Power distribution network fault detection method and system based on multi-source data
By fusing multi-source data and extracting features, and combining the distribution network topology and load level, we have achieved highly sensitive and adaptable detection of distribution network faults, solving the problems of insufficient sensitivity and weak adaptability in existing technologies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAANXI SIRUI TOMORROW INTELLIGENT EQUIP CO LTD
- Filing Date
- 2026-03-19
- Publication Date
- 2026-04-17
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing fault detection methods for power distribution networks are not sensitive enough to high-impedance grounding faults and low-current grounding faults, and the models are not adaptable to different operating modes, leading to misjudgments or missed judgments.
By acquiring multi-source data from SCADA systems, micro phasor measurement units, and environmental monitoring terminals, multi-dimensional time-series features of wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis are extracted. The sameness, difference, and opposition components are calculated. Combined with the real-time topology and load level of the distribution network, fault identification is performed using the comprehensive opposition index and connection number.
It improves the accuracy and reliability of fault detection in power distribution networks, and can better adapt to changes in fault characteristics under different operating modes, reducing misjudgments and missed judgments.
Smart Images
Figure CN121880893A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of fault detection, specifically relating to a method and system for detecting faults in power distribution networks based on multi-source data. Background Technology
[0002] As the final link in the power system, the distribution network directly serves a large number of users, and its importance is self-evident. However, distribution networks are characterized by long lines, numerous branches, and complex operating environments, making them susceptible to faults caused by severe weather, external damage, and equipment aging. Therefore, the rapid and accurate detection and location of distribution network faults are crucial for ensuring power supply reliability.
[0003] Existing fault detection methods for distribution networks mainly rely on relay protection devices, which make judgments by detecting over-limit information of electrical quantities such as current and voltage, such as overcurrent protection and zero-sequence current protection. However, for complex fault types such as high-impedance grounding faults and low-current grounding faults, where fault characteristics are not obvious, there are problems of insufficient sensitivity and easy failure to operate or false operation.
[0004] With the integration of new loads such as distributed power sources and electric vehicle charging stations, the operating characteristics and fault transient processes of the power distribution network are becoming more complex, placing higher demands on the real-time performance and accuracy of fault detection technology.
[0005] Some studies have used wavelet transform, Hilbert-Huang transform, and other methods to extract time-frequency features of transient signals, and combined them with support vector machines, artificial neural networks, deep learning models for fault identification. However, these methods have failed to fully integrate the advantages of different data sources, and have not fully considered the impact of changes in distribution network topology and load levels on fault feature boundaries, resulting in poor adaptability of the models under different operating conditions and a tendency to misjudge or miss faults. Summary of the Invention
[0006] This invention provides a method and system for detecting distribution network faults based on multi-source data, in order to solve the technical problems of low sensitivity and poor adaptability in the existing technology of distribution network fault detection.
[0007] In a first aspect, the present invention provides a method for detecting faults in a distribution network based on multi-source data, comprising the following steps: S1. Acquire electrical quantity data from the SCADA system in the distribution network, synchronous phasor data from the micro phasor measurement unit, and environmental data from the environmental monitoring terminal to obtain multi-source operation data; for the multi-source operation data within a predetermined time window, extract multi-dimensional time-series features including wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis to obtain a real-time state feature sample set; and based on historical normal operation data, establish a normal state feature reference set using the same feature extraction method. S2, calculate the sameness component, difference component, and opposition component between the real-time state feature sample set and the normal state feature reference set, and calculate the distance between the centroid of the real-time state feature sample set and the normal state feature reference set; based on this distance, perform nonlinear correction on the basic opposition influence matrix to obtain the opposition influence matrix of the current working condition; use the opposition influence matrix of the current working condition to weight the feature components of the real-time state feature sample set to obtain the comprehensive opposition index. S3 uses the comprehensive opposition index as an adjustment factor, determines the opposition coefficient through a preset functional relationship, and calculates the connection number by combining the preset difference coefficient, the same degree component, the difference component and the opposition component; according to the real-time topology of the distribution network and the total active load level, it queries the preset relationship table to determine the fault judgment threshold: when the connection number is less than the fault judgment threshold and the comprehensive opposition index is greater than the preset opposition threshold, the distribution network is judged to have a fault.
