Excitation inrush current identification model training method, relay protection method and related device

Abstract

The embodiment of the invention discloses a training method of an excitation surge current identification model, a relay protection method and a related device. The training method comprises the following steps: acquiring a plurality of pieces of real operation state characteristic data for reflecting whether excitation surge current exists in a transformer or not; obtaining a plurality of first operation state feature data by using the real operation state feature data and a preset Gaussian mixture model; using the first operation state feature data, the real operation state feature data and a preset generative adversarial network to obtain second operation state feature data; obtaining a target training sample based on the second and real operation state feature data; and utilizing the target training sample and a preset compression and excitation network model to carry out excitation surge current identification training to obtain a trained target compression and excitation network model for identifying whether the predicted operation state characteristic data is excitation surge current data or not. Therefore, the magnetizing inrush current identification accuracy is improved, and relay protection misoperation caused by the magnetizing inrush current can be avoided.

Description

Technical Field

[0001] This invention relates to the field of relay protection technology, and in particular to a training method for an excitation inrush current identification model, a relay protection method, and related devices. Background Technology

[0002] With the rapid development of power systems, relay protection, as a key technology for ensuring the safe and stable operation of power systems, has become increasingly important. The core task of a relay protection system is to quickly and accurately identify faults and take corresponding protective measures when a fault occurs in the power system, in order to prevent the fault from escalating or causing more serious consequences. However, in actual operation, various factors often pose challenges to the relay protection system, leading to malfunctions. Among these, the inrush current phenomenon generated when a transformer is closed under no-load conditions is one of the important factors causing misjudgments and malfunctions in the relay protection system, thus affecting the reliability of the entire power system.

[0003] Inrush current is a transient current phenomenon generated when a transformer is switched on under no-load conditions due to the core magnetic flux entering the saturation region. Because its waveform characteristics are similar to internal fault currents, traditional relay protection methods often struggle to accurately distinguish between inrush current and fault current, leading to malfunctions in protection devices. Therefore, effectively identifying inrush current and preventing malfunctions in relay protection systems has become a pressing problem in the field of relay protection. Summary of the Invention

[0004] The main objective of this invention is to provide a training method for an inrush current identification model, a relay protection method, and related devices, which can solve the problem in the prior art that it is difficult to accurately distinguish between inrush current and fault current, leading to malfunction of protection devices.

[0005] To achieve the above objectives, the first aspect of the present invention provides a training method for an excitation inrush current identification model, the training method comprising:

[0006] Acquire several real operating state characteristic data of the transformer; the operating state characteristic data is used to reflect whether the transformer has inrush current;

[0007] Using the actual operating state feature data and the preset Gaussian mixture model, preliminary data generation is performed to obtain several first operating state feature data.

[0008] The second running state feature data is obtained by using the first running state feature data, the real running state feature data, and the preset generative adversarial network to generate data again.

[0009] A target training sample is obtained based on the second operating state feature data and the real operating state feature data. The target training sample includes the correspondence between the target operating state feature data and the pre-determined real type label. The target operating state feature data includes the second operating state feature data and the real operating state feature data. The type label is used to reflect whether the operating state feature data is excitation inrush current feature data.

[0010] The target training samples and the preset compression and excitation network model are used to train the inrush current identification, and the trained target compression and excitation network model is obtained. The target compression and excitation network model is used to identify and predict the true type label of the operating state feature data.

[0011] To achieve the above objectives, a second aspect of the present invention provides a relay protection method, the relay protection method comprising:

[0012] Obtain the current operating status characteristic data of the transformer;

[0013] The current operating state feature data is input into the target compression and excitation network model to identify the inrush current and obtain the second prediction type label. The target compression and excitation network model is trained using the training method described in the first aspect.

[0014] The current operating status of the transformer is determined by identifying the current operating status characteristics based on the current operating status feature data and the preset relay protection judgment rules.

[0015] If the second prediction type label is inrush current or the current operating state is a non-fault state, then a blocking signal is output to the relay protection device corresponding to the transformer.

[0016] To achieve the above objectives, a third aspect of the present invention provides a training apparatus for an inrush current identification model, the training apparatus comprising:

[0017] Data acquisition module: used to acquire several real operating status characteristic data of the transformer; the operating status characteristic data is used to reflect whether the transformer has inrush current;

[0018] First generation module: used to generate preliminary data using the real running state feature data and the preset Gaussian mixture model, to obtain several first running state feature data;

[0019] The second generation module is used to generate data again using the first running state feature data, the real running state feature data, and the preset generative adversarial network to obtain the second running state feature data.

[0020] Sample determination module: used to obtain target training samples based on the second operating state feature data and the real operating state feature data. The target training samples include the correspondence between the target operating state feature data and the pre-determined real type labels. The target operating state feature data includes the second operating state feature data and the real operating state feature data. The type labels are used to reflect whether the operating state feature data is excitation inrush current feature data.

[0021] Model training module: used to perform excitation inrush current identification training using the target training samples and a preset compression and excitation network model, to obtain a trained target compression and excitation network model, which is used to identify and predict the true type label of the operating state feature data.

[0022] To achieve the above objectives, a fourth aspect of the present invention provides a relay protection device, the relay protection device comprising:

[0023] Data acquisition module: used to acquire the current operating status characteristic data of the transformer;

[0024] Inrush current identification module: used to input the current operating state feature data into the target compression and excitation network model to identify the excitation inrush current and obtain a second predicted type label. The target compression and excitation network model is trained using the training method described in the first aspect.

[0025] Status determination module: used to identify the status based on the current operating status feature data and preset relay protection judgment rules, and determine the current operating status of the transformer;

[0026] Anti-maloperation module: If the second prediction type label is inrush current or the current operating state is a non-fault state, it outputs a blocking signal to the relay protection device corresponding to the transformer.

[0027] To achieve the above objectives, a fifth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform steps of the training method as described in the first aspect or steps of the relay protection method as described in the second aspect.

[0028] To achieve the above objectives, a fourth aspect of the present invention provides a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform steps of the training method as described in the first aspect or steps of the relay protection method as described in the second aspect.

[0029] The embodiments of the present invention have the following beneficial effects:

[0030] This invention provides a training method for an inrush current identification model. The training method includes: acquiring several real operating state feature data of a transformer; the operating state feature data is used to reflect whether the transformer has inrush current; using the real operating state feature data and a preset Gaussian mixture model to perform preliminary data generation, obtaining several first operating state feature data; using the first operating state feature data, the real operating state feature data, and a preset generative adversarial network to perform further data generation, obtaining second operating state feature data; obtaining target training samples based on the second operating state feature data and the real operating state feature data, the target training samples including the correspondence between target operating state feature data and a pre-determined true type label, the target operating state feature data including the second operating state feature data and the real operating state feature data, the type label being used to reflect whether the operating state feature data is inrush current feature data; using the target training samples and a preset compression and excitation network model to perform inrush current identification training, obtaining a trained target compression and excitation network model, the target compression and excitation network model being used to identify and predict the true type label of the operating state feature data.

