A clustering detection method and device for power distribution network single-phase ground fault time
By performing low-pass filtering, principal component analysis, and Hilbert transform dimensionality reduction on the zero-sequence current of the distribution network, combined with density clustering, the lag problem in single-phase grounding fault identification was solved, enabling accurate location of the fault moment and improving the accuracy of line selection and fault early warning capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD
- Filing Date
- 2023-03-29
- Publication Date
- 2026-06-30
AI Technical Summary
In existing technologies, the identification methods for single-phase grounding faults in distribution networks suffer from a lag in the start-up time, which leads to a decrease in the reliability of protection devices and makes it difficult to accurately identify the occurrence time of transient grounding faults, thus affecting the accuracy of fault location.
The zero-sequence current is preprocessed by constructing a low-pass filter, and the feature spectrum is extracted by empirical mode decomposition and fast Fourier transform. Dimensionality reduction is performed by combining principal component analysis and Hilbert transform. Finally, density clustering is used to identify the separation points between fault data and non-fault data to determine the time of fault occurrence.
It enables rapid identification of single-phase grounding faults, improves the accuracy of line selection, reduces the misjudgment rate of transient grounding faults, and can provide reliable early warning within ±10ms of the actual occurrence of the fault.
Smart Images
Figure CN116466179B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power distribution network technology, and specifically to a clustering detection method and apparatus for single-phase grounding faults in power distribution networks. Background Technology
[0002] Single-phase ground faults are highly probable in distribution networks and are a major factor in inducing forest fires and causing electric shock injuries and deaths. The State Grid Corporation's Q / GDW 10370-2016 "Technical Guidelines for Distribution Networks" proposes the principle of "safe arc suppression for transient faults and rapid isolation for permanent faults" for handling distribution network ground faults. With the introduction of line selection methods based on transient quantities and traveling waves, the accuracy of line selection devices has significantly improved. However, the start-up time identified by the zero-sequence voltage over-limit start-up criterion often lags behind the actual occurrence time of the ground fault. This affects the reliability of single-phase ground fault protection devices based on transient quantity principles, because the first half-wave signal that reflects the characteristics of the ground fault has already passed several cycles. Currently, the mainstream approach is to increase the waveform recording data window and trace back several cycles (5-10 cycles) to search for clearer characteristic signals. However, how to effectively determine the occurrence time of transient characteristics currently lacks targeted research. Similarly, for transient ground faults, the zero-sequence voltage after the fault will drop below the threshold after several cycles. Therefore, after each fault is detected, a certain period of time is required to determine whether the fault is a permanent ground fault. However, the threshold for the time of delay is also difficult to determine. Therefore, a method is needed to identify the period of transient fault occurrence. Summary of the Invention
[0003] To address the aforementioned technical problems, this invention provides a clustering detection method for single-phase ground faults in distribution networks, comprising:
[0004] The zero-sequence current of each feeder in the system is used as the original dataset. The original dataset is filtered by a pre-constructed low-pass filter to obtain a multi-dimensional feature dataset of the zero-sequence current of the feeders.
[0005] The multidimensional feature dataset is analyzed to extract the principal components of the zero-sequence current of each feeder in the system, and a two-dimensional feature dataset containing the time dimension is obtained.
[0006] The two-dimensional feature dataset is upgraded into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; the three-dimensional feature dataset is then reduced to a two-dimensional feature dataset without the time dimension through transfer projection.
[0007] Density-based clustering operations divide the two-dimensional feature dataset, which does not contain a time dimension, into fault data and non-fault data, and the separation point between the fault data and non-fault data is determined as the time of fault occurrence.
[0008] Furthermore, the step of filtering the original dataset using a pre-constructed low-pass filter to obtain a multi-dimensional feature dataset of the feeder zero-sequence current includes:
[0009] Low-pass filters are pre-constructed based on Empirical Mode Decomposition (EMD).
[0010] The original dataset is decomposed into multiple signals of different frequencies using the low-pass filter.
