Resonant grounding power distribution network single-phase grounding fault line selection method and system

Abstract

The invention discloses a resonant grounding power distribution network single-phase grounding fault line selection method and system. The method comprises the following steps: when a single-phase earth fault occurs in the power distribution network, acquiring real-time zero-sequence current data of each feeder line; and inputting the real-time zero-sequence current data into the fault line selection model to judge a fault line and output a result. The training process of the fault line selection model comprises the following steps: acquiring a noisy source domain data set from an electromagnetic transient simulation platform and a target domain data set from a real power distribution network; constructing a noise reduction automatic encoder as a feature extractor, and performing unsupervised pre-training by using the source domain data set to minimize a reconstruction error; pre-trained feature extractor parameters are frozen, and a trainable fault classifier is connected to form a transfer learning model; and carrying out supervised training on the model by using the source domain and target domain data sets and taking the fault line selection accuracy as a target to obtain a final model. By implementing the method provided by the invention, the accuracy and generalization ability of the line selection model in a real noise environment can be effectively improved through transfer learning and noise reduction feature extraction.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection technology, and in particular to a method and system for locating single-phase grounding faults in resonant grounding distribution networks. Background Technology

[0002] Single-phase ground faults are the most common type of fault in power distribution systems, especially in AC power distribution grounding systems, where they account for more than 80% of all line faults. These faults not only directly cause insulation safety problems in the power system but can also trigger more serious multi-phase faults, causing greater damage. Therefore, timely and accurate identification and location of faulty lines are crucial for ensuring the safe and stable operation of the power grid.

[0003] However, due to their large scale, wide distribution, and proximity to users, power distribution networks operate in complex and variable environments, making high-impedance grounding faults unavoidable. These faults, characterized by high resistance, result in low fault currents, making them difficult for conventional protection devices to detect and thus hindering effective fault location. Traditional methods such as empirical mode decomposition, Hilbert-Huang transform, wavelet transform, and mathematical morphology can extract fault features to some extent, but each has limitations, such as mode aliasing or the need to experimentally determine the number of components. With the development of information technology, data-driven methods, particularly those based on machine learning and deep learning, have been introduced into power system fault protection, demonstrating the potential to solve these problems. Nevertheless, this approach also faces challenges: first, acquiring high-quality labeled data is difficult, especially in the complex environment of power distribution networks; second, existing models are sensitive to noise and perform poorly in noisy environments; and finally, discrepancies between simulation data and actual power grid data limit the practical application of the models.

[0004] Therefore, it is necessary to design a new method with strong noise suppression capabilities, which can effectively utilize simulation data and achieve efficient and accurate fault location with only a small amount of real data. This will help improve the reliability and safety of the distribution network and reduce the economic losses and social impact caused by faults. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for selecting single-phase grounding faults in resonant grounding distribution networks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network, comprising: When a single-phase ground fault occurs in the distribution network, real-time zero-sequence current data of each feeder is obtained; The real-time zero-sequence current data is input into the fault line selection model to identify the faulty line and obtain the identification result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets by combining a pre-trained noise reduction feature extractor with a fault classifier. Output the discrimination result.

[0007] The further technical solution is as follows: the training process of the fault selection model includes: The source domain dataset and the target domain dataset from the target real distribution network are obtained from a simplified electromagnetic transient simulation platform of the actual distribution network. The source domain dataset and the target domain dataset are respectively the zero-sequence current data of each feeder and the corresponding fault line label after a single-phase ground fault occurs in the distribution network. Construct a denoising autoencoder and use it as a denoising feature extractor; The denoising autoencoder is pre-trained in unsupervised using noisy zero-sequence current data from the source domain dataset. The training objective is to minimize the reconstruction error between the output signal and the clean simulation signal in order to obtain the parameters of the pre-trained denoising feature extractor. Freeze the parameters of the pre-trained denoising feature extractor, and connect the output of the denoising feature extractor to a trainable fault classifier to form a transfer learning line selection model. Using the source domain dataset and the target domain dataset, the transfer learning line selection model is trained in a supervised manner with the goal of achieving fault line selection accuracy. The parameters of the noise reduction feature extractor are kept frozen or only partially fine-tuned to obtain the fault line selection model.

[0008] The further technical solution is as follows: the electromagnetic transient simulation platform is mainstream power system electromagnetic transient simulation software; the target real distribution network includes the actual operating distribution network or the distribution network under real-world testing.

[0009] The further technical solution is as follows: the source domain dataset is constructed by injecting Gaussian white noise into the pure zero-sequence current data generated by simulation.

[0010] The further technical solution is as follows: the noise reduction autoencoder includes an encoder and a decoder, wherein the encoder includes a one-dimensional convolutional neural network for extracting the noise-reduced feature representation from the noisy zero-sequence current input; the decoder includes a one-dimensional deconvolutional neural network for reconstructing a clean signal based on the feature representation.

[0011] The further technical solution is as follows: the one-dimensional convolutional neural network includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the one-dimensional deconvolutional neural network includes a one-dimensional deconvolutional layer, an activation function layer, and a pooling layer.

