Wind turbine replacement schedule

Abstract

A method and an arrangement for estimating a replacement schedule of a component, in particular a mechanical component, of a wind turbine is described, the method comprising estimating a remaining producible energy until the component is replaced (5, 11, 13, 15).

Description

TECHNICAL FIELD

[0001] The present invention relates to a method of estimating a replacement schedule of a component of a wind turbine, to a method of operating a wind turbine, and also to a device for estimating a replacement schedule of a component of a wind turbine. BACKGROUND

[0002] A wind turbine comprises several components which are prone to wear due to operation, or even to failure or damage. In particular, mechanical components such as bearings, in particular of the main rotating shaft, can be subject to wear and eventually fail. Therefore, when controlling the operation of a wind turbine, parameters related to the (health) state of one or more wind turbines or one or more wind turbine components have to be taken into account. These parameters can include temperature, noise, vibrations and other state-related measurements, e.g. performed directly at the component. An analysis of these measurements can allow an operator to assess at which point in time the considered component reaches a critical state and has to be shut down. When this happens, a maintenance task needs to be performed.

[0003] However, the detection of a future time at which a wind turbine or a wind turbine component needs to be shut down is generally not solved. Conventional approaches predict the remaining useful life (RUL) of the turbine or of its components. However, this RUL only provides a time scale for the developing critical state. It has been observed, however, that the prediction of the remaining useful life does not in all cases allow for an appropriate scheduling of maintenance or replacement, and does not in all cases allow for an optimized control of the wind turbine.

[0004] Therefore, there can be a need for a method of estimating a replacement schedule of a component of a wind turbine, and possibly correspondingly for a corresponding device for estimating a replacement schedule of a component of a wind turbine, wherein at least partly the drawbacks or problems associated with conventionally known methods and devices are overcome. SUMMARY

[0005] This need can be met by the subject-matter according to the independent claims. Advantageous embodiments of the present invention are described by the dependent claims.

[0006] According to an embodiment of the present invention, a method of estimating a replacement schedule of a component, in particular a mechanical component, of a wind turbine is provided, the method comprising estimating a remaining producible energy until the component will be replaced, also referred to as remaining generated power, RGP.

[0007] The method can be performed, for example, by a module or a part of a wind turbine controller or any processor comprised within the wind turbine.

[0008] According to other embodiments of the present application, the method can be performed by a processor or computer located outside the wind turbine or not belonging to a part of the wind turbine, e.g. in a monitoring office, in a wind farm controller or running in the cloud.

[0009] The component can be a mechanical component, e.g. a bearing (part of) of a shaft, in particular of a main rotation shaft on which a plurality of wind turbine blades are mounted and which can further be coupled with a generator. The main rotation shaft can be attached (e.g. by a hub) on one side with the rotor blades and at the other end it can be directly coupled with the generator or it can be coupled with an optional gearbox.

[0010] Another example of the component can be a bearing of a gearbox or of a part of a gearbox or of a counter shaft, e.g. connected at one end to the gearbox and at the other end coupled with the generator. As yet another example of the component, there can be another bearing in order to allow a yaw rotation of the wind turbine nacelle on top of the wind turbine tower. Another example of the component can be a bearing of a variable pitch system for pitching the individual rotor blades by rotating them around a longitudinal axis of the rotor blades. In other embodiments, the component can comprise electrical and / or electronic components as well, e.g.

[0011] The method can be started when it is observed that the component under consideration shows signs of damage or failure. The method can allow to plan for a replacement of the component in the future. However, unlike conventionally known methods, the method does not (mainly) output or estimate the remaining useful life (RUL) but can output or estimate the remaining energy producible until the component will be replaced. The remaining energy producible until the component will be replaced is an estimate of the energy (e.g. integral of the output power) that can still be produced before the component is so severely damaged that it needs to be replaced. The remaining energy producible estimated by the method can also be referred to as remaining generated power (RGP) in the present application.

[0012] The estimation can involve the application of machine learning techniques or nonlinear estimators.

[0013] As the name suggests, the conventionally estimated remaining useful life only provides a time scale for the critical state in the development of the component of the wind turbine. However, this time scale is not a physical parameter in the operation of the wind turbine and has no value in terms of power output. For example, according to one scenario, during operation, a critical state of a rotating part is detected according to a conventional method that will necessarily occur within 30 days. However, it turns out that the next 20 days are windless and the effective remaining useful life of the turbine should obviously have been estimated to be 50 days. Thus, the conventional method estimates a wrong date or replacement time of the component. As a result, the wind turbine can not be operated to optimize the power output or to optimize other criteria.

