An intelligent impedance acquisition method for a wind farm combining simulation data and measured data

Abstract

This invention relates to the field of wind power technology, and more particularly to an intelligent method for acquiring wind farm impedance by combining simulation and measured data. The steps include: constructing a simulation sample set using a wind turbine with parameters similar to those of the actual wind turbines; constructing a labeled measured sample set and an unlabeled sample set using the actual wind turbines; building an artificial neural network to form a pre-trained model; fine-tuning the pre-trained model using a hybrid loss function to form a wind turbine impedance identification model; acquiring the phase angle at each wind turbine using a SCADA system; inputting the steady-state operating data and scanning frequency of each wind turbine into the wind turbine impedance identification model to obtain the dq impedance of each wind turbine in its own coordinate system; performing coordinate transformation using the phase angle at the wind turbine to convert the dq impedance to the admittance in the global xy coordinate system; and calculating the overall wind farm impedance based on the extended connection matrix and combined with the wind turbine admittance. This invention combines simulation and measured data, effectively improving the accuracy of the obtained impedance.

Description

Technical Field

[0001] This invention relates to the field of wind power technology, and in particular to a method for intelligently acquiring wind farm impedance by combining simulation and measured data. Background Technology

[0002] In recent years, with the continuous expansion of wind farm scale, the centralized layout of wind farms has reduced the system's short-circuit ratio. This may induce a series of small-disturbance stability issues. If the impedance of the wind farm can be obtained online and in a timely manner, it is possible to further assess small-disturbance stability, providing crucial assistance in maintaining the safe operation of the system.

[0003] Due to the diversity of wind turbine operating points and the commercial confidentiality of turbine structural parameters, wind farms exhibit "black box" characteristics, making it difficult to establish analytical impedance models. Therefore, analytical methods are often used to analyze oscillation mechanisms rather than for online applications. To address the "black box" problem, parameter identification impedance methods are widely adopted. These methods identify wind turbine parameters by measuring input and output quantities, thereby establishing a wind farm impedance model for online evaluation. However, most of these parameter identification methods require determining the wind turbine structure to set the transfer function order and improve identification accuracy. Therefore, they essentially solve a "grey box" problem rather than a "black box" problem.

[0004] In recent years, data-driven methods such as machine learning have been applied to wind farm impedance identification and online stability assessment. Existing methods utilize large amounts of offline data to train neural networks for online identification of wind turbine impedance. These methods only require impedance measurement data, making them suitable for solving "black box" problems. However, on the one hand, these methods mainly focus on single-unit impedance identification and do not study the impedance identification of the entire wind farm. On the other hand, these methods require a large amount of impedance data from dozens or even hundreds of wind turbine operating conditions to train the impedance identification model. This is relatively easy in simulation software and even in laboratories, but it is difficult to achieve in wind farms due to measurement costs and time constraints. Furthermore, due to aging, capacity, parameter settings, and other factors, there are differences between actual wind turbines in wind farms and simulation models, making it unsuitable to directly apply impedance measurement samples from simulation models to actual wind farms. Summary of the Invention

[0005] This invention provides a method for intelligently acquiring wind farm impedance by combining simulation and measured data, which can effectively solve the problem of poor accuracy in impedance measurement relying solely on simulation models in the prior art.

[0006] This invention provides a method for intelligently acquiring wind farm impedance by combining simulation and measured data, the steps of which include: S1: Using a simulated wind turbine with parameters similar to those of the actual wind turbine, inject harmonic signals to measure impedance and form a simulation sample set; S2: Measure the impedance of the wind turbine by injecting harmonic signals on site, and construct a labeled measured sample set and an unlabeled sample set; S3: Construct an artificial neural network and pre-train it using a simulation sample set to form a pre-trained model; S4: Based on the conditional distribution adaptive method, the pre-trained model is fine-tuned by combining labeled and unlabeled sample sets through a hybrid loss function to form a wind turbine impedance identification model. S5: Use the SCADA system to collect steady-state data such as active power, reactive power, and voltage of each wind turbine, and combine the constraint equations to perform state estimation and obtain the phase angle at each wind turbine. S6: Input the steady-state operating data and scanning frequency of each wind turbine into the wind turbine impedance identification model to obtain the dq impedance of each wind turbine in its own coordinate system; S7: Use the phase angle at the wind turbine to perform coordinate transformation, and convert the dq impedance to the admittance in the global coordinate xy system; S8: Construct the wind farm network admittance matrix based on the extended connection matrix, and calculate the overall impedance of the wind farm by combining the wind turbine admittance in the global coordinate system.