[0012] The further technical solution is as follows: the reconstruction error is calculated using a mean square error function, wherein the expression of the mean square error function is: ; Where X = (x1, x2, ..., x...) n The input data is Y = (y1, y2, ..., y3). n ) represents the output data, n represents the data length, and W and b represent the weights and biases of the one-dimensional convolutional neural network in the encoder and the one-dimensional deconvolutional neural network in the decoder, respectively. opt and b opt For the optimal weights and biases of the one-dimensional convolutional neural network in the encoder, and The optimal weights and biases are given by the encoder for the one-dimensional convolutional neural network.

[0013] The further technical solution is as follows: the fault classifier includes a neural network with a Softmax output layer, and the output fault probability distribution is calculated by the Softmax function. , where m i P(c=i|x) is the output value of the classifier for the i-th line, and P(c=i|x) represents the probability that the fault occurs on the i-th line when x is input.

[0014] The further technical solution is as follows: Using the source domain dataset and the target domain dataset, with the fault selection accuracy as the objective, the transfer learning fault selection model is trained in a supervised manner, wherein the parameters of the denoising feature extractor are kept frozen or only partially fine-tuned to obtain the fault selection model, including: Using the source domain dataset and the target domain dataset, with the goal of improving fault selection accuracy, a sample-weighted strategy is adopted to assign higher weights to samples in the target domain dataset than to samples in the source domain dataset. The transfer learning fault selection model is then trained in a supervised manner, wherein the parameters of the noise reduction feature extractor are kept frozen or only partially fine-tuned to obtain the fault selection model.

[0015] This invention also provides a single-phase grounding fault location system for resonant grounding distribution networks, comprising: The acquisition unit is used to acquire real-time zero-sequence current data of each feeder when a single-phase ground fault occurs in the distribution network. The discrimination unit is used to input the real-time zero-sequence current data into the fault line selection model to discriminate the fault line and obtain the discrimination result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets through a pre-trained noise reduction feature extractor combined with a fault classifier. The output unit is used to output the discrimination result.

[0016] The advantages of this invention compared to existing technologies are as follows: This invention acquires real-time zero-sequence current data from each feeder and inputs it into a fault line selection model that combines a pre-trained denoising feature extractor and a fault classifier for fault line identification, thereby achieving efficient and accurate fault identification. This model utilizes a one-dimensional convolutional denoising autoencoder to inherently process noise interference, enhancing the extraction capability of weak fault features. Simultaneously, it employs a transfer learning strategy to fully utilize a large amount of simulation data for pre-training, requiring only a small amount of real data for fine-tuning to achieve excellent performance, significantly reducing the dependence on expensive real label data. This method not only possesses strong noise suppression capabilities but also effectively improves the reliability and safety of the distribution network, reducing economic losses and social impacts caused by faults.

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the single-phase grounding fault location method for a resonant grounding distribution network provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the equivalent topology of the arc suppression coil grounding system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the simulation topology of a distribution network with an arc suppression coil grounding provided in an embodiment of the present invention; Figure 4 This is a typical waveform diagram of a single-phase ground fault in a distribution network with an arc suppression coil grounding, provided by an embodiment of the present invention. Figure 5 A schematic diagram of the fault route selection model provided in an embodiment of the present invention; Figure 6 This is a diagram illustrating the noise reduction effect of simulated fault data provided in an embodiment of the present invention. Figure 7 This is a diagram illustrating the noise reduction effect of waveform fault data provided in an embodiment of the present invention. Figure 8 This is a diagram showing the line selection results of the waveform data provided in an embodiment of the present invention.

[0020] Figure 9 This is a schematic block diagram of a single-phase grounding fault location system for a resonant grounding distribution network provided in an embodiment of the present invention. Figure 10A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0021] 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, not all, of the embodiments of the present invention. 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.

[0022] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] Please see Figure 1 , Figure 1 This is a flowchart illustrating the single-phase grounding fault location method for resonant grounding distribution networks provided in this embodiment of the invention. This method is applied in a server. By constructing and pre-training a denoising autoencoder as a denoising feature extractor, and combining it with a fault classifier for supervised transfer learning training on simulation and a small amount of real-world datasets, strong noise suppression capabilities and efficient fault line identification are achieved. This method utilizes a source domain dataset containing Gaussian white noise for unsupervised pre-training to optimize denoising feature extraction. Subsequently, the entire model is fine-tuned by assigning higher weights to the target domain dataset, ensuring accurate identification of faulty lines even with limited actual operating data. This effectively improves the reliability and safety of the distribution network and reduces economic losses and social impacts caused by faults.

[0026] Figure 1 This is a flowchart illustrating the single-phase grounding fault location method for resonant grounding distribution networks provided in this embodiment of the invention. Figure 1 As shown, the method includes the following steps S110 to S130.

[0027] S110. When a single-phase ground fault occurs in the distribution network, acquire the real-time zero-sequence current data of each feeder.

