Single-phase ground fault line selection method and system for active power distribution network

Abstract

This invention belongs to the field of power transmission and distribution technology, and specifically relates to a method and system for selecting the line of a single-phase grounding fault in an active power distribution network. The method includes: collecting the zero-sequence current of each feeder within 20ms after a single-phase grounding fault occurs using simulation software; adding noise and performing equalization and normalization processing to convert the data into two-dimensional image samples; inputting these samples into a KEMT network model for training; and then inputting the zero-sequence current data of each feeder from the actual field into the trained model. The model then infers the feeder where the single-phase grounding fault occurred. This method can obtain a sufficient number of fault samples for model training without affecting power supply, and it also solves the problem of weak characteristic signals of the faulty line being masked by noise. It is particularly suitable for the selection and determination of single-phase high-resistivity grounding faults in power transmission and distribution.

Description

Technical Field

[0001] This invention belongs to the field of power transmission and distribution technology, and discloses a method, recording medium and system for selecting single-phase grounding faults in active power distribution networks. Background Technology

[0002] The power distribution network is a critical component of the power system, and its reliable operation is of great significance for energy security, economic development, and the improvement of power supply quality. Studies have shown that over 80% of faults in the distribution network are single-phase grounding faults. Among these, single-phase high-resistance grounding faults caused by factors such as broken distribution lines falling to the ground or trees touching the ground can easily lead to electric shock risks and forest fires, highlighting the urgent need to improve the reliability and accuracy of fault detection and fault location.

[0003] High-resistance grounding faults in power distribution lines present problems such as weak fault current and susceptibility to harmonic interference. Existing high-resistance grounding fault detection methods are mostly based on the identification of voltage and current distortion waveforms during the high-resistance grounding fault. Once the on-site high-resistance grounding fault conditions change or a large amount of noise is introduced, the voltage and current waveforms of the high-resistance grounding fault exhibit low signal-to-noise ratio characteristics, and the measured waveforms will change significantly, causing the reliability of traditional detection methods to drop rapidly.

[0004] Artificial intelligence (AI) is a new trend in fault diagnosis. Those skilled in the art attempt to input fault signals into neural networks and classify them through multiple rounds of training, hoping to obtain line selection results with high reliability and validity. However, when a single-phase high-resistance ground fault occurs, the zero-sequence current is very weak and mixed with noise signals. Designing the right neural network structure to accurately identify the faulty line is a problem that those skilled in the art need to consider. Furthermore, training requires a large number of samples; where can a sufficient number of fault samples be obtained without affecting power supply? Summary of the Invention

[0005] To address the above problems, this invention provides a method for selecting the line of a single-phase grounding fault in an active power distribution network, comprising the following steps: S1. Based on the actual parameters of the active distribution network, set the distribution network topology parameters, fault initial phase angle, fault node location and transition resistance value in the electromagnetic transient simulation model, and collect the zero-sequence current of each feeder within 20ms after a single-phase ground fault occurs. S2. Add Gaussian noise of not less than 5dB to the zero-sequence current, and use undersampling technology to make the zero-sequence current sample formed by the single-phase ground fault of the feeder itself and the zero-sequence current sample formed by the single-phase ground fault of the other feeder uniformly distributed after noise processing, and classify them according to the two labels of single-phase ground fault of this line and single-phase ground fault of other lines. S3. After normalizing and preprocessing all zero-sequence current samples, map them to the polar coordinate system, construct the Gram summation field matrix through trigonometric function operations, and then convert the matrix into a two-dimensional image. Collect all two-dimensional images and divide them into training set, validation set and test set as sample inputs for subsequent model training. S4. Each two-dimensional image is converted into an embedding vector and then input into a neural network model to extract features related to the classification task. The association rules between features and labels are obtained through model training. In this process, different hyperparameter models are trained using the training set to obtain model parameters. The optimal model hyperparameters are evaluated and selected using the validation set. The model is tested using the test set, and the association rules are solidified by the model that passes the test. S5. When the zero-sequence current of the busbar is detected to be non-zero during actual operation, the fault signal acquisition device is activated to collect the zero-sequence current data within 20ms after the fault of each feeder. The data is converted into a two-dimensional image and then into an embedded vector according to step S3. The data is then input into the model that has passed the test. The classification results of the zero-sequence current of each feeder are output. The feeder whose zero-sequence current data is assigned to the single-phase grounding tag of this line is the detected single-phase grounding feeder.

