On the importance of assessing the operational context impact on maintenance management for life cycle cost of wind energy projects

On the importance of assessing the operational context impact on maintenance management for life cycle cost of wind energy projects J.Izquierdo a,b, ∗ , A.Crespo Márquez b, ∗∗ , J.Uribetxebarria a , A.Erguido a a Ikerlan Technology Research Centre, Basque Research and Technology Alliance (BRTA), 20500 Gipuzkoa, Spain b Industrial Organization and Business Management I, School of Engineering, University of Seville, 41092 Seville, Spain 5 Abstract The increasing demand for energy from renewable sources is entailing the development of technologies oriented to increase the protability of such projects and thus the attractiveness for potential investors. Wind power constitutes one of the most relevant renewable energy sources; however, the costs of the wind farms associated with Operations & Maintenance are prominent along the life-cycle. This paper proposes an approach intended to reduce these costs and lower the Levelized Cost of Energy. In this context, it is presented an opportunistic maintenance policy based on more accurate reliability estimates of the wind turbines components. The reliability of the components is estimated through a model based on Articial Neural Networks that dynamically calculates the impact of operational conditions on the failures of the wind turbines. The approach has been validated through a case study based on real eld data which proposes a multi-objective optimization of the maintenance strategy for the life-cycle of a wind farm. The obtained results provide interesting ndings from the perspective of wind farms investors, operators, and owners. Keywords: Wind energy, Maintenance management, Life-cycle, Articial Neural Network, Operational context 1. Introduction The deteriorating environment along with global warming and the shortage of fossil fuels is a current 10 issue rising pressure levels in governments around the world (and in the European Union, EU) [1]. These concerns are propitiating policies like binding targets on greenhouse emissions and are urging a shift towards renewable energy sources [2, 3]. The attention drawn by renewable energy has increased over the recent years nurturing an important growth that has been especially prominent in the wind energy sector [4, 5]. For instance, in the EU, wind power installed more capacity than any other form of power 15 generation in 2018, rising from 12% in 2017 to 14% of covered energy demand [6]. However, to keep up with the increasing demand for renewable and aordable energy, the protability of wind energy projects should be guaranteed by reducing the Levelized Cost of Energy (LCoE) to its minimum [7]. In the literature, a considerable amount of works aimed at reducing the LCoE by addressing the Operations and Maintenance costs (O&M), see for instance [810] among other works which will be later on reviewed. 20 The costs associated with O&M are known to be prominent [11, 12]. They may account for 12-30% of onshore wind farms (WFs) rising up to 32% in oshore projects [13, 14]. These costs are uncertain and inuence the economic feasibility of wind energy projects; a potential investor will rather allocate resources in a project not susceptible to risks [15]. In this context, to increase the cost-eectiveness of WFs it is necessary to reduce the cost derived from O&M activities [16, 17]. Nonetheless, the prob25 lem of the present objective of reducing the LCoE by cutting the O&M costs is a two-fold challenge [18]. If maintenance activities are insucient, the failure rate will increase lowering system's reliability. Otherwise, if maintenance activities are performed too often, the system's maintenance costs increase to undesirable levels [16]. Besides, it is necessary to minimize the lost energy production at down-times caused by failures or maintenance activities for the entire life-cycle [19], which oscillates around 20 years 30 [20]. ∗ Principal corresponding author, Tel.: +34 943 712400. ∗∗ Corresponding author, Tel.: +34 954 487215. Email addresses: [email protected] (J.Izquierdo), [email protected] (A.Crespo Márquez), juribetxebar[email protected] (J.Uribetxebarria), [email protected] (A.Erguido) Preprint submitted to Renewable Energy January 13, 2020 1.1. Related works Considering the data provided by the International Renewable Energy Agency and other works [2123], maintenance activities account for a considerable fraction of the LCoE. On these grounds, the optimization of the O&M strategy acquires an important role [24]. The evolution of the models and approaches to 35 optimize the maintenance strategy have evolved along with the steady technological development of the wind turbines (WTs) [7]. The determination of WFs operators for maximizing the protability of the investment drives the development of new techniques and decision-support tools for optimal maintenance strategies [18]. The maintenance management works are mainly focused on two main objectives, the minimization of the costs whilst maximizing the availability of the WTs [7, 25]. 