Wind power plant wake effect cooperative control method based on big data

Abstract

The invention relates to the technical field of wind power control, in particular to a wind power plant wake effect cooperative control method based on big data. The method comprises the steps that fan multi-source data are collected and preprocessed, the preprocessed data are subjected to normalization processing, and fan feature vectors are constructed; constructing a wind power plant directed graph, and calculating an edge weight based on the preprocessed data; based on the edge weights, constructing an improved graph neural network model to obtain hidden state vectors of the fan nodes, and calculating wake flow influence intensity between the fan nodes; and generating a wake flow coupling control factor based on the wake flow influence intensity between the fan nodes, and adjusting the original control action to realize the cooperative control based on the wake flow effect. The problems that a traditional wind power plant wake flow effect cooperative control method lacks explicit modeling of wake flow propagation paths between fans, space information and historical disturbance information between the fans are not fully utilized, and the fans are not guided to adjust control responses according to the current wake flow relation are solved.

Description

A collaborative control method for wind farm wake effect based on big data Technical Field

[0001] This invention relates to the field of wind power control technology, and in particular to a collaborative control method for wind farm wake effect based on big data. Background Technology

[0002] With the rapid development of renewable energy, wind power, as a crucial component, is playing an increasingly important role in the global energy structure. Large wind farms typically consist of multiple wind turbine units, leading to complex aerodynamic wake effects between the turbines. The operation of the leading turbines causes wind speed disturbances and enhanced turbulence, which in turn affects the intake conditions of downstream turbines, resulting in decreased wind energy utilization efficiency and power output fluctuations. In severe cases, this can even lead to turbine overload or shutdown. Therefore, effectively modeling the wake effects within a wind farm and using this model to achieve coordinated control of power, pitch angle, and other parameters among multiple turbines is one of the key technical challenges in the current wind power control field.

[0003] Existing wind farm wake control methods mainly fall into two categories: The first is centralized control methods based on classical physical models, which rely on pre-defined wake mathematical models (such as the Jensen model and the Frandsen model) to suppress wake effects by optimizing turbine layout, pitch angle, or yaw angle. However, these methods are sensitive to wind farm layout and wind condition changes, and the model accuracy is difficult to adapt to complex real-world environments. The second category is multi-agent cooperative methods based on distributed control strategies. In recent years, with the development of deep reinforcement learning technology, some studies have attempted to apply reinforcement learning to wind turbine control. However, the following limitations still exist in the perception of disturbances between wind turbines: (1) The lack of explicit modeling of the wake propagation path between wind turbines weakens the control coupling relationship between upstream and downstream wind turbines, making it difficult to accurately reflect the propagation direction and intensity of wake disturbances; (2) Control strategies are usually optimized based on the local state of a single unit, failing to fully utilize the spatial structure information and historical disturbance information between wind turbines, resulting in limited collaborative control effects; (3) Existing reinforcement learning models are difficult to dynamically adjust the influence weights between wind turbines, lacking a mechanism to guide wind turbines to adjust their control responses according to the current wake relationship, leading to insufficient strategy generalization ability. Therefore, it is urgent to provide a collaborative control method for wind farm wake effects based on big data to address the above shortcomings. Summary of the Invention

[0004] This invention provides a big data-based collaborative control method for wind farm wake effects, which addresses the technical problems of traditional wind farm wake effect collaborative control methods, such as lack of explicit modeling of wake propagation paths between wind turbines, insufficient utilization of spatial information and historical disturbance information between wind turbines, and failure to guide wind turbines to adjust control responses according to the current wake relationship.

