New energy unit full black box impedance identification method and system considering phase angle correction

Abstract

The present application belongs to the field of stability analysis and control technology of new energy grid-connected system, and specifically discloses a new energy unit full black box impedance identification method and system considering phase angle correction. The method obtains the steady-state operating condition parameters of the new energy unit, extracts high-fidelity dq-axis electrical quantities by using positive and negative sequence component complex value superposition technology; after standardization preprocessing, the characteristic data is input into the pre-trained deep feedforward neural network for forward reasoning, and the real part and imaginary part of the positive and negative sequence impedance under the corresponding frequency are quickly output, and they are converted into impedance amplitude and phase angle; for the low amplitude frequency band, a linear interpolation algorithm is introduced for phase angle correction to eliminate the phase angle jump error caused by the inverse tangent function when the amplitude approaches zero, and finally the smooth and accurate wideband impedance characteristic curve is output. The present application gets rid of the dependence on the internal white box parameters of the new energy unit, and can still achieve high-precision impedance identification when the internal control topology and parameters of the new energy unit are unknown.

Description

Technical Field

[0001] This invention belongs to the field of stability analysis and control technology of new energy grid-connected systems, specifically involving a method and system for identifying the impedance of a new energy unit in a black box considering phase angle correction. Background Technology

[0002] With the rapid increase in the penetration rate of new energy sources such as wind power and photovoltaic power in the power system, the interaction between new energy units and the weak power grid is becoming increasingly complex. Wideband oscillation problems such as subsynchronous oscillations and supersynchronous oscillations have become a major challenge threatening the safe and stable operation of the power grid. Impedance analysis, with its clear physical meaning and applicability to black-box systems, has become one of the core methods for assessing and solving wideband oscillation problems. However, the engineering application of impedance analysis is highly dependent on obtaining an accurate wideband impedance model of the new energy units.

[0003] In existing technologies, methods for obtaining broadband impedance models of new energy power units are mainly divided into physical analysis modeling and active disturbance measurement methods. Physical analysis modeling, based on the detailed topology and control parameters of the converter, establishes an analytical expression for impedance through theoretical derivation. This method relies on a complete understanding of the white-box parameters within the unit and is thus called white-box modeling. Active disturbance measurement methods utilize a high-power disturbance source to inject a broadband sweep harmonic signal into the grid connection point, collect the response voltage and current, and calculate the impedance value at each frequency using Fourier transform. This is currently a relatively direct measurement method in engineering. In recent years, some research has also attempted to use data-driven methods for impedance modeling, constructing simple regression models by collecting electrical quantity data under a limited number of operating conditions.

[0004] However, the aforementioned existing technologies all have significant drawbacks in practical applications. For physical analysis modeling, new energy generating units often behave as black boxes in actual engineering projects, making it difficult to obtain their internal control parameters and detailed topology. Furthermore, for large wind farms containing multiple units, the derivation process of the aggregated analytical model is extremely complex, making it difficult to guarantee engineering accuracy. For active disturbance measurement methods, this method requires expensive additional hardware, the frequency sweep measurement process is time-consuming, and injecting harmonics into the grid can interfere with the normal operation of the units, failing to meet the real-time online impedance monitoring requirements under all operating conditions. As for existing data-driven methods, their sample coverage is usually limited to a single operating condition or a few frequency points, lacking a systematic training and verification process, and the model's generalization ability and reliability are insufficient to support practical engineering applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the reliance on white-box parameters for impedance modeling of new energy units, difficulty in adapting to black-box environments, complex site aggregation calculations, and high cost and interference with system operation of traditional active measurement methods. This invention proposes a new energy unit full black-box impedance identification method and system that considers phase angle correction.

[0006] The technical solution of this invention is as follows: This invention provides a method for identifying the impedance of a new energy power unit in a completely black box, taking into account phase angle correction, including: Acquire real-time characteristic data of the new energy unit under the current operating conditions. The real-time characteristic data includes the frequency, operating parameters, and dq axis electrical quantities of the new energy unit. The real-time feature data is input into a trained deep feedforward neural network for forward inference to obtain the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance of the new energy unit at the corresponding frequency point. The positive sequence impedance amplitude and phase angle at the corresponding frequency point are obtained from the real and imaginary parts of the positive sequence impedance; the negative sequence impedance amplitude and phase angle at the corresponding frequency point are obtained from the real and imaginary parts of the negative sequence impedance. For the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold, linear interpolation is used to correct it, and the corrected positive sequence impedance phase angle is obtained. Based on the positive sequence impedance amplitude, the corrected positive sequence impedance phase angle, the negative sequence impedance amplitude, and the negative sequence impedance phase angle at each frequency point under the current operating conditions, the impedance amplitude-frequency and phase frequency characteristics of the new energy unit are obtained.

[0007] Preferably, the operating parameters include: d-axis reference current, phase-locked loop proportional coefficient, grid voltage per unit value, and current loop proportional coefficient; The dq axis electrical quantities include: d-axis voltage, q-axis voltage, d-axis current, q-axis current, d-axis active power, q-axis active power, d-axis reactive power, and q-axis reactive power.