[0011] The characteristic spectra of the multiple signals at different frequencies are obtained by fast Fourier transform, and the dataset composed of the characteristic spectra of the signals at a specified frequency is used as the multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0012] Furthermore, the original dataset is decomposed into multiple signals of different frequencies using the low-pass filter, including:
[0013] The filter decomposes the original dataset into narrowband IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal.
[0014]
[0015] in, The sum of n narrowband component IMF signals; r n (n) represents the residual signal; imf1(t) ~ imf n (t) correspond to the decomposed signals from high frequency to low frequency, respectively.
[0016] Furthermore, the characteristic spectra of the multiple signals at different frequencies are obtained through Fast Fourier Transform, and the dataset composed of the characteristic spectra of the signals at a specified frequency is used as a multidimensional characteristic dataset of the zero-sequence current of the feeder, including:
[0017] The narrowband component IMF signal mf is analyzed using the Fast Fourier Transform as shown in the following formula. i (t) Calculate the characteristic spectrum and retain the narrowband component of the IMF signal with the main spectrum within 25–200 Hz. i (t);
[0018] The narrowband component IMF signal with the main spectrum in the range of 25-200Hz. i The characteristic spectrum composed of (t) serves as a multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0019] Furthermore, the principal components of the zero-sequence current of each feeder in the system are extracted, and the multidimensional feature dataset is analyzed to obtain a two-dimensional feature dataset containing the time dimension, including:
[0020] For the principal components of the zero-sequence current of each feeder in the system within the two-dimensional feature dataset, assuming there are m feeders and n zero-sequence current sampling points, a high-dimensional dataset X is constructed. n×m There are m n-dimensional samples, where column data are feeder numbers and row data are samples;
[0021] Sample data x for X (i) =(x (1) ,x (2) ,...,x (m) Centralization, that is
[0022] Deneutralized sample data x (i) The projection in a one-dimensional coordinate system is z (i) =w T x (i) , where w T is the projection operator, representing the coordinates of the projection hyperplane;
[0023] z (i) Together with the time dimension, they form a two-dimensional feature dataset containing the time dimension.
[0024] Furthermore, the two-dimensional feature dataset is upgraded into a three-dimensional feature dataset including a time dimension through Hilbert transform, including:
[0025] will be by z (i) The resulting one-dimensional time-series feature dataset is represented as a time-based signal z(t).
[0026]
[0027] Where ω is the frequency of the signal, Let A(ω) be the initial phase angle at that frequency, and let A(ω) be the amplitude at that frequency.
[0028] Applying a Hilbert transform to the signal z(t) and delaying its phase by 90° yields the transformed signal y(t).
[0029]
[0030] From the feature dataset z (i) Hilbert transform dataset y (i) and time t (i) To form a three-dimensional dataset q (i) =(t (i) ,y (i) ,z (i) Project this dataset onto the YZ plane, and perform a second dimensionality reduction to obtain a two-dimensional feature dataset f that does not include the time dimension. (i) =(z (i) ,j·y(i) ), where j is the imaginary unit.
[0031] This invention also provides a clustering detection device for single-phase ground faults in distribution networks, comprising:
[0032] The multidimensional feature dataset acquisition unit is used to take the zero-sequence current of each feeder in the system as the original dataset, and filter the original dataset through a pre-built low-pass filter to obtain the multidimensional feature dataset of the zero-sequence current of the feeder.
[0033] The first dimensionality reduction unit is used to analyze the multidimensional feature dataset, extract the principal components of the zero-sequence current of each feeder in the system, and obtain a two-dimensional feature dataset containing the time dimension.
[0034] The second dimensionality reduction unit is used to upgrade the two-dimensional feature dataset into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; and to reduce the three-dimensional feature dataset into a two-dimensional feature dataset without the time dimension through transfer projection.
[0035] The fault occurrence time determination unit is used for density-based clustering operations to divide the two-dimensional feature dataset without time dimension into fault data and non-fault data, and to determine the separation point of the fault data and non-fault data as the fault occurrence time.