[0008] Furthermore, the electrical quantity data includes the effective values of three-phase voltage and current, and the switching status; the synchronous phasor data includes voltage and current synchronous phasors, frequency and frequency change rate; and the environmental data includes temperature, humidity and wind speed.
[0009] Furthermore, multidimensional time-series features, including wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis, are extracted, including the following steps: The current signal in the electrical quantity data is decomposed into 5-level wavelet packets using the db4 wavelet basis. Calculate the proportion of energy in each decomposition layer frequency band to the total energy of the current signal, and form an energy distribution sequence; The wavelet energy entropy is calculated based on the Shannon entropy formula; A fast Fourier transform is performed on the zero-sequence current signal calculated based on electrical quantity data to extract the maximum amplitude in the 1kHz to 2kHz frequency band, and the maximum amplitude is taken as the transient zero-sequence current amplitude. The fourth central moment of the sampling point sequence in the electrical quantity data is calculated and divided by the fourth power of the standard deviation of the sampling point sequence to obtain the signal kurtosis.
[0010] Furthermore, S2 includes the following steps: Calculate the centroid vector and standard deviation vector of the normal state feature reference set; for a single feature sample in the real-time state feature sample set, determine its first... 3D feature vector Does it belong to the sameness feature, the difference feature, or the opposite feature? Then the feature components are counted as features of the same degree; if Then the feature components are counted as difference features; if Then the feature components are counted as the degree of contrast features, where, It is the first in the normal state characteristic reference set The average value across each feature dimension It is the first in the normal state characteristic reference set Standard deviation of each feature dimension; Divide the number of similarity features, difference features, and opposition features by the total feature dimension to obtain the normalized similarity components, difference components, and opposition components.
[0011] Furthermore, based on this distance, the fundamental opposing influence matrix is nonlinearly corrected to obtain the opposing influence matrix for the current operating condition, including the following steps: Using the correction function Calculate the nonlinear correction coefficient ,in, It is the adjustment coefficient of the correction function. The Mahalanobis distance of a single feature sample in the real-time state feature sample set relative to the centroid of the normal state feature reference set. The mean Mahalanobis distance between all samples in the normal state feature reference set and the centroid of the normal state feature reference set; Multiply each element in the fundamental opposition influence matrix by the aforementioned nonlinear correction coefficient. This yields the opposing influence matrix of the current operating condition.
[0012] Furthermore, the feature components of the real-time state feature sample set are weighted using the opposing influence matrix of the current operating condition to obtain a comprehensive opposing degree index, including the following steps: Construct an opposing feature indicator vector When the first feature sample in the real-time state feature sample set... When a feature component is determined to be a complementarity feature, the complementarity feature indicator vector... The One element is assigned the value 1, and the rest are assigned the value 0; Through calculation The comprehensive opposition index is obtained, where M is the opposition influence matrix of the current working condition. For the opposite feature indicator vector The transpose of .
[0013] Furthermore, using the comprehensive opposition index as a regulating factor, the opposition coefficient is determined through a pre-defined functional relationship, including the following steps: The degree of opposition coefficient is calculated using a correction function. ,in, This is the adjustment coefficient for the correction function. To comprehensively assess the degree of opposition, This is the preset benchmark value for the comprehensive degree of opposition.
[0014] Furthermore, the correlation coefficient is calculated by combining the preset difference coefficient, the sameness component, the difference component, and the oppositeness component, including the following steps: According to the formula Calculate the number of contacts ,in They are the same degree component. It is the difference component. It is a component of the degree of opposition. The preset difference coefficient, It is the coefficient of opposition.
[0015] Furthermore, based on the real-time topology of the distribution network and the total active load level, a pre-set relationship table is queried to determine the fault discrimination threshold, including the following steps: The real-time topology is divided into three operating modes: backbone power supply, tie line power supply, and ring network operation. The total active load level is divided into three levels according to the percentage of rated capacity: light load with a load rate of less than 30%, normal load with a load rate between 30% and 70%, and heavy load with a load rate of more than 70%. Construct a 3-row, 3-column two-dimensional relational table, using the operating mode as the row index and the load level as the column index, and query and determine the fault discrimination threshold.
[0016] Secondly, the present invention provides a distribution network fault detection system based on multi-source data, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned distribution network fault detection method based on multi-source data is implemented.