[0008] Preferably, the deep feedforward neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer cascaded in sequence, wherein each hidden layer is followed by a ReLU activation function.

[0009] Preferably, the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold is corrected by linear interpolation to obtain the corrected positive sequence impedance phase angle, including: Traverse all frequency points and identify the frequency points where the positive sequence impedance amplitude is lower than the preset threshold as distortion points; For the positive sequence impedance phase angle at each distortion point, linear interpolation is performed using the positive sequence impedance phase angles of the nearest non-low amplitude points on the left and right sides of the distortion point to obtain the corrected positive sequence impedance phase angle. The non-low amplitude points are frequency points where the positive sequence impedance amplitude is not lower than the preset threshold.

[0010] Preferably, the formula for calculating the corrected positive-sequence impedance phase angle is: ; In the formula, This is the corrected positive sequence impedance phase angle. The frequency of the nearest non-low amplitude point on the right. The frequency of the distortion point, The frequency of the nearest non-low amplitude point on the left. The positive sequence impedance phase angle is the point on the left that is not a low-amplitude point. The positive sequence impedance phase angle is the closest non-low amplitude point on the right.

[0011] Preferably, the training process of the deep feedforward neural network includes: S1: Set multiple variable operating condition parameters and perform frequency sweep simulation on the new energy unit under each operating condition. Based on the electrical response at each frequency point, obtain multiple original sample data to form an original sample set. The original sample data includes feature data and impedance label data. S2: The original sample set is preprocessed and then divided into a training set and a validation set; S3: The deep feedforward neural network is trained using the training set to obtain a trained deep neural network.

[0012] Preferably, S1 includes: S11: Set variable operating condition parameters, and perform gridded value acquisition on each operating condition parameter to form multiple sets of operating condition combinations. The variable operating condition parameters include at least the d-axis reference current, phase-locked loop proportional coefficient, grid voltage per unit value and current loop proportional coefficient. S12: Under the operating conditions corresponding to each operating condition combination, perform frequency band sweep simulation with variable step size for the frequency range of 10Hz-5000Hz. The 10Hz-1500Hz band uses a step size of 2Hz, and the 1500Hz-5000Hz band uses a step size of 10Hz. S13: Inject positive-sequence perturbation voltage and negative-sequence perturbation voltage at each frequency point, calculate the positive-sequence impedance and negative-sequence impedance corresponding to each frequency point, and calculate the dq-axis electrical quantity corresponding to each frequency point using the complex value superposition method of positive and negative sequence components. S14: Use the frequency, operating parameters and dq axis electrical quantities corresponding to each frequency point as the feature data, and use the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance corresponding to each frequency point as the impedance label data to obtain the original sample data corresponding to each frequency point. Multiple original sample data constitute the original sample set.

[0013] Preferably, S2 includes: S21: Perform deduplication and outlier removal on the original sample set; S22: Divide the original sample data into four frequency bands according to frequency, and stratify each frequency band according to the combination of operating parameters. Within each stratum, the training set and the validation set are randomly divided in a 7:3 ratio. The four frequency bands include: 10-100Hz, 100-500Hz, 500-1500Hz and 1500-5000Hz. S23: Perform Z-score normalization on the feature data and impedance label data of the training set and the validation set, respectively.

[0014] Preferably, S3 includes: The Adam optimizer is used, and the initial learning rate, learning rate decay strategy, mini-batch sample size, and maximum number of iterations are set. The deep feedforward neural network is trained using the training set. During the training process, the mean squared error is calculated periodically using the validation set according to the preset training batch. If the mean squared error corresponding to the validation set does not decrease for a preset number of consecutive times, the training is terminated early, and the model parameters with the smallest mean squared error in the validation set are saved as the optimal model parameters, thus obtaining the trained deep neural network.

[0015] This invention also provides a black-box impedance identification system for new energy generating units that considers phase angle correction, applicable to the black-box impedance identification method for new energy generating units that considers phase angle correction described in any of the above embodiments, including: The data acquisition module is used to acquire real-time characteristic data of the new energy unit under the current operating conditions. The real-time characteristic data includes the frequency, operating parameters and dq axis electrical quantities of the new energy unit. The impedance identification module is used to input the real-time feature data into a trained deep feedforward neural network for forward inference to obtain the real and imaginary parts of the positive-sequence impedance and the real and imaginary parts of the negative-sequence impedance of the new energy unit at the corresponding frequency point. The data processing correction module is used to obtain the positive sequence impedance amplitude and positive sequence impedance phase angle at the corresponding frequency point based on the real and imaginary parts of the positive sequence impedance, and to obtain the negative sequence impedance amplitude and negative sequence impedance phase angle at the corresponding frequency point based on the real and imaginary parts of the negative sequence impedance; and to correct the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold using linear interpolation to obtain the corrected positive sequence impedance phase angle. The impedance characteristic acquisition module is used to obtain the impedance amplitude-frequency and phase-frequency characteristics of the new energy unit based on the positive sequence impedance amplitude, the corrected positive sequence impedance phase angle, the negative sequence impedance amplitude, and the negative sequence impedance phase angle at each frequency point under the current operating conditions.

