams is an important factor in addition to the CTV constraints. Besides this, maintenance task costs play a role in daily operations.

5.1.1 Maintenance teams

As mentioned in Section 3.3, it is assumed that categorical skills can be found in offshore wind turbine maintenance. Three types of technicians were identified by literature: electrical, mechanical, and electromechanical. However, interviewee D argued that "... electrical, mechanical and hydraulic technicians work on a tur-bine... an electrical technician is a specialist and a mechanical technician is a specialist. However, an electrical technician will never become a mechanical tech-nician, and a mechanical technician will never become an electrical technician." Interviewee E agreed. Therefore, it is impossible to say that electromechanical technicians work on a turbine. According to D "... they (electromechanical) ex-ist, but are almost no where to be found". Besides this, all interviewees agreed that technicians are indicated by their expertise and certificates. For each type of maintenance activity, different certificates are required and in most cases a train-ing matrix was applied (C,D,E). D stressed that:

"We qualify personnel based on experience and training. ...after basic training the technician has reached level D. At level C, a technician will get to know the turbine in a broader sense, but all turbine specific. If you are really skilled, we send you on level B training, which is really troubleshooting. We certify our employees based on such kind of things ... and this of course is a constraint for the supervisor who puts all the teams together."

(23)

Bruecker et al., 2015). This is stressed by D who indicates that "the technician with the highest level is also the leader and ensures that the turbine is locked down and that tasks are distributed". According to D "the cost for electrical technicians is overall the same as for mechanical technicians, they only cannot perform the same tasks." However, a distinction needs to be made within technician type pools. D stated that "... the most experienced and qualified employee earns the most" and "technicians are given salary according to their expertise and certificates."

(24)

they work either week by week or two weeks"(E).

The scheduling problem becomes more complex as schedulers are restricted to the maximum offshore time of technicians. If technicians stay offshore longer than the restrictions permit, the maintenance service provider risks paying a penalty. This could be imposed either by the OWF owner or a governmental institution. Concluding remarks. Maintenance team configuration is more complex than presented in the literature so far. In contrast to Irawan et al. (2017), the inter-views identified hierarchical skill levels in the technician pools. A maintenance team is restricted to several protocols, which can only be initiated by a technician with the proper experience and certificates. Schedulers cannot randomly select technicians from a technician pool to form a maintenance team. Additionally, Ir-awan et al. (2017) indicate that every maintenance task requires technicians with different skills. However, the interviews found that the maintenance team is formed depending on the maintenance task and its priority. The team can be formed with multiple technician types or, as in the case of E, the maintenance team consists of one technician type. Furthermore, the cost structure presented in Dinwoodie et al. (2015) and Irawan et al. (2017) does not reflect reality. In real life, the cost structure is different within technician pools; a less skilled worker costs less than a skilled worker. Finally, the working shift hours found in the literature, (Dinwoodie et al., 2015; Irawan et al., 2017), are in line with the present findings.

5.1.2 Crew Transfer Vessel constraints

(25)

parts. Interviewee B, an external logistics provider, indicated that a CTV has a load capacity of 4 to 24 tonnes depending on its size. CTVs with a higher load capacity are usually larger and more expensive, as they use more fuel and have a higher day rate (B). CTVs can transfer a wide variety of spare parts and mainten-ance tools. Additionally, it was noted by B and E that the maximum spare part weight for small maintenance tasks is limited by the David’s crane available on the transfer platform. This crane is suitable to lift the daily tool bags (B,D,E) and smaller spare parts. If multiple lifting operations are needed, the transfer time for technicians and tools increases (D).

Concluding remarks. The personnel capacity of CTVs discussed above is in line with the available capacity limits in the literature (Dai et al., 2015; Dinwoodie et al., 2015; Irawan et al., 2017). On average, a CTV can hold up to twelve tech-nicians, or four maintenance teams. However, no consensus can be found in the literature regarding the extent to which CTVs can transfer spare parts. For in-stance, Irawan et al. (2017) defines four different CTVs with load capacities varying from 1.5 to 26 tonnes. On the other hand, Dai et al. (2015) lists two CTVs with a capacity of 1.5 and 0.5 tonnes. Dalgic et al. (2015) and Dinwoodie et al. (2015) do not use a load capacity at all. Interviewees indicated that that the load capacity of a CTV is between 4 and 24 tonnes.

5.1.3 Costs of performing offshore wind turbine maintenance

The interviewees mentioned several parameters that affect the cost of performing offshore maintenance: technician costs, vessel costs and maintenance costs (spare parts cost). D stated that "... the technician cost for one OWF consisting of eleven technicians is 1.2 million Euro per year. This includes training for all the technicians." According to C, "if you only use CTVs, the cost level of technicians is very important, because on such a small boat the impact is way bigger than on a bigger boat." D agreed with this as on a daily basis, personnel cost is the biggest cost component. On the long term, however, "... the biggest cost component is the expenditures on main components (gear boxes, etc.) and after that the salary component. The third biggest cost component is that of vessels"(D).