[0036] Furthermore, the multidimensional feature dataset acquisition unit includes:
[0037] A low-pass filter construction subunit is used to pre-construct low-pass filters based on Empirical Mode Decomposition (EMD).
[0038] The signal decomposition subunit is used to decompose the original dataset into multiple signals of different frequencies through the low-pass filter.
[0039] The feature spectrum acquisition subunit is used to acquire the feature spectra of the multiple signals of different frequencies through fast Fourier transform, and to use the dataset composed of the feature spectra of the signals of a specified frequency as the multidimensional feature dataset of the zero-sequence current of the feeder.
[0040] Furthermore, the signal decomposition subunit includes:
[0041] The signal representation subunit is used by the filter to decompose the original dataset into narrowband component IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal.
[0042]
[0043] in, The sum of n narrowband component IMF signals; r n(n) represents the residual signal; imf1(t) ~ imf n (t) correspond to the decomposed signals from high frequency to low frequency, respectively.
[0044] Furthermore, the characteristic spectrum extraction sub-unit includes:
[0045] The filtering subunit is used to filter the narrowband component IMF signal mf according to the Fast Fourier Transform shown below. i (t) Calculate the characteristic spectrum and retain the narrowband component of the IMF signal with the main spectrum within 25–200 Hz. i (t);
[0046] Multidimensional feature dataset construction sub-unit, used to convert narrowband component IMF signals with main spectrum within 25-200Hz into IMFs. i The characteristic spectrum composed of (t) serves as a multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0047] This invention provides a clustering detection method and apparatus for single-phase grounding faults in distribution networks. It identifies potential single-phase grounding faults in the system, providing early warning for the safe operation of the distribution network. It can be used in conjunction with single-phase grounding fault location methods in distribution networks to address the delayed start-up problem caused by damping effects in transient-based single-phase grounding fault location methods, thus improving location accuracy. It has good identification capabilities for transient single-phase grounding faults, effectively reducing misjudgments. By processing the zero-sequence current of the distribution network feeders, the actual time of fault occurrence can be identified. Attached Figure Description
[0048] Figure 1 This is a flowchart illustrating a clustering detection method for single-phase grounding faults in power distribution networks provided by the present invention.
[0049] Figure 2 This is a flowchart of the EMD algorithm involved in this invention;
[0050] Figure 3 This is a flowchart of the density clustering algorithm involved in this invention;
[0051] Figure 4 This invention relates to the experimental space frame structure;
[0052] Figure 5 This invention relates to the zero-sequence current of a strong noise tree-blocking arc-grounding fault;
[0053] Figure 6 This invention relates to the EMD and spectrum analysis of zero-sequence current;
[0054] Figure 7 This is a comparison chart of the filtering effects involved in this invention;
[0055] Figure 8 This invention relates to a first-order dimensionality reduction based on principal component analysis;
[0056] Figure 9 This invention relates to a second dimensionality reduction based on Hilbert transform mapping;
[0057] Figure 10 This invention relates to fault occurrence timing identification based on density clustering;
[0058] Figure 11 This is a schematic diagram of the structure of a clustering detection device for single-phase grounding faults in power distribution networks provided by the present invention. Detailed Implementation
[0059] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0060] Currently, there are two commonly used start-up criteria for fault location devices: one is that the zero-sequence voltage exceeds 15% of the phase voltage, and the other is that the zero-sequence current surge is greater than 1A. These criteria set a relatively high threshold to avoid false start-up of the fault location device. However, the zero-sequence voltage of high-resistance grounding faults is often concentrated around 4-6% of the phase voltage, and the zero-sequence current surge may not reach the start-up threshold. Therefore, a major reason limiting further improvement in the accuracy of actual fault location is the difficulty in achieving sensitive identification of permanent faults while maintaining stability.
[0061] To address one or more of the aforementioned technical problems, this invention provides a clustering detection method for single-phase ground faults in distribution networks, the process of which is as follows: Figure 1 As shown, it includes the following steps:
[0062] Step S101: The zero-sequence current of each feeder in the system is used as the original dataset. The original dataset is filtered by a pre-constructed low-pass filter to obtain a multi-dimensional feature dataset of the zero-sequence current of the feeders.
