Method for training a machine learning model to determine the remaining useful life of another wind turbine

Description

[0001] The application relates to a method for training a machine learning model, particularly usable for determining, in particular estimating, a system condition of at least one component of a wind turbine, particularly preferably usable for determining, in particular estimating, a remaining service life of the wind turbine. The application further relates to a method for using a machine learning model and a computing device.

[0002] Electrical energy is increasingly being generated using renewable energy sources. For example, (onshore and / or offshore) wind farms with at least one wind turbine are being installed. A wind turbine is designed to convert wind energy, i.e., the kinetic energy of moving air, into electrical energy.

[0003] Typically, a wind turbine consists of a tower and a nacelle mounted on the tower. Wind energy is converted into electrical energy by means of a rotor, a generator, and the like.

[0004] A wind turbine is exposed to high (mechanical) loads after installation and during operation. These loads lead to a finite technical service life of every wind turbine. This is primarily due to fatigue phenomena in certain (non-repairable or non-replaceable) components of a wind turbine. These components are also referred to as load-bearing structural components (e.g., tower, nacelle, foundation, etc.).

[0005] In practice, such structural components cannot be replaced or repaired with spare parts. This means, in particular, that replacing or repairing such components is not carried out in practice for cost reasons, even if it would be theoretically possible, compared to a new installation.

[0006] The basic objective is to operate a wind turbine, if possible, until the end of its actual technical service life. The end of the technical service life is, in particular, the end of the safe operating time of a wind turbine. The actual service life refers in particular to the point in time at which the actual condition of the wind turbine reaches, falls below, or exceeds a certain condition limit. The certain condition limit can be predetermined. The certain condition limit can, in particular, indicate a condition of the wind turbine (in particular of one of the structural components of the wind turbine) beyond which the risk of damage (in particular fatigue damage) to one of the non-repairable or non-replaceable components of the wind turbine is no longer acceptable.

[0007] In other words, the operation of a wind turbine may no longer be permitted if at least one specific turbine condition limit value is reached.

[0008] Therefore, with the ever-increasing number of wind turbines and wind farms, there is a fundamental need to determine, and in particular, predict, the technical service life, also known as the "Remaining Useful Life" (RUL), of an operating wind turbine with the highest possible accuracy. The RUL of wind turbines is an important key figure in the evaluation of a wind farm.

[0009] The RUL can be determined by the fatigue damage introduced into the structure of the wind turbine, i.e. the at least one structural component of the wind turbine, up to the time of assessment (cf. current turbine condition) in comparison to the permissible damage (for example defined by the at least one turbine condition limit value of the at least one structural component).

[0010] It is understood that the permissible damage may include appropriate safety factors. In other words, the actual system condition of at least one structural component can be determined in comparison to a permissible system condition, i.e., compared with at least one specific system condition limit.

[0011] It is known from the prior art to calculate the damage to a wind turbine from the wind turbine loads or stresses using a computer system based on a suitable damage model for the wind turbine. In particular, a model can be used for each component or structural component under consideration. The most accurate approach would therefore be to continuously measure the component loads on all wind turbines over the entire service life of the respective wind turbine and then make a prediction based on the damage model. Due to the enormous costs associated with this approach, which would require performing such measurements on each wind turbine over a period of at least three months, preferably at least six months, this approach is simply not feasible in practice.

[0012] Therefore, state-of-the-art simulation models are used to estimate the RUL of a wind turbine. For example, so-called aeroelastic simulation models are used to estimate the component loads of a wind turbine. A simulation model specifically represents a specific turbine type. However, this representation is associated with considerable effort. The simulation models are then calibrated and validated using available operating data from a wind turbine, such as the power curve, thrust curve, foundation loads under design conditions, etc.

[0013] Load measurements during operation of a wind turbine are generally not performed. The uncertainty of such a simulation model is therefore usually very high. Due to the high uncertainty, correspondingly high safety factors are taken into account, resulting in very conservative RUL results. However, with the current state of technology, the high uncertainty and the already high effort involved can only be improved by investing considerable resources.

[0014] A rarely used option in practice to at least somewhat reduce the uncertainty is based on an estimation of the loads acting on a wind turbine using operating data from the wind turbine, in particular in the form of SCADA (Supervisory Control and Data Acquisition) operating data from the wind turbine.

[0015] The problem with a damage accumulation estimation process based on deriving loads from SCADA operating data is that it is highly non-linear. This problem regularly leads to highly unreliable results in existing simulation models.

[0016] In summary, it can be stated that the estimation of the remaining service life of a wind turbine is very complex and, at the same time, involves a high degree of uncertainty.

[0017] US 2021 / 123416 A1 is a relevant example of the prior art.

[0018] Therefore, the application is based on the object of providing a possibility for determining the remaining useful life of a wind turbine, which enables a more reliable determination of the remaining useful life with, in particular, less effort.

[0019] According to a first aspect of the application, the object is achieved by a method according to claim 1, in particular a computer-implemented method, for training a machine learning model used to determine a remaining service life of another wind turbine. The method comprises: Providing a plurality of (time-dependent) operating data sets of a reference wind turbine, providing a plurality of (time-dependent) load data sets of the reference wind turbine, wherein a load data set is based on at least one load parameter measured on the reference wind turbine, and generating a plurality of wind turbine training data sets for training a machine learning model by synchronously assigning a respective operating data set to a respective load data set.

[0020] According to a further aspect of the application, the object is achieved by a method according to claim 12 for using a machine learning model trained according to the described method. The method comprises: Inputting at least one operating data set of the further wind turbine into the (trained) machine learning model, and outputting, by the (trained) machine learning model, at least one turbine status data set.

[0021] The method is used in particular to determine the current and, in particular, actual (structural and mechanical) condition of at least one structural component of the additional wind turbine. Based on this (at least one) condition, a remaining service life of the additional wind turbine can then be determined, in particular predicted with increased accuracy.

