Leakage current fault monitoring and electric leakage point segmentation positioning method

Abstract

The invention discloses a leakage current fault monitoring and electric leakage point segmentation positioning method, which comprises a monitoring method and a segmentation positioning method, and is characterized in that the monitoring method comprises the following steps: data acquisition: sampling point arrangement: arranging sampling equipment at the head end, the tail end and key branch points of each line, and acquiring transient data of zero sequence current; denoising data: carrying out noise suppression by using wavelet transform or an adaptive filter to obtain a pure zero-sequence current signal; and ITD decomposition: decomposing the non-stationary signal into a plurality of inherent rotation components and monotonic trend components. According to the leakage current fault monitoring and electric leakage point segmentation positioning method, complex working conditions such as single-phase grounding, line tail end high-resistance grounding, arc suppression coils with different compensation degrees and noise interference are included, and the positioning accuracy reaches 95% or above and reaches 100% under some conditions.

Description

Technical Field

[0001] The invention relates to the technical field of leakage current fault monitoring, in particular to a leakage current fault monitoring and leakage point segmented positioning method. Background Art

[0002] At present, there are many grounding methods in the power system, mainly including neutral point direct grounding system, neutral point arc suppression coil grounding system (resonance grounding system), neutral point resistor grounding system and neutral point ungrounded system. These grounding systems may have the following fault problems in actual operation:

[0003] Leakage fault: In the grounding system, due to insulation aging, external force damage, equipment defects and other reasons, the insulation performance of the line to the ground may decrease, resulting in a leakage fault.

[0004] Grounding fault: If the grounding resistance of the grounding system is too large or the grounding wire is aging, it may cause a grounding fault and affect the safe operation of the system.

[0005] Resonant overvoltage: In a resonant grounding system, due to changes in line parameters, resonant overvoltage may occur, causing damage to system equipment.

[0006] Intermittent arcing: In a ground fault, intermittent arcing may occur, making the fault difficult to locate and eliminate quickly.

[0007] At present, the following methods are mainly used for leakage current monitoring and segmented location of leakage points in grounding systems:

[0008] Monitoring method based on current transformer: Leakage current is monitored by installing current transformers in the system, but this method is limited by the accuracy and response speed of the transformer.

[0009] Fault indicator method: A fault indicator is installed on the feeder. When a leakage fault occurs, the indicator can display the fault location, but it is greatly affected by environmental factors and has limited positioning accuracy.

[0010] Distributed fiber optic sensor method: Using distributed fiber optic sensors to monitor leakage current has high accuracy, but the cost is high and the installation and maintenance are complex.

[0011] Signal processing-based method: Fault detection and location are achieved by processing and analyzing the leakage current signal, such as Fourier transform, wavelet transform, etc., but the algorithm is highly complex and has poor real-time performance.

[0012] The existing monitoring and leakage point segmentation positioning methods have the following shortcomings: poor adaptability: different grounding systems require different monitoring methods, and the existing methods are difficult to adapt to a variety of grounding systems; low positioning accuracy: in a complex power grid environment, the leakage point positioning accuracy of the existing methods is often not high; poor real-time performance: when a fault occurs, the existing methods may not be able to quickly and accurately detect and locate the fault; cost and complexity: some high-precision monitoring methods are expensive and the system is complex, which is not conducive to large-scale promotion and application; Therefore, in response to the above problems, a leakage current fault monitoring and leakage point segmentation positioning method is designed to better meet the actual use needs. Summary of the invention

[0013] The object of the present invention is to provide a leakage current fault monitoring and leakage point segmentation location method to solve the problems raised in the above background technology.