[0063] This step primarily uses the zero-sequence current of each feeder as the original dataset. Leveraging the frequency adaptive capability of the Empirical Mode Decomposition (EMD) algorithm, a low-pass filter is constructed to filter the dataset, resulting in a multi-dimensional feature dataset with a dominant frequency below 25–200 Hz. Because the high-resistance ground fault current is relatively small, it is easily affected by noise. The main sources of noise are twofold: firstly, system background noise, primarily white noise; secondly, measurement noise from the current transformer (CT), where measurement errors are superimposed when the current is below the CT's minimum precision current; and thirdly, the potential presence of a DC component in the actual zero-sequence current, which has a significant impact, especially on short lines with small fault currents.
[0064] The EMD algorithm is a frequency-adaptive time-series signal decomposition method. Its basic idea is to transform an irregularly frequencyed signal into a superposition of multiple signals with approximate frequencies. Due to its adaptive frequency decomposition, this method is suitable for processing non-stationary zero-sequence current data generated by high-impedance grounding faults. The EMD algorithm implementation flowchart is shown below. Figure 2 As shown. First, a low-pass filter is pre-constructed based on Empirical Mode Decomposition (EMD). Then, the original dataset is decomposed into multiple signals of different frequencies using the low-pass filter. The characteristic spectra of the multiple signals of different frequencies are obtained by Fast Fourier Transform (FFT). The dataset composed of the characteristic spectra of the signals of a specified frequency is used as the multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0065] The filter decomposes the original dataset into narrowband IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal.
[0066]
[0067] in, The sum of n narrowband component IMF signals; r n (n) represents the residual signal; the EMD algorithm is a reversible decomposition, meaning it can be decomposed using IMF. i The original signal is restored by recombination of (t), where imf1(t) ~ imf n (t) correspond to the decomposed signals from high frequency to low frequency, respectively.
[0068] According to the Fast Fourier Transform shown in Equation (2), the narrowband component IMF signal mf i (t) Calculate the characteristic spectrum and retain the narrowband component of the IMF signal with the main spectrum within 25–200 Hz. i (t); Low-pass filtering effect is achieved by combining specific low-frequency IMF components.
[0069] The narrowband component IMF signal with the main spectrum in the range of 25-200Hz. iThe characteristic spectrum composed of (t) serves as a multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0070] Although the EMD algorithm has some frequency leakage issues, the purpose of applying this algorithm in this paper is to remove high-frequency noise and DC components, so frequency leakage has little impact on the algorithm proposed in this paper.
[0071]
[0072] Among them, X k For narrowband components IMF n The discrete spectrum, where N represents imf n The signal length.
[0073] Step S102: Analyze the multidimensional feature dataset, extract the principal components of the zero-sequence current of each feeder in the system, and obtain a two-dimensional feature dataset containing the time dimension.
[0074] Compared to zero-sequence voltage, zero-sequence current exhibits better transient characteristics after fault clearance, and can more significantly distinguish between permanent and transient ground faults. However, typical distribution networks have multiple feeders, making analysis difficult. It is necessary to extract the principal component that can characterize this feature from the zero-sequence current of all feeders.
[0075] The principal component can be defined as a one-dimensional time-series data feature, representing the feature that contributes the most to the variance of the zero-sequence current data of all feeders. For the principal components of the zero-sequence current of each feeder in the system within the two-dimensional feature dataset, assuming there are m feeders and n zero-sequence current sampling points, a high-dimensional dataset X is constructed. n×m There are m n-dimensional samples, where the column data are the categories, i.e., the feeder numbers, and the row data are the samples, i.e., the sampling data of the zero-sequence current;
[0076] Sample data x for X (i) =(x (1) ,x (2) ,...,x (m) Centralization, that is
[0077] Reducing the data from n dimensions to one dimension, i.e., decentralizing the sample data x (i) The projection in a one-dimensional coordinate system is z (i) =w T x (i) , where w T Let z be the projection operator, representing the coordinates of the projection hyperplane; if z is used... (i) To restore the original data x (i) The recovered data can be obtained.