40 In this context, it is essential to consider and combine physical and statistical models with technical know-how [26, 27]. The WFs operators are compelled to implant new techniques and decision-support tools for optimal maintenance strategies if they seek to maximize the protability of their investment [18]. However, the reality nowadays is that the most applied strategies are corrective maintenance (CM) and time-based minor preventive maintenance (PM) [28, 29]. Additionally, Condition Based Mainte45 nance (CBM) is a popular method which has proven to be cost-eective [30, 31] and it has been widely researched [32, 33]. Notwithstanding the eectiveness of CBM methods, it is crucial to consider in the maintenance strategy that WTs are multi-components systems constituted by subsystems with dependencies among them conditioning the adequacy of the maintenance strategy [34, 35]. These dependencies have been classied as (i) structural, when to perform maintenance actions on one system some actions 50 are required on others [36]; (ii) economic, when performing simultaneous activities entails dierent costs than performing them separately [37] ; and (iii) stochastic, for those systems whose failure rates are not independent [38]. The most studied maintenance policies dealing with the aforementioned dependencies of the WTs are the opportunistic maintenance and group maintenance [34]. In the case of wind energy, the opportunistic 55 maintenance is especially interesting since it takes advantage of short-term circumstances performing maintenance actions on non-failed systems when failure happened in another one [25]. Traditionally, this policy has not been implemented in the wind energy sector [39], however, some recent works have demonstrated its potential for reaping important benets due to economic dependencies among WTs, e.g. [4043]. 60 Among the reviewed publications, the work of Besnard et al. [40] focuses on reducing maintenance and opportunity costs jointly performing corrective and preventive actions at low wind speed periods. The work of Tian et al. [39] is a step forward in an opportunistic maintenance policy supported by condition monitoring indicators, the work proposes a reliability threshold based on systems' remaining useful life. Ding and Tian [35] consider perfect and imperfect maintenance actions in their work, these actions are 65 triggered by an age indicator based on the Mean Time To Failure of the WTs systems. Another important contribution of the same authors is an extension of the previous work [28] where they consider dierent age thresholds for systems belonging to failed and running WTs. It is interesting to consider as well the work of Atashgar and Abdollahzadeh [44] in which they address the two-fold challenge previously mentioned of reducing the costs and maximizing the energy production by implementing a multi-objective 70 optimization of the opportunistic policy. And the work of Abdollahzadeh et al. [41] determines the optimal maintenance activities according to reliability thresholds calculated for each component. Finally, the work of Zhu et al. [43] is based on the study of three dierent maintenance strategies consisting of periodic routines, reactive maintenance, and opportunistic maintenance. According to the reviewed works, it is important to notice that the decision of whether to maintain 75 or to not maintain a system is taken according to dierent thresholds regarding the system's age, reliability or health condition. It is therefore essential to estimate those indicators as accurately as possible [19]. In particular, reliability indicators are estimated with traditional models which involve assumptions and simplications [45]; these are the reason underlying the inability of the reliability models to properly describe the true behaviour of the systems [46]. An important assumption that induces consid80 erable uncertainty is the consideration that the operational conditions and external factors inuencing the assets are constant [47]. It has been already stated in the literature that more realistic reliability estimates, through a model integrating operational context information, will enable more eective and better-customized maintenance strategies [48]. There have been several authors who have studied the aection that the operational context may have 85 on reliability engineering and thus on maintenance management. The operational context is explicitly taken into account in