[0014] To achieve the above object, the present invention provides the following technical solution: a leakage current fault monitoring and segmented location method of leakage point, including a monitoring method and a segmented location method, characterized in that the monitoring method includes the following steps:

[0015] Data collection:

[0016] Sampling point arrangement: Sampling equipment is arranged at the beginning, end and key branch points of each line to collect transient data of zero-sequence current;

[0017] Data denoising: Use wavelet transform or adaptive filter to suppress noise and obtain pure zero-sequence current signal;

[0018] ITD decomposition: decompose the non-stationary signal into multiple intrinsic rotation components and monotonic trend components;

[0019] Signal decomposition formula: Decompose the input signal x(t):

[0020] x(t)=PRC i (t))+RTC(t)

[0021] Baseline extraction: extract the baseline component of the signal;

[0022] Recursive decomposition: iteratively extract the rotation component;

[0023] Stop condition: When all extreme points of the residual component RTC(t) disappear, the iteration stops;

[0024] Feature extraction:

[0025] Extract zero-sequence current rotation component: extract the first rotation component of each line for calculating similarity;

[0026] Construct feature matrix: Construct feature matrix S according to the cosine similarity calculation results;

[0027] Fault feature classification: Using the feature matrix S as input, fault identification is performed on each line segment;

[0028] Data normalization: normalize each column of the matrix S;

[0029]

[0030] Feature labeling: The label of each feature matrix S[i] is the fault segment number;

[0031] Parameter adjustment suggestion: gain control parameter k

[0032] k = 0.5: suitable for most nonlinear signals;

[0033] k=0.3 or k=0.7: Adjust when the signal fluctuates violently;

[0034] The segmented positioning method includes the following steps:

[0035] Convolutional deep belief network is used to classify fault features and realize fault segmentation location. CDBN extracts fault features through multi-layer convolution restricted Boltzmann machine, optimizes feature representation layer by layer, and finally outputs fault segmentation classification results;

[0036] CDBN consists of multiple convolutional restricted Boltzmann machine (CRBM) layers, each of which contains the following components:

[0037] Visible layer: input feature matrix;

[0038] Hidden layer: learn local features;

[0039] Pooling layer: reduces computational complexity;

[0040] The energy function of each layer of CRBM is defined as:

[0041]

[0042] The CDBN construction process is as follows:

[0043] Input data: The feature matrix S obtained by ITD decomposition, with a size of N×M, where N is the number of samples and M is the feature dimension;

[0044] Network layer design: First layer convolution:

[0045] Convolution kernel size: k1×k1

[0046] Output feature number: C1

[0047] Convolution operation:

[0048]

[0049] Pooling layer: Use p×p pooling window for maximum pooling (Max Pooling):

[0050]

[0051] Deepen the network layer by layer:

[0052] The feature dimension of each layer gradually decreases, and the final output is the input of the fully connected layer;

[0053] Fully connected layer and classification:

[0054] Fully connected layer: Flatten the pooled features of the last layer and pass it through a fully connected layer;

[0055] Algorithm solution steps:

[0056] Data preprocessing: construct feature matrix S as input of CDBN;

[0057] Data normalization is performed to ensure that each eigenvalue is in the range of [0,1];

[0058]

[0059] Unsupervised pre-training: Each layer of CRBM performs unsupervised training and initializes weights and biases:

[0060] Randomly initialize W,b,c; update parameters using contrastive divergence;

[0061] Supervised fine-tuning: Use fully connected layers and target labels for supervised fine-tuning;

[0062] Training and validation:

[0063] Training set: used to train CDBN and optimize model parameters;

[0064] Validation set: evaluate the performance of the model in real time to prevent overfitting;

[0065] Fault segmentation location Input test sample (feature matrix StestS_{\text{test}}Stest) and output classification result;

[0066] Parameter setting and tuning:

[0067] Convolution kernel size: first layer k1 = 3 × 3, second layer k2 = 3 × 3;

[0068] Pooling window: p = 2, i.e. 2x downsampling;

[0069] Learning rate: initial value 0.001, using exponential decay strategy;

[0070] Dataset Construction: Data Generation

[0071] Simulation platform: Use PSCAD / EMTDC simulation to generate fault data;

[0072] Distribution network structure: The system is a 10kV distribution network, including 4 feeders;

[0073] Each feeder line consists of multiple sections, and the length of the sections is randomly distributed between 2 and 3 km;

[0074] The sampling frequency of data division is 10kHz, and the characteristic matrix is: through ITD decomposition, the first rotating component PRC1 of each zero-sequence current is extracted.