(26)

day rate, the only influencing factor on minimising the maintenance cost would be the fuel expenses. However, in many cases CTVs with different capabilities and day rates are used (B).

Concluding remarks. The finding that performing maintenance offshore con-sists of several parameters confirms the assumptions made in Dalgic et al. (2015); Dinwoodie et al. (2015); Irawan et al. (2017). Furthermore, the assumption made by Dai et al. (2015) to not consider the cost of technicians on the service vessels is misplaced. On a small vessel, the impact of technician costs is large. If larger vessels would be considered, the impact of technician cost would decrease. Addi-tionally, it is important to distinguish between fixed and variable CTV costs. A day rate has to be paid and for every litre of fuel used an extra payment is needed. A fixed cost is already taken into account by Dinwoodie et al. (2015), however, they neglect the variable fuel cost. On the other hand, Irawan et al. (2017) base the cost of a vessel solely on the time offshore and the fuel cost parameter. It is important to combine both parameters to route vessels as economically as possible.

5.2 Disturbances in offshore wind turbine maintenance

Several stochastic, disturbances can be identified in the offshore turbine mainten-ance supply chain. This section highlights the influence of weather forecasts, wind farm accessibility and seasonal effects of the weather on maintenance scheduling. Additionally, failure rates of turbines are rather unpredictable and highly influence the availability of OWFs.

5.2.1 Impact of weather

The major disturbance in the supply chain of offshore wind turbine maintenance is the impact of weather on the routing and scheduling of maintenance. All inter-viewees agreed that weather conditions are the most influential factor for offshore wind turbine maintenance. According to B,

(27)

firm. For instance, E uses two weather forecast systems and decides at 1 PM what the weather window for the next day will be. However, the weather condi-tions can change overnight and the vessel master can say in the morning, based on experience, that weather conditions are too bad too take to vessel out (E). Furthermore, B highlighted that "forecasts cannot distinguish between locations in the OWF. The forecasts are not so good that they can say at this side of the OWF maintenance is possible while at the other side it is not." This was confirmed by C, who noted that a margin on top of weather reports and forecasts has to be taken into account. Moreover, some forecasters use weather buoys in the OWF to measure the wave height and get real-time data (B, E). However, B questions the reliability of these buoys. E stresses that the reliability depends on the position of the buoy in the wind farm. If, for instance, the wave buoy is stationed close to a wind turbine, its measurements would be less accurate as the wind turbine influences the nearby waves.

It is hard to say whether weather forecasts are reliable enough to plan maintenance tasks. All interviewees agreed that human experience should be given weight in planning daily maintenance. This is underlined by C: "if we are offshore and the wind comes from the West, the report will say that in two days the weather will be bad. However, we know that that won’t be in two days but tomorrow afternoon. That is from experience." It can happen that a weather report says that a transfer is possible, but in real-time this is not the case (B, E). According to B and C, tech-nicians and vessel crews visit the site every day and will start to recognise weather patterns. By making decisions based on their experience, fuel costs can be saved as vessels will be less likely to visit the OWF in unsuitable weather conditions. Concluding remarks. A weather window defined by a forecasting mechanism is not necessarily the same as weather conditions on-site. Weather forecasts are not accurate enough to forecast weather conditions per turbine location. OWF op-erators depend on the experience of the technicians and sailors to safely perform maintenance tasks. A future challenge is to visualise these weather behaviours by measurement tools on the CTV, a development seen by B.

(28)

to D, the accessibility is also linked to the wind speed: "the wind speed is not allowed to be more than 12 m/s, the maximum significant wave height is set at 1.5 metres and the visibility has to be at least 50 metres." If conditions exceed these limits, the CTV will not leave the port or leave the OWF. B states that their limits are 1.5 metres significant wave height and 15m/s wind speed. This im-plies that thresholds are set per OWF and are based on the local weather patterns. The wind speed limit is set according to the ability to safely transfer tools to a turbine using the David’s crane. Based on these limits, a decision is made in the morning whether the vessel leaves. When a CTV does not leave to service the OWF, weather conditions can be so uncertain that technicians have to be on stand-by. In case of weather condition improvement, the technicians might be sent to the turbine (D). Because of these limits, the weather window can vary from per day. According to A, "The weather window in which a CTV can travel, mainten-ance can be performed and the CTV is back in time varies. It can be 6, 8 or 10 hours. This highly depends on the season of the year."

