A method for identifying a misoperation mode of a relay protection based on cluster analysis

Abstract

The application belongs to the technical field of cluster analysis, and particularly relates to a relay protection misoperation mode identification method based on cluster analysis, which comprises the following steps: collecting fault recording data, extracting current mutation, zero sequence component proportion and second harmonic ratio, and constructing a feature set through normalization; calculating sample local sparsity based on a KNN algorithm, combining a logarithmic difference of the sparsity with an energy gating term to construct a density adjustment factor; generating a dynamic radius by using the density adjustment factor to replace a fixed search threshold in a clustering algorithm, and introducing a shared neighbor check to remove pseudo-connected samples to construct a time sequence cluster; and based on the time sequence cluster statistical characteristics, if the second harmonic ratio is over the limit, it is determined to be a magnetic inrush current and is blocked, otherwise, according to the current mutation and the zero sequence component proportion, a metallic or high resistance grounding fault is determined according to the duration. The application solves the problems of difficulty in capturing sparse fault characteristics and misoperation identification by dynamically adjusting the clustering search range.

Description

Technical Field

[0001] This invention relates to the field of cluster analysis technology. More specifically, this invention relates to a method for recognizing malfunction patterns in relay protection based on cluster analysis. Background Technology

[0002] In the operation and maintenance support system of power systems, relay protection malfunction identification is a key execution link to verify the correctness of device behavior. Its core function is to convert the electrical quantity data collected by the fault recording device into fault nature criteria, and use feature analysis technology to distinguish between real fault signals and non-fault disturbances, so as to ensure the stability and safety of power grid operation.

[0003] Existing methods for identifying malfunctions generally employ density clustering techniques based on a fixed neighborhood radius. This assumes that various fault features are uniformly distributed in the data space, thereby setting a globally unified search radius and density threshold. During the analysis process, the recorded waveform feature data are clustered and divided strictly according to preset fixed geometric parameters.

[0004] However, in actual operation, the distribution density of fault characteristic data is not constant and uniform, but rather exhibits significant non-uniformity and sparsity due to the constraints of the fault's physical mechanism. Specifically, high-impedance grounding faults are limited by high impedance, resulting in a discrete and sparse distribution of characteristic points in space, while metallic short-circuit faults have highly concentrated characteristics, and background noise is randomly distributed.

[0005] When the algorithm uses a globally fixed search radius to process samples with varying density, if the focus is on high-precision narrow-radius search, sparsely distributed high-resistivity grounding faults will be cut into isolated noise points because they cannot meet connectivity requirements, leading to missed fault detections. If the search radius is forcibly expanded to capture sparse faults, the physical characteristic boundaries will be broken, and background noise or inrush currents that are originally discrete will be mistakenly absorbed into the fault cluster, causing misjudgments. This adaptive deficiency caused by the lack of a dynamic radius adjustment mechanism makes it difficult for existing technologies to achieve accurate automated identification under complex and ever-changing power grid conditions. Summary of the Invention

[0006] To address the technical problem of existing technologies using a fixed search radius when processing non-uniformly distributed fault features, which cannot simultaneously capture sparse high-resistivity faults and eliminate background noise, this invention provides a relay protection malfunction pattern recognition method based on clustering analysis. The method includes: obtaining samples to be analyzed by slicing waveform data using a sliding window; obtaining the current surge of the samples using a time-domain analysis algorithm; obtaining the proportion of the zero-sequence component of the samples using a vector decomposition algorithm; obtaining the second harmonic ratio of the samples using a frequency-domain transformation algorithm; normalizing the current surge, the proportion of the zero-sequence component, and the second harmonic ratio of each sample to construct a feature set; determining the nearest neighbor set of the features of the samples using the KNN algorithm; and determining the nearest neighbor set of the features of the samples based on their nearest neighbors. The local sparsity of the sample to be analyzed is obtained by measuring the distance differences between the features of neighboring samples in the neighbor set; the density adjustment factor is obtained based on the difference between the current mutation amount and the start-up threshold of the sample to be analyzed, and the logarithmic difference between the local sparsity and the historical benchmark mean; the adaptive gain coefficient is obtained based on the density adjustment factor; a dynamic radius is generated based on the preset base radius, density adjustment factor and adaptive gain coefficient; the DBSCAN clustering algorithm is improved by replacing the fixed search threshold in the DBSCAN clustering algorithm with the dynamic radius, and the improved DBSCAN clustering algorithm is used to cluster the feature set, and the shared nearest neighbor verification is applied to the clustering results to complete the construction of time-series clusters; the false action pattern recognition results are determined and output based on the second harmonic ratio, current mutation amount and zero-sequence component ratio in the time-series clusters.