[0006] Preferably, the sampling frequency for a 50Hz AC power feeder is not less than 1000Hz.

[0007] Preferably, the two-dimensional image to embedding vector conversion is to divide the two-dimensional image into 16×16 pixel blocks, stitch together classification labels and superimpose position codes, generate embedding vectors through linear transformation, use a transformer network model, perform multi-layer feature extraction with a twelve-layer encoder, use an exponential moving average mechanism to plan the attention score smoothing function, and integrate historical and current batch data; and set a local bias term in the self-attention module to highlight the spatial neighborhood information weight.

[0008] Preferably, the feedforward sublayer in the transformer network model is changed from an MLP structure to a KAN structure.

[0009] Preferably, the initial parameters of the model are downloaded from the network, and the initial hyperparameters of the model are set as follows: learning rate = 5e-5, batch size = 64, embedding dimension = 768, number of attention heads = 12, and weight decay coefficient = 0.0001.

[0010] Another aspect of the present invention is to provide a non-transient readable recording medium for storing one or more programs containing multiple instructions, which, when executed, cause the processing circuit to perform the above-described method for selecting a single-phase ground fault in an active power distribution network.

[0011] Another aspect of the present invention provides a single-phase grounding fault location system for an active power distribution network, including a processing circuit and a memory electrically coupled thereto. The memory is configured to store at least one program, the program containing multiple instructions. The processing circuit runs the program and can execute the above-mentioned single-phase grounding fault location method for an active power distribution network.

[0012] Compared with existing technologies, the single-phase ground fault location method, recording medium, and system for active power distribution networks provided by this invention have the following advantages: To address the challenge of selecting high-impedance grounded feeders in low signal-to-noise ratio environments of active power distribution networks, innovative designs are implemented focusing on three main aspects: signal preprocessing, feature extraction, and improved attention mechanisms. I. By converting one-dimensional time-series signals into two-dimensional image features, the inherent anti-interference capability of the image spatial structure is relied upon, thereby improving the quality of the original signal in low signal-to-noise ratio environments. Second, at the feature extraction level, the feedforward sublayer in the transformer network model is changed from an MLP structure to a KAN structure with strong nonlinear mapping capabilities, thereby enhancing the discriminability of weak fault features and improving feature extraction. Third, at the attention mechanism level, a stable attention structure relying on a smoothing factor is planned, and a moving average is applied to the attention score to significantly reduce the model's sensitivity to noise, effectively solving the problem that traditional self-attention mechanisms are prone to fluctuations and loss of key features under low signal-to-noise ratio conditions.

[0013] These three core modules work together to form a complete KEMT (Kolmogorov–Arnold Enhanced Multi-layer Vision Transformer) model architecture, achieving accurate identification and reliable judgment of high-resistance grounding faults in active distribution networks.

[0014] In addition, the original sample data input to the KEMT model comes from the electromagnetic transient simulation model, which is obtained by adding noise to simulate the data. It is not the actual active power distribution network, so a sufficient number of fault samples can be obtained without affecting the power supply. Attached Figure Description

[0015] Figure 1 A comparison chart of the transient information content of zero-sequence current at different sampling frequencies; Where (a) corresponds to a sampling frequency of 100kHz and (b) corresponds to a sampling frequency of 1kHz; Figure 2 A schematic diagram of zero-sequence current transient information under different signal-to-noise ratios; Figure 3This is a flowchart of the single-phase high-resistance grounding fault feeder selection method in an embodiment of the present invention; Figure 4 This is a structural diagram of the KEMT model in an embodiment of the present invention; Figure 5 A schematic diagram of converting a zero-sequence current sample into a two-dimensional image; (a) corresponds to a faulty line with a single-phase grounding fault on this line, and (b) corresponds to a healthy line with a single-phase grounding fault on another line. Figure 6 This is a comparison diagram of the embodiments of the present invention with other intelligent route selection models; Figure 7 This is a graph showing the accuracy and loss of the base model trained under ideal conditions. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without innovative effort are within the scope of protection of the present invention.