[0075] Preferably, the sampling frequency of data acquisition is 10kHz to capture the transient characteristics of fault occurrence.

[0076] Preferably, the sampling data range is a complete power frequency cycle before and after the fault occurs.

[0077] Prior to this, the sampling frequency ensures that the sampling frequency satisfies the Nyquist sampling theorem.

[0078] First, the back-propagation algorithm is used to update the parameters of the entire network.

[0079] Preferably, the classification result is the number of the fault segment.

[0080] Prior to this, the fault types are divided into single-phase to ground fault and fault point interval.

[0081] Preferably, the sample content of the data division is that each segment of fault data contains the zero-sequence current waveform of one power frequency cycle before and after the fault.

[0082] Prioritized, unsupervised pre-training: 50 contrastive divergence optimizations are performed on each CRBM layer.

[0083] Prioritized, supervised fine-tuning: fully connected layers are concatenated with the true labels and back-propagation optimized.

[0084] Compared with the prior art, the beneficial effects of the present invention are as follows: using the methods of intrinsic time scale decomposition (ITD) and convolutional deep belief network (CDBN) to analyze the transient zero-sequence current of the distribution network, so as to accurately monitor the leakage current fault and locate the leakage point in sections; using advanced signal processing technology to improve the accuracy of the segmented positioning of the leakage point, optimizing the algorithm, improving the real-time performance of fault monitoring and positioning, and combining artificial intelligence technology to realize intelligent monitoring and automatic positioning of leakage current faults;

[0085] Supports a wide range of fault types: including complex working conditions such as single-phase grounding, high-resistance grounding at the end of the line, arc suppression coils with different compensation degrees, and noise interference. The positioning accuracy rate is above 95%, and in some cases it reaches 100%. Noise robustness: When the signal-to-noise ratio (SNR) is 20dB, the accuracy rate remains above 95%; even under extreme working conditions (SNR = 10dB), the accuracy rate is still over 75%. High-resistance grounding fault: When the transition resistance is as high as 1kΩ, it can still maintain an accuracy rate of 94%, which is significantly better than traditional methods.

[0086] It can adapt to various complex grounding systems in the distribution network, including different compensation degrees of arc suppression coils, access to distributed power sources, etc. It has good scalability and is suitable for different distribution network topologies, such as radial, ring and complex branch distribution networks;

[0087] Through ITD decomposition and efficient feature extraction of CDBN, the fault location time can be significantly reduced. For the standard 10kV distribution network model, the location time from fault occurrence to result output is less than 1 second, which meets the actual engineering needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 It is a schematic diagram of the working process of the present invention. DETAILED DESCRIPTION

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] See also Figure 1 The present invention provides a technical solution: a leakage current fault monitoring and segmented positioning method of leakage points, including a monitoring method and a segmented positioning method, characterized in that the monitoring method includes the following steps: data collection and preprocessing:

[0091] Data collection:

[0092] Sampling point arrangement: Sampling equipment is arranged at the beginning, end and key branch points of each line to collect transient data of zero-sequence current.

[0093] Sampling frequency: 10kHz, ensuring the capture of transient characteristics of faults.

[0094] Sampling data range: a complete power frequency cycle (20ms) before and after the fault occurs.

[0095] Data denoising: Use wavelet transform or adaptive filter to suppress noise and obtain pure zero-sequence current signal:

[0096] Wavelet threshold denoising formula:

[0097]

[0098] in:

[0099] ·ψ k (t): wavelet basis function;

[0100] c k : wavelet decomposition coefficients;

[0101] λ: threshold (calculated based on signal-to-noise ratio).