[0028] In this embodiment, real-time zero-sequence current data refers to the zero-sequence current values ​​measured from the beginning of each feeder when a single-phase ground fault occurs in the distribution network. These data are important parameters reflecting the internal state of the power system, especially crucial for detecting and locating single-phase ground faults. Specifically: Real-time zero-sequence current data is typically acquired using current transformers (CTs) installed on each feeder and transmitted to a central processing unit via a digital substation automation system or dedicated protection device. For resonant grounding systems (such as arc suppression coil grounding systems), accurate measurement and analysis of the zero-sequence current component is particularly critical because the system is designed to limit single-phase ground fault current.

[0029] Zero-sequence current refers to the current component generated when the sum of the three-phase current vectors is not zero. It mainly occurs under asymmetrical fault conditions, such as single-phase ground faults. Real-time zero-sequence current data not only contains direct evidence of the fault occurrence, but may also be affected by factors such as system operating conditions, load changes, and external electromagnetic interference, resulting in noise mixed into the data.

[0030] The collected real-time zero-sequence current data is input into a pre-trained fault location model for processing. This model utilizes deep learning techniques, particularly combining a noise-reducing autoencoder with transfer learning algorithms, to effectively identify and filter out noise components in the data, thereby accurately determining which feeder has failed. This method is particularly suitable for high-noise environments and situations where field tag data is scarce, significantly improving the accuracy and reliability of fault location.

[0031] In summary, timely and accurate acquisition of real-time zero-sequence current data of each feeder is a fundamental step in achieving efficient fault detection and location when a single-phase ground fault occurs in the distribution network, and is of great significance for ensuring the safe and stable operation of the power system.

[0032] S120. Input the real-time zero-sequence current data into the fault line selection model to identify the faulty line and obtain the identification result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets through a pre-trained noise reduction feature extractor combined with a fault classifier.

[0033] In this embodiment, the discrimination result refers to the information output after processing real-time zero-sequence current data through a pre-trained fault selection model, indicating which feeder experienced a single-phase ground fault. Specifically, the discrimination result includes the following aspects: Fault line identification: This is the most direct and crucial part, identifying which feeder experienced a single-phase ground fault. This information is usually given in the form of feeder number or name, such as "Feeder 3 has a fault".

[0034] Fault Probability Distribution: In addition to explicitly identifying the feeder where the fault occurred, the fault location model also provides a probability distribution for all possible fault locations. This means that for each feeder, the model provides a numerical value indicating the likelihood that the feeder is the source of the fault. This helps operators understand the confidence level of their judgment and may be used for further risk assessment and decision-making.

[0035] Reliability metrics: To measure the reliability of the discrimination results, the model may output additional metrics such as confidence scores and classification accuracy. These metrics can help technicians quickly evaluate the quality of the model's output, especially important in high-noise environments or when the amount of available data is limited.

[0036] Timestamp and Serial Number: To facilitate subsequent analysis and recording, each judgment result is usually accompanied by a timestamp (i.e. the time when the fault occurred) and a unique serial number to facilitate the tracing and management of historical data.

[0037] Recommended measures: Based on the specific judgment results, the system may also provide a series of recommended operating procedures or emergency response measures, such as disconnecting specific switches or notifying the maintenance team, to ensure that the power grid can be quickly restored to normal operation while ensuring personnel safety.

[0038] In summary, by inputting the real-time zero-sequence current data of each feeder into a carefully designed and fully trained fault location model, rapid and accurate location of single-phase grounding faults can be achieved, thus providing strong support for the stable operation of the power system. This process not only relies on advanced deep learning algorithms but also requires the optimization of model parameters based on practical engineering experience to ensure that it maintains efficient and stable performance in complex and ever-changing real-world environments.

[0039] In one embodiment, the training process of the fault selection model described above includes steps S121 to S125.

[0040] S121. Obtain the source domain dataset from a simplified electromagnetic transient simulation platform of the actual distribution network and the target domain dataset from the target real distribution network. The source domain dataset and the target domain dataset respectively contain the zero-sequence current data of each feeder and the corresponding fault line label after a single-phase ground fault occurs in the distribution network.

[0041] The electromagnetic transient simulation platform is a mainstream power system electromagnetic transient simulation software; the target real distribution network includes the actual operating distribution network or the distribution network in a real-world test.

[0042] The source domain dataset is constructed by injecting Gaussian white noise into the pure zero-sequence current data generated by simulation, in order to simulate a real-world noise environment.

[0043] S122. Construct a noise reduction autoencoder and use it as a noise reduction feature extractor.

[0044] In this embodiment, the noise reduction autoencoder includes an encoder and a decoder. The encoder includes a one-dimensional convolutional neural network for extracting noise-reduced feature representations from a noisy zero-sequence current input. The decoder includes a one-dimensional deconvolutional neural network for reconstructing a clean signal based on the feature representations.

[0045] The one-dimensional convolutional neural network includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the one-dimensional deconvolutional neural network includes a one-dimensional deconvolutional layer, an activation function layer, and a pooling layer.

[0046] S123. The denoising autoencoder is pre-trained in unsupervised using the noisy zero-sequence current data in the source domain dataset. The training objective is to minimize the reconstruction error between the output signal and the clean simulation signal in order to obtain the parameters of the pre-trained denoising feature extractor.

