A method of controlling wind turbines of a wind farm using a trained AI model

Abstract

A method of controlling wind turbines of a wind farm using a trained AI model is disclosed. Incident signal data is obtained from a plurality of data-providing wind turbines, the incident signal data comprising incident signals generated by the data-providing wind turbines, and the incident signal data is fed to an artificial intelligence model, by means of which patterns are identified in the incident signals generated by the data-providing wind turbines from the incident signal data. One or more actions are associated with the identified patterns based on identified actions performed by the data-providing wind turbines in response to the generated incident signals. During operation of the wind turbines of the wind farm, one or more incident signals from one or more wind turbines of the wind farm are detected and compared to the patterns identified by the AI model. In the event that the detected incident signals match at least one of the identified patterns, the wind turbines of the wind farm are controlled by performing the action associated with the matching pattern.

Description

Technical Field

[0001] This invention relates to a method for controlling wind turbines in a wind farm using a trained artificial intelligence (AI) model. The method according to the invention allows wind turbines to react early to extreme events, such as extreme weather events, thereby reducing the risk of damage to the wind turbines. Background Technology

[0002] When wind turbines experience extreme conditions (such as high wind speeds exceeding their design limits), they typically need to be shut down to protect them. To ensure this, wind turbines are equipped with safety systems that automatically shut down the turbine or perform other relevant actions when certain limits are exceeded, such as when the average wind speed exceeds a predefined threshold for a specified time interval (usually 10 minutes), or when other types of alarms are triggered.

[0003] Sometimes, the weather changes rapidly and is unpredictable. In such cases, the wind turbine's safety system can detect when other design parameters of the wind turbine are exceeded, and will then shut down the wind turbine accordingly.

[0004] When incidents such as anomalies, malfunctions, or exceeding design limits are detected during wind turbine operation, an incident signal corresponding to the detected incident is generated. Incident signals may take the form of warnings to alert operators, alarms that may cause the wind turbine to shut down automatically, or any other suitable type of incident signal. Incident signals typically have a unique code that identifies the type of incident that triggered the signal.

[0005] In rare cases, when extreme weather events move across a wind farm, some of the wind turbines may detect an alarm condition that would cause them to shut down, while other turbines may not detect such an alarm condition and continue operating. This could damage the turbines that continue operating, for example, until those turbines also detect the alarm condition and shut down. Summary of the Invention

[0006] The purpose of embodiments of the present invention is to provide a method for controlling wind turbines in a wind farm, wherein the risk of damage to wind turbines during extreme weather events is reduced.

[0007] This invention provides a method for controlling wind turbines in a wind farm, wherein the wind farm includes multiple wind turbines, and the method includes the following steps:

[0008] - Accident signal data is obtained from multiple data-providing wind turbines, the accident signal data including accident signals generated by the data-providing wind turbines.

[0009] - The accident signal data is fed into an artificial intelligence (AI) model, and the AI ​​model is trained using the accident signal data to identify patterns in accident signals generated by the wind turbines providing the data.

[0010] - Identify actions performed by wind turbines in response to generated accident signals, based on which one or more actions are associated with the identified patterns.

[0011] - During the operation of wind turbines in a wind farm, detect one or more fault signals from one or more wind turbines in the wind farm.

[0012] - Compare the detected accident signals with patterns identified by the AI ​​model, and

[0013] - If a detected accident signal matches at least one of the identified patterns, the wind turbines of the wind farm are controlled by executing an action associated with the matching pattern.

[0014] Therefore, the method according to the invention is a method for controlling wind turbines in a wind farm. In the current context, the term "wind farm" should be interpreted as a collection of two or more wind turbines arranged within a limited geographical area, which share various infrastructures (such as access roads, communication lines, grid connections, substations, etc.). Thus, a wind farm comprises multiple wind turbines.

[0015] In the method according to the invention, incident signal data is initially obtained from a plurality of data-providing wind turbines. In the current context, the term "data-providing wind turbine" should be interpreted as any wind turbine that contributes to the incident signal data obtained or collected during the first step of the method. Thus, the incident signal data being collected originates from a plurality of data-providing wind turbines.

[0016] The wind turbines that provide data are typically other wind turbines besides those that form part of a wind farm, such as wind turbines that form part of another wind farm. However, it is not excluded that some or all of the wind turbines that form part of a wind farm (i.e., wind turbines controlled by means of the method according to the invention) provide accident signal data, thus forming a group of multiple wind turbines that provide data.