[0078] z(i) Together with the time dimension, they form a two-dimensional feature dataset containing the time dimension.
[0079] Considering the entire sample set, we want all sample data to be close enough to the projection hyperplane, i.e., to take the minimum value of equation (5).
[0080]
[0081] Organizing can yield
[0082]
[0083] Minimizing the above expression is equivalent to
[0084]
[0085] We can obtain the Lagrange function.
[0086] J(W) = -tr(W) T XX T W+λ(W T WI)) (6)
[0087] Taking the derivative with respect to W, we get
[0088] XX T W=λW (7)
[0089] For the original dataset, you only need to use z (i) =W T x (i) This can reduce multidimensional zero-sequence current data into one-dimensional time-series characteristic data that does not include the time domain.
[0090] Step S103: The two-dimensional feature dataset is upgraded into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; the three-dimensional feature dataset is reduced to a two-dimensional feature dataset without the time dimension through transfer projection.
[0091] One-dimensional time-series feature data, excluding the time domain, has a uniform distribution in the time domain, making it difficult to achieve density-based feature clustering. Therefore, it needs to be mapped and transformed.
[0092] will be by z (i) The resulting one-dimensional time-series feature dataset is represented as a time-based signal z(t).
[0093]
[0094] Where ω is the frequency of the signal, Let A(ω) be the initial phase angle at that frequency, and let A(ω) be the amplitude at that frequency.
[0095] Applying a Hilbert transform to the signal z(t) and delaying its phase by 90° yields the transformed signal y(t).
[0096]
[0097] From the feature dataset z (i) Hilbert transform dataset y (i) and time t (i) To form a three-dimensional dataset q (i) =(t (i) ,y (i) ,z (i) Project this dataset onto the YZ plane, and perform a second dimensionality reduction to obtain a two-dimensional feature dataset f that does not include the time dimension. (i) =(z (i) ,j·y (i) ), where j is the imaginary unit.
[0098] Step S104: Based on density clustering, the two-dimensional feature dataset without time dimension is divided into fault data and non-fault data, and the separation point of the fault data and non-fault data is determined as the time of fault occurrence.
[0099] Two-dimensional feature dataset f can reflect the feature differences before and after a fault, but the features are not the same under different power distribution systems or grounding fault conditions, so an adaptive fault demarcation method is needed.
[0100] Because the data obtained from the second-order dimensionality reduction projection exhibits characteristics where non-fault data cluster together with high density, while data after a fault gradually moves away from non-fault data, this paper employs a density-based clustering method to distinguish between faulty and non-faulty data, thereby obtaining the fault time. The algorithm implementation flowchart is as follows: Figure 3 As shown.
[0101] Density clustering was used to divide the dimensionality-reduced feature data into two classes: non-fault data and fault data. The adjacent points of the two classes represent the detected fault time. Since the transient features used for line selection mainly exist within the first half-wave (10ms) after the fault occurs, the detected fault time is within ±10ms of the actual fault time and can be considered reliable.
[0102] Application examples are as follows:
[0103] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. It should be noted that the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0104] The specific embodiments of the present invention are described below with reference to the accompanying drawings.
[0105] A two-step dimensionality reduction-based clustering detection method for single-phase grounding faults in distribution networks
[0106] In such Figure 4 The diagram illustrates a simulated single-phase ground fault in a 10kV distribution network system. The system has three feeders, and the neutral point is resonantly grounded. A single-phase ground fault is simulated on one of the feeders. A complete sample period is 0.2s, and the data sampling frequency is Fs = 12.8kHz. The system has three feeders, and the zero-sequence current of each feeder is as follows: Figure 5 As shown, the selected sample data is an n×m dimensional dataset with n=2560 and m=3.
[0107] Preprocessing: The raw data is decomposed using the EMD algorithm, and spectral analysis is performed on each of the decomposed IMFs. Taking I30 as an example, its EMD decomposition results and corresponding spectrum are as follows: Figure 6 As shown.