[0017] The beneficial effects are as follows: This invention corrects the asymmetric correlation between features by measuring the distance between the centroid of the real-time state feature sample set and the centroid of the normal state feature reference set, calculates the comprehensive opposition index, and then determines the opposition coefficient through a preset functional relationship. This can better discover the connection and mutual influence between different fault feature components. It combines the fault judgment threshold with the real-time topology of the distribution network and the total active load level, and uses a dual verification mechanism of connection number and comprehensive opposition index for decision-making, thereby improving the accuracy and reliability of distribution network fault detection. Attached Figure Description
[0018] Figure 1 The flowchart shows a method for fault detection in distribution networks based on multi-source data. Figure 2 This is a schematic diagram illustrating the correction of the basic opposing influence matrix; Figure 3 This is a structural diagram of a power distribution network fault detection system based on multi-source data. Detailed Implementation
[0019] An embodiment of the power distribution network fault detection method based on multi-source data provided by this invention: like Figure 1 As shown, the distribution network fault detection method based on multi-source data includes the following steps: S1. Acquire electrical quantity data from the SCADA system in the distribution network, synchronous phasor data from the micro phasor measurement unit, and environmental data from the environmental monitoring terminal to obtain multi-source operation data; for the multi-source operation data within a predetermined time window, extract multi-dimensional time-series features including wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis to obtain a real-time state feature sample set; and based on historical normal operation data, use the same feature extraction method to establish a normal state feature reference set.
[0020] Specifically, electrical quantity data uploaded by the SCADA system in the distribution network is acquired, including the effective values of three-phase voltage and current, and switch status; synchronization phasor data uploaded by micro phasor measurement units (PMUs) deployed at key nodes is collected, including voltage and current synchronization phasors, frequency, and rate of change of frequency; and environmental data uploaded by environmental monitoring terminals along the line corridor is collected, including temperature, humidity, and wind speed. After data cleaning and time-stamp alignment, a unified time-series multi-source operational dataset is formed.
[0021] In an optional embodiment, extracting multidimensional time-series features including wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis includes the following steps: The current signal in the electrical quantity data is decomposed into 5-level wavelet packets using the db4 wavelet basis. Calculate the proportion of energy in each decomposition layer frequency band to the total energy of the current signal, and form an energy distribution sequence; The wavelet energy entropy is calculated based on the Shannon entropy formula; A fast Fourier transform is performed on the zero-sequence current signal calculated based on electrical quantity data to extract the maximum amplitude in the 1kHz to 2kHz frequency band, and the maximum amplitude is taken as the transient zero-sequence current amplitude. The fourth central moment of the sampling point sequence in the electrical quantity data is calculated and divided by the fourth power of the standard deviation of the sampling point sequence to obtain the signal kurtosis.
[0022] Specifically, for a current signal containing 1024 sampling points, a 5-level decomposition using the db4 wavelet basis is performed to obtain signal components in 32 frequency bands. The energy of each frequency band is calculated; for example, the energy of the first frequency band is 10, the energy of the second frequency band is 5, and so on up to the 32nd frequency band. Assuming the total energy is 100, this constitutes an energy distribution sequence P, which may be 0.1, 0.05, etc. This energy distribution sequence P is substituted into the Shannon entropy formula to calculate a value; for example, the calculation result is 2.8. This value is the wavelet energy entropy feature.
[0023] The three-phase currents are synthesized into a zero-sequence current signal, and a spectrum is obtained through a fast Fourier transform. The highest peak value in the 1kHz to 2kHz interval of the spectrum is found. For example, if the amplitude corresponding to the peak value is 1.5A, then 1.5A is taken as the transient zero-sequence current amplitude characteristic.
[0024] For the 1024 sampling points of the original current signal, calculate its mean and standard deviation. Then, calculate the average of the sum of the fourth power of the difference between each sampling point and the mean to obtain the fourth central moment. Divide the fourth central moment by the fourth power of the standard deviation, for example, if the result is 4.5, then the signal kurtosis characteristic is obtained.
[0025] Ultimately, the three values of 2.8, 1.5, and 4.5 constitute the multidimensional feature vector at the current moment.
[0026] S2, calculate the sameness component, difference component, and opposition component between the real-time state feature sample set and the normal state feature reference set, and calculate the distance between the centroid of the real-time state feature sample set and the normal state feature reference set; based on this distance, perform nonlinear correction on the basic opposition influence matrix to obtain the opposition influence matrix of the current working condition; use the opposition influence matrix of the current working condition to weight the feature components of the real-time state feature sample set to obtain the comprehensive opposition index.