[0016] The beneficial effects of this invention are: This invention presents a black-box impedance identification method for new energy generating units that considers phase angle correction. Compared to traditional frequency sweep instrument measurement methods, this method utilizes a trained model for online inference, eliminating the need to inject harmonic disturbances into the grid and thus not affecting normal unit power generation. Furthermore, it boasts extremely fast calculation speed, meeting the requirements for real-time online stability analysis. Addressing the singularity problem in arctangent function phase angle calculation caused by the amplitude approaching zero during complex-to-polar coordinate conversion, a linear interpolation post-processing algorithm for phase angle is proposed, completely eliminating phase angle jumps and ensuring the smoothness and accuracy of the final output impedance curve. Compared to traditional physical modeling, this method avoids complex analytical derivations and site aggregation calculations, and can be directly applied to the rapid acquisition of impedance from new energy generating units, providing an efficient tool for oscillation analysis. Attached Figure Description

[0017] Figure 1 The flowchart shown is a process for identifying the impedance of a new energy unit in a black box considering phase angle correction. Figure 2 The diagram shown is a schematic of a direct-drive wind turbine generator system. Figure 3 The diagram shows the topology of a direct-drive wind power grid-connected converter. Figure 4 The diagram shows the overall flowchart of training and prediction for a deep feedforward neural network. Figure 5 The diagram shown is a schematic of a BP neural network structure. Figure 6 The figure shows the positive sequence impedance amplitude frequency response of a direct-drive wind turbine under nine operating conditions. Figure 7 The figure shows the positive sequence impedance phase angle frequency response of a direct-drive wind turbine under nine operating conditions. Figure 8 The figure shows the output prediction effect of the impedance BP network for a direct-drive wind turbine. Figure 9 The image shown is a visualization of the prediction error. Figure 10 The diagram shown is a comparison of impedance amplitude distribution under all operating conditions. Figure 11 The diagram shown is a comparison of impedance phase angle distribution under all operating conditions. Figure 12 The diagram shown is a structural block diagram of a black-box impedance identification system for new energy generating units that takes phase angle correction into account. Detailed Implementation

[0018] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0019] In a first aspect, embodiments of the present invention provide a black-box impedance identification method for new energy generating units that considers phase angle correction, which can achieve uninterrupted, high-precision online prediction of wideband impedance using only measurable external electrical quantities.

[0020] Please see Figure 1 , Figure 1 The diagram shows a flowchart of a black-box impedance identification method for new energy generating units considering phase angle correction. Figure 1 As shown, the black-box impedance identification method for new energy generating units considering phase angle correction in this embodiment includes the following steps: Step 1: Obtain real-time characteristic data of the new energy unit under the current operating conditions. The real-time characteristic data includes the frequency of the new energy unit. Operating parameters and electrical quantities of the dq axis.

[0021] In this embodiment, the operating parameters include: d-axis reference current. Phase-locked loop proportional coefficient Grid voltage per unit value and current loop proportionality coefficient The electrical quantities of the dq axis include: d-axis voltage. q-axis voltage d-axis current q-axis current d-axis active power q-axis active power d-axis reactive power and q-axis reactive power .

[0022] For example, real-time feature data can be represented as .

[0023] Step 2: Input the real-time feature data into the trained deep feedforward neural network for forward inference to obtain the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance of the new energy unit at the corresponding frequency point.

[0024] In this embodiment, the deep feedforward neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer that are cascaded in sequence, wherein each hidden layer is followed by a ReLU activation function.

[0025] Optionally, the deep feedforward neural network can be a backpropagation (BP) neural network. In this embodiment, the input layer of the BP neural network has 13 neurons, corresponding to 13-dimensional feature data. The first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the third hidden layer has 32 neurons. Each hidden layer is followed by a ReLU activation function. The output layer has 4 neurons and uses a linear regression layer to output a 4-dimensional predicted impedance vector. That is, the real and imaginary parts of the positive-sequence impedance and the real and imaginary parts of the negative-sequence impedance; the network is finally connected to a regression layer, and the entire network can be represented as a mapping .

[0026] In this embodiment, the width of the hidden layers decreases progressively, which helps extract high-level features and prevents overfitting; the ReLU activation function alleviates the gradient vanishing problem and accelerates convergence. The output layer uses a linear regression layer to directly output continuous values.

[0027] It is worth noting that when using a trained deep feedforward neural network to perform forward inference on real-time feature data, it is necessary to use the feature mean saved in the training set during the training process of the deep feedforward neural network. with standard deviation The real-time feature data is normalized using standard methods, and the mean of the labels stored in the training set is used. with standard deviation The predicted values ​​of the real and imaginary parts of the positive and negative sequence impedances output by the deep feedforward neural network are denormalized to obtain the actual physical values.

