New energy station sequence impedance identification model training method and power system stability analysis method

Abstract

This invention belongs to the technical field of power system analysis, and discloses a training method for a sequence impedance identification model of new energy power plants and a power system stability analysis method. The training method includes: acquiring a clean dataset and a noisy dataset; using the noisy dataset as the training set and the clean dataset as the validation set, training the original identification model using the training set until its impedance identification performance on the validation set reaches a preset level; using both the clean and noisy datasets to perform adversarial training on the model, with different weights for noisy samples in different harmonic frequency bands in the loss function used for adversarial training; using the clean dataset to perform adversarial training on the fine-tuned identification model, and outputting the trained identification model. This training method, by incorporating statistical knowledge of the noise distribution of impedance samples and adversarial training strategies, offline construction of a single-machine impedance identification model can effectively improve the model's identification accuracy and generalization ability.

Description

Technical Field

[0001] This invention belongs to the technical field of power system analysis, and more specifically, relates to a training method for a sequence impedance identification model of new energy power plants and a method for power system stability analysis. Background Technology

[0002] Currently, the large-scale grid connection of new energy power plants, represented by wind and solar power, is fundamentally changing the dynamic characteristics of the power system. This grid connection mode leads to a simultaneous weakening of system inertia and damping, causing the power grid to change from "strong" to "weak," and the risk of small-signal instability increases sharply. New energy power plants typically have multiple generating units, which have two control structures: grid-connected and grid-connected. Different generating units in different new energy power plants may use different control structures. Therefore, efficient and accurate online stability assessment of each power plant is an urgent need to ensure the safe operation of the power grid.

[0003] The key to stable assessment of a power plant lies in identifying the impedance parameters of each unit within the plant. Existing offline analysis methods (such as time-domain simulation, open-loop modal analysis, and state-space methods) have limited applicability due to the unknown internal parameters and control structures of many renewable energy devices, resulting in "black box" characteristics in the specific models. Existing online identification technologies essentially fall into the "grey box" category, their performance heavily reliant on prior knowledge of the equipment's control framework; such methods fail when faced with completely "black box" equipment.

[0004] To break free from reliance on physical models, pure data-driven impedance identification methods based on deep neural networks have emerged. However, in engineering practice, their identification accuracy is severely limited by "double noise," meaning that inherent noise in both training labels (impedance measurement data) and input features (power grid measurement data) significantly weakens the model's accuracy and generalization ability. Therefore, improving the identification accuracy and generalization ability of the identification model is a pressing technical challenge. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a training method for the sequence impedance identification model of new energy power plants and a power system stability analysis method, the purpose of which is to improve the identification accuracy and generalization ability of the identification model.

[0006] To achieve the above objectives, the following technical solution is proposed.

[0007] According to a first aspect of the present invention, a training method for a sequence impedance identification model for new energy power plants is provided, comprising: Obtaining Datasets: Obtain a clean dataset and a noisy dataset. The clean dataset contains input samples and corresponding noise-free sample labels, while the noisy dataset contains input samples and corresponding noise sample labels. Each input sample contains steady-state information of the power station unit, state variables of the control loop, and harmonic frequencies. Each sample label contains impedance information of the power station unit. Model coarse tuning: Using the noisy dataset as the training set and the clean dataset as the validation set, the original identification model is trained using the training set until its impedance identification performance on the validation set is improved to a preset level. The original identification model is a neural network. Model fine-tuning: The clean dataset and the noisy dataset are used together to perform adversarial training on the coarse-tuned recognition model; in the loss function of adversarial training used in the fine-tuning stage, the weight of clean samples is greater than that of noisy samples, and noisy samples of different harmonic frequency bands have different weights. The lower the harmonic frequency band, the smaller the weight of the noisy sample. Model calibration: The finely tuned recognition model is adversarially trained using the clean dataset. Model output: Outputs the recognition model after training.

[0008] According to a second aspect of the present invention, a method for power system stability analysis is provided, comprising: The steady-state information of the station unit, the state variables of the control loop, and the harmonic frequency of the new energy power station are used as input features and input into the new energy power station sequence impedance identification model obtained by the above training method to obtain the impedance information of the corresponding station unit. The impedance information of each unit in the power station is aggregated based on the real-time acquired grid strength and station topology to obtain the aggregated admittance matrix of the power station, and the back ratio matrix of the power system is constructed based on the aggregated admittance matrix. The stability of the system is analyzed based on the back ratio matrix.

