Automated wind turbine maintenance scheduling

Automated Wind Turbine Maintenance Scheduling Nurseda Y. Y¨ur¨u¸sena, Paul N. Rowleyb, Simon J. Watsonc, Julio Meleroa,∗ aInstituto Universitario de Investigaci´on CIRCE (Universidad de Zaragoza - Fundaci´on CIRCE), C/ Mariano Esquillor 15, 50018, Zaragoza, Spain bCREST, Loughborough University, Holywell Park, Loughborough, LE113TU, UK cDUWIND, Delft University of Technology, Kluyverweg 1, 2629 HS Delft, Netherlands Abstract While many operation and maintenance (O&M) decision support systems (DSS) have been already proposed, a serious research need still exists for wind farm O&M scheduling. O&M planning is a challenging task, as maintenance teams must follow specific procedures when performing their service, which requires working at height in adverse weather conditions. Here, an automated maintenance programming framework is proposed based on real case studies considering available wind speed and wind gust data. The methodology proposed consists on finding the optimal intervention time and the most effective execution order for maintenance tasks and was built on information from regular maintenance visit tasks and a corrective maintenance visit. The objective is to find possible schedules where all work orders can be performed without breaks, and to find out when to start in order to minimise revenue losses (i.e. doing maintenance when there is least wind). For the DSS, routine maintenance tasks are grouped using the findings of an agglomerative nesting analysis. Then, the task execution windows are searched within pre-planned maintenance day. Keywords:Wind Turbine, O&M, Maintenance, Scheduling 1. Introduction The cost of maintenance is a major contributor to the total levelized cost of energy (LCOE) from wind farms accounting for a share of around ∗Corresponding author: [email protected] Preprint submitted to Reliability Engineering & System Safety March 20, 2020 20%-25% [1]. Minimisation of the maintenance costs requires as precise as possible maintenance scheduling. Extended downtime, which can occur when maintenance interventions are delayed during poor weather, incurs financial costs for wind farm owners. Wind farm operational scheduling has been found in the literature to be a function of a range of factors such as the energy demand [2], electricity market price and wind speed [3]. When the constraints are investigated, wind farm accessibility normally comes in first place in terms of importance. Farm accessibility depends on the variability of the wind speed and the associated health-safety and environment regulations (HSE) [4–9]. In other words, finding an appropriate weather window is a major criterion for any type of intervention for wind turbines and there is a research need closing the gap between academic models and application in practice [10]. According to the literature, maintenance weather windows are dependent on the wind speed for an onshore wind farm, while the wave height is also a decisive factor for offshore wind farms where accessibility depends on the type of maintenance vessel utilized [11–14]. For offshore wind farm operations, the location (distance to shore and water depth), meteorological and oceanographic variables influence the site accessibility, it is highlighted in the literature that there is a trend of moving from near-shore to deep water for offshore wind farm installations, which results in lower site accessibility and higher costs for the executions of corrective maintenance actions [15]. In addition to the measured mean wind speed, industry practice highlights wind gust as an important parameter when considering access to a wind turbine [16]1. However, thus far, this has not been referred to in the literature concerning scheduling studies as a constraint which affects either operational scheduling or downtime. Previous academic work in this field has considered only wind speed and output power as the decisive parameters when generating a feasible maintenance plan in onshore [18, 19] and significant wave height, wave peak period and wind speed in offshore [7, 15]. It is already noted in both onshore and offshore crane manuals and safe working guides that working height and wind gust speed influence executions of crane operations [20, 21]. When a crane operation cannot be performed, the cor1In this study, two major wind turbine manufacturers’ O&M guidelines are used. These two companies are also leading original equipment manufacturers in the wind sector. According to 2017 statistics, the original equipment manufacturers of wind turbine have the highest market share among the wind farm O&M service providers [17]. 2 responding wind turbine maintenance action can also not be performed as well and results in delay for O&M actions. This delay contributes to weather related downtime, which is resulting from coarse maintenance planning and insufficient accessibility of both wind farm site and wind turbine component. In the present study, the authors consider both mean wind speed and wind gust as limiting factors for accessibility to an onshore wind turbine and demonstrate the applications of wind speed measurements in determining task execution sequence whilst minimising downtime due to adverse weather conditions during periods of intended maintenance. The goal is automated scheduling of tasks to be performed within a workday, such that tasks with strict requirements are scheduled when conditions are most benign. The normal practice depends on two weeks ahead maintenance service team booking with a single call entailing which alarm