[0017] Accident signal data includes accident signals generated by the wind turbines providing the data. In the current context, the term "accident signal" should be interpreted as a signal generated by one of the data-providing wind turbines in response to an accident experienced by the data-providing wind turbine. For example, an accident could be one or more sensor readings falling outside specified limits. In this case, the sensors could, for example, measure environmental or weather conditions such as wind speed, wind speed variation, wind direction, wind direction variation, turbulence, gust conditions, ambient temperature, humidity, precipitation, air density, etc. Alternatively or additionally, the sensors could measure various parameters of the wind turbine, such as the temperature of various parts of the wind turbine, blade deflection, tower deflection, tower oscillation, yaw error, etc.

[0018] Therefore, an incident signal is a signal generated by the data-providing wind turbine and indicates that the data-providing wind turbine has experienced a specific incident, such as high winds, a malfunction, or excessive deflection. Thus, incident signals can take the form of warnings and / or alarms generated by the data-providing wind turbine. The incident signal thereby provides information that a given data-providing wind turbine has experienced a specific incident that triggered the generation of the corresponding incident signal by the data-providing wind turbine.

[0019] Because accident signals triggered by a given incident typically identify potential incidents (e.g., by providing a unique code corresponding to the potential incident for the accident signal), accident signals from various types of wind turbines can be applied, even if design constraints regarding environmental conditions and operating parameters vary from one type of wind turbine to others. Therefore, it is not important which specific environmental conditions or sensor values ​​triggered the accident signal. The key is the conditions experienced by the given wind turbine providing the data, which caused it to generate a specific type of accident signal. This significantly increases the number of wind turbines eligible as data providers for this purpose.

[0020] Accident signal data may also include information about the actions taken by the wind turbine that provided the data in response to the generated accident signal.

[0021] Accident signal data can be retrieved from historical operating data from a large number of wind turbines (e.g., from a fleet of wind farms). Alternatively or additionally, training data can be specifically obtained from a specified number of wind turbines (e.g., wind turbines in a wind farm).

[0022] Next, the accident signal data is fed into an artificial intelligence (AI) model, which is then trained using the accident signal data. This allows the model to identify patterns in the accident signals generated by the wind turbines that provide the data.

[0023] The identified patterns may include, for example, combinations and / or sequences of various incident signals generated by data-providing wind turbines. Furthermore, these patterns may include combinations of incident signals generated by data-providing wind turbines arranged close to each other (e.g., data-providing wind turbines arranged within the same wind farm).

[0024] Next, the actions taken by the wind turbines providing the data in response to the generated accident signals are identified. As mentioned above, this information can be incorporated into the accident signal data fed into the AI ​​model. Alternatively, this information can be added when the AI ​​model identifies a pattern.

[0025] Actions may include, for example, shutting down the wind turbine, changing the upper and / or lower limits of one or more operating parameters (e.g., by narrowing the operating range of one or more parameters), or switching the wind turbine to a safe mode.

[0026] Based on the identified actions, one or more actions are associated with patterns identified by the AI ​​model. Thus, a connection or correlation is established between accident signal patterns, such as accident signal sequences, and actions performed by the wind turbine. Since this is accomplished based on actual actions performed by the wind turbine in response to the generation of actual accident signals, it can be assumed that the actions associated with a given accident signal pattern are appropriate for the wind turbine to perform when it experiences that accident signal pattern. This association between actions and patterns can be called "labeling," and patterns with associated actions can be called "labeled patterns."

[0027] Next, during the operation of the wind turbines in the wind farm (i.e., wind turbines controlled according to the method of the present invention), one or more accident signals are detected from one or more of the wind turbines in the wind farm.

[0028] The detected accident signals are compared with patterns identified by the AI ​​model. If a match exists between the detected accident signal and at least one of the patterns identified by the AI ​​model, it can be assumed that the wind turbine generating the accident signal is experiencing conditions similar to those experienced by the data-providing wind turbine that generated the accident signal leading to that pattern. Therefore, it can be assumed that the wind turbine now generating a similar accident signal can appropriately perform actions also performed by the data-providing wind turbine. Thus, in this case, the wind turbines of the wind farm are controlled by performing actions associated with the matching pattern.

[0029] Therefore, once an accident signal generated by a wind turbine is identified as following a pattern that is likely to cause the wind turbine to automatically activate in the near future, the given wind turbine can be controlled to perform such an appropriate action, such as shutting it down. Thus, actions to protect the wind turbine from damage in extreme weather events can be implemented at an early stage, significantly reducing the risk of severe damage to the wind turbine.