[0009] According to a third aspect of the invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0010] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described above.

[0011] Overall, compared with the prior art, the technical solutions conceived in this invention have the following beneficial effects.

[0012] 1. This invention provides a training method for a sequence impedance identification model of a new energy power station. Under the circumstances where the control structure and parameters of the generating units in the new energy power station are kept commercially confidential, and impedance measurement noise and steady-state information are present, the model is trained in three stages: First, considering that noise labels are not completely random, but exhibit bounded oscillations around the true value on a macroscopic scale, with their deviation constrained by the physically feasible region, this characteristic ensures that the noisy sample set still contains true data distribution information in a statistical sense. Therefore, a large-scale noisy dataset covering a wide range of operating conditions is first used to perform preliminary parameter optimization and coarse tuning of the identification model. This allows full utilization of noise samples, enabling the model to learn effective feature-label mapping relationships from the noise samples. Second, considering that real-time input data is inevitably affected by measurement and communication noise... The process involves generating input feature noise and combining it with clean and noisy datasets. A free adversarial training mechanism is then introduced for adversarial training. This mechanism improves the model's stability and generalization ability when facing information noise or malicious attacks by applying small adversarial perturbations to the model's input. At this stage, weights are assigned to the loss function based on the different credibility levels of each sample, guiding the model to prioritize learning the feature-label mapping relationships inherent in high-quality samples. Finally, to ensure optimal model performance under ideal conditions and to perform final calibration of its decision boundaries, only clean data is used, and adversarial training continues. The aim is to eliminate any minor biases the model might have learned from noisy data in the second stage, ensuring that its decision-making behavior is fully aligned with the most credible real data distribution while maintaining adversarial robustness. This training method, by incorporating statistical knowledge of impedance sample noise distribution and adversarial training strategies, constructs a single-machine impedance identification model offline, effectively improving the identification accuracy and generalization ability of the model under the influence of dual noise.

[0013] 2. This invention also provides a power system stability analysis method. Utilizing the new energy power plant sequence impedance identification model trained by this invention, the impedance information of each unit within the power plant is automatically identified. Then, the impedance information of each unit is aggregated to obtain the aggregated admittance matrix of the power plant. Based on the aggregated admittance matrix, the retracement matrix of the power system is constructed. After obtaining the retracement matrix, system stability can be analyzed using existing methods. Because the new energy power plant sequence impedance identification model has high identification accuracy and generalization ability, it can output high-precision unit impedance information, thereby improving the accuracy of system stability analysis.

[0014] 3. Furthermore, when assessing station stability, distinguishing between grid-connected and network-structured stations and constructing different forms of system back-comparison matrices can avoid misjudgments caused by using a simplified version of the generalized Nyquist criterion. Attached Figure Description

[0015] Figure 1 This is a flowchart of the steps of the training method for the sequence impedance identification model of a new energy power station in one embodiment of the present invention; Figure 2 This is a diagram illustrating the information processing of a training method for a sequence impedance identification model for new energy power stations according to an embodiment of the present invention. Figure 3 This is a flowchart of the steps of a power system stability analysis method according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the steps of model training and applying the trained model to perform power system stability assessment in one embodiment of the present invention. Figure 5 This is a schematic diagram of the equivalent aggregate impedance of a power station in one embodiment of the present invention; Figure 6 This is a power system topology and control framework in one embodiment of the present invention; Figure 7 These are two methods used in one embodiment of the present invention to evaluate the stability of a power station under different operating conditions. (a) Stability evaluation result of a power station connected to a strong grid; (b) Stability evaluation result of a power station connected to a weak grid; (c) Stability evaluation result of a power station connected to a weak grid; (d) Stability evaluation result of a power station connected to a strong grid. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Example 1 This invention provides a training method for a sequence impedance identification model for new energy power plants, such as... Figure 1 The diagram shown is a flowchart of the steps in the training method for the sequence impedance identification model of a new energy power station according to an embodiment of the present invention. Figure 2 The diagram shown is an information processing flow diagram of the training method for the sequence impedance identification model of a new energy power station according to an embodiment of the present invention. The following is a combination of... Figure 1 and Figure 2 The training method for the sequence impedance identification model of new energy power plants proposed in this invention is described in detail.