[0031] By introducing a Gaussian mixture model into the generative adversarial network (GAN) in the above manner, the quality of the data input to the GAN is improved, the convergence speed of the GAN is increased, and the gradient vanishing problem is alleviated. Furthermore, by using the data samples expanded by the GAN, the compression and excitation network model is trained for inrush current identification, which can improve the model's ability to capture inrush current characteristics and improve the accuracy of inrush current identification. In the relay protection process, it can avoid the relay protection device from malfunctioning due to inrush current. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] in:

[0034] Figure 1 This is a flowchart of a training method for an excitation inrush current identification model in an embodiment of the present invention;

[0035] Figure 2 This is a block diagram of a generative adversarial network in an embodiment of the present invention;

[0036] Figure 3This is a block diagram of the attention mechanism of a compression and activation network model in an embodiment of the present invention;

[0037] Figure 4 This is a graph showing the variation of the Wasserstein distance with parameter K in a generative adversarial network according to an embodiment of the present invention.

[0038] Figure 5 This is a flowchart of a relay protection method according to an embodiment of the present invention;

[0039] Figure 6 This is a graph showing the variation of the malfunction rate of a relay protection system with parameter n in an embodiment of the present invention;

[0040] Figure 7 This is a structural block diagram of a training device for an excitation inrush current identification model according to an embodiment of the present invention;

[0041] Figure 8 This is a structural block diagram of a relay protection device according to an embodiment of the present invention;

[0042] Figure 9 This is a structural block diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 , Figure 1 This is a flowchart illustrating a training method for an excitation inrush current identification model according to an embodiment of the present invention. This method can be applied to either a terminal or a server. The terminal can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet computer, or laptop computer. The server can be a standalone server or a server cluster composed of multiple servers. This embodiment uses a terminal application as an example for illustration. Figure 1 The method shown includes the following steps:

[0045] 101. Obtain several real operating state characteristic data of the transformer; the operating state characteristic data is used to reflect whether the transformer has inrush current;

[0046] It should be noted that, in order to reduce malfunctions of relay protection equipment under non-fault conditions, this application will identify non-fault conditions and notify the relay protection equipment to maintain lockout when a non-fault condition is identified, thereby reducing malfunctions. To achieve the identification of non-fault conditions, a large amount of power operation data needs to be collected to learn from the power performance under non-fault conditions and identify patterns. Specifically, this application identifies the non-fault condition of inrush current generated during no-load closing. The collected power operation data can be transformer operating state characteristic data, thereby obtaining several real operating state characteristic data of the transformer to prepare for subsequent inrush current identification. The operating state characteristic data reflects whether the transformer has inrush current, and the real operating state characteristic data can be historical operating state characteristic data of the transformer, including various operating state characteristics that can reflect the transformer's operating state.

[0047] It should be noted that transformers, as electromagnetic energy conversion devices, rely on their iron cores to convert voltage into current. Under normal operating conditions, the transformer's magnetic flux linkage remains in the linear region, and the flux linkage and current have a linear relationship, thus achieving efficient energy transmission. However, when a transformer is energized under no-load conditions, the instantaneous voltage surge causes the magnetic flux linkage to enter the saturation region, triggering a sharp increase in current; this phenomenon is called inrush current. Many factors influence inrush current, primarily including the core saturation characteristics, the closing phase angle, and circuit impedance. For single-phase transformers, theoretically, no inrush current should occur when the closing phase angle is 90°. However, in actual power systems, transformers are mostly three-phase structures. Due to the 120° phase difference between the three phases, regardless of the closing time, inrush current will inevitably occur when a three-phase transformer is energized under no-load conditions.

[0048] The inrush current generated by a three-phase transformer when it is energized under no-load conditions has the following typical characteristics: 1) Due to the 120° phase shift between the three phases, when the transformer is energized under no-load conditions, at least two phases will experience inrush current simultaneously, and the inrush currents of each phase also maintain a 120° phase difference relationship; 2) Under certain specific conditions, symmetrical inrush current phenomena may occur, characterized by a symmetrical distribution of the current waveform above and below the zero axis; 3) From the perspective of harmonic analysis, there are significant differences in the second harmonic components in the inrush current waveforms of phases A, B, and C; 4) Compared with a single-phase transformer, the waveform of the three-phase inrush current exhibits a smaller discontinuity angle, but still maintains obvious discontinuities and peak characteristics, while the positive peak value and the reverse peak value exhibit a 120° phase difference.

[0049] Based on the above analysis, to achieve high-precision inrush current identification, the following features are extracted in this application: 1) voltage and current data of phases A, B, and C of the three-phase transformer; 2) second harmonic components of phases A, B, and C of the three-phase transformer; 3) discontinuity angles of inrush current in phases A, B, and C of the three-phase transformer. In other words, the three-phase current, three-phase voltage, second harmonic components, and discontinuity angles of the transformer are used as operating status features for subsequent inrush current identification.

[0050] 102. Using the actual operating state feature data and the preset Gaussian mixture model, preliminary data generation is performed to obtain several first operating state feature data.

[0051] 103. Using the first running state feature data, the real running state feature data, and the preset generative adversarial network, data is generated again to obtain the second running state feature data;

[0052] It should be noted that current datasets for power systems are relatively small. Designing relay protection schemes based on traditional machine learning models struggles to accurately capture data features, leading to high malfunction rates in relay protection equipment. However, directly using deep learning models can result in overfitting. Therefore, the method proposed in this application utilizes Generative Adversarial Networks (GANs) to expand the scale of existing datasets. First, the original GAN ​​and a Wasserstein distance-based GAN are introduced, followed by an improved GAN.

[0053] Generative Adversarial Networks (GANs) are a type of generative model. GANs generate new data that resembles the distribution of real-world data through adversarial training of two neural networks. GANs have achieved significant results in areas such as image generation, video synthesis, and natural language processing, becoming one of the hottest research topics in the field of deep learning in recent years.

[0054] The core idea of ​​GANs is to enable the generator to produce samples that are highly similar to the distribution of real data through adversarial training between the generator and the discriminator. The generator's goal is to generate data that is as realistic as possible, while the discriminator's goal is to distinguish between real data and generated data as accurately as possible. Through this adversarial training, the generator and discriminator are continuously optimized, ultimately enabling the generator to produce high-quality samples.

[0055] As described above, a GAN network has two important components: a generator and a discriminator. The generator (G) takes a random noise vector (z) as input (typically sampled from noise samples following a Gaussian or uniform distribution). The generator maps the random noise z to the data space using a neural network, generating samples G(z) that are similar in distribution to the real data. The generator's goal is to generate data that is as realistic as possible, making it difficult for the discriminator (D) to distinguish between generated and real data. The discriminator (D) takes real data x and generated data G(z) as input. The discriminator maps the input data to a scalar value representing the probability that the input data is real data using a neural network. The discriminator's goal is to distinguish between real and generated data as accurately as possible.

[0056] The training process of GAN can be viewed as a minimax game, and its objective function can be expressed as:

[0057]

[0058] In equation (1), P data (x) is the distribution of the real data, P0 z (z) is the distribution of random noise, D(x) represents the output of the discriminator, G(z) represents the output of the generator, and E(·) represents the mathematical expectation.