[0030] Pattern recognition by AI models is an advantage because it significantly reduces the amount of manual work required. Furthermore, it allows for the identification of patterns that are not necessarily apparent to the human eye. For example, discovering or describing complex patterns and / or correlations in generated accident signals can require tremendous effort, potentially rendering the task impossible to perform manually.

[0031] The method may also include the step of monitoring wind conditions at the site of the wind farm, and the step of controlling the wind turbines of the wind farm may also be based on the monitored wind conditions.

[0032] According to this embodiment, when selecting an action to be performed by the wind turbine in response to a detected match between an accident signal generated by the wind turbine and one or more identified patterns, the prevailing wind conditions at the wind farm site are also considered. For example, wind conditions may be considered as part of a pattern formation. In this case, an identified pattern may be a combination of wind speeds exceeding a certain threshold and one or more warnings and / or alarms generated based on sensor signals from the wind turbine. Furthermore, certain wind conditions (such as high wind speeds) may generate the generated accident signal, in which case such detected wind conditions would naturally form part of the accident signal pattern.

[0033] Examples of wind conditions include, but are not limited to, wind speed exceeding a certain threshold, wind speed increase exceeding a certain threshold, gust conditions, wind direction, wind direction change exceeding a certain threshold, etc.

[0034] As mentioned above, incident signals can include warnings and / or alarms. In the current context, the term "warning" should be interpreted as a signal generated by the wind turbine to remind the operator that something needs attention, but the detected anomaly is not currently critical to the continued operation of the wind turbine. Therefore, in response to a warning, the wind turbine will typically not shut down, but it may operate in a protected or degraded mode.

[0035] In the current context, the term "alarm" should be interpreted as a signal generated by the wind turbine indicating that a critical fault or condition has been detected. An alarm is likely to cause the wind turbine to shut down automatically.

[0036] Actions associated with the identified pattern may include shutting down one or more wind turbines. Alternatively or additionally, actions may include operating one or more wind turbines in protected and / or degraded modes, changing the limits of the parameter range of one or more control parameters and / or sensor parameters, and / or any other suitable type of action.

[0037] The steps for controlling the wind turbines of a wind farm may include shutting down one or more wind turbines at the wind farm before initiating a shutdown procedure at one or more wind turbines based on a warning or alarm generated by the wind turbines themselves.

[0038] According to this embodiment, upon detecting that a wind turbine is generating an accident signal matching a pattern that could cause it to shut down, the wind turbine is immediately shut down, rather than waiting for a subsequent accident signal that would likely cause the wind turbine's safety system to shut it down. Therefore, the wind turbine shuts down at an earlier stage, reducing the risk of damage, such as due to severe weather conditions.

[0039] The method may also include the following steps:

[0040] - Obtain fault signal data from the wind turbines of the wind farm during operation.

[0041] - Feed the acquired accident signal data into the AI ​​model, and

[0042] - Retrain the AI ​​model based on accident signal data to improve the AI ​​model.

[0043] According to this embodiment, the AI ​​model is continuously retrained and improved based on data obtained from operating wind turbines.

[0044] The method may also include the step of determining that an extreme weather event is occurring based on comparing the detected accident signal with a pattern identified by an AI model.

[0045] According to this embodiment, historical accident signal data can reveal that when certain extreme weather events are occurring, certain accident signal patterns are likely to be generated by wind turbines experiencing those events. Therefore, when an accident signal pattern previously associated with an extreme weather event is detected, it is determined that such an extreme weather event is likely to occur, allowing for preventative measures to be taken at an early stage. For example, if an extreme weather event is determined to be occurring, the wind turbine that generated the associated accident signal pattern can be shut down. Furthermore, depending on the type of extreme weather event, nearby wind turbines, wind turbines directly downstream of the wind turbine, or possibly all wind turbines in a wind farm may be additionally shut down.

[0046] Examples of extreme weather events include, but are not limited to, extremely high wind speeds, rapid changes in wind direction, extreme turbulence conditions, and tornadoes.

[0047] The patterns identified by the AI ​​model can include the same or similar accident signals generated by two or more wind turbines providing data, located in the same wind farm.

[0048] According to this embodiment, a given fault signal pattern is not necessarily based on fault signals generated by a single wind turbine, but includes fault signals generated by two or more wind turbines arranged in the same wind farm and thus near each other.

[0049] For example, this pattern could include the same type of warning or alarm generated by two or more wind turbines located in the same wind farm, possibly with a time delay between them. For instance, an upstream wind turbine might initially detect wind speeds exceeding design limits and generate a corresponding warning or alarm. Subsequently, high wind speeds will travel through the wind farm, causing downstream wind turbines to generate the same warning or alarm in sequence. When this pattern is detected, it can be assumed that further downstream wind turbines will also detect high wind speeds in the near future. Therefore, actions taken by the upstream wind turbines in response to the generated warnings or alarms can be immediately also performed on the downstream wind turbines, thereby reducing the risk of damage to these downstream wind turbines.

