Power distribution network cable early fault identification method based on adaptive and principal component dimension reduction

Abstract

The application discloses a power distribution network cable early fault identification method based on self-adaption and principal component dimension reduction, and the method comprises the following steps: acquiring a fault phase current under a cable early operation state; extracting a multi-domain feature of the fault phase current; performing feature dimension reduction on the multi-domain feature of the fault phase current through a principal component analysis method, and obtaining effective features according to the size of component cumulative contribution rate; and inputting the obtained effective features into a convolutional neural network model after normalization processing, and performing fast and accurate classification identification. Through the extraction of the multi-domain effective features of the fault phase current, the application can realize the fast and effective identification of the cable early fault under the power distribution network background. The identification method is from the perspective of the cable multi-domain features, the effective multi-domain features after dimension reduction are selected as the basis for identifying the early fault and are subjected to classification identification, and meanwhile, a plurality of fault conditions are set, so that the accuracy and reliability of the classification identification are ensured.

Description

Technical Field

[0001] This invention belongs to the field of early fault identification and protection of cables, and particularly relates to an early fault identification method for distribution network cables based on adaptive and principal component dimensionality reduction. Background Technology

[0002] Cables, as one of the most important pieces of equipment in power distribution networks, are widely used in urban power distribution networks due to their advantages such as being typically buried underground, occupying little space, and being unaffected by weather conditions. However, after installation, cables are susceptible to partial discharge due to factors such as the operating environment, external damage, and improper operation by maintenance personnel. This partial discharge can develop into early faults, eventually leading to permanent faults and causing irreversible losses to the power system. The short duration of partial discharges makes detection or identification relatively difficult. Early faults, being repetitive transient faults preceding permanent failures, can be addressed by analyzing the wealth of information contained within them. This allows for proactive early warning and handling, overcoming the shortcomings and limitations of traditional reactive repair and inspection methods. Currently, early fault identification mainly involves feature extraction and classification. Feature extraction typically utilizes time-frequency analysis methods such as wavelet transform and empirical mode decomposition to extract single-domain features, which may result in the loss of multi-domain information. Excessive feature variables can also lead to information redundancy and the curse of dimensionality. Therefore, extracting effective features for subsequent classification is essential for accurate and effective identification of early cable faults. Summary of the Invention

[0003] The purpose of this invention is to provide an early fault identification method for distribution network cables based on adaptive S-transform and principal component dimensionality reduction, which solves the problem of accurately and effectively identifying and protecting early faults in distribution network cables from the perspective of extracting effective features of fault phase current.

[0004] This invention is implemented according to the following technical solution:

[0005] An early fault identification method for distribution network cables based on adaptive and principal component dimensionality reduction, the method comprising:

[0006] Obtain the fault phase current under the early operating conditions of the cable;

[0007] Extracting multi-domain features of the fault phase current, the multi-domain features including: time-domain features, frequency-domain features, and time-frequency-domain features;

[0008] Principal component analysis was used to reduce the dimensionality of the multi-domain features of the fault phase current, and the effective features were obtained by sorting the components according to their cumulative contribution rate.

[0009] The obtained effective features are normalized and then input into a convolutional neural network model for classification and identification.

[0010] In one embodiment, the fault phase current includes: fault phase currents under six cable operating conditions: early half-cycle fault, early multi-cycle fault, motor switching, load switching, capacitor switching, and fixed impedance grounding.

[0011] In one embodiment, the time-domain features include dimensional features and dimensionless features; the frequency-domain features include centroid frequency and frequency standard deviation; the time-frequency domain features are obtained by extracting the singular entropy of the mode-time-frequency matrix after the adaptive S-transform of the fault phase current.

[0012] Preferably, the adaptive S-transform includes:

[0013] The S-transform of any time series data x(t) can be defined as:

[0014]

[0015] In the formula, For a window function, the corresponding standard deviation is σ(f) = 1 / |f|;

[0016] After introducing the adaptive S-transform, its window function standard deviation is defined as:

[0017]

[0018] In the formula, a, b, c, and f are all adjustment parameters. By adjusting these parameters, the standard deviation of the window function is controlled, thereby controlling the time-frequency resolution of the S-transform. The window function is defined as:

[0019]

[0020] The corresponding adaptive S-transform is:

[0021]

[0022] Preferably, the singular entropy is:

[0023]

[0024] Where SE is the singular entropy and n is the number of singular values ​​in the modulus-time frequency matrix after the S-transform.