[0059] The goal of generator G is to minimize log(1-D(G(z))), that is, to generate samples that are as close as possible to the real data, making it difficult for discriminator D to distinguish between generated and real data. The generator optimizes its parameters to make the distribution P of its generated data G(z) conform to... G (z) Approximate the true data distribution P as closely as possible data (x). The goal of the discriminator D is to maximize log(D(x)) + log(1 - D(G(z))), that is, to distinguish between real data and generated data as accurately as possible. The discriminator optimizes its parameters so that for real data x, the discriminator output D(x) is close to 1; and for generated data G(z), the discriminator output D(G(z)) is close to 0.

[0060] The training process of GANs is a dynamic adversarial process. The generator and discriminator alternately optimize until Nash equilibrium is reached, meaning that the samples generated by the generator cannot be distinguished from the real data distribution by the discriminator. The training process can be summarized in two parts. The first part involves fixing the generator G and optimizing the discriminator D. In this stage, the discriminator D distinguishes between real and generated data by maximizing the objective function. Specifically, the discriminator D receives data from the real data distribution P. dataThe first part describes the process of classifying samples (x) and generated samples G(z) from the discriminator G, attempting to categorize them as either "real" or "generated." By maximizing the objective function, the discriminator D can better distinguish between real and generated data. The second part involves fixing the discriminator D and optimizing the generator G. In this stage, the generator G generates samples that are closer to real data by minimizing the objective function. The goal of the generator G is to generate samples that can "fool" the discriminator D, making it difficult for the discriminator D to distinguish between generated and real data. By minimizing the objective function, the generator G can generate more realistic samples.

[0061] As mentioned above, the optimization objective of GAN is to minimize the generated data distribution P. G (z) and the true data distribution P data The difference between (x) and (x). Traditional GANs measure the difference between two distributions by minimizing the Jensen-Shannon divergence (JS divergence). However, JS divergence can lead to the vanishing gradient problem during training. To address the training instability caused by JS divergence, the Wasserstein distance is introduced to measure the difference between two distributions. The Wasserstein distance is defined as:

[0062]

[0063] In equation (2), Π(p) data (x),p G (x) represents the set of joint distributions of the real data distribution and the generated data distribution. The Wasserstein distance can better reflect the difference between the two distributions and avoid the gradient vanishing problem.

[0064] Although WGAN alleviates the problems of training instability and gradient vanishing to some extent by introducing Wasserstein distance, the low quality of the initially generated samples due to the random noise input to the generator in the early stages of network training can still lead to gradient vanishing, making model training still quite difficult. To address these issues, this application proposes a network model based on the combination of Gaussian Mixture Model (GMM) and GAN, namely GMM-GAN (hereinafter referred to as iGAN).

[0065] Gaussian Mixture Models (GMMs) are generative models based on probability distributions, widely used in data clustering, probability density estimation, and generation tasks. GMMs model complex data distributions through linear combinations of multiple Gaussian distributions. Compared to models with a single Gaussian distribution, GMMs better capture the inherent structure of the data and are suitable for more complex tasks.

[0066] Gaussian Mixture Models (GMMs) assume that the data is a mixture of multiple Gaussian distributions, with each distribution corresponding to a cluster. The probability density function of a GMM can be expressed as:

[0067]

[0068] In equation (3), K is the number of Gaussian distributions, and π k It is the mixing coefficient of the k-th Gaussian distribution, and satisfies μ k and Σ k These are the mean and covariance matrix of the k-th Gaussian distribution, respectively, N(x|μ k ,Σ k Let π represent the probability density function of the k-th Gaussian distribution. The mixing coefficient π is... k This represents the weight of the k-th Gaussian distribution in the model, reflecting its contribution to the overall data distribution. Mean μ k The covariance matrix Σ represents the center position of the k-th Gaussian distribution. k The shape and orientation of the distribution are described.

[0069] The parameters of a Gaussian Mixture Model (GMM) are typically estimated using the Expectation-Maximization (EM) algorithm. The EM algorithm is an iterative optimization method that maximizes the log-likelihood function of the data by alternately performing the expectation step (E-step) and the maximization step (M-step).

[0070] In E-step, x is calculated for each data point. i The posterior probability γ of belonging to the k-th Gaussian distribution ik ,Right now:

[0071]

[0072] Equation (4), γ ik Represents data point x i The probability that it belongs to the k-th Gaussian distribution.

[0073] In the M-step, the posterior probability γ is calculated based on the E-step. ik Update model parameters π k μ k and Σ k :

[0074]

[0075] In equations (5) to (7), N represents the total number of sample points. By iterating through E-step and M-step, the parameters of GMM gradually converge, and finally the optimal model parameters are obtained.

[0076] The process of generating data using Gaussian Mixture Models (GMMs) involves two steps: 1) selecting a Gaussian distribution; and 2) sampling from the Gaussian distribution. In the first step, based on the mixing coefficient π... k A Gaussian distribution is randomly selected. Specifically, a random number q is first generated that is uniformly distributed between [0,1]. Assume q falls within the interval [0,1]. If the k-th Gaussian distribution is selected, then in the second step, the k-th Gaussian distribution N(x|μ) is chosen. k ,Σ k Data is generated in the covariance matrix Σ. Specifically, the data is first generated from the covariance matrix Σ. k Performing Cholesky decomposition yields the lower triangular matrix L. k To satisfy Among them, the superscript (·) T This represents the transpose of the matrix. Then, a matrix with μ is generated. k A standard normal distribution q ~ N(0, I) with the same dimension is used, where I represents the identity matrix. Finally, a linear transformation is performed to generate a Gaussian distribution N(x|μ) that conforms to the k-th Gaussian distribution. k ,Σ k Data:

[0077] x = μ k +L k q (8)

[0078] Repeat step 1) Select a Gaussian distribution; step 2) Sample from the Gaussian distribution to generate multiple data points until the required number of data points are generated.

[0079] The GMM (Gaussian Mixture Model) data generation process only requires sampling from multiple Gaussian distributions, resulting in low computational complexity. Furthermore, the generation process does not involve complex neural network training, making it particularly suitable for the relatively small datasets related to relay protection issues in this application. However, the data generated by GMM is limited by the Gaussian distribution assumption, making it difficult to characterize complex nonlinear data distributions. Therefore, this application uses the GMM-generated data as input noise data for the GAN (Generative Adversarial Network), improving the quality of its generator input data, accelerating the convergence speed of the GAN network, alleviating its training instability and gradient vanishing problems, thereby improving the quality of the final generated data. (See also...) Figure 2 , Figure 2 The diagram below shows a generative adversarial network (GAN) in an embodiment of the present invention. In iGAN, initial data, namely the first running state feature data, is first generated using the GMM model. Then, it is input into the generator of GAN. Finally, GAN generates high-quality generated data, namely the second running state feature data.

[0080] In one feasible implementation, step 102 includes steps A01 to A04:

[0081] A01. Using the actual operating state feature data and the Expectation-Maximization (EM) algorithm, the optimal model parameters of the Gaussian Mixture Model (GMM) are estimated to obtain the optimal model parameters of the GMM. The optimal model parameters include at least the covariance matrices Σ of the k Gaussian distributions corresponding to the GMM. k and mean μ k For details, please refer to the iterative process of formulas (3) to (7) above.