[0014] The present invention proposes a new measure, which is related to the operation and energy or power production of a wind turbine, instead of the remaining useful lifetime. In particular, the estimated remaining producible energy can be continuously monitored and used for determining a control scheme for operation and maintenance scheduling.

[0015] According to embodiments of the invention, the method further comprises detecting a value of a physical quantity indicative of a failure of the component, and starting the estimating of the remaining producible energy, if the value exceeds a threshold value, in particular the method further comprises estimating the remaining producible energy further based on the detected value of the quantity.

[0016] The value of the physical quantity can be measured or can be estimated or inferred from other measured or inferred or estimated quantities. The physical quantity can for example comprise a temperature, a vibration, a noise, any electrical quantity related to the power output of the wind turbine, a rotational speed of a main or a secondary rotational shaft, a friction, etc. The physical quantity can differ depending on the component considered. When it is determined that the value of the physical quantity exceeds a threshold value, it can be an indication that the component considered suffers from partial damage or has problems in terms of functionality.

[0017] For detecting the value of the physical quantity, one or more sensors can be provided, for example microphones, acceleration sensors, rotational speed sensors, electrical measuring devices, temperature sensors, etc. It is also considered that the detected value of the quantity used for estimating the remaining producible energy can improve the estimation. In particular, the higher the value of the physical quantity, the higher the degree of damage of the component. Before the value of the physical quantity exceeds a threshold value, the method of estimating can not be performed, because it can be assumed that the component is in a proper health condition. Thereby, it can be avoided that the method involving computational effort is unnecessarily performed.

[0018] According to embodiments of the invention, the physical quantity indicative of a failure of the component comprises a vibration and / or a noise, in particular measured by a sensor close to the component, wherein the component in particular comprises a bearing, in particular a bearing of a main rotational shaft.

[0019] The vibration or noise can for example be measured by a microphone or in an accelerometer. This can be particularly useful when the component comprises a bearing. The degree of damage of the bearing can be properly monitored by monitoring the vibration of the component or the noise produced by the component. In other embodiments, the estimating of the remaining producible energy can not consider the detected value of the quantity indicative of a failure of the component as an input. In some cases, there can be no physical quantity available which is indicative of a failure of the component.

[0020] According to embodiments of the invention, the estimating of the remaining producible energy is performed during the electricity production of the wind turbine.

[0021] When the method is performed during power production of the wind turbine, i.e. during normal operation of the wind turbine, the operation of the wind turbine can be adjusted in dependence of the estimated remaining producible energy. Thereby, the power output can be improved and / or the load experienced by the components of the wind turbine can be optimized, e.g. minimized.

[0022] According to embodiments of the present application, estimating the remaining producible energy uses machine learning, comprising: using a neural network, in particular a recurrent neural network, that has been trained with training data from a plurality of, e.g. between 50 and 1000, training wind turbines, the training data comprising training values of input parameters and training values of at least one output parameter, the input parameters comprising at least one wind turbine operation parameter and / or at least one environmental condition parameter; the at least one output parameter being indicative of the remaining producible energy until the component will be replaced.

[0023] The neural network can comprise a plurality of network nodes and links between these nodes. Neural networks are known to the skilled person. The neural network can comprise an input layer, one or more hidden layers and an output layer. A recurrent neural network (RNN) is a class of artificial neural network where connections between nodes form a directed graph along time. This can allow temporal dynamic behavior to be exhibited. RNNs can use their internal state (memory) to process sequences of inputs. The term "recurrent neural network" is used to refer to two broad classes of networks with similar overall structure, one of which is a finite impulse, and the other of which is an infinite impulse. Both classes of networks exhibit temporal dynamic behavior. Finite impulse recurrent networks are directed acyclic graphs that can be unfolded and replaced with a strict feed-forward neural network, while infinite impulse recurrent networks are directed cyclic graphs that cannot be unfolded.

[0024] Thereby, a conventionally available neural network can be utilized, thereby simplifying the implementation of the method.

[0025] The training wind turbines can comprise the same or similar components with the same or similar faults as the wind turbine under consideration. For example, the components of the training wind turbines can be of the same type, model, size configuration, age as the components of the wind turbine under consideration. Thereby, the accuracy of the method can be improved. When there are the same or similar components in the training wind turbines and in the wind turbine under consideration, it can be expected that the behavior or evolution of the components' constitution during further operation is similar or even identical.

[0026] According to embodiments of the present application, estimating the remaining producible energy comprises using the trained neural network, comprising: providing test values of input parameters to the trained neural network, the test values relating to the wind turbine; outputting, by the trained neural network, a probability distribution (e.g. parameterized thereof) of the remaining producible energy relating to the wind turbine.