the work of Tang et al. [49] for cable failures, and so it is in the work of Lin et al. [50] for traction transformers. More specically in the wind energy sector, the operational context eect has been considered to model the failure rate in WTs components [51]. Besides, the work of Mazidi et 2 al. [52] explores how dierent operational parameters aect the stress condition of the WTs. 90 Whilst reviewing relevant related works, it is inevitable to encounter with research making use of the Proportional Hazards Model to relate the reliability of the components with operational parameters, e.g. [53, 54]. Another recurrent approach, which has attracted considerable attention lately (see [55]), is the application of Articial Neural Networks (ANN) due to their capabilities to represent non-linear relationships [56], and also due to the fact that no "a priori" assumption of the model is required [57]. 95 Several authors present interesting research works in the application of ANN to determine components reliability according to operational context; Al-Garni et al. [58] compare their performance against traditional Weibull regression model, Fink et al. [59] provides a time-series perspective to predict reliability through ANN and in Beg et al. [60] several ANN-based models are compared. Besides, in the specic context of wind energy the ANN methods have also proven very useful predicting reliability [33, 61]. 100 It is especially interesting for the scope of the present research the work of Izquierdo et al. [48] where statistical models are combined with ANN methods to provide a novel model with dynamic capabilities. 1.2. Motivation and scientic contribution It has been already stated the role that O&M activities have on the costs associated with a wind energy project and therefore, the importance of optimizing the maintenance management in such scenarios. The 105 literature review shows evidence that the opportunistic policy has the potential of minimizing maintenance costs whilst maximizing energy production. However, this policy should be supported by estimates which undoubtedly must be accurate in order to ensure the eectiveness of the maintenance actions. Nevertheless, the traditional reliability models, which render the estimates that trigger maintenance actions, involve assumptions and simplications which may jeopardize the accuracy of the estimates. An 110 important assumption often found in traditional reliability models is the operation under constant working conditions, but recent works have provided tools, technologies, and methods capable of overcoming the aforementioned simplication. These works show how their proposals render better estimates than traditional reliability models. However, to the best of authors' knowledge, the benets of the proposed models have not been integrated with the benets of advanced maintenance policies such as opportunistic 115 maintenance. The integration of opportunistic maintenance with models integrating the eect of working conditions on assets' reliability is a novel proposition that should be compared with the same policy supported by traditional reliability models to see if it provides any improvement. The research herein presented provides considerable improvements that will set the foundations to explore the literature gap of combining reliability models integrating the operational context with advanced maintenance policies. 120 Accordingly, the research herein presented aims at unifying an operational-context aware reliability model with an opportunistic maintenance policy. The present work contributes by demonstrating, through a case study in the wind energy sector, that a maintenance policy should be supported by accurate reliability estimates considering operational context; and vice versa, an advanced reliability model, which takes into account the operational context, provides an important potential if it is integrated within an 125 opportunistic maintenance policy. The link-up of an opportunistic maintenance policy with a reliability model considering operational context is a novel proposal intended to provide less uncertainty for potential investors in wind energy projects for two reasons: • The opportunistic maintenance policy is optimized according to a multi-objective logic, i.e. the costs are not optimized regardless of other business objectives, organizational goals such as the 130 maximization of energy production are considered as well. Such multi-objective logic entails a trade-o among the objectives since they often imply competing scenarios. In such a context, the multi-objective optimization of the opportunistic policy provides a wide spectrum of solutions so dierent trade-os may be considered and the one more suitable for the business goals selected. • Since the reliability