[0082] A02. Randomly select a target Gaussian distribution from the k Gaussian distributions of the Gaussian Mixture Model (GMM);

[0083] A03. Generate the first operating state feature data using the covariance matrix and mean of the target Gaussian distribution, and increment the sampling count by 1;

[0084] A04. If the number of samplings is less than the preset threshold, then return to the step of randomly selecting a target Gaussian distribution until the number of samplings is not less than the preset threshold, and obtain several first running state feature data. Specifically, refer to the two steps of the process of generating data using GMM: 1) selecting a Gaussian distribution; 2) sampling from the Gaussian distribution, and formula (8).

[0085] In one feasible implementation, step 103 includes steps B01 to B10:

[0086] B01. Randomly sample the first operating state feature data to obtain the third operating state feature data; randomly sample the actual operating state feature data to obtain the fourth operating state feature data;

[0087] B02. Input the third running state feature data into the generator of the generative adversarial network to generate data and obtain the fifth running state feature data.

[0088] B03. Input the fifth running state feature data and the fourth running state feature data into the discriminator of the generative adversarial network to train the distinction between real data and generated data, and obtain the first loss value of the discriminator.

[0089] B04. If the first loss value reflects that the discriminator has not converged, then the parameters of the discriminator are updated using the first loss value, and the process of randomly sampling the first running state feature data to obtain the third running state feature data and randomly sampling the real running state feature data to obtain the fourth running state feature data is returned.

[0090] B05. If the first loss value reflects the convergence of the discriminator, then the trained target discriminator is obtained; and the first running state feature data is randomly sampled to obtain the sixth running state feature data.

[0091] B06. Input the sixth running state feature data into the generator of the generative adversarial network to generate data, and obtain the seventh running state feature data.

[0092] B07. Input the seventh operating state feature data into the target discriminator to distinguish between real data and generated data, and obtain the probability that the target discriminator judges the seventh operating state feature data as real data;

[0093] B08. Determine the second loss value of the generator based on the judgment probability;

[0094] B09. If the second loss value reflects that the generator has not converged, then the parameters of the generator are updated using the second loss value, and the process of randomly sampling the first running state feature data to obtain the sixth running state feature data is returned.

[0095] B010. If the second loss value reflects the convergence of the generator, then the trained target generator and the seventh running state feature data are obtained; the second running state feature data includes the seventh running state feature data.

[0096] For example, using the data generated by GMM as the input noise data for GAN, the specific process of iGAN training is as follows:

[0097] 1. Step B01: Prepare input: Gaussian noise and real data

[0098] Generator input: from Gaussian distribution The random sampling noise z (e.g., sampling a 100-dimensional vector, each dimension of which follows a standard normal distribution) is included in the third running state feature data;

[0099] Real data: from the target true distribution p data (x) samples real samples x real (For example, real images, text vectors, etc.), the fourth running state feature data includes x real .

[0100] 2. Phase 1 includes steps B02 to B05: fix the generator G and optimize the discriminator D;

[0101] Generating fake samples: Input Gaussian noise into the generator to obtain x fake =G(z); The fifth operating state characteristic data includes x fake;

[0102] Discriminator learns to distinguish: D receives x real (real sample) and x fake (Generate samples) by maximizing the objective function max D V(D,G) learns the "features of real samples" and the "defects of current fake samples", that is, lets D(x) learn the "features of real samples" and the "defects of current fake samples". real )≈1 (determined as true), D(x) fake )≈0 (judged as false);

[0103] Update the parameters of D: Keep the parameters of G fixed, and only perform gradient ascent (or transform into gradient descent of the loss function) on the parameters of D.

[0104] Phase 2 includes steps B05 to B10: fixing the discriminator D and optimizing the generator G;

[0105] Generate new fake samples: Sample new noise z′ again from the Gaussian distribution to obtain x′fake=G(z′); where the sixth running state feature data includes the new noise z′; the seventh running state feature data includes x′fake.

[0106] Generator learning to deceive: The goal of G is to make x′fake as close as possible to the real data, such that D(x′fake)≈1 (classified as true by D), and the parameters are optimized by minimizing the objective function minGV(D,G);

[0107] Update the parameters of G: Keep the parameters of D fixed, and perform gradient descent only on the parameters of G.

[0108] Iterate until convergence, then repeat steps 2 and 3 until G generates a sample x′. fake Compared with the real sample x real The distributions are difficult to distinguish (D's output for both is close to 0.5), reaching a Nash equilibrium.

[0109] Finally, the data distribution P is generated. G (z) and the true data distribution P data The difference between (x) can be evaluated using either JS divergence or Wasserstein distance, and no particular choice is made here.

[0110] 104. A target training sample is obtained based on the second operating state feature data and the real operating state feature data. The target training sample includes the correspondence between the target operating state feature data and the pre-determined real type label. The target operating state feature data includes the second operating state feature data and the real operating state feature data. The type label is used to reflect whether the operating state feature data is excitation inrush current feature data.

[0111] 105. Using the target training samples and the preset compression and excitation network model, perform excitation inrush current identification training to obtain the trained target compression and excitation network model. The target compression and excitation network model is used to identify and predict the true type label of the operating state feature data.

[0112] Finally, the generated data and real data are used as training samples. The sample label is the true type label of the sample data, which includes either inrush current or non-inrush current. The compression and excitation network model is trained using these training samples. Specifically, target training samples are obtained based on the second operating state feature data and the real operating state feature data. The target training samples include the correspondence between target operating state feature data and pre-determined true type labels. The target operating state feature data includes the second operating state feature data and the real operating state feature data. The type label reflects whether the operating state feature data is inrush current feature data; the type label includes inrush current and non-inrush current. The true type label includes either inrush current or non-inrush current; the predicted type label includes either inrush current or non-inrush current.

[0113] Based on improved GANs, large-scale datasets can be generated. This section proposes a transformer inrush current identification method using a deep learning model. Among current deep learning models, convolutional neural networks have received considerable attention due to their high feature capture capabilities. SENet, in particular, has demonstrated superior performance in various tasks, such as image classification, object detection, and semantic segmentation, by introducing a channel attention mechanism. The core idea of ​​SENet is to adaptively recalibrate channel feature responses by explicitly modeling the dependencies between channels, thereby improving the network's representational capabilities. Based on its excellent performance and versatility, SENet is adopted as the basic network architecture in this application.

[0114] SENet achieves feature channel reweighting by embedding the Squeeze-and-Excitation Block (SE Block) into the base network. The core operation of the SE Block consists of two steps: Squeeze and Excitation. In the Squeeze stage, global average pooling compresses the spatial dimension of each channel into a single scalar, generating channel descriptors. In the Excitation stage, fully connected layers and non-linear activation functions (such as ReLU and Sigmoid) are used to learn the non-linear relationships between channels, generating channel weights. Ultimately, these weights are used to rescale the original features, thereby enhancing the feature responses of important channels and suppressing less important channels.

[0115] In the SENet network architecture, SE-ResNet-50 and SE-ResNet-101 are two commonly used network structures. SE-ResNet-50 contains 50 layers, including 48 convolutional layers and 2 fully connected layers; SE-ResNet-101 contains 101 layers, including 99 convolutional layers and 2 fully connected layers. As the network depth increases, the model's representational power is further enhanced, but the computational complexity also increases accordingly. Existing experiments show that SE-ResNet-50 achieves a better balance between the number of parameters and computational efficiency; therefore, SE-ResNet-50 is selected as the base model in this application.

