WO2023083426A1

WO2023083426A1 - Wind power plant control scheme

Info

Publication number: WO2023083426A1
Application number: PCT/DK2022/050237
Authority: WO
Inventors: Mahmood MIRZAEI; Ewan MACHEFAUX; Poul Brandt Christensen
Original assignee: Vestas Wind Systems A/S
Priority date: 2021-11-12
Filing date: 2022-11-10
Publication date: 2023-05-19

Abstract

A control system, for controlling a flow of wind through a wind power plant comprising a plurality of wind turbine generators, comprising: a first controller, local to the wind power plant, configured to: receive operational data relating to the wind power plant and, in 5response, to send control signals to the plurality of wind turbine generators based on an implemented wind power plant control function, and a second controller. The second controller is configured to: receive operational data from the wind power plant; train a wind power plant model using the received operational data; generate a trained wind power plant control function based on the trained wind power plant model; and transmit the 10trained wind power plant control function to the first controller. Following the transmission of the trained wind power plant control function to the first controller, the first controller is further configured to: update the wind power plant control function with the trained wind power plant control function. The invention has particular utility in the context of controlling the flow of wind through the wind power plant, for example by controlling the wake steering 15and/or axial induction factor of the wind turbines in the power plant. A complex dynamic relationship exists between the control actions of wind turbines and the resultant wind flow through the power plant which is difficult to control accurately through conventional control approaches. The invention implements a machine learning approach to improve the operational performance of the control function over time, thereby to optimise the flow of 20wind through the wind park by controlling the yaw offset/axial induction factors of the wind turbines, which may achieve improved power production.

Description

WIND POWER PLANT CONTROL SCHEME

TECHNICAL FIELD

The invention relates to a system and method for controlling a wind farm using an open loop control or feed forward method. In particular, though not exclusively, the system and method involve controlling wind turbines in the wind farm collectively so as to mitigate wake effects and improve overall power production.

BACKGROUND

An important consideration for wind power plant operators is how to control the wind turbines in the wind power plant to maximise overall power production of the wind power plant whilst avoiding excessive dynamic loading. This is a difficult challenge. Conventionally, wind turbines within wind power plants have been operated at a local level, in which each wind turbine has a controller that is responsible for maximising the power output of the wind turbine in accordance with a power demand reference or setpoint that has been transmitted to it by a higher-level power plant controller. However, the proximity of wind turbines to each other in a wind power plant means that downstream wind turbines are affected by the wakes of upstream wind turbines. As is known in the art, wake effects are detrimental to the power production of downstream wind turbines.

WO2018/007012 discloses an approach of controlling wake effects which involves steering wake centres of wind turbines using closed loop feedback control. In this disclosure, LiDAR technology is used to determine the position of the wake centre of a wind turbine and a controller adjusts suitable wind turbine control variables such that the wake is redirected with respect to downstream wind turbines.

It is against this background to which the present invention is set.

SUMMARY OF THE INVENTION

According to an aspect of the invention there is provided a a control system, for controlling a flow of wind through a wind power plant comprising a plurality of wind turbine generators, comprising: a first controller, local to the wind power plant, configured to: receive operational data relating to the wind power plant and, in response, to send control signals to the plurality of wind turbine generators based on an implemented wind power plant control function, and a second controller. The second controller is configured to: receive operational data from the wind power plant; train a wind power plant model using the received operational data; generate a trained wind power plant control function based on the trained wind power plant model; and transmit the trained wind power plant control function to the first controller. Following the transmission of the trained wind power plant control function to the first controller, the first controller is further configured to: update the wind power plant control function with the trained wind power plant control function.

The invention also resides in a computer-implemented method of controlling a wind power plant comprising a plurality of wind turbines, the method comprises: receiving operational data relating to the wind power plant and, in response, sending control signals to the plurality of wind turbine generators based on an implemented wind power plant control function; sending operational data to a model training system remote from the wind power plant, receiving a trained wind power plant control function from the model training system; and updating the wind power plant control function with the trained wind power plant control function.

The invention extends to a computer program product comprising instructions which, when executed in a computing environment, are configured to carry out the method as defined herein.