[0025] In one implementation, the cumulative contribution rate is expressed as:

[0026]

[0027] In one embodiment, prior to the principal component analysis, the extracted multi-domain features are subjected to KMO and Bartlett factor correlation tests, and the initial features are judged to be suitable for principal component analysis based on the test results.

[0028] In one implementation, the step of normalizing the obtained effective features and then inputting them into the convolutional neural network model specifically includes:

[0029] Initialize the parameters of the convolutional neural network;

[0030] After initialization, the effective features are normalized and then the training and test sets are divided according to the ratio.

[0031] The training set is fed into the convolutional neural network to train the model and update the model parameters.

[0032] After training is complete, the test set data is input into the model for validation.

[0033] In one implementation, the parameters for initializing the convolutional neural network include: network depth and number of neurons per layer, convolutional kernel size and dimensions, initial learning rate, and number of iterations.

[0034] Compared with existing technologies, the advantages of this invention are:

[0035] 1. This invention utilizes a power distribution network simulation model that combines cables and overhead lines, multi-domain feature extraction of fault phase currents for early faults and transient disturbances, and principal component dimensionality reduction of high-dimensional features to facilitate the extraction of effective features. Furthermore, it proposes a corresponding network model for the identification and detection of early faults and transient disturbances, enabling the classification of effective features, completing the identification of early faults, and improving the reliability and stability of cable operation.

[0036] 2. This invention obtains multi-domain initial features by calculating the time-domain, frequency-domain, and time-frequency-domain features of the fault phase current under different cable conditions. Furthermore, to address the limitation of the S-transform in time-frequency analysis where the window function is fixed, an adaptive S-transform is introduced into the time-frequency domain feature extraction.

[0037] 3. This invention uses principal component analysis to perform dimensionality reduction on the data by using initial features, thereby reducing information redundancy, simplifying the data structure, and avoiding the curse of dimensionality. By selecting appropriate principal component features based on the cumulative contribution rate of the principal components, high-dimensional features can be simplified without affecting their effectiveness, thus obtaining reliable features.

[0038] 4. This invention has good sensitivity and reliability. Selecting time-domain, frequency-domain, and time-frequency-domain multi-domain features of the fault phase current can more completely characterize all information of the cable's operating status. Using principal component analysis to reduce the dimensionality of the multi-domain features to obtain effective features can reduce information redundancy and improve subsequent identification efficiency. At the same time, classifying the effective features can effectively and quickly identify early faults from disturbances, which has high reliability. Furthermore, the feature classification and identification after dimensionality reduction has better effectiveness and accuracy. Attached Figure Description

[0039] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0040] Figure 1 This is a schematic diagram of the fault identification method provided by the present invention;

[0041] Figure 2 A flowchart illustrating the specific implementation of the fault identification method provided by this invention;

[0042] Figure 3 The simplified encapsulation diagram for the power distribution network simulation model is shown below (in the diagram, Bus represents a busbar, gray square modules represent cables, white square modules represent overhead lines, and triangles represent load modules).

[0043] Figure 4 Waveform diagram of electrical characteristics of multi-cycle arc fault;

[0044] Figure 5 Principal component analysis factor loading heatmap;

[0045] Figure 6 A comparison chart of classification and identification results for multiple features.

[0046] It should be noted that these accompanying drawings and textual descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art by referring to specific embodiments. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0049] This invention considers the randomness of fault time and location in actual field conditions, analyzes voltage and current characteristics under different neutral grounding methods, and proposes an early fault identification method for distribution network cables based on adaptive S-transform and principal component dimensionality reduction. This method classifies early faults and transient disturbances to protect cables. It is used for early fault identification of cables and safe operation of the power grid. The fault identification method steps are as follows:

[0050] Step S1: Obtain the fault phase current under the early operating conditions of the cable.

[0051] Furthermore, the fault phase current includes: fault phase currents under six cable operating conditions, such as cable half-cycle early fault, multi-cycle early fault, motor switching, load switching, capacitor switching, and fixed impedance grounding.

[0052] Step S2: Extract the multi-domain characteristics of the fault phase current.

[0053] Specifically, the fault phase currents under six cable operating conditions were used as raw data samples to extract multi-domain features of the fault phase currents. By analyzing the electrical quantity characteristics of arc faults, which conform to the characteristics of early cable faults, the time-domain, frequency-domain, and time-frequency-domain multi-domain features of the early fault phase currents were extracted as initial features.

[0054] Furthermore, the time-domain features include dimensional and dimensionless features, the frequency-domain features include centroid frequency and frequency standard deviation, and the time-frequency domain features are obtained by extracting the singular entropy of the mode time-frequency matrix after the adaptive S-transform of the fault phase current.