is activated for which turbine. These work orders are lacking detailed planning of the maintenance day and the tasks to be performed. Therefore, there are coarse planning and weather related waiting periods in the wind farm site. The structure of this paper is as follows: in the following section, an overview of a typical wind farm maintenance policy is described. General characteristics of mean wind speed, wind gust, maintenance log books and task completion duration data are presented in Section 3. The next section describes the methodology, and the proposed framework. Case-studies are then presented to show the value of a proposed maintenance planning methodology, which demonstrates how an optimal sequence for maintenance interventions can be devised. Then, in Section 6, practical explanation of the findings, the limitations and the assumptions are presented. The final section summarises the main outcomes of this study. 2. Maintenance plans & problem statement Wind turbine maintenance can consist of both corrective and preventive actions. Long term maintenance policies must cover both of these. Corrective maintenance is normally carried out once a fault has been detected, whereas, preventive maintenance is generally performed according to calendar-based pre-determined intervals such as biannual, annual, biennial and quinquennial periods [4, 5, 22–24]. The number of tasks and the duration of a scheduled maintenance action are different from one case to another and depend on specific sub-assembly, components, manufacturer, model and capacity of the wind turbine. Scientific literature and manufacturers’ maintenance guides 3 give figures for the required duration of a range of maintenance tasks that vary from lubrication which typically takes few hours to other more lengthy which last up to 18 hours [23, 24] during a biannual maintenance visit. In addition, working practices may differ from one operator to another. These factors must be taken into account in terms of defining a comprehensive maintenance strategy and provide a challenge when developing a model for maintenance optimisation. (a) 4 (b) Figure 1: Maintenance scheduling procedure (a) preventative policy and (b) corrective intervention. Preventative maintenance is usually planned a year in advance on an annual basis for onshore wind farms [25]. A typical flow diagram of this type of advanced planning is shown in Figure 1a where account needs to be taken of the weather and electricity market prices as well as the availability of a maintenance team [16]. Requirements of preventative maintenance can also be seen even when planning corrective actions as shown in Figure 1b. Both, regular and corrective maintenance involve uncertainties, particularly concerning the weather related limitations. The typical limiting factor for executing maintenance actions is the wind speed. Regulations and manufacturers’ good practices set the maximum values of the wind speed which allow work at different locations on the turbine. This information has been used in previous research works to develop maintenance frameworks. For example, one study fixed the wind speed limit as 10 m/s for accessing the whole turbine [5], while another based the safe working limit on cut in wind speed, i.e., the turbine was only considered maintainable when the wind speed was lower than cut-in [6]. Furthermore, current regulations and maintenance guides include dynamic safety limits taking into account not only the mean 10-minute wind speed value but also the gust value, when a crane usage is required for such a case like major component replacement. The definition of gust is a short-duration (seconds) maximum of the fluctuating wind speed [26]. The maximum permissible wind gust speed for crane usage depends on 5 various factors such as mean wind speed, intervention height and weight of the load [20]. Therefore, corresponding wind gust restriction for any intervention requires timely and case based controls. Moreover, high gust values cause more restrictive wind turbine component specific accessibility rules reducing the highest allowed mean wind speed. Taking into account only wind speed limits, the safe working rules are also different depending on turbine model and size. For example, in the case of MADE AE 46 turbines, preventative maintenance requires wind speeds below 20 m/s at the nacelle, however changing the whole nacelle requires the wind speed to be not more than 5 m/s. If we check the requirements for NEG Micon NM 52 turbines, working in the hub requires wind speeds below 15 m/s while working in the nacelle roof is allowable until 12 m/s and generator alignment should not be performed for wind speeds above 10 m/s. Finally, for the Vestas V 90 3.0 MW model, generator alignment intervention can’t be done for wind speeds above 8 m/s, changing pitch angle requires wind speed values smaller than 6 m/s and working in the drive train is allowed up to 7 m/s [16]. Within a work shift, various tasks must be completed on a wind turbine according to the prevailing time and labour force restrictions. As stated in [11], it is almost impossible to generate a flawless maintenance plan in terms of avoiding production loss, since it is difficult to find a period where the turbine is not producing due to low wind speeds. What can be done in this sense is to schedule the maintenance with an acceptable uncertainty [27, 28]. 