[0116] In the SE-ResNet-50 network, SE modules are embedded in each residual block. Each residual block contains multiple convolutional layers with a kernel size of 3x3 and a stride of 1. Each SE module first generates channel descriptors through global average pooling, then learns channel weights through fully connected layers, and finally applies the weights to the original features using the sigmoid function. This design not only enhances the network's ability to focus on important features but also significantly improves the model's performance on complex tasks.

[0117] In the SENet network, the channel attention mechanism evaluates the importance of each feature channel, enhancing the weights of key features and suppressing the weights of unimportant features. However, existing SENets rely solely on global average pooling to extract feature information, resulting in insufficient feature diversity. To improve network performance, this application introduces global standard deviation pooling on top of global average pooling to enhance the diversity and comprehensiveness of feature channels, thereby more fully aggregating channel information. Subsequently, by utilizing fully connected layers to capture the dependencies between channels, an improved SENet, iSENet, is proposed. iSENet enhances the network's ability to capture global features from the channel dimension by combining global average pooling and global standard deviation pooling of input features, thereby improving the diversity of channel features.

[0118] In one feasible implementation, the compression and excitation network model includes at least a compression network and an excitation network, then step 105 includes steps C01 to C07:

[0119] C01. Input the target operating state feature data into the compression network, and obtain the global average pooling feature vector and the global standard deviation pooling feature vector of each feature channel included in the target operating state feature data through preset global average pooling and global standard deviation pooling.

[0120] C02. Input the global average pooling feature vector and the global standard deviation pooling feature vector into the activation network. Construct the dependency relationship between feature channels through the two fully connected layers of the activation network. Generate corresponding weights for the feature mapping of each feature channel to obtain the weight of each feature channel. The weight is used to reflect the importance of each feature channel.

[0121] C03. Using the weights and the target running state feature data, perform weighted processing to obtain the output feature tensor;

[0122] C04. Based on the output feature tensor, perform inrush current identification to obtain a first prediction type label;

[0123] C05. Determine the third loss value based on the first predicted type label and the true type label;

[0124] C06. If the third loss value reflects that the compression and excitation network model has not converged, then update the model parameters of the compression and excitation network model based on the third loss value, and return to the step of inputting the target running state feature data into the compression network and obtaining the global average pooling feature vector and global standard deviation pooling feature vector of each feature channel included in the target running state feature data through preset global average pooling and global standard deviation pooling.

[0125] C07. If the third loss value reflects the convergence of the compression and excitation network model, then the trained target compression and excitation network model is obtained.

[0126] The specific implementation method is as follows:

[0127] (1) First, execute step C01, which maps the original features X (i.e., input features) to the output feature space Z∈R through feature compression. C×H×W The method is as follows:

[0128]

[0129] Equation (9) represents global average pooling, and Equation (10) represents global standard deviation pooling. Z1 represents the global average pooling feature vector; Z2 represents the global standard deviation pooling feature vector; C, H, and W represent the number of feature channels, height, and width, respectively; x C ∈R H×W This represents the feature of input feature X on feature channel C, where input feature X includes target running state feature data; x C (i,j) represents the value of input feature X in the i-th row and j-th column of feature channel c; F sq (x C ) and F s ' q(x C All of these are feature compression operations. Among them, F... sq (x C ) and F s ' q (x C All of these are feature compression operations. As shown in equations (9) and (10), iSENet compresses each feature channel to a dimension of C×1×1 through global average pooling and global standard deviation pooling.

[0130] (2) Next, step C02 is executed. The computational complexity is reduced through activation operations, and the dependencies between channels are built using two fully connected layers, generating corresponding weights for the feature maps of each channel. The expression is as follows:

[0131] S = F ex (Z,w)

[0132] =σ[w1(δ([w0Z1))]+σ[w1(δ([w0Z2))] (11)

[0133] In equation (11), F ex (Z,w) represents the activation map, w0 and w1 represent the parameter vectors of the first and second layers in the fully connected layer, respectively, and δ and σ represent the ReLU and Sigmoid activation functions, respectively.

[0134] (3) Then, step C03 is executed to perform a weighted operation on S and the input feature X to obtain the output feature tensor Y = [y1, y2, ..., y]. C ]∈R C×H×W The method is as follows:

[0135] y C =F scale (x C ,s C )=x C ×s C (12)

[0136] In equation (12), F scale (x C ,s C ) represents a weighted mapping function, used to apply weights to the input features, s C ∈R H×W This represents the characteristics of S on channel C.

[0137] From the above process (see reference) Figure 3 , Figure 3As shown in the block diagram of the attention mechanism of a compression and excitation network model (iSENet) in an embodiment of the present invention, the number of channels of the input feature X is C. After processing by iSENet, the importance of each channel of the output feature Y is adjusted, so that the network can focus more on the channels with higher weights.

[0138] (4) Finally, perform steps C04 to C07 to identify and train the inrush current, predict the true label of the output feature Y, until the trained target compression and excitation network model is obtained.

[0139] This invention provides a training method for an inrush current identification model. By introducing a Gaussian mixture model into a generative adversarial network (GAN) as described above, the quality of the data input to the GAN is improved, the convergence speed of the GAN is increased, and the gradient vanishing problem is alleviated. Then, the data samples expanded by the GAN are used to train a compression and excitation network model for inrush current identification. This can improve the model's ability to capture inrush current features and enhance the accuracy of inrush current identification. In the relay protection process, it can prevent relay protection devices from malfunctioning due to inrush current.

[0140] Furthermore, a series of targeted experiments were conducted to verify the performance of the method proposed in this application, with the experimental setup as follows:

[0141] The hardware platform used in the experiment was configured as follows: Intel Xeon E7-3280 V5 processor (32 cores), 128GB DDR4 memory, and six NVIDIA GeForce GTX 4090 graphics cards. The software operating environment was Ubuntu 18.01, configured with a CUDA 13.2 parallel computing architecture, and PyTorch 2.5.1 was deployed. Regarding the dataset, a total of 400 fault samples were collected. Each sample was collected from three sampling cycles before and after the relay protection device malfunction, and the data was normalized. These 400 fault samples included 200 samples each for no-load closing and external fault clearing. The dataset was divided proportionally as follows: 80% for training and 20% for testing.

[0142] To evaluate the performance of the algorithm proposed in this application, the experimental results of the following nine algorithms were compared: 1) Support Vector Machine (SVM), 2) Random Forest (RF), 3) Recurrent Neural Network (RNN), 4) Long Short-Term Memory (LSTM), 5) SENet, 6) iSENet, 7) GAN+SENet, 8) GAN+iSENet, and 9) iGAN+iSENet.