The invention has particular utility in the context of controlling the flow of wind through the wind power plant, for example by controlling the wake steering and/or axial induction factor of the wind turbines in the power plant. A complex dynamic relationship exists between the control actions of wind turbines and the resultant wind flow through the power plant which is difficult to control accurately through conventional control approaches. The invention implements a machine learning approach to improve the operational performance of the control function over time, thereby to optimise the flow of wind through the wind park by controlling the yaw offset/axial induction factors of the wind turbines, therefore to improve power production and, therefore, profitability.

A particular benefit of the invention is that the model optimisation process is carried out by cloud-based computing resources. This means that such computing resources does not have to be provided at the power plant level. The model training can therefore be carried out with the most effective techniques and most powerful resources as available in the cloud without requiring updates of local computing hardware, firmware and software at the power plant level. Scalability of the solution is also benefitted because the same cloudbased computing resources can be used for multiple wind power plants.

Preferred and/or optional features of the invention are provided in the appended claims.

BRIEF DESCRIPTION OF THE DRAWINGS

One or more examples of the invention will now be described with reference to the accompanying drawings, in which:

Figure 1 schematically illustrates a wind power plant or wind park and an associated control system, including a main controller, and a controller training module;

Figure 2 illustrates schematically a control function of the main controller which links a plurality of inputs to output signals of the controller, which in this example are shows as wind turbine yaw offsets;

Figure 3 illustrates schematically a model of the wind power plant and its relationship to a set of input data and correspond output date, which is shown as the power production of the wind turbines in the wind power plant.

DETAILED DESCRIPTION

The invention provides a method and system for controlling total power output of a wind power plant or ‘wind park’ more effectively using an open loop or feed forward control regime. Aerodynamic interactions between the individual wind turbines within a wind power plant are complex. Wakes from one wind turbine in the power plant can affect others nearby. Induction also has a similar unpredictable effect between wind turbines. This means that it is a complex challenge to optimise the overall power production of a wind plant. As has been explained above, a feedback or ‘closed loop’ control approach is challenging to implement which is largely due to the time delays between adjusting control parameters and the effects being seen in power production. However, it is also challenging to evaluate whether the power generation changes are due to the adjustments to the control inputs or simply down to changes in the flow disturbances, such as wind speed and direction changes. The invention adopts an open loop control approach in which a control function implemented in the wind power plant provides suitable control action outputs, which may be yaw offset setpoints and/or collective pitch offset setpoints for each of the wind turbines in the wind power plant based on an input stream of operational data. The operational data may include sensed parameters regarding wind behaviour such as wind speed and wind direction, and also composite values such as turbulence, wind shear and wind veer, all of which may be estimated by software-driven estimator functions.

The objective of this is to control the wind turbines in the wind power plant based on an open loop or feedforward control methodology to improve overall power production over as wide a range of input conditions as possible. However, the control function implemented in the controller inherently has a degree of inaccuracy and therefore is prone to result in suboptimal power production of the wind power plant. The invention proposes an approach to address this challenge by implementing a training process such that the control laws/function in the power plant controller can be updated periodically with a revised, adapted or optimised control function in order to maximise the overall power production of the wind power plant.

The training process may make use of machine learning technology to analyse historical data related to the operation of the wind power plant thereby to gain insights in how wind turbine settings such as yaw offsets, which are linked to wake control, and collective pitch settings, which are linked to induction control, affect the total power production of the wind power plant. Using established model optimisation techniques, a wind power plant model can be refined in order to provide updated control settings to the controller.

Figure 1 is a schematic view that represents a control system 2 for a wind power plant 4 or ‘wind park’. The wind power plant 4 comprises a plurality of wind turbines 6 which are controlled by a power plant controller 8.

As has been mentioned, the invention concerns an approach for controlling a wind power plant in such a way to take account of the aerodynamic interactions between the plurality of wind turbines in the wind power plant in a better way to improve annual energy production (AEP). The terms ‘wind power plant’ and ‘wind park’ are therefore to be interpreted in this context, and to exclude wind installations including a single wind turbine. Typically, a wind farm to which this invention applies will have several tens of wind turbines and may exceed a hundred wind turbines. More specifically, irrespective of the size of wind park, the invention has utility in wind parks which have a low wake efficiency, which may be due to geographical topography, wind turbine layout and other factors.

