A COMPUTER-IMPLEMENTED METHOD FOR GENERATING A PREDICTION MODEL FOR PREDICTING ROTOR BLADE DAMAGES OF A WIND TURBINE

Abstract

A computer-implemented method for generating a prediction model for predicting rotor blade damages of a wind turbine is provided, wherein the method provides data including data sets for wind turbines, where each data set includes respective values of turbine variables(s), weather variable(s) and damage variable(s) wherein the method includes: a) discretizing the values, resulting in modified data sets; b) structure learning of a plurality of Bayesian networks based on the modified data sets, where each Bayesian network is learned by another learning method; c) determining an optimum Bayesian network based on a performance measure reflecting the prediction quality of a respective Bayesian network, where the optimum Bayesian network has the best performance measure; d) parameter learning of the optimum Bayesian network based on the modified data sets, resulting in conditional probabilities, where the optimum Bayesian network combination with the conditional probabilities is the prediction model.

Claims

1. A computer-implemented method for generating a prediction model for predicting rotor blade damages of a wind turbine, wherein the method processes previously acquired data, the data comprising data sets for a plurality of wind turbines, where each data set refers to a specific wind turbine and comprises respective values of variables, the variables including one or more turbine variables defining characteristics of the specific wind turbine, one or more weather variables defining weather conditions averaged over the operation time of the specific wind turbine and one or more damage variables defining damages having occurred on at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine, wherein the method comprises the following: a) discretizing the values of those variables which are numerical variables, resulting in modified data sets; b) structure learning of a plurality of Bayesian networks based on the modified data sets, where each Bayesian network is a probabilistic directed acyclic graph comprising the variables as nodes and defining the probabilistic dependencies between nodes by directed edges and where each Bayesian network is learned by another learning method; c) determining an optimum Bayesian network out of the plurality of Bayesian networks based on a performance measure reflecting the prediction quality of a respective Bayesian network, where the optimum Bayesian network has the best performance measure; and d) parameter learning of the optimum Bayesian network based on the modified data sets, resulting in conditional probabilities between variables representing nodes linked by respective directed edges the optimum Bayesian network, where the optimum Bayesian network in combination with the conditional probabilities is the prediction model.

2. The method according to claim 1, wherein the one or more turbine variables comprise one or more of the following variables: a turbine variable specifying the generator type used in the specific wind turbine; a turbine variable specifying the rotational speed for which the rotor of the specific wind turbine is configured; a turbine variable specifying the rotor diameter of the specific wind turbine; a turbine variable specifying a category to which the specific wind turbine belongs; a turbine variable specifying the altitude of the specific wind turbine; a turbine variable specifying the type of the specific wind turbine; a turbine variable specifying the rotor tip speed for which the specific wind turbine is configured; a turbine variable specifying the age of the specific wind turbine; a turbine variable specifying the total amount of time where the specific wind turbine experiences a wind speed over a predetermined value during its operation; and a turbine variable specifying whether the specific wind turbine is an onshore or offshore wind turbine.

3. The method according to claim 1, wherein the one or more weather variables comprise one or more of the following variables: a weather variable specifying the average wind speed at the location of the specific wind turbine during the operation time of the specific wind turbine; a weather variable specifying the average air humidity at the location of the specific wind turbine during the operation time of the specific wind turbine; a weather variable specifying the average lightning density which is the average number of lightning strikes per time unit and per area unit around the location of the specific wind turbine during the operation time of the specific wind turbine; a weather variable specifying the average precipitation per time unit and per area unit around the location of the specific wind turbine during the operation time of the specific wind turbine.

4. The method according to claim 1, wherein the one or more damage variables comprise one or more of the following variables: one or more erosion occurrence variables, each specifying the number of erosion occurrences with a respective severity level and/or of a respective type in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; one or more erosion area variables each specifying the total area of all erosions with a respective severity level and/or of a respective type occurred in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; one or more superficial damage occurrence variables, each specifying the number of superficial damage occurrences with a respective severity level and/or of a respective type in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; one or more superficial damage area variables, each specifying the total area of all superficial damages with a respective severity level and/or of a respective type occurred in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; a crack variable specifying the number of cracks occurred on at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine an accessory loss variable specifying the number of accessory parts lost from at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; a contamination occurrence variable specifying the number of contamination occurrences in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; a contamination area variable specifying the total area of all contaminations occurred in a predetermined region of at least one rotor blade of the specific wind turbine during the operation time of the specific wind turbine; a lightning protection system failure variable specifying the number of failures of the lightning protection system of at least one rotor blade of the specific wind turbine which occurred during the operation time of the specific wind turbine.