[0007] Furthermore, in step S1, generating a large number of simulation samples to form a simulation sample set specifically involves: Multiple operating points are set for the fan used in simulation measurement. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U, and multiple frequencies s are set under each operating point. The multiple operating points need to cover the common operating area of ​​the fan. At each operating point, inject harmonic currents of various frequencies s into the fan, calculate the harmonic voltages at each frequency s, and calculate the corresponding dq impedances. Record all operating points and frequency points and their corresponding dq impedances to form a simulation sample set D. s .

[0008] Furthermore, in step S2, the construction of the labeled test sample set and the construction of the unlabeled sample set are specifically as follows: Multiple operating points are set for the on-site wind turbines. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U. Multiple frequencies s are set under each operating point. At each operating point, a harmonic current of each frequency s is injected into the fan, the harmonic voltage at each frequency s is measured, and the corresponding dq impedance is calculated. Record all operating points and frequency points and their corresponding dq impedances to form a measured sample set D. m ; From the measured sample set D s A set number of samples are randomly selected as the validation set, and the remaining samples are divided into a set of labeled test samples D according to a set ratio.L With unlabeled sample set D U Label sample set D L The number of samples is less than the number of unlabeled sample sets D. U The number of samples.

[0009] Furthermore, in step S3, forming the pre-trained model specifically involves: Construct an artificial neural network with an input layer, an output layer, and several hidden layers; The input layer is set with four dimensions: active power P, reactive power Q, voltage amplitude U, and frequency s. The output layer is set with 8 dimensions, which are the real and imaginary parts of the four complex elements of the dq impedance matrix; The hidden layers have 4 layers, each with 128 neurons, and the ReLU activation function is used. The simulation sample set D s The system is divided into a test set and a training set. The artificial neural network is trained using the training set. After training, the model with the smaller loss on the test set is saved as the final pre-trained model.

[0010] Furthermore, in step S4, the formation of the wind turbine impedance identification model specifically involves: The weights and biases of each neuron in the pre-trained model are extracted and assigned to a new model. The structure of the new model is completely consistent with that of the pre-trained model. Using labeled sample set D L With unlabeled sample set D U Through the hybrid loss function L H The new model is trained, and after a set number of training rounds, the wind turbine impedance identification model is finally obtained. Hybrid loss function L H Specifically: L H =β1·MSE+β2·L c ; In the formula, β1 and β2 are weighting coefficients; MSE is the mean squared error between the predicted values ​​of all labeled samples obtained by the new model and the measured values ​​of the labeled samples. L c The conditional distribution difference between labeled and unlabeled sample sets is calculated using regenerated kernel Hilbert space embedding.

[0011] Furthermore, in step S7, the conversion of the dq impedance to the admittance in the global coordinate system is specifically as follows: For the i-th wind turbine in the wind farm, the steady-state operating data and scanning frequency of the turbine are input into the turbine impedance identification model to obtain the dq impedance Y of the turbine in its own coordinate system. WTidq for: ; The phase angle at the i-th fan is θ. WTi ; Transform the dq impedance to the admittance Y in the global xy coordinate system. WTixy for: .

[0012] Furthermore, in step S8, the overall impedance of the wind farm is calculated as follows: Set the connection matrix T ac In the matrix, the elements in the first row and first column, the second row and second column, the third row and third column, ... are all -I, and the elements in the second row and first column, the third row and second column, the fourth row and third column, ... are all I, where I is a 2×2 identity matrix; all other elements in the matrix are 2×2 zero matrices. Calculate the network matrix Y of the wind farm area net And the wind power field impedance matrix Zc: ; In the formula, y B =[y B1 ,y B2 ,......,y Bb ] T For the admittance of each line; y N =[y N1 ,y N2 ,......,y Nb ] T y represents the admittance at each node; Gxy =[y G1xy ,y G2xy ,...,y Gmxy ] T For wind turbine admittance; y WTixy To guide Y WTixy All elements in the matrix; each element in each matrix is ​​a 2×2 matrix.

