Method for computer-implemented forecasting of wind phenomena having an impact on wind turbines

Abstract

The invention relates to a method for computer-implemented forecasting of wind phenomena having an impact on one or more wind turbines of a wind farm. At each of one or more time steps during the operation of the wind farm, the following steps are performed: in a first step, a digital image (IM) from an operational forecasting system based on a high-resolution simulation performed with a numerical weather prediction (NWP) model is obtained, the digital image (IM) being provided from the area of ​​the wind turbines. In a second step, a forecast of a class with the highest probability among a plurality of predefined classes is determined by processing the digital image (IM) by a trained data-driven model (MO), wherein the digital image (IM) is fed as a digital input to the trained data-driven model (MO), and the trained data-driven model (MO) provides the class with the highest probability as a digital output, wherein the number of classes corresponds to different meteorological conditions.

Description

[0001] The present invention relates to a method and apparatus for computer-implemented forecasting of wind phenomena having an impact on wind turbines.

[0002] The wind turbines of a wind farm comprise an upper part with a rotor and a nacelle on top of a tower, wherein the upper part can be rotated around a vertical yaw axis in order to change the yaw angle of the respective turbine. The yaw angle of the turbine is usually adjusted so that the rotor of the wind turbine faces the wind. For this purpose, a wind sensor (i.e., anemometer) is mounted on the respective wind turbine to estimate the wind direction. On the one hand, yaw misalignment can cause power losses in terms of electrical energy production in a wind farm and should be avoided. On the other hand, if the wind is strong or becomes fast, a wrong yaw angle may cause damage to the wind turbines of the wind farm.

[0003] In particular, changing wind conditions may be the result of local wind phenomenology mainly associated with complex terrain (i.e., complex earth surface in the surroundings of an onshore wind farm). Two well-known topographic wind phenomena that may have a direct impact on energy production, machine life, safety and maintenance of wind turbines are, among others, Downslope Windstorm (DSWS) and Hydraulic Jump (HJ). Such wind phenomena are not detectable by means of sensors of wind turbines of a wind farm.

[0004] It is therefore an object of the present invention to provide a simple method in order to be able to forecast wind phenomena having an impact on wind turbines within a wind park.

[0005] This object is solved by the independent patent claims. Preferred embodiments of the invention are defined in the dependent claims.

[0006] The invention provides a method for computer-implemented forecasting of wind phenomena having an impact on one or more wind turbines of a wind farm. The wind farm comprises at least one wind turbine. In the case of a plurality of wind turbines, the number of wind turbines may be arbitrary. Each of the wind turbines comprises an upper part on top of a tower, the upper part being pivotable about a vertical axis and having a nacelle and a rotor with rotor blades. The rotor is attached to the nacelle, and the rotor blades are rotatable by the wind about a horizontal rotor axis.

[0007] According to the method of the invention, at each time step of one or more time steps during operation of the wind farm, the following steps are performed.

[0008] In a first step, a digital image from an operational forecasting system based on a high-resolution simulation performed using a numerical weather prediction (NWP) model is obtained. In the following, the term "image" refers to a digital image. The term "obtaining an image" means that the image is received by a processor implementing the method of the present invention. The digital image is not an image taken by a camera, but is based on multiple high-resolution simulations of an operational forecasting system, which are based on NWP models, such as the well-known Weather Research and Forecasting (WRF) model. WRF is a NWP system designed to serve the needs of atmospheric research and operational forecasting. Numerical weather prediction refers to the simulation and prediction of the atmosphere using computer models, and WRF is a collection of software used for this. The obtained digital image is provided from the area of ​​the wind turbine.

[0009] The time interval between two consecutive time steps can be chosen arbitrarily and in particular regular. The time interval can be chosen from minutes to hours. The smaller the time interval chosen, the more accurate the forecast of the wind phenomenon will be.

[0010] In a second step, a prediction of a class with the highest probability among a plurality of predefined classes is determined by processing the digital image by a trained data-driven model. The digital image is fed as a digital input to the trained data-driven model, and the trained data-driven model provides the class with the highest probability as a digital output, wherein the number of classes corresponds to different meteorological conditions. The different meteorological conditions correspond to wind phenomena that may or may not have an impact on one or more wind turbines of the wind farm.