[0007] This invention constructs a feature set of the samples to be analyzed using time-domain analysis, vector decomposition, and frequency-domain transformation algorithms, and performs standardized preprocessing to suppress waveform distortion caused by magnetic saturation of current transformers, ensuring the purity of feature data in multi-dimensional physical space. By calculating local sparsity to measure the distribution density of samples in the feature space, it distinguishes between weak fault points and uniform background noise, filling the blind spot in relay protection devices when identifying sparsely distributed fault features. Density adjustment is constructed by introducing a Heaviside step function and a Sigmoid nonlinear mapping. The algorithm avoids erroneously absorbing background noise due to an excessively large search range. By constructing a density adjustment factor that integrates sparsity bias and energy gating, and combining it with an adaptive gain coefficient to generate a dynamic radius, it implements differentiated adjustments to the search range of samples with different confidence levels, optimizing the adaptability of the fixed-radius DBSCAN algorithm under multi-density distribution conditions. By applying shared nearest neighbor verification to eliminate pseudo-connected samples and performing full-time correlation decision based on time-series clusters, it distinguishes between transient disturbances such as excitation inrush current and real faults such as high-resistance grounding, providing an objective basis for fault assessment under complex power system conditions.

[0008] Preferably, the step of normalizing the current mutation, zero-sequence component ratio, and second harmonic ratio of each sample to be analyzed and constructing a feature set includes: using the Z-score normalization algorithm and Sigmoid nonlinear mapping to uniformly map the current mutation, zero-sequence component ratio, and second harmonic ratio corresponding to each sample to be analyzed to a numerical range of 0 to 1; and summarizing all samples to be analyzed that contain the normalized current mutation, zero-sequence component ratio, and second harmonic ratio to construct a feature set.

[0009] Preferably, the step of obtaining the local sparsity of the sample to be analyzed based on the distance difference between the features of the sample to be analyzed and its neighboring samples in the nearest neighbor set includes: traversing the feature set using the standard K-nearest neighbor algorithm, calculating the Euclidean distance between each sample to be analyzed and all other samples, and selecting the k nearest samples to form the nearest neighbor set of the sample to be analyzed. The Euclidean distance between the sample to be analyzed and the k-th nearest neighbor sample is obtained and defined as the k-distance index of the sample to be analyzed; simultaneously, the nearest neighbor set is traversed, and for each sample in the set, its corresponding k-th nearest neighbor sample is determined and the Euclidean distance between them is obtained and defined as the k-distance index of the neighboring sample; the local sparsity of the sample to be analyzed is calculated based on the k-distance index of the sample to be analyzed and the k-distance index corresponding to each sample in the set.

[0010] Preferably, the local sparsity satisfies the following relationship: ; In the formula, Let be the local sparsity of the i-th sample to be analyzed; Let be the Euclidean distance between the i-th sample to be analyzed and its k-th nearest neighbor sample; Let be the number of samples in the nearest neighbor set of the i-th sample to be analyzed; Let k be the k-distance index of the j-th neighbor sample in the nearest neighbor set of the i-th sample to be analyzed; This represents the operation of an exponential function with the natural constant e as the base. To prevent constants with a denominator of zero.

[0011] This invention constructs local sparsity by calculating the Euclidean distance between the sample to be analyzed and other samples and combining it with the k-distance index. This characterizes the edge characteristics of the sample transitioning from the dense fault feature region to the sparse background noise region in the feature space. It measures the dispersion and local density difference of the sample to be analyzed, providing a statistical basis for identifying weak and sparsely distributed high-resistance grounding faults from strong noise backgrounds, and reducing the situation where weak fault features are misjudged as background noise.

[0012] Preferably, the density adjustment factor satisfies the following relationship: ; In the formula, For the first Density adjustment factor for each sample to be analyzed; For shape parameters; Let be the local sparsity of the i-th sample to be analyzed; This is the historical benchmark average. Let Heaviside be the unit step function; Let be the current mutation amount of the i-th sample to be analyzed; This is the activation threshold; It is the natural logarithm function.

[0013] This invention constructs a density adjustment factor by fusing the logarithmic difference between local sparsity and historical benchmark mean, and an energy gating term based on current mutation, to characterize the degree of conformity of the current sample to atypical fault modes, and to measure the density and energy level of the sample's environment; by using the energy gating term to forcibly disable the adjustment function under low-energy noise, and driving the search radius to adaptively expand according to sparsity when the gating is open, it maintains the purity of the cluster boundary while covering sparsely distributed fault edge points.

[0014] Preferably, the adaptive gain coefficient satisfies the following relationship: ; In the formula, Let be the adaptive gain coefficient for the i-th sample to be analyzed; Let be the density adjustment factor for the i-th sample to be analyzed; This is the arithmetic mean of the density adjustment factors for all samples in the feature set; is the basis smoothing constant.

[0015] This invention constructs an adaptive gain coefficient by comparing the density adjustment factor of the i-th sample to be analyzed with the arithmetic mean of the density adjustment factors of all samples in the feature set. This coefficient characterizes the discrete significance and fault confidence of the sample relative to the global background, and measures the necessity of clustering expansion of the sample in the current environmental context. This coefficient is used to grant greater search authority to sparse fault samples with high confidence to break spatial limitations, and to implement strict search restrictions on low-confidence samples belonging to environmental interference to keep them silent. This achieves differentiated processing of potential fault points and random noise points, and prevents erroneous actions caused by blindly expanding the search range.

