Intelligent detection method, system and equipment for high-resistance grounding fault of power distribution network and medium

Abstract

The invention relates to the technical field of power distribution network fault detection, and discloses a power distribution network high-resistance grounding fault intelligent detection method, system and device and a medium, and the method comprises the steps: collecting a power distribution network line transient signal which comprises a transient zero-sequence voltage signal and a transient zero-sequence current signal; preprocessing the transient signal and extracting a multi-dimensional feature set for representing high-resistance grounding fault characteristics, wherein the multi-dimensional feature set comprises a transient energy feature, a wavelet entropy feature and a harmonic component feature; and inputting the extracted multi-dimensional feature set into a pre-trained machine learning model, and simultaneously outputting a judgment result of the high-resistance grounding fault and an estimated value of the grounding resistance through the machine learning model. According to the method, the transient energy, the multi-scale wavelet entropy, the odd harmonic amplitude and the phase, which are derived from different analysis domains and have complementary physical meanings, are deeply fused, so that the problem that a single feature is weak in characterization capability and easy to fail under the conditions of high noise and high resistance is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault detection technology, specifically to an intelligent detection method, system, equipment, and medium for high-impedance ground faults in distribution networks, which is particularly suitable for high-impedance ground fault (HIF) scenarios in distribution networks where fault characteristics are weak and traditional detection methods are not sensitive enough. Background Technology

[0002] When a single-phase ground fault occurs in a distribution network (especially a neutral-point non-effectively grounded system), if the fault point is exposed to tree contact, contaminated insulation surfaces, or arcing, a transition resistance as high as several thousand to tens of thousands of kiloohms will be formed, which is called a high-resistance ground fault. This type of fault has the following significant characteristics:

[0003] 1. Weak fault current: Due to the large transition resistance, the fault steady-state current is often less than 10A, or even lower than the load current, which causes conventional devices based on zero-sequence current overcurrent or directional protection to fail to start or operate sensitively.

[0004] 2. Complex and unstable characteristics: The fault process is often accompanied by intermittent electric arcs, generating abundant transient harmonics and complex nonlinear characteristics. However, the signal amplitude is small and the duration is short, making it easily overwhelmed by system noise and load fluctuations.

[0005] 3. The hazards are hidden but serious: Although high-resistance grounding faults will not immediately cause large-scale power outages, their long-term existence may lead to further deterioration of the insulation at the fault point, developing into phase-to-phase short circuits; continuous grounding arcs may cause fires and cumulative damage to line insulation and connected equipment, seriously threatening the safe operation of the distribution network.

[0006] Currently, the main detection methods for high-resistance grounding faults include:

[0007] Steady-state characteristic method: such as detecting whether the zero-sequence voltage amplitude exceeds a set value, analyzing the change of the 5th harmonic, etc. This method is simple and easy to implement, but it has extremely low sensitivity to high-impedance faults (>10kΩ), is easily affected by system imbalance and load harmonic interference, and has a high false alarm rate.

[0008] Transient characteristic methods: such as utilizing transient zero-sequence current amplitude, first half-wave polarity, wavelet transform energy, etc. Compared with steady-state methods, they are effective for detecting faults with low to medium resistance values, but their ability to capture transient characteristics of faults with higher resistance values ​​(>30kΩ) is insufficient. They also heavily rely on threshold settings, have poor noise immunity, and have limited adaptability in real-world complex electromagnetic environments.

[0009] Artificial intelligence-assisted methods: In recent years, some studies have attempted to use machine learning for fault classification, but most of them use features from a single domain (such as only wavelet coefficients or harmonic amplitudes), which have limited feature representation capabilities and generally lack the ability to quantitatively estimate the fault transition resistance, thus failing to provide key basis for subsequent risk assessment and graded disposal.

[0010] In summary, existing technologies for high-resistance grounding fault detection generally suffer from three major challenges: difficulty in balancing sensitivity and reliability, lack of effective multi-dimensional feature characterization methods, and inability to achieve accurate online estimation of grounding resistance. These limitations make it difficult to meet the demands of modern power distribution networks for high power supply reliability and intelligent operation and maintenance. Therefore, there is an urgent need for a new intelligent detection method that can comprehensively utilize multi-dimensional information from transient signals, possess strong anti-interference capabilities and high identification accuracy, and quantitatively assess the severity of faults. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent detection method, system, device, and medium for high-resistance grounding faults in power distribution networks, thereby solving the technical problems of low sensitivity, high false alarm rate, and inability to quantitatively assess the severity of faults in traditional methods for detecting high-resistance grounding faults.