[0011] The method of the invention provides a simple and direct method for determining / forecasting wind phenomena with high accuracy, which wind phenomena may have an impact, in particular a negative impact, on one or more wind turbines with respect to energy production, machine life, safety and maintenance. To this end, a trained data-driven model is used. This model is trained by training data comprising a plurality of images from an operational forecasting system based on a plurality of NWP high-resolution simulations, together with information on the categories of different meteorological conditions to which the respective images belong.

[0012] Any known data-driven model learned by machine learning can be used in the method according to the invention. In a particularly preferred embodiment, the trained data-driven model is a neural network, preferably a convolutional neural network, which is particularly suitable for processing image data. However, other trained data-driven models, such as models based on decision trees or support vector machines, can also be implemented in the method of the invention.

[0013] In a preferred embodiment of the present invention, information based on the category of the corresponding meteorological condition with the highest probability is output via a user interface. For example, the category itself can be output via a user interface. Alternatively or additionally, a specific meteorological condition or wind phenomenon can be provided via a user interface. In the case where the meteorological condition or wind phenomenon corresponds to a condition or phenomenon that may have an impact on one or more wind turbines of a wind farm, a warning can be provided via a user interface. Thus, a human operator is informed of an upcoming critical condition so that he can initiate appropriate countermeasures by, for example, adjusting the yaw angle of the wind turbine or cutting back or stopping the turbine before the wind condition causes extreme loads on turbine components. Preferably, the user interface comprises a visual user interface, but it may also include another type of user interface (e.g., an acoustic user interface).

[0014] In another particularly preferred embodiment, the method of the invention generates a control command for the wind farm (i.e., the plurality of wind turbines of the wind farm) if the category with the highest probability corresponds to a meteorological condition that is considered to have an impact on one or more wind turbines of the wind farm. The control command enables some or all of the wind turbines to be triggered to change, for example, the yaw angle, etc., so that potential dangers can be minimized, or so that the generated electrical power of the wind farm can be enhanced. This embodiment enables automatic alignment of at least some of the wind turbines in advance of upcoming wind conditions. This reduces power losses due to yaw misalignment, or reduces the risk of reduced machine life or increased maintenance due to specific wind phenomena and meteorological conditions, respectively.

[0015] In another embodiment, the image is obtained from a high-resolution simulated cross section through the site of the wind farm or through a place close to the site of the wind farm. The place close to the wind farm may be, for example, a weather station near the wind farm, which monitors weather conditions for general control operations. The digital image may be, for example, a vertical cut through the wind farm. However, the cross section may also be selected in a different way.

[0016] In particular, the image illustrates the wind intensity and behavior in a cross section of a high-resolution simulation. In particular, the cross section of the high-resolution simulation lies in a plane perpendicular to the surface of the earth and extending along the prevailing wind direction. Since the wind speed (or wind intensity) is the only information necessary for forecasting wind phenomena having an impact on one or more wind turbines of a wind farm, the image may be a grayscale image, where the brightness corresponds to the wind speed. However, if other parameters will also be used for the analysis, the image may contain colors containing information about the selected parameters (e.g. temperature or turbulent kinetic energy).

[0017] This image is based on multiple high-resolution simulations that are the result of nesting a corresponding number of domains provided by the global WRF model. Each of the domains may use a different model to simulate and predict the atmosphere. Nesting of simulation data is a well-known means for increasing the resolution in a particular region of interest (i.e., the region where the wind farm is located). It is also known as "performing physical downscaling".

[0018] In addition to the above method, the invention also relates to an apparatus for computer-implemented forecasting of wind phenomena having an impact on one or more wind turbines of a wind farm. The apparatus is configured to perform the method according to the invention or one or more preferred embodiments of the method according to the invention.

[0019] Furthermore, the invention relates to a computer program product with a program code stored on a non-transitory machine-readable carrier for implementing the method according to the invention or one or more preferred embodiments thereof, when the program code is executed on a computer.

[0020] Furthermore, the invention relates to a computer program having a program code for implementing the method according to the invention or one or more preferred embodiments thereof when the program code is executed on a computer.

[0021] Embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] Figure 1 is a schematic diagram of an operation prediction system for providing high-resolution images for image recognition under deep learning technology according to the present invention.

[0023] Figure 2 A schematic diagram of a controller for performing an embodiment of the present invention is shown.