[0018] S11. Obtain the dataset: Obtain a clean dataset and a noisy dataset. The clean dataset contains input samples and corresponding noise-free sample labels, while the noisy dataset contains input samples and corresponding noise sample labels. Each input sample contains the steady-state information of the station unit, the state variables of the control loop, and the harmonic frequency. Each sample label contains the impedance information of the station unit.

[0019] In this invention, the goal of training the identification model is to enable it to predict the impedance information of the power station unit based on known parameter information, thereby automating the impedance identification of the unit. Therefore, when training the model, its input consists of features containing key parameters such as the steady-state information of the power station unit, the state variables of the control loop, and harmonic frequencies, while its output and sample labels consist of features containing the impedance information of the power station unit.

[0020] Furthermore, in this invention, training the identification model requires both a clean dataset (free of noise) and a noisy dataset (containing noise). Therefore, two types of datasets need to be constructed. The unit port impedance data provided by the equipment manufacturer and the high-precision measurement data from the pre-grid connection commissioning phase can both constitute a clean dataset (free of noise). After the unit is actually connected to the grid, the port impedance under normal operation is monitored, constituting a noisy dataset. The process of acquiring unit data through measurement is as follows: when the unit operates to a steady state, voltage harmonics at the target frequency are injected and the corresponding response current is measured to calculate the impedance dataset; the target frequency and steady-state operating point are changed, and the above operation is repeated to obtain the impedance data corresponding to each target frequency when the unit is under different steady-state operating conditions. The measurement data from the pre-grid connection commissioning phase can be considered clean data, while the measured impedance data after grid connection generally contains noise; therefore, the data measured during this phase is considered noisy data.

[0021] It should be noted that the amount of clean data provided by equipment manufacturers is limited, and the pre-grid connection commissioning phase cannot cover all operating conditions of the unit. Some operating conditions can only be achieved after grid connection. Therefore, the clean data obtained during the pre-grid connection commissioning phase is not complete enough. If the model is trained based solely on clean data, it will seriously impair the model's generalization ability. Furthermore, the impedance data measured after grid connection generally contains noise (which can be called noise label). If the noisy data is directly used to train the model together with the clean data, the presence of noise will lead to insufficient model prediction accuracy. This invention reveals that existing noise labels are not completely random, but rather exhibit bounded oscillations around the true value on a macroscopic level. Their deviation is constrained by the physically feasible region. This characteristic ensures that the noisy sample set still contains statistically accurate data distribution information, allowing the model to learn effective feature-label mapping relationships. Furthermore, the noise distribution shows significant frequency dependence; impedance measurement errors are larger in the low-to-mid-frequency range, dominated by control components, while the impact is minimal in the high-frequency range, determined by physical hardware such as filters. Therefore, even with noise, impedance information in the high-frequency range has a much higher reliability than data from the low-to-mid-frequency range. Based on these findings, this invention employs a unique design for utilizing noisy and clean samples, thereby training an impedance identification model with strong generalization ability and high prediction accuracy.

[0022] Each input sample contains steady-state information of the power station unit, state variables of the control loop, and harmonic frequency f. In specific implementation, the steady-state information of the input sample includes the active power P, reactive power Q, voltage amplitude U, and current amplitude I at the unit port. The state variables of the control loop include the d-axis voltage obtained by transforming the voltage U and current I at the unit port through the dq axis. q-axis voltage d-axis current and q-axis current Among them, the state variables of the control loop Steady-state information at the unit port The coordinate transformation relationship between them is as follows:

[0023] Based on the above transformation relationship, it is only necessary to measure the steady-state information at the unit port. Then the state variables of the control loop can be calculated. The harmonic frequency is the frequency f of the currently injected voltage harmonic.

[0024] At this point, the input sample X can be represented as: .

[0025] Furthermore, the voltage, current, and power in the input samples can be normalized separately.

[0026] Each sample tag contains impedance information for the station's generating units. In practice, this impedance information includes the unit's port admittance matrix. The amplitude and phase angle corresponding to the four complex elements in the equation. At this point, the sample label Y can be represented as:

[0027] In the formula, Represent the admittance matrix respectively The four complex elements, These represent taking the magnitude and phase angle of the element, respectively.