[0017] Figure 3 This is a flowchart of the single-phase high-resistance grounding fault feeder selection method in an embodiment of the present invention, which, after refinement, includes the following steps: Step 101: Create an electromagnetic transient simulation model of a high-resistance grounding fault in the active distribution network of the target area, and obtain dynamic sequence data of zero-sequence current under typical operating conditions. The specific steps are: first, set the distribution network topology parameters, initial phase angle of the fault, fault node location, and transition resistance value in the electromagnetic transient simulation model according to the actual parameters of the distribution network in the target area; then, obtain the zero-sequence current transient data of each feeder 20ms after the fault occurs. (Refer to...) Figure 1 To compare the transient information content of zero-sequence current at different sampling frequencies, a typical sampling rate of not less than 1kHz was chosen. Assuming the feeder... i In the feeder j The occurrence of the first k The zero-sequence current transient sampling dataset of the secondary fault is denoted as I. i,j,k The fault labels corresponding to this dataset L i,j,k The calculation formula is as follows:

[0018] feeder i In the feeder j The occurrence of the first k The zero-sequence current transient sampling dataset of the secondary fault is denoted as I. i,j,k Fault labels corresponding to this dataset Li,j,k Merge the results and use them as input for subsequent steps.

[0019] Step 102: Figure 2 This is a schematic diagram of zero-sequence current transient information under different signal-to-noise ratios, based on the zero-sequence current transient sampling dataset I formed in step 101. i,j,k Gaussian noise of at least 5 dB was added to the data to accurately represent the actual operating conditions. The resulting noise-processed zero-sequence current transient dataset is as follows: Subsequently, based on undersampling techniques, the noise-processed zero-sequence current transient dataset was improved. The fault label is L i,j,k =0 and L i,j,k The number of samples with a value of 1 is evenly distributed, addressing the imbalance in the sample distribution of the dataset. This is achieved by processing the zero-sequence current transient dataset after equalization and noise reduction. The training set was divided into two parts according to a 7:2:1 ratio. Validation set and test set This serves as the data sample input for subsequent steps.

[0020] Step 103: Process the zero-sequence current transient dataset obtained in Step 102 after equalization and noise processing. Convert to a color 2D image set Img i,j,k First, the zero-sequence current transient dataset after equalization and noise processing is analyzed. Normalization preprocessing is performed to obtain Then, the normalized preprocessed zero-sequence current transient dataset is... Mapped to polar coordinates, an angle accumulation algorithm is used to construct a matrix, and finally, a zero-sequence current transient dataset after equalization and noise processing is generated. The corresponding two-dimensional image set Img i,j,k training set Validation set and test set This serves as the sample input for subsequent model training steps.

[0021] Step 104: Construct a KEMT model suitable for high-resistivity fault location in distribution networks. Its main architecture can be summarized as follows: Integrating image block embedding technology to combine the two-dimensional image set Img generated in step 103. i,j,kThe data is used as input to the KEMT model and divided into 16×16 pixel blocks. An embedding vector is generated through linear transformation. In the feedforward neural network stage, the VAT (an abbreviation of VIT with MLP replaced by KAN, where MLP is the abbreviation for Multilayer Perceptron Network and KAN is the abbreviation for Kolmogorov-Arnold Network) structure mapping is used to enhance the nonlinear expression capability. Then, multilayer feature extraction is carried out with the help of a twelve-layer encoder. During this process, an exponential moving average mechanism is used to plan the attention score smoothing function, and the historical and current batch data are fused and weighted. In addition, a local bias term is set in the self-attention module to highlight the importance of spatial neighborhood information.

[0022] Step 105: Configure the parameters and initial weights of the KEMT model suitable for high-resistivity fault location in distribution networks. Set the learning rate to 5e-5, batch size to 64, embedding dimension to 768, and attention heads to 12. Build the AdamW optimizer and set the weight decay coefficient to 0.0001. Download the model weight file vit_base_patch16_224_imagenet1k.pth from the official open-source website and load the weight file to assign initial weights to each neuron. Use the training set obtained in Step 103... and verification set Input the KEMT model constructed in step 104 (see...) Figure 4 The input end of the model integrates image block embedding technology. Based on the pre-trained weights, high-resistivity fault location in the distribution network is used as a downstream sub-task for further training. This updates the weight coefficients of each neuron in the model and the coefficients of newly trained neurons, thereby improving the accuracy of high-resistivity fault location in the distribution network based on the KEMT model. The trained KEMT model's weight re-file KEMT.pth is saved, and the test set obtained in step 103 is used... The model, loaded with the weight file KEMT.pth, is fed into the test set for detection. If the model can classify the zero-sequence current transient data samples obtained from the faulty line on the test set as... L i,j,k =1 (i=j), the zero-sequence current transient data samples obtained from the sound circuit are classified as follows: L i,j,k If i = 0 (i ≠ j), then save the weight file as best.pt as the weight file to be loaded for the subsequent high-resistivity fault selection task in the distribution network.