[0012] This invention is achieved through the following technical solution:

[0013] A smart detection method for high-resistance grounding faults in power distribution networks includes:

[0014] The transient signals of the distribution network lines are collected, including transient zero-sequence voltage signals and transient zero-sequence current signals;

[0015] The transient signal is preprocessed and a multi-dimensional feature set for characterizing the high-resistance grounding fault is extracted. The multi-dimensional feature set includes transient energy features, wavelet entropy features, and harmonic component features.

[0016] The extracted multi-dimensional feature set is input into a pre-trained machine learning model, which simultaneously outputs the judgment result of high-resistance grounding fault and the estimated value of grounding resistance.

[0017] As an optimization, the transient energy characteristics are obtained by calculating the integral of the product of transient zero-sequence voltage and transient zero-sequence current over the fault transient period.

[0018] As an optimization, the wavelet entropy features are extracted in the following way:

[0019] Wavelet transform is performed on the transient zero-sequence current signal to obtain wavelet coefficients at multiple scales;

[0020] Calculate the information entropy of wavelet coefficients at each scale;

[0021] Information entropy at various scales is combined to form wavelet entropy features.

[0022] As an optimization, the harmonic component features are extracted in the following way:

[0023] Perform a Fast Fourier Transform on the transient zero-sequence voltage signal or the transient zero-sequence current signal;

[0024] The amplitude and phase values ​​of the 3rd, 5th, and 7th harmonics are extracted from the transformation results and together constitute the harmonic component characteristics.

[0025] As an optimization, the specific process for preprocessing the transient signal is as follows:

[0026] Synchronization technology based on the Global Positioning System or a precision clock source is used to time-align the acquired transient zero-sequence voltage signal and transient zero-sequence current signal, and the synchronization error is controlled within a set time threshold range.

[0027] Wavelet transform threshold denoising method is used to denoise the synchronized signal in order to suppress power frequency harmonics, high frequency interference and random noise;

[0028] Statistical discrimination criteria are used to remove outlier data points whose amplitudes exceed the effective range in the denoised signal.

[0029] As an optimization, the machine learning model is one of support vector machine, deep neural network or gradient boosting tree; the machine learning model is trained through a high-resistance grounding fault sample library, which includes fault samples generated based on transient signals collected under different grounding resistance values ​​and different fault locations, and normal samples generated based on transient signals collected during normal operation of the distribution network.

[0030] As an optimization, after outputting the estimated value of the grounding resistance, if the estimated value is less than or equal to a first preset threshold, a fault isolation command is generated and executed; if the estimated value is greater than the first preset threshold, a fault warning command is generated.

[0031] This invention also discloses an intelligent detection system for high-resistance grounding faults in distribution networks, used to execute the aforementioned intelligent detection method for high-resistance grounding faults in distribution networks, comprising:

[0032] The acquisition module is used to acquire transient signals of the distribution network lines, including transient zero-sequence voltage signals and transient zero-sequence current signals;

[0033] The signal preprocessing module is used to preprocess the transient signal;

[0034] The feature extraction module is used to extract a multi-dimensional feature set from the preprocessed transient signal to characterize the high-resistance grounding fault. The multi-dimensional feature set includes transient energy features, wavelet entropy features, and harmonic component features.

[0035] The intelligent judgment module is used to input the extracted multi-dimensional features into a pre-trained machine learning model, and the machine learning model simultaneously outputs the judgment result of high-resistance grounding fault and the estimated value of grounding resistance.

[0036] The instruction generation module is used to generate fault isolation instructions or fault warning instructions based on the judgment result and the estimated value.

[0037] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned intelligent detection method for high-resistance grounding faults in a power distribution network.

[0038] The present invention also discloses a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned intelligent detection method for high-resistance grounding faults in power distribution networks.

[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0040] This invention creatively integrates three types of features originating from different analytical domains and possessing complementary physical meanings: time-domain global energy (transient energy), time-frequency domain local complexity (multi-scale wavelet entropy), and frequency-domain specific components (odd harmonic amplitude and phase). This fusion forms a high-dimensional, robust feature space. This feature set can comprehensively capture the weak transient essence of high-resistance grounding faults from multiple dimensions, including energy intensity, signal structure, and spectrum. It effectively solves the problem of weak characterization and easy failure of single features under high noise and high resistance conditions, laying a solid foundation for high-precision intelligent identification.