[0108] The original signal was decomposed into 10 IMF components and one residual component. Only the IMF components and the residual component Res with the main spectrum within the range of 25–200 Hz were retained.
[0109]
[0110] Figure 6 The original signal and the filtered signal were compared. This method filters out most of the low-frequency noise and DC bias components.
[0111] First dimensionality reduction based on principal component analysis: The feature extraction algorithm based on principal component analysis proposed in this paper is used to perform dimensionality reduction on the filtered dataset X. 11×2560 Dimensionality reduced to one-dimensional feature data T 1×2560 ,like Figure 8 These are the dimensionality-reduced feature time series data.
[0112] The second dimensionality reduction based on Hilbert transform mapping: The feature time-series data after the first dimensionality reduction is then increased in dimensionality using the Hilbert transform, and then projected back onto the YZ platform for further dimensionality reduction. The dimensionality reduction process is as follows: Figure 9 As shown.
[0113] A fault occurrence time identification method based on density clustering: This paper uses the proposed density clustering method to classify the feature data after secondary dimensionality reduction. Under the condition that the sampling rate of the measured data is 25600Hz, and the number of neighboring samples δ = 100 is set, the experimental results show that setting the neighborhood ε = 0.35 can take into account both local linearity and global structural properties. The classification results are as follows: Figure 10 As shown in the figure, the boundary time between the fault moment and the non-fault moment can be clearly displayed.
[0114] Based on the same inventive concept, this invention also provides a clustering detection device 100 for single-phase ground fault detection in power distribution networks, such as... Figure 11 As shown, it includes:
[0115] The multidimensional feature dataset acquisition unit 110 is used to take the zero-sequence current of each feeder in the system as the original dataset, and filter the original dataset through a pre-built low-pass filter to obtain a multidimensional feature dataset of the zero-sequence current of the feeder.
[0116] The first dimensionality reduction unit 120 is used to analyze the multidimensional feature dataset, extract the principal components of the zero-sequence current of each feeder in the system, and obtain a two-dimensional feature dataset containing the time dimension.
[0117] The second dimensionality reduction unit 130 is used to upgrade the two-dimensional feature dataset into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; and to reduce the three-dimensional feature dataset into a two-dimensional feature dataset without the time dimension through the transfer projection method.
[0118] The fault occurrence time determination unit 140 is used for density-based clustering operations to divide the two-dimensional feature dataset without time dimension into fault data and non-fault data, and to determine the separation point of the fault data and non-fault data as the fault occurrence time.
[0119] Furthermore, the multidimensional feature dataset acquisition unit includes:
[0120] A low-pass filter construction subunit is used to pre-construct low-pass filters based on Empirical Mode Decomposition (EMD).
[0121] The signal decomposition subunit is used to decompose the original dataset into multiple signals of different frequencies through the low-pass filter.
[0122] The feature spectrum acquisition subunit is used to acquire the feature spectra of the multiple signals of different frequencies through fast Fourier transform, and to use the dataset composed of the feature spectra of the signals of a specified frequency as the multidimensional feature dataset of the zero-sequence current of the feeder.
[0123] Furthermore, the signal decomposition subunit includes:
[0124] The signal representation subunit is used by the filter to decompose the original dataset into narrowband component IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal.
[0125]
[0126] in, The sum of n narrowband component IMF signals; r n (n) represents the residual signal; imf1(t) ~ imf n (t) correspond to the decomposed signals from high frequency to low frequency, respectively.
[0127] Furthermore, the characteristic spectrum extraction sub-unit includes:
[0128] The filtering subunit is used to filter the narrowband component IMF signal mf according to the Fast Fourier Transform shown below. i (t) Calculate the characteristic spectrum and retain the narrowband component of the IMF signal with the main spectrum within 25–200 Hz. i (t);
[0129] Multidimensional feature dataset construction sub-unit, used to convert narrowband component IMF signals with main spectrum within 25-200Hz into IMFs. i The characteristic spectrum composed of (t) serves as a multidimensional characteristic dataset of the zero-sequence current of the feeder.