[0027] Specifically, each feature vector in the real-time state feature sample set is compared with the normal state feature reference set. If a feature vector in the real-time state feature sample set falls within the 95% confidence interval of the feature distribution in the normal state feature reference set, it is denoted as "same". The ratio of the number of feature vectors in the real-time state feature sample set that are judged as "same" to the total number of feature vectors in the real-time state feature sample set is the same-degree component. If a value falls outside the 95% confidence interval but within the preset fuzzy boundary, it is denoted as "dissimilar". The ratio of the number of feature vectors in the real-time state feature sample set that are judged as dissimilar to the total number of feature vectors in the real-time state feature sample set is the dissimilarity component. If it exceeds the fuzzy boundary, it is denoted as the opposite. The ratio of the number of feature vectors in the real-time state feature sample set that are judged as the opposite to the total number of feature vectors in the real-time state feature sample set is the degree of oppositeness component. Simultaneously, the mean vector and covariance matrix of all samples in the normal state feature reference set are calculated, and then the Mahalanobis distance from the feature vector in each real-time state feature sample set to the mean vector is calculated, and the mean of the distance is taken as the distance.
[0028] In a more specific embodiment, calculating the identity component, difference component, and opposition component between the real-time state feature sample set and the normal state feature reference set includes the following steps: Calculate the centroid vector and standard deviation vector of the normal state feature reference set; For a single feature sample in the real-time state feature sample set, determine its first... 3D feature vector Does it belong to the sameness feature, the difference feature, or the oppositeness feature? like If so, the feature components are counted as features of the same degree; like If so, the feature components are counted as difference features; like If so, the feature component is counted as the degree of opposition feature; in, It is the first in the normal state characteristic reference set The average value across each feature dimension It is the first in the normal state characteristic reference set Standard deviation of each feature dimension; Divide the number of similarity features, difference features, and opposition features by the total feature dimension to obtain the normalized similarity components, difference components, and opposition components.
[0029] Specifically, assume the normal-state feature reference set contains 1000 samples, each with three features: wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis. The mean and standard deviation of each of the 1000 samples in each dimension are calculated to obtain the centroid vector. ,For example The values are [2.5, 0.1, 3.0]; and the standard deviation vector. ,For example [0.2, 0.05, 0.4].
[0030] Obtain feature samples from a real-time state feature sample set. Its value is [2.6, 0.4, 4.5], for the first dimension feature ( The absolute value of its difference from the center of gravity is =0.1, less than the standard deviation =0.2, therefore it is determined to be a similarity feature. For the second feature ( The absolute value of its difference from the center of gravity is =0.3, greater than the standard deviation The value is 0.05, which is greater than three times the standard deviation of 0.15, therefore it is judged as a feature of contrast. Regarding the third dimension feature ( The absolute value of its difference from the center of gravity is =1.5, greater than the standard deviation =0.4, and greater than 3 times the standard deviation of 1.2, therefore it is also judged as an opposite feature. Statistically, the number of identical features is 1, the number of dissimilar features is 0, the number of opposite features is 2, and the total dimension is 3. After normalization, the identical component is obtained. The difference component is 1 / 3. The value is 0 / 3, representing the degree of opposition. It is 2 / 3.
[0031] In an optional embodiment, the basic opposing influence matrix is nonlinearly corrected based on this distance to obtain the opposing influence matrix for the current operating condition, including the following steps: Using the correction function Calculate the nonlinear correction coefficient ,in, It is the adjustment coefficient of the correction function. The Mahalanobis distance of a single feature sample in the real-time state feature sample set relative to the centroid of the normal state feature reference set. The mean Mahalanobis distance between all samples in the normal state feature reference set and the centroid of the normal state feature reference set; Multiply each element in the fundamental opposition influence matrix by the aforementioned nonlinear correction coefficient. This yields the opposing influence matrix of the current operating condition.
[0032] A 3×3 basic opposition influence matrix is established in advance based on expert experience and offline data analysis. Its elements represent the strength of the interaction between features. For example, for: ; in the matrix The term represents the degree of influence of wavelet energy entropy on transient zero-sequence current. The term represents the degree of influence of the transient zero-sequence current on the wavelet energy entropy, and Greater than This reflects the asymmetry of influence.