[0047] The reconstruction error is calculated using the mean square error function, the expression of which is: ; Where X = (x1, x2, ..., x...) n The input data is Y = (y1, y2, ..., y3). n ) represents the output data, n represents the data length, and W and b represent the weights and biases of the one-dimensional convolutional neural network in the encoder and the one-dimensional deconvolutional neural network in the decoder, respectively. opt and b opt For the optimal weights and biases of the one-dimensional convolutional neural network in the encoder, and The optimal weights and biases are given by the encoder for the one-dimensional convolutional neural network.

[0048] S124. Freeze the parameters of the pre-trained denoising feature extractor, and connect the output of the denoising feature extractor to a trainable fault classifier to form a transfer learning line selection model.

[0049] The fault classifier includes a neural network with a Softmax output layer, and the output fault probability distribution is calculated by the Softmax function: , where m iP(c=i|x) is the output value of the classifier for the i-th line, and P(c=i|x) represents the probability that the fault occurs on the i-th line when x is input.

[0050] S125. Using the source domain dataset and the target domain dataset, with the fault selection accuracy as the objective, supervised training is performed on the transfer learning line selection model, wherein the parameters of the denoising feature extractor are kept frozen or only partially fine-tuned to obtain the fault selection model.

[0051] Specifically, using the source domain dataset and the target domain dataset, with the fault selection accuracy as the objective, a sample-weighted strategy is adopted to assign higher weights to samples in the target domain dataset than to samples in the source domain dataset, and the transfer learning selection model is trained in a supervised manner. The parameters of the noise reduction feature extractor are kept frozen or only partially fine-tuned to obtain the fault selection model.

[0052] The transfer learning-based line selection model is trained in a supervised manner using both source and target domain datasets. Specifically, samples from the target domain dataset are given higher weights to address the inconsistency in data distribution between the source and target domains. To improve fault line selection accuracy, model parameters are adjusted, while allowing fine-tuning of some parameters of the denoising feature extractor, thereby enhancing the model's adaptability to new data.

[0053] When training the transfer learning line selection model, all weight parameters of the noise reduction feature extractor are either completely frozen or fine-tuned, and only the parameters of the fault classifier are updated.

[0054] This process not only effectively solves the problem of low reliability in fault location using traditional methods in high-noise environments, but also significantly improves the model's learning efficiency and generalization ability on limited labeled data. By combining deep learning and transfer learning techniques, the method proposed in this embodiment can provide more reliable fault detection and location services in complex real-world application environments.

[0055] S130. Output the discrimination result.

[0056] The above discrimination results are output to the terminal for display.

[0057] In this embodiment, as Figure 2 As shown, in a single-supply radial topology, there are n feeders connected to the same AC bus. The neutral point of the system is grounded through an arc suppression coil LP. Each feeder also has some branch lines. 0i Let R be the magnitude of the zero-sequence current measured at the beginning of the i-th feeder line, where i ∈ [1, n]. At t=0s, a single-phase ground fault occurs on feeder n, with a transition resistance of R. f The current flowing through the fault point is i fFor the transient characteristic analysis after a single-phase ground fault, a third-order equivalent circuit model was established, which can obtain relatively accurate simulation results. However, the third-order equation is not easy to solve. In order to reduce the order, the single-phase ground fault was divided into low-resistance ground fault and high-resistance ground fault.

[0058] For details, please refer to [link / reference] Figure 2 The phase voltages in the figure are U a U b and U c The three-phase voltages are symmetrical, C a C b and C c Let I represent the equivalent capacitance to ground for each phase, L be the loss coil, and UN be the voltage to ground at neutral point N. After a ground fault occurs in phase A, phases B and C change from phase voltages to line voltages, and the resulting ground currents are IB, IC, and UN be the voltage to ground at neutral point N, respectively. bd and I cd Their combined vector is Id, and the fault current at the A-phase grounding point is I. L .

[0059] A simulation platform for electromagnetic transients was built, such as... Figure 3 The simulation model of the power distribution network with grounded coil shown in the figure has a total of six feeders, each with some branch lines, such as L11, L12, etc. Cable lines and overhead lines make up the feeders. The type and length of each feeder are as follows: Figure 2 As shown, in the simulation parameters for cable lines and overhead lines, the positive-sequence reactance, resistance, and capacitance of the cable lines are 0.255 mH / km, 0.27 Ω / km, and 0.339 μF / km, respectively. The zero-sequence reactance, resistance, and capacitance are 1.02 mH / km, 2.7 Ω / km, and 0.28 μF / km, respectively. For overhead lines, the positive-sequence reactance, resistance, and capacitance are 1.2 mH / km, 0.17 Ω / km, and 0.0097 μF / km, respectively. The zero-sequence reactance, resistance, and capacitance are 5.48 mH / km, 0.23 Ω / km, and 0.006 μF / km, respectively. The simulation system is grounded through an arc suppression coil LP, and the system is in an overcompensated state with a compensation rate of ρ%. Based on the total zero-sequence capacitance of the system, we can obtain... ; where: L P C is the equivalent reactance of the arc suppression coil. 0∑ Let ω0 be the total zero-sequence capacitance of all lines and ω0 be the power frequency. The equivalent resistance of the arc suppression coil is η% of its impedance, calculated as L. P The equivalent resistance is R P =ηω0L P%. In this report, ρ% = 110% and η% = 3%. The active power of feeders L1-L6 ranges from 0.5MW to 1.5MW. Because the transformer is delta-connected, the magnitude of active power has little impact on zero-sequence current and zero-sequence voltage.