[0130] This invention provides a clustering detection method and apparatus for single-phase grounding faults in distribution networks. It identifies potential single-phase grounding faults in the system, providing early warning for the safe operation of the distribution network. It can be used in conjunction with single-phase grounding fault location methods in distribution networks to address the delayed start-up problem caused by damping effects in transient-based single-phase grounding fault location methods, thus improving location accuracy. It has good identification capabilities for transient single-phase grounding faults, effectively reducing misjudgments. By processing the zero-sequence current of the distribution network feeders, the actual time of fault occurrence can be identified.
[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.
Claims
1. A clustering detection method for single-phase ground faults in distribution networks, characterized in that, include: The zero-sequence current of each feeder in the system is used as the original dataset. The original dataset is filtered by a pre-constructed low-pass filter to obtain a multi-dimensional feature dataset of the zero-sequence current of the feeders. The multidimensional feature dataset is analyzed to extract the principal components of the zero-sequence current of each feeder in the system, and a two-dimensional feature dataset containing the time dimension is obtained. The two-dimensional feature dataset is upgraded into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; the three-dimensional feature dataset is then reduced to a two-dimensional feature dataset without the time dimension through transfer projection. Density-based clustering operations divide the two-dimensional feature dataset that does not contain a time dimension into fault data and non-fault data, and the separation point between the fault data and non-fault data is determined as the time of fault occurrence. The principal components of the zero-sequence current of each feeder in the system are extracted, and the multidimensional feature dataset is analyzed to obtain a two-dimensional feature dataset containing the time dimension, including: For the principal components of the zero-sequence current of each feeder in the system within the two-dimensional feature dataset, assuming there are m feeders and n zero-sequence current sampling points, a high-dimensional dataset is constructed. There are m n-dimensional samples, where column data are feeder numbers and row data are samples; Sample data for X Decentralization, that is ; Decentralized sample data Projected in a one-dimensional coordinate system as ,in is the projection operator, representing the coordinates of the projection hyperplane; Will Together with the time dimension, they form a two-dimensional feature dataset containing the time dimension; The two-dimensional feature dataset is upgraded into a three-dimensional feature dataset including a time dimension through Hilbert transform, including: Will be The constructed one-dimensional time-series feature dataset is represented as a time-based signal. , in, For the frequency of the signal, This is the initial phase angle at that frequency. This represents the amplitude at that frequency. For signal Perform a Hilbert transform to delay the phase by 90° to obtain the transformed signal. , From feature dataset Hilbert Transform Dataset and time To form a three-dimensional dataset Projecting this dataset onto the YZ plane, a second dimensionality reduction yields a two-dimensional feature dataset that does not include the time dimension. , where j is the imaginary unit.
2. The method according to claim 1, characterized in that, The process of filtering the original dataset using a pre-constructed low-pass filter to obtain a multi-dimensional feature dataset of the feeder zero-sequence current includes: Low-pass filters are pre-constructed based on Empirical Mode Decomposition (EMD). The original dataset is decomposed into multiple signals of different frequencies using the low-pass filter. The characteristic spectra of the multiple signals at different frequencies are obtained by fast Fourier transform, and the dataset composed of the characteristic spectra of the signals at a specified frequency is used as the multidimensional characteristic dataset of the zero-sequence current of the feeder.
3. The method according to claim 2, characterized in that, The original dataset is decomposed into multiple signals of different frequencies using the low-pass filter, including: The filter decomposes the original dataset into narrowband IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal. in, for The sum of narrowband component IMF signals; It is a residual signal; These correspond to the decomposed signals from high frequency to low frequency, respectively.
4. The method according to claim 2, characterized in that, The characteristic spectra of the multiple signals at different frequencies are obtained by Fast Fourier Transform. The dataset composed of the characteristic spectra of the signals at a specified frequency is used as a multidimensional characteristic dataset of the zero-sequence current of the feeder, including: The narrowband component IMF signal is analyzed using the Fast Fourier Transform as shown in the following formula. Calculate the characteristic spectrum and retain the narrowband IMF signal with the main spectrum within the range of 25~200Hz. ; Narrowband component IMF signal with main spectrum within 25~200Hz The resulting characteristic spectrum serves as a multidimensional characteristic dataset of the zero-sequence current in the feeder.