3. Data source, wind farm maintenance procedure and data Maintenance logs, service work orders and SCADA data were obtained from a Spanish wind farm. In this analysis, O&M service reports, which cover a 3 years window, are used to define the list of actions and the needed duration for each type of intervention and activity in the studied wind farm. Regarding the meteorological data, 10-minute wind speed and wind gust data are collected for year 2019. In the final analysis, accessibility investigations for 24 hours windows are provided for the example cases. According to the information gathered, the average duration of the biannual, annual, biennial and quinquennial visits are approximately 21, 26, 15 and 18 hours respectively. The total number of different tasks to be performed in maintenance visits is 169. Most of them, 117, are included in the biannual visit actions while the others are distributed over the other visits. However, not all main6 tenance actions are carried out during each planned visit as some of them have priority based on the findings of previous service visits and the needs of the wind turbine. Table 1: Executed tasks for the scheduled visit Turb. Work Zone Sub System Task Numbers A-Ground Tower 1 to 2 A-Ground Electrical Parts 3 A-Ground Rotor-Blades 4 B-Platform Electrical Parts 5 to 7 C-Tower Yaw System 8 to 14 D-Nacelle Main Shaft and Bearing 15 to 17 D-Nacelle Gearbox 18 to 27 D-Nacelle Generator 28 D-Nacelle Base Structure and Cover 29 to 31 D-Nacelle Electrical Parts 32 E-Hub Rotor 33 to 34 F-Outside of Nacelle Sensors 35 to 36 Figure 2: Example of turbine working zones Figure 2 shows the considered turbine working zones, while the task numbers associated to these zones are listed in Table 1. In this work, for the sake 7 of simplicity, only the tasks numbers listed in Table 1 are used to define a case study considering a regular service visit. A second case study is based on a major intervention, which requires a crane usage. More specifically, a generator replacement is studied and more information will be provided regarding the corresponding task. To explain the working environment of the service personal for performing either a regular service or a major intervention, the seasonal and general characteristics of the subject wind farm are shown in Figures 3 and 4. In Figure 3, the wind speed seasonal histograms from the case Spanish wind farm are presented. The annual histogram is included in each graphic to highlight the seasonal contribution. It can be seen that the majority of wind speed observations lie between 0 and 10 m/s in summer months. Then, summer looks the best season for maintenance actions, but there are still a significant number of wind speed observations with values higher than 15 m/s. Figure 3: Annual versus seasonal wind speed histograms using 10-minute averaged mean wind speeds Figure 3 shows the seasonal characteristics of the nacelle wind speed obtained from the analysed wind farm. It is known that the seasonal wind 8 speed behaviour is dependent on the location of the wind farm. The annual maintenance plan must be prepared considering the seasonal wind behaviour and the electricity market prices of the country where the analysed wind farm is located. Then, the seasonal wind behaviour is an important factor for long term scheduling, which is not the aim of this study. The resulting program from the annual maintenance plan is an input to decision making support tool. Therefore, this input must be modified, when the analysed wind farm is changed. Figure 4 illustrates the diurnal behaviour of the wind speed for each season during 2019 comparing the maximums recorded in hourly data per seasons. It can be seen that the day shift (08:00 to 18:00) in summer, with wind speed maximums lower than 20 m/s, indicates relatively reasonable wind farm accessibility to perform a maintenance visit. Figure 4: Seasonal wind speed trends as hourly maximums. This figure is obtained calculating the maximums per hour of each day over a season in 2019. The window, which is shaded in yellow represents the day shift from 08:00 to 18:00. The majority of scheduled maintenance interventions are planned in summer and autumn months in the case study maintenance log. For this reason, 9 In dendrogram visualisation, height represents the value of the Euclidean distance between clusters. To estimate this distance, input data must be scaled. As an example, for an input consisting of 100 rows and 2 columns, the first column indicates the working zone and the second one stands for the corresponding wind speed restriction. After scaling the input data each observation is firstly assigned to a temporary cluster. Following this procedure, in the first step there exist 100 clusters (100 tasks) and, for instance, the Euclidean distance between Cluster 1 and Cluster 2 can be obtained as; Eucdist =q(HSEvC1−HSEvC2)2+ (WZC1−WZC2)2(5) where HSEvrepresents the wind speed restriction and W Z represents the working zone. The same calculation is repeated for all 100 clusters. Afterwards, the Ward algorithm groups these clusters according to the minimisation principle of Euclidean distances. 4.4. Proposed framework The proposed methodology is graphically explained in Figure 7. The initial step is to provide information on the type of the intervention, initial safe working rules and wind forecasts. Then, it is required to decide if wind gust measurements and estimations are needed as decision variables. The corresponding answer depends on the specific requirements of the planned intervention, such that intervention may require a crane usage. 