[0022] In contrast to the prior art, the application provides a method for determining the current state of a wind turbine and, based thereon, the remaining service life of the wind turbine. This method enables a more reliable determination of the remaining service life with, in particular, less effort by training a machine learning model with training data sets, each of which contains synchronized operating data and load data from a reference wind turbine. The trained machine learning model can then be used to determine the state of another wind turbine. Based on this determined state, the remaining service life of the other wind turbine can then be determined (for example, in a conventional manner).

[0023] According to the application, a method for training a machine learning model is provided. In other words, a machine learning model can be trained using the method according to the application. The trained machine learning model can then be used, in particular indirectly, to determine the remaining service life of a wind turbine.

[0024] The machine learning model (also referred to as machine learning model or machine learning model) may preferably be and / or comprise an artificial neural network.

[0025] An artificial neural network comprises artificial neurons. Such a neuron is particularly configured to receive at least one input from at least two other artificial neurons. The at least two inputs can each be received via at least one connection with a specific (trainable) weighting.

[0026] An artificial neuron may further be configured to integrate the at least one input into at least one output, for example, by summing at least two received inputs and applying a function, such as a sigmoid transformation, to the sum to generate an output. The output of an artificial neuron, in particular multiplied by a connection weight of a further connection to another neuron, may form an input for this further neuron.

[0027] According to the application, a plurality of operating data sets of this at least one reference wind turbine and a plurality of (mechanical and / or structural) load data sets of this at least one reference wind turbine are provided for at least one reference wind turbine.

[0028] As already described, the current system condition of a structural component can be determined, in particular estimated, from load parameters or the load parameter values ​​measured on a wind turbine during a measurement period. However, determining the current system condition based solely on the operating data (preferably SCADA operating data) acquired during a measurement period is not possible according to the state of the art.

[0029] According to the application, it has been recognized that there is a fundamental relationship between the load parameter values ​​and the operating data values. In particular, it has been recognized that, using special training data sets containing both load parameter values ​​and operating data values ​​of a (reference) wind turbine, a machine learning model can be trained in such a way that this relationship or relationships can be detected and mapped in the machine learning model. In particular, a machine learning model can recognize patterns between load parameter values ​​and operating data values. These can be mapped in the machine learning model.

[0030] This subsequently makes it possible to determine, in particular estimate, the remaining service life of another wind turbine (at least indirectly) using (exclusively) the operating data of this other wind turbine. Thus, at least one operating data set, preferably a plurality of operating data sets, of another wind turbine can be input into the machine learning model, i.e., in particular, made available as input(s). As output, at least one turbine condition data set can be provided by the machine learning model, which may already contain an (estimated) remaining service life or at least one turbine condition indicator (e.g., at least one fatigue load indicator) from which the remaining service life can be determined, in particular, estimated in a conventional manner.

[0031] As already described, a plurality of operating data sets and a plurality of load data sets of a reference wind turbine are provided to create a plurality of training data sets. A reference wind turbine refers in particular to an operating wind turbine, particularly of a wind farm, whose operating parameter values ​​and load parameter values ​​are recorded to create training data sets. A reference wind turbine can be selected in any manner from a plurality of wind turbines of a wind farm.

[0032] An operating data record contains at least one operation-related parameter value of the reference wind turbine. An operating data record preferably contains a plurality of these operating parameters (values). An operating parameter refers, in particular, to an operating parameter of the wind turbine recorded during operation of a wind turbine (for the operation of the wind turbine) that relates to the operation of the wind turbine. Preferably, a provided operating data record of the reference wind turbine is a SCADA operating data record of the reference wind turbine or an operating data record of the reference wind turbine with at least one SCADA operating parameter of the reference wind turbine.

[0033] A load data set, in particular in the form of a fatigue load data set, contains at least one structural and / or mechanical load parameter or load parameter value of a (previously described) structural component of the reference wind turbine.

[0034] In particular, during the operation of a reference wind turbine, load parameter values ​​of at least one load parameter of the reference wind turbine and operating parameter values ​​of at least one operating parameter of the reference wind turbine are recorded simultaneously.

[0035] A plurality of time-dependent operating data records can then be formed from the recorded operating parameter values ​​of the at least one operating parameter, in particular in a conventional manner. Each operating data record can in particular be assigned an operating data duration (e.g., a timestamp) that relates to the measurement period of the measurement or recording of the at least one operating parameter. For example, an operating data duration can comprise a start time (formed in particular by the date and time (e.g., 1 / 1 / 21, 1:00 a.m.)), a time length (preferably between 5 and 30 minutes, in particular (in a conventional manner) 10 minutes) and an end time (formed in particular by the date and time (e.g., 1 / 1 / 21, 1:10 a.m.)).

[0036] A plurality of time-dependent load data sets can then be created from the recorded load parameter values ​​of the at least one load parameter. In particular, to enable a (temporally) synchronized assignment of operating data sets to load data sets, the creation of load data sets can preferably be based on or correspond to the creation of the operating data sets (e.g., already specified by a SCADA system).

[0037] In particular, each load data record can be assigned a load data duration (e.g., a timestamp) that relates to the measurement period of the measurement or recording of the at least one load parameter. This load data duration can be selected according to the operating data duration. For example, a load data duration can comprise a start time (formed in particular by the date and time (e.g., 1 / 1 / 21, 1:00 a.m.)), a time length (preferably between 5 and 30 minutes, in particular (conventionally) 10 minutes), and an end time (formed in particular by the date and time (e.g., 1 / 1 / 21, 1:10 a.m.)).

[0038] To generate wind power training datasets, according to the application, load datasets are assigned to operating datasets in a synchronous or synchronized manner, in particular depending on the respective assigned load data period or operating data period. In other words, a temporally synchronous or synchronized assignment of temporally synchronous datasets, i.e., datasets with the same time dependency (e.g., the same time stamp), to one another can be performed.

[0039] For example, an operating data record with a start time t A1 (e.g. 1.1.21, 1:00 a.m.), time length T L1 (e.g. 10 min) and an end time point t E1 (e.g. 1.1.21, 1:10 a.m.) is assigned a load data record with a corresponding start time t A1 (e.g. 1.1.21, 1:00 a.m.), a corresponding time length T L1 (e.g. 10 min) and a corresponding end time point t E1 (e.g. 1.1.21, 1:10 a.m.).