[0028] S12. Model coarse tuning: Use a noisy dataset as the training set and a clean dataset as the validation set. Train the original identification model using the training set until its impedance identification performance on the validation set is improved to a preset level. The original identification model is a neural network.

[0029] In practice, the original neural network can be composed of two convolutional neural networks (CNN), one long short-term memory network (LSTM), and one multilayer perceptron (MLP) module connected in series. The CNN is used to extract features, the LSTM is used to extract frequency domain features, and the MLP is used to form the final predicted amplitude and phase angle.

[0030] This step aims to perform preliminary parameter optimization and coarse tuning of the recognition model using a large-scale, noisy dataset covering a wide range of operating conditions. During this process, the training set consists entirely of the noisy dataset, while the clean dataset is used as the validation set to monitor the model's generalization performance. When the model's performance metrics on the clean validation set reach saturation (i.e., no longer showing significant improvement), the pre-training phase terminates, and the model proceeds to the next training phase. Loss Function as follows:

[0031] In the formula, This represents the mean squared error loss function. This indicates the number of samples in a batch. Indicates the first in the batch Index of each sample, and This represents the i-th input sample and its corresponding noise sample label. Indicates will The model prediction value obtained after inputting the data into the model.

[0032] S13. Model fine-tuning: The coarse-tuned recognition model is jointly trained adversarially using a clean dataset and a noisy dataset. In the adversarial training loss function used in the fine-tuning stage, the weight of clean samples is greater than that of noisy samples, and noisy samples of different harmonic frequency bands have different weights. The lower the harmonic frequency band, the smaller the weight of noisy samples.

[0033] In addition to label noise, the model also needs to address input feature noise. Real-time input sample data acquired and transmitted by phasor measurement units (PMUs) are inevitably subject to measurement and communication noise. If the impedance identification model is not robust enough to these types of input feature disturbances, its identification results may deviate significantly even under similar system operating conditions, leading to misjudgments of station stability. Therefore, high robustness to feature noise is crucial for data-driven impedance identification models.

[0034] Therefore, to enhance the model's robustness to input perturbations, this stage aims to combine clean and noisy datasets and introduce a free adversarial training mechanism. This mechanism improves the model's stability and generalization ability when facing information noise or malicious attacks by applying small adversarial perturbations to the model's input. Similarly, when the model's performance metrics reach saturation, it proceeds to the next training stage.

[0035] In this invention, considering the uneven statistical distribution of label noise across different harmonic frequency bands—typically exhibiting larger errors in the low-to-mid frequency band and smaller errors in the high-frequency band—the noise level of a noisy dataset can be determined based on its frequency band, thereby determining the sample's credibility. Different weights are assigned based on the sample's credibility: the lower the frequency band, the greater the noise, the lower the credibility, and the smaller the weight; conversely, clean samples have the highest credibility and the highest weight. For example, the harmonic frequency band can be divided into two bands: the first band not exceeding 300Hz and the second band exceeding 300Hz. The weight of noisy samples in the first band is less than that in the second band. Understandably, the frequency band division can be flexibly chosen based on actual circumstances. In this invention, the loss function is assigned corresponding weights based on the different credibility levels of each sample, thereby guiding the model to prioritize learning the feature-label mapping relationship inherent in high-quality samples. Loss Function It can be represented as follows:

[0036] In the formula, This indicates the number of samples in a batch. Indicates the first in the batch Index of each sample, Indicates assignment to the first The weights of each sample, This represents the mean squared error loss function. and This represents the i-th input sample and its corresponding sample label. Indicates the application applied to the sample On the opposing perturbation, Indicates will Input the model and apply a perturbation The obtained model predictions; Indicates to Find the p-norm. Indicates the search for a perturbation Its size does not exceed a preset threshold. However, it can maximize the model's loss function.

[0037] S14. Model calibration: Use a clean dataset to perform adversarial training on the finely tuned recognition model.

[0038] After the first two stages, the model has developed good generalization ability and robustness under mixed data distributions. However, to ensure optimal performance under ideal conditions and to perform final calibration of its decision boundaries, this invention establishes this final stage. This stage utilizes only clean data and continues to introduce adversarial training. Its purpose is to eliminate the small biases that the model may have learned from noisy data in the second stage, ensuring that its decision-making behavior is fully aligned with the most reliable real data distribution while maintaining adversarial robustness.