[0041] This invention trains a machine learning model (such as SVM) that integrates classification and regression functions. This invention can not only accurately determine whether a high-resistance grounding fault has occurred, but also simultaneously estimate the grounding resistance value of the fault point online, with the estimation error controlled within 10%. This is a major breakthrough in traditional fault detection methods, providing for the first time key parameters for quantitatively assessing the severity of a fault. It provides a scientific and direct basis for subsequent risk classification, fault location, and response strategy formulation, greatly enhancing the engineering application value of the detection results.

[0042] This invention is not limited to the detection process but integrates detection results with operation and maintenance control logic. Based on the estimated grounding resistance value, a risk threshold (e.g., 30kΩ) is set, and differentiated handling instructions are automatically generated: high-risk faults (low resistance) are immediately isolated to prevent the accident from escalating; low-risk faults (high resistance) receive timely warnings to guide precise inspections. This mechanism achieves precise and intelligent fault response, minimizing unnecessary power outages and improving power supply reliability while ensuring grid safety.

[0043] The preprocessing steps employed in this invention (high-precision synchronization, wavelet denoising, and outlier removal) effectively suppress on-site noise. The model trained on a massive and diverse sample library (covering different resistances, locations, and operating conditions) exhibits strong generalization capabilities. The entire method requires minimal hardware modifications and can be easily embedded into existing distribution terminals or master station systems. It is applicable to distribution networks with different voltage levels and line structures, and is particularly suitable for complex scenarios with stringent requirements for detection accuracy and real-time performance. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0045] Figure 1 This is a flowchart of the intelligent detection method for high-resistance grounding faults in distribution networks according to an embodiment of the present invention. The flowchart clearly shows the complete process from signal acquisition, signal preprocessing, feature extraction, model training, fault determination to instruction generation. The steps are logically connected, and the sequence and core links of the technology implementation are clearly defined.

[0046] Figure 2 This is a structural block diagram of the intelligent detection system for high-resistance grounding faults in power distribution networks according to an embodiment of the present invention. The diagram details the six major components of the system, including a signal acquisition module, a signal preprocessing module, a feature extraction module, a model training module, a fault determination module, and an instruction generation module. It clarifies the functional positioning and data transmission path of each module and intuitively reflects the overall architecture of the system.

[0047] Figure 3 This is a schematic diagram of transient signal feature extraction according to an embodiment of the present invention. The diagram shows the extraction process of transient signals from raw data to a 13-dimensional feature vector, including key steps such as transient energy calculation, wavelet entropy feature generation, and harmonic amplitude and phase extraction, clearly presenting the construction logic of multi-dimensional features. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0049] Example 1 discloses an intelligent detection method for high-resistance grounding faults in distribution networks, such as... Figure 1 , 3 As shown, it includes:

[0050] S1. Acquire transient signals of the distribution network lines, including transient zero-sequence voltage signals and transient zero-sequence current signals.

[0051] At the outgoing or neutral point of the 10kV distribution network line, a zero-sequence voltage transformer (TV) with an accuracy of 0.2 and a zero-sequence current transformer (TA) with an accuracy of 0.5 are installed. The outputs of the TV and TA are connected to a DAQ-2000 high-speed data acquisition card. To accurately capture the transient process at the moment of fault, the sampling frequency needs to be ≥5kHz; this invention preferably sets it to 10kHz, with a sampling window set to 0.1 seconds before and after the fault occurrence, totaling 0.2 seconds. A built-in GPS module ensures that the synchronization error of multi-channel acquisition is less than 0.5 milliseconds, thereby obtaining a high-quality transient zero-sequence voltage signal. With transient zero-sequence current signal This ensures the complete capture of the transient processes before and after the fault occurs, providing comprehensive data support for subsequent feature extraction.

[0052] S2. The transient signal is preprocessed and a multi-dimensional feature set for characterizing the high-resistance grounding fault is extracted. The multi-dimensional feature set includes transient energy features, wavelet entropy features, and harmonic component features.

[0053] The multi-dimensional feature set is the core of this invention. It creatively integrates three complementary features from the time domain, time-frequency domain, and frequency domain to comprehensively characterize the weak and complex transient characteristics of high-resistance grounding faults.

[0054] S2.1 Signal Preprocessing: The purpose of preprocessing is to improve signal quality and lay the foundation for feature extraction. Specifically, it includes:

[0055] Synchronization: A synchronization technique based on the Global Positioning System (GPS) or a precision clock source is employed to time-align the acquired transient zero-sequence voltage and current signals, controlling the synchronization error within a set time threshold range. This synchronization is already implemented via GPS in S1, ensuring... and The time is strictly aligned.