All the interviewees agreed that local weather conditions have huge effects on the accessibility of turbines. For instance, A indicated that "each turbine has its own coordinates and the weather can allow maintenance to be performed at one turbine while a kilometre next to the turbine maintenance is not possible" and according to C, "at the turbine itself you can only say if it is possible to safely transfer people". Local weather conditions can be influenced by sandbanks in the sea. It is possible that on one side of the sandbank, maintenance is possible for all turbines, while on the other side turbine maintenance is not possible at all (B, C). Besides this, the tides and wind influence the water flow, which influences the accessibility of wind turbines (B). If CTVs are pushed sideways by the water flow while technicians are transferred, safety cannot always be guaranteed. This is in contrast with a water flow from the front of the CTV, as that mostly does not limit turbine accessibility. According to B, the wave period can also be seen as a limiting factor: "On a short wave period, the maximum significant wave height is indeed 1.5. However, on a long period we also transferred people with significant wave heights of 2.5m up to 3m". The wave period is the time between two waves and depends on the local conditions. A shorter wave period indicates that waves of the same magnitude follow each other in rapid succession.

(29)

in line with the identified weather window limits by the interviews, CTVs with a 1.5m and 1.8m significant wave height limit. However, simply assuming weather conditions are good enough to perform maintenance or basing weather windows on significant wave heights is not enough. Dalgic et al. (2015) recognise this and set a maximum operational wind speed limit of 25m/s. This seems reasonable as it is the cut-off speed for many offshore turbines. Nevertheless, this limit is set too high, as a maximum wind speed for lifting of between 12m/s and 15m/s is used by the interviewees.

Besides this, OWF accessibility is highly influenced by local weather conditions and is far more complex than thought so far in the literature. The assumption made by Stålhane et al. (2015) that the weather is good enough that no restric-tions are imposed on the time where maintenance activities can be started is therefore too simple. Irawan et al. (2017) take different weather windows into account, but do not state how these windows are generated. To use their model as an operational decision support tool, the uncertainty and variability associated with weather conditions should be taken into account. This makes it important to model the weather pattern such that weather windows can be calculated with sufficient accuracy.

Seasonal effects. A identified a seasonal pattern in performing maintenance: "In July, 100% of the planned maintenance activities can be performed, while in January, up to 100% of the planned maintenance activities cannot be performed." However, some nuances have to be placed on this statement, as according to B, "on average, the winters are bad, but sometimes the weather in the winter is good. Of course it can be seen that you’re more days on the sea in the summer than in the winter." D mentions that the overall maintenance schedule is adjusted according to the seasonal effects: "... our season is from April to September." E confirmed this: "you can say you have the vessel season between March and October. Let’s say the CTV season in that respect that you have fewer weather downtimes and that you have to stay in port because of the weather conditions. And from the middle to end of October until the beginning of March you have a lot of weather days where the CTVs have to stay in port." Therefore, if the scheduling and rout-ing of maintenance vessels is executed in the winter, the model should take the corresponding weather windows into account.

(30)

5.2.2 Failure rates and maintenance tasks

D indicated that the failure rates of turbines are unpredictable: "... we visit older turbines around 3 times per year unplanned." However, this is an estimation and E added that so called ’Monday turbines’ exist, turbines with higher failure rates than other turbines. Due to systems like SCADA (supervisory control and data acquis-ition), technicians already know in which part of the turbine the failure occurred. According to D, "An alarm code is generated with x amount of options. Based on those options, the technicians get a tool kit that covers all the failure modes. So we basically do a trial and error process." Experience determines approximately how much time the maintenance will take. However, as troubleshooting has to be done on-site, sometimes the maintenance takes more time than expected. D indicated that troubleshooting, especially in the electrical systems, is rather dif-ficult and is performed by the best technicians. Besides this, the experience of the maintenance team highly influences the service times (D, E). Less experienced maintenance teams are more likely to require more time than a more experienced team to perform the same maintenance task.

(31)

6 Scientific model inputs

As discussed before, this thesis will put the identified basic conditions and disturb-ances for offshore wind turbine maintenance into practice via multi-period service planning and routing model. This section discusses the weather subsystem through which weather windows are simulated. Furthermore, it discusses the control and experimental set-up of the model.

6.1 Weather subsystem

Weather patterns are difficult to model as the variation of wind speeds highly depends on the given site and atmospheric conditions (Gavriluta et al., 2012). Besides this, significant wave heights depend on wind conditions and the ocean floor. Therefore, wind speed and wave height must be modelled with sufficient accuracy to calculate the weather windows. This section describes the modelling of weather patterns based on real weather data. The weather data used in this thesis is obtained from the FINO11 _{offshore research platform, which is situated}

approximately 45km off the coast of Germany. According to Dinwoodie et al. (2015), this platform’s data can be considered representative of Central North Sea conditions.

Wind speed modelling. The wind speed distribution is estimated based on data from 2005-2015. The data was measured at 80 metres, the approximate height of a turbine. However, due to the harsh climate and length of the data set there were variations in recording interval and periods where gaps exist in the data. Therefore, gaps were filled using a cubic spline interpolation and the data was processed into hourly resolutions. This method is similar to the method used by Dinwoodie et al. (2015).