[0005] This invention discloses a collaborative control method for wind farm wake effect based on big data, comprising the following steps: S1. Collecting multi-source data from wind turbines and preprocessing it to obtain preprocessed data; normalizing the preprocessed data and constructing wind turbine feature vectors based on the normalized data; constructing a directed graph of the wind farm and calculating edge weights based on the preprocessed data; based on the edge weights, introducing an edge weight perception mechanism and a historical state convolution fusion mechanism to construct an improved graph neural network model, obtaining the hidden state vectors of wind turbine nodes, and calculating the wake influence intensity between wind turbine nodes; S2. Introducing a multi-agent reinforcement learning collaborative control mechanism based on wake perception, generating a wake coupling control factor based on the wake influence intensity between wind turbine nodes; adjusting the original control actions based on the wake coupling control factor to obtain the final control actions; mapping the final control actions to actual control commands to achieve collaborative control based on the wake effect.

[0006] Preferably, step S1 specifically includes: the multi-source data of the wind turbine includes inflow wind speed, wind direction, active power and blade pitch angle; at the same time, the original coordinates of the wind turbine are obtained, encoded using a unified projection coordinate system, and missing value filling and outlier removal are performed to obtain the final coordinates of the wind turbine.

[0007] Preferably, S1 specifically includes: constructing a directed graph of the wind farm by taking the wind turbine as a node and the path between the wind turbines that has wake influence as an edge; calculating the wake propagation factor based on the preprocessed data and the final coordinates of the wind turbine, taking into account three factors: spatial distance attenuation, wind direction consistency and actual wind speed deviation; and using the wake propagation factor as the edge weight.

[0008] Preferably, S1 specifically includes: the improved graph neural network model, which integrates graph structure propagation with time series convolution, uses the historical feature vector sequence of the wind turbine node itself as the input of a one-dimensional time convolutional network, and combines edge weights to obtain the hidden state vector of the wind turbine node.

[0009] Preferably, S1 specifically includes: performing a dot product operation on the hidden state vectors of the two wind turbine nodes to quantify the similarity between the hidden state vectors of the wind turbine nodes and obtain the wake influence intensity between the wind turbine nodes.

[0010] Preferably, S2 specifically includes: in the process of implementing the multi-agent reinforcement learning collaborative control mechanism based on wake perception, the wind turbine is regarded as an agent, and the wake coupling control factor is generated by comprehensively considering the dual factors of wake influence intensity between wind turbine nodes and spatial distance attenuation.

[0011] Preferably, S2 specifically includes: in the process of implementing the multi-agent reinforcement learning collaborative control mechanism based on wake perception, introducing the historical final control action of the neighboring wind turbine, combining the wake coupling control factor to generate a wake perception gating factor, and adjusting the original control action to obtain the final control action; and performing inverse normalization processing on the final control action to map it into an actual control command.

[0012] Preferably, S2 specifically includes: the original control action is obtained by forward reasoning of the joint observation vector of the wind turbine through a reinforcement learning policy network; the joint observation vector of the wind turbine includes the feature vector of the wind turbine at the current moment, the historical feature vector sequence, and the final control action of the neighboring wind turbine at the previous moment.

[0013] The beneficial effects of the technical solution of the present invention are as follows: 1. The present invention proposes an improved graph neural network model that introduces an edge weight perception mechanism and a historical state convolution fusion mechanism, which integrates graph structure propagation and time series convolution, so that the state representation of wind turbine nodes in the graph neural network has both spatial wake dependency perception capability and time state fusion capability, which significantly improves the accuracy of wake impact modeling and the consistency of upstream and downstream disturbance response.

[0014] 2. This invention proposes a multi-agent reinforcement learning collaborative control mechanism based on wake perception. By considering both the wake influence intensity and spatial distance attenuation between wind turbine nodes, a wake coupling control factor is introduced. This allows the wind turbines to assign higher weights only to the actions of wind turbines with significant wake influence and close spatial distance when generating control actions. This effectively suppresses the ineffective interference of distant or weak wake-related wind turbines on control actions, thereby improving the pertinence and stability of multi-unit collaborative control.

[0015] 3. In the process of generating the final control action of the wind turbine, the present invention constructs a wake perception gating factor and explicitly introduces a wake coupling control factor so that the control action of each wind turbine not only depends on its own operating state, but also comprehensively considers the historical control behavior of other wind turbines that have a wake influence relationship with it, thereby realizing coordinated control based on the wake effect. Attached Figure Description

[0016] Figure 1 is a flowchart of a wind farm wake effect collaborative control method based on big data according to the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a collaborative control method for wind farm wake effects based on big data, provided by this invention.