5. The method according , wherein the discretization step a) is based on the Hartemink's information-preserving algorithm.

6. The method according to wherein the learning methods used in step b) comprise one or more of the following methods: PC algorithm, Grow-Shrink; Incremental Association Markov Blanket; Fast Incremental Association; Interleaved Incremental Association; Hill Climbing; Tabu Search.

7. The method according to claim 1, wherein in step c) one or more cross validations are performed for each Bayesian network of the plurality of Bayesian networks, where in each cross validation a parameter learning of the respective Bayesian network based on a first part of the modified data sets is performed, resulting in conditional probabilities between variables representing nodes linked by respective directed edges in the respective Bayesian network, where a prediction of the values of the one or more damage variables of a second part of the modified data sets being different from the first part is performed by the respective Bayesian network in combination with the conditional probabilities based on the values of the one or more turbine variables and weather variables of the second part of the modified data sets, where a prediction quality parameter is determined for the respective cross validation by comparing the predicted values with the actual values of the one or more damage variables of the second part of the modified data sets, where the prediction quality parameter is the performance measure in case of a single cross validation and where the average of the prediction quality parameters over the cross validations is the performance measure in case of several cross validations.

8. The method according wherein the performance measure is based the F1 score or the Precision or the Recall.

9. The method according to claim 1, wherein newly acquired data sets are added to the previously acquired data, where the steps a) to d) are performed for the previously acquired data additionally comprising the newly acquired data sets, thus resulting in an updated prediction model.

10. A computer-implemented method for predicting rotor blade damages of a wind turbine, where the method processes the prediction model which is generated by the method according to claim 1, or which has been generated beforehand by the method, where the value of at least one damage variable is predicted based on known values of at least one turbine variable and at least one weather variable valid for the wind turbine.

11. A computer program product comprising a computer readable hardware storage device having computer readable program code stored therein, said program code executable by a processor of a computer system to implement a method with program code, which is stored on a non-transitory machine-readable carrier, for carrying out a method according to claim 1 when the program code is executed on a computer.

12. A computer program with program code for carrying out a method according to claim 1 when the program code is executed on a computer.

Description

BRIEF DESCRIPTION

[0051] Some of the embodiments will be described in detail, with reference to the following figures, wherein like designations denote like members, wherein:

[0052] FIG. 1 is a flowchart illustrating the steps performed in a variant of the method according to embodiments of the invention; and

[0053] FIG. 2 is a diagram illustrating an optimum Bayesian network determined based on the method as shown in FIG. 1.

DETAILED DESCRIPTION

[0054] The method as described in the following processes data DA in the form of a plurality of data sets DS, where each data set refers to a specific wind turbine. In other words, there are a plurality of data sets referring to different specific wind turbines designated as WT in FIG. 1. Each data set comprises respective values of variables of a specific wind turbine. The variables include a plurality of turbine variables T1 to T10 defining characteristics of the specific wind turbine WT, a plurality of weather variables W1 to W4 defining weather conditions averaged over the operation time of the specific wind turbine WT and a plurality of damage variables D1 to D19 defining damages having occurred on a rotor blade (i.e., a specific rotor blade) of the specific wind turbine WT during the operation time of the specific wind turbine. In an alternative, the damages may also cover the damages of several rotor blades or all (three) rotor blades of the specific wind turbine. The above variables will be described in detail in the following. Nevertheless, those variables are examples and other types of variables for describing characteristics of the wind turbine, averaged weather conditions and damages may be used in modified embodiments.

Turbine Variable T1=WTG Type:

[0055] This variable specifies the generator type used in the specific wind turbine.