[0056] Noise Reduction: Wavelet transform threshold denoising is used to reduce noise in the synchronized signal to suppress power frequency harmonics, high-frequency interference, and random noise. More specifically, a wavelet threshold denoising method with a three-level decomposition of the signal using the Db5 wavelet basis function is employed. This method is the preferred solution for the characteristics of transient signal noise in distribution networks (including power frequency harmonics and high-frequency electromagnetic interference). By processing the detail coefficients with a soft threshold function, noise can be effectively suppressed while preserving the transient characteristics of faults to the greatest extent.

[0057] Invalid data removal: Statistical discrimination criteria are used to remove outlier data points in the denoised signal whose amplitudes exceed the valid range. Specifically, the Raida criterion is used (…). (Criteria). Calculate the mean of the denoised signal data. and standard deviation Automatically remove values ​​whose amplitude falls within the range To ensure data validity, identify any abnormal data points other than those mentioned above.

[0058] 2.2 Basic Multi-Dimensional Feature Set Extraction

[0059] 2.2.1 Extracting transient energy features, which are time-domain features.

[0060] The transient energy characteristics are obtained by integrating the product of transient zero-sequence voltage and transient zero-sequence current over the transient fault period. The formula is as follows:

[0061] (The integration interval is the transient period of the fault).

[0062] The energy integral during the fault transient period is calculated using this formula. This feature quantifies the transient power intensity of the fault injection system and is sensitive to faults with medium to low resistance values.

[0063] 2.2.2 Extracting wavelet entropy features, which are time-frequency domain features.

[0064] a) The preprocessed transient zero-sequence current signal Wavelet transform is performed. In this embodiment, the Db5 wavelet basis is preferred, and a 5-level decomposition is performed to obtain 5 detail coefficients at different scales (d1~d5), which correspond to different frequency bands.

[0065] b) Calculate the Shannon information entropy for each scale detail coefficient. The magnitude of the information entropy value reflects the complexity and uncertainty of the signal components in that frequency band. High-impedance faults can cause regular changes in the entropy value of a specific frequency band.

[0066] c) Arrange the information entropy values ​​of these five scales in order to form a 5-dimensional wavelet entropy feature vector [H1, H2, H3, H4, H5]. This feature reveals the changes in signal structure caused by faults from the perspective of time-frequency domain complexity and is robust to noise.

[0067] 2.2.3 Extracting harmonic component features, which are frequency domain features.

[0068] The harmonic component features are extracted in the following way:

[0069] a) The preprocessed transient zero-sequence voltage signal Perform a Fast Fourier Transform (FFT) to obtain its spectrum.

[0070] b) Accurately extract the amplitudes (A3, A5, A7) and phases (A3, A5, A7) of the 3rd, 5th, and 7th harmonics from the spectrum. , , These specific harmonics were chosen because when the distribution network is grounded with high resistance (especially through arc grounding), the asymmetry of the zero-sequence loop will significantly excite these odd harmonics, and their amplitude-phase relationship contains rich fault information.

[0071] c) These six quantities (3 amplitudes + 3 phases) together constitute the harmonic component characteristics. This characteristic provides fault fingerprint information from the perspective of specific components in the frequency domain.

[0072] In summary, this invention yields a 13-dimensional multi-dimensional feature set: 1-dimensional energy feature + 5-dimensional wavelet entropy feature + (3+3)-dimensional harmonic feature. These three types of features corroborate and complement each other from different physical dimensions, jointly constituting a highly recognizable fault feature space, completely solving the problem that traditional single features (such as using only zero-sequence voltage amplitude or a certain harmonic) have weak characterization ability and are easily interfered with under high impedance conditions.

[0073] In some embodiments, it further includes S2.3, adaptive feature selection and weight allocation.

[0074] This step aims to address the issue that some feature dimensions may fail or introduce interference when a fixed feature set faces different fault types (such as stable high resistance and intermittent arcing) and different background noise levels. To improve the robustness of the basic feature set under different fault types and noise environments, an adaptive optimization mechanism can be introduced, the process of which is as follows:

[0075] 2.3.1 Preliminary evaluation of feature effectiveness: 13-dimensional feature vectors extracted from the current sample Calculate its feature mean vector with that of historical normal samples. Standardized Euclidean distance For each feature dimension The greater the distance, the more significantly the feature deviates from the normal state in the current sample, and the stronger its potential indication of failure may be.