[0023] Step 106: When a non-zero-sequence current is detected on the busbar, the fault signal acquisition device is activated to acquire the zero-sequence current signal within 20ms after each feeder fault. This zero-sequence current signal is then converted into an image. The obtained fault image data is input into the distribution network high-resistance grounding fault selection method, which has been loaded with the model weight file best.pt obtained from steps 103-105. By outputting the distribution network high-resistance fault feeder classification results, the distribution network high-resistance fault selection is achieved. This method realizes an end-to-end distribution network high-resistance grounding fault selection method from signal acquisition to feeder fault judgment.

[0024] Step 103 also includes a transformation process for converting one-dimensional sequence data into an image, which involves standardizing the zero-sequence current signal, polar coordinate mapping, matrix construction, and generating a two-dimensional image that integrates spatial features. See [link to relevant documentation]. Figure 5 .

[0025] Adaptive time series window slicing:

[0026] Where S and ε represent the start and end indices of the time series slice, respectively, and D represents the image resolution. f s Indicates the sampling frequency; k 2 represents the sampling end time. k 1 represents the fault start time.

[0027] Mapping the zero-sequence current to polar coordinates, we get:

[0028] In the formula, The phase shift caused by grounding resistance was captured. r i The fault variation over time is marked, where N represents the length of the zero-sequence current sequence. I 0 This represents the time series composed of zero-sequence currents. I 0i Indicates time t i The corresponding zero-sequence current value.

[0029] The matrix capturing global time relationships is:

[0030] in, This represents the normalized zero-sequence current. express transpose, Represents the outer product operation, where each element in the matrix... GASF ijThe value directly determines the brightness or color of the corresponding pixel in the image, reflecting the first value in the zero-sequence current sequence. i The point and the first j The angular relationship between the points.

[0031] In step 104, the key framework of the KEMT model suitable for high-resistance fault location in distribution networks needs to integrate image block embedding technology, VAT structure, and stable attention mechanism to significantly optimize nonlinear fitting ability and enhance anti-interference performance. This allows for the development of an effective algorithm for high-resistance grounding fault location in distribution networks. The core expression of VAT is:

[0032] In the formula, It is the outer activation function of VAT. These are the inner basis functions of VAT, typically in B-spline form, where n represents the number of VAT layers and 2n+1 represents the number of basis functions per input dimension. In this study, VAT is applied to both the feedforward network and the final classification head.

[0033] Specifically, for input VAT first generates a set of inner basis functions. Through inner layer summation and outer layer activation function The operation yields the output. ,in and These are the dimensions of the input and output, respectively.

[0034] A smoothing factor and a buffer are introduced into the attention module to dynamically update and store the smoothed attention score. The attention score for the current batch is:

[0035] In the formula, Indicates the first h The first attention point in the t Attention score matrix for each time step; and These are the query matrix and key matrix for the current batch, respectively. The dimension of the attention head is used to scale the dot product results for stable training.

[0036] Subsequently, a global attention matrix was introduced. The attention information from historical and current batches is fused using the EMA method, with the following formula:

[0037] In the formula: This represents the global attention matrix of the h-th attention head, updated at the t-th time step. This represents the current attention score along the 0th dimension (here referring to the batch dimension). Take the average value. This is a smoothing factor that controls the weighting ratio between historical and current attention scores. This refers to the global attention matrix at the previous time step.

[0038] Next, it is necessary to determine the weight coefficients of each component in the constructed KEMT model of high-resistance grounding fault in the distribution network. These weight values ​​will become the key basis for subsequent feature weighting calculations, and their specific expression is as follows:

[0039] In the formula, This represents the normalized attention weight of the h-th attention head at time step t. This indicates that softmax normalization is performed along the last dimension (here referring to the dimension of the key).

[0040] Step 105 involves establishing a three-level training process to complete the high-impedance grounding fault location in the distribution network. First, during the KEMT model training and optimization phase, 100 rounds of iterative training are conducted on a massive dataset using the cross-entropy loss function, and a cosine annealing strategy is employed to dynamically adjust the learning rate. Then, in the online fault location module, the thoroughly validated KEMT algorithm is embedded into the distribution network high-impedance grounding fault location model. This model continuously receives and processes zero-sequence current signals, transforms them, and submits them to the neural network for accurate identification and judgment. Finally, in the distribution network high-impedance grounding fault location stage, standardized high-impedance grounding fault location results are output based on the generated fault line identification information.