[0040] The synchronous assignment according to the application creates, in particular, a wind power training data set which (always) contains (exactly) one load data set (or the corresponding parameter values) and (exactly) one operating data set (or the corresponding parameter values) or is created from them.

[0041] According to one embodiment of the method according to the application, the machine learning model can be trained (in a conventional manner) with the generated wind power training data sets during a training period. In other words, the generated wind power training data sets are provided to the machine learning model, in particular for training purposes.

[0042] According to a further embodiment of the method according to the application, a portion (for example, between 10% and 30%, preferably between 15% and 25%) of the generated wind power training data sets can be used as validation data sets during a training period. In particular, the training success of a machine learning model can be checked, i.e., validated, using validation data sets. The validation data sets can preferably be randomly selected from the generated wind power training data sets.

[0043] According to a preferred embodiment of the method according to the application, a provided operating data set can contain at least one operating parameter (or a recorded operating parameter value) of the reference wind turbine. The at least one operating parameter can be selected from the group comprising: Tower top acceleration (also called tower top acceleration), pitch angle (also called pitch angle), pitch speed (in particular the speed at which the blade is adjusted), rotor speed (also called rotor speed), electrical power, in particular electrical active power (also called active power), nacelle wind speed(s).

[0044] The operating data set can comprise only one of the aforementioned operating parameters, preferably two of the aforementioned operating parameters, particularly preferably all of these operating parameters. It is understood that additional operating parameters not explicitly mentioned here can also be provided alternatively or additionally. In particular, at least one sensor module, in particular a plurality of (different) sensor modules, can be provided to measure the at least one operating parameter or its operating parameter values ​​(preferably continuously). The measured operating parameter values ​​and / or signals of an operating parameter can then be provided, for example, by a SCADA control module.

[0045] Preferably, an operating data record has an operating data duration as previously described.

[0046] According to a further embodiment of the method according to the application, an operating data record can comprise, as an operating parameter value of an operating parameter, at least one operating parameter value selected from the group comprising: maximum operating parameter value recorded during (the time length) of the operating data period of the operating data set, minimum operating parameter value recorded during (the time length) of the operating data period of the operating data set, operating parameter mean value determined from the operating parameter values ​​recorded during (the time length) of the operating data period of the operating data set, standard deviation determined from the operating parameter values ​​recorded during (the time length) of the operating data period of the operating data set.

[0047] An operating data set can comprise two of the aforementioned operating parameter values ​​for the at least one operating parameter, preferably all of these operating parameter values. This allows higher-quality training data sets to be created.

[0048] A preferred example of an operational data set according to the present application is given in Table 1. Table 1 Operating parameters Operating parameter value Turret head acceleration - max. acceleration value - min. acceleration value - averaged acceleration value - Standard deviation of acceleration values Pitch angle - max. angle value - min. angle value - averaged angle value - Standard deviation of the angle values Rotor speed - max. speed value - min. speed value - average speed value - Standard deviation of the speed values Electrical power, - max. power value especially electrical - min. power value Active power - averaged performance value - Standard deviation of performance values Nacelle wind speed - max. speed value - min. speed value - average speed value - Standard deviation of the speed values

[0049] The preferred operating parameters described above refer in particular to onshore wind turbines.

[0050] According to the application, it has been recognized that in the case of an offshore wind turbine, an operating data record contains at least one further (offshore-related)

[0051] Operating parameters should include, or a further operating data set with at least one further (offshore-related) operating parameter should be provided. In particular, according to the application, it has been recognized that, for an offshore wind turbine, an operating data set (or another offshore operating data set that can be additionally assigned according to the described synchronous assignment) can contain at least one water status parameter. A water status parameter means, in particular, a parameter that indicates the status of the body of water surrounding the offshore wind turbine (in particular, a sea).

[0052] The at least one water body condition parameter is preferably a hydrodynamic parameter. The at least one water body condition parameter can be a wave parameter (e.g., wave height, wave direction, and / or wave frequency) and / or a flow parameter (e.g., flow velocity and / or flow direction).

[0053] An example of an (offshore) operational data set according to the present application with offshore-related operational parameters is given in Table 2. Table 2 Operating parameters Operating parameter value Wave height - max. altitude value - min. altitude value - averaged altitude value - Standard deviation of the altitude values Wave frequency - max. frequency value - min. frequency value - averaged frequency value - Standard deviation of the frequency value Current velocity and / or wave velocity - max. speed value - min. speed value - average speed value - Standard deviation of the speed values

[0054] It is understood that Table 1 and Table 2 can be combined with each other.

[0055] According to a further preferred embodiment of the method according to the application, the provided operating parameter values ​​and / or load parameter values ​​and / or wind power training data parameter values ​​can be preprocessed. In particular, the quality of the training data sets can be improved if the quality of the aforementioned values ​​is improved (e.g., erroneous values ​​due to measurement errors are identified in advance and then, in particular, remain unconsidered or are replaced by more suitable values ​​using suitable means).

[0056] Pre-processing may in particular include at least one of the following measures: Removal of invalid operating parameter values ​​and / or load parameter values ​​and / or wind turbine training data parameter values ​​(or corresponding outliers). This is preferably done via direct indicators of the wind turbine condition (wind turbine condition signal in SCADA, if available) and / or indirectly via the power signal in SCADA in conjunction with the power curve specification. Extraction of a suitable set of features / parameter values ​​by taking into account collinearity of the features / parameter values. This can be done by calculating the correlations between the parameters and comparing them with a permissible upper bound. Alternatively, all features / parameter values ​​(except for the highly collinear ones) can be retained and a machine learning method with implicit feature selection can be used (e.g., Bayesian LASSO).Trevor Park & ​​George Casella (2008), The Bayesian Lasso, Journal of the American Statistical Association) or MARS). This can be done, for example, using a regression technique such as multivariate regression (cf., e.g., Friedman, JH (1991) "Multivariate Adaptive Regression Splines"). Scaling of the operating parameter values ​​and / or load parameter value and / or wind turbine training data parameter values, e.g., standardizing the operating parameter values ​​with a zero mean and, e.g., constant variance (usually equal to 1).