model assesses the WTs failure probability considering the operational context 135 and the changes happening in it, the estimates are more accurate and reect better the real failure behaviour of the dierent components which may be inuenced by dierent parameters. The in- uence of the operational context is considered in the estimates that will trigger the maintenance actions and this fact is remarkable because it provides a universality character to the maintenance plans guaranteeing that the output of certain maintenance plans will not dier for dierent opera140 tional contexts. The research presented consist of theoretical and practical foundations which gather and combine state-of-the-art contributions of the authors to the elds of reliability engineering, asset management and O&M in the wind energy sector. The scientic contribution of the work is the novel conjoint consideration 3 of a recent opportunistic maintenance policy with a recent reliability model making use of ANN to consider 145 the operational context. This combination is validated through a case study in which it is opposed to the traditional approach based on the Weibull reliability model. In order to provide a solid validation, the case study is based on real-eld data and consists of extensive experimentation. 1.3. Overview In order to provide a holistic and comprehensive overview of the contribution here presented, the clas150 sication framework proposed by Shaee and Sørensen [3] is utilized. In the framework ve classication criteria are decomposed into several categories, Table 1 contains the information positioning the research here presented according to the framework's classication criteria. Table 1: Positioning of the research according to criteria in Shaee and Sørensen [3] Criteria Description Present research positioning System conguration The type of wind power asset and the level of system modelling WT at component level. Decision-making attribute Planning horizon, the decision-maker and the availability of eld data Finite time horizon considering time as continuous variable. The decision-maker are considered to be the WF owners and operators. The data for the case study comes from eld failure data. System failure modelling Include the type of damage/failure and the failure modelling approach Both, minor and major failures are considered with grey-box models. Optimization model Optimality criterion and the solution technique Two optimally criteria are considered, the minimization of cost and the maximization of power output. The solution technique is a multi-objective model. Maintenance strategy The maintenance policy and the eectiveness of the repair actions The opportunistic maintenance is the chosen policy considering both imperfect and perfect repair eectiveness. The remaining of the paper consists of several sections that cover the theoretical aspects, the case study and the conclusions withdrawn from the obtained results. Section 2 depicts the theoretical aspects 155 covering the maintenance strategy and the calculation of the life-cycle costs. This section explains the opportunistic logic contemplated under a life-cycle perspective as opposed to current practices of optimizing the rst years of the maintenance of the WTs. Then section 3 introduces the reliability model which will support the maintenance strategy, in the section the strengths of the model are detailed. The case study is described in section 4 which comprehends the description of the data utilized and the adopted 160 approach along with the obtained results. Finally, section 5 comprises the key conclusions obtained from the study and its results. 2. Maintenance strategy Considering the insights provided by the literature review, an opportunistic maintenance policy is proposed to make optimal maintenance decisions beneting from the economic dependencies among the 165 WTs' components. This maintenance strategy is intended to maximize the energy outcome of the WF whilst minimizing the costs not only for the rst operating years but the whole life-cycle. The maintenance model has been adapted from recent literature, and while the interested readers are addressed to see the original model in [25], the essential aspects and characteristics of the proposed policy are hereunder presented. 170 The generic problem to be considered can be dened as a WF involving the maintenance of the WTs ( h= 1,2, ..., H ) and their systems ( i= 1,2, ..., N ) arranged in a serial disposition for failing purposes. 