[0027] The probability distribution can be represented in different ways. For example, a parameterization of the probability distribution can be output. Thus, the estimation method can estimate parameters of the probability distribution. The probability distribution can for example be described by a sum of basis functions weighted by coefficients. The coefficients can be obtained by the estimation method. A basis function is an element of a subset (basis) of functions in the space of all functions mapping between a fixed set.

[0028] Whenever the neural network is trained, only the test value of the at least one input parameter has to be provided to the network. Then, the neural network outputs a probability distribution of the remaining producible energy as resulting from the test value of the at least one input parameter. The probability distribution can indicate for each energy value how probable it is that the energy value equals or corresponds to the remaining producible energy. When outputting a probability distribution instead of a single estimated remaining producible energy, it is made possible to generate a plurality of different statistical parameters from the probability distribution, such as mean, median, quantile, etc. Different statistical parameters can be utilized depending on the application.

[0029] According to embodiments of the present invention, a probability distribution is given for a plurality of time points in the future. Thereby, the replacement scheduling can be performed in an improved way.

[0030] According to embodiments of the present invention, one of the following is used as an estimate of the remaining producible energy at a desired time point, namely: the mean at the desired time point; the median; the mode; the probability density function.

[0031] The mode is a term in statistics denoting a measure of the most common value in a set of data, i.e. the position at which the probability distribution takes its maximum value. The probability density function is a function which provides the relative likelihood of a value at any given sample of a random variable.

[0032] Thereby, different statistical factors or measures can be derived from the probability distribution, and a particular statistical factor or parameter can be selected based on a particular application.

[0033] According to embodiments of the present invention, training the neural network comprises weighting the input parameters according to their importance for reaching the respective training values of the output parameter from the training data by applying a cost function. The cost function can for example involve the neural network weights to be adjusted, the parameterization of the probability distribution, the turbine state, the input data and the remaining power production. The combination of parameters and their importance varies with the underlying statistical model defining the cost function.

[0034] The neural network can involve the weights to be adjusted, the parameterization of the probability distribution, the turbine state, the input data and the remaining power production. The combination of parameters and their importance can vary with the underlying statistical model defining the cost function.

[0035] According to embodiments of the present application, the at least one wind turbine operating parameter comprises at least one of: power output; rotational speed of a rotational shaft on which a plurality of rotor blades are mounted; torque of a generator; energy diffusion; constructional properties of a component.

[0036] Energy diffusion is an exchange of energy from a high energy area to a low energy area. In this particular case, energy diffusion can involve a heat transfer (from hot to cold) inside and outside the wind turbine as indicated by temperature measurements.

[0037] Thereby, conventionally available wind turbine operating parameters can be supported. The operational behavior of a component can depend on one or more of these wind turbine operating parameters. Thereby, an estimation of the remaining producible energy can be achieved and improved.

[0038] According to embodiments of the present application, the at least one environmental condition parameter comprises at least one of: wind speed; wind turbulence; humidity; temperature.

[0039] Moreover, at least one of the environmental condition parameters can also influence the operational behavior of a component. As such, at least one of the environmental condition parameters can thus enable to infer the operational behavior of a component from this environmental condition parameter.

[0040] According to embodiments of the present application, the method further comprises: indicating a replacement of the component if the remaining producible energy is less than an energy threshold, in particular substantially zero.

[0041] In particular, a point in time at which a component can be replaced can be indicated, wherein the point in time at which a replacement is suggested can be derived from the remaining producible energy. Thus, the point in time at which a component is replaced can be a quantity which can be derived from the initially estimated remaining producible energy, e.g. by an extrapolation method.

[0042] According to embodiments of the present application, a method of operating a wind turbine is provided, comprising: performing the method according to one of the preceding embodiments; operating the wind turbine so as to optimize a power output based on the estimated remaining producible energy.

[0043] It is considered that the estimated remaining producible energy can enable to optimize the operation of a wind turbine, e.g. for optimizing a power output and / or improving a load or a combination thereof.

[0044] According to embodiments of the present application, a device for estimating a replacement schedule of a component, in particular a mechanical component, of a wind turbine, in particular a replacement schedule of a wind turbine, is provided, the device being adapted to control or perform the method according to any of the preceding embodiments.

[0045] It should be understood that features which have been described, explained or provided separately or in any combination for the method of estimating a replacement schedule for a component of a wind turbine can also be applied to the device for estimating a replacement schedule for a component of a wind turbine according to embodiments of the present application, either separately or in any combination, and vice versa.