[0057] Other transformations are also conceivable. The goal is always that the variables are approximately standard normally distributed after the transformation.

[0058] According to a further preferred embodiment of the method according to the application, a provided load data set can be based on at least one measured load parameter of the reference wind turbine. "Based" means, in particular, that the load data set does not directly comprise the (measured) load parameter values ​​of the at least one load parameter, but rather fatigue load indicators or indicator values ​​derived or determined therefrom.

[0059] The at least one load parameter may preferably be selected from the group comprising: Blade root load parameters, rotor load parameters, tower load parameters, tower torsion parameters, tower top moment parameters.

[0060] The load dataset can be based on only one of the aforementioned load parameters. The load dataset can be based on two of the aforementioned load parameters, preferably on all of these load parameters.

[0061] The at least one blade root loading parameter can in particular be a blade bending parameter (also referred to as flap bending) and / or an edge bending parameter (also referred to as edge bending). Furthermore, the at least one rotor loading parameter can be a tilt moment parameter (also referred to as tilt moment), a yaw moment and / or a rotor torque (also referred to as rotor torque). The at least one tower loading parameter can be a tower root bending parameter (in two directions). The at least one tower head moment parameter can be a tilt moment parameter (also referred to as tilt moment), a yaw moment and / or a roll moment parameter (also referred to as roll moment).

[0062] From the load parameters mentioned, a system condition or a component condition of a specific component, in particular a structural component, of the wind turbine can be determined.

[0063] The load parameter values ​​of a load parameter can be (continuously) measured and, in particular, provided by at least one load sensor of a sensor arrangement. In particular, the at least one load sensor can measure with a high resolution (e.g., 20 Hz). Buffering of the measured values ​​for further processing is possible. For this purpose, the at least one load sensor can be arranged at a suitable position in or on the wind turbine.

[0064] According to a particularly preferred embodiment of the method according to the application, the method may further comprise: Measuring at least one load parameter of the reference wind turbine, wherein the measurement of at least one load parameter is carried out in particular in accordance with the standard IEC 61400-13 [2].

[0065] In particular, at least the (key) stress parameters referred to in section 3.3.2 of the standard IEC 61400-13 [2] can be measured, in particular in accordance with the said standard.

[0066] According to a further embodiment of the method according to the application, a load parameter value selected from the group comprising: maximum load parameter value measured during (the time length) of the load data period, minimum operating parameter value measured during (the time length) of the load data period, load parameter mean value determined from the load parameter values ​​measured during (the time length) of the load data period, standard deviation determined from the load parameter values ​​measured during (the time length) of the load data period.

[0067] A stress dataset can be based on at least two of the specified stress parameter values, preferably on all of these stress parameter values. This allows for the creation of higher-quality training datasets.

[0068] A preferred example of stress parameters and stress parameter values ​​on which a stress data set can be based is given in particular in Table 3. Table 3 Load parameters Load parameter value Blade bending parameters - max. bending value - min. bending value - averaged bending value - Standard deviation of the bending values Edge bending parameters - max. bending value - min. bending value - averaged bending value - Standard deviation of the bending values Rotor tilting moment parameters - max. torque value - min. torque value - averaged moment value - Standard deviation of the moment values

[0069] According to a further embodiment of the method according to the application, the method may further comprise: Measuring the at least one load parameter of the reference wind turbine during a measuring period, and recording the operating data of the reference wind turbine during the measuring period, wherein the measuring period is in particular at least 3 months, preferably at least 6 months (and e.g. at most 36 months).

[0070] Each measured (digital) load parameter value can be assigned a timestamp of the measurement time. This can be taken into account (at least indirectly) to create a load data set and, in particular, for the synchronized assignment described above.

[0071] During a measurement period, at least one operating parameter can be continuously recorded and, in particular, measured. For example, the measured data can be temporarily stored.

[0072] As already described, corresponding datasets can be created and then provided, in particular to generate a large number of wind power training datasets for training a machine learning model. To obtain a sufficient number of training datasets (especially with an assumed time length of 10 minutes for each operating dataset (and load dataset), the measurement period can be at least 3 months, preferably at least 6 months.

[0073] For example, approximately 100,000 wind power training datasets can be generated. Preferably, approximately 20,000 of these wind power training datasets can be used as validation datasets.

[0074] According to a further embodiment of the method according to the application, the method may further comprise: Forming a (time-dependent) load data set by converting the at least one measured load parameter value of the reference wind turbine into at least one fatigue load indicator, wherein the conversion is based in particular on a rainflow counting method (also called rainflow counting method).

[0075] As already described, a load data set can be based on at least one measured load parameter value of at least one load parameter of the reference wind turbine. In particular, a load data set can be based on at least one converted load parameter value, thus in particular containing at least one fatigue load indicator.

[0076] Particularly preferably, the at least one measured load parameter value of the reference wind turbine can be converted into at least one fatigue load indicator according to the standard ASTM E1049 - 85 (2017).

[0077] To obtain a high-quality machine learning model, it has been recognized, as stated in the application, that so-called overfitting should be avoided during the training process. This specifically means ensuring that the machine learning model only recognizes patterns in the training data that actually exist, rather than patterns that occurred randomly (once).

[0078] Therefore, according to a further preferred embodiment of the method according to the application, the application proposes that at least one regularization technique be applied during the training period. This can at least reduce the risk of overfitting.

[0079] The at least one regularization technique can in particular be selected from the group comprising: Linear Bayesian regression (also called Bayesian linear regression), Bayesian neural network with concrete dropout (also called Bayesian neural network by concrete dropout) or Bayesian by variational inference, Adaptive Bayesian spline regression (also called Bayesian adaptive spline regression).

[0080] Such regularization techniques have proven to be particularly advantageous.