[0102] ITD (Intrinsic Time-scale Decomposition) decomposition: ITD is a method to decompose a non-stationary signal into multiple intrinsic rotation components (PRCs) and monotonic trend components (RTCs).

[0103] Signal decomposition formula:

[0104] Decompose the input signal x(t):

[0105] x(t)=PRC i (t)+RTC(t)

[0106] Similarity calculation: Cosine similarity formula:

[0107]

[0108] in:

[0109] A, B: rotation components of two signals;

[0110] Sim(A,B): represents the correlation between two signals.

[0111] Construct feature matrix: According to the cosine similarity calculation results, construct feature matrix S:

[0112] S[i][j]=Sim(PRC 1，i , PRC 1，j )

[0113] in:

[0114] S[i][j]: represents the similarity between the i-th and j-th segments.

[0115] Fault feature classification: The feature matrix S is used as input to identify faults for each line segment. CDBN network structure: CDBN consists of multiple convolutional restricted Boltzmann machine (CRBM) layers, each of which contains the following components: visible layer: input feature matrix; hidden layer: learn local features; pooling layer: reduce computational complexity.

[0116] CRBM basic formula: The energy function of each layer is defined as:

[0117]

[0118] in:

[0119] υ i : the i-th visible unit;

[0120] h j : jth hidden unit;

[0121] W ij : weight matrix;

[0122] b i 、c j : Bias term.

[0123] Conditional probability:

[0124]

[0125] in is the sigmoid function.

[0126] CDBN construction process: Input data: The input data is the feature matrix S obtained by ITD decomposition, with a size of N×M, where N is the number of samples and M is the feature dimension.

[0127] Network layer design: First layer convolution: Convolution kernel size: k1×k1; Output feature number: C1 Convolution operation:

[0128]

[0129] Pooling layer: Use p×p pooling window for maximum pooling (Max Pooling):

[0130]

[0131] Deepen the network layer by layer: the feature dimension of each layer gradually decreases, and the final output is the input of the fully connected layer.

[0132] Fully connected layer and classification:

[0133] Fully connected layer: Flatten the pooled features of the last layer and pass them through a fully connected layer

[0134] y=softmax(Wx+b)

[0135] in:

[0136] W: weight matrix;

[0137] b: bias vector;

[0138] · For multi-classification tasks.

[0139] Algorithm solution steps:

[0140] Data preprocessing: construct feature matrix S as input of CDBN;

[0141] Data normalization ensures that each eigenvalue is in the range of [0,1]:

[0142]

[0143] Unsupervised pre-training: Each layer of CRBM performs unsupervised training and initializes weights and biases:

[0144] Randomly initialize W, b, c; use contrastive divergence (CD) to update parameters:

[0145]

[0146] △b i = i > data -〈v i > model

[0147] △c j = <h j > data -〈h j ) model

[0148] Supervised fine-tuning: Use fully connected layers with target labels for supervised fine-tuning:

[0149] Define the cross entropy loss function:

[0150]

[0151] Use the backpropagation algorithm to update the parameters of the entire network.

[0152] Training and validation:

[0153] Training set: used to train CDBN and optimize model parameters;

[0154] Validation set: evaluate the performance of the model in real time to prevent overfitting.

[0155] Fault segmentation location:

[0156] Input the test sample (feature matrix StestS_{\text{test}}Stest) and output the classification result; the classification result is the number of the fault segment.

[0157] Parameter setting and tuning:

[0158] Convolution kernel size: first layer k1 = 3 × 3, second layer k2 = 3 × 3;

[0159] Pooling window: p = 2, i.e. 2x downsampling;

[0160] Learning rate: initial value 0.001, using exponential decay strategy:

[0161] η=η0·exp(-λ·t)

[0162] Where λ = 0.1 and t is the number of iterations.