Even though there is no visible trend (see Figure 4), there is a seasonal com-ponent within the data. The visualisation of the seasonal comcom-ponent can be found in Figure B1 (Appendix B). During the spring and summer, the likelihood of ex-treme wind speeds decreases, while during autumn and winter it increases. Besides seasonality, a cyclical pattern can be observed. This indicates that the wind speed at time t is related to the wind speed at time t − 1. Therefore, random sampling from the data is impossible. To correctly model the wind speed, the data first must be broken down into its components. The first step in doing this is calculating the monthly seasonal wind speed factor St and remove it from the data by means of seasonal indices. To calculate a seasonal index, the average of a particular season

(32)

Figure 4: Empirical wind speed time series 2012/2013

is divided by the average of the entire time series. The next step is to divide each data point by its respective seasonal index, which removes the seasonal factor from the data. This is common practice in time series forecasting Chopra and Meindl (2010).

After removing seasonality from the time series, the data is fitted to a Weibull distribution. Wind speeds are often characterised by Weibull distributions (Karki et al., 2006), and methods using Weibull distributions have previously been used to estimate wind speed data in wind power studies (Abouzahr and Ramakumar, 1991). The Weibull distribution is fitted using the maximum likelihood estima-tion (MLE) (Fisher, 1922), a common method to calculate the optimal fit between sample data and population distribution (Andrawus, 2008). The Weibull distribu-tion (provided in Appendix B Figure B2) gives a shape parameter α of 2.32 and a scale parameter β of 10.52. To include a cyclical pattern in the simulated wind speeds, an ARMA time-series model was used. An ARMA time-series model has been used by researchers as an accurate method to forecast wind speed data at any particular location (Karki et al., 2006). The wind speed at any given time for a particular geographic location can be simulated by using equation 1

SWt= µt+ σt× yt (1)

where SWtis the simulated wind speed for hour t from historical mean wind speed

µtand standard deviation σt. Time-series values of ytcan be obtained by equation

2

yt= φ1yt−1+ φ2yt−2+ ... + φnyt−n+ αt+ αt−1θ1+ αt−2θ2+ ... + αt−mθm (2)

(33)

mean and a variance of σ2

a (i.e. αt ∈ N ID(0, σa2)), where NID stands for normally

independently distributed. The order [n, m] and the values of the auto-regressive and moving average parameters shown in equation 2 can be obtained from many years of collected wind speed data using non-linear least-square method (Karki and Hu, 2005). In this case, the NumXL add-in for Excel was used to obtain these values using a non-linear least-square method. The hourly mean and standard deviation of wind speeds from 10 years were obtained to determine the ARMA model for the weather subsystem. The respective mean is 9.332 and the standard deviation is 4.243. The parameters of the ARMA time series model can be found in equations 3 and 4

yt= 0.9621yt−1+ αt+ 0.1606αt−1 (3)

αt ∈ N ID(0, 1.05902) (4)

Hourly wind speeds were repeatedly simulated for a period of 10 years. After deseasonalised wind speeds were generated, the monthly seasonal factor was added to ensure a seasonal pattern. This is confirmed by Figure B3 (Appendix B). Table 4 shows the mean and standard deviation values of the 10 years of collected data. The simulated values are close to those obtained from the actual data. Additionally, Figure 5 provides visual confirmation that the ARMA model is an accurate method to simulate wind speed data. However, randomly sampling from any distribution would result in positive and negative values of wind speeds. This also holds for the ARMA time-series modelling (Karki et al., 2006). Negative values have no physical meaning in this thesis and could be set to zero, as recommended in Karki and Hu (2005). This would, however, result in a significant wave height with the value 0. As this value has not been observed in the actual data, it is incorrect to model a wave height of 0. Therefore, a random positive wind speed is generated for simulated wind speeds below 0.

The simulated wind speed is generated at hub level. However, to calculate a weather window at time t, the wind speed needs to be extrapolated to a different level, as the David’s crane is positioned about 50 metres above the sea surface. Because lifting operations are needed to transfer tools from a CTV to the turbine,

Table 4: Statistical characteristics of simulated and actual wind speed data Simulation run Mean value µ Standard deviation σ

Simulated Actual Simulated Actual

Run 1 9.277 9.332 3.991 4.243

Run 2 9.386 9.322 3.962 4.243

(34)

Figure 5: Sampling monthly average

a wind speed limit for lifting is set at 12m/s. This was the lowest identified, via interviews, limit for lifting tools and spare parts. If a wind speed limit would not be set, the lifting operation might cause serious damage to the turbine as spare parts might start to circle in the air. To extrapolate the wind speed to a different level, the power law of Justus and Mikhail (1976) is applied. Equation 5 and 6 provide the mathematical formulation

ws2 ws1 = (h2 h1 )α (5) ws1 = hα1( ws2 h2 )α (6)

The wind speed at hub level h2 is denoted by ws2. An α of 0.1 has been taken to

extrapolate the wind speed to the David’s crane level.