[0081] According to a further embodiment of the method according to the application, a model for estimating the aleatoric and / or epistemic uncertainties can be applied during the training period.

[0082] Furthermore, during the training process, a so-called loss function can be used to evaluate the machine learning model, at least partially, based on the output. A loss function can, for example, contain a metric comparing the output with a reference value. A loss function can, for example, be defined when a reference output is known for a given input. For example, the loss function can be a so-called "Gaussian negative log likelihood function."

[0083] In a training step of the training process, the machine learning model can be adapted. Adapting can, in particular, involve changing at least one parameter of the machine learning model.

[0084] A wind power training dataset can preferably be in the form of text data, from which a suitable representation can be generated. The generated representation can, for example, be formed from a vector, a matrix, and / or a tensor.

[0085] As already described, the trained machine learning model can be used to determine the turbine condition of another wind turbine, in particular to estimate it with a high accuracy.

[0086] A machine learning model can, for example, be any algorithm that receives an input (in this case, in particular at least one operating data set, preferably a plurality of historical and stored operating data sets of a further wind turbine) and returns an output (in this case, in particular at least one system condition data set, preferably containing at least one fatigue load indicator of at least one component of the further wind turbine), wherein the output depends both on the input and on at least one parameter of the machine learning model.

[0087] As described, the machine learning model is previously trained to learn a relationship between inputs and outputs. The trained machine learning model can then be used to generate a new output from a new input.

[0088] The input for the trained machine learning model in this case is in particular operational data sets, which can preferably be transformed into a vector, a matrix and / or a tensor for input.

[0089] According to a further embodiment of the method according to the application, the additional wind turbine can be a wind turbine type that is identical to the wind turbine type of the reference wind turbine. In other words, if the reference wind turbine is of type V1, then the trained machine learning model can preferably be used to determine, in particular to estimate, the turbine state and in particular the RUL of additional wind turbines of the same or a similar type (i.e., type Vx). It is understood that the machine learning model trained according to the application can still deliver acceptable results even for similar types (for example, with an upgrade).

[0090] Alternatively or additionally, the additional wind turbine and the reference wind turbine can be part of the same wind farm. A wind farm can comprise a plurality of wind turbines. In order to determine the turbine state and, in particular, the RUL of preferably all of the wind turbines in a wind farm, at least one wind turbine in the wind farm (in variants of the applications, more, for example, two, can be selected) can be selected as the reference wind turbine. After training a machine learning model, as described above, the turbine state and, in particular, the RUL of at least one additional wind turbine, preferably all of the additional wind turbines, in the wind farm can be determined using the trained machine learning model.

[0091] According to a preferred embodiment of the method according to the application, the method may further comprise: Determining the remaining operating time of the additional wind turbine based on at least one turbine status data set of the additional wind turbine.

[0092] As already described, the remaining operating time or RUL of the other wind turbine can be estimated with a high degree of accuracy from the output turbine status data set (especially using already known estimation methods).

[0093] For some components or structural elements of a wind turbine, such as tubular steel towers, the damage (and thus the RUL) can be estimated using structural models of the tower shell and welds (and the turbine condition dataset). For this purpose, weld details, tower geometry (e.g., thicknesses, diameters), and material properties may be known or (conservatively) estimated.

[0094] However, for structural components such as rotor blades, machine frames, generator baseplates, and / or hubs, this would require considerable effort to determine material quality and geometries, create FEM models, and calculate fatigue strength. Such an approach is only useful if more detailed information can be obtained from the turbine manufacturer or if the number of turbines of the type under consideration is sufficiently large.

[0095] Alternatively, attempting to obtain reference / design loads might be an option. For the tower base, such loads can be found in publicly available foundation load data; for other components, collaboration with the OEM may be unavoidable.

[0096] A variation of the further option could be to select wind conditions from the measurements that are close to the design conditions of the turbine and to take the reference loads from the measurements themselves and extrapolate them to 20 years.

[0097] Yet another aspect of the application is a computing device according to claim 15, comprising at least one data memory containing computer program code and at least one processor, wherein the program code and the processor are configured such that the computing device is caused to generate at least one plant status data set based on at least one provided operational data set and at least partially using a machine learning model trained according to any one of claims 1 to 11.

[0098] In particular, the machine learning model can be stored in the data store.

[0099] A previously described module, element, etc. may at least partially comprise hardware elements (e.g. processor, memory means, etc.) and / or at least partially comprise software elements (e.g. executable code).

[0100] There are now numerous possibilities for designing and further developing the method according to the application for training a machine learning model usable for determining the remaining service life of a wind turbine, the method according to the application for using a trained machine learning model, and the computing device according to the application. Reference is made, on the one hand, to the patent claims subordinate to the independent patent claims and, on the other hand, to the description of exemplary embodiments in conjunction with the drawing. The drawing shows: Fig. 1 is a schematic view of an example of a wind farm with a plurality of wind turbines, Fig. 2 is a diagram of an embodiment of a method according to the present application, Fig. 3 is a diagram of a further embodiment of a method according to the present application, Fig. 4 is a diagram of a further embodiment of a method according to the present application, Fig. 5 is an exemplary plant state diagram, Fig. 6 is an exemplary diagram for checking collinearity by correlation, and Fig. 7 is a schematic view of an embodiment of a computing device according to the present application.

[0101] The Figure 1shows a schematic view of an example of a wind farm 100 with a plurality of wind turbines 102, 104. Here, the reference number 102 denotes the (selected) reference wind turbine and the reference number 104 denotes the further wind turbines of the wind farm 100. The wind farm 100 can be an onshore wind farm and / or an offshore wind farm.

[0102] The wind farm 100 may preferably have a (central) control device 106, in particular comprising at least one control module 108, for example a SCADA control module 108.

[0103] In particular, the operating parameter values ​​of at least one operating parameter of at least one wind turbine 102, 104, preferably all wind turbines, of the wind farm 100 can be recorded and transmitted, for example, via a (wireless and / or wired) communication network 110 to the (central) control device 106. Preferred operating parameters are, for example, the parameters listed in Table 1.