The power plant controller 8 may have several functionalities. As would be understood by those skilled in the art, one of these functionalities is to act as a dispatcher to dispatch active and reactive power references or setpoints to the wind turbines 6 in order to satisfy the demands of a transmission system operator or ‘TSO’. This is functionality that is known in the art and is not the focus of this discussion.

The functionality of the power plant controller 8 that is the focus of the discussion is its role in controlling the wind turbines 6 in the wind power plant to adjust their yaw angle offsets (hereinafter referred to as “yaw offsets”) and collective pitch angle offsets (hereinafter referred to as “pitch offsets”) in order to control the flow of wind through the wind power plant and therefore optimise power production of the wind power plant as a whole. This functionality may be implemented by the same software environment of the power plant controller 8 or a separate software environment, for example the implementation may be on the same hardware in different software partitions, or the functionalities may be implemented in different hardware units. The term ‘power plant controller’ can therefore be considered to relate to a ‘wind flow controller’ in this context, rather than other control functionality such as control of active and reactive power output in response to grid demands. The power plant controller will therefore be referred to as the ‘wind flow controller’ from now on.

As is understood in the art, yaw offsets are used to steer the wakes of wind turbines to minimise the wake effects on downstream wind turbines. As is also understood, pitch angle offsets are used to adjust the axial induction of wind turbines which is also known to affect the individual power production of wind turbines in the wind power plant. Here, the discussion will focus on the adjustment of yaw offsets, but it should be understood that the techniques discussed here also apply to the adjustment of pitch angle offsets either alone or in combination with yaw offset adjustments.

In Figure 1 , the wind flow controller 8 implements an open loop or feed forward control approach. As such, the wind flow controller 8 is configured to output one or more control parameters in dependent on one or more input variables in accordance with a control function implemented by the wind flow controller 8. In a practical implementation, the wind flow controller 8 receives a plurality of input variables representing operational data of the wind power plant. The input variables may come from different sources. Firstly, the wind turbines 6 are equipped with a variety of sensors to monitor their operation. The turbine specific data may include data relating to the wind speed and/or wind direction at that location, wind turbine power production, local air temperature and so on. In some examples, the operational data includes only environmental data such as wind related conditions.

Secondly, the wind flow controller 8 may receive composite data which is not derived solely from sensor data. For example, software-based estimators may receive that sensor data and implement appropriate software algorithms to derive other variables of use to the wind flow controller 8. Such composite variables may be values of turbulence intensity, wind sheer and wind veer. The estimation of these values is generally known in the art and is not within the scope of the invention discussed here.

Thirdly, the wind flow controller 8 may receive data from external sources. Examples of external data are current energy price and wind/weather forecasts.

As shown in Figure 1 , the incoming data signals 10 from the wind turbines and data 12 from external sources are fed into a data interface module 14. It should be appreciated that it is optional that the signals which are used directly by the wind flow controller 8 are received by the data interface module 14 as an intermediate step. However, it is shown here for convenience. In Figure 1 , the wind turbine data is also shown directly input to the wind flow controller 8 by a dashed arrow for completeness.

In turn, the data interface module 14 provides input data 20 to the wind flow controller 8.

The function of the data interface module 14 is to be an intermediary between the wind flow controller 8 and the variety of data sources. As such the data interface module 14 is configured to transfer the incoming data form the data sources into a form that is required by the wind flow controller 8.

Figure 2 depicts a schematic view of the wind flow controller 8 associated with its input data 20 and its output data 22. As can be seen, the input data 20 comprises environmental data: wind speed, wind direction, turbulence intensity, wind shear, wind veer, and wind turbine status. It should be noted that all of these data items are related to each respective wind turbine in the wind power plant.

The wind flow controller 8 acts on the input data to provide an output data 22 as appropriate signals to respective wind turbines. Here, that output data 22 is a yaw offset signal for each of the plurality of wind turbines in the wind power plant 6. The skilled person will appreciate that the wind flow controller 8 acts in an open loop manner since it does not have any feedback data with which to minimise specific errors. Instead, the wind flow controller 8 implements suitable control function that maps input parameters to the yaw offset output.