[0060] After a single-phase ground fault occurs on feeder 2, Gaussian white noise is superimposed on the zero-sequence currents of each feeder to obtain a typical noisy fault zero-sequence current, as shown below. Figure 4 As shown. Figure 4 In certain fault conditions, the zero-sequence current of a healthy feeder may increase over time, exceeding the amplitude of the faulty feeder, and the phase is basically the same, making accurate fault location more difficult.

[0061] The above simulation steps not only yielded the voltage and current changes after a single-phase ground fault, but also allowed for the acquisition of fault line signals with varying levels of noise. In reality, cable lines with higher capacitive current levels exhibit lower noise amplitudes, while overhead lines with very low capacitive currents have higher noise amplitudes. The arc suppression coil compensation can lead to situations where the phase of the zero-sequence current in the faulty feeder and the normal feeder is essentially the same. Furthermore, in this case, the amplitude of the zero-sequence current in the faulty feeder is not necessarily greater than that in the normal feeder, significantly increasing the difficulty of fault location. To overcome the limitations of signal data processing in practical environments, a fault location model is developed, such as... Figure 5 As shown.

[0062] Figure 5 One-dimensional Convolutional Neural Network (1D CNN): 1D CNNs are widely used in one-dimensional signal feature extraction and signal processing. They can efficiently extract features from signals and are deep feedforward neural networks with characteristics such as local connectivity and weight sharing. They are widely used in pattern classification. In each layer of a 1D CNN, the forward propagation can be represented as... ; Where: x k Let z be the input of a certain layer in a 1D CNN, z be the output of that layer, and w be the output of that layer. k Let be the kernel of the 1D CNN, and b be the bias. The dimensions of the input and output of a 1D CNN can be represented as: ; where: L x and L zLet k be the input and output lengths of a one-dimensional vector, p be the padding size, d be the dilation size, and s be the stride. Pooling layers reduce the dimensionality of data samples by summing local features, thus improving the robustness of CNNs. Max pooling layers are used to reduce data dimensionality by calculating local maxima after each convolutional computation.

[0063] Figure 5 In Chinese, an autoencoder is a neural network model that learns latent features from input data using an unsupervised approach. This encoder network structure can be divided into an encoder network that learns the latent features of the input data and a decoder network that uses the learned features to reconstruct the data. For example... Figure 5 As shown, the autoencoder network learns the effective features of the original data x and generates latent variables z. The decoding network attempts to reconstruct the generated data using the latent variables z. The purpose of learning through this neural network is to reduce reconstruction error; the optimization objective of its network parameters is... ; The training process of an autoencoder essentially involves using the input of sample data as both input and output of the neural network. By minimizing the reconstruction error, it learns the effective latent features of the samples. This effectiveness allows autoencoders to be used for the pre-training of deep models; after pre-training, the autoencoder's output can be used as input for subsequent models. Encoding utilizes activation functions to nonlinearly map high-dimensional input data to a low-dimensional space, resulting in a new feature expression as shown in the formula: In the formula, X is the input layer vector; Y is the hidden layer vector; W is the weight; b is the bias; and s(x) is the activation function for encoding and decoding. Generally, the Sigmoid function is chosen, i.e., f(x) = 1 / (1+ex). Decoding uses the activation function to inversely and nonlinearly map the low-dimensional feature expression to a high-dimensional space, thus reconstructing the input data. The formula is: In the formula, Z is the output layer vector; W′ is the weight; and b′ is the bias. Based on the reconstruction error function and the penalty term, the corresponding loss function is obtained as follows: In the formula, θ is the transformation parameter; L(W,b,W′,b′) is the reconstruction error; β is the weight of the sparsity penalty factor; m is the total number of units in the hidden layer; KL is the relative entropy, and here the Kullback-Leibler (KL) divergence is chosen; ρ is the target sparsity parameter; and ρj is the average activation of the j-th unit in the hidden layer. By comparing the difference between the reconstructed data and the input data, the weights and biases of each layer of the network are adjusted.

[0064] Figure 5In programming, high-dimensional input data typically contains some information redundancy. The purpose of autoencoders is to achieve more efficient data representation. Besides minimizing reconstruction errors, effective data representation should also possess other properties, such as robustness to partial data destruction. A denoising autoencoder is an autoencoder that increases encoding robustness by introducing a certain amount of noise. The core idea is to add random noise to the input layer before training the constructed autoencoder. By mapping the features of the contaminated data back to the original uncontaminated data, the encoder can learn more robust and generalizable feature representations.