[0013] Furthermore, in step S8, when the wind turbine stops operating, the y corresponding to the stopped wind turbine will be... WTixy Replace with a 2×2 matrix of zeros.

[0014] Furthermore, in step S8, before constructing the wind farm network admittance matrix based on the extended connection matrix, the connection matrix T is automatically generated according to the actual electrical wiring diagram or topology configuration file of the wind farm. ac To adapt to the physical connection relationships of the collection lines in different wind farms.

[0015] Furthermore, between steps S6 and S7, the rationality of the dq impedance result output by the wind turbine impedance identification model is verified, checking whether the dq impedance meets the set constraints, and marking or correcting abnormal results.

[0016] The technical solution of this invention can achieve the following technical effects: 1. By adopting a conditional distribution adaptive method, a large amount of simulation data and a small amount of measured data are fully integrated, which solves the problem of the dependence of data-driven methods on measured samples. Even when the proportion of labeled samples in the measured data is low, it can still maintain high identification accuracy of wind turbine impedance, thereby reducing measurement costs and time costs.

[0017] 2. The phase angle of the wind turbine is obtained by combining SCADA data with state estimation. No additional detection equipment is required. Without significantly reducing the accuracy of impedance acquisition, the installation and maintenance costs of the equipment are greatly reduced, and it is compatible with existing wind farm monitoring systems.

[0018] 3. By extending the connection matrix to adapt to changes in wind farm topology, the method can cope with scenarios such as wind turbine shutdown and line switching, thus improving its engineering applicability. At the same time, it uses steady-state data to achieve online evaluation, avoiding the lag of traditional methods and providing early warning of oscillation risks. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The flowchart outlines a method for intelligently acquiring wind farm impedance by combining simulation and measured data. Figure 2 A schematic diagram of the logic principle of a method for intelligently acquiring wind farm impedance by combining simulation and measured data; Figure 3 A schematic diagram of the logic principle for training an artificial neural network. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] 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. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] This invention relates to a method for intelligently acquiring wind farm impedance by combining simulation and measured data, such as... Figures 1-2 As shown, this method aims to address the dependence of data-driven model training on massive amounts of measured samples and to achieve online and accurate construction of wind farm hierarchical impedance. The entire method consists of two main stages: an offline model training stage and an online impedance construction stage.

[0024] Offline model training phase: This step includes four steps, S1 to S4. It involves offline sample collection and training of an actual wind turbine impedance identification model. The goal is to obtain a data-driven impedance model of the actual wind turbine. The core idea is to use a large amount of simulation data from a wind turbine simulation model to train a pre-trained impedance model. Then, using a small amount of measured impedance data from actual wind turbines, the pre-trained impedance model is fine-tuned to obtain the data-driven impedance model of the actual wind turbine.

[0025] The specific steps are as follows: S1: Using a simulated wind turbine with parameters similar to those of the actual wind turbine, a harmonic signal is injected to measure the impedance, forming a simulation sample set to provide simulation data for the subsequent construction of the identification model; S2: Measure the impedance of the wind turbine by injecting harmonic signals on site, and construct a labeled and an unlabeled sample set to provide measured data for the subsequent construction of the identification model; S3: Construct an artificial neural network and pre-train it using a simulated sample set, such as... Figure 3 As shown, the artificial neural network initially learns the general laws and complex nonlinear mapping relationships of the wind turbine impedance as a function of operating point and frequency point from the simulation data, forming a pre-trained model with a good generalization foundation. S4: Based on the conditional distribution adaptive method, the pre-trained model is fine-tuned by combining labeled and unlabeled sample sets through a hybrid loss function to form a wind turbine impedance identification model. This step optimizes the model's fitting accuracy to labeled measured data while also making the predicted distribution of the pre-trained model for unlabeled measured data consistent with its predicted distribution for labeled data. Thus, the pre-trained model is optimized using both simulation and measured data, effectively improving the model's credibility.