Wave height modelling. Accurate modelling of significant wave height requires historical data on the wind speed, duration of the wind and the fetch (the length of the water of which wind flows). However, a correlation analysis between wind speed and wave height produced a high Pearson correlation of 0.69 (see Figure B4). Therefore, the correlation is assumed to be 1 so that wave height modelling could be based on the wind speed only. Because of this correlation, it is not possible to randomly sample from historical data, as there would be no direct relationship between wind speed and wave height. Therefore, the relationship is presented by the factor β. Besides this, the seasonal factor (Sh

t) is calculated from the historical

data and added to the equation 7 to calculate wave height. SSWt =

SWt

β S

h

(35)

The β is optimised by means of the minimum mean squared error (MMSE) method. Excel Solver is used to minimise the mean squared error (MSE) and in this case it resulted in β = 7.2. Figure B5 provides visual confirmation that the method accurately models significant wave heights. It can be observed that the empirical data has steeper peaks, however, the model is sufficiently accurate with respect to the access criteria of CTVs.

6.2 Experiments on new dataset with multiple OWFs

In this section, newly generated datasets and different scenarios for scenario ana-lysis are presented. The procedure used in this paper is derived from Schrotenboer et al. (2017) in their working paper and the heuristic works as follows: from all the maintenance tasks in a specific instance, a random row is generated. After this, jobs, in the order shown on the random row, are put on the best place in the solution. However, it is possible that the random row of jobs is infeasible and, therefore, the heuristic creates 50,000 solutions. As 50,000 feasible solutions are generated per instance, it is possible to derive the cheapest feasible insertion. The results of the best solutions per instance per scenario are used for further analyses. Note that, it is not claimed that the scenarios and values of the input parameters accurately represent the O&M of a typical OWF. However, the datasets are based on three wind farms close to the Dutch coast (see Figure 6)2 _{and interviews}

con-ducted with people involved in the O&M of several OWFs. The first OWF consists of 60 turbines, the second of 43 turbines and the third of 36 turbines.

6.2.1 Control set-up

A base case was generated in which the three aforementioned OWFs are serviced from one O&M base. Three vessels are available, however, they are limited to ac-cessing one OWF each. This thesis considers three types of technicians: electrical, mechanical and hydraulic with an availability of 3 of each technician per OWF. Unfortunately, it is not possible to incorporate hierarchical skills in the model. Furthermore, two types of datasets, namely Dataset BC1 and Dataset BC2, were constructed. Each dataset consists of 20 instances and the planning horizon is set to 3 and 7 days respectively. For each instance, the number of turbines that need to be maintained is evenly distributed over the OWFs. with 24 turbines for Dataset BC1 and 45 for Dataset BC2. Furthermore, some parameters of the new dataset, like service times and spare part weights, are similar to those in Irawan et al. (2017) and Dai et al. (2015).

(36)

Figure 6: Dutch offshore wind farms in the North Sea

Table 5 presents the specification of the vessels used in the base case. The data provides the O&M base per vessel, personnel capacity, load capacity, significant wave height limit of the vessel and the fuel cost/hour. Like in Irawan et al. (2017), it is assumed that the vessels are available and that the vessels are equipped to transport spare parts to the turbines. Based on the interview findings this is jus-tified as CTVs can carry a wide variety of spare parts.

The weather windows for dataset BC1 and BC2 are given in Table 6 and retrieved from the weather subsystem in section 6.1. A random date within the CTV season is selected from the weather subsystem as a starting point. Note that the weather subsystem can be used to predict future states of weather conditions.

Table 5: Specification of the vessels used in the generated data Vessel OWF Personnel

(37)

Table 6: Weather window for each vessel Vessel Period 1 2 3 1 7 7 7 2 10 10 10 3 10 10 10 4 9 9 9 5 10 10 10 6 10 10 10 7 7 7 7

Table 7 provides an example of instance 1 of the BC1 dataset. It gives the main-tenance time for each turbine that needs to be maintained in each OWF. It also gives the weight of spare parts per maintenance job and a penalty cost per main-tenance task, which has to be paid every day that the mainmain-tenance task is fulfilled after the latest period. The penalty cost and latest period per maintenance task are randomly generated and the values are similar to those of Irawan et al. (2017). Table 7 also shows the number of technicians per technician type for each mainten-ance task. The cost per day, unrelated to the worked hours, ise350 for technician type 1, e300 for technician type 2 and e250 for technician type 3. It is assumed that the vessels do not have to stay at the turbine while maintenance is performed, however, the possibility is available to let a vessel stay at a turbine. Besides this, it is assumed that spare parts are available at the O&M hub and the time to transfer technicians and tools from the vessels onto the turbines is set to 15 minutes (0.25 hours). This accords with the interview results.