[0065] Let X = (x1, x2, ..., xn) be the original signal. By randomly setting X to zero or adding noise, we "contaminate" X and generate a noisy signal X′ = (x1′, x2′, ..., xn′). This process can be represented as follows: In the formula, N is a distribution determined by the original input layer and the type of random noise added to X. Then, the encoder network f maps X′ to the hidden layer. In the formula: z represents the effective features of the hidden layer; W represents the weight parameters mapped to the hidden layer network; and b represents the bias term mapped to the hidden layer network. After being reverse-mapped by the decoder, z ultimately constructs an output identical to the original signal. for. In the formula: W′ represents the weight parameters of the effective features inversely mapped to the output layer network; b′ represents the bias term inversely mapped to the network layer output. The objective result is to make the output value... To approximate the original signal X as closely as possible, a squared reconstruction error function is constructed, and this function is minimized to obtain the optimized network parameters such as weights and biases. 2; In the formula, the subscript opt ​​represents the optimal values ​​of the network weights and biases.

[0066] In currently popular supervised learning models, deep learning-based neural network models can generally achieve high line selection accuracy when the number of samples (m) in the training set is sufficiently large. However, limitations may arise in real-world environments. The primary issue is the lack of sufficient labeled training samples; obtaining fault-labeled training samples in real-world environments is relatively difficult and costly. Additionally, it's challenging to obtain training samples under various fault conditions, including different transition resistances, grounding media, initial fault phase angles, and single-phase grounding faults at different fault locations. Furthermore, there is the imbalance phenomenon during fault occurrence. In real-world environments, the fault probability varies for each line; some feeders may have a very large number of samples, while others may have a very small number. This sample imbalance will significantly impact the trained model.

[0067] Based on the topology and parameters of the actual distribution network, a simplified simulation model was established in simulation software. This model simulated single-phase grounding faults under various conditions, efficiently and easily obtaining a large number of labeled training samples, which constitute the source domain (DS). After appropriately adjusting the simulated distribution network parameters, the obtained single-phase grounding fault data samples constitute the target domain (DT). Most existing works assume that the two distributions are identical, i.e., P(DT) = P(DS). However, since the simulation model is simplified and the fault conditions in the actual distribution network are very complex, this assumption has some problems. This paper proposes a sample-based transfer learning method that assumes that samples in DT and DS have different importance when training the neural network. This method utilizes the different weights of samples in the dataset to update the importance of samples. This method addresses the problem of P(DT) ≠ P(DS), improving the model's performance.

[0068] Let x∈Rd represent the input of the neural network model, where the zero-sequence currents of each feeder collectively constitute x. Let c={0, 1, 2, ..., n} represent the data label, indicating the label information of a fault occurring on a certain feeder or bus. Therefore, the training sample set can be represented as D={(xi, ci)|1≤i≤s}, where xi∈x, ci∈c, and s is the total number of samples in the training sample set. The subsequently designed neural network model is equivalent to a multi-label classifier model, whose output is an n-dimensional vector, which can be represented as {m1, ..., mn}. The probability that the input x belongs to the i-th label information is: In a neural network model, the last layer is typically a softmax function. Training samples are iterated to minimize the difference between the probability of each output line and the input fault label. The loss function is: In the formula, f is the loss function, and basic loss functions include cross-entropy and mean squared error. A data-driven model is trained using a large number of labeled training samples. After training, the test samples are input into this data-driven model to obtain the line selection information for single-phase grounding faults.

[0069] Waveforms before and after noise reduction are as follows Figure 6As shown, an autoencoder-convolutional neural network (ANN) approach using deep learning can solve the signal noise reduction problem of single-phase grounding faults in power distribution networks. The autoencoder performs noise reduction processing on the collected power distribution network signals, and noise feature extraction is used for fault location. This overcomes the challenges of high signal noise, weak noise signals, and low signal-to-noise ratio (SNR) in real-world environments, thus improving the reliability of fault location. Figure 6 The denoising results obtained using the proposed denoising autoencoder are shown. Figure 6 (a) The variation of zero-sequence current in six feeders is depicted without any added noise. The current in each feeder is normalized, and the amplitude is limited to 0-1. Figure 6 (b) shows the waveform after adding noise at a rate of 0.08. It is clear from the figure that the transient characteristics of the fault are masked after adding noise, thus complicating the selection of the faulty line. Figure 6 (c) shows the denoising effect of the trained autoencoder introduced in this study. It is clear that the denoised waveform is closer to the original signal.

[0070] Actual waveform recording data and noise reduction effect of actual power distribution network as follows Figure 7 This study demonstrates the effectiveness of the proposed method in denoising and feature extraction for typical high-impedance faults. Visual comparisons clearly highlight the successful extraction of weak fault signals from noisy environments, providing reliable features for accurate fault identification. Side-by-side comparisons with traditional methods intuitively confirm the superior performance in effectively reducing noise while preserving essential fault features. Figure 8 The results of fault line selection using actual recorded waveform data are shown; even in high-noise real-world environments, the proposed method can still accurately select lines. Field verification results ultimately demonstrate the practical advantages of this method in real-world distribution network applications.

[0071] In summary, this embodiment provides a method for single-phase grounding fault location in resonant grounding distribution networks to address the low reliability of fault location in high-noise environments and under conditions of scarce field tag data. This embodiment aims to achieve enhanced extraction and accurate identification of weak fault features through an end-to-end deep learning model integrating signal denoising and transfer learning.