5. A clustering detection device for single-phase ground faults in power distribution networks, characterized in that, include: The multidimensional feature dataset acquisition unit is used to take the zero-sequence current of each feeder in the system as the original dataset, and filter the original dataset through a pre-built low-pass filter to obtain the multidimensional feature dataset of the zero-sequence current of the feeder. The first dimensionality reduction unit is used to analyze the multidimensional feature dataset, extract the principal components of the zero-sequence current of each feeder in the system, and obtain a two-dimensional feature dataset containing the time dimension. The second dimensionality reduction unit is used to upgrade the two-dimensional feature dataset into a three-dimensional feature dataset containing the time dimension through Hilbert transformation; and to reduce the three-dimensional feature dataset into a two-dimensional feature dataset without the time dimension through transfer projection. The fault occurrence time determination unit is used for density-based clustering operations to divide the two-dimensional feature dataset without time dimension into fault data and non-fault data, and to determine the separation point of the fault data and non-fault data as the fault occurrence time. The principal components of the zero-sequence current of each feeder in the system are extracted, and the multidimensional feature dataset is analyzed to obtain a two-dimensional feature dataset containing the time dimension, including: For the principal components of the zero-sequence current of each feeder in the system within the two-dimensional feature dataset, assuming there are m feeders and n zero-sequence current sampling points, a high-dimensional dataset is constructed. There are m n-dimensional samples, where column data are feeder numbers and row data are samples; Sample data for X Decentralization, that is ; Decentralized sample data Projected in a one-dimensional coordinate system as ,in is the projection operator, representing the coordinates of the projection hyperplane; Will Together with the time dimension, they form a two-dimensional feature dataset containing the time dimension; The two-dimensional feature dataset is upgraded into a three-dimensional feature dataset including a time dimension through Hilbert transform, including: Will be The constructed one-dimensional time-series feature dataset is represented as a time-based signal. , in, For the frequency of the signal, This is the initial phase angle at that frequency. This represents the amplitude at that frequency. For signal Perform a Hilbert transform to delay the phase by 90° to obtain the transformed signal. , From feature dataset Hilbert Transform Dataset and time To form a three-dimensional dataset Projecting this dataset onto the YZ plane, a second dimensionality reduction yields a two-dimensional feature dataset that does not include the time dimension. , where j is the imaginary unit.
6. The apparatus according to claim 5, characterized in that, The multidimensional feature dataset acquisition unit includes: A low-pass filter construction subunit is used to pre-construct low-pass filters based on Empirical Mode Decomposition (EMD). The signal decomposition subunit is used to decompose the original dataset into multiple signals of different frequencies through the low-pass filter. The feature spectrum acquisition subunit is used to acquire the feature spectra of the multiple signals of different frequencies through fast Fourier transform, and to use the dataset composed of the feature spectra of the signals of a specified frequency as the multidimensional feature dataset of the zero-sequence current of the feeder.
7. The apparatus according to claim 6, characterized in that, The signal decomposition subunit includes: The signal representation subunit is used by the filter to decompose the original dataset into narrowband component IMF signals with different frequencies, specifically represented as multiple narrowband component signals and a residual signal. in, for The sum of narrowband component IMF signals; It is a residual signal; These correspond to the decomposed signals from high frequency to low frequency, respectively.
8. The apparatus according to claim 6, characterized in that, The characteristic spectrum extraction sub-unit includes: The filtering subunit is used to filter the narrowband component IMF signal according to the Fast Fourier Transform shown below. Calculate the characteristic spectrum and retain the narrowband IMF signal with the main spectrum within the range of 25~200Hz. ; Multidimensional feature dataset construction sub-units are used to construct narrowband component IMF signals with the main spectrum within the range of 25~200Hz. The resulting characteristic spectrum serves as a multidimensional characteristic dataset of the zero-sequence current in the feeder.