[0026] Online Impedance Construction Stage: This step includes four steps, S5 to S8. The purpose of constructing the overall impedance of the wind farm online is to obtain the impedance of the actual wind farm. The core idea is to use the wind turbine impedance model and combine it with SCADA system data to obtain the impedance of each wind turbine in the wind farm. Then, based on the impedance of each wind turbine and combined with the topology, line impedance and other information of the wind farm, a digital-driven impedance model of the actual wind farm is obtained.

[0027] The specific steps are as follows: S5: Use the SCADA system to collect steady-state data such as active power, reactive power, and voltage of each wind turbine, and combine them with constraint equations to perform state estimation and obtain the phase angle at each wind turbine; the specific measurement method for this step is existing technology and will not be described in detail here. S6: Input the steady-state operating data and scanning frequency of each wind turbine into the wind turbine impedance identification model, so that the dq impedance of each wind turbine in its own coordinate system can be quickly calculated using the model. S7: Use the phase angle at the wind turbine to perform coordinate transformation and convert the dq impedance to the admittance in the global coordinate xy system. This step is mainly to eliminate the coordinate system differences caused by different rotor positions or control references of each wind turbine, and to create conditions for the parallel superposition of the impedances of multiple turbines. S8: Construct the wind farm network admittance matrix based on the extended connection matrix, and calculate the overall impedance of the wind farm by combining the wind turbine admittance in the global coordinate system.

[0028] Preferably, in step S1, a large number of simulation samples are generated to form a simulation sample set, specifically as follows: Multiple operating points are set for the wind turbine used in simulation measurements. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U, meaning that at least one of these parameters must be different between any two operating points. The number of operating points should be as large as possible (e.g., 510) to ensure that the parameters cover the common operating range of the wind turbine. Multiple frequencies s are set for each operating point, for example, 100 frequencies from 1Hz to 100Hz.

[0029] At each operating point, harmonic currents of various frequencies (s) are injected into the wind turbine. The harmonic voltages at each frequency (s) are calculated, along with the corresponding dq impedances. The dq impedance refers to the impedance value obtained when performing impedance analysis on a load or equipment in a three-phase power system in a synchronous rotating coordinate system (dq coordinate system). It is typically represented as a 2x2 matrix containing four components: dd-axis impedance, dq-axis impedance, qd-axis impedance, and qq-axis impedance. It is important to note that the dq impedance calculation here is performed using electrical principles and simulation software. All operating points and frequency points, along with their corresponding dq impedances, can be automatically batch-generated using scripts in the simulation software, efficiently generating tens or even hundreds of thousands of samples.

[0030] Record all operating points and frequency points and their corresponding dq impedances to form a simulation sample set D. s Taking the above data as an example, with 510 operating points and 100 frequency values ​​for each operating point, the simulation sample set D... s A total of 51,000 dq impedance samples will be generated.

[0031] Preferably, in step S2, the construction of the labeled measured sample set and the construction of the unlabeled sample set are specifically as follows: Multiple operating points are set for the on-site wind turbines. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U, and multiple frequencies s are set under each operating point. The setting principle here is the same as described above and will not be repeated here. However, since these are measured values, the number of on-site wind turbines and operating points will be significantly reduced compared to the number set in the simulation.

[0032] At each operating point, harmonic currents of various frequencies s are injected into the fan, the harmonic voltages at each frequency s are measured, and the corresponding dq impedances are calculated. It should be noted that the dq impedances are calculated using actual measured data.