[0020] Referring to Figure 1, a flowchart of a collaborative control method for wind farm wake effect based on big data according to an embodiment of the present invention is shown. The method includes the following steps: S1. Collecting multi-source data from wind turbines and preprocessing it to obtain preprocessed data; normalizing the preprocessed data and constructing wind turbine feature vectors based on the normalized data; constructing a directed graph of the wind farm and calculating edge weights based on the preprocessed data; based on the edge weights, introducing an edge weight perception mechanism and a historical state convolution fusion mechanism to construct an improved graph neural network model, obtaining the hidden state vectors of wind turbine nodes, and calculating the wake influence intensity between wind turbine nodes; collecting data through wind turbine sensor devices. Time Fan Multi-source data, including inflow wind speed (Unit: m / s) Wind direction (Unit: °) Active power (Unit: kW) and pitch angle (Unit: °), and simultaneously obtain wind turbine information through the wind farm's GPS positioning system. The original coordinates were encoded using a unified projected coordinate system, and missing value imputation and outlier removal were performed to obtain the wind turbine. final coordinates The multi-source data and final coordinates of the wind turbine are stored in the wind turbine database. Further, the multi-source data is preprocessed, and missing data is filled in using the nearest neighbor interpolation method to obtain data with missing values. The data with missing values ​​is then processed using... Outliers are removed according to principle, resulting in outlier-removed data. This outlier-removed data is then time-aligned using linear interpolation to obtain pre-processed data, including the pre-processed inflow velocity. (Unit: m / s) Wind direction after pretreatment (Unit: °) Active power after preprocessing (Unit: kW) and pre-treated pitch angle (Unit: °); The preprocessed data is scaled using the min-max normalization method to reduce the numerical values ​​to °. The interval was used to obtain normalized data, including the normalized inflow velocity. Normalized wind direction Normalized active power and normalized pitch angle Based on the normalized data, construct Time Fan eigenvectors The data is then stored in the wind turbine database. The techniques used in the preprocessing and normalization processes described above are well-known to those skilled in the art and will not be elaborated upon here. Furthermore, a directed graph of the wind farm is constructed. Treating wind turbines as a directed graph of a wind farm The nodes in the graph represent the paths between wind turbines that have wake effects, forming a directed graph of the wind farm. The edge weights are calculated by comprehensively considering three factors: spatial distance attenuation, wind direction consistency, and actual wind speed deviation. Specifically, an exponential function is used to establish the attenuation relationship of the wake with distance, and this is multiplied by a max function that determines whether the wake's influence propagates downstream to the wind turbine, as well as the relative quantification of the wind speed loss of the downstream wind turbine, to obtain the wake propagation factor. The wake propagation factor is used as the edge weight, as shown in the following formula: ,in, express Time Fan For the fan The wake propagation factor; Indicates wind turbine For the fan The wake effect decreases exponentially with distance; the greater the distance, the smaller the wake effect. , where is the distance attenuation coefficient, in meters. -1 The method is used to control the degree of wake attenuation with distance. It is obtained by fitting the distance between wind turbines with historical wind speed loss using the nonlinear least squares method. The least squares method is a well-known technique in the art and will not be described in detail here. Indicates wind turbine With wind turbine Distance between, in meters; Indicates wind turbine The final coordinates, Indicates wind turbine The final