Turbine Variable T2=RPM:

[0056] This variable specifies the rotational speed for which the rotor of the specific wind turbine is designed.

Turbine Variable T3=Diameter:

[0057] This variable specifies the rotor diameter of the specific wind turbine.

Turbine Variable T4=Platform:

[0058] This variable defines a category of the specific wind turbine which is a category higher than the category “Turbine Type” defined below.

Turbine Variable T5=Turbine_Altitude:

[0059] This variable specifies the altitude of the specific wind turbine.

Turbine Variable T6=Turbine_Type:

[0060] This category specifies the type of the wind turbine which can be specified by a respective type identifier.

Turbine Variable T7=Tip_Speed:

[0061] This variable specifies the rotor tip speed for which the specific wind turbine is designed.

Turbine Variable T8=Age:

[0062] This variable specifies the age of the specific wind turbine.

Turbine Variable T9=Stress_Hours:

[0063] This variable specifies the total amount of time where the specific wind turbine experiences a wind speed over a speed of 23 m/s during its operation.

Turbine Variable T10=Park_Type:

[0064] This variable specifies whether the specific wind turbine is an onshore or offshore wind turbine.

Weather Variable W1=MAWS:

[0065] This variable specifies mean annual wind speed at the location of the specific wind turbine during the operation time of the specific wind turbine.

Weather Variable W2=Humidity:

[0066] This variable specifies the average air humidity at the location of the specific wind turbine during the operation time of the specific wind turbine.

Weather Variable W3=Lightning:

[0067] This variable specifies the average lightning density which is the average number of lightning strikes per year within 1 km.sup.2 around the location of the specific wind turbine during the operation time of the specific wind turbine.

Weather Variable W4=Precipitation:

[0068] This variable specifies the average precipitation (referring to rain and snow) per year and per m.sup.2 around the location of the specific wind turbine during the operation time of the specific wind turbine.

Damage Variable D1=Erosion2:

[0069] This variable specifies the number of erosion occurrences with a medium severity level (i.e., level 2) at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D2=Erosion2 Area:

[0070] This variable specifies the total area of all erosions with a medium severity level (i.e., level 2) occurred at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D3=Erosion3:

[0071] This variable specifies the number of erosion occurrences with a high severity level (i.e., level 3) at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D4=Erosion3 Area:

[0072] This variable specifies the total area of all erosions with a high severity level (i.e., level 3) occurred at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D5=Erosion4:

[0073] This variable specifies the number of erosion occurrences with a very high severity level (i.e., level 4) at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D6=Erosion4 Area:

[0074] This variable specifies the total area of all erosions with a very high severity level (i.e., level 4) occurred at the leading edge of a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D7=Surface Voids:

[0075] This variable specifies the number of surface void occurrences at a rotor blade of the specific wind turbine during the operation time of the specific wind turbine, where surface voids refer to erosions at other parts of the rotor blade than the leading edge.

Damage variable D8=Surface Voids Area:

[0076] This variable specifies the total area of all surface voids occurred at one rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D9=Paint Damage:

[0077] This variable specifies the number of paint damage occurrences at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine. Paint damages refer to damages in the form of peeling and chipping paint which do not go deeper inside the rotor blade as it is the case for erosions or surface voids.

Damage variable D10=Paint Damage Area:

[0078] This variable specifies the area of all paint damages occurred at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D11=Cracks:

[0079] This variable specifies the number of cracks occurred at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D12=Pinholes:

[0080] This variable specifies the number of pinhole occurrences at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine. Pinholes are well-known damages and refer to small holes within the paint of the rotor blade which are different from the above paint damages referring to peeling and chipping paint.

Damage variable D13=Pinholes Area:

[0081] This variable specifies the total area of all pinholes occurred on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D14=Add On:

[0082] This variable specifies the number of accessory parts lost from a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D15=Contamination:

[0083] This variable specifies the total area of all contaminations occurred at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D16=Contamination Area:

[0084] This variable specifies the total area of all contaminations occurred on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

Damage variable D17=LPS:

[0085] This variable specifies to the number of failures of the lightning protection system (i.e., of the lightning conductor) of a rotor blade of the specific wind turbine which occurred during the operation time of the specific wind turbine.