4 Each of the systems may fail in k dierent failure modes (FMs) entailing dierent consequences, and therefore requiring dierent CM actions ( k= 1,2, ..., K ). A FM in this study is considered according to the denition introduced by Rausand and Høyland [62] and by Crespo [63] in which it is the manifestation 175 of a failure entailing the termination of one or more functions. Besides, not only the WF managers can decide to preventively maintain the WTs' systems before failure occurs, but it is possible to perform PM at dierent levels ( j= 1,2, ..., J ) depending on the restoration factor ( q ). If the preventive action restores the system to a state as-good-as-new it is considered as a perfect maintenance; on the contrary, if the PM action leaves the system in a better state than before but still worse than new, it is considered to be 180 an imperfect maintenance being j= 1 for the present model the most imperfect maintenance and j=J the perfect one (see [64] for further details). The opportunistic maintenance is intended to address the previously described problem and it is optimized according to two main objectives, these objectives consider business implications and have a life-cycle perspective. For wind energy projects, it is essential to optimize the operational costs of the 185 WF whilst also minimizing the lost energy production due to downtimes in the WTs. To mathematically describe these objectives it is important to dene corrective and preventive costs, subject to the restoration eect ( q ). The corrective cost (CC) and the preventive cost (PC) can be seen in Equation 1 and Equation 2 respectively, they consider the materials and tools needed to perform the actions ( cc ik , cpr ik ), energyproduction opportunity cost ( cna ) and penalty cost ( cp ) in case the supplied energy does not meet the 190 committed level. Also, it is important to consider some binary decision variables associated to the model in order to understand the model: zhikt determines if CM action k is performed in system i of WT h in period t , yhikjt does the same with PM actions but the subindex j determines the type of preventive action, θt establishes if a maintenance team is correctively dispatched to the WF in period t and γt determines identical action but for a preventive dispatch. 195 CC =X hX iX kX t zhikt hcc ik (qc ik)2+mc ik ·GPt(cna +cp)i (1) P C =X hX iX kX jX t yhikjt cpr ik qpr ikj2 +mpr ikj ·GPt·cna (2) Another major cost element related to the costs-minimization objective is derived from the maintenance resources (MC) required to attend the WF, see Equation 3. They consider the number of maintenance teams (NT), their xed costs ( cteam ), and the costs associated with their dispatches( cdisp ) either preventively ( γt ) or correctively ( θt ). MC =NT ·cteam +X t (γt+θt)·cdisp (3) According to the second objective of the maintenance strategy optimization, the lost production (LP) 200 is calculated as described in Equation 4. To calculate the lost in every period, the maintainability of CM and PM ( mc ik and mpr ik respectively) are considered along with the power that would have been generated in that time ( GPt calculated as in [65]). LP =X t GPt X hX iX k mc ik ·zhikt +X hX iX kX j mpr ikj ·yhikjt  (4) Having dened the components of the two objective functions of the optimization, it is vital to also dene the constraints. The decision of whether preventively maintain or not a non-failed system is taken according to reliability thresholds: DRTikt is the threshold that compulsory dispatches a maintenance team to perform PM and SRTikjt is the threshold to determine certain PM action (according to j) once there is at least one maintenance team in the WF. Every SRTj threshold must be higher than the DRT , and they should be sorted according to their level, being the lowest the most imperfect action threshold (see constraint in Equation 7). Another important constraint is the availability and working time of the maintenance teams dened as Twt and regarded in Equation 8. Finally, it is important to consider in the model that only one maintenance action in the same WT is allowed for a single time period (see Equation 5 9). Therefore, having dened the terms comprising the objectives functions, as well as the ones in the constraints, the formulation of the model is expressed by the following equations, where ka is the rate of the time value for money: OFOpex =min ([MC +CC + +P C]·(1 + ka)−t (5) OFLP =min (LP ) (6) S.T. 0≤DRTikt ≤SRTik1t≤... ≤SRTikjt ≤≤ ... ≤SRTikJt ≤1iI, kK, jJ;tT (7) X iX kX j mpr ikj ·yikjt +X iX k mc ik ·zikt ≤NT ·Twt ∀tT (8) X j yhikjt +zhikt ≤1hH, iI, kK, tT (9) zhikt, yhikjt{0,1}hH, iI, kK, tT, ∀j= 1,2 3. Reliability model The ultimate goal of the optimization process is to nd the values of the reliability thresholds (i.e. 205 DRTikt and SRTikjt ) that maximizes the energy production and minimizes the maintenance costs. As the optimized thresholds will launch maintenance activities, it is vital to ensure that the estimates of components' reliability are as accurate as possible. Therefore, as stated by several authors mentioned in the literature review, it is required to integrate operational context information in the reliability models. To such aim, a dynamic ANN-based reliability model is adopted from the work by Izquierdo et al. [48]. 