[0041] Figure 6 This simulation scenario is based on a 10kV distribution network with distributed generation. The sampling frequency is set to six levels from 1kHz to 100kHz (increasing sequentially by step size; signal-to-noise ratio covering four levels: no noise, 20dB, 10dB, and 5dB; initial phase angle within the range of [0°, 180°]; and transition resistance values ​​are set within the range of [1000Ω, 8000Ω]. One-dimensional, two-dimensional, and traditional machine learning benchmark models are developed, and comparative experiments are conducted. The performance differences of each scheme are evaluated using quantitative results, and a final conclusion is drawn. Figure 6 The content presented highlights the noise immunity of this invention in the field of high-resistivity fault location in actual power distribution networks.

[0042] The performance of sequence-based and image-based AI models in fault location accuracy at low sampling frequencies was compared. The results show that as the signal-to-noise ratio decreases, the accuracy of sequence-based AI models in high-resistance fault location drops significantly. This is highly susceptible to impacts in actual power grid operation, leading to fault location accuracy failing to meet the requirements of distribution network protection accuracy. Converting one-dimensional signals into images can enhance noise immunity to some extent. Adaptive image transformation strengthens the spatial feature representation of the signal, making image-based AI models more robust to noise. Image-based signal processing enables the application of image-based AI algorithms in the field of high-resistance fault location in distribution networks, accelerating the development of AI algorithms in this area.

[0043] Figure 7 To enhance the benchmark performance evaluation results of the artificial intelligence model under noise-free conditions and to verify its fault feeder identification performance, this study selected a 1kHz sampling rate and actual distribution network simulation test. It can be seen that after 40 cycles, the accuracy and loss stabilized, and the accuracy reached 100% accurate classification.

[0044] When studying the fault diagnosis effect and performance of the basic model algorithm under various complex working conditions, the sampling accuracy is shown in Table 1: Table 1: Accuracy at different sampling frequencies and signal-to-noise ratios

[0045] It is known that when the sampling frequency is 1 kHz and the signal-to-noise ratio is only 5 dB, the accuracy of the basic model is only 81.44%, a significant decline, indicating that this algorithm is inadequate in dealing with more diverse operating conditions. The low sampling frequency severely weakens the expressive power of transient features. Based on this premise, the low signal-to-noise ratio further amplifies the impact of noise on signal processing, making it difficult for the self-attention module in the basic model to accurately extract key feature information. Under such a combination of adverse factors, there is still significant room for improvement in the performance of fault feeder identification, and it is urgent to improve its adaptability by means of model optimization or the introduction of new feature enhancement methods.

[0046] Ablation analysis was conducted to verify the roles and interactions of each module in the KEMT artificial intelligence model for high-resistivity fault location in distribution networks. The differences in performance between the stable attention mechanism and the VAT structure, used individually or in combination, were investigated in detail. Experimental data show that under low sampling frequency and low signal-to-noise ratio conditions, the KEMT model combining the stable attention mechanism and the VAT structure enhances the advantages of nonlinear feature capture while improving feature smoothing. This improves the robustness of the artificial intelligence model under noisy conditions and enhances the extraction of high-resistivity fault features in the distribution network, significantly improving the location accuracy of high-resistivity faults. This design significantly enhances the feasibility of applying the artificial intelligence model to high-resistivity fault location in distribution networks and can greatly promote the practical implementation of artificial intelligence algorithms.

[0047] From the above results, we can conclude that: This embodiment aims to improve the accuracy and reliability of feeder fault diagnosis, focusing on three key aspects: improved spatial feature extraction, enhanced multi-scale representation capabilities, and improved decision stability. Experiments have shown that: 1) Adaptive image transformation is used to process the zero-sequence current signal, and spatial filtering effectively reduces noise interference; 2) A VAT structure is established to expand the range of nonlinear mapping of the model, thereby achieving better results in feature learning; 3) An exponentially weighted moving average strategy is adopted to dynamically adjust the allocation mechanism of attention weights, which can greatly enhance the robustness of the system in complex operating environments.

[0048] Furthermore, this embodiment also provides a multi-stage fault location method for single-phase high-resistance grounding faults in active distribution networks, used to execute the above-described multi-stage fault location method for single-phase high-resistance grounding faults in active distribution networks.