Damage variable D18=Scratches Gouges:

[0086] This variable specifies the number of occurrences of scratches and gouges at any location on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine. Scratches and gouges are well-known superficial damages on a rotor blade which can be clearly distinguished from other damages.

Damage variable D19=Scratches Gouges Area:

[0087] This variable specifies the total area of all scratches and gouges occurred on a rotor blade of the specific wind turbine during the operation time of the specific wind turbine.

[0088] With respect to the above damage variables, the variables D1, D3, D5 and D7 can be regarded as erosion occurrence variables in the sense of the patent claims. Furthermore, the variables D2, D4, D6 and D8 can be regarded as erosion area variables in the sense of the patent claims. Moreover, the variables D9, D12 and D18 can be regarded as superficial damage occurrence variables in the sense of the patent claims whereas the variables D10, D13 and D19 can be regarded as superficial damage area variables in the sense of the patent claims.

[0089] In the method as shown in FIG. 1, the data DA are at first discretized in a discretization step S1. To do so, the values of those variables in data sets DS referring to numerical variables with numerical values are assigned to a predetermined number of levels. In the embodiment described herein, all variables of the respective data sets DS refer to numerical variables except the variables T1, T4, T6 and T10 which are categorical variables defining corresponding categories with respect to the wind turbine. The discretization step S1 is necessary to enable the learning of Bayesian networks as described below.

[0090] Different methods for the discretization step S1 can be used. In the embodiment herein, the discretized data are generated by the well-known Hartemink's information-preserving discretization which retains as much information as possible. The Hartemink's information-preserving discretization is well-known for a skilled person and thus will not be described in detail herein.

[0091] When applying the Hartemink's information-preserving discretization, a corresponding number of levels for each variable to be discretized needs to be defined beforehand. In the embodiment herein, the discretization is performed such that all damage variables have three levels of discretization. In case that a damage variable represents a number (e.g. a number of occurrences), one level refers to a high number, another level to a medium-sized number and yet another level to a low number. In case that a damage variable represents an area (e.g. the total area of all erosions), one level refers to a large area, another level to a medium-sized area and yet another level to a low area. Contrary to that, for the numerical levels referring to turbine variables and weather variables, a number of seven levels was used for the Hartemink's information-preserving discretization.

[0092] The discretization step Si will result in modified data sets DS′ which include the corresponding levels for the discretized variables. Based on these modified data sets DS′, a well-known structure learning of a plurality of Bayesian networks is performed in step S2. A Bayesian network refers to a directed acyclic graph comprising the variables as nodes and defining the probabilistic dependencies between nodes by respective directed edges. This structure learning is performed several times using the data sets DS′ based on different learning methods. For example, each learning method results in another Bayesian network of the plurality of Bayesian networks. Examples of such learning methods have been mentioned above. In the embodiment described herein, learning methods from the R package “bnlearn” also mentioned above has been applied to the data sets DS′. In a variant, about 20 or more Bayesian networks, e.g. 25 Bayesian networks, are generated in step S2 by different learning methods. Those Bayesian networks are indicated as BN1, BN2, , BNN in FIG. 1.

[0093] The Bayesian networks BN1 to BNN are thereafter processed by step S3 in which several cross validations are performed for each of the Bayesian networks BN1 to BNN. In a cross validation performed for a corresponding Bayesian network, parameter learning of this Bayesian network based on a first part of the modified data sets is performed, resulting in conditional probabilities between variables representing nodes linked by respective directed edges in the Bayesian network. Thereafter, a prediction of the values of the damage variables of a second part of the modified data sets different from the first part is performed by the corresponding Bayesian network in combination with the condition probabilities based on the values of the turbine variables and the weather variables of the second part of the modified data sets.

[0094] As mentioned above, several cross validations are performed for each Bayesian network. However, the method may also be used by only performing one cross validation for each Bayesian network. In case of several cross validations, the first part of the modified sets used for parameter learning and the second part of the modified sets used for prediction are chosen differently between the cross validations.