[0072] This embodiment provides an innovative method for single-phase grounding fault location in resonant grounding distribution networks, addressing the challenges of existing technologies. By integrating signal denoising and transfer learning into an end-to-end deep learning model, this invention aims to enhance the extraction and accurate identification of weak fault features. A one-dimensional convolutional denoising autoencoder is used as the core component for feature extraction. Through unsupervised pre-training, this encoder can recover and enhance weak fault transient features in noisy environments. Compared to traditional external filters, the "inside-out" denoising approach offers superior performance. A "pre-training-freeze-fine-tuning" transfer learning paradigm is proposed. First, the encoder is pre-trained using massive amounts of simulation data to retain general denoising and feature representation knowledge; then, in practical applications, by training only the classifier or fine-tuning only some layers, it can quickly and efficiently adapt to a small amount of target domain data. This method significantly reduces the reliance on expensive and scarce real fault label data.

[0073] Especially in high-resistance grounding faults and high-noise environments, the method in this embodiment can effectively extract fault features from noise-inundated data, thereby significantly improving line selection accuracy. Compared to traditional methods and ordinary deep learning models, the method in this embodiment demonstrates a superior performance improvement.

[0074] The method in this embodiment significantly reduces the need for real fault label data, utilizing low-cost simulation data to build a robust foundational model. Final deployment requires only a small amount of real data, making it highly valuable for engineering applications.

[0075] By deeply integrating a noise-reducing automatic encoder with transfer learning technology, the method in this embodiment effectively solves the reliability problems in fault location caused by weak high-resistance grounding fault signals, high noise interference, and scarce field tag data. This not only significantly improves the accuracy of fault location but also enhances its practicality, providing strong support for the reliable operation of complex power systems.

[0076] The aforementioned method for selecting fault lines in resonant grounded distribution networks with single-phase grounding faults acquires real-time zero-sequence current data from each feeder and inputs it into a fault selection model that combines a pre-trained denoising feature extractor and a fault classifier for fault line identification, thereby achieving efficient and accurate fault identification. This model utilizes a one-dimensional convolutional denoising autoencoder to inherently process noise interference, enhancing the extraction capability of weak fault features. Simultaneously, it employs a transfer learning strategy to fully utilize a large amount of simulation data for pre-training, requiring only a small amount of real data for fine-tuning to achieve excellent performance, significantly reducing reliance on expensive real label data. This method not only possesses strong noise suppression capabilities but also effectively improves the reliability and safety of the distribution network, reducing economic losses and social impacts caused by faults.

[0077] Figure 9 This is a schematic block diagram of a single-phase grounding fault location system 300 for a resonant grounding distribution network provided in an embodiment of the present invention. Figure 9 As shown, corresponding to the above-described method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network, the present invention also provides a system 300 for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network. This system 300 includes a unit for executing the above-described method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network, and the system can be configured in a server. Specifically, please refer to... Figure 9 The single-phase grounding fault selection system 300 for the resonant grounding distribution network includes an acquisition unit 301, a discrimination unit 302, and an output unit 303.

[0078] The acquisition unit 301 is used to acquire real-time zero-sequence current data of each feeder when a single-phase ground fault occurs in the distribution network; the discrimination unit 302 is used to input the real-time zero-sequence current data into the fault line selection model to perform fault line discrimination and obtain the discrimination result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets through a pre-trained noise reduction feature extractor combined with a fault classifier; the output unit 303 is used to output the discrimination result.

[0079] In one embodiment, the system further includes a model training unit; the model training unit includes: The data acquisition subunit is used to acquire source domain datasets from a simplified electromagnetic transient simulation platform of the actual distribution network and target domain datasets from the target real distribution network. The source domain datasets and target domain datasets respectively contain zero-sequence current data of each feeder and the corresponding fault line label after a single-phase ground fault occurs in the distribution network. The construction subunit is used to construct a noise reduction autoencoder, which serves as a noise reduction feature extractor. The unsupervised pre-training subunit is used to perform unsupervised pre-training of the noise reduction autoencoder using the noisy zero-sequence current data in the source domain dataset, with the training objective being to minimize... The reconstruction error between the output signal and the clean simulation signal is used to obtain the parameters of the pre-trained denoising feature extractor; a freezing subunit is used to freeze the parameters of the pre-trained denoising feature extractor, and the output of the denoising feature extractor is connected to a trainable fault classifier to form a transfer learning line selection model; a supervised training subunit is used to perform supervised training on the transfer learning line selection model with the source domain dataset and the target domain dataset, with the fault line selection accuracy as the objective, wherein the parameters of the denoising feature extractor are kept frozen or only partially fine-tuned to obtain the fault line selection model.

[0080] In one embodiment, the supervised training subunit is used to supervise the transfer learning line selection model by using the source domain dataset and the target domain dataset, with the goal of fault line selection accuracy, and by employing a sample weighting strategy to assign higher weights to samples in the target domain dataset than to samples in the source domain dataset. The parameters of the denoising feature extractor are kept frozen or only partially fine-tuned to obtain the fault line selection model.