The control function governing the wind flow controller 8 may be implemented in different ways, depending on the number in independent input variables. In a simple implementation for example with only two input variables, the function that maps the outputs to the inputs may be expressed as a simple two-dimensional array or look up table. However, in practice there are many input variables such that the function mapping inputs to outputs is required to be a type of complex algorithm such as a Support Vector Machine/Hyperplane or a neural network.

An initial model may be implemented as a comparatively low fidelity model of the wind power plant using engineering wake models to model the aerodynamic interactions between wind turbines. This initial model can then be modified and improved by the process described herein.

Returning to Figure 1 , from the above discussion, it will therefore be appreciated that the power plant controller 8 implements an open loop approach for controlling the yaw offsets of the wind turbines in the wind power plant 6.

To improve the performance of the wind flow controller 8, the system of the invention is provided with a learning system 24. The learning system 24 comprises a model training module 26 and an optimiser module 28. In overview, the learning module 24 receives operational data relating to the wind power plant and carries out an appropriate form of machine learning in order to optimise, or train, a software-implemented model of the wind power plant 8. Following the training of the model, the improved model can therefore be translated or converted into a trained control function suitable for being implemented into the power plant controller 8, as will be described in more detail later.

The learning system 24 also includes a control interface 29 that allow a user to interact with and control the model training module 26 and the optimiser module 28.

The model training module 26, the optimiser module 28 and the control interface 29 may be implemented in suitable computing equipment provided at the wind power plant. For example, that computing equipment may be housed within a substation of the power plant together with the computer equipment on which the wind flow controller 8 is implemented. It is also envisaged that the computing environment for the model training module 26 and the model optimiser 28 may be shared in the wind flow controller 8.

However, it is currently preferred that the computing resource on which the model training module 26, the optimiser module 28 and the control interface 29 is based is provided remotely from the wind flow controller 8 and, more specifically, that it is cloud-based.

Therefore, in the illustrated example, the learning system 24 is configured to received data from the wind power plant 2 over an internet-based communication network 30. Similarly, the learning system 30 is configured to transmit data to the wind power plant 2 over the same communications network 30. Appropriate communications protocols may be used for data communications over the network 30, as would be understood by the skilled person.

Being based in the cloud confers advantages for the learning system 24. Principally, basing the learning system 24 in the cloud enables the appropriate computing resources to be dedicated to the software functionality independent from the computing resource that is available at the wind power plant. This enables more sophisticated modelling and learning approaches to be used than would otherwise be practical. A further benefit is that the latest modelling approaches can be used as time goes on without the need to update every wind park with which they are used, since the training module is centralised in the cloud. Scalability of this approach is therefore particularly advantageous.

The model training module 26 is configured to execute a model of the wind power plant 6 based on a set of suitable training data. That training data is indicated as ’32’ in Figure 1. The training data is provided to it from a first memory store 34. The first memory store 34 is also implemented as a cloud-based computing resource and communicates over the network 30 with the wind power plant6. More specifically, data is provided to the first memory store 34 by the data interface module 14.

The function of the data interface module 14 is to provide an intermediate layer of data analysis and transformation between the learning system 24 and the data that flows from the wind power plant 2. As has been discussed above, the wind flow controller 8 uses a variety of data which is processed by its control function into suitable yaw offset output signals. Some of this data is direct sensor readings from the wind turbines. Other data comprises composite variables as would be suitably provided by estimators. However, the data rate of the data provided by the sensors in the wind turbines 8 may not be appropriate for processing by the learning system 24. For example, the data sampling frequency may be in the range of 1 Hz to 30Hz. The data interface module 14 therefore performs a suitable transformation of the data into a type of data that is appropriate to be used by the learning system 24, in addition to providing the operational data to the learning system 24 at an acceptable data rate, which may be at a lower data rate than the data is being provided to the wind flow controller 8.

For instance, the data interface module 14 may resample data relating to flow properties such as wind speed, wind direction, and turbulence intensity for each of the wind turbine locations. It may also resample power production values and various load parameters that are measured at each individual wind turbine. Resampling may be done at a lower rate than the generated sensor data. In terms of resampling, the relatively high frequency data rate in the context of the wind power plant (sample rate in the order of a few Hz for example) may be resampled by the data interface module 14 at a rate of between 1 and 5 minutes, by way of example. This provides a much lower transmission bandwidth that is required between the wind power plant and the training module 24.