[0143] In the experiments, the parameter settings for each comparison algorithm were as follows. In the SVM method, a Gaussian kernel function was used, with a regularization parameter of 0.8, a kernel bandwidth of 1.0, an upper limit of 100 iterations, and a maximum tolerable error of 0.001. In the RF method, the number of decision trees was set to 100, the maximum depth of a single tree was set to 5 layers, the minimum number of samples for node splits was 4, the minimum number of samples for leaf nodes was set to 2, the random number seed was fixed at 30, and a 10% dropout rate was set to prevent overfitting. In the RNN method, three hidden layers were used, each containing 256 neurons, the learning rate was set to 0.005, the Adam optimizer was used for parameter updates, the regularization coefficient was set to 0.3, the dropout rate was set to 10%, and ReLU was selected as the activation function. Considering that the LSTM method is an improved architecture of the RNN method, to maintain the fairness of the comparison experiments, all parameters were kept consistent with the RNN method except for setting the bidirectional mode to true. In SENet and its improved version iSENet, SE-ResNet-50 is used as the basic network architecture. Each residual module contains multiple 3×3 convolutional kernels (stride 1) and is coupled with 2×2 pooling layers (stride 2). The dropout rate is set to 10%, the optimizer uses the Adam algorithm, and the activation function is LeakyReLU. For the GAN+SENe, GAN+iSENet, and iGAN+iSENet methods that combine generative adversarial networks, in the GAN network, the generator network uses a Gaussian distribution as the prior distribution of input noise. In the iGAN network, the parameters of the GMM model are set as follows: maximum number of iterations is 100, convergence threshold is set to 0.001, and regularization factor is set to 1.0e-6. In the common part of GAN and iGAN, a fully connected network structure is used, the activation function is LeakyReLU, the optimizer also uses the Adam algorithm, and the learning rate is set to 0.005.

[0144] The experimental results are as follows: In the experiments of this application, the performance of the proposed method is evaluated by three indicators, including: 1) Wasserstein distance (WD), used to evaluate the performance of GAN and iGAN networks; 2) Recognition Accuracy Rate (RAR), used to measure the model's recognition accuracy of inrush current; 3) Relay protection system malfunction rate (R), used to evaluate the effectiveness of the relay protection scheme.

[0145] In improved generative adversarial networks (GANs), the choice of the number of Gaussian distributions, K, is crucial. Therefore, in the first experiment, we focused on investigating how the performance of the iGAN network varies with the parameter K. (See also...) Figure 4 , Figure 4 This is a graph showing the variation of the Wasserstein distance with parameter K in a generative adversarial network according to an embodiment of the present invention. Figure 4 The experimental results are presented in [the document]. Figure 4 As can be observed, with the gradual increase of parameter K, the performance of the iGAN network initially improves significantly (i.e., WD decreases rapidly), and then gradually declines. The performance of the iGAN network reaches its optimum when parameter K = 17. Therefore, in subsequent experiments, parameter K was fixed at 17. For comparison, Figure 3 The paper also presents the data distributions (WD) of the generated data and the real sample distributions for the original GAN. It can be seen that when parameter K=17, the WD of iGAN is significantly better than that of the original GAN, demonstrating the effectiveness of introducing the GMM model. Furthermore, due to the introduction of the GMM model, the network convergence speed is accelerated, and the training efficiency of iGAN is significantly higher than that of the original GAN. In this experiment, when parameter K=17, the training time of iGAN is 23.7 minutes, while the training time of the original GAN ​​network is 45.3 minutes.

[0146] Please see Figure 5 , Figure 5 This is a flowchart illustrating a relay protection method in an embodiment of the present invention. This method can be applied to either a terminal or a server. The terminal can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet computer, or laptop computer. The server can be a standalone server or a server cluster composed of multiple servers. This embodiment uses a terminal application as an example. Figure 5 The relay protection methods shown include:

[0147] 501. Obtain the current operating status characteristic data of the transformer;

[0148] The relay protection method adopts Figure 1The method of using the target compression and excitation network obtained from the training shown is as follows: the data dimension of the transformer current operating state feature data obtained in step 501 is the same as the dimension used when training the network, which will not be repeated here. Please refer to the above content. The difference is that the current data is obtained in real time so as to monitor the changes in the operating state in real time during the operation of the transformer and distinguish between faults and non-faults.

[0149] 502. Input the current operating state feature data into the target compression and excitation network model to identify the inrush current and obtain the second prediction type label;

[0150] Furthermore, the current operating status feature data can be input into the trained target compression and excitation network model to identify inrush current and obtain a second prediction type label. This determines whether inrush current exists and whether it is an unloaded closing operation. The target compression and excitation network model employs a method such as... Figure 1 The training method shown is used to obtain the results.

[0151] 503. Based on the current operating status characteristic data and the preset relay protection judgment rules, perform status identification to determine the current operating status of the transformer;

[0152] The relay protection judgment rule can be a traditional relay protection judgment rule, which can identify the existence of a fault and whether relay protection is required based on feature data. This is achieved by using current operating status feature data and preset relay protection judgment rules to determine the current operating status of the transformer, such as a fault state or a non-fault state. A fault state requires relay protection, while a non-fault state does not.

[0153] 504. If the second prediction type label is inrush current or the current operating state is a non-fault state, then output a blocking signal to the relay protection device corresponding to the transformer.

[0154] Based on iSENet-based inrush current identification, this application proposes the following relay protection scheme to improve the reliability of relay protection. Specifically, based on the existing relay protection scheme, a module for continuous monitoring of transformer operating status via iSENet is added. When no-load closing characteristics are detected, a blocking command is sent to the protection device; conversely, if the iSENet network identifies a non-no-load state n times consecutively (the specific value is set based on experience or experimental results), the blocking is released. This scheme adopts a dual confirmation mechanism, performing a logical "OR" operation between the iSENet blocking judgment and the existing relay protection blocking signal. The protection device will only execute the action command when neither of them issues a blocking signal. This design significantly improves the reliability of the protection system and effectively reduces the risk of malfunction caused by transformer inrush current.

[0155] That is, if the second prediction type label is inrush current or the current operating state is a non-fault state, a blocking signal is output to the relay protection device corresponding to the transformer to disable the relay protection; if the second prediction type label is non-inrush current and the current operating state is a fault state, an action signal is output to the relay protection device corresponding to the transformer to enable the relay protection. Furthermore, the blocking can be released when the second prediction type label is non-inrush current n times consecutively.

[0156] The online inrush current identification process based on iSENet, as described above, is as follows: During each sampling period, data from each channel is collected through the protection element, and corresponding features are extracted. After each sampling, the system appends new data to the tail of the queue while removing historical data from the head of the queue. Subsequently, this preprocessed data is fed into the iSENet model for state identification. To meet the real-time requirements of relay protection during inrush current identification, a sliding window mechanism is adopted, meaning only one sampling point is updated each time. Therefore, the data processing volume remains constant and is unaffected by the monitoring duration. This method maintains a low computational load and a fast response speed.

[0157] In the second experiment, the influence of parameter n on the effectiveness of the relay protection scheme proposed in this application was investigated. In the experiment, a traditional differential protection scheme was selected as the basis, and the relay protection scheme was implemented by combining it with the iGAN+iSENet inrush current identification method proposed in this application. The experimental results are as follows: Figure 6 As shown, Figure 6 This diagram illustrates the variation of the false trip rate of a relay protection system with parameter n in an embodiment of the present invention. The experimental results show that the false trip rate of the relay protection system initially decreases rapidly with the increase of parameter n, and then tends to stabilize. When parameter n = 7, the system's false trip rate reaches a relatively low level. Therefore, to ensure the real-time performance of the system, parameter n = 7 is set in subsequent experiments.