[0055] Specifically, the calculation method for multi-domain features is as follows:

[0056] (I) Temporal characteristics:

[0057]

[0058] Where x(t) represents time series data.

[0059] (II) Frequency Domain Characteristics:

[0060]

[0061] Where s(k) represents the spectrum of the original signal x(t), k = 1, 2, ..., K, K is the number of spectral lines, and f(k) represents the frequency value of the kth spectral line.

[0062] (III) Time-frequency domain characteristics:

[0063] First, an adaptive S-transform is performed on the fault phase current to obtain an m×n time-frequency matrix. Taking the modulus of this matrix yields the modulo-time-frequency matrix, where row vectors represent frequencies and column vectors represent times. This modulo-time-frequency matrix reflects the time-frequency characteristics of the signal. Then, singular value decomposition is performed on the matrix to obtain a set of singular values. Larger singular values ​​correspond to the principal components of the signal, while smaller singular values ​​are mainly noise components. Therefore, singular values ​​can characterize the original signal to some extent. Finally, to quantitatively describe the distribution characteristics of the signal, the singular entropy SE is defined as:

[0064]

[0065] In the formula: n is the number of singular values ​​of the modulus time-frequency matrix after S-transformation.

[0066] Considering the problem that the fixed size of the S-transform window makes it difficult to accurately filter out fixed-frequency signals, this invention introduces an adaptive S-transform based on the S-transform.

[0067] Preferably, the S-transform of any time-series data x(t) can be defined as:

[0068]

[0069] In the formula, For the window function, the corresponding standard deviation is σ(f) = 1 / |f|

[0070] After introducing the adaptive S-transform, its window function standard deviation is defined as:

[0071]

[0072] In the formula, a, b, c, and f are all adjustment parameters. Adjusting these parameters controls the standard deviation of the window function, thereby controlling the time-frequency resolution of the S-transform. The window function can then be defined as:

[0073]

[0074] The corresponding adaptive S-transform is:

[0075]

[0076] Step S3: Perform feature dimensionality reduction on the multi-domain features of the fault phase current using principal component analysis, and sort the effective features according to the cumulative contribution rate of the components.

[0077] Specifically, this invention introduces principal component analysis to reduce the dimensionality of the initial features, and obtains the effective features by sorting them according to the cumulative contribution rate of the components, as follows:

[0078] (1) If the number of data samples is m, the sample dataset is: X = {x1, x2, ..., x} m The feature value of each sample is represented as x. m ={x m1 ,x m1 ,…,x md} indicates that each sample has d feature values. The original data is arranged into an m x d matrix M, specifically represented as:

[0079]

[0080] (2) After standardizing the sample set M, M' is expressed as:

[0081]

[0082] In the formula: Let be the mean of each column, and var be the variance of the feature data in each column. The expressions are as follows:

[0083]

[0084] (3) Calculate the corresponding coefficient matrix R based on the feature data:

[0085]

[0086] (4) Find the eigenvalues ​​and eigenvectors of the coefficient matrix R.

[0087] |λI-R|=0

[0088] The calculated eigenvalues, sorted from smallest to largest, are represented as: λ1, λ2, ..., λ d The corresponding feature vectors are represented as: p1, p2, ..., p d

[0089] (5) Select the top k principal components based on the cumulative contribution rate to obtain the dimensionality-reduced feature data, where the cumulative contribution rate ζ is represented as:

[0090]

[0091] This feature extraction method can overcome the limitations of single-domain features and solve the problem of dimensionality curse or information redundancy caused by too many types of multi-domain features. The extracted effective features can comprehensively reflect the characteristic information of cable operation status and have strong reliability.

[0092] Optionally, before principal component analysis, the extracted multi-domain features may also be subjected to KMO and Bartlett factor correlation tests, and the initial features may be judged as suitable for principal component analysis based on the test results.

[0093] After dimensionality reduction, the influence of the initial feature variables on the principal component features is characterized by the factor loading heatmap, and the effective features of the principal components are calculated from the matrix component table.

[0094] Step S4: After normalizing the obtained effective features, input them into the convolutional neural network model for classification and identification.

[0095] Furthermore, the formation of a convolutional neural network model specifically includes:

[0096] After normalizing the effective features, the training set and the test set are divided according to the proportion.

[0097] Initialize the parameters of the convolutional neural network;

[0098] The training set is fed into the convolutional neural network to train the model and update the model parameters.

[0099] After completing the training of the convolutional neural network, a network model is generated;

[0100] The test set data is input into the network model for validation.