[0050] Alternatively or additionally, one or more of the identified patterns may be based on the same or similar accident signals generated by two or more wind turbines providing data, located in two or more wind farms.

[0051] The method may further include the step of controlling at least one of the wind turbines in the wind farm by performing an action in response to a match between an accident signal detected by another wind turbine in the wind farm and at least one of the identified patterns.

[0052] As mentioned above, if the matching pattern corresponds to a situation that may affect other wind turbines in the wind farm besides the wind turbine that generated the accident signal, then it may be relevant to actively perform relevant actions on such other wind turbines, such as shutting them down, degrading them, or changing the range of operating parameters. Attached Figure Description

[0053] The invention will now be described in more detail with reference to the accompanying drawings, in which:

[0054] Figure 1 The illustration depicts an extreme weather event in the form of a tornado traveling through a wind farm, and

[0055] Figure 2The illustration shows the control of a wind turbine in a wind farm according to an embodiment of the present invention. Detailed Implementation

[0056] Figure 1 The diagram shows a wind farm 1 comprising multiple wind turbines 2. An extreme weather event in the form of a tornado 3 is traveling through the wind farm 1 in the direction indicated by arrow 4.

[0057] Tornado 3 has passed over wind turbines 2a and 2b. In response, wind turbines 2a and 2b have generated one or more accident signals, such as warnings and / or alarms. In response to the generated accident signals, wind turbines 2a and 2b have performed one or more related actions to protect themselves from the effects of tornado 3. For example, wind turbines 2a and 2b may have shut down.

[0058] Since tornado 3 has passed both wind turbines 2a and 2b, the sequence or pattern of incident signals generated by the two wind turbines 2a and 2b is likely to be the same or at least very similar. However, since tornado 3 passed wind turbine 2a before passing wind turbine 2b, there may be a time delay between the incident signal generated by wind turbine 2a and the corresponding incident signal generated by wind turbine 2b.

[0059] Tornado 3 is about to pass by wind turbine 2c. Therefore, wind turbine 2c may have generated at least some of the accident signals generated by wind turbines 2a and 2b. For example, in the manner already described above and / or referred to below. Figure 2 In a further detailed description, the accident signal generated by the wind turbine 2c is compared with the accident signal pattern identified by the aid of an AI model.

[0060] This comparison can reveal a match between one of the identified patterns and an incident signal generated by wind turbine 2c. It can also be further revealed that the matching pattern matches incident signals generated by wind turbines 2a and 2b, and that the generated incident signals are compatible with the passage of tornado 3. Therefore, once such a match is identified, wind turbine 2c can be controlled based on this (e.g., by shutting down wind turbine 2c), even if warnings and / or alarms generated by wind turbine 2c itself have not yet caused wind turbine 2c to shut down.

[0061] Furthermore, since tornado 3 is moving towards wind turbine 2d, wind turbine 2d may have already generated, for example, the first in a sequence of accident signals corresponding to those generated by wind turbines 2a, 2b, and 2c. As mentioned above, it has been determined that wind turbines 2a, 2b, and 2c generate the same or similar accident signal sequences, and this pattern is compatible with tornado 3 passing through wind farm 2. The fact that the three wind turbines 2a, 2b, and 2c have generated the same or similar accident signal sequences can also be considered an accident signal pattern. Therefore, once wind turbine 2d generates the first accident signal of the accident signal pattern, it can be concluded that tornado 3 is likely to reach wind turbine 2d soon. Therefore, relevant actions, such as shutting down wind turbine 2d, can be performed at this stage.

[0062] Figure 2 The illustration shows the control of a wind turbine 2 in a wind farm 1 according to a method based on an embodiment of the present invention.

[0063] Multiple data-providing wind turbines 5 (three of which are shown) arranged in multiple data-providing wind farms 6 (one of which is shown) collect data during operation of the data-providing wind turbines 5. The collected data includes incident signal data, which includes incident signals (such as warnings and alarms) generated by the data-providing wind turbines 5. The incident signal data may also include information about actions performed by the data-providing wind turbines in response to the generated incident signals.

[0064] The data-providing wind turbine 5 provides fault signal data to the local data center (Hub) 7 of the data-providing wind farm 6. Fault signal data from the local data centers 7 of each of the data-providing wind farms 6 are provided to the central data center 8, which then feeds the fault signal data to the AI ​​model 9.