[0024] The method as described below provides a simple method for forecasting wind phenomena (also referred to as meteorological conditions) which have a particularly negative impact on one or more wind turbines of a wind farm. To this end, a plurality of images IM from an operational forecasting system based on high-resolution simulations are obtained. The operational forecasting system can be based, for example, on the well-known Weather Research and Forecasting (WRF) model. The WRF model is a numerical weather prediction (NWP) system designed to serve both atmospheric research and operational forecasting needs. NWP refers to the simulation and prediction of the atmosphere using computer models, and WRF is a collection of software for this. The model is used for a wide range of meteorological applications across scales ranging from a few meters to thousands of kilometers. WRF allows the generation of simulations that reflect real data (such as, observations or analysis). WRF provides operational forecasts.

[0025] exist Figure 1In , D1 represents the first domain provided by the global model in the WRF model. D2, D3 and D4 refer to different domains obtained from different simulations that achieve progressively better resolution for the region of interest. Figure 1 In the present example, the region of interest is represented by D4. The increase in resolution of the simulation from domains D1 to D2, D2 to D3 and D3 to D4 is achieved by means of known nesting.

[0026] Merely as an example, in the region of interest indicated by D4, two wind farms indicated by MET1, MET2 are located. Images for forecasting wind phenomena are obtained from sections CS1, CS2 of a high-resolution NWP simulation through the sites of the wind farms (i.e. MET1, MET2) or through places close to the sites of the wind farms. Such places close to the sites of the wind farms may be meteorological stations near the wind farms.

[0027] The image obtained from the sections CS1, CS2 of the high-resolution simulation in the region of interest (i.e., domain D4) shows the wind intensity in the sections CS1, CS2 of the high-resolution NWP simulation. Since the wind speed is the most relevant parameter, a grayscale image is used, where the brightness corresponds to the wind speed. In the image IM on the right side of the domain D4, the higher the wind speed, the lighter the grayscale color. The sections CS1, CS2 of the high-resolution NWP simulation from which the image IM for forecasting wind phenomena is obtained are located in a plane perpendicular to the earth's surface (i.e., at height H) and extending along the dominant wind direction DWD.

[0028] Although the cross sections CS1 , CS2 of the wind farms MET1 , MET2 are parallel in this example, this is not necessary. There may also be an angle different from 0 degrees between the cross sections CS1 , CS2.

[0029] For each location of the wind farm MET1, MET2, a plurality of images IM greater than 1 are obtained at predetermined time intervals. In the present example, the time interval between two images IM is chosen to be 10 minutes (starting from 00:00 in the case of the first image, from 00:10 in the case of the second image, from 00:20 in the case of the third image, etc.). However, it is to be understood that the time intervals may be chosen in different ways.

[0030] The corresponding image IM obtained from the operational forecast system based on the high-resolution NWP simulation is transmitted to the controller 10 of the wind farm ( Figure 2 ). The controller 10 comprises a processor PR implementing a trained data-driven model MO which receives as digital input the respective image IM and provides as digital output the class CL with the highest probability of the upcoming meteorological condition.

[0031] In the embodiments described herein, the trained data-driven model MO is based on a convolutional neural network (CNN) that has been pre-learned by means of training data. The training data comprises a plurality of images IM obtained from an operational forecasting system based on high-resolution images of the respective sites of the wind farm, together with information on the category of the current wind phenomenon or meteorological conditions. According to the prior art, convolutional neural networks are well known and are particularly suitable for processing digital images. A convolutional neural network typically comprises a convolutional layer, followed by a pooling layer and a fully connected layer in order to determine at least one category of the respective image or series of images, wherein the category according to the invention is one of a plurality of different meteorological conditions.

[0032] As an example, three different categories corresponding to normal weather conditions, downslope storm (DSWS), and hydraulic jump (HJ) may be selected.

[0033] In the embodiment presented herein, for each image IM, a category CL corresponding to a specific meteorological condition or wind phenomenon generated as an output of the model MO is generated. Figure 2 The output on the user interface UI is only schematically shown in FIG. Preferably, the user interface UI comprises a display. The user interface provides information for a human operator. The output based on the category CL of the meteorological condition may be the category CL or the meteorological condition itself, so that the operator is informed of the upcoming meteorological condition or wind phenomenon.

[0034] Alternatively or additionally, in case the category CL corresponds to a meteorological condition or a wind phenomenon having an expected negative impact on at least one wind turbine of the wind park, the output may be a warning.