coordinates; Indicates the pre-treated wind direction and fan. The angle between the lines; It is obtained by calculating the dot product between the preprocessed wind direction vector and the wind turbine direction vector; Used to determine whether the pre-treated wind direction is favorable for the wake from the fan. wind turbine If spread, This indicates that the wind is roughly following the direction of the fan. wind turbine The wind blows in the direction of [the blast], and there is a wake effect. At this time... Values Conversely, it indicates that the wind is coming from the fan. wind turbine The direction of the blower Unaffected by wind turbine Wake effect, at this time The value is 0; , representing the wind speed influence adjustment factor, is obtained by fitting historical wind speed loss and preprocessed active power using the linear least squares method; express Time Fan Pre-treated inflow velocity; This is obtained by inputting the pre-processed inflow velocity into the existing Jensen wake model. Time Fan Predicted inflow wind speed; The smaller the value, the better the fan. The greater the wake effect; For wind turbine The relative quantification of inflow wind speed loss, i.e., the relative quantification of downstream wind turbine wind speed loss, is achieved by considering the greater the wake influence, the larger the wake propagation factor value, and the larger the edge weight. Furthermore, based on the directed graph of the wind farm... Based on the traditional graph neural network model, an edge weight awareness mechanism and a historical state convolution fusion mechanism are introduced to construct an improved graph neural network model. The historical feature vector sequence of the wind turbine node itself is used as the input of the one-dimensional temporal convolutional network in the improved graph neural network model to obtain the hidden state vector of the wind turbine node. The update formula of the hidden state vector of the wind turbine node is as follows: ,in, For wind turbine nodes exist The hidden state vector at time step; It is a ReLU nonlinear activation function; Indicates the wind turbine node The set of neighboring wind turbine nodes that have wake effects; express Wind turbine nodes For wind turbine nodes The wake propagation factor; This is the graph adjacency propagation weight matrix, used to divide neighboring wind turbine nodes. eigenvectors Mapped to the hidden state space, participating in the wind turbine nodes The hidden state update is obtained through backpropagation algorithm, which is a well-known technique in the art and will not be described in detail here. Indicates wind turbine node exist The feature vector at time step; The weight matrix for the temporal convolution path is used to... A linear mapping to the hidden state space is obtained through backpropagation. Indicates wind turbine node exist Time before A sequence of historical feature vectors at each time step; Indicates to Time before Time step fan node The historical feature vector sequence is subjected to one-dimensional temporal convolution to extract historical change trends. The one-dimensional temporal convolutional network has 16 convolutional kernels, a kernel size of 3, and a stride of 1. A same-padded strategy is used to maintain the sequence length. After the convolution operation, a ReLU nonlinear activation function is introduced to enhance expressive power. Subsequently, max pooling (window size 2) is added to extract significant temporal features. Batch normalization and Dropout regularization mechanisms are introduced to improve training stability. The one-dimensional temporal convolutional network is part of a graph neural network model, and its parameters are obtained through joint training with the graph neural network model using backpropagation. The loss function is either mean squared error (MSE) or mean absolute error (MAE). The training-validation-test set split ratio is 70%:15%:15%, and the initial learning rate is... Furthermore, a cosine annealing strategy is employed for dynamic adjustment. One-dimensional temporal convolutional networks are a well-known technique in the field and will not be elaborated upon here.