[0158] In the third set of experiments, the performance and training efficiency of the aforementioned nine methods on the training and test sets were compared in detail, and the specific results are shown in Table 1. It should be noted that for algorithms incorporating GAN and iGAN modules, the original dataset size was expanded from 400 to 4000 samples (all test data were actual samples) to enhance model performance and generalization ability, which directly led to a longer training time. Therefore, in terms of training time, these methods are not fairly comparable to the others. As can be seen from Table 1, the performance of SVM, RF, RNN, LSTM, SENet, and iSENet on the test set is significantly lower than their performance on the training set. This indicates that when the dataset is small, these methods are prone to severe overfitting. Among the above methods, iSENet outperforms SENet, demonstrating the effectiveness of introducing global standard deviation pooling. As shown in Table 1, on the test set, the performance of GAN+SENet, GAN+iSENet, and iGAN+iSENet significantly outperforms the other six methods, demonstrating the effectiveness of expanding the training set using generative adversarial networks. Among the three methods, iGAN+iSENet exhibits the best performance and the highest training efficiency, highlighting the effectiveness of introducing the GMM model.

[0159] Table 1 shows the performance of different algorithms.

[0160]

[0161] The main innovations of this application are as follows: 1) A dataset augmentation method based on an improved generative adversarial network (GAN) is proposed. By introducing a Gaussian mixture model into the GAN, the quality of the initial generated data is improved, thereby increasing the convergence speed and performance of the GAN; 2) Based on SENet, global standard deviation pooling is introduced, resulting in an improved SENet (iSENet), which enhances the diversity of feature channels and the network's ability to capture global features; 3) A relay protection scheme based on iSENet recognition results is proposed. Through a dual confirmation mechanism, the malfunction rate of relay protection devices caused by inrush current is effectively reduced, improving system reliability. Simulation results show that the proposed method exhibits significant advantages in both inrush current recognition and relay protection performance, providing a new technical means for the safe and stable operation of power systems.

[0162] This invention provides a relay protection method, comprising: acquiring current operating state characteristic data of a transformer; inputting the current operating state characteristic data into a target compression and excitation network model for inrush current identification to obtain a second predicted type label, wherein the target compression and excitation network model is trained using the training method described in the first aspect; performing state identification based on the current operating state characteristic data and preset relay protection judgment rules to determine the current operating state of the transformer; and outputting a blocking signal to the relay protection device corresponding to the transformer if the second predicted type label is inrush current or the current operating state is a non-fault state. This method effectively reduces the malfunction rate of relay protection devices caused by inrush current through a dual confirmation mechanism, thereby improving system reliability.

[0163] Please see Figure 7 , Figure 7 This is a structural block diagram of a training device for an excitation inrush current identification model according to an embodiment of the present invention, as shown below. Figure 7 The training device shown includes:

[0164] Data acquisition module 701: used to acquire several real operating status characteristic data of the transformer; the operating status characteristic data is used to reflect whether the transformer has inrush current;

[0165] First generation module 702: used to generate preliminary data using the real running state feature data and the preset Gaussian mixture model to obtain several first running state feature data;

[0166] The second generation module 703 is used to generate data again using the first running state feature data, the real running state feature data and the preset generative adversarial network to obtain the second running state feature data.

[0167] Sample determination module 704: used to obtain target training samples based on the second operating state feature data and the real operating state feature data, wherein the target training samples include the correspondence between the target operating state feature data and the pre-determined real type labels, the target operating state feature data includes the second operating state feature data and the real operating state feature data, and the type labels are used to reflect whether the operating state feature data is excitation inrush current feature data;

[0168] Model training module 705: used to perform excitation inrush current identification training using the target training samples and a preset compression and excitation network model, to obtain a trained target compression and excitation network model, which is used to identify and predict the true type label of the operating state feature data.

[0169] It should be noted that, Figure 7The functions of each module in the device shown are as follows: Figure 1 The steps in the method shown are similar, and to avoid repetition, they will not be elaborated here. Please refer to the relevant documentation for details. Figure 1 The content of each step in the method shown.

[0170] This invention provides a training device for an inrush current identification model. By introducing a Gaussian mixture model into a generative adversarial network (GAN), the quality of the data input to the GAN is improved, the convergence speed of the GAN is increased, and the gradient vanishing problem is alleviated. Furthermore, the data samples expanded by the GAN are used to train a compression and excitation network model for inrush current identification. This improves the model's ability to capture inrush current characteristics and enhances the accuracy of inrush current identification. In relay protection processes, this can prevent relay protection devices from malfunctioning due to inrush current.

[0171] Please see Figure 8 , Figure 8 This is a structural block diagram of a relay protection device according to an embodiment of the present invention, such as... Figure 8 The relay protection device shown includes:

[0172] Data acquisition module 801: Used to acquire the current operating status characteristic data of the transformer;

[0173] Inrush current identification module 802: used to input the current operating state feature data into the target compression and excitation network model to identify the excitation inrush current and obtain a second predicted type label. The target compression and excitation network model is adopted as follows: Figure 1 The training method described above was used to obtain the results.

[0174] Status determination module 803: used to identify the current operating status of the transformer based on the current operating status feature data and preset relay protection judgment rules;

[0175] Anti-maloperation module 804: If the second prediction type label is inrush current or the current operating state is a non-fault state, it outputs a blocking signal to the relay protection device corresponding to the transformer.

[0176] It should be noted that, Figure 8 The functions of each module in the device shown are as follows: Figure 5 The steps in the method shown are similar, and to avoid repetition, they will not be elaborated here. Please refer to the relevant documentation for details. Figure 5 The content of each step in the method shown.

[0177] This invention provides a relay protection device, comprising: a data acquisition module for acquiring current operating status characteristic data of a transformer; an inrush current identification module for inputting the current operating status characteristic data into a target compression and excitation network model to identify inrush current and obtain a second predicted type label, wherein the target compression and excitation network model is trained using the training method described in the first aspect; a status judgment module for performing status identification based on the current operating status characteristic data and preset relay protection judgment rules to determine the current operating status of the transformer; and a maloperation prevention module for outputting a blocking signal to the relay protection device corresponding to the transformer if the second predicted type label is inrush current or the current operating status is a non-fault state. This device effectively reduces the maloperation rate of relay protection devices caused by inrush current through a dual confirmation mechanism, thereby improving system reliability.

[0178] Figure 9 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 9 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program, which, when executed by the processor, causes the processor to perform the aforementioned methods. The internal memory may also store a computer program, which, when executed by the processor, causes the processor to perform the aforementioned methods. Those skilled in the art will understand that… Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0179] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform... Figure 1 or Figure 5 The steps of the method shown.

[0180] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform... Figure 1 or Figure 5 The steps of the method shown.

[0181] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0182] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0183] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for training an inrush identification model, characterized in that, The training method comprises: obtaining a plurality of real operating state feature data of a transformer; the operating state feature data is used to reflect whether the transformer has a magnetizing inrush current; performing preliminary data generation by using the real operating state feature data and a preset Gaussian mixture model to obtain a plurality of first operating state feature data; performing re-data generation by using the first operating state feature data, the real operating state feature data and a preset generative adversarial network to obtain second operating state feature data; obtaining target training samples based on the second operating state feature data and the real operating state feature data, wherein the target training samples comprise a corresponding relationship between target operating state feature data and a pre-determined real type label, the target operating state feature data comprises the second operating state feature data and the real operating state feature data, and the type label is used to reflect whether the operating state feature data is magnetizing inrush current feature data; performing magnetizing inrush current identification training by using the target training samples and a preset compression and excitation network model to obtain a target compression and excitation network model after training, wherein the target compression and excitation network model is used to identify and predict the real type label of the operating state feature data.