[0101] Furthermore, the initialization of network parameters for a convolutional neural network includes: network depth and number of neurons per layer, convolutional kernel size and dimensions, initial learning rate, and number of iterations.

[0102] Specifically, effective features of various states such as early faults and disturbances are extracted and normalized. Then, training and test sets are divided according to proportions, the parameters of the convolutional neural network are initialized, the training set data is input into the network model for learning and training, and the test set data is input into the convolutional neural network for testing and verification. The trained model can classify early faults and transient disturbances of distribution network cables, thereby achieving the purpose of early fault detection and comprehensive condition monitoring of cables.

[0103] Figure 2 The specific implementation flowchart of the fault identification method provided by the present invention is as follows:

[0104] A PSCAD simulation model of an IEEE 13-node distribution network with a mix of cable and overhead lines was established, and the simplified encapsulation is as follows: Figure 3 As shown in the diagram, gray squares represent cable modules, white squares represent overhead line modules, and triangles represent load modules. Based on the waveform characteristics of a multi-cycle arc fault with the neutral point grounded via an arc suppression coil, which match the waveform characteristics of an early-stage multi-cycle fault, combined with... Figure 4 It can be concluded that the present invention is feasible to use an arc module to simulate early faults. The simulation simulated six cases: half-cycle arc fault, multi-cycle arc fault, fixed impedance fault, capacitor switching, motor switching, and load switching. Different fault conditions such as neutral grounding method, fault line, and fault angle were set, and the fault phase current under different conditions was collected as raw data samples.

[0105] The time-domain, frequency-domain, and time-frequency-domain multi-domain features of the fault phase current under each state are extracted as initial features. The time-domain features include dimensional and dimensionless features, the frequency-domain features include centroid frequency, and the time-frequency-domain features include the singular entropy of the modulus-time-frequency matrix after the adaptive S-transform of the fault phase current.

[0106] By performing KMO and Bartlett factor correlation analysis on the extracted multi-domain features, the results are shown in Table 1. A KMO > 0.8 indicates that principal component analysis is highly suitable, while a p < 0.05 indicates rejection of the null hypothesis and suitability for principal component analysis. *** represents a 1% significance level. Table 1 shows a high degree of correlation among the initial features. This invention introduces principal component analysis to reduce the dimensionality of the initial features, and the effective features are obtained by sorting them according to their cumulative contribution rate.

[0107] Table 1 KMO test and Bartlett's test

[0108]

[0109] Combination Figure 5 The factor loading heatmap of principal component analysis shows that the numbers and color intensity represent the importance of the original features in each principal component feature.

[0110] Table 2 Principal Component Weight Results

[0111]

[0112] The results of principal component weight analysis based on information such as load coefficients are shown in Table 2. The variance explained rate reflects the explanatory power of the principal components for the original variables; its magnitude is directly proportional to the explanatory power and effectiveness of the principal components. The cumulative variance explained rate, also known as the cumulative contribution rate, is equal to the sum of the variance explained rates of each principal component. The weight is calculated as: variance explained rate / rotated cumulative variance explained rate. Table 2 shows that principal component feature 1 has the largest weight, reaching 65.885%, while principal component feature 7 has the smallest weight, at 0.496%. This feature extraction method overcomes the limitations of single-domain features and solves the problem of dimensionality curse or information redundancy caused by too many types of multi-domain features. The extracted effective features can comprehensively reflect the characteristic information of the cable's operating status and have strong reliability.

[0113] Effective features from various states, including early faults and disturbances, were extracted and normalized. The training and test sets were then divided in an 85:15 ratio. The parameters of the convolutional neural network (CNN) were initialized, with the network depth and number of neurons in each layer (C1-C2-C3-C4-F1) set to 16-32-64-128-2, the kernel size set to 3×3, the initial learning rate set to 0.005, and the number of iterations set to 300. The time-domain, frequency-domain, time-frequency-domain, time-domain+frequency-domain+time-frequency-domain, and time-domain+frequency-domain+time-frequency-domain dimensionality-reduced features of half-cycle early faults, fixed impedance grounding faults, and capacitor switching were then imported into the CNN for training. The comparison of the recognition performance of various features after training is shown in Table 3.