[0065] Wind turbine 2 of wind farm 1 can also act as a data-providing wind turbine in the manner described above. In this case, the accident signal data generated by wind turbine 2 is also provided to AI model 9 in the same manner as the accident signal data generated by wind turbine 5 of wind farm 6.

[0066] The AI ​​model is trained based on accident signal data to identify patterns in the accident signals generated by the data-providing wind turbine 5. Furthermore, the pattern is labeled by associating the actions performed by the data-providing wind turbine 5 in response to the generated accident signals. Therefore, the labeled patterns indicate which combinations of accident signals were generated by the data-providing wind turbine 5, and what actions it performed in response.

[0067] The trained model is provided to a local detector 10 at the wind farm 1 to be controlled. During the operation of the wind turbines 2 at the wind farm 1, the wind turbines 2 collect data (including fault signals generated by the wind turbines 2) and provide it to the local data center 11 of the wind farm 1. The generated fault signals are then compared with patterns identified by the AI ​​model. If a match exists between a fault signal generated by at least one of the wind turbines 2 and at least one of the identified patterns, the detector 10 instructs the relevant wind turbine 2 to perform an action associated with the matching pattern, such as shutting down one or more of the wind turbines 2.

[0068] In addition, the local data center 11 provides the accident signal generated by the wind turbine 2 to the central data center 8, which in turn feeds the data into the AI ​​model. The AI ​​model is then retrained based on the additional accident signal to obtain an updated and improved model.

[0069] The trained data model can be further assigned to other wind farms (such as wind farm 6 that provides the data). In this case, the trained model is provided to the detector 12 of the wind farm 6 that provides the data, and the wind turbine 5 that provides the data can be controlled in the same manner as described with reference to the wind turbine 2 of wind farm 1.

Claims

1. A method for controlling wind turbines (2) in a first wind farm (1), the method comprising the following steps: - Accident signal data is obtained from multiple data-providing wind turbines (5) located in multiple wind farms other than the first wind farm, wherein the accident signal data includes accident signals generated by the data-providing wind turbines (5). - The accident signal data is fed into an artificial intelligence (AI) model (9) and trained using the accident signal data to identify patterns in the accident signals generated by the data-providing wind turbines (5), wherein the patterns identified by the AI ​​model include combinations of the same or similar accident signals generated by two or more data-providing wind turbines located in the second wind farm. - Identify one or more actions performed by the wind turbine providing the data in response to an accident signal generated by the wind turbine (5) providing the data, and based on this, associate the one or more actions with the identified pattern. - During the operation of each wind turbine (2) in the first wind farm (1), detect fault signals from one or more wind turbines (2) in the first wind farm (1). - Compare the fault signals from one or more wind turbines in the first wind farm with the patterns identified by the AI ​​model (9), and - In response to determining that an accident signal from one or more wind turbines in the first wind farm matches at least one of the patterns identified by the AI ​​model, control the one or more wind turbines (2) in the first wind farm (1) by performing one or more actions associated with the at least one pattern.

2. The method according to claim 1, further comprising the step of monitoring wind conditions at the site of the first wind farm (1), wherein the step of controlling the one or more wind turbines (2) of the first wind farm (1) is also based on the wind conditions monitored at the site of the first wind farm (1).

3. The method of claim 1 or 2, wherein, The incident signals include warnings and / or alarms.

4. The method of claim 1 or 2, wherein, One or more actions associated with the identified pattern include shutting down one or more wind turbines (5).

5. The method of claim 1 or 2, wherein, The steps of controlling the one or more wind turbines (2) in the first wind farm (1) include shutting down the one or more wind turbines (2) in the first wind farm (1) before initiating a shutdown procedure at the one or more wind turbines (2) based on a warning or alarm generated by the one or more wind turbines (2) themselves.

6. The method according to claim 1 or 2, further comprising the following step: - During the operation of the first wind farm (1), fault signal data is obtained from one or more wind turbines (2) in the first wind farm (1). - Feed the accident signal data to the AI ​​model (9), and - The AI ​​model (9) is retrained based on the accident signal data, thereby improving the AI ​​model (9).

7. The method according to claim 1 or 2, further comprising the following step: The determination of an extreme weather event is based on a step of comparing accident signals from one or more wind turbines in the first wind farm with patterns identified by an AI model.

8. The method according to claim 1 or 2, further comprising the following step: Control at least one of the wind turbines (2) in the first wind farm (1) by performing actions in response to a match between an accident signal detected by another wind turbine (2) in the first wind farm (1) and at least one of the identified patterns.

Images