6.2.2 Experimental set-up

In the experimental set-up, specific variables and their effects on outcomes of the model are analysed. This means that the experimental cases are similar to the control set-up; the only difference involves the variable that is analysed. Several scenarios have been created to analyse these effects:

(38)

(39)

the OWFs. New O&M hub coordinates have been generated via equation 8,

N CB = CB + (CB − CTn) (8)

where N CB stands for the newly generated coordinate of the O&M base, CB indicates the coordinate of the current location of the O&M base and CB − CTn is the difference between the coordinates for a random turbine

CTn and the base CB. This equation enables positioning the O&M base

such that the distance between the O&M base and OWFs is doubled. Note that equation 8 is used to generate both new X and Y coordinates. As travel times increase, it is expected that less maintenance tasks can be performed within one weather window. This might lead to higher penalty costs, higher fuel costs or even no feasible solution at all.

• Scenario 2. In the base case, three vessels are used to service the three OWFs. In scenario 2, four CTVs will be used to service all the OWFs. The first two vessels serve OWF 1 and 2, the last two serve OWF 2 and 3. It is expected that the penalty cost will decrease while the fuel cost will increase. As CTVs are allowed to serve two OWFs, the technicians are reallocated over the OWFs. Five of each of technician types 1, 2 and 3 are available for CTVs 1 and 2. CTVs 3 and 4 have an availability of 4 of each technician type. • Scenario 3. In scenario 3, the effect of a different fleet composition is

ana-lysed. During interviews held with industry professionals, various CTV types have been identified. These CTVs all have their own specifications, signific-ant wave height limit and cost structure. However, what happens if the significant wave height limit, and consequently the maximum offshore time, increases? The vessels used and their specifications in this scenario can be found in Table 8. As the CTVs can withstand higher significant wave heights, the weather windows are adjusted accordingly. It is expected that the total cost for performing maintenance will decrease, even though the fuel cost/hour is higher. The maximum offshore time for the CTVs is higher than in the base case, which will most likely result in less penalty costs as more vessels can be used to perform maintenance.

Table 8: Specification of the vessels used in scenario 3 Vessel OWF Personnel

(40)

• Scenario 4. In the base case, the maximum offshore time highly depends on the maximum offshore time for the technicians. Interviewee D indicated that the technicians sail out at 7am and leave the OWF at 4pm. This indicates a maximum offshore time of 10 hours, including travelling time from the OWF to the O&M base. In this scenario, the effect of staying 2 more hours in the OWF (if weather conditions allow this) to perform maintenance tasks is analysed. Two more hours offshore is a possible case, as it is done at the OWFs in Germany, during the installation phase of OWFs and as regulations in Germany and the Netherlands allow it. The latter means that the maximum offshore time of 12 hours conforms to the Dutch ’mijnenwet’ discussed in Section 5.1.1.

(41)

7 Computational analysis

In this section, the results of the base case and the five scenarios are presented. First, results of the base case are examined followed by scenario analysis to evaluate the different scenarios as identified in Section 6.2.2. The outputs are based on 50,000 runs per instance to cover many possible solutions. The cheapest solution per instance is used in the remainder of this section.

7.1 Base case analysis

Table 9 provides an overview of the BC1 outputs in which a planning horizon of 3 days is considered with 8 jobs per OWF. Three types of costs are considered: travel, personnel and penalty cost. The total for all the instances, the average for the instances and the proportion of the cost components are given. On average, the personnel cost contributes the highest proportion to the total cost, at around 62%. This is in line with interview findings. The travel and penalty costs con-tribute to the total cost in approximately the same proportion. However, with a longer planning horizon and more jobs to schedule, the difference in proportion between the travel and penalty cost becomes more visible (Table 10). Besides this, the proportion of personnel cost slightly increases. This could be because the proportion of penalty cost decreases as more time is available to schedule the jobs. This allows scheduling jobs such that the total penalty cost decreases per instance. Additionally, the travel cost contributes to the total cost with approximately the same proportion as in BC1.

Table 9: BC1 outputs (planning horizon 3 days, 24 jobs)

BC1 Total cost Travel cost Personnel cost Penalty cost Total 20 instances e676,800.70 e128,500.65 e423,800 e124,500

Average cost e33,840.04 e6,425.03 e21,190 e6,225

Proportion N/A 18.99% 62.62% 18.40%

Table 10: BC2 outputs (planning horizon 7 days, 45 jobs)

BC2 Total cost Travel cost Personnel cost Penalty cost Total 20 instances e1,184,181 e227,581 e776,600 e180,000

Average cost e59,209.05 e11,379.05 e38,830 e9000

(42)

7.2 Scenario analysis

For each scenario, the same job and technician instances are used as in the base case. The only changes are one or several parameters per scenario (see Section 6.2.2). As illustrated by Figure 7, the impact of different scenarios varies per plan-ning horizon. Therefore, this section discusses the outputs per planplan-ning horizon and identifies possible cost savings.