16 Figure 7: Flowchart of proposed solution, HSE: Health-Safety and Environment regulations In the proposed methodology iis the user defined limit for initiating the agglomerative nesting process, as shown in Figure 7. Here, we assumed that a maintenance task can be done within a minimum of four stages such as: 17 access to working area, access to failed component, remove failed component and placement of the new component. Then, for a case that each stage requires a unique safe working rule, the minimum total number of safe working rules is 4. Therefore, predetermined comparison value, i, is set to 4 . For an intervention consisting of more than four tasks or requiring the fulfilment of more than four safety rules related to wind speed, forecasts must be used along with the outcomes of the agglomerative nesting as input in the search process. The gust forecasts are necessary if the intervention is performed using a crane, which requires reduction of the wind speed limits due to the high gust values. Lastly, the search process scans the available time windows during the intended maintenance day to find the optimal time window for the work to take place. If the maintenance intervention can be executed during the pre-planned day, optimal execution time and order of the tasks are determined. If not, a change in the pre-planned day is suggested. This methodology can also be used for offshore applications, but it is very important to update HSE requirements considering wave height and offshore operations specific rules. Moreover, intervention type, required duration, outputs of annual maintenance, etc. must be updated considering the technology type and the working environment. 5. Results The trials with the proposed approach for two distinct maintenance visits are reported in this section. Case 1 is an application test for a routine maintenance visit, whereas Case 2 focuses on a major component replacement. 5.1. Case 1: Routine Maintenance 5.1.1. Clustering The problem of planning a high number of tasks is simplified by applying the agglomerative nesting methodology to the pool of 36 tasks. Clustering was performed using the Euclidean distance as similarity criterion. It was calculated using the wind speed limit and the corresponding turbine working zone of each operation. Figure 8 shows how the tasks are grouped forming a total number of 4 clusters (represented with different colors) as a function of the restrictions, wind turbine working zone and wind speed. 18 Figure 8: Graphical representation of the clustering process. Different colours represent the different clusters of tasks. (The dendrogram needs to be regenerated for different technologies considering maintenance intervention lists.). A summary of the clustering results is given in Table 2 where the cluster duration and its wind speed limit are shown. As maintenance tasks are usually accomplished by two technicians, which will require half the time, and the required resolution for the planning schedule is based on 10-minute steps, the rounded duration per person on a 10-minute scale is also provided. Table 2: Clustering results Cluster Duration (min) Per person (10 mins) vlim (m/s) 1 66 4 20 2 106 6 15 3 491 25 12 4 50 3 10 5.1.2. 24 hours evaluation for executable/not executable windows Now by applying the procedure, explained in Section 4.2, with measured wind speed data of test days (the summer day was 27th June 2019 and the autumn day, 08th November 2019) executable and not executable periods for the maintenance clusters are determined. Figure 9 and Figure 10 show the allowed intervention starting times for each of the clusters found in the previous section. 19 Execution of the maintenance service is only possible, if the starting time of the intervention is within the green dots. Here green dots represent valid periods for both wind speed safe working limit and the availability of a window to accomplish the task within its minimum required completion duration. In these figures, vlim represents wind speed limit and Dur stands for the required duration for the execution of the corresponding cluster. Figures 9 and 10 are given in order to display the complexity of programming with dynamic weather restrictions. The decision maker must consider all the intervention specific accessibility windows and generate a maintenance program combining them. Figure 9: Routine maintenance evaluation with actual input data for the summer day (a) cluster 1, (b) cluster 2 (c) cluster 3 (d) cluster 4 20 Figure 10: Routine maintenance evaluation with actual input data for the autumn day (a) cluster 1, (b) cluster 2 (c) cluster 3 (d) cluster 4 The results for each of the clusters were: Cluster 1: Tasks are executable during both analysed days, since the corresponding wind speed restriction is very flexible and its duration is relatively low, see Figures 9a and 10a. Cluster 2: Tasks are mostly executable for both days, see Figures 9b and 10b. Although, there are short non-executable windows in the autumn day, see Figure 9b. Cluster 3: Tasks are executable for the calmer summer day and tasks are non-executable for the windier autumn day, see Figures 9c and 10c. Cluster 3 tasks are the most challenging group, because they require a longer time with major wind speed restrictions. Cluster 4: The execution of Cluster 4 tasks depends mostly on the most restrictive wind speed limit. Nevertheless, it can be seen that there exist some executable windows, since the execution of this cluster requires the lowest duration, see Figures 9d and 10d. These preliminary analysis shows that in the summer day all tasks can be performed, whereas in the autumn day there is no suitable time window to perform Cluster 3 tasks. Therefore, in the next analysis only