210 The present model is characterized by a failure rate decomposed in two terms, a baseline hazard dependent on time and an exponential part in which the operational context variables are considered as inputs of an ANN. The mathematical formulation of the hazard function can be seen in Equation 10; the h0(t) term corresponds to the baseline hazard; and the neural network is denoted as the function G(X,W,B) where X is the input covariates vector, W are the weights of the connections between the nodes and B 215 collects the bias parameters of the ANN. h(t, X) = h0(t)·exp (G(X,W,B)) (10) The ANN embedded in the hazard function employs as activation function the hyperbolic tangent described by Equation 11, and the input values of the operational context are normalized through Equation 12. As the ANN is integrated into a statistical model, the traditional training methods are of no use in this case; however, in order to obtain the optimal weights ( W ) and bias ( B ) values, the concept of max220 imum partial likelihood (introduced by Cox [66]) is employed. The optimization consists of nding the values that maximize the partial likelihood ( L ) described by Equation 13 in which p are all the historical failure data. As the complex architecture of ANN involves numerous parameters and thus considerably complicates the optimization process, a genetic algorithm is employed in order to nd the optimal weights ( W ) and bias ( B ). 225 g(x) = exp(x)−exp(−x) exp(x) + exp(−x) (11) xnorm =2(x−xmin) (xmax −xmin)−1 (12) 6 L= p Y i=1 exp(G(Xi,W,B)) Pl  Riexp(G(Xl,W,B)) (13) Once the optimal parameters of the ANN have been obtained, if the baseline hazard rate is considered to follow a Weibull distribution, the reliability may be expressed by Equation 14 in which α is the scale parameter and β is the shape parameter. If the integration is solved for an asset operating with changes in the operational parameters, the reliability model is expressed by Equation 15 in which Ci is the integration constant for the dierent i operational contexts in which the asset works. 230 R(t, Xi) = exp −ˆβ αt αβ−1 exp(G(X,W,B))dt! (14) R(t, Xi) = exp −t αγ exp(G(Xi,W,B)) + Ci (15) Paying special attention at the integration constant Ci and assuming that the reliability is a continuous function with a value of 1 at t= 0 , i.e. no failure probability before start operating, the value of the constant can be described by Equation 16. It is important to notice that the value of the constant Ci∀i6= 0 , can be decomposed into three terms: •ti αγ . It is reasonable to reckon that the changes in the operational context do not aect the asset 235 equally along the span of its operating time. This term explains such logic, the aection that a change may have on the asset depends on its technical characteristics, thus the scale ( α ) and shape ( β ) parameters, and on the moment at which the changes happen ( ti ). •(exp(G(Xi,W,B)) −exp(G(Xi−1,W,B)) . The impact that an operational context change may have on an asset is going to depend on how dierent the new conditions are from the previous, such 240 information is integrated into this term. • Ci−1 . By the recurrence of taking into account its previous value, the model is integrating information from previous operational context changes. It means that the system's current probability of failure is also aected by the operational changes suered in its past operating time. Ci=   0∀i= 0 ti αγ(exp(G(Xi,W,B)) −exp(G(Xi−1,W,B))) + Ci−1∀i6= 0 (16) As can be seen, the model integrates the changes in the operational context, which are not specied 245 but dened according to time periods allowing dynamic calculation of reliability in dierent time intervals, which don't have to be of the same length. Besides, by embedding into a statistical model the architecture of ANN it is possible to integrate information regarding interactions among operational context's variables and some other hidden phenomena without 'a priori' dening them. 4. Wind energy case study 250 To test if the integration of the operational context-aware reliability model with opportunistic maintenance policy provides a solid approach to maintenance management, a case study based on real eld data is proposed. However, due to the stochastic processes entailed by maintenance management, it is dicult to adopt an analytical resolution method. Therefore a simulation-based is here presented since it has proven to eectively characterize the maintenance processes and its optimization [41, 42, 67]. The logic 255 underlying the simulation is represented in the owchart