[0035] It is to be understood that for each site of a wind farm, the forecast of wind phenomena or meteorological conditions is or can be performed separately. If wind farms are located close together, the forecast of wind phenomena or meteorological conditions can be done in a combined process, processing images from two sites of a wind farm in the model MO.

[0036] The category CL determined by the model MO also generates control commands CC which are provided to the controllers of at least some of the wind turbines of the wind park(s) in order to adjust, for example, the yaw angle of the turbines. The control commands CC cause the wind turbines to rotate, for example, about their respective yaw axes, by an angle sufficient to avoid a critical situation with respect to a critical load or in a manner that increases the generated energy production.

[0037] The invention as described in the foregoing has several advantages.

[0038] The method described above enables earlier detection of critical weather patterns. Wind turbines at a wind farm can be activated before the weather pattern damages the turbines. Thus, the safety and security of maintenance workers are increased.

[0039] The method as described above can also be applied in wind farm siting projects, and not just existing wind farms.

[0040] In particular, in wind farm siting projects, the method enables avoiding suboptimal placement of wind turbines during the siting process, as the method can be used not only for operational forecasts, but also for studying future wind farm locations. Given an area of ​​interest, it is possible to determine in advance how the wind flow behaves (e.g., daily simulations for a year combined with deep learning), and detect critical locations that are highly affected by extreme wind phenomena. In this way, dangerous locations can be avoided and help design more productive wind farms.

[0041] Knowing critical weather phenomena in advance enables reduced downtime of wind turbines of existing wind farms. Due to the possibility of reducing critical loads on wind turbines, an extended life of wind turbines can be achieved.

Claims

1. A method for computer-implemented forecasting of wind phenomena having an impact on one or more wind turbines of a wind farm, wherein at each of one or more time steps during operation of the wind farm, the following steps are performed: obtaining a digital image (IM) from an operational forecasting system based on a plurality of high-resolution simulations performed using a numerical weather prediction (NWP) model, the digital image (IM) being provided from the area of ​​the wind turbines; A prediction of a class with the highest probability among a plurality of predefined classes is determined by processing the digital image (IM) by a trained data driven model (MO), wherein the digital image (IM) is fed as a digital input to the trained data driven model (MO), and the trained data driven model (MO) provides the class with the highest probability as a digital output, wherein the plurality of classes correspond to different meteorological conditions, wherein the image is obtained from a high resolution simulated cross section through the site of the wind farm or through a place close to the site of the wind farm.

2. The method of claim 1, wherein the trained data-driven model (MO) is a neural network.

3. The method according to claim 1 or 2, wherein information based on the category with the highest probability is output via a user interface (UI).

4. The method according to claim 1 or 2, wherein a control command (CO) for the wind park is generated if the category with the highest probability corresponds to a meteorological condition which is considered to have a negative impact on one or more wind turbines of the wind park.

5. Method according to claim 1, wherein said image (IM) shows wind intensity in a cross section of a high resolution simulation.

6. The method of claim 1 or 5, wherein the cross section of the high-resolution simulation lies in a plane perpendicular to the earth's surface and extending along the prevailing wind direction.

7. The method of claim 1 or 2 or 5, wherein the image is a grayscale image, wherein brightness corresponds to wind speed.

8. The method of claim 2, wherein the neural network is a convolutional neural network.

9. An apparatus for computer-implemented forecasting of wind phenomena having an impact on one or more wind turbines of a wind farm, wherein the apparatus comprises a processor (PR) configured to perform the following steps at each of one or more time steps during operation of the wind farm: obtaining a digital image (IM) from an operational forecast system based on a plurality of high resolution simulations, the digital image (IM) being provided from an area of ​​the wind turbine; A prediction of a class with the highest probability among a plurality of predefined classes is determined by processing the digital image (IM) by a trained data driven model (MO), wherein the digital image (IM) is fed as a digital input to the trained data driven model (MO), and the trained data driven model (MO) provides the class with the highest probability as a digital output, wherein the plurality of classes correspond to different meteorological conditions, wherein the image is obtained from a high resolution simulated cross section through the site of the wind farm or through a place close to the site of the wind farm.

10. The apparatus according to claim 9, wherein the apparatus is configured to perform the method according to one of claims 2 to 8.

11. A computer program product having a program code stored on a non-transitory machine-readable carrier for implementing the method according to one of claims 1 to 8 when said program code is executed on a computer. 12 . A computer program product having a program code for implementing the method according to claim 1 when the program code is executed on a computer.

Images