2. The training method of claim 1, wherein, The preliminary data generation by using the real operating state feature data and a preset Gaussian mixture model to obtain a plurality of first operating state feature data comprises: estimating optimal model parameters of the Gaussian mixture model by using the real operating state feature data and an expectation maximization algorithm to obtain the optimal model parameters of the Gaussian mixture model, wherein the optimal model parameters at least comprise covariance matrices and means of k Gaussian distributions corresponding to the Gaussian mixture model; randomly selecting a target Gaussian distribution from the k Gaussian distributions of the Gaussian mixture model; generating first operating state feature data by using the covariance matrix and the mean of the target Gaussian distribution, and adding 1 to a sampling number; if the sampling number is less than a preset number threshold, returning to perform the step of randomly selecting a target Gaussian distribution until the sampling number is not less than the preset number threshold to obtain a plurality of first operating state feature data.

3. The training method of claim 1, wherein, The re-data generation by using the first operating state feature data, the real operating state feature data and a preset generative adversarial network to obtain second operating state feature data comprises: randomly sampling the first operating state feature data to obtain third operating state feature data, and randomly sampling the real operating state feature data to obtain fourth operating state feature data; inputting the third operating state feature data into a generator of the generative adversarial network to generate data and obtaining fifth operating state feature data; inputting the fifth operating state feature data and the fourth operating state feature data into a discriminator of the generative adversarial network to perform real data and generated data distinguishing training and obtaining a first loss value of the discriminator; If the first loss value reflects that the discriminator has not converged, parameters of the discriminator are updated using the first loss value, and the step of randomly sampling the first operating state feature data to obtain third operating state feature data is performed again; the step of randomly sampling the real operating state feature data to obtain fourth operating state feature data is performed again; If the first loss value reflects that the discriminator has converged, a target discriminator trained is obtained; and the step of randomly sampling the first operating state feature data to obtain sixth operating state feature data is performed again; The sixth operating state feature data is input into a generator of a generative adversarial network to generate data, and seventh operating state feature data is obtained; The seventh operating state feature data is input into the target discriminator to distinguish real data from generated data, and a judgment probability of the target discriminator for the seventh operating state feature data being real data is obtained; A second loss value of the generator is determined according to the judgment probability; If the second loss value reflects that the generator has not converged, parameters of the generator are updated using the second loss value, and the step of randomly sampling the first operating state feature data to obtain sixth operating state feature data is performed again; If the second loss value reflects that the generator has converged, a target generator trained is obtained, and the seventh operating state feature data is obtained; the second operating state feature data includes the seventh operating state feature data.

4. The training method of claim 1, wherein, The compression and excitation network model at least includes a compression network and an excitation network, and the step of performing excitation inrush current identification training using the target training sample and the preset compression and excitation network model to obtain a target compression and excitation network model trained includes: The target operating state feature data is input into the compression network, and a global average pooling and a global standard deviation pooling are performed to obtain a global average pooling feature vector and a global standard deviation pooling feature vector of each feature channel included in the target operating state feature data; The global average pooling feature vector and the global standard deviation pooling feature vector are input into the excitation network, a dependency relationship between feature channels is constructed through two fully connected layers of the excitation network, a corresponding weight is generated for feature mapping of each feature channel, a weight of each feature channel is obtained, and the weight is used to reflect importance of each feature channel; The weight and the target operating state feature data are used for weighted processing to obtain an output feature tensor; Excitation inrush current identification is performed based on the output feature tensor to obtain a first predicted type label; A third loss value is determined according to the first predicted type label and the real type label; If the third loss value reflects that the compression and excitation network model has not converged, model parameters of the compression and excitation network model are updated based on the third loss value, and the step of inputting the target operating state feature data into the compression network, obtaining global average pooling feature vectors of each feature channel included in the target operating state feature data and global standard deviation pooling feature vectors through preset global average pooling and global standard deviation pooling is performed again. If the third loss value reflects that the compression and excitation network model has converged, a target compression and excitation network model trained is obtained.

5. The training method of claim 4, wherein, The global average pooling and global standard deviation pooling include the following mathematical expression: In the formula, Z1 represents a global average pooling feature vector; Z2 represents a global standard deviation pooling feature vector; C, H, W respectively represent a feature channel number, a height and a width; x C ∈R H×W represents a feature of an input feature X on a feature channel C, the input feature X including target running state feature data; x C (i,j) denotes the value of input feature X on feature channel c at the i-th row and j-th column; F sq (x C ) and F' sq (x C ) are both feature compression operations.

6. A method of relaying protection, characterized by, The relay protection method includes: obtaining current operating state feature data of a transformer; inputting the current operating state feature data into a target compression and excitation network model to perform excitation inrush current identification and obtain a second prediction type label, the target compression and excitation network model being trained by using the training method according to any one of claims 1 to 5; performing state identification according to the current operating state feature data and a preset relay protection judgment rule to determine a current operating state of the transformer; if the second prediction type label is excitation inrush current or the current operating state is a non-fault state, outputting a blocking signal to a relay protection device corresponding to the transformer.

7. A device for training an inrush identification model, characterized by, The training device includes: a data acquisition module configured to obtain a plurality of real operating state feature data of a transformer, the operating state feature data being used to reflect whether the transformer has excitation inrush current; a first generation module configured to perform preliminary data generation by using the real operating state feature data and a preset Gaussian mixture model to obtain a plurality of first operating state feature data; a second generation module configured to perform secondary data generation by using the first operating state feature data, the real operating state feature data and a preset generative adversarial network to obtain second operating state feature data; a sample determination module configured to obtain target training samples based on the second operating state feature data and the real operating state feature data, the target training samples including a corresponding relationship between target operating state feature data and a pre-determined real type label, the target operating state feature data including the second operating state feature data and the real operating state feature data, and the type label being used to reflect whether the operating state feature data is excitation inrush current feature data; a model training module configured to perform excitation inrush current identification training by using the target training samples and a preset compression and excitation network model to obtain a target compression and excitation network model trained, the target compression and excitation network model being used to identify and predict a real type label of the operating state feature data.

8. A protective relay device, characterized by The relay protection device includes: a data acquisition module configured to obtain current operating state feature data of a transformer; The inrush identification module is configured to input the current operating state feature data into a target compression and excitation network model to identify a second predicted type label of the inrush current, the target compression and excitation network model being trained by the training method in any one of claims 1 to 5. The state judgment module is configured to identify the state according to the current operating state feature data and a preset relay protection judgment rule to determine the current operating state of the transformer. The anti-misoperation module is configured to output a blocking signal to a corresponding relay protection device of the transformer if the second predicted type label is the inrush current or the current operating state is a non-fault state.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to enable the processor to perform the steps of the training method in any one of claims 1 to 5 or the steps of the relay protection method in claim 6. 10.A computer device, comprising a memory and a processor, and characterized in that, The memory stores the computer program, and the computer program is executed by the processor to enable the processor to perform the steps of the method in any one of claims 1 to 5 or the steps of the relay protection method in claim 6.

Images