[0104] The at least one control module 108 can preferably control and / or regulate the wind farm in a conventional manner, based at least also on the received (SCADA) operating data sets of the wind turbines 102, 104.

[0105] Furthermore, the operating data records can be recorded or stored (in a data storage device (not shown) (e.g., of the control device 106). This can be provided for documentation purposes anyway. Preferably, the operating data records of the wind turbine 102, 104 can be stored and used to determine a respective system status of the respective wind turbine 102, 104 (as will be described below).

[0106] As can also be seen, the reference wind turbine 102 has a sensor arrangement 112 with at least one load sensor (in particular, a plurality of load sensors). The at least one load sensor is configured to measure at least one load parameter (e.g., blade root load parameter, rotor load parameter, tower load parameter, tower torsion parameter, and / or tower tip moment parameter).

[0107] The measured load parameter values ​​of the at least one load parameter can preferably be stored in a data memory (not shown) (e.g., the control device 106) for subsequent processing, as will be described.

[0108] Preferably, the wind turbines 102, 104 of the wind farm 100 can be of the same wind turbine type.

[0109] The Figure 2 shows a diagram of an embodiment of a method according to the present application. The method is, in particular, a computer-implemented method for training a machine learning model. The machine learning model can be used to indirectly determine a remaining service life of a wind turbine, in particular to determine the structural and / or mechanical condition of the wind turbine.

[0110] In a step 201, a plurality of operating data sets of a reference wind turbine are provided. An operating data set is, in particular, a SCADA operating data set. For an onshore wind turbine, the operating data set can be formed according to the exemplary Table 1. It is understood that other, additional, or fewer operating parameters and / or values ​​can be included. For an offshore wind turbine, the operating data set can be formed according to the exemplary Tables 1 and 2. This includes, in particular, that an operating data set can be formed from two separate sub-operating data sets. It is also understood here that other, additional, or fewer operating parameters and / or values ​​can be included.

[0111] In particular, each operating data record can be assigned an operating data period that refers to the measurement period for the measurement of the at least one operating parameter. For example, an operating data period can include a start time (e.g., 1 / 1 / 21, 1:00 a.m.), a time length (e.g., 10 minutes), and an end time (e.g., 1 / 1 / 21, 1:10 a.m.).

[0112] In particular, a large number of operational data sets can be provided for the entire recording or measurement period.

[0113] In step 202, a plurality of load data sets of the reference wind turbine are provided, wherein a load data set is based on at least one load parameter measured on the reference wind turbine.

[0114] In particular, each load data record can be assigned a load data period that refers to the measurement period for the measurement of the at least one load parameter. This load data period can be selected according to the operating data period.

[0115] For example, a load data period may include a start time (e.g., 1 / 1 / 21; 1:00 a.m.), a time length (e.g., 10 min), and an end time (e.g., 1 / 1 / 21, 1:10 a.m.). In other variants (particularly if the time length is always fixed), an operational data period and / or a load data period may also be sufficiently determined by a single timestamp.

[0116] In a further step 203, a plurality of wind power training data sets are generated for training a machine learning model. The generation occurs by synchronously assigning a respective operating data set to a respective load data set.

[0117] The (temporally) synchronized assignment of an operating data record to a load data record is carried out, in particular, depending on the respectively assigned load data durations or operating data durations. In particular, the start time and / or end time can be evaluated for this purpose. For example, if, as in the example above, the start time (e.g., 1 / 1 / 21, 1:00 a.m.) and / or end time (e.g., 1 / 1 / 21, 1:10 a.m.) of an operating data record and a load data record are the same, then these operating data records can be assigned to each other.

[0118] In particular, the matching includes forming a wind power training dataset containing the data from the operational dataset and the associated load dataset. In other words, an operational dataset can be correlated with a load dataset to form a wind power training dataset.

[0119] Example wind power training datasets WTD 1 , WTD 2 , ..., WTD n (n is a natural number for different training dataset periods to represent different operating data periods) are shown in Table 4. The parameter values ​​mentioned are illustrative examples. Table 4 WTD1 WTD2 WTD n Blade_Load_Ind 4.002E+03 4.344E+03 4.172E+03 Blade_Load_dev_Ind 2.444E+02 3.050E+02 1.811E+02 Power_Active_Mean 1.300E+03 1.465E+03 1.370E+03 Power_Active_Min 1.067E+03 1.175E+03 1.195E+03 Power_Active_Max 1.502E+03 1.757E+03 1.568E+03 Power_Active_dev 1.081E+02 1.707E+02 6.498E+01 Rotor_Speed_Mean 1.179E+01 1.223E+01 1.205E+01 Rotor_Speed_Min 1.117E+01 1.170E+01 1.153E+01 Rotor_Speed_Max 1.224E+01 1.258E+01 1.240E+01 Rotor_Speed_dev 3.198E-01 2.755E-01 1.818E-01 ... ... ... ...

[0120] Here, Mean represents the mean value, Min represents the minimum value, Max represents the maximum value, Dev represents the standard deviation, and Ind represents a fatigue load indicator. It is understood that (as indicated by ...) a variety of other data may be included.

[0121] In an (optional) step 204, the machine learning model is trained using the generated wind power training data sets. Preferably, a portion of the generated wind power training data sets can be used as validation data sets. In a particularly conventional manner, a machine learning model can preferably be trained in the form of a neural network. For input, a wind power training data set can be transformed into a vector, a matrix, and / or a tensor.

[0122] The Figure 3shows a diagram of another embodiment of a method according to the present application. To avoid repetition, only the differences from the embodiment according to Figure 2 described and otherwise referred to the previous explanations.

[0123] In order to provide the at least one load data set, in step 301, at least one load parameter of the reference wind turbine (i.e., of at least one component of the reference wind turbine) can be measured during a measurement period (preferably at least 6 months).

[0124] In particular, at least almost continuous measurement and thus monitoring can be performed. As already described, the measurement of at least one load parameter can be carried out in particular in accordance with the IEC 61400-13 standard [2]. Preferably, the load parameter values ​​of a plurality of load parameters of a plurality of components of the reference wind turbine can be measured.