of Figure 1. It can be seen how the dynamic ANN-based reliability models calculate the reliability of the components in every period and then the values are compared with the opportunistic threshold to trigger maintenance actions if necessary. Considering that the research pursues more than one objective, i.e. minimization of costs and lost energy production, a multi-objective algorithm must be implemented. To this aim, the Non-Sorted 260 7 Figure 1: Simulation chart Genetic Algorithm II (NSGA-II) has been implemented as it has proven to be useful for providing highquality non-dominated solutions on the Pareto front [19, 25, 68]. Accordingly, the optimal maintenance strategies will be obtained by the joint use of the simulation and the optimization. On the one hand, The simulation allows evaluating in every iteration of the NSGA-II algorithm the results of certain maintenance strategies in terms of cost key performance indicator (KPI) and energy lost KPI. Whilst, on 265 the other hand, the optimization will guide the maintenance thresholds towards their optimal according to the NSGA-II logic. Since the scope of this research is to demonstrate the added value provided by a reliability model with operational context information integrated into the maintenance policy, this scenario has been compared with one applying a traditional Weibull reliability model. Besides, this research is also intended to 270 evidence that the optimization should be done considering the whole life-cycle of the project as opposed to current practices of optimizing the rst 5-10 years of the project as a part of the warranty plans. Therefore, according to these aims the case study described in Figure 2 presents the optimization of 4 scenarios: Scenario A - Weibull reliability model and 7 years optimization. In this scenario, a traditional 275 two-parameter Weibull model is adjusted for the failure data of every FM of the WTs studied. Having the parameters of the Weibull models, the opportunistic policy is optimized for the rst years of the WF operating time. Then it is evaluated how the optimal strategy for the 7 years behaves in a lifecycle, i.e. 20 years, perspective, obtaining several cost and energy lost KPIs for the dierent strategies which comprehend the Pareto-front. 280 Scenario B - Weibull reliability model and 20 years optimization. In this scenario, a Pareto front is obtained from the optimization of the maintenance strategy for 20 years. The reliability estimates triggering maintenance actions are calculated with the adjusted Weibull models for every FM. Scenario C - Dynamic ANN-based reliability model and 7 years optimization. In this 285 scenario, a dynamic ANN-based reliability model is adjusted for every FM and then embedded in the simulation to optimize the maintenance for 7 years. Having optimized the reliability thresholds for 7 years, it is tested the output these thresholds render in a 20 years life period. Scenario D - Dynamic ANN-based reliability model and 20 years optimization. In this scenario, the simulation is with the same models of Scenario C but optimizing the thresholds for 20 290 years ensuring that the maintenance policy is not optimal for rst years of operating time but for the life-cycle of the project. The comparison of the four scenarios will be later on presented. Firstly, the optimizations for the 7 years are compared (scenarios A and C) to see if in the short term the model integrating the opera8 tional context provides better solutions than the Weibull model. The strategies provided by these two 295 optimizations then are projected in the long term to see their performance in spite of not being optimal. Then, the optimizations for 20 years (scenarios B and D) are compared among them and with the 20 years projections of the previous 7-years optimizations. It is done so it can be seen how a life-cycle perspective provides better solutions even with a traditional reliability model (scenario B compared with scenario A). Then both are compared with a maintenance 300 plan optimized for the early years but with a model integrating operational context information (scenario C compared with B and A). Finally, scenario D will be compared with the previous to see how the dynamic ANN-based model in a life-cycle perspective provides the best results in terms of costs and energy loss. Figure 2: Case study procedure chart 4.1. Case description and data considerations 305 The case study presented in Figure 2 is based on real eld data provided by a wind energy OEM and comes from over 300 onshore WTs of 1.67 megawatts(MW) which are operating in dierent locations in the north of Spain, and therefore they are exposed to multiple operational conditions. The eld data comprises a period of over 12 years and includes two databases coming