[0021] The above formula proposes an improved graph neural network model that incorporates an edge weight perception mechanism and a historical state convolution fusion mechanism. By fusing graph structure propagation with time series convolution, the hidden state vectors of wind turbine nodes in the graph neural network simultaneously possess spatial wake dependency perception capability and temporal state fusion capability, significantly improving the accuracy of wake impact modeling and the consistency of upstream and downstream disturbance response.

[0022] Furthermore, calculate the wind turbine nodes. For wind turbine nodes Wake effect intensity Intensity of wake effect between wind turbine nodes The larger the value, the stronger the wake effect; The similarity between the hidden state vectors of wind turbine nodes is quantified by dot product operation, thereby reflecting the coupling strength between the two wind turbines in the wake influence propagation path; For wind turbine nodes exist Hidden state vector at time step Transpose of; For wind turbine nodes exist The hidden state vector at time step; through The activation function maps the intensity of the wake effect between wind turbine nodes to ; This indicates transpose.

[0023] S2. Introduce a multi-agent reinforcement learning collaborative control mechanism based on wake perception. Generate a wake coupling control factor based on the wake influence intensity between wind turbine nodes. Adjust the original control action based on the wake coupling control factor to obtain the final control action. Map the final control action to the actual control command to realize collaborative control based on the wake effect.

[0024] After obtaining the wake influence intensity between wind turbine nodes, a multi-agent reinforcement learning cooperative control mechanism based on wake perception is proposed to achieve cooperative control of the wake effect between wind turbines in a wind farm. Using multi-agent reinforcement learning as the basic architecture, a reinforcement learning policy network is constructed, treating each wind turbine as an agent. By considering both the wake influence intensity between wind turbine nodes and the spatial distance attenuation, a wake coupling control factor is introduced. The formula is as follows: ,in, , is the wake coupling control factor, used to describe the current Time Fan Control actions on the fan The intensity of the wake coupling disturbance generated by the control action; , representing the wake adjustment factor, and representing the degree to which historical control effects are retained at the current moment, are obtained by fitting historical wind turbine control actions with the wake effect intensity using the least squares method; Indicates wind turbine For the fan The intensity of the wake effect; Used for quantizing fans With wind turbine The spatial distance attenuation is greater the closer the two wind turbines are; Indicates wind turbine With wind turbine The distance between; Indicates wind turbine The final coordinates, Indicates wind turbine The final coordinates; The maximum distance between any two wind turbines in the wind farm is used for normalization. The above formula introduces a wake coupling control factor that considers both the wake influence intensity between wind turbine nodes and the spatial distance attenuation. This ensures that when generating control actions, wind turbines only give higher weight to actions of wind turbines with significant wake influence and close spatial distance. This effectively suppresses the ineffective interference of long-distance or weak wake-related wind turbines on control actions and improves the pertinence and stability of multi-unit collaborative control.

[0025] To further enhance the adaptability of the control actions, the historical final control actions of neighboring wind turbines are weighted and fused using a wake coupling control factor, and then normalized to generate a wake-aware gating factor. This wake-aware gating factor is then multiplied by the original control action output by the reinforcement learning policy network to adjust the original control action, thus obtaining the final control action for the wind turbine. exist The final control action at any given moment is given by the following formula: ,in, Indicates wind turbine exist The final control actions at any given moment, such as pitch angle commands and active power settings, are stored in the wind turbine database. The wake perception gating factor serves as an adjustment gating for the final control action, used to regulate the execution amplitude of the original control action output by the reinforcement learning policy network. Indicates the wind turbine node The set of neighboring wind turbine nodes that have wake effects; This is the wake coupling control factor; , indicating wind turbine exist The final control action at any given moment is obtained through the wind turbine database; , recorded as , The original control action is represented by a reinforcement learning policy network. Input fan exist Joint observation vector at time 1 Obtained through forward reasoning; For parameters The reinforcement learning policy network consists of 2 to 3 fully connected layers, each containing 64 to 128 neurons, and uses ReLU as the activation function. The policy network parameters, including the weight matrix and bias vector, are obtained by adjusting the wind turbine. exist The joint observation vector at time step, the initial control action taken, and the reward based on the feedback from the control action. The joint observation vectors at each time point are obtained through reinforcement learning training, and the rewards based on the feedback from the control actions include improvements in active power, etc. Indicates wind turbine exist The joint observation vector at each moment, including the wind turbine exist Moment Feature vector ,forward Historical feature vector sequence at each time step Japanese-style fan exist The final control action at any moment The above formula constructs a wake perception gating factor during the generation of the final control action of the wind turbine. By explicitly introducing a wake coupling control factor, the control action of each wind turbine not only depends on its own operating state, but also takes into account the historical control behavior of other wind turbines that have a wake influence relationship with it, thereby realizing coordinated control based on the wake effect.

[0026] Finally, the final control action will be... After inverse normalization, the values ​​are mapped to actual control commands, such as the pitch angle command and the active power setpoint, which are given by the following formulas: Active power setting ,in Indicates wind turbine exist Actual pitch angle command at any given time (unit: °). , They represent the fans respectively. The minimum and maximum values ​​of the pitch angle are obtained from the wind turbine database; Indicates wind turbine exist Active power setpoint at any given time (unit: kW). , They represent the fans respectively. The minimum and maximum active power values ​​are set through the wind turbine database.