[0163] Result verification: Data set construction, data generation;

[0164] Simulation platform: Use PSCAD / EMTDC simulation to generate fault data.

[0165] Distribution network structure: The system is a 10kV distribution network consisting of 4 feeders.

[0166] Each feeder consists of multiple sections, and the lengths of the sections are randomly distributed in 2 ～ 3km.

[0167] Positive sequence and zero sequence parameters (unit: Ω / km):

[0168] Overhead lines: positive sequence (0.175+j1.210), zero sequence (0.230+j5.478).

[0169] Cable: positive sequence (0.270+j0.354), zero sequence (2.700+j1.109).

[0170] Fault type: single-phase grounding fault (A, B, C phases are all faulty). Fault point interval: one fault point is set every 100m.

[0171] Fault parameters: Initial angle: 0°, 30°, 60°, 90°, 145°, 175°.

[0172] Transition resistance: 0.01Ω, 10Ω, 100Ω, 500Ω, 1kΩ.

[0173] Data partitioning:

[0174] Sampling frequency: 10kHz.

[0175] Sample content: Each segment of fault data contains the zero-sequence current waveform of one power frequency cycle (20ms) before and after the fault.

[0176] Feature matrix:

[0177] Through ITD decomposition, the first rotating component PRC1 of each zero-sequence current is extracted.

[0178] Calculate similarity and construct feature matrix Where N=1000 (number of samples), M=50 (feature dimension).

[0179] Dataset division:

[0180] Training set: 70% samples (700,700,700 pieces of data).

[0181] Validation set: 15% samples (150,150,150 pieces of data).

[0182] Test set: 15% samples (150,150,150 pieces of data).

[0183] Model training: hyperparameter setting

[0184] Convolution kernel size: 3×3.

[0185] Pooling window: 2×2.

[0186] Hidden layer dimensions: 16 for the first layer and 32 for the second layer.

[0187] Learning rate: Initial value is 0.001, decaying by 10% every 10 rounds.

[0188] Optimizer: Adam.

[0189] Loss function: cross entropy.

[0190] Training process:

[0191] Unsupervised pre-training:

[0192] Each layer of CRBM is optimized 50 times using contrastive divergence (CD-1).

[0193] Initialize parameters layer by layer, and randomly initialize weights from Gaussian distribution N(0,0.01).

[0194] Supervised fine-tuning: The fully connected layer is connected to the true label (fault section number), and back-propagation optimization is performed. The number of training rounds is 100.