[0076] 2.3.2 Introduction of a Temporal Attention Mechanism: Considering the temporal evolution characteristics of the transient process of high-resistivity faults, this embodiment performs special processing on the transient energy features and wavelet entropy features (both of which implicitly contain temporal information). For the transient energy integration process containing multiple data points and the wavelet transform coefficients at each scale, a lightweight one-dimensional convolutional attention module is used to automatically learn and focus on the most critical abrupt change periods and frequency bands in the fault transient process, and output the corresponding attention weights. This is equivalent to intelligently weighting the contributions of different time periods / frequency bands.

[0077] 2.3.3 Feature Redundancy Analysis: Calculate the Spearman rank correlation coefficient matrix between the 13-dimensional features. If the absolute value of the correlation coefficient between two features exceeds a preset threshold (e.g., 0.85), they are considered to carry highly redundant information. The system will retain the feature with a higher overall correlation to the historical fault sample database (determined through canonical correlation analysis) and appropriately reduce the weight of the other feature to avoid excessive amplification of redundant information in the model.

[0078] 2.3.4 Dynamic Weight Vector Generation and Fusion: Combining the results of the above three steps (significance weights based on normal deviation, key weights based on attention mechanisms, and independence weights based on redundancy analysis), a 13-dimensional dynamic weight vector is generated through a learnable weighting function (such as using a very small fully connected neural network or a combination of fixed formulas). ,in Ultimately, the input to the machine learning model is not the original feature vector F, but a weighted feature vector. ( This represents the Hadamard product, which is element-wise multiplication.

[0079] In some embodiments, to delve deeper into the physical mechanisms behind harmonic data and improve the interpretability of features and the distinguishability of arc-type faults, the harmonic features in S2.2 can be physically enhanced and transformed:

[0080] 2.2.4 Establishing the Arc Model: A simplified Cassie-Mayr combined model suitable for grounding arcs in distribution networks is adopted. This model treats the arc as a dynamic impedance, and the change in its admittance g(t) follows the differential equation: ,in, It is static admittance. It is arc current. It is arc dissipation power. It is to maintain voltage. It is the time constant. This model can simulate the nonlinear and dynamic characteristics of an electric arc relatively well. Its dynamic characteristics are determined by the parameter set. describe.

[0081] 2.2.5 Harmonic-Physical Parameter Inversion: The amplitudes of the fundamental wave and the 3rd, 5th, and 7th harmonics extracted from S2.2 are... With phase This is considered as the output of the electric arc model under specific system conditions. The optimal electric arc physical parameters are then derived by solving the following nonlinear optimization problem. In other words, the acquired fundamental and harmonic amplitude and phase information is considered as the steady-state response output of the arc model under specific power supply conditions (system voltage, line parameters). An optimal set of arc model parameters is found by constructing and solving a nonlinear optimization problem (inverse problem). This minimizes the error between the harmonic amplitude phase calculated by substituting the parameter set into the model and the harmonic amplitude phase extracted by actual measurement.

[0082] Specifically, we need to find the minimum value of the following objective function:

[0083] ;

[0084] in, and These are the measured amplitude and phase of the h-th harmonic, respectively. and It is the arc parameters After substituting into the Cassie-Mayr arc model and the power distribution network system equations, the amplitude and phase of the h-th harmonic are calculated. , These are weighting coefficients set for the amplitude and phase error terms of each harmonic, used to adjust the relative importance of amplitude matching and phase matching in optimization.

[0085] This optimization problem can be solved using the Levenberg-Marquardt algorithm (gradient-based, fast convergence) or the particle swarm optimization algorithm (strong global search capability, suitable for nonlinear problems).

[0086] 2.2.6 Constructing Physical Enhancement Features: The electric arc physical parameters obtained from the optimization solution are... (Unit: V, reflecting the dielectric properties of the electric arc channel) (Unit: W, reflecting the energy dissipation rate of the arc column) and the total harmonic distortion (THD), which characterizes the nonlinear intensity, together constitute a new harmonic characteristic component. This characteristic directly characterizes the physical state of the arc at the fault point. That is:

[0087] ;

[0088] It can be directly calculated from the extracted harmonic amplitude, complementing the physical parameters.

[0089] S2.2.4 Forming the final enhanced feature set: The multi-dimensional feature set used in this embodiment consists of the following parts:

[0090] Transient energy characteristics: E, 1-dimensional.