[0028] Step 3: Based on the real and imaginary parts of the positive sequence impedance, obtain the positive sequence impedance amplitude and phase angle at the corresponding frequency point. Based on the real and imaginary parts of the negative sequence impedance, obtain the negative sequence impedance amplitude and phase angle at the corresponding frequency point.

[0029] In this embodiment, the positive / negative sequence impedance magnitude and phase angle can be obtained through polar coordinate transformation.

[0030] It should be noted that when the impedance amplitude approaches zero, both the real and imaginary parts are close to 0. In this case, the calculated impedance phase angle will exhibit random jumps (around ±π or 0), resulting in a discontinuous phase angle curve. Therefore, linear interpolation correction of the phase angle is required in the low-amplitude frequency band. Since the negative sequence impedance amplitude is usually always large, only the positive sequence impedance phase angle needs to be corrected.

[0031] Step 4: Correct the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than the preset threshold by using linear interpolation to obtain the corrected positive sequence impedance phase angle.

[0032] In this embodiment, step 4 includes: Step 4.1: Traverse all frequency points and identify the frequency points where the positive sequence impedance amplitude is lower than the preset threshold as distortion points; Step 4.2: For the positive sequence impedance phase angle of each distortion point, perform linear interpolation correction using the positive sequence impedance phase angles of the nearest non-low amplitude points on the left and right sides of the distortion point to obtain the corrected positive sequence impedance phase angle. Among them, the non-low amplitude points are the frequency points where the positive sequence impedance amplitude is not lower than the preset threshold.

[0033] Specifically, the formula for calculating the corrected positive-sequence impedance phase angle is as follows: ; In the formula, This is the corrected positive sequence impedance phase angle. The frequency of the nearest non-low amplitude point on the right. The frequency of the distortion point, The frequency of the nearest non-low amplitude point on the left. The positive sequence impedance phase angle is the point on the left that is not a low-amplitude point. The positive sequence impedance phase angle is the closest non-low amplitude point on the right.

[0034] Optionally, a preset threshold can be set. .

[0035] Step 5: Based on the positive sequence impedance amplitude, the corrected positive sequence impedance phase angle, the negative sequence impedance amplitude, and the negative sequence impedance phase angle at each frequency point under the current operating conditions, obtain the impedance amplitude-frequency and phase frequency characteristics of the new energy unit.

[0036] In this embodiment, the impedance amplitude-frequency characteristics of the new energy unit can be obtained based on the positive-sequence impedance amplitude and negative-sequence impedance amplitude at each frequency point, and the impedance phase-frequency characteristics of the new energy unit can be obtained based on the corrected positive-sequence impedance phase angle and negative-sequence impedance phase angle at each frequency point. After correction, the phase angle curve across the entire frequency band is smooth without jumps, ultimately outputting accurate broadband impedance amplitude-frequency and phase-frequency characteristics.

[0037] Furthermore, the training process of the deep feedforward neural network in this embodiment will be explained.

[0038] In this embodiment, the training process of the deep feedforward neural network includes the following steps: S1: Set multiple variable operating condition parameters and perform frequency sweep simulation on the new energy unit under each operating condition. Based on the electrical response at each frequency point, obtain multiple original sample data to form an original sample set. The original sample data includes feature data and impedance label data.

[0039] In this embodiment, S1 includes: S11: Set variable operating condition parameters, and perform gridded value acquisition for each operating condition parameter to form multiple operating condition combinations. The variable operating condition parameters shall include at least the d-axis reference current. Phase-locked loop proportional coefficient Grid voltage per unit value and current loop proportionality coefficient .

[0040] S12: Under the operating conditions corresponding to each operating condition combination, perform frequency band sweep simulation with variable step size for the frequency range of 10Hz-5000Hz. The 10Hz-1500Hz band uses a step size of 2Hz, and the 1500Hz-5000Hz band uses a step size of 10Hz.

[0041] In this embodiment, the resonant sensitive region of 10Hz-1500Hz is subjected to encrypted scanning.

[0042] S13: Inject positive-sequence perturbation voltage and negative-sequence perturbation voltage at each frequency point, calculate the positive-sequence impedance and negative-sequence impedance corresponding to each frequency point, and calculate the dq-axis electrical quantity corresponding to each frequency point using the complex value superposition method of positive and negative sequence components.

[0043] Specifically, for each frequency point Inject positive-sequence small perturbation voltages respectively and negative sequence small disturbance voltage Calculate the corresponding positive sequence impedance. and negative sequence impedance Simultaneously, the dq-axis voltage is calculated by superimposing the complex values ​​of the positive and negative sequence components. , dq axis current , dq axis active power , and reactive power , .

[0044] S14: Use the frequency, operating parameters and dq axis electrical quantities corresponding to each frequency point as feature data, and use the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance corresponding to each frequency point as impedance label data to obtain the original sample data corresponding to each frequency point. Multiple original sample data constitute the original sample set.

[0045] In this embodiment, the frequency Current operating parameters and , , , , , , , As feature data, it is represented as: ; positive sequence impedance real part With the imaginary part Negative sequence impedance real part With the imaginary part As impedance tag data, it is represented as: .