[0046] The above defined aspects of the present application, as well as further aspects, will be apparent from the examples of embodiments to be described hereinafter and will be explained with reference to these examples of embodiments. The present application will be described in more detail hereinafter with reference to examples of embodiments but the present application is not limited to these examples of embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 schematically illustrates a neural network as used in a method of estimating a replacement schedule according to embodiments of the present application;

[0048] Figure 2 illustrates a plot representing a probability distribution of the remaining producible energy as obtained according to embodiments of the present application. DETAILED DESCRIPTION

[0049] According to one embodiment, at least one failure state of at least one component of a wind turbine, e.g. a bearing, is detected. After the failure state is identified, e.g. using values of vibrations or noise as measured by microphones or accelerometers, the remaining producible energy, also referred to as the remaining generation of power (RGP), is estimated. In this context, the RGP can be considered as a distribution of remaining kWh at a given instance of time until the component fails. In contrast to the commonly considered remaining useful life (RUL), the RGP is independent of the need for continuous operation. Thus, according to embodiments of the present application, the RGP can be directly linked to performance gains and, thereby, to costs involved in operation and maintenance.

[0050] The basic method of estimating the remaining producible energy (RGP) can be based on a deep learning architecture, known as a recurrent neural network (RNN), as schematically illustrated in Figure 1 The network 1 is designed to analyze one or more data sets, i.e. input data 3, to estimate the remaining producible energy of a failure, which is output by the neural network 1 as a probability distribution 5. In case of a main bearing, which supports the main rotating shaft of a wind turbine, the input data 3 can be defined on a feature set from wind turbine operational data and environmental condition data, i.e. measurements associated with environmental conditions, energy dissipation and rotating / movable path, and further other health indicators, e.g. vibration-based indicators, can be considered as well.

[0051] The output space, i.e. the output 5, can be defined as a parameterization of the tractable distribution on the RGP (satisfying the underlying statistical model) enabling to express the RGP in terms of probabilities rather than point measures. The training objective of such an RNN can be defined by a cost function and expressed in terms of the RGP of the trained wind turbine.

[0052] The cost function can be a function depending on the wind turbine state, the remaining power production (remaining producible energy) and the parameterization of the probability distribution, where the last mentioned implicit dependence on the neural network weights. When training the neural network, i.e. adjusting the weights, the cost function will provide how to adjust to comply with the properties described by the cost function, e.g. error minimization.

[0053] All training can be done on empirical and / or historical data. Such data can come from other turbines or components of the same or similar type that experienced the same or similar failure. The power production yield (probability distribution of the remaining producible energy) is output by the network 1 at the output layer as output 5.

[0054] An example of an estimated probability distribution of the remaining producible energy is shown in Figure 2 a graph with an abscissa 7 representing the days to failure or recommended replacement and with an ordinate 9 on the left representing the remaining producible energy in arbitrary units. The probability distribution of the remaining producible energy is illustrated as a shaded area and labeled with reference numeral 11. From this probability distribution 11, the median is illustrated as a curve 13 and the mode is represented as a curve 15. The true remaining producible energy is illustrated as a curve 17.

[0055] A curve 19 represents the measured vibrations, where the ordinate 21 on the right represents the vibration level in arbitrary units.

[0056] At a point in time 23 (120 days to failure or recommended replacement), it is detected that the main bearing suffers from mechanical problems or partial damage based on the vibration curve 19, since the vibration curve 19 is above a threshold 24. From that time on, the method of estimating the replacement schedule according to embodiments of the invention is started using the neural network shown in Figure 1

[0057] The higher values of the vibration curve 19 represent a higher failure risk. The method provides a distribution over the remaining producible energy, i.e. each energy is actually a probability of the remaining producible energy. This distribution is provided for each instance in time. The production yield can be retrieved from the median or mode or set density shown as curve 13 or 15.

[0058] ​When the estimated remaining energy that can be generated is essentially zero, the component will be replaced, which can be indicated, for example, at time point 25 (0 days before failure).

[0059] Then, the control scheme according to embodiments of the invention can adjust the power output to optimize operation constrained by maintenance schedules. Wind forecasting allows for the estimation of revenue from electricity production during periods of inactivity. This can be done without losing the generality of the RGP framework, compared to a fixed time range (estimated quantity RUL). Therefore, the RGP more accurately reflects the time and costs associated with wind turbine operation. Simultaneously, the RGP can also be used as a factor in addressing seasonal high and low flow conditions.