[0081] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned resonant grounding distribution network single-phase grounding fault selection system 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0082] The aforementioned single-phase ground fault location system 300 for resonant grounding distribution networks can be implemented as a computer program, which can be used in various ways, such as... Figure 10 It runs on the computer device shown.

[0083] Please see Figure 10 , Figure 10 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0084] See Figure 10 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0085] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for selecting a single-phase ground fault in a resonant grounding distribution network.

[0086] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0087] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for selecting a single-phase ground fault in a resonant grounding distribution network.

[0088] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 10The 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 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0089] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the single-phase grounding fault selection method for the resonant grounding distribution network.

[0090] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be 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 general-purpose processor may be a microprocessor or any conventional processor.

[0091] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0092] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the single-phase ground fault location method for the resonant grounding distribution network.

[0093] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0094] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0095] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0096] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0097] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network, characterized in that, include: When a single-phase ground fault occurs in the distribution network, real-time zero-sequence current data of each feeder is obtained; The real-time zero-sequence current data is input into the fault line selection model to identify the faulty line and obtain the identification result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets by combining a pre-trained noise reduction feature extractor with a fault classifier. Output the discrimination result.

2. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 1, characterized in that, The training process of the fault selection model includes: The source domain dataset and the target domain dataset from the target real distribution network are obtained from a simplified electromagnetic transient simulation platform of the actual distribution network. The source domain dataset and the target domain dataset are respectively the zero-sequence current data of each feeder and the corresponding fault line label after a single-phase ground fault occurs in the distribution network. Construct a denoising autoencoder and use it as a denoising feature extractor; The denoising autoencoder is pre-trained in unsupervised using noisy zero-sequence current data from the source domain dataset. The training objective is to minimize the reconstruction error between the output signal and the clean simulation signal in order to obtain the parameters of the pre-trained denoising feature extractor. Freeze the parameters of the pre-trained denoising feature extractor, and connect the output of the denoising feature extractor to a trainable fault classifier to form a transfer learning line selection model. Using the source domain dataset and the target domain dataset, the transfer learning line selection model is trained in a supervised manner with the goal of achieving fault line selection accuracy. The parameters of the noise reduction feature extractor are kept frozen or only partially fine-tuned to obtain the fault line selection model.

3. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The electromagnetic transient simulation platform is a mainstream power system electromagnetic transient simulation software; the target real distribution network includes the actual operating distribution network or the distribution network in a real-world test.

4. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The source domain dataset is constructed by injecting Gaussian white noise into the pure zero-sequence current data generated by simulation.

5. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The noise reduction autoencoder includes an encoder and a decoder. The encoder includes a one-dimensional convolutional neural network for extracting noise-reduced feature representations from a noisy zero-sequence current input. The decoder includes a one-dimensional deconvolutional neural network for reconstructing a clean signal based on the feature representations.

6. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 5, characterized in that, The one-dimensional convolutional neural network includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the one-dimensional deconvolutional neural network includes a one-dimensional deconvolutional layer, an activation function layer, and a pooling layer.

7. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The reconstruction error is calculated using the mean square error function, the expression of which is: ; Where X = (x1, x2, ..., x...) n The input data is Y = (y1, y2, ..., y3). n ) represents the output data, n represents the data length, and W and b represent the weights and biases of the one-dimensional convolutional neural network in the encoder and the one-dimensional deconvolutional neural network in the decoder, respectively. opt and b opt For the optimal weights and biases of the one-dimensional convolutional neural network in the encoder, and The optimal weights and biases are given by the encoder for the one-dimensional convolutional neural network.

8. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The fault classifier includes a neural network with a Softmax output layer, and the output fault probability distribution is calculated by the Softmax function: , where m i P(c=i|x) is the output value of the classifier for the i-th line, and P(c=i|x) represents the probability that the fault occurs on the i-th line when x is input.

9. The method for selecting the fault location of a single-phase grounding fault in a resonant grounding distribution network according to claim 2, characterized in that, The process of using the source domain dataset and the target domain dataset to perform supervised training on the transfer learning line selection model with the goal of achieving fault line selection accuracy, wherein the parameters of the denoising feature extractor are kept frozen or only partially fine-tuned to obtain the fault line selection model, includes: Using the source domain dataset and the target domain dataset, with the goal of improving fault selection accuracy, a sample-weighted strategy is adopted to assign higher weights to samples in the target domain dataset than to samples in the source domain dataset. The transfer learning fault selection model is then trained in a supervised manner, wherein the parameters of the noise reduction feature extractor are kept frozen or only partially fine-tuned to obtain the fault selection model.

10. A single-phase ground fault location system for resonant grounding distribution networks, characterized in that, include: The acquisition unit is used to acquire real-time zero-sequence current data of each feeder when a single-phase ground fault occurs in the distribution network. The discrimination unit is used to input the real-time zero-sequence current data into the fault line selection model to discriminate the fault line and obtain the discrimination result; wherein, the fault line selection model is a model obtained by supervised training on simulation and real datasets through a pre-trained noise reduction feature extractor combined with a fault classifier. The output unit is used to output the discrimination result.

Images