[0049] The detailed information on each module or unit of the multi-stage fault location method for single-phase high-resistance grounding faults in active distribution networks has been explained in detail in the previous section on feeder fault location method for single-phase grounding faults, so it will not be repeated here.

[0050] Assembling the above methods and steps into a program and storing it on a hard disk or other non-transitory storage medium constitutes an embodiment of the present invention's "a non-transitory readable recording medium"; while electrically connecting the storage medium to a computer processor and performing data processing to complete the selection of a single-phase ground fault in an active power distribution network constitutes an embodiment of the present invention's "a single-phase ground fault selection system for an active power distribution network".

[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computers or available storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A single-phase ground fault line selection method for an active distribution network, characterized in that, Includes the following steps: S1. Based on the actual parameters of the active distribution network, set the distribution network topology parameters, fault initial phase angle, fault node location and transition resistance value in the electromagnetic transient simulation model, and collect the zero-sequence current of each feeder within 20ms after a single-phase ground fault occurs. S2. Add Gaussian noise of not less than 5dB to the zero-sequence current, and use undersampling technology to make the zero-sequence current sample formed by the single-phase ground fault of the feeder itself and the zero-sequence current sample formed by the single-phase ground fault of the other feeder uniformly distributed after noise processing, and classify them according to the two labels of single-phase ground fault of this line and single-phase ground fault of other lines. S3. After normalizing and preprocessing all zero-sequence current samples, map them to the polar coordinate system, construct the Gram summation field matrix through trigonometric function operations, and then convert the matrix into a two-dimensional image. Collect all two-dimensional images and divide them into training set, validation set and test set as sample inputs for subsequent model training. S4. Each two-dimensional image is converted into an embedding vector and then input into a neural network model to extract features related to the classification task. The association rules between features and labels are obtained through model training. In this process, different hyperparameter models are trained using the training set to obtain model parameters. The optimal model hyperparameters are evaluated and selected using the validation set. The model is tested using the test set, and the association rules are solidified by the model that passes the test. S5. When the zero-sequence current of the busbar is detected to be non-zero during actual operation, the fault signal acquisition device is activated to collect the zero-sequence current data within 20ms after the fault of each feeder. The data is converted into a two-dimensional image and then into an embedded vector according to step S3. The data is then input into the model that has passed the test. The classification results of the zero-sequence current of each feeder are output. The feeder whose zero-sequence current data is assigned to the single-phase grounding tag of this line is the detected single-phase grounding feeder.

2. The single-phase ground fault line selection method of an active power distribution network according to claim 1, characterized in that, The sampling frequency for a 50Hz AC power feeder should be no less than 1000Hz.

3. The single-phase ground fault line selection method of an active power distribution network according to claim 2, characterized in that, The process of converting a 2D image into an embedding vector involves dividing the 2D image into 16×16 pixel blocks, stitching together classification labels and superimposing positional codes, and generating embedding vectors through a linear transformation. The neural network model is a transformer network model, which uses a twelve-layer encoder to perform multi-layer feature extraction. During this process, an exponential moving average mechanism is used to plan the attention score smoothing function, and historical and current batch data are integrated. In the self-attention module, a local bias term is set to highlight the weight of spatial neighborhood information.

4. The single-phase ground fault line selection method of an active power distribution network according to claim 3, characterized in that, In the transformer network model, the feedforward sublayer is changed from an MLP structure to a KAN structure.

5. The single-phase earth fault line selection method of an active distribution network according to claim 4, characterized in that, The initial parameters of the model were downloaded from the network. The initial hyperparameters of the model were set as follows: learning rate = 5e-5, batch size = 64, embedding dimension = 768, number of attention heads = 12, and weight decay coefficient = 0.0001.

6. A non-transitory, readable recording medium storing one or more programs including a plurality of instructions, wherein the plurality of instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 5. When the instruction is executed, the processing circuit will perform a single-phase grounding fault location method for an active distribution network as described in any one of claims 1-5.

7. A single-phase ground fault line selection system for an active distribution network, comprising a processing circuit and a memory electrically coupled to the processing circuit, wherein the processing circuit is configured to: The memory is configured to store at least one program, the program containing multiple instructions, and the processing circuit runs the program to execute a single-phase grounding fault location method for an active power distribution network according to any one of claims 1-5.

Images