[0033] Simultaneously, based on the normal state feature reference set, the average Mahalanobis distance from all sample points to their centroids is calculated. , assuming It is 3.5.
[0034] When a new real-time feature sample arrives, calculate its Mahalanobis distance to the centroid of the normal-state feature reference set. Assuming the calculation yields It is 8.0. Then, and Substitute into the correction function to calculate the nonlinear correction coefficient. ,set up Taking 0.5, the nonlinear correction coefficient is approximately 0.91. Multiplying this correction coefficient of 0.91 by the fundamental opposing influence matrix... For each element, a new opposition influence matrix for the current operating condition is obtained, such as... Figure 2 As shown, for example, the basic opposing influence matrix The element with a value of 0.6 is corrected to 0.546.
[0035] In a more specific embodiment, the feature components of the real-time state feature sample set are weighted using the opposition influence matrix of the current operating condition to obtain a comprehensive opposition index, including the following steps: Construct an opposing feature indicator vector When the first feature sample in the real-time state feature sample set... When a feature component is determined to be a complementarity feature, the complementarity feature indicator vector... The One element is assigned the value 1, and the rest are assigned the value 0; Through calculation The comprehensive opposition index is obtained, where M is the opposition influence matrix of the current working condition. For the opposite feature indicator vector The transpose of .
[0036] Following the previous example, after determining the real-time feature samples... The second and third dimensions of the feature vector [2.6, 0.4, 4.5] represent the degree of contrast. A 3-dimensional contrast feature indicator vector is constructed. Its second and third elements are 1, and the first element is assigned the value 0, that is... column vector .
[0037] Use the opposing influence matrix of the current operating condition obtained in the previous step. Perform matrix quadratic operations to calculate... transpose (i.e., row vector [0, 1, 1]) and its opposing influence matrix The product of and , and then the resulting row vector is multiplied by the column vector V= Multiplying them yields the comprehensive index of opposition. For example, if M is: M The result is 1.5.
[0038] S3 uses the comprehensive opposition index as an adjustment factor, determines the opposition coefficient through a preset functional relationship, and calculates the connection number by combining the preset difference coefficient, the same degree component, the difference component and the opposition component; according to the real-time topology of the distribution network and the total active load level, it queries the preset relationship table to determine the fault judgment threshold: when the connection number is less than the fault judgment threshold and the comprehensive opposition index is greater than the preset opposition threshold, the distribution network is judged to have a fault.
[0039] For example, the on / off status of all circuit breakers is acquired in real time via SCADA to determine the current distribution network topology and classify it into operating modes such as mainline power supply, tie-line power supply, and ring network operation. The active power at all feeder outlets is summarized and classified into three load levels: heavy load, normal load, and light load. Based on the combination of operating mode and load level, the corresponding tie count threshold is determined by looking up the corresponding threshold in a pre-established two-dimensional lookup table. For example, the threshold is -0.5 for a ring network under heavy load and -0.2 for a radial network under light load.
[0040] A dual judgment is made using the found connection number threshold and a preset opposition threshold: First, the currently calculated connection number is less than the connection number threshold found in the table; second, the aforementioned calculated comprehensive opposition index is greater than the preset opposition threshold, for example, 3.0. When both conditions are met simultaneously, a fault is determined to have occurred in the distribution network, and an alarm signal is issued.
[0041] In an optional embodiment, the comprehensive opposition index is used as an adjustment factor, and the opposition coefficient is determined through a preset functional relationship, including the following steps: The degree of opposition coefficient is calculated using a correction function. ,in, This is the adjustment coefficient for the correction function. To comprehensively assess the degree of opposition, This is the preset benchmark value for the comprehensive degree of opposition.
[0042] Assuming the comprehensive opposition index is calculated through the aforementioned steps The preset benchmark value for the comprehensive antagonism index is 1.5. This represents the critical point where the degree of opposition goes from acceptable to unacceptable, for example... It is 1.0. =1.5 and Substituting 1.0 into the above function, assuming... =2.0, the resulting opposition coefficient The opposition coefficient is approximately -0.731, serving as a negative moderating factor. The evaluation score is reduced in the state assessment fusion formula to reflect the failure risk caused by the conflict between features.