[0039] Since this stage only uses clean data of equal importance, sample weighting is no longer needed. The loss function for this stage... It can be represented as follows:

[0040] In the formula, and Specifically refers to clean samples.

[0041] S15, Model Output: Outputs the recognition model after training.

[0042] After three stages of noise immunity training, a single-machine sequence impedance identification model was constructed.

[0043] The present invention, through the above-mentioned adversarial dual-noise training method, can train an accurate single-machine sequence impedance identification model on a sample set containing labeled noise, thereby reducing the dependence on high-fidelity samples, reducing the cost of impedance identification, and improving its practicality.

[0044] Example 2 This invention also provides a method for power system stability analysis, such as... Figure 3 The diagram shown is a flowchart of the steps in a power system stability analysis method according to an embodiment of the present invention. Figure 4 The diagram illustrates the steps of model training and applying the trained model to assess power system stability in one embodiment of the present invention. The following is a related explanation. Figure 3 and Figure 4 The methods for power system stability analysis are explained in detail.

[0045] S21. Input the steady-state information of the station units, the state variables of the control loop, and the harmonic frequencies of the new energy power station into the new energy power station sequence impedance identification model to obtain the impedance information of the corresponding station units.

[0046] Specifically, the steady-state information of each generator unit at the new energy power station—active power, reactive power, voltage amplitude, and current amplitude—is measured in real time, and the d-axis voltage is calculated. q-axis voltage d-axis current and q-axis current After normalization, the input feature quantities are obtained. The impedance of a single unit in normal operation within the power station is obtained by inputting it into a pre-trained single-unit sequence impedance identification model.

[0047] In addition, the steady-state phase angle of each unit relative to the common coupling point is measured as the basis for the next step of equivalent aggregation.

[0048] S22. Based on the real-time acquired grid strength and station topology, the impedance information of each station unit in the station is aggregated to obtain the station's aggregated admittance matrix, and the back-comparison matrix of the power system is constructed based on the aggregated admittance matrix.

[0049] like Figure 5 The diagram shown is an equivalent aggregated impedance diagram of a power station according to an embodiment of the present invention. First, based on the impedance information of each unit in the power station, a aggregated admittance matrix of the power station can be constructed using conventional techniques, and then a lag ratio matrix can be constructed based on the aggregated admittance matrix.

[0050] S221. Based on the impedance information of the station units, the single-unit sequence admittance matrix of the corresponding unit j can be reconstructed. j is the group index.

[0051] Specifically, based on the impedance information obtained from the new energy power plant sequence impedance identification model, the corresponding single-machine sequence admittance matrix, i.e., the local sequence admittance matrix, can be reconstructed. Since the single-machine sequence admittance matrix is ​​based on the local coordinate system, it needs to be transformed to the global coordinate system according to the following mathematical transformation:

[0052] In the formula, j is the index of the station unit. This represents the steady-state phase angle of the j-th unit. This is the coordinate transformation matrix. Indicates the first The single-unit ordered admittance matrix of the tandem generator set. Indicates the first The global order admittance matrix of the generator unit is zero for units that have been shut down, so no coordinate transformation is required.

[0053] S222. Using the generating units as nodes, based on multi-port network theory, the global ordered admittance matrices of each generating unit within the station are equivalently aggregated to obtain the node ordered admittance matrices within the station. .

[0054] Nodal Admittance Matrix It can be represented as:

[0055] In the formula, n is the number of nodes, and the off-diagonal elements are... Represents a node With nodes The negative value of the admittance matrix of the connecting branches, and the diagonal elements. Represents nodes The sum of the ordered admittance matrices of all connected branches, where all elements in the matrix are equal to the values ​​in the formula. Form. The collector line is Equivalent sequence impedance model.

[0056] S223, Based on the node sequential admittance matrix Constructing the aggregated admittance matrix of the site .

[0057] Assuming the injection frequency at the grid connection point current Since the generator power supply only generates the fundamental frequency component, the equivalent injected current at all generator nodes is zero at this frequency. Therefore, the relationship between the station voltage and current is as follows:

[0058] In the formula, , , , These are the node-order admittance matrices. The block matrix obtained by partitioning. , , , , These are the voltage responses for nodes 0 and nodes 1 through n, respectively.