[0195] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A leakage current fault monitoring and segmented location method for leakage points, comprising a monitoring method and a segmented location method, characterized in that: The monitoring method includes the following steps: Data collection: Sampling point arrangement: Sampling equipment is arranged at the beginning, end and key branch points of each line to collect transient data of zero-sequence current; Data denoising: Use wavelet transform or adaptive filter to suppress noise and obtain pure zero-sequence current signal; ITD decomposition: decompose the non-stationary signal into multiple intrinsic rotation components and monotonic trend components; Signal decomposition formula: Decompose the input signal x(t): x(t)=PRC i (t)+RTC(t) Baseline extraction: extract the baseline component of the signal; Recursive decomposition: iteratively extract the rotation component; Stop condition: When all extreme points of the residual component RTC(t) disappear, the iteration stops; Feature extraction: Extract zero-sequence current rotation component: extract the first rotation component of each line for calculating similarity; Construct feature matrix: Construct feature matrix S according to the cosine similarity calculation results; Fault feature classification: Using the feature matrix S as input, fault identification is performed on each line segment; Data normalization: normalize each column of the matrix S; Feature labeling: The label of each feature matrix S[i] is the fault segment number; Parameter adjustment suggestion: gain control parameter k k = 0.5: suitable for most nonlinear signals; k=0.3 or k=0.7: Adjust when the signal fluctuates violently; The segmented positioning method includes the following steps: Convolutional deep belief network is used to classify fault features and realize fault segmentation location. CDBN extracts fault features through multi-layer convolution restricted Boltzmann machine, optimizes feature representation layer by layer, and finally outputs fault segmentation classification results; CDBN consists of multiple convolutional restricted Boltzmann machine (CRBM) layers, each of which contains the following components: Visible layer: input feature matrix; Hidden layer: learn local features; Pooling layer: reduces computational complexity; The energy function of each layer of CRBM is defined as: The CDBN construction process is as follows: Input data: The feature matrix S obtained by ITD decomposition, with a size of N×M, where N is the number of samples and M is the feature dimension; Network layer design: First layer convolution: Convolution kernel size: k1×k1 Output feature number: C1 Convolution operation: Pooling layer: Use p×p pooling window for maximum pooling (Max Pooling): Deepen the network layer by layer: The feature dimension of each layer gradually decreases, and the final output is the input of the fully connected layer; Fully connected layer and classification: Fully connected layer: Flatten the pooled features of the last layer and pass it through a fully connected layer; Algorithm solution steps: Data preprocessing: construct feature matrix S as input of CDBN; Data normalization is performed to ensure that each eigenvalue is in the range of [0,1]; Unsupervised pre-training: Each layer of CRBM is trained unsupervised, and weights and biases are initialized: W, b, c are randomly initialized; parameters are updated using contrastive divergence; Supervised fine-tuning: Use fully connected layers and target labels for supervised fine-tuning; Training and validation: Training set: used to train CDBN and optimize model parameters; Validation set: evaluate the performance of the model in real time to prevent overfitting; Fault segment location Input test samples (feature matrix StestS_{\text{test}}Stest) and output classification results; Parameter setting and tuning: Convolution kernel size: first layer k1 = 3 × 3, second layer k2 = 3 × 3; Pooling window: p = 2, i.e. 2x downsampling; Learning rate: initial value 0.001, using exponential decay strategy; Dataset Construction: Data Generation Simulation platform: Use PSCAD / EMTDC simulation to generate fault data; Distribution network structure: The system is a 10kV distribution network, including 4 feeders; Each feeder line consists of multiple sections, and the length of the sections is randomly distributed between 2 and 3 km; The sampling frequency of data division is 10kHz, and the characteristic matrix is: through ITD decomposition, the first rotating component PRC1 of each zero-sequence current is extracted.

2. A leakage current fault monitoring and leakage point segmentation location method according to claim 1, characterized in that: The sampling frequency of data acquisition is 10kHz, which captures the transient characteristics of fault occurrence.

3. A leakage current fault monitoring and leakage point segmentation location method according to claim 2, characterized in that: The sampling data range is a complete power frequency cycle before and after the fault occurs.

4. A leakage current fault monitoring and leakage point segmentation location method according to claim 3, characterized in that: The sampling frequency ensures that the sampling frequency satisfies the Nyquist sampling theorem.

5. A leakage current fault monitoring and leakage point segmentation location method according to claim 4, characterized in that: The back-propagation algorithm is used to update the parameters of the entire network.

6. A leakage current fault monitoring and leakage point segmentation location method according to claim 5, characterized in that: The classification result is the number of the fault segment.

7. A leakage current fault monitoring and leakage point segmentation location method according to claim 6, characterized in that: The fault types are divided into single-phase grounding fault and fault point interval.

8. A leakage current fault monitoring and leakage point segmentation location method according to claim 7, characterized in that: The sample content of data division is that each segment of fault data contains the zero-sequence current waveform of one power frequency cycle before and after the fault.

9. A leakage current fault monitoring and leakage point segmentation location method according to claim 8, characterized in that: Unsupervised pre-training: 50 contrastive divergence optimizations are performed on each CRBM layer.

10. A leakage current fault monitoring and leakage point segmentation location method according to claim 9, characterized in that: Supervised fine-tuning: The fully connected layer is connected to the true label and back-propagation optimization is performed.

Images