[0091] Wavelet entropy features: [H1,H2,H3,H4,H5], 5-dimensional.

[0092] Physically enhanced harmonic characteristics: , 3D.

[0093] To preserve the critical phase relationship, additional phase differences of the 3rd and 5th harmonics can be added. , 1-dimensional.

[0094] Therefore, this embodiment constructs a 9-dimensional or 10-dimensional fault feature set that deeply integrates physical mechanisms. This feature set has a more concise dimension, and each feature has a clear physical or mathematical interpretation.

[0095] Through this embodiment, the present invention not only extracts features at the signal level, but also combines these features with the core physical process of high-resistance grounding faults (arc nonlinear characteristics). By establishing an arc mathematical model, the extracted harmonic data is inverted into arc parameters with clear physical meaning (such as sustaining voltage and dynamic resistance). This innovation transforms traditional "black box" data features into "white box" physical features, which not only enhances the ability of features to characterize the essence of the fault, but also significantly improves the robustness and generalization ability of the model in complex and variable arc scenarios, making the detection results physically interpretable.

[0096] In summary, this embodiment provides two optional, progressively related multi-dimensional feature set construction schemes:

[0097] Option 1 (Basic Option): A 13-dimensional basic multi-dimensional feature set is directly formed by the 1-dimensional transient energy feature, 5-dimensional wavelet entropy feature and 6-dimensional original harmonic component feature (3 amplitudes + 3 phases) extracted in the previous steps.

[0098] Option 2 (Physical Enhancement Scheme): To further explore the physical connotation of harmonic information, the previously extracted 6-dimensional original harmonic component features can be converted into 3-dimensional physically enhanced harmonic features through the physical inversion process based on the arc model described in Section S2.2.4 below. This 3D physical feature, together with the 1D transient energy feature and the 5D wavelet entropy feature, constitutes a 9-dimensional physically enhanced multi-dimensional feature set. To preserve key phase information, a 1D harmonic phase difference feature can also be added to form a 10-dimensional feature set.

[0099] In practical applications, depending on the detection accuracy, real-time requirements, and computing resources, the feature set constructed by Scheme 1 or Scheme 2 can be selected to be input into the subsequent machine learning model.

[0100] S3. Input the extracted multi-dimensional features into the pre-trained machine learning model, and output the judgment result of high-resistance grounding fault and the estimated value of grounding resistance through the machine learning model.

[0101] 3.1 Model Training and Sample Library Construction

[0102] The model's performance depends on a high-quality sample library. The sample library constructed in this embodiment covers a wide range of operating conditions:

[0103] Fault Sample: The grounding resistance was obtained through simulation and on-site waveform recording. , , , , Five gradients are used to cover transient signals at the beginning, middle and end of the line, with 50 samples generated for each case, for a total of 750 fault samples.

[0104] Normal sample: 200 sets of transient signals of normal operation of the power distribution network under different load and weather conditions were collected.

[0105] All samples are processed in step S2 to generate corresponding 13-dimensional feature vectors as input to the model.

[0106] In some embodiments, the machine learning model is one of support vector machines, deep neural networks, or gradient boosting trees; in this embodiment, a support vector machine is selected as the machine learning model, and a radial basis function is used as the kernel function. The key parameters of the SVM (penalty coefficient C, kernel parameters) are evaluated using a grid search algorithm. The optimal parameters were determined to be C=10 and γ=0.1 after optimization. The samples were then divided into training and test sets in a 7:3 ratio for training.

[0107] 3.2 Fault Diagnosis and Resistance Estimation

[0108] The feature vector extracted by S2 is input into the trained SVM model. This model performs the following tasks:

[0109] Binary classification: The result of determining whether the output is a high-impedance grounding fault or normal operation.

[0110] Regression estimation: The model outputs the estimated value of grounding resistance (Rest) synchronously through the built-in regression function.

[0111] Tests showed that the method in this embodiment achieved a fault identification accuracy of ≥90% on the test set and an average relative error of ≤10% in grounding resistance estimation, ensuring the accuracy and practicality of the detection results and providing a reliable basis for fault handling. This method is significantly superior to traditional steady-state detection methods (which typically have an accuracy of less than 70%) and single transient feature methods.

[0112] S4. Generate a disposal instruction based on the judgment result and the estimated resistance value.