[0125] In step 303, which can be performed at least partially in parallel with step 301, the at least one measured load parameter value of the reference wind turbine can be converted into at least one fatigue load indicator. Preferably, all load parameter values ​​can be converted. The conversion can be based, in particular, on a rainflow counting method. Particularly preferably, the at least one measured load parameter value of the reference wind turbine can be converted into at least one fatigue load indicator in accordance with the ASTM E1049-85(2017) standard.

[0126] In step 303, the load data sets can also be created. For this purpose, the fatigue load indicators (in particular according to the timestamp provided with the load parameter values ​​(or the corresponding sample values)) can be divided into a plurality of load data sets, each with a previously described load data period. The load data period can preferably correspond to the operating data period. In particular, an operating data period is assigned to each provided (SCADA) operating data set.

[0127] The operating parameter values ​​of the at least one operating parameter can be recorded during the measurement period and in particular (in a known manner) provided in the form of operating data records in step 302 (see, for example, Table 1 and / or 2).

[0128] In step 304, a plurality of wind power training data sets (see Table 4) are generated for training a machine learning model by synchronously assigning a respective operating data set to a respective load data set, as previously described.

[0129] In particular, preprocessing or pretreatment of the generated wind power training data sets can be performed in step 305. For example, invalid values ​​and / or indicators can be removed. Feature selection can also be performed in the manner described above in step 305. Examples include sequential feature selection (e.g., Ferri, FJ & Pudil, Pavel & Hatef, M. (2001). Comparative Study of Techniques for Large-Scale Feature Selection. Pattern Recognition in Practice, IV: Multiple Paradigms, Comparative Studies and Hybrid Systems. 16. 10.1016 / B978-0-444-81892-8.50040-7), recursive feature selection (e.g., Guyon, I, J Weston, S Barnhill, and V Vapnik. 2002. "Gene Selection for Cancer Classification Using Support Vector Machines." Machine Learning 46 (1): 389-422), linear regression with L1 regulation (Lasso) and / or Bayesian LASSO (e.g.,Trevor Park & ​​George Casella (2008) The Bayesian Lasso, Journal of the American Statistical Association, 103:482, 681-686, DOI: 10.1198 / 016214508000000337).

[0130] In step 306, the generated wind power training data sets are divided into training data sets (e.g., 80%) and validation data sets (e.g., 20%). Furthermore, the wind power training data sets are transformed into data representations that are suitable as input for the machine learning model to be trained, i.e., that can depend in particular on the machine learning model to be trained (in a known manner). Preferably, a wind power training data set is transformed into a vector, a matrix, and / or a tensor.

[0131] In step 307, the machine learning model is trained or learned during a training period. Preferably, a neural network can be trained. Preferably, so-called supervised training or learning takes place using the training data sets and the validation data sets.

[0132] During the training period, a model can be applied to estimate the aleatoric and / or epistemic uncertainties, whereby the estimation can be achieved by applying at least one machine learning method. During the training period or during the training process, at least one regularization technique can be applied. This can at least reduce the risk of overfitting. The at least one regularization technique (in particular the at least one machine learning method) can in particular be selected from the group comprising: Linear Bayesian regression (also called Bayesian linear regression), Bayesian neural network with concrete dropout (also called Bayesian neural network by concrete dropout) or Bayesian neural network with variational inference (Bayesian neural network by variational inference), adaptive Bayesian spline regression (also called Bayesian adaptive spline regression).

[0133] The methods mentioned already include an estimation of aleatory and / or epistemic uncertainties. Bayesian neural networks, in particular, allow the representation of heteroskedastic uncertainties, i.e., uncertainties whose magnitude depends on the expected value of the estimated target variable (plant condition).

[0134] Alternatively (or additionally), these can be determined during the training period or during the training process using other methods (especially bootstrapping).

[0135] After training, a trained machine learning model can be made available for further use in step 308.

[0136] The Figure 4 shows a diagram of an embodiment of a method for using a machine learning model which is implemented according to the present application, for example according to the embodiment according to Figure 2 and / or 3 was trained.

[0137] In a first step 401, at least one operating data set (preferably a plurality of operating data sets) of a further wind turbine is entered into the machine learning model. Preferably, the operating data sets recorded over a specific period (in particular of at least 2 months, preferably at least 3 months (and e.g., at most 24 months)) can be entered. The further wind turbine is preferably of the same, or at least a similar, type as the reference wind turbine. Preferably, the further wind turbine can additionally be part of the same wind farm as the reference wind turbine (cf. e.g., Fig. 1 ). It is understood that the additional wind turbine may also come from another wind farm.

[0138] In step 402, the machine learning model outputs at least one system condition data set. The system condition data set may, for example, contain at least one determined fatigue load indicator of the additional wind turbine. It is understood that a different load parameter indicator may also be output.

[0139] In an optional step 403, the remaining operating time (RUL) of the additional wind turbine can be determined based on the at least one output system status data set of the additional wind turbine. This includes, in particular, determining, in particular estimating, the RUL of at least one component of the additional wind turbine.

[0140] By training a machine learning model with load data and operating data from a reference wind turbine, according to the application, to recognize patterns between this data and "store" them accordingly in the trained machine learning model, at least the condition of another wind turbine can be determined, in particular estimated with a high degree of reliability, using the trained machine learning model based solely on the operating data of this other wind turbine (i.e., without actually measuring a load parameter). As already described, the condition of the wind turbine includes the condition of one (or more) structural component(s) of the wind turbine.

[0141] The Figure 5 shows an example plant status diagram, with the help of which the RUL can be determined, i.e. in particular estimated.

[0142] The y-axis shows the system condition WK cond of the wind turbine or the component condition WK cond of a structural component of the wind turbine and the time t is shown on the x-axis.

[0143] First, it can be assumed that at the (current) time t 1 the (current) plant state WK is cond_1. This can be determined in particular according to the method of Figure 4 be determined.