[0114] Table 3. Comparison of diagnostic performance for various features

[0115]

[0116] As shown in Table 3, when frequency domain features and time-frequency domain features are used as feature vector inputs for the convolutional neural network, the training times are 68.72s and 53.84s, respectively, which is much less than the training time of other features. This is because the dimensionality of frequency domain features and time-frequency domain features is smaller than other cases, and they occupy relatively less computer memory during training. However, the recognition accuracy of these two types of features is 48.52% and 82.01%, respectively, which is poor. When time domain features are used as input feature vectors, the recognition accuracy is too low. When time domain + frequency domain + time-frequency domain features are used as input feature vectors, although the recognition accuracy can reach the highest, the model training time is also the most. However, the recognition accuracy of the feature vectors after dimensionality reduction by time domain + frequency domain + time-frequency domain is only 0.6% lower than the highest accuracy, but the training time is reduced by 324.03s. Taking into account both identification accuracy and training time, the multi-domain features reduced by principal component analysis have better identification performance. Therefore, using time-domain + frequency-domain + time-frequency-domain multi-domain dimensionality-reduced features as the input feature vector of the convolutional neural network can accurately and efficiently complete the early fault identification of cables.

[0117] By sequentially adding motor switching, load switching, and early multi-cycle faults as state types 4, 5, and 6, and inputting single-domain features, multi-domain features, and multi-domain dimensionality-reduced features into a convolutional neural network for classification and identification, the accuracy can be obtained as follows: Figure 6 As shown. Combined with Figure 6 It can be seen that the trained model can classify early faults and transient disturbances of distribution network cables under various conditions. Considering both the identification accuracy and efficiency, the dimensionality-reduced features are more effective in realizing early fault detection and comprehensive condition monitoring of cables.

[0118] The cable early fault identification method provided in this invention is more conducive to extracting all information representing the operating status by extracting initial features from multiple domains. Through principal component analysis of the initial features, more accurate and effective features can be extracted for subsequent classification and identification, ensuring the effectiveness and reliability of early fault identification.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction, characterized in that: The method is used for early fault identification in cables and safe operation of power grids. Specific methods include: Obtain the fault phase current under the early operating conditions of the cable; Extracting multi-domain features of the fault phase current, the multi-domain features including: time-domain features, frequency-domain features, and time-frequency-domain features; Principal component analysis was used to reduce the dimensionality of the multi-domain features of the fault phase current, and the effective features were obtained by sorting the components according to their cumulative contribution rate. The obtained effective features are normalized and then input into a convolutional neural network model for classification and identification. The time-domain features include dimensional features and dimensionless features; The frequency domain features include the centroid frequency and the frequency standard deviation; The time-frequency domain features are obtained by extracting the singular entropy of the modulus time-frequency matrix after the adaptive S-transform of the fault phase current; The adaptive S-transform includes: The S-transform of any time series data x(t) is defined as: In the formula, For a window function, the corresponding standard deviation is σ(f) = 1 / |f|; After introducing the adaptive S-transform, its window function standard deviation is defined as: In the formula, a, b, c, and f are all adjustment parameters. By adjusting these parameters, the standard deviation of the window function is controlled, thereby controlling the time-frequency resolution of the S-transform. The window function is defined as: The corresponding adaptive S-transform is: In the formula, σ is the standard deviation and τ is the time delay; The singular entropy is: Where SE is the singular entropy, n is the number of singular values ​​in the modulus-time frequency matrix after the S-transform, and λ i and λ j Let i and j be the i-th and j-th eigenvalues ​​in the modulus-time-frequency matrix after S-transformation.

2. The method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction according to claim 1, characterized in that: The fault phase currents include: fault phase currents under six cable operating conditions: early half-cycle fault, early multi-cycle fault, motor switching, load switching, capacitor switching, and fixed impedance grounding.

3. The method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction according to claim 1, characterized in that: The cumulative contribution rate is expressed as: Where, λ i and λ j Let be the i-th and j-th eigenvalues ​​in the modulus-time-frequency matrix after S-transformation, k be the number of extracted eigenvalues, and d be the total number of eigenvalues.

4. The method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction according to claim 1, characterized in that: Before the principal component analysis, the KMO and Bartlett factor correlation tests are performed on the extracted multi-domain features, and the test results are used to determine whether the initial features are suitable for principal component analysis.

5. The method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction according to claim 1, characterized in that: After normalizing the effective features, the training set and the test set are divided according to the proportion. Initialize the parameters of the convolutional neural network; The training set is fed into the convolutional neural network to train the model and update the model parameters. After completing the training of the convolutional neural network, a network model is generated; The test set data is input into the network model for validation.

6. The method for early fault identification of distribution network cables based on adaptive and principal component dimensionality reduction according to claim 1, characterized in that: The initialization of the network parameters of the convolutional neural network includes: network depth and number of neurons per layer, convolutional kernel size and dimensions, initial learning rate and number of iterations.

Images