[0113] To achieve intelligent processing, this embodiment sets the first preset threshold as follows: Develop a tiered strategy:

[0114] If a fault is identified and Rest ≤ 30kΩ, it indicates that the transition resistance at the fault point is low and the risk of development is high. The system immediately generates a fault isolation command and notifies the line circuit breaker to trip within 100 milliseconds via the IEC 61850 protocol.

[0115] If a fault is identified but Rest > 30kΩ, it indicates that the fault current is extremely weak and the risk is controllable in the short term. The system generates a fault warning command and pushes the faulty line, estimated resistance, and suspected section (roughly located based on impedance method) to the distribution automation master station and the mobile terminal of maintenance personnel to guide precise inspection.

[0116] This invention, for the first time, synergistically fuses three heterogeneous features—transient energy, multi-scale wavelet entropy, and specific odd-order harmonic amplitude and phase—to construct a 13-dimensional feature set capable of comprehensively and robustly characterizing high-resistance grounding faults. By combining this feature set with an SVM model, the detection accuracy of high-resistance grounding faults (especially >10kΩ) is increased from less than 70% using traditional methods to over 95% without increasing hardware costs. Furthermore, it achieves online, rapid grounding resistance estimation (error <10%) for the first time, providing crucial quantitative evidence for risk assessment and tiered response.

[0117] Meanwhile, the integrated detection-estimation-tiered handling process proposed in this invention solves the closed-loop problem from perception to execution, and can be directly integrated into the existing power distribution network automation system to improve the level of intelligent operation and maintenance.

[0118] This invention employs a machine learning model to achieve intelligent detection. Combined with a carefully constructed and diverse sample library, it not only enables accurate identification of high-resistance grounding faults but also allows for quantitative estimation of grounding resistance, improving the practicality of the detection results and providing a scientific basis for fault handling.

[0119] This invention establishes a graded handling mechanism based on grounding resistance values, generating corresponding handling instructions for faults of different risk levels, achieving precise fault response, effectively reducing the risk of accident escalation, and ensuring the safe and stable operation of the power distribution network.

[0120] Example 2

[0121] Based on Example 1, this example further discloses an intelligent detection system for high-resistance grounding faults in distribution networks, used to efficiently execute the above method. For example... Figure 2 As shown, the system includes:

[0122] The data acquisition module, physically composed of a high-precision TV, TA, and a high-speed data acquisition card, is responsible for step S1. Specifically, it consists of a high-precision zero-sequence voltage transformer (error ≤ 0.2%), a zero-sequence current transformer (error ≤ 0.5%), and a high-speed data acquisition card (sampling frequency 10kHz, resolution 16-bit), ensuring the accuracy and timeliness of the acquired data.

[0123] Preprocessing module: Implemented using an embedded DSP chip, with built-in GPS synchronous decoding, wavelet denoising (Db5, 3 layers), and... The elimination algorithm is responsible for the preprocessing part of S2, with a processing delay of <5ms, which meets the real-time requirements of power distribution network fault detection.

[0124] Feature extraction module: Integrated into the preprocessing module or host computer, it has embedded transient energy integration, 5-layer Db5 wavelet transform and entropy calculation, and FFT harmonic analysis algorithm, and is responsible for automatically generating 13-dimensional feature vectors or 9- or 10-dimensional features.

[0125] Model training module: It has a built-in sample library management unit and machine learning model training unit, which supports sample updates and iterative optimization of models to adapt to the operating characteristics of different power distribution networks.

[0126] Intelligent decision-making module: Equipped with a high-performance processor, it stores and runs the trained SVM model (parameters) After receiving the feature vector, the fault determination and resistance estimation are completed within 10ms.

[0127] Instruction generation module: Based on the output of the intelligent judgment module, according to... Threshold logic generates and sends corresponding isolation or warning commands via communication interface (RS485 / Ethernet) to ensure the timeliness and effectiveness of fault handling.

[0128] This invention effectively solves the problems of low sensitivity and high false judgment rate of traditional detection methods, eliminates the reliance on human experience, and improves the accuracy, real-time performance and practicality of high-resistance grounding fault detection in power distribution networks. It is applicable to the safe operation monitoring of various power distribution networks.

[0129] Example 3

[0130] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in Embodiment 1. This electronic device can be a dedicated fault detection device installed in a power distribution switchgear, or it can be a server in a power distribution network dispatching master station.