Figure 7: Cheapest feasible insertion per instance per planning horizon

(43)

Table 11: Proportion distribution per cost component for each scenario Proportion

distribution 3 day planning horizon

Scenario BC 1 2 3 4 5

Travel cost 18.99% 21.50% 20.06% 32.44% 20.11% 19.44% Personnel cost 62.62% 55.34% 67.51% 59.39% 61.82% 63.52% Penalty cost 18.40% 23.16% 12.42% 8.17% 18.07% 17.04%

7 day planning horizon

Scenario BC 1 2 3 4 5

Travel cost 19.22% 21.57% 19.22% 31.78% 21.01% 16.15% Personnel cost 65.58% 58.35% 66.73% 60.87% 67.41% 54.67% Penalty cost 15.20% 20.08% 14.05% 7.35% 11.58% 29.19%

Moreover, the proportion for each cost parameter per scenario is approximately the same between the two datasets. This, however, does not hold for scenario 5, where the contribution of penalty costs is higher for the 7 day planning horizon than for the 3 day planning horizon. This also results in different contributions of the personnel cost to the overall cost. Additionally, the distribution of costs also differs in scenario 4 for each planning period. This is mainly because, on average, less penalty costs have to be paid when considering a longer planning horizon. 7.2.1 Planning horizon 3 days

Table 12 provides the cost savings per scenario compared to the base case. A positive value denotes a cost increase and a negative value is a cost saving. As the distinction is made between the three cost components, more in-depth analysis is possible.

Table 12: Cost savings different scenarios (3 day planning horizon)

Scenario 1 2 3 4 5

Total cost 22.51% -5.70% -5.39% -10.61% -2.54% Travel cost 38.75% -0.36% 61.65% -5.30% -0.20% Personnel cost 8.27% 1.68% -10.26% -11.75% -1.14% Penalty cost 54.22% -36.31% -57.99% -12.21% -9.72%

(44)

this scenario, the distance between the onshore O&M base and the OWFs is doubled (see Section 6.2.2). Besides this, the total cost increases by 22.51% compared to the base case (Table 12). It should be noted that these values are obtained from the cost outputs available for scenario 1, twelve instead of twenty. As expected, the travel cost increases with the increase in the travel distance to the OWF. Furthermore, the penalty cost in scenario 1 increases by 54.22%. This is because the total time in the OWF decreases as more time is spent travelling on the CTV. Therefore, it is more difficult to schedule maintenance tasks in their preferable position, which results in higher penalty costs.

• In scenario 2, the effect of using four vessels instead of 3 is analysed. An overall cost saving of -5.70% is reached. Even though more CTVs are used, the travel cost slightly decreases. This could be because not all the CTVs are utilised every period. Besides this, the personnel cost increases as it is likely that more technicians are utilised. However, using four CTVs results in a significant decrease in penalty cost, as most jobs can be scheduled before their final period of maintenance.

• In scenario 3, a different set of CTVs with more offshore time and higher fuel cost parameters is used. Subsequently, the travel cost increases by 61.65%. However, as more time can be spent working on the turbines, this scenario results in a cost saving on penalty costs of -57.99%. Besides this, as it is likely that technicians are not used every day, more maintenance jobs can be scheduled in one maintenance window. Thus, the personnel cost decreases by -10.26% as compared to the base case. This results in an overall cost saving of -5.39%.

• The highest overall cost savings are obtained in scenario 4. In this scenario, technicians are allowed to stay 2 more hours offshore than in the base case, resulting in an overall cost saving of -10.69%. In contrast to the previous scenarios, this scenario shows a cost saving for each cost component.

• The final scenario simulates winter weather conditions and obtains cost sav-ings. This is because in the first three days, the weather conditions were good enough to perform all the maintenance tasks. It can be seen that over-all, a better schedule was obtained for scenario 5. However, the question remains whether this would be true for a longer planning period with more small weather windows.

(45)

This leads to savings on all the cost components taken into consideration. Besides this, as expected, the total cost increases if the distance between the O&M base and the OWFs increases. In particular, more penalty cost and travel cost have to be paid in that case. Furthermore, using four CTVs (scenario 3) or three CTVs that can handle harder weather conditions (scenario 2) also results in overall cost savings. The average costs per scenario for each cost component can be found in Figure 8 . The figure confirms that scenario 4 is the preferred solution while scenario 2 results in the highest average costs.