Furthermore, the data interface module 14 carries out statistical processing on the received data in order to calculate other variables that may be useful for implementation in the training system 24. Such variables may be bulk wind direction estimation of any point in the wind park based on the turbine-specific wind direction data, wind shear data at various heights above ground level, dynamic loading variables based on turbine-specific loading data, and wake data such as wake position and direction information. Whereas the learning system 24 is envisaged as being implemented as a cloud-based computing resource, the data interface module 14 may be embodied as an edge-based computing resource. As shown in Figure 1 , the data interface module 14 is located geographically as part of the wind power plant 2. However, the computing resource that provides the data interface module 14 is connected by a suitable data communications channel to the wind turbines 6. The interconnections between the wind turbines 6 and the data interface module 14 may be direct or ‘point to point’ connections, or may be part of a local area network (LAN) operated under a suitable protocol (CAN-bus or Ethernet for example). Also, it should be appreciated that rather than using cabling, the control commands may be transmitted wirelessly over a suitable wireless network, for example operating under WiFi™ or ZigBee™ standards (IEEE802.11 and 802.15.4 respectively).

The data interface module 14 can therefore be considered to be an edge computing resource because it acts as a high-performance intermediary between the data source (the wind turbines) and the cloud-based resource. The benefit of this is that can reduce the volume of data that needs to be sent to the learning system 24 for analysis. Instead, the data interface module 14 performs suitable resampling data compression, statistical analysis techniques and so on in order to optimise the data prior to transmission to the learning system 24 in the cloud.

The data provided by the data interface module 14 is packaged and stored at the first storage module 34 in the cloud. The packaged data may therefore be configured to be used as a training data set for the model training module 26.

The model implemented by the model training module 26 may take various forms, as would be understood by the skilled person. For example, the model may be implemented as a neural network in a suitable format comprising an input neuron layer comprising a plurality of nodes, an output neuron layer comprising a plurality of output nodes, and one or more hidden layers of nodes. The precise form of model is not within the scope of the discussion and would be within the ambit of the skilled person with knowledge of machine learning technologies to specific and implement a suitable model. Another option is to use a so- called ‘grey box’ modelling approach which is based on an initial theoretical model structure which is then optimised using known data analysis techniques.

Refer to Figure 3, the training model module 26 is shown linked to input and output data 40, 42. As can be seen, the model processes turbine-related data provided by the first storage module 34 including: wind speed, wind direction, turbulence intensity estimate, wind shear estimate, wind veer estimate, wind turbine status data, and yaw offsets of the individual turbines. Other data may be included.

The data may relate to a suitable time period. For example, it is envisaged that the data is collected over a period of a few days to a few weeks, and stored as a training data set and a test data set as required. Therefore, the training data set can be considered to be historical operational data of the wind power plant.

The power plant model is then executed based on the training data. The output of the training process is a collection of values indicating the power production of the individual wind turbines.

As an initial step in the model training process, the output of the power plant model that is executed with the training data set is compared with the stored test data set in order to evaluate the accuracy of the power plant model. Known statistical analysis techniques can be used for this purpose as would be understood by the skilled person.

Following this, the power plant model can be trained to reduce the mismatch between the actual plant behaviour as observed in the test data, and the modelled behaviour based on the training data. The optimisation can be achieved using conventional training algorithms such as TensorFlow, SGD (Stochastic Gradient Descent) and other applicable tools.

Once the power plant model has been trained to match the behaviours of the power plant as close as possible based on the same input flow conditions, an optimisation step may be performed. The optimisation process may involve changing control inputs of the trained power plant model whilst the model is executed with a range of input flow conditions. The control inputs may be, for example, yaw offset settings for the wind turbines and/or axial induction settings for the wind turbines.

The result of the optimisation step is that control input settings are determined for an array of flow condition inputs which can be grouped into specific flow input ranges. This generates an association or mapping between the optimal control inputs and the flow conditions. The mapping may thus be implemented as the control function in the power plant controller. Returning to Figure 1 , once the optimiser module 26 has validated the trained model and generated the corresponding control function for thewind flow controller 8, it is operable to then transmit the generated control function to the wind flow controller 8 over the communications network 30. This is shown in Figure 1 by the data flow arrow 44.