[0131] Example 4

[0132] This embodiment discloses a non-transitory computer-readable storage medium (such as a USB flash drive, hard disk, SSD, or cloud storage space) storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0133] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. For example, the wavelet denoising algorithm can be replaced with other types of wavelet functions (such as db4, sym5, etc.); the machine learning model can be replaced with deep learning neural networks (such as CNN, LSTM) or gradient boosting trees (such as XGBoost, LightGBM); the sampling frequency can be adjusted to values ​​≥5kHz, such as 5kHz or 8kHz, according to the actual distribution network conditions; the grounding resistance gradient and the number of samples can be expanded or adjusted according to the actual application scenario; the resistance threshold (30kΩ) for fault handling can be flexibly adjusted according to the safe operation requirements of the distribution network. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

[0134] The above technical solutions can also be extended to high-resistance grounding fault detection in distribution networks of different voltage levels (such as 6kV and 20kV distribution networks), or other similar weak signal fault detection scenarios (such as poor contact fault detection in low-voltage distribution networks). They are especially suitable for scenarios where fault characteristics are not obvious, noise interference is severe, and the requirements for detection accuracy and real-time performance are high.

Claims

1. A method for intelligent detection of high-resistance grounding faults in power distribution networks, characterized in that, include: The transient signals of the distribution network lines are collected, including transient zero-sequence voltage signals and transient zero-sequence current signals; The transient signal is preprocessed and a multi-dimensional feature set for characterizing the high-resistance grounding fault is extracted. The multi-dimensional feature set includes transient energy features, wavelet entropy features, and harmonic component features. The extracted multi-dimensional feature set is input into a pre-trained machine learning model, which simultaneously outputs the judgment result of high-resistance grounding fault and the estimated value of grounding resistance.

2. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, The transient energy characteristics are obtained by calculating the integral of the product of transient zero-sequence voltage and transient zero-sequence current over the transient fault period.

3. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, The wavelet entropy features are extracted in the following way: Wavelet transform is performed on the transient zero-sequence current signal to obtain wavelet coefficients at multiple scales; Calculate the information entropy of wavelet coefficients at each scale; Information entropy at various scales is combined to form wavelet entropy features.

4. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, The harmonic component features are extracted in the following way: Perform a Fast Fourier Transform on the transient zero-sequence voltage signal or the transient zero-sequence current signal; The amplitude and phase values ​​of the 3rd, 5th, and 7th harmonics are extracted from the transformation results and together constitute the harmonic component characteristics.

5. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, The specific process for preprocessing the transient signal is as follows: Synchronization technology based on the Global Positioning System or a precision clock source is used to time-align the acquired transient zero-sequence voltage signal and transient zero-sequence current signal, and the synchronization error is controlled within a set time threshold range. Wavelet transform threshold denoising method is used to denoise the synchronized signal in order to suppress power frequency harmonics, high frequency interference and random noise; Statistical discrimination criteria are used to remove outlier data points whose amplitudes exceed the effective range in the denoised signal.

6. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, The machine learning model is one of support vector machine, deep neural network or gradient boosting tree; the machine learning model is trained through a high-resistance grounding fault sample library, which includes fault samples generated based on transient signals collected under different grounding resistance values ​​and different fault locations, and normal samples generated based on transient signals collected during normal operation of the distribution network.

7. The intelligent detection method for high-resistance grounding faults in distribution networks according to claim 1, characterized in that, After outputting an estimated value of the grounding resistance, if the estimated value is less than or equal to a first preset threshold, a fault isolation command is generated and executed; if the estimated value is greater than the first preset threshold, a fault warning command is generated.

8. A smart detection system for high-resistance grounding faults in a distribution network, used to execute the smart detection method for high-resistance grounding faults in a distribution network as described in any one of claims 1-7, characterized in that, include: The signal acquisition module is used to acquire transient signals of the power distribution network lines, including transient zero-sequence voltage signals and transient zero-sequence current signals. The signal preprocessing module is used to preprocess the transient signal; The feature extraction module is used to extract a multi-dimensional feature set from the preprocessed transient signal to characterize the high-resistance grounding fault. The multi-dimensional feature set includes transient energy features, wavelet entropy features, and harmonic component features. The intelligent judgment module is used to input the extracted multi-dimensional features into a pre-trained machine learning model, and the machine learning model simultaneously outputs the judgment result of high-resistance grounding fault and the estimated value of grounding resistance. The instruction generation module is used to generate fault isolation instructions or fault warning instructions based on the judgment result and the estimated value.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a method for intelligent detection of high-resistance grounding faults in a power distribution network as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for intelligent detection of high-resistance grounding faults in a power distribution network as described in any one of claims 1 to 7.

Images