[0046] S2: After preprocessing the original sample set, it is divided into a training set and a validation set.

[0047] In this embodiment, S2 includes: S21: Perform deduplication and outlier removal on the original sample set.

[0048] In this embodiment, outliers include NaN / Inf outliers and samples with an absolute power value greater than 2 pu.

[0049] S22: Divide the original sample data into four frequency bands according to frequency, and stratify each frequency band according to the combination of operating parameters. Within each stratum, randomly divide the data into training set and validation set in a 7:3 ratio.

[0050] In this embodiment, the four frequency bands include: 10-100Hz, 100-500Hz, 500-1500Hz, and 1500-5000Hz.

[0051] S23: Perform Z-score standardization on the feature data and impedance label data of the training set and validation set respectively.

[0052] In this embodiment, the feature mean can be calculated using the training set. with standard deviation and label mean with standard deviation Z-score normalization is performed on both the training and validation sets, as shown in the following process: ; In the formula, This represents the feature data after standard normalization. This represents the impedance tag data after standard normalization. The feature data of the original sample, This is the impedance label data for the original sample.

[0053] S3: Train the constructed deep feedforward neural network using the training set to obtain a trained deep neural network.

[0054] In this embodiment, S3 includes: using the Adam optimizer, setting the initial learning rate, learning rate decay strategy, mini-batch sample size and maximum iteration rounds, training the deep feedforward neural network using the training set, periodically calculating the mean squared error using the validation set according to the preset training batch during the training process, and terminating the training early if the mean squared error corresponding to the validation set does not decrease for a preset number of consecutive times, and saving the model parameters with the smallest mean squared error of the validation set as the optimal model parameters, thereby obtaining the trained deep neural network.

[0055] For example, the initial learning rate The learning rate decays to 0.5 every 10 training epochs. The mini-batch size is set to 1024, and the maximum number of training epochs is set to 30. The mean squared error (MSE) is calculated on the validation set every 20 batches. Training is terminated early if the validation set loss does not decrease for 5 consecutive validations.

[0056] Optionally, several representative operating conditions can be selected, and the theoretical impedance values ​​can be obtained through frequency sweep simulation calculations using a physical model. And the predicted impedance value is obtained through a trained deep feedforward neural network. Compare the amplitude and phase angle of the two to calculate the absolute error. and relative error ,in To prevent division by zero by a preset positive decimal, a comparison curve and an error bar chart are plotted to evaluate the model accuracy of the deep feedforward neural network in this embodiment.

[0057] This invention generates a large amount of multi-condition data through frequency sweep simulation to construct neural network training samples, solving the problem of not being able to obtain internal parameters in a black-box environment, and realizing impedance prediction based solely on externally measurable quantities (such as voltage, current, and power). When constructing the training set features, an innovative method of superimposing complex values ​​of positive and negative order components is used, strictly preserving the phase interaction information of electrical signals and avoiding data distortion caused by premature modulus taking in traditional methods. A hierarchical partitioning strategy is adopted, dividing the samples into layers according to frequency segments and combinations of operating condition parameters, and independently dividing the training and validation sets within each layer. This ensures that the operating condition distribution and frequency distribution of the training and validation sets are completely consistent, significantly improving the model's generalization ability across all operating conditions and frequency bands.

[0058] Furthermore, specific simulation experiments are used to exemplarily illustrate the training and model performance of the deep feedforward neural network in this embodiment of the invention. Please refer to [link / reference]. Figure 4 , Figure 4 The diagram shown is an overall flowchart of the training and prediction process for a deep feedforward neural network. Figure 4 It demonstrates the entire process from network construction, training, and prediction.

[0059] Taking a direct-drive permanent magnet synchronous wind turbine as an example, its system structure is as follows: Figure 2 As shown, the grid-connected converter topology is as follows: Figure 3 As shown, Figure 2 The diagram shown is a schematic of a direct-drive wind turbine generator system. Figure 3 The diagram shows the topology of a direct-drive wind turbine grid-connected converter. The basic parameters of the direct-drive permanent magnet synchronous wind turbine are shown in Table 1.

[0060] Table 1

[0061] When constructing the original sample set, the operating condition parameter range is set as follows: d-axis reference current. Use 0.5~1.2 pu (step size 0.1) for the phase-locked loop proportional gain. Use 0.5 to 1.5 times the reference value (step size 0.1 times), per-unit value of grid voltage. Use a value of 0.9~1.1 pu (step size 0.1), current loop proportional coefficient Use a value 0.5 to 1.5 times the baseline (step size 0.1 times). Perform mesh generation on the above four types of parameters to generate... Group operating conditions combination. Frequency scanning range 10~5000Hz, with a step size of 2Hz in the 10~1500Hz range (resonance region densification) and a step size of 10Hz in the 1500~5000Hz range, generating a total of 2341 frequency points.