[0059] eliminate The aggregate admittance matrix of the station can be obtained. for: .

[0060] S224, Based on Converged Admittance Construct the system back ratio matrix.

[0061] Typically, based on aggregate admittance When constructing the system return ratio matrix, directly assuming the station type is a grid-type station reveals that applying simplified criteria without distinguishing station types can lead to catastrophic evaluation errors when analyzing specific systems. Therefore, this embodiment distinguishes between grid-type and network-structured station types, constructing corresponding system return ratio matrices. When the station is a grid-type station, the corresponding return ratio matrix is... When the station is a network-type station, the corresponding back-to-back ratio matrix is: ;

[0062] In the formula, s represents , represents an imaginary number, Indicates the target angular frequency. , These are the grid impedance matrix and admittance matrix, respectively, which can be obtained by real-time measurement of grid strength. For the aggregate admittance of the station. for The inverse matrix.

[0063] S23. System stability analysis based on back ratio matrix.

[0064] Among these, the stability margin and potential oscillation frequency can be analyzed based on the system back ratio matrix.

[0065] Specifically, stability assessment can be based on a simplified version of the generalized Nyquist criterion. Assuming the station satisfies self-stabilizing characteristics, the assessment checks whether the Nyquist curve of the hysteresis matrix encloses the station. If the system is surrounded, determine that it is unstable and calculate the corresponding oscillation frequency; if it is not surrounded, determine that the system is not unstable and calculate the corresponding potential oscillation frequency and stability margin.

[0066] Example 3 The present invention also relates to an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0067] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0068] Example 4 The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0069] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0070] Example 5 In this embodiment, the technical solution of the present invention is verified by simulation. Specifically, taking the construction of a grid-connected substation with 6 units and a grid-forming substation with 8 units connected to a power grid of different strengths as an example, the system topology and control framework are as follows. Figure 6 As shown. During the sample set generation stage, electromagnetic transient simulations were performed on MATLAB software under different operating conditions and target frequencies. For stability testing, four cases were set up to test the integration of grid-type power stations into strong / weak networks and grid-type power stations into weak / strong networks, respectively, and Gaussian white noise was added to the steady-state information.

[0071] Stability assessment results were tested for different grid strengths and different types of power plants, and compared with theoretical values ​​and traditional methods. Specifically: 1) Theoretical truth value: The actual stable evaluation value is obtained based on theoretical analysis; 2) Conventional deep learning methods do not consider label noise distribution and input feature noise; 3) The method proposed in this invention takes into account the label noise distribution and input characteristic noise, and undergoes three-stage anti-dual-noise training.

[0072] The results are as follows Figure 7 Figures (a), (b), (c), and (d) are shown in the diagram. (a) shows the stability assessment results of a grid-connected power station integrated into a strong network; (b) shows the stability assessment results of a grid-connected power station integrated into a weak network; (c) shows the stability assessment results of a network-structured power station integrated into a weak network; and (d) shows the stability assessment results of a network-structured power station integrated into a strong network. , These are two eigenvalues ​​of the back-ratio matrix. In the figure, the closer the eigenvalues ​​predicted by different methods are to the theoretical eigenvalues, the more accurate the method is. The figure mainly focuses on... Local methods are applied to the curved region.

[0073] Table 1 below shows a comparison of stability assessment results when grid-type power stations are integrated into a weak grid, and Table 2 below shows a comparison of stability assessment results when grid-type power stations are integrated into a strong grid: Table 1. Comparison of stability assessment results when grid-type substations are integrated into a weak grid.

[0074] Table 2 Comparison of Stability Assessment Results for Grid-Type Power Stations Integrated into Strong Grid

[0075] From Tables 1 and 2 and Figure 7 As can be seen, the method proposed in this invention can more accurately evaluate the stability, stability margin and potential oscillation frequency of power stations under different power grid intensities and different types of power stations compared with conventional deep learning methods.

[0076] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" are intended to illustrate the present invention and are not intended to limit the present invention.