[0060] As from Figure 2 As can be seen, the actual remaining reproducible energy, according to curve 17, decreases monotonically from the first indication of failure (time point 23) to the actual time point 25 when the component is replaced. Curves 13 and 15, derived from the probability distribution of the estimated remaining reproducible energy, are higher than the actual remaining reproducible energy 17 until approximately 60 days before failure. After this time point, the median and mode of the probability distributions 13 and 15, respectively, approach the actual remaining reproducible energy (curve 17) and are slightly lower than it. However, the estimated remaining reproducible energy increasingly matches the actual remaining reproducible energy, demonstrating the good reliability and accuracy of the method.

[0061] It should be noted that the term "comprising" does not exclude other elements or steps, and the wording "a," "an," or "a kind" does not exclude multiple. Furthermore, elements described in different embodiments may be combined. It should also be noted that the reference numerals in the claims should not be construed as limiting the scope of the claims.

Claims

1. A method of estimating a replacement schedule of a component of a wind turbine, the method comprising: detecting a value of a physical quantity indicative of a failure of the component of the wind turbine; estimating a remaining producible energy until the component is replaced (5, 11, 13, 15), wherein estimating the remaining producible energy (5, 11, 13, 15) is started if the value of the physical quantity exceeds a threshold value (24), wherein estimating the remaining producible energy (5, 11, 13, 15) uses machine learning, comprising: using a neural network (1) that has been trained with training data from a plurality of training wind turbines, the training data comprising training values of input parameters and training values of at least one output parameter, the input parameters comprising at least one wind turbine operating parameter and / or at least one environmental condition parameter; the at least one output parameter being indicative of the remaining producible energy until the component is replaced.

2. The method according to claim 1, further comprising: estimating the remaining producible energy further based on the detected value of the physical quantity.

3. The method of claim 1, wherein, the physical quantity indicative of the failure of the component comprises a vibration.

4. The method of claim 3, wherein, the vibration is measured by a sensor close to the component.

5. The method of claim 3, wherein, the component comprises a bearing.

6. The method of claim 5, wherein, the bearing is a bearing of a main rotating shaft.

7. The method of any one of claims 1-6, wherein, estimating the remaining producible energy (5, 11, 13, 15) is performed during power production of the wind turbine.

8. The method according to claim 1, the training wind turbines comprising identical or similar components with identical or similar failures.

9. The method of claim 1, wherein, estimating the remaining producible energy comprises using a neural network, comprising: providing test values of the input parameters to the trained neural network, the test values relating to the wind turbine; outputting by the trained neural network a probability distribution of the remaining producible energy relating to the wind turbine.

10. The method of claim 9, wherein, giving the probability distribution for a plurality of future points in time.

11. The method of any one of claims 9 and 10, wherein, using one of the following as an estimate of the remaining producible energy (5, 11, 13, 15) at a desired point in time: a mean of the probability distribution at the desired point in time; a median; a mode; a set density.

12. The method of claim 1, wherein, training the neural network comprises: weighting the input parameters according to their importance for arriving at the respective training values of the output parameter by applying a cost function.

13. The method of any one of claims 1-6, wherein, the at least one wind turbine operating parameter comprises at least one of: a power output; a rotational speed of a rotating shaft on which a plurality of rotor blades are mounted; a torque of a generator; an energy spread; a construction property of the component.

14. The method of any one of claims 1-6, wherein, the at least one environmental condition parameter comprises at least one of: a wind speed; a wind turbulence; a humidity; a temperature.

15. The method according to any of claims 1-6, further comprising: indicating to replace the component if the remaining producible energy is less than an energy threshold value.

16. The method according to any of claims 1-6, further comprising: indicating to replace the component if the remaining producible energy is zero.

17. The method of claim 1, wherein, the component of the wind turbine is a mechanical component.

18. The method of claim 1, wherein, the neural network is a recurrent neural network.

19. A method of operating a wind turbine, comprising: performing the method according to any of the preceding claims; operating the wind turbine so as to optimize power output based on the estimated remaining producible energy.

20. An apparatus for estimating a replacement schedule of a component of a wind turbine, the apparatus being adapted to control or perform the method according to any of the preceding claims.

21. The apparatus of claim 20, wherein, the component of the wind turbine is a mechanical component.

21. A computer program product comprising computer readable instructions which, when executed by a computer, cause the computer to perform the method according to any of the preceding claims.

22. A computer program product comprising computer readable instructions which, when executed by a computer, cause the computer to perform the method according to any of the preceding claims.

23. A computer program product comprising computer readable instructions which, when executed by a computer, cause the computer to perform the method according to any of the preceding claims.

24. A computer program product comprising computer readable instructions which, when executed by a computer, cause the computer to perform the method according to any of the preceding claims.

25. A computer program product comprising computer readable instructions which, when executed by a computer, cause the computer to perform the method

Images