[0043] In an optional embodiment, the connection coefficient is calculated by combining preset difference coefficients, identity components, difference components, and opposition components, including the following steps: According to the formula Calculate the number of contacts ,in They are the same degree component. It is the difference component. It is a component of the degree of opposition. The preset difference coefficient, It is the coefficient of opposition.
[0044] For example, the preset difference coefficient A value of 0 indicates neutral uncertainty. The calculated value... , and j and Substituting 0 into the formula ,Right now Calculate the number of connections The value of .
[0045] In an optional embodiment, based on the real-time topology of the distribution network and the total active power load level, a pre-set relationship table is queried to determine the fault discrimination threshold, including the following steps: The real-time topology is divided into three operating modes: backbone power supply, tie line power supply, and ring network operation. The total active load level is divided into three levels according to the percentage of rated capacity: light load with a load rate of less than 30%, normal load with a load rate between 30% and 70%, and heavy load with a load rate of more than 70%. Construct a 3-row, 3-column two-dimensional relational table, using the operating mode as the row index and the load level as the column index, and query and determine the fault discrimination threshold.
[0046] The current distribution network is determined to be operating in ring network mode through topology analysis. Simultaneously, the total active load of the current line, collected from SCADA, is 5MW, while the rated capacity of the line is known to be 10MW, resulting in a calculated load factor of 50%. According to the classification rules, 50% falls under the normal load level. A pre-set 3x3 two-dimensional quantitative relationship table is used, with rows corresponding to three operating modes: main line power supply, tie line power supply, and ring network operation, and columns corresponding to three load levels: light load, normal load, and heavy load. Ring network operation is used as the row index, and normal load as the column index to find the value at the intersection in the table. For example, if the threshold for the corresponding ring network operation and normal load position in the table is 0.95, then the fault discrimination threshold under the current operating condition is 0.95. If the final calculated connection number... If the value is below 0.95, it is considered a fault condition.
[0047] An embodiment of the power distribution network fault detection system based on multi-source data provided by this invention: like Figure 3 As shown, the distribution network fault detection system based on multi-source data includes a processor and a memory. The memory stores computer program instructions, which are executed by the processor to implement the above-mentioned distribution network fault detection method based on multi-source data.
[0048] The distribution network fault detection system based on multi-source data also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.
[0049] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.
Claims
1. A method for detecting faults in a distribution network based on multi-source data, characterized in that, Includes the following steps: S1. Obtain electrical quantity data of SCADA system in distribution network, synchronous phasor data of micro phasor measurement unit and environmental data of environmental monitoring terminal to obtain multi-source operation data; for multi-source operation data within a predetermined time window, extract multi-dimensional time series features including wavelet energy entropy, transient zero-sequence current amplitude and signal kurtosis to obtain real-time state feature sample set. Based on historical normal operation data, the same feature extraction method is used to establish a normal state feature reference set; S2, calculate the sameness component, difference component, and opposition component between the real-time state feature sample set and the normal state feature reference set, and calculate the distance between the centroid of the real-time state feature sample set and the normal state feature reference set; based on this distance, perform nonlinear correction on the basic opposition influence matrix to obtain the opposition influence matrix of the current working condition; use the opposition influence matrix of the current working condition to weight the feature components of the real-time state feature sample set to obtain the comprehensive opposition index. S3 uses the comprehensive opposition index as an adjustment factor, determines the opposition coefficient through a preset functional relationship, and calculates the connection number by combining the preset difference coefficient, the sameness component, the difference component and the opposition component. Based on the real-time topology of the distribution network and the total active load level, the fault discrimination threshold is determined by querying the preset relationship table: when the number of connections is less than the fault discrimination threshold and the comprehensive opposition index is greater than the preset opposition threshold, the distribution network is judged to have a fault.
2. The distribution network fault detection method based on multi-source data according to claim 1, characterized in that, The electrical quantity data includes the effective values of three-phase voltage and current, and switch status; the synchronous phasor data includes voltage and current synchronous phasors, frequency and frequency change rate; and the environmental data includes temperature, humidity and wind speed.