[0062] Under each operating condition, a small positive-sequence perturbation voltage is injected at each frequency point. and negative sequence small disturbance voltage Calculate the positive sequence impedance based on the analytical expression of sequence impedance. and negative sequence impedance Simultaneously, the dq-axis voltage is calculated using the complex value superposition method of positive and negative sequence components. Current The active and reactive power data constitute 13-dimensional feature data, and the real and imaginary parts of the positive-sequence impedance and negative-sequence impedance are used as 4-dimensional impedance label data. Theoretically, approximately 6.798 million samples are generated. After data cleaning (removing duplicate samples, NaN / Inf outliers, and absolute power values ​​greater than a certain threshold),... After obtaining the initial sample, 3.18 million valid samples were obtained.

[0063] The samples were divided into four frequency bands: 10-100Hz, 100-500Hz, 500-1500Hz, and 1500-5000Hz. Layered labels were generated based on operating parameters. Within each layer, the training and validation sets were randomly divided in a 7:3 ratio, with 2.227 million training samples and 954,000 validation samples. Then, the feature data and impedance label data of both the training and validation sets were Z-score standardized.

[0064] The structure of a BP neural network as a deep feedforward neural network is as follows: Figure 5 As shown, Figure 5 The diagram shows the structure of a backpropagation (BP) neural network. The input layer has 13 neurons, corresponding to 13-dimensional feature data; the first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the third hidden layer has 32 neurons. Each hidden layer uses the ReLU activation function; the output layer has 4 neurons, no activation function, and uses linear regression for output.

[0065] The BP neural network is trained using the acquired training set, specifically using the Adam optimizer, with an initial learning rate of... The learning rate decays to 0.5 every 10 training epochs. The mini-batch size is set to 1024, and the maximum number of training epochs is set to 30. The mean squared error (MSE) is calculated on the validation set after every 20 batches. If the validation set loss does not decrease for five consecutive validations, training is terminated early, and the model with the smallest validation set loss is saved. The training process took approximately 906.8 minutes on a CPU, and the model converged stably without triggering early termination. After training, the mean relative error is calculated on the validation set: 3.51% for the real part and 5.58% for the imaginary part of positive-sequence impedance; 0.29% for the real part and 3.71% for the imaginary part of negative-sequence impedance; and the coefficient of determination is calculated. All are greater than 0.99.

[0066] Three representative operating conditions were selected to test the trained deep feedforward neural network. The specific operating conditions included: Operating Condition 1: Operating Condition 2: Operating Condition 3: Operating conditions 1, 2, and 3 represent three typical operating conditions: the unit's output power is below the lower limit, the rated power, and the upper limit, respectively.

[0067] Impedance data for a wide bandwidth of 10-5000Hz was obtained through analytical calculations using a mechanistic model and predictions using a backpropagation neural network. The data was then compared after amplitude-phase conversion of the real and imaginary parts. (See [link to documentation]). Figures 6-9 , Figure 6 and Figure 7The amplitude and phase angle frequency characteristics of the positive sequence impedance of the direct-drive wind turbine are shown in nine representative operating conditions, including the lower limit of the unit output power, medium and low output, rated, medium and high output, upper limit, low voltage ride-through, overvoltage, weak grid adaptation, and control parameter mismatch. Figure 8 The image shows the output prediction effect of the direct-drive wind resistance BP network; Figure 9 The image shown is a visualization of the prediction error. Among them, Figure 6 The frequency characteristics of the positive sequence impedance amplitude under nine representative operating conditions are shown. Figure 7 The frequency characteristics of the positive sequence impedance phase angle under nine representative operating conditions are shown. Figure 8 The network's prediction performance for four outputs is demonstrated. Figure 9 The distribution of prediction errors is shown. As can be seen from the figure, the predicted curve and the theoretical curve are highly consistent across the entire frequency band, with only minor deviations in the low amplitude range. The maximum absolute error of amplitude is less than 40Ω, and the maximum absolute error of phase angle is less than 0.6rad. Moreover, the errors are mainly concentrated in non-critical frequency bands with amplitudes less than 2Ω, and have no substantial impact on stability analysis.

[0068] Furthermore, to verify the model's continuous generalization capability across the entire power output range and frequency band of the wind turbine, a reference current along the d-axis was selected. The positive sequence impedance amplitude was obtained by continuously varying the parameter from 0.4 pu to 1.2 pu, covering the full parameter space from 10 Hz to 5000 Hz, through analytical calculation using a mechanistic model and prediction using a BP neural network. and phase angle The distribution surface, such as Figure 10 and Figure 11 As shown. Figure 10 The diagram shown is a comparison of impedance amplitude distribution under all operating conditions. Figure 11 The diagram shown is a comparison of impedance phase angle distribution under all operating conditions.

[0069] As shown in the figure, the impedance amplitude surface predicted by the neural network is basically consistent with the surface calculated analytically by the mechanistic model within the covered operating conditions and frequency bands. There are no significant deviations in the resonant peak characteristics of the low-frequency band, the attenuation trend of the mid-frequency band, and the inductance-dominated characteristics of the high-frequency band. The predicted phase angle distribution surface is also basically consistent with the calculation results of the mechanistic model, with no trend deviation across the entire frequency band. A slight deviation exists in non-critical frequency bands where the amplitude is below 2Ω, but this deviation can be eliminated after post-processing linear interpolation correction. These results demonstrate that the trained deep feedforward neural network can effectively predict impedance characteristics within the covered operating conditions and frequency range.