[0144] The specific turbine condition limit WK cond_grenz can be specified. Then, using models as described, the time t 2 can be determined. In particular, the time t 2 is the point in time beyond which the wind turbine is expected to no longer be able to operate (for safety reasons). The RUL or T RUL can then generally be calculated as follows: T RUL = t 2 - t1.

[0145] The Figure 6shows an exemplary diagram to illustrate the checking of collinearity by correlation, as can be carried out, for example, in step 305.

[0146] As can be seen in particular from the Figure 6 As can be seen, the correlations (e.g., expressed by the Pearson correlation coefficient in the interval [-1 .., 1]) of the operating parameters intended for use are calculated. Operating parameters with a high correlation coefficient (typically >= 0.8 to 0.9 or <= -0.8 to -0.9) are referred to as collinear and should not be used together for training machine learning methods.

[0147] The Figure 7shows a schematic view of an embodiment of a computing device 700 according to the present application. The computing device 700 comprises at least one processor (e.g., microprocessor, DSP, FPGA, and / or the like) and at least one data memory 720 containing computer program code. The data memory 720 contains, in particular, the trained or yet-to-be-trained machine learning model 750.

[0148] Furthermore, at least one communication interface 730 and / or at least one user interface 740 is provided, configured for inputting and / or outputting data, such as operating data sets, training data sets and / or system status data sets.

[0149] The data memory 720 and the processor 710 are configured such that the computing device 700 is caused to generate at least one plant status data set, based on at least one provided operational data set and at least partially using a machine learning model 750, which, for example, is implemented in accordance with the embodiment according to Figure 2 and / or Figure 3 was trained.

Claims

1. A method, in particular a computer-implemented method, for training a machine learning model (750) used for determining a remaining useful life of a further wind turbine (104), comprising: - providing a plurality of operation data sets of a reference wind turbine (102), - providing a plurality of load data sets of the reference wind turbine (102), wherein a load data set is based on at least one load parameter measured at the reference wind turbine (102), and - generating a plurality of wind turbine training data sets for training the machine learning model (750) by synchronously assigning a respective operation data set with a respective load data set.

2. The method according to claim 1, characterized in that - the machine learning model (750) is trained with the generated wind power training data sets during the training time period.

3. The method according to claim 2, characterized in that - a portion of the generated wind power training data sets is used as validation data sets during a training time period.

4. The method according to any one of the previous claims, characterized in that - a provided operation data set contains at least one operation parameter of the reference wind turbine (102), - wherein the at least one operation parameter is selected from the group comprising: - tower top acceleration, - pitch angle, - pitch speed, - rotor speed, - electrical power, in particular active electrical power, - nacelle wind speed.

5. The method according to claim 4, characterized in that - an operation data set comprises as operation parameter value of an operation parameter at least one operation parameter value selected from the group comprising: - maximum operation parameter value detected during the operating data period, - minimum operation parameter value detected during the operation data time period, - operation parameter mean value determined from the operation parameter values detected during the operation data time period, - standard deviation determined from the operation parameter values detected during the operation data time period.

6. The method according to one of the previous claims, characterized in that - a provided load data set is based on at least one measured load parameter of the reference wind turbine (102), - wherein the load parameter is selected from the group comprising: - blade root load parameter, - rotor load parameter, - tower load parameter, - tower torsion parameter, - tower top moment parameter.

7. The method according to any of the previous claims, characterized in that the method comprises: - measuring the at least one load parameter of the reference wind turbine (102), - wherein measuring the at least one load parameter is in particular performed according to the IEC 61400-13 standard [2].

8. The method according to claim 6 or 7, characterized in that - as load parameter value of a load parameter a load parameter value is provided, selected from the group comprising: - maximum load parameter value measured during the load data time period, - minimum operation parameter value measured during the load data time period, - load parameter mean value determined from the load parameter values measured during the load data time period, - standard deviation determined from the load parameter values measured during the load data time period.

9. The method according to any of the previous claims, characterized in that the method comprises: - measuring the at least one load parameter of the reference wind turbine (102) during a measurement time period; and - detected the operational data parameter values of the reference wind turbine (102) during the measurement time period, - wherein the measurement time period is in particular at least 3 months, preferably at least 6 months.

10. The method according to any one of the previous claims, characterized in that the method comprises: - forming a load data set by converting the at least one measured load parameter value of the reference wind turbine (102) into at least one fatigue load indicator, - wherein the converting is based in particular on a Rainflow Counting method.

11. The method according to any one of the previous claims, characterized in that - at least one regularization technique is applied during the training time period, - wherein the regularization technique is in particular selected from the group comprising: - linear Bayesian regression, - Bayesian neural network with concrete dropout, - Bayesian neural network with variational inference - adaptive Bayesian spline regression, and / or - during the training time period, a model for estimating the aleatory and / or epistemic uncertainties is applied, wherein the estimation is achieved by applying at least one machine learning method, wherein the machine learning method is in particular selected from the group comprising: - linear Bayesian regression, - Bayesian neural network with concrete dropout, - Bayesian neural network with variational inference - adaptive Bayesian Spline Regression, and / or - wherein the estimation is implemented by bootstrapping.

12. A method of using a machine learning model (750) trained according to any one of claims 1 to 11, comprising: - inputting at least one operation data set of the further wind turbine (104) into the machine learning model (750), and - outputting, by the machine learning model (750), at least one turbine condition data set.

13. The method according to claim 12, characterized in that - the further wind turbine (104) is a wind turbine type identical to the wind turbine type of the reference wind turbine (102), and / or - the further wind turbine (104) and the reference wind turbine (102) are comprised by the same wind farm (100).

14. The method according to claim 12 or 13, characterized in that the method comprises: - determining the remaining runtime of the further wind turbine (104) based on the at least one turbine condition data set of the further wind turbine (104).

15. A computing device (700) comprising at least one data memory (720) containing computer program code and at least one processor (710), wherein the data memory (720) and the processor (710) are configured such that the computing device (700) is caused to generate at least one turbine condition data set based on at least one provided operation data set and at least in part using a machine learning model (750) trained according to any one of claims 1 to 11.