[0033] Record all operating points and frequency points and their corresponding dq impedances to form a measured sample set D. m From the measured sample set D s A predetermined number of samples (e.g., 10%) are randomly selected as the validation set for final model performance evaluation and are not used in training. The remaining test samples are divided into two parts: one part serves as the labeled test sample set D. L One part contains complete input-output pairs, i.e., complete operating point and frequency points and their corresponding dq impedance information; the other part only retains the operating point and frequency points, discarding their impedance values ​​to form an unlabeled sample set D. U Tag sample set D L The number of samples is less than the number of unlabeled sample sets D. U The number of samples, the labeled sample set D L Typically, this accounts for 10% to 15% of the remaining portion, to simulate situations in reality where only a small amount of precise measurement data can be obtained.

[0034] In the preferred step S3, the formation of the pre-trained model specifically involves: Construct an artificial neural network with an input layer, an output layer, and several hidden layers; The input layer is set with four dimensions: active power P, reactive power Q, voltage amplitude U, and frequency s. The output layer is set with 8 dimensions, which are the real and imaginary parts of the four complex elements of the dq impedance matrix; The hidden layers have 4 layers, each with 128 neurons, and the ReLU activation function is used. The simulation sample set D s The dataset is split into a test set and a training set, with 15% of the samples designated as the test set and 85% as the training set. The artificial neural network is trained using the training set. The training iterations are set to 1000, and the learning rate is 1e. -4 After training, the model with the smaller loss is saved on the test set. At this point, the model has learned a fairly accurate mapping from "operating condition + frequency" to "impedance," and serves as the final pre-trained model. It should be noted that the neural network structure and the proportion of the test set mentioned in this step are not limited and can be modified as needed.

[0035] In the preferred step S4, the formation of the wind turbine impedance identification model is specifically as follows: The weights and biases of each neuron in the pre-trained model are extracted and assigned to a new model. The structure of the new model is completely consistent with that of the pre-trained model, and the new model serves as the starting point for fine-tuning.

[0036] Using labeled sample set D L With unlabeled sample set D U Through the hybrid loss function L H The new model was trained for 500 epochs with a learning rate of 1e. -4 After setting the number of training rounds, the final wind turbine impedance identification model is obtained.

[0037] Hybrid loss function L H Specifically: L H =β1·MSE+β2·L c ; In the formula, β1 and β2 are weighting coefficients.

[0038] MSE is the mean squared error between the predicted values ​​of all labeled samples obtained through the new model and the measured values ​​of the labeled samples; after expanding MSE, L H The expression will then change to: ; In the formula, there is a labeled sample set D. L This can be expressed as: ; Unlabeled sample set D U This can be expressed as: ; f WTN (x Li The information x represents the operating point and frequency information corresponding to the i-th sample in the labeled samples. Li (Including active power P, reactive power Q, voltage amplitude U, and frequency s) Predicted values ​​obtained through the new model; yLi That is, the measured value corresponding to the i-th sample in the labeled samples.

[0039] L c The conditional distribution difference between labeled and unlabeled sample sets is calculated using regenerating kernel Hilbert spatial embedding. c The expression is: ; In the formula, For the Gram matrix, calculate the elements using the Gaussian kernel trick: ; In the formula, α is the Gaussian kernel coefficient; n2 is the number of unlabeled samples; Tr refers to the trace of the matrix; ; ψ(Y L )=[ψ(y L1 ), ψ(y L2 ), ......, ψ(y Ln1 )]; φ(X L )=[φ(x L1 ), φ(x L2 ), ......, φ(x Ln1 )]; And φ(x) and ψ(y) are respectively in RKHS, and this function maps the elements therein from the corresponding sample space to the H of RKHS. φ and H ψ middle; n1 is the number of labeled samples; I is the identity matrix; η is the regularization coefficient to ensure that the inverse matrix is ​​well-posed.

[0040] Preferably, in step S7, the conversion of the dq impedance to the admittance in the global coordinate system is specifically as follows: For the i-th wind turbine in the wind farm, the steady-state operating data and scanning frequency of the turbine are input into the turbine impedance identification model to obtain the dq impedance Y of the turbine in its own coordinate system. WTidq for: ; The phase angle at the i-th fan is θ. WTi ; Transform the dq impedance to the admittance Y in the global xy coordinate system. WTixy for: .