[0070] The black-box impedance identification method for new energy generating units, considering phase angle correction in this invention, compared to traditional frequency sweep instrument measurement methods, utilizes a trained model for online inference. This eliminates the need to inject harmonic disturbances into the grid, thus not affecting normal unit power generation, and offers extremely fast calculation speed, meeting the requirements of real-time online stability analysis. To address the singularity problem in arctangent function phase angle calculation caused by the amplitude approaching zero during complex-to-polar coordinate conversion, a linear phase angle interpolation post-processing algorithm is proposed, completely eliminating phase angle jumps and ensuring the smoothness and accuracy of the final output impedance curve. Compared to traditional physical modeling, this method avoids complex analytical derivations and site aggregation calculations, and can be directly applied to the rapid acquisition of impedance from new energy generating units, providing an efficient tool for oscillation analysis.

[0071] Secondly, embodiments of the present invention provide a black-box impedance identification system for new energy generating units that considers phase angle correction, applicable to the black-box impedance identification method for new energy generating units that considers phase angle correction provided in the first aspect.

[0072] Please see Figure 12 , Figure 12 The diagram shown is a structural block diagram of a black-box impedance identification system for new energy generating units considering phase angle correction. Figure 12 As shown, the all-black-box impedance identification system for new energy generating units considering phase angle correction in this embodiment includes: a data acquisition module, an impedance identification module, a data processing and correction module, and an impedance characteristic acquisition module.

[0073] The data acquisition module acquires real-time characteristic data of the new energy generator unit under current operating conditions. This real-time characteristic data includes the unit's frequency, operating parameters, and dq-axis electrical quantities. The impedance identification module inputs the real-time characteristic data into a trained deep feedforward neural network for forward inference, obtaining the real and imaginary parts of the positive-sequence impedance and the negative-sequence impedance at corresponding frequency points. The data processing and correction module obtains the positive-sequence impedance amplitude and phase angle at corresponding frequency points based on the real and imaginary parts of the positive-sequence impedance, and the negative-sequence impedance amplitude and phase angle based on the real and imaginary parts of the negative-sequence impedance. For frequency points where the positive-sequence impedance amplitude is lower than a preset threshold, linear interpolation is used to correct the positive-sequence impedance phase angle, resulting in a corrected positive-sequence impedance phase angle. The impedance characteristic acquisition module obtains the impedance amplitude-frequency and phase-frequency characteristics of the new energy generator unit based on the positive-sequence impedance amplitude, corrected positive-sequence impedance phase angle, negative-sequence impedance amplitude, and negative-sequence impedance phase angle at each frequency point under current operating conditions.

[0074] For details regarding the specific content and corresponding beneficial effects of the black-box impedance identification system for new energy generating units that considers phase angle correction, please refer to the relevant content of the first aspect providing the black-box impedance identification method for new energy generating units that considers phase angle correction; it will not be elaborated here.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0076] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0077] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction, characterized in that, include: Acquire real-time characteristic data of the new energy unit under the current operating conditions. The real-time characteristic data includes the frequency, operating parameters, and dq axis electrical quantities of the new energy unit. The real-time feature data is input into a trained deep feedforward neural network for forward inference to obtain the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance of the new energy unit at the corresponding frequency point. The positive sequence impedance amplitude and phase angle at the corresponding frequency point are obtained from the real and imaginary parts of the positive sequence impedance; the negative sequence impedance amplitude and phase angle at the corresponding frequency point are obtained from the real and imaginary parts of the negative sequence impedance. For the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold, linear interpolation is used to correct it, and the corrected positive sequence impedance phase angle is obtained. Based on the positive sequence impedance amplitude, the corrected positive sequence impedance phase angle, the negative sequence impedance amplitude, and the negative sequence impedance phase angle at each frequency point under the current operating conditions, the impedance amplitude-frequency and phase frequency characteristics of the new energy unit are obtained.

2. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction as described in claim 1, characterized in that, The operating parameters include: d-axis reference current, phase-locked loop proportional coefficient, grid voltage per unit value, and current loop proportional coefficient; The dq axis electrical quantities include: d-axis voltage, q-axis voltage, d-axis current, q-axis current, d-axis active power, q-axis active power, d-axis reactive power, and q-axis reactive power.

3. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction as described in claim 1, characterized in that, The deep feedforward neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer, which are cascaded in sequence. Each hidden layer is followed by a ReLU activation function.

4. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 1, characterized in that, For the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold, linear interpolation is used to correct it, resulting in a corrected positive sequence impedance phase angle, including: Traverse all frequency points and identify the frequency points where the positive sequence impedance amplitude is lower than the preset threshold as distortion points. For the positive sequence impedance phase angle at each distortion point, linear interpolation is performed using the positive sequence impedance phase angles of the nearest non-low amplitude points on the left and right sides of the distortion point to obtain the corrected positive sequence impedance phase angle. The non-low amplitude points are frequency points where the positive sequence impedance amplitude is not lower than the preset threshold.

5. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 1, characterized in that, The formula for calculating the corrected positive-sequence impedance phase angle is as follows: ； In the formula, This is the corrected positive sequence impedance phase angle. The frequency of the nearest non-low amplitude point on the right. The frequency of the distortion point, The frequency of the nearest non-low amplitude point on the left. The positive sequence impedance phase angle is the point on the left that is not a low-amplitude point. The positive sequence impedance phase angle is the closest non-low amplitude point on the right.

6. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 1, characterized in that, The training process of the deep feedforward neural network includes: S1: Set multiple variable operating condition parameters and perform frequency sweep simulation on the new energy unit under each operating condition. Based on the electrical response at each frequency point, obtain multiple original sample data to form an original sample set. The original sample data includes feature data and impedance label data. S2: The original sample set is preprocessed and then divided into a training set and a validation set; S3: The deep feedforward neural network is trained using the training set to obtain a trained deep neural network.

7. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 6, characterized in that, S1 includes: S11: Set variable operating condition parameters, and perform gridded value acquisition on each operating condition parameter to form multiple sets of operating condition combinations. The variable operating condition parameters include at least the d-axis reference current, phase-locked loop proportional coefficient, grid voltage per unit value and current loop proportional coefficient. S12: Under the operating conditions corresponding to each operating condition combination, perform frequency band sweep simulation with variable step size for the frequency range of 10Hz-5000Hz. The 10Hz-1500Hz band uses a step size of 2Hz, and the 1500Hz-5000Hz band uses a step size of 10Hz. S13: Inject positive-sequence perturbation voltage and negative-sequence perturbation voltage at each frequency point, calculate the positive-sequence impedance and negative-sequence impedance corresponding to each frequency point, and calculate the dq-axis electrical quantity corresponding to each frequency point using the complex value superposition method of positive and negative sequence components. S14: Use the frequency, operating parameters and dq axis electrical quantities corresponding to each frequency point as the feature data, and use the real and imaginary parts of the positive sequence impedance and the real and imaginary parts of the negative sequence impedance corresponding to each frequency point as the impedance label data to obtain the original sample data corresponding to each frequency point. Multiple original sample data constitute the original sample set.

8. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 6, characterized in that, S2 includes: S21: Perform deduplication and outlier removal on the original sample set; S22: Divide the original sample data into four frequency bands according to frequency, and stratify each frequency band according to the combination of operating parameters. Within each stratum, the training set and the validation set are randomly divided in a 7:3 ratio. The four frequency bands include: 10-100Hz, 100-500Hz, 500-1500Hz and 1500-5000Hz. S23: Perform Z-score normalization on the feature data and impedance label data of the training set and the validation set, respectively.

9. The method for identifying the impedance of a new energy unit in a completely black box considering phase angle correction according to claim 6, characterized in that, S3 includes: The Adam optimizer is used, and the initial learning rate, learning rate decay strategy, mini-batch sample size, and maximum number of iterations are set. The deep feedforward neural network is trained using the training set. During the training process, the mean squared error is calculated periodically using the validation set according to the preset training batch. If the mean squared error corresponding to the validation set does not decrease for a preset number of consecutive times, the training is terminated early, and the model parameters with the smallest mean squared error in the validation set are saved as the optimal model parameters, thus obtaining the trained deep neural network.

10. A black-box impedance identification system for new energy generating units considering phase angle correction, characterized in that, The all-black-box impedance identification method for new energy generating units considering phase angle correction, applicable to any one of claims 1-9, includes: The data acquisition module is used to acquire real-time characteristic data of the new energy unit under the current operating conditions. The real-time characteristic data includes the frequency, operating parameters and dq axis electrical quantities of the new energy unit. The impedance identification module is used to input the real-time feature data into a trained deep feedforward neural network for forward inference to obtain the real and imaginary parts of the positive-sequence impedance and the real and imaginary parts of the negative-sequence impedance of the new energy unit at the corresponding frequency point. The data processing correction module is used to obtain the positive sequence impedance amplitude and positive sequence impedance phase angle at the corresponding frequency point based on the real and imaginary parts of the positive sequence impedance, and to obtain the negative sequence impedance amplitude and negative sequence impedance phase angle at the corresponding frequency point based on the real and imaginary parts of the negative sequence impedance; and to correct the positive sequence impedance phase angle corresponding to the frequency point where the positive sequence impedance amplitude is lower than a preset threshold using linear interpolation to obtain the corrected positive sequence impedance phase angle. The impedance characteristic acquisition module is used to obtain the impedance amplitude-frequency and phase-frequency characteristics of the new energy unit based on the positive sequence impedance amplitude, the corrected positive sequence impedance phase angle, the negative sequence impedance amplitude, and the negative sequence impedance phase angle at each frequency point under the current operating conditions.

Images