The wind flow controller 8 is operable to update current control function with the newly received control function from the optimiser module 26. In one envisaged configuration, parallel instances of the control function may be executable within wind flow controller 8 to provide redundancy during the update procedure. To avoid the need to pause operation of the control function during the update, control authority may be taken by one of the parallel control functions whilst the other is being updated. Following completion of the update, authority may then be switched back to the updated control function by way of a smooth bumpless transfer

Provision may be made for version control of the control functions that are generated by the learning system 24 by way of a second memory module 36. Each generated control function may be stored in the second memory module 36, and be suitably identified by creation date and metrics which enable its effectiveness to be evaluated against other stored control functions.

The second memory module 36 is also envisaged to be a cloud-based computing resource.

As well as storing the generated control functions, the optimiser may also be configured to store the power plant model versions in the second memory module 36.

Storing the generated control functions for the wind flow controller, and also the power plant models, is advantageous as it allows those models to be analysed in the future. It also provides for reversion to earlier versions of the power plant model or generated control function if it is required.

Claims

1. A control system for controlling the flow of wind through a wind power plant including a plurality of wind turbine generators, the control system comprising: a first controller, local to the wind power plant, configured to: receive operational data relating to the wind power plant and, in response, to send control signals to the plurality of wind turbine generators based on an implemented control function; a second controller configured to: receive operational data from the wind power plant; train a wind power plant model using the received operational data; generate a trained control function based on the trained wind power plant model; transmit the trained control function to the first control module; wherein the first controller is further configured to: update the control function with the trained control function.

2. The control system of Claim 1 , wherein the second controller is implemented as a cloud computing resource that is remote from the wind power plant.

3. The control system of Claim 1 or Claim 2, further providing a data interface module configured to: monitor a plurality of operational parameter signals relating to the wind power plant; process the plurality of operational parameter signals to generate the operational data; and transmit the operational data to the second controller.

4. The control system of Claim 3, wherein the data interface module is further configured to: monitor the plurality of operational parameter signals at a first sampling rate, and resample at least some of the operational parameter signals at a second sampling rate that is lower than the first sampling rate.

5. The control system of Claim 4, wherein the data interface module is provided as an edge computing resource, such that the data interface module is configured to monitor the plurality of operational parameter signals over a communications network local to the wind power plant.

6. The control system of any preceding claim, wherein the control signals output from the first controller to the plurality of wind turbine generators include at least one of i) a yaw offset set point, and ii) a pitch offset set point.

7. The control system of Claim 6, wherein the operational data received by the control function comprise at least two or more of: wind speed, wind direction, turbulence intensity estimate, estimated wind shear, estimated wind veer, wind turbine status.

8. The control system of any one of the preceding claims, wherein the second controller receives said operational data and transmits the trained plant control function over a telecommunications network.

9. The control system of Claim 9, wherein the telecommunications network is the internet.

10. The control system of any one of the preceding claims, wherein the first controller is implemented within computing apparatus that forms part of the wind power plant.

11 . The control system of Claim 10, wherein the first controller is implemented within a wind turbine generator of the wind power plant. 15

12. The control system of Claim 10, wherein the first controller is implemented within a power plant controller of a wind power plant.

13. A computer-implemented method of controlling the flow of wind through a wind power plant comprising a plurality of wind turbines, the method comprising: receiving operational data relating to the wind power plant and, in response, sending control signals to the plurality of wind turbine generators based on an implemented wind power plant control function; sending operational data to a model training system remote from the wind power plant, receiving a trained wind power plant control function from the model training system; and update the wind power plant control function with the trained wind power plant control function.

14. The method of Claim 13, wherein the trained wind power plant control function is received from the learning system over a wide-area telecommunications network.

15. The method of Claims 13 or 14, further comprising transmitting, over a local-area telecommunications network of the wind power plant, a plurality of operational parameter signals to a data interface module connected to the local-area telecommunications network.

16. The method of Claim 15, further comprising: processing the plurality of operational parameter signals, by the data interface module, to derive the operational data that is sent to the model training system.

17. A computer program product comprising instruction which, when executed in a computing environment, are configured to carry out the method of Claims 13 to 16.