[0095] After having performed the cross validations, the well-known F1 score is determined for each cross validation of a respective Bayesian network. The F1 score refers to a higher prediction quality the higher the score is. The F1 score is calculated based on the predicted and actual values of the damage variables of the second part of the modified data sets obtained for a corresponding cross validation. All F1 scores of the cross validations performed for a respective Bayesian network are averaged and form a prediction measure PM. In case that only one cross validation is performed for each Bayesian network, the F1 score calculated for this cross validation directly corresponds to the performance measure PM.

[0096] Out of the plurality of Bayesian networks BN1 to BNN, the network is identified having the best performance measure PM (i.e., the highest average F1 score) reflecting the best prediction quality. This network is chosen as the optimum Bayesian network OBN and forms the result of step S3 in FIG. 1.

[0097] In a next step S4, a parameter learning of the optimum Bayesian network OBN is performed based on all modified data sets DS′, resulting in conditional probabilities between variables representing nodes linked by respective directed edges in the optimum Bayesian network. As mentioned above, parameter learning methods for Bayesian networks are well-known for a skilled person. The optimum Bayesian network in combination with those conditional probabilities forms the prediction model which is the result of the method of FIG. 1.

[0098] FIG. 2 shows an example of an optimum Bayesian network OBN obtained by the method as described above. As can be seen in FIG. 2, the Bayesian network OBN comprises a plurality of nodes where each node corresponds to one of the variables T1 to T10, W1 to W4 and D1 to D19 as described before. Probabilistic dependencies between variables are expressed by directed edges. For the sake of clear presentation, only a few of those edges are indicated with reference numeral E in FIG. 2. For each node, there exists a conditional probability table with respect to its parent nodes, i.e., the nodes from which edges terminating in the respective node originate. Within a table, each conditional probability refers to a probability that the variable presented by the node assumes a specific value in case that the parent nodes assume specific values. In other words, conditional probabilities for all possible values of the variable of a respective node under all possible conditions of values of variables representing parent nodes are defined for the nodes in the Bayesian network.

[0099] With the prediction model obtained in step S4 of FIG. 1, damage values for an arbitrary wind turbine based on corresponding values of turbine variables specifying the turbine and weather data around the turbine can be predicted where prediction means that the optimum Bayesian network OBN provides corresponding probabilities for the levels (i.e., discretized values) of the respective damage variables. This information is used in order to evaluate whether damages and particularly which damages are likely to occur at a wind turbine. The wind turbine may refer to a turbine already in use. Nevertheless, the prediction may also be applied to wind turbines not yet built or installed at the corresponding site in order to evaluate the risk of damages beforehand and in order to determine which wind turbine is suitable at a specific location in order to minimize the risk of damages.

[0100] The predictions by the above prediction model are generated in combination with a suitable user interface (particularly a display) making it possible for a user to specify corresponding prediction tasks for wind turbine damages via the user interface and to receive the output of the prediction tasks via the user interface.

[0101] In an embodiment, the prediction model provided by the method described above is updated in regular intervals. To do so, new available data sets are downloaded from a cloud system to the computer which generates the prediction model. After having downloaded new available data, the method steps S1 to S4 as described with respect to FIG. 1 are performed once again based on the old data sets in combination with the new available data sets. As a result, an updated prediction model is generated which can thereafter be used for the prediction of damage variables.

[0102] Embodiments of the invention as described in the foregoing has several advantages. Particularly, an expanded and improved rotor blade damage prediction can be achieved. This prediction does also consider the weather conditions around wind turbines besides other turbine parameters. Contrary to that, previous theoretical models cannot account for such real-world factors and only will end up with imprecise static results. The prediction model generated by embodiments of the invention can lead to a better planning of logistics and services provided for a wind turbine. Furthermore, the risk of rotor blade damages can be predicted even before installing a wind farm making it possible to choose appropriate wind turbines for the farm at the site where the farm shall be installed.

[0103] Although the present invention has been disclosed in the form of embodiments and variations thereon, it will be understood that numerous additional modifications and variations could be made thereto without departing from the scope of the invention.

[0104] For the sake of clarity, it is to be understood that the use of “a” or “an” throughout this application does not exclude a plurality, and “comprising” does not exclude other steps or elements.