from the historic maintenance records and the SCADA information. The SCADA data contained information regarding the alarms and states of 310 the WTs along with sensors information, and the maintenance records contained the carried maintenance activities with the corresponding associated costs. Both databases were combined by identifying the patterns of the failures in the alarms and states and linking them with the corresponding maintenance actions in order to build a single RAM (Reliability, Availability, and Maintainability) database containing the dierent Times Between Failure (TBFs), the associated costs and the operational conditions. The 315 obtained RAM database was validated and contrasted with the know-how and opinion of experts in WTs operators and WTs OEM, and it is nowadays used to support data-driven decision-making processes. For the simulation, a WF consisting of 62 WTs is considered. Every WT consist of four main components (N=4) and each one of the components has minor and major failures (K=2). The considered components correspond to four critical systems from the maintenance perspective: gearbox, blades, yaw, 320 and pitch. Therefore there are two FMs for every component, and for every of them it is considered perfect and imperfect maintenance levels with corresponding restoration factors of qpr ik2= 1 and qpr ik1= 0.75 respectively. Regarding the cost structure, the costs considered in the simulation are those regarded as relevant to analyse the maintenance management from a life-cycle perspective. The cost of a maintenance team 325 consisting of 2 workers is considered to be 800 e /day, 105 e /MWh as an opportunity cost and 35 e /MWh as the penalization cost. Likewise, each component has its corresponding material cost for which the interested readers are addressed to [69]. The cost of PM is considered to be 30% lower than CM and a discount rate of 5% is considered for the annualized life-cycle cost analysis. 9 URL https://onlinelibrary.wiley.com/doi/abs/10.1002/qre.1114 635 [59] O. Fink, E. Zio, U. Weidmann, Predicting component reliability and level of degradation with complex-valued neural networks, Reliability Engineering & System Safety 121 (2014) 198  206. doi:https://doi.org/10.1016/j.ress.2013. 08.004 . URL http://www.sciencedirect.com/science/article/pii/S0951832013002391 [60] A. Beg, F. Awwad, W. Ibrahim, F. Ahmed, On the reliability estimation of nano-circuits using neural networks, 640 Microprocessors and Microsystems 39 (8) (2015) 674  685. doi:https://doi.org/10.1016/j.micpro.2015.09.008 . URL http://www.sciencedirect.com/science/article/pii/S0141933115001507 [61] Y. Yan, J. Li, D. W. Gao, Condition parameter modeling for anomaly detection in wind turbines, Energies 7 (5) (2014) 31043120. doi:10.3390/en7053104 . URL https://www.mdpi.com/1996-1073/7/5/3104 645 [62] M. Rausand, A. Höyland, System reliability theory: models, statistical methods, and applications, Vol. 396, John Wiley & Sons, 2004. URL https://www.wiley.com/en-us/System+Reliability+Theory%3A+Models%2C+Statistical+Methods%2C+and+ Applications%2C+2nd+Edition-p-9780471471332 [63] A. Crespo, The maintenance management framework: models and methods for complex systems maintenance, 650 1st Edition, Relibaility Engineering, Springer Science & Business Media, 2007. doi:https://doi.org/10.1007/ 978-1-84628-821-0 . URL https://link.springer.com/book/10.1007/978-1-84628-821-0#about [64] H. Pham, H. Wang, Imperfect maintenance, European Journal of Operational Research 94 (3) (1996) 425438. doi: 10.1016/s0377-2217(96)00099-9 . 655 URL http://dx.doi.org/10.1016/S0377-2217(96)00099-9 [65] R. Karki, J. Patel, Reliability assessment of a wind power delivery system, Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability 223 (1) (2008) 5158. doi:10.1243/1748006xjrr218 . URL http://dx.doi.org/10.1243/1748006XJRR218 [66] D. R. Cox, Regression Models and Life-Tables, Springer New York, New York, NY, 1992, pp. 527541. doi:10.1007/ 660 978-1-4612-4380-9_37 . URL https://doi.org/10.1007/978-1-4612-4380-9_37 [67] V. Zille, C. Bérenguer, A. Grall, A. Despujols, Modelling multicomponent systems to quantify reliability centred maintenance strategies, Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability 225 (2) (2011) 141160. doi:10.1177/1748006X11402269 . 665 [68] D. Salazar, C. M. Rocco, B. J. Galván, Optimization of constrained multiple-objective reliability problems using evolutionary algorithms, Reliability Engineering & System Safety 91 (9) (2006) 10571070. doi:10.1016/j.ress. 2005.11.040 . [69] M. Martin-Tretton, M. Reha, M. Drunsic, M. Keim, Data collection for current us wind energy projects: Component costs, nancing, operations, and maintenance, Contract 303 (2012) 2753000. doi:10.2172/1034220 . 670 URL https://www.osti.gov/biblio/1034220 16