[0027] In summary, a collaborative control method for wind farm wake effect based on big data has been developed.

[0028] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0029] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0030] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A collaborative control method for wind farm wake effect based on big data, characterized in that, Includes the following steps: S1. Collect multi-source data from the wind turbine and perform preprocessing to obtain preprocessed data; The preprocessed data is normalized, and a wind turbine feature vector is constructed based on the normalized data. Construct a directed graph of the wind farm and calculate edge weights based on the preprocessed data; Based on edge weights, an improved graph neural network model is constructed by introducing an edge weight perception mechanism and a historical state convolution fusion mechanism to obtain the hidden state vectors of wind turbine nodes and calculate the wake influence intensity between wind turbine nodes. S2. Introduce a multi-agent reinforcement learning collaborative control mechanism based on wake perception. Generate a wake coupling control factor based on the wake influence intensity between wind turbine nodes. Adjust the original control action based on the wake coupling control factor to obtain the final control action. Map the final control action to the actual control command to realize collaborative control based on the wake effect.

2. The method for coordinated control of wind farm wake effect based on big data as described in claim 1, characterized in that, S1 specifically includes: the multi-source data of the wind turbine includes inflow wind speed, wind direction, active power and blade pitch angle; at the same time, the original coordinates of the wind turbine are obtained, encoded using a unified projection coordinate system, and missing value filling and outlier removal are performed to obtain the final coordinates of the wind turbine.

3. The method for coordinated control of wind farm wake effect based on big data according to claim 2, characterized in that, S1 specifically includes: constructing a directed graph of the wind farm by taking wind turbines as nodes and paths between wind turbines that have wake effects as edges; calculating the wake propagation factor based on preprocessed data and the final coordinates of the wind turbines, taking into account three factors: spatial distance attenuation, wind direction consistency and actual wind speed deviation; and using the wake propagation factor as edge weights.

4. The method for coordinated control of wind farm wake effect based on big data as described in claim 1, characterized in that, S1 specifically includes: the improved graph neural network model, which integrates graph structure propagation with time series convolution, uses the historical feature vector sequence of the wind turbine node itself as the input of a one-dimensional time convolutional network, and combines edge weights to obtain the hidden state vector of the wind turbine node.

5. The method for coordinated control of wind farm wake effect based on big data according to claim 4, characterized in that, S1 specifically includes: performing a dot product operation on the hidden state vectors of the two wind turbine nodes to quantify the similarity between the hidden state vectors of the wind turbine nodes and obtain the wake influence intensity between the wind turbine nodes.

6. The method for coordinated control of wind farm wake effect based on big data according to claim 1, characterized in that, S2 specifically includes: in the process of implementing the multi-agent reinforcement learning collaborative control mechanism based on wake perception, the wind turbine is regarded as an agent, and the wake coupling control factor is generated by comprehensively considering the dual factors of wake influence intensity between wind turbine nodes and spatial distance attenuation.

7. The method for coordinated control of wind farm wake effect based on big data according to claim 6, characterized in that, S2 specifically includes: in the process of implementing the multi-agent reinforcement learning collaborative control mechanism based on wake perception, the historical final control action of the neighboring wind turbine is introduced, and a wake perception gating factor is generated by combining the wake coupling control factor, and the original control action is adjusted to obtain the final control action; the final control action is denormalized and mapped to the actual control command.

8. The method for coordinated control of wind farm wake effect based on big data according to claim 7, characterized in that, S2 specifically includes: the original control action is obtained by forward reasoning of the joint observation vector of the wind turbine through a reinforcement learning policy network; the joint observation vector of the wind turbine includes the feature vector of the wind turbine at the current moment, the historical feature vector sequence, and the final control action of the neighboring wind turbine at the previous moment.

Images