[0041] Preferably, in step S8, the overall impedance of the wind farm is calculated as follows: Set the connection matrix T ac In the matrix, the elements at row 1, column 1, row 2, column 2, row 3, column 3, ... are all -I, and the elements at row 2, column 1, row 3, column 2, row 4, column 3, ... are all I, where I is a 2×2 identity matrix; all other elements in the matrix are 2×2 zero matrices; T ac The specific form of the matrix is ​​as follows: .

[0042] Calculate the network matrix Y of the wind farm area net And the wind power field impedance matrix Zc: ; In the formula, y B =[y B1 ,y B2 ,......,y Bb ] T For the admittance of each line; y N =[y N1 ,y N2 ,......,y Nb ] T y represents the admittance at each node; Gxy =[y G1xy ,y G2xy ,...,y Gmxy ] T For wind turbine admittance; y WTixy To guide Y WTixy All elements in the matrix; each element in each matrix is ​​a 2×2 matrix; the inv function is used to find the inverse of a matrix, and the diag function is used to form a diagonal matrix by sequentially selecting the corresponding elements within the parentheses.

[0043] Preferably, in step S8, when the wind turbine stops operating, the y corresponding to the stopped wind turbine is... WTixy Replace it with a 2×2 zero matrix to enable the connection matrix to quickly adapt to changes in the wind farm's operating mode and ensure the accuracy of the wind farm impedance matrix calculation.

[0044] Preferably, in step S8, before constructing the wind farm network admittance matrix based on the extended connection matrix, the connection matrix T is automatically generated according to the actual electrical wiring diagram or topology configuration file of the wind farm. ac To adapt to the physical connection relationships of the collection lines in different wind farms.

[0045] Preferably, between steps S6 and S7, the rationality of the dq impedance result output by the wind turbine impedance identification model is verified. For example, it is checked whether the impedance matrix meets physical constraints such as passivity or causality, and abnormal results are marked or corrected by interpolation.

[0046] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for intelligently acquiring wind farm impedance by combining simulation and measured data, characterized in that the steps include: include: S1: Using a simulated wind turbine with parameters similar to those of the actual wind turbine, inject harmonic signals to measure impedance and form a simulation sample set; S2: Measure the impedance of the wind turbine by injecting harmonic signals on site, and construct a labeled measured sample set and an unlabeled sample set; S3: Construct an artificial neural network and pre-train it using a simulation sample set to form a pre-trained model; S4: Based on the conditional distribution adaptive method, the pre-trained model is fine-tuned by combining labeled and unlabeled sample sets through a hybrid loss function to form a wind turbine impedance identification model. S5: Use the SCADA system to collect steady-state data such as active power, reactive power, and voltage of each wind turbine, and combine the constraint equations to perform state estimation and obtain the phase angle at each wind turbine. S6: Input the steady-state operating data and scanning frequency of each wind turbine into the wind turbine impedance identification model to obtain the dq impedance of each wind turbine in its own coordinate system; S7: Use the phase angle at the wind turbine to perform coordinate transformation, and convert the dq impedance to the admittance in the global coordinate xy system; S8: Construct the wind farm network admittance matrix based on the extended connection matrix, and calculate the overall impedance of the wind farm by combining the wind turbine admittance in the global coordinate system.

2. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 1, characterized in that, In step S1, a large number of simulation samples are generated to form a simulation sample set, specifically as follows: Multiple operating points are set for the fan used in simulation measurement. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U, and multiple frequencies s are set under each operating point. Multiple operating points need to cover the common working areas of the wind turbine; At each operating point, inject harmonic currents of various frequencies s into the fan, calculate the harmonic voltages at each frequency s, and calculate the corresponding dq impedances. Record all operating points and frequency points and their corresponding dq impedances to form a simulation sample set D. s .

3. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 1, characterized in that, In step S2, the construction of the labeled test sample set and the construction of the unlabeled sample set are specifically as follows: Multiple operating points are set for the on-site wind turbines. Each operating point is defined by active power P, reactive power Q, and voltage amplitude U. Multiple frequencies s are set under each operating point. At each operating point, a harmonic current of each frequency s is injected into the fan, the harmonic voltage at each frequency s is measured, and the corresponding dq impedance is calculated. Record all operating points and frequency points and their corresponding dq impedances to form a measured sample set D. m ; From the measured sample set D s A set number of samples are randomly selected as the validation set, and the remaining samples are divided into a set of labeled test samples D according to a set ratio. L With unlabeled sample set D U Label sample set D L The number of samples is less than the number of unlabeled sample sets D. U The number of samples.

4. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 2, characterized in that, In step S3, the pre-trained model is formed as follows: Construct an artificial neural network with an input layer, an output layer, and several hidden layers; The input layer is set with four dimensions: active power P, reactive power Q, voltage amplitude U, and frequency s. The output layer is set with 8 dimensions, which are the real and imaginary parts of the four complex elements of the dq impedance matrix; The hidden layers have 4 layers, each with 128 neurons, and the ReLU activation function is used. The simulation sample set D s The system is divided into a test set and a training set. The artificial neural network is trained using the training set. After training, the model with the smaller loss on the test set is saved as the final pre-trained model.

5. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 3, characterized in that, In step S4, the formation of the wind turbine impedance identification model is specifically as follows: The weights and biases of each neuron in the pre-trained model are extracted and assigned to a new model. The structure of the new model is completely consistent with that of the pre-trained model. Using labeled sample set D L With unlabeled sample set D U Through the hybrid loss function L H The new model is trained, and after a set number of training rounds, the wind turbine impedance identification model is finally obtained. Hybrid loss function L H Specifically: L H =β1·MSE+β2·L c ; In the formula, β1 and β2 are weighting coefficients; MSE is the mean squared error between the predicted values ​​of all labeled samples obtained by the new model and the measured values ​​of the labeled samples. L c The conditional distribution difference between labeled and unlabeled sample sets is calculated using regenerated kernel Hilbert space embedding.

6. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 1, characterized in that, In step S7, the specific steps for converting the dq impedance to the admittance in the global coordinate system are as follows: For the i-th wind turbine in the wind farm, the steady-state operating data and scanning frequency of the turbine are input into the turbine impedance identification model to obtain the dq impedance Y of the turbine in its own coordinate system. WTidq for: ； The phase angle at the i-th fan is θ. WTi ; Transform the dq impedance to the admittance Y in the global xy coordinate system. WTixy for: 。 7. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 6, characterized in that, In step S8, the overall impedance of the wind farm is calculated as follows: Set the connection matrix T ac In the matrix, the elements in the first row and first column, the second row and second column, the third row and third column, ... are all -I, and the elements in the second row and first column, the third row and second column, the fourth row and third column, ... are all I, where I is a 2×2 identity matrix; all other elements in the matrix are 2×2 zero matrices. Calculate the network matrix Y of the wind farm area net And the wind power field impedance matrix Zc: ； In the formula, y B =[y B1 ,y B2 ,......,y Bb ] T For the admittance of each line; y N =[y N1 ,y N2 ,......,y Nb ] T y represents the admittance at each node; Gxy =[y G1xy ,y G2xy ,...,y Gmxy ] T For wind turbine admittance; y WTixy To guide Y WTixy All elements in the matrix; each element in each matrix is ​​a 2×2 matrix.

8. The intelligent method for acquiring wind farm impedance combining simulation and measured data according to claim 7, characterized in that, In step S8, when the wind turbine stops operating, the y corresponding to the stopped wind turbine will be... WTixy Replace with a 2×2 matrix of zeros.

9. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 7, characterized in that, In step S8, before constructing the wind farm network admittance matrix based on the extended connection matrix, the connection matrix T is automatically generated according to the actual electrical wiring diagram or topology configuration file of the wind farm. ac To adapt to the physical connection relationships of the collection lines in different wind farms.

10. The intelligent method for acquiring wind farm impedance by combining simulation and measured data according to claim 1, characterized in that, Between steps S6 and S7, the rationality of the dq impedance result output by the wind turbine impedance identification model is verified, checking whether the dq impedance meets the set constraints, and marking or correcting abnormal results.

Images