Figure 8: Average cost per scenario

So far, we only discussed the results of the several scenarios for a shorter planning horizon. The next section discusses and analyses the results of the scheduling of maintenance tasks for a planning horizon of 7 days.

7.2.2 Planning horizon 7 days

Figure 7 indicates that the cost for performing maintenance is higher for scenario 1 and 5. This is confirmed by Table 13. This table provides the cost savings for a 7 day planning horizon per scenario compared to the base case. As before, a positive value denotes a cost increase and a negative value is a cost saving.

(46)

Table 13: Cost savings different scenarios (7 day planning horizon) Scenario 1 2 3 4 5 Total cost 24.24% 1.21% -1.69% -11.17% 25.69% Travel cost 39.44% 1.21% 62.56% -2.90% 5.60% Personnel cost 10.55% 2.99% -8.76% -8.69% 4.78% Penalty cost 64.11% -6.44% -52.44% -32.33% 141.33%

In comparison to the base case, cost increases by 24.24%. First, because the travel cost increases as more distance has to be covered and second, the pen-alty costs increase by 64.11%. As already identified in Section 7.2.1, this is because travelling time is a burden on the total time spent in the OWF. • In the short term, using four instead of three CTVs resulted in a cost saving

of -5.70% for scenario 2 (see Table 12). However, in the long term, using more CTVs results in a slight increase of the total cost. This is because the influence of the decreased penalty cost is not as big as in BC1.

• Additionally, in scenario 3, only a minor improvement in terms of total cost is obtained compared to the base case. Again, the travel costs increase significantly as the CTVs used are more expensive than those in the base case. However, as more time can be spent in the OWF, the penalty cost decreases by -52.44%.

• As already observed for BC1, the preferred solution would also be scenario 4 for BC2. The technicians are allowed to spend 2 more hours offshore which results in a total cost saving of -11.17% in just 7 days.

• Finally, as expected, the total cost for performing maintenance in the winter increases by 25.69%. The winter weather conditions were so extreme that the planning period was set to 12 days. The input parameters, like latest period, penalty cost and maintenance time were kept the same. If the planning horizon would not have been extended, the scenario would be unsolvable as not enough time could be spent in the OWF. This results in an increase in 141.33% for the penalty costs and 5.60% and 4.78% for the travel and personnel costs, respectively.

(47)

influence the outcome of the model, as can be seen in scenario 5, where the overall cost increases by 25.69%. Finally, scenario 1 resulted in slightly different values than those obtained for a planning horizon of 3 days. To visualise the different scenarios, Figure 9 provides the average cost per scenario. This figure confirms that scenarios 1 and 5 result in much higher costs, while scenario 4 is the overall cheapest scenario.

Figure 9: Average cost per scenario

7.2.3 Concluding remarks

As expected, each scenario delivered different results in terms of costs and cost component distributions. For the influence of the planning horizon on the total cost savings, it was found that in a short term planning horizon, scenarios 2, 3 and 4 result in cost savings of up to 10.61%. However, in a long term planning horizon, cost savings for scenario 2 and 3 decrease and become almost irrelevant Therefore, the most likely scenario, for both planning horizons, is scenario 4, in which technicians are allowed to spend 2 more hours offshore.

7.3 Validation

(48)

Several limitations on the input data for the computational analysis were identified in this extra interview. First, it was noted that the maintenance times could be, on average, higher than the current input. For corrective maintenance tasks, the maintenance times reflect real life situations. However, for preventive maintenance tasks, one would expect to spend several complete days visiting one turbine, be-cause it takes two to four days to complete preventive maintenance tasks. Second, the weight of spare parts is high. There are maintenance tasks where the spare parts weigh 1000 kilograms. However, this does not happen often and the average weight of spare parts per maintenance task is around 100 kilograms. Further-more, the penalty cost depends on the contract between the maintenance provider and the OWF owner. For instance, in the first months after installation, the pen-alty costs is different than after two years, when the turbines are running smoothly. Additionally, it was rather difficult for the interviewee to quantify the penalty costs, as was done in the base case and scenarios. The interviewee indicated that one would expect that the costs of downtime for each turbine would be constant for each hour, however, it would depend on the wind conditions if an energy based contract is in place. Regarding the weather windows, the interviewee indicated that one should expect days where no weather window is available at all, that is, a weather window of 0. This is the case even in summer. However, the inter-viewee noted that this depends on the wind farm sites taken into consideration and the amount of historical data. Finally, as in the current model the fuel cost is multiplied with the time offshore, a rather high total fuel cost per day arises. Interviewee D stressed that on average, for an OWF half an hour from shore, the fuel cost would be e500 per day. This depends on the total kilometres travelled by the CTV, which implies that a cost parameter should be used that multiplies the amount of kilometres travelled with the price per kilometre.

Identification of Basic Conditions and Disturbances for Offshore Wind Turbine Maintenance