3. The distribution network fault detection method based on multi-source data according to claim 1, characterized in that, Extracting multidimensional time-series features, including wavelet energy entropy, transient zero-sequence current amplitude, and signal kurtosis, includes the following steps: The current signal in the electrical quantity data is decomposed into 5-level wavelet packets using the db4 wavelet basis. Calculate the proportion of energy in each decomposition layer frequency band to the total energy of the current signal, and form an energy distribution sequence; The wavelet energy entropy is calculated based on the Shannon entropy formula; A fast Fourier transform is performed on the zero-sequence current signal calculated based on electrical quantity data to extract the maximum amplitude in the 1kHz to 2kHz frequency band, and the maximum amplitude is taken as the transient zero-sequence current amplitude. The fourth central moment of the sampling point sequence in the electrical quantity data is calculated and divided by the fourth power of the standard deviation of the sampling point sequence to obtain the signal kurtosis.
4. The distribution network fault detection method based on multi-source data according to claim 1, characterized in that, S2 includes the following steps: Calculate the centroid vector and standard deviation vector of the normal state feature reference set; for a single feature sample in the real-time state feature sample set, determine its first... 3D feature vector Does it belong to the sameness feature, the difference feature, or the opposite feature? Then the feature components are counted as features of the same degree; if If so, the feature components are counted as difference features; like Then the feature components are counted as the degree of contrast features, where, It is the first in the normal state characteristic reference set The average value across each feature dimension It is the first in the normal state characteristic reference set Standard deviation of each feature dimension; Divide the number of similarity features, difference features, and opposition features by the total feature dimension to obtain the normalized similarity components, difference components, and opposition components.
5. The distribution network fault detection method based on multi-source data according to claim 4, characterized in that, Based on this distance, the fundamental opposing influence matrix is nonlinearly corrected to obtain the opposing influence matrix for the current operating condition, including the following steps: Using the correction function Calculate the nonlinear correction coefficient ,in, It is the adjustment coefficient of the correction function. The Mahalanobis distance of a single feature sample in the real-time state feature sample set relative to the centroid of the normal state feature reference set. The mean Mahalanobis distance between all samples in the normal state feature reference set and the centroid of the normal state feature reference set; Multiply each element in the fundamental opposition influence matrix by the aforementioned nonlinear correction coefficient. This yields the opposing influence matrix of the current operating condition.
6. The distribution network fault detection method based on multi-source data according to claim 5, characterized in that, The comprehensive opposition index is obtained by weighting the feature components of the real-time state feature sample set using the opposition influence matrix of the current operating condition, including the following steps: Construct an opposing feature indicator vector When the first feature sample in the real-time state feature sample set... When a feature component is determined to be a complementarity feature, the complementarity feature indicator vector... The One element is assigned the value 1, and the rest are assigned the value 0; Through calculation The comprehensive opposition index is obtained, where M is the opposition influence matrix of the current working condition. For the opposite feature indicator vector The transpose of .
7. The distribution network fault detection method based on multi-source data according to any one of claims 1-6, characterized in that, Using the comprehensive degree of antagonism index as an adjustment factor, the degree of antagonism coefficient is determined through a pre-defined functional relationship, including the following steps: The degree of opposition coefficient is calculated using a correction function. ,in, This is the adjustment coefficient for the correction function. To comprehensively assess the degree of opposition, This is the preset benchmark value for the comprehensive degree of opposition.
8. The distribution network fault detection method based on multi-source data according to claim 7, characterized in that, The correlation coefficient is calculated by combining the preset difference coefficient, the sameness component, the difference component, and the oppositeness component, including the following steps: According to the formula Calculate the number of contacts ,in They are the same degree component. It is the difference component. It is a component of the degree of opposition. The preset difference coefficient, It is the coefficient of opposition.
9. The distribution network fault detection method based on multi-source data according to claim 8, characterized in that, Based on the real-time topology of the distribution network and the total active load level, the fault discrimination threshold is determined by querying a preset relationship table, including the following steps: The real-time topology is divided into three operating modes: backbone power supply, tie line power supply, and ring network operation. The total active load level is divided into three levels according to the percentage of rated capacity: light load with a load rate of less than 30%, normal load with a load rate between 30% and 70%, and heavy load with a load rate of more than 70%. Construct a 3-row, 3-column two-dimensional relational table, using the operating mode as the row index and the load level as the column index, and query and determine the fault discrimination threshold.
10. A distribution network fault detection system based on multi-source data, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the distribution network fault detection method based on multi-source data as described in any one of claims 1-9 is implemented.