the results obtained from the summer day are presented. Considering the hourly electricity market price, it is possible to combine 21 it with the energy losses for each plan, estimated from measured wind speed values and manufacturer’s power curve, to obtain the revenue prioritised decision pools. Figure 11 shows the day-ahead electricity market prices for 27th June 2019 and 08th November 2019 [35]. It can be clearly seen that in Summer, 27th June 2019, the electricity market prices higher than 50 EUR/MWh, while in Autum, 08th November 2019, most of the prices are between 30 and 40 EUR/MWh, following the common trend observed in the spanish market [35]. Figure 11: Day-ahead hourly electricity market price Figure 12 shows the corresponding revenue losses of the prioritised maintenance plans, labelled as “low” when they are below the mean of the revenue losses estimated for the day under consideration. When they are lesser than the third quartile and greater than the mean, the label is “medium”. Lastly, for the plans with the revenue losses greater than the third quartile, the label is “high”. This DSS is prepared as a computational tool and the visualisation of the reporting module is given in the Appendix 1, where the alternative plans and the revenue evaluation procedure are exemplified. 22 Figure 12: Decision pool for routine maintenance visit scheduling for the summer day, the second y axis stands for the grouping according to the revenue losses, yellow shaded window shows the day shift. In Figure 12, only selected alternative maintenance plans are plotted when a cost-wise clear separation can be observed among the 1093 alternatives (14 of 1093) for the analysed summer day. Each alternative represents a program, which confirms that weather related downtime is minimised. According to Figure 12, the early hours of the day are more preferable in order to perform preventative intervention considering the revenue losses. Although electricity market prices are high during these hours, the limited wind resource availability, reduces power production losses and corresponding revenue losses. 5.2. Case 2: Generator replacement 5.2.1. 24 hours evaluation for executable/not executable windows In this case study, the generator replacement is investigated for the proposed scheduling process. To replace the generator, a crane must be used. 23 Firstly, the nacelle cover must be removed and then the failed generator must be taken out. These removals are followed by installation of the new generator and re-installation of the original nacelle cover. In other words, this intervention requires two types of lifting /unloading tasks. Safety requirements with regards to wind speed vary due to the gust values. The mean wind speed limit for safe working has to be decreased by 2 m/s when the wind speed gust is above 5 m/s for operation requiring a crane usage [16], from 10 m/s, for a gust lower than 5 m/s, to 8 m/s for a gust higher than 5 m/s in the case of nacelle cover and from 8 m/s to 6 m/s, for the same gust values, in the case of the generator. It is worth mentioning here that the gust limit, to the authors knowledge, has never been considered in previous scientific studies. Another difference, regarding routine maintenance plan, is the requirement to follow a fixed task order, as obviously, it would not be possible to perform removal of old generator before removing the nacelle cover. Therefore, the maintenance execution order is fixed for this problem. The obtained results for a corrective maintenance visit in the previously selected Summer day are shown in Figure 13. Executable (green) and not executable (red) time windows are shown for the four main tasks of a corrective intervention. Here, a 120 minutes window is searched for the removal of the old generator and another 120 minutes window for placement of the new one. In these searches, the permissible wind speed reduces from 8 m/s to 6 m/s, when the wind gust value exceeds 5 m/s. The remaining tasks require a 90 minutes window search for the removal of the nacelle cover and another 90 minutes for the placement. In these searches, the permissible wind speed reduces from 10 m/s to 8 m/s, when the wind gust value exceeds 5 m/s. 24 Figure 13: Evaluation of generator replacement when considering dynamic safe working limits due to gust (Summer day). It is rather easy to highlight the impact of the gust variable with a simple comparison between Figure 13 and Figure 10. Due to the gust related restrictions, a corrective intervention cannot be performed in this case, although it was possible to perform a preventative maintenance intervention. 6. Discussion and Limitations When a decision maker uses only the mean wind speed characteristics, any day from the summer season is a good candidate in order to prepare the maintenance plans. This study presented that each candidate day must be analysed profoundly. Because, while the power losses resulting from the maintenance interventions could be limited, the revenue losses could be severe due to the electricity market prices and vice versa. As it is shown in this study, not only the mean wind speeds, but also the wind gusts are the limiting factors for performing some major maintenance activities. The implementation of other environmental limiting factors (fog, rain, etc.) was not possible due to data unavailability. The practicality of such a DSS highly depends on the input data. Uncertainties in regards to duration of tasks and in relation to weather forecasts are not considered in the present study. It must be noted that in order to 25