[0077] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for training a new energy station sequence impedance identification model, characterized in that, include: Obtaining Datasets: Obtain a clean dataset and a noisy dataset. The clean dataset contains input samples and corresponding noise-free sample labels, while the noisy dataset contains input samples and corresponding noise sample labels. Each input sample contains steady-state information of the power station unit, state variables of the control loop, and harmonic frequencies. Each sample label contains impedance information of the power station unit. Model coarse tuning: Using the noisy dataset as the training set and the clean dataset as the validation set, the original identification model is trained using the training set until its impedance identification performance on the validation set is improved to a preset level. The original identification model is a neural network. Model fine-tuning: The coarse-tuned recognition model is jointly trained adversarially using the clean dataset and the noisy dataset. In the adversarial training loss function used in the fine-tuning stage, the weight of clean samples is greater than that of noisy samples, and noisy samples in different harmonic frequency bands have different weights. The lower the harmonic frequency band, the smaller the weight of the noisy samples. The harmonic frequencies are divided into low-frequency bands and high-frequency bands. The frequency of the low-frequency band does not exceed 300Hz, and the frequency of the high-frequency band is greater than 300Hz. The weight of noisy samples in the high-frequency band is greater than that of noisy samples in the low-frequency band. Model calibration: The finely tuned recognition model is adversarially trained using the clean dataset. Model output: Outputs the recognition model after training; In the model fine-tuning stage, the loss function is: f = 1 / 2 * (y - y') ； wherein, denotes the number of samples in a batch, denotes the index of the -th sample in the batch, denotes the weight assigned to the -th sample, denotes the mean squared error loss function, and denotes the i-th input sample and its corresponding sample label, denotes the adversarial perturbation applied to the sample , denotes the model prediction obtained by inputting the perturbed sample into the model, denotes the p-norm of , denotes finding a perturbation whose size does not exceed a pre-defined threshold but maximizes the loss function of the model.​ 2. The training method for the sequence impedance identification model of new energy power stations as described in claim 1, characterized in that, In each input sample, the steady-state information includes active power P, reactive power Q, voltage amplitude U and current amplitude I of the unit port, and the state variables of the control loop include d-axis voltage , q-axis voltage , d-axis current and q-axis current obtained by dq-axis transformation of the voltage U and current I of the unit port; and in each sample label, the impedance information includes the amplitudes and phase angles of the four complex elements in the admittance matrix of the unit port.

3. The training method for the sequence impedance identification model of new energy power stations as described in claim 1, characterized in that, Loss function during the coarse-tuning phase of the model for: ； In the formula, This represents the mean squared error loss function. This indicates the number of samples in a batch. Indicates the first in the batch Index of each sample, and This represents the i-th input sample and its corresponding sample label. Indicates will The model prediction value obtained after inputting the data into the model.

4. The training method for the sequence impedance identification model of new energy power stations as described in claim 1, characterized in that, Loss function during model calibration for: ； In the formula, This indicates the number of samples in a batch. Indicates the first in the batch Index of each sample, This represents the mean squared error loss function. and This represents the i-th input sample and its corresponding sample label. Indicates the application applied to the sample On the opposing perturbation, Indicates will Input the model and apply a perturbation The obtained model predictions; Indicates to Find the p-norm. Indicates the search for a perturbation Its size does not exceed a preset threshold. However, it can maximize the model's loss function.

5. A method for power system stability analysis, characterized in that, include: The steady-state information of the station unit, the state variables of the control loop, and the harmonic frequency of the new energy power station are used as input features. These are then input into the new energy power station sequence impedance identification model obtained by the training method described in any one of claims 1 to 4 to obtain the impedance information of the corresponding station unit. The impedance information of each unit in the power station is aggregated based on the real-time acquired grid strength and station topology to obtain the aggregated admittance matrix of the power station, and the back ratio matrix of the power system is constructed based on the aggregated admittance matrix. The stability of the system is analyzed based on the back ratio matrix.

6. The power system stability analysis method as described in claim 5, characterized in that, When constructing the back ratio matrix of a power system based on the aggregated admittance matrix, it is necessary to distinguish between grid-connected and grid-connected power stations. When the power station is a grid-connected power station, the corresponding back ratio matrix is: When the station is a network-type station, the corresponding back-to-back ratio matrix is: ; ； In the formula, s represents , represents an imaginary number, Indicates the target angular frequency. , These are the grid impedance matrix and admittance matrix, respectively. For the aggregate admittance of the station. for The inverse matrix.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images