[0016] Preferably, the dynamic radius satisfies the following relationship: ); In the formula, Let be the dynamic radius of the i-th sample to be analyzed; Base radius; Let be the density adjustment factor for the i-th sample to be analyzed; Let be the adaptive gain coefficient for the i-th sample to be analyzed; It is the natural logarithm function.

[0017] This invention generates a dynamic radius by nonlinearly stretching the base radius using a density adjustment factor and an adaptive gain coefficient. The dynamic radius represents the optimal search horizon of the i-th sample to be analyzed under the current environmental background, and measures the sample's ability to break through the sparsity space limitation and trace potential similar fault points. The radius is automatically reduced in the feature-dense region to avoid misjudging the normal operating state, and automatically expanded in the feature-sparse region to capture similar samples that are spatially distant, thus realizing the adaptive definition of the physical boundary of faults and interference.

[0018] Preferably, the step of applying shared nearest neighbor verification to the clustering results to complete the construction of time-series clusters includes: applying shared nearest neighbor verification to the clustering results, traversing sample pairs with density direct access relationships in the initial cluster set, calculating the number of shared nearest neighbors between each pair of samples, and if the number does not exceed a preset connectivity threshold, it is determined to be pseudo-connectivity and the association is forcibly severed, and isolated samples in the initial set that have lost connectivity due to the association being severed are removed, thus completing the construction of time-series clusters.

[0019] Preferably, the step of determining and outputting the malfunction pattern recognition result based on the second harmonic ratio, current surge, and zero-sequence component proportion in the time series cluster includes: calculating the proportion of samples in the time series cluster whose unnormalized second harmonic ratio exceeds the inrush current blocking threshold; if the proportion exceeds a preset confidence threshold, the time series cluster is determined to be an excitation inrush current interference; for time series clusters that do not meet the excitation inrush current interference criteria, if the average value of their unnormalized current surge exceeds a preset short-circuit surge threshold, the time series cluster is determined to be a metallic fault; if the average value of their unnormalized current surge is less than the short-circuit surge threshold, but the average value of their unnormalized zero-sequence component proportion exceeds a preset high-resistance grounding threshold, and the duration exceeds a preset continuity determination threshold, the time series cluster is determined to be a high-resistance grounding fault.

[0020] Preferably, the steps of obtaining the current mutation of the sample to be analyzed using a time-domain analysis algorithm, obtaining the proportion of the zero-sequence component of the sample to be analyzed using a vector decomposition algorithm, and obtaining the second harmonic ratio of the sample to be analyzed using a frequency-domain transformation algorithm include: the time-domain analysis algorithm is the cycle difference algorithm, the vector decomposition algorithm is the symmetric component method, and the frequency-domain transformation algorithm is the fast Fourier transform algorithm.

[0021] The beneficial effects of this invention are as follows: This invention utilizes time-domain analysis, vector decomposition, and frequency-domain transformation algorithms to construct multi-dimensional features including current surges, zero-sequence components, and second harmonic ratios, eliminating the masking of fault modes by system background and establishing a standardized data foundation for identifying malfunction modes; by calculating local sparsity based on Euclidean distance, it measures the dispersion and local density variation trend of the analyzed samples, stripping weak fault modes from strong noise environments and suppressing the tendency of relay protection devices to miss high-resistance grounding faults; by constructing a density adjustment factor that integrates sparsity bias and energy gating, it achieves close environmental control. The degree of nonlinear mapping suppresses low-energy noise while adaptively adjusting the analysis scale, overcoming the insufficient adaptability of fixed thresholds in the face of dispersed faults. By using density adjustment factors and adaptive gain coefficients to generate dynamic radii and combining them with shared nearest neighbor verification to establish search thresholds, adaptive adjustment of the clustering search range is achieved, breaking the limitations of the fixed-radius DBSCAN algorithm under multi-density distributions. Based on time-series clusters, time-series multi-mode clustering analysis and full-time correlation decision are performed, mining the difference between interference signals and real faults from the dynamic trend dimension, reducing the relay protection maloperation rate caused by transient interferences such as inrush current. Attached Figure Description

[0022] Figure 1 This is a flowchart of a relay protection malfunction pattern recognition method based on cluster analysis; Figure 2 This is a clustering result diagram of the fixed radius clustering algorithm; Figure 3 This is a clustering result diagram of the dynamic radius clustering algorithm. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] This invention discloses a method for recognizing relay protection malfunction patterns based on cluster analysis, referring to... Figure 1 This includes steps S1 to S5: S1. Collect fault waveform data and construct the feature set of the sample to be analyzed using relativization processing and feature extraction algorithms.

[0026] It should be noted that while relay protectors determine the fault type based on the fault current amplitude, current transformers can cause magnetic saturation, leading to waveform distortion. This makes it impossible for the relay protector to distinguish between fault signals and system fluctuations, resulting in malfunctions. Therefore, this invention extracts relative values ​​to construct multi-dimensional features, eliminating the masking effect of system background on fault modes and providing a standardized data foundation for identifying malfunction modes.

[0027] Specifically, fault recording devices acquire waveform data at a preset sampling frequency, and a sliding window is used to slice the waveform data to generate N samples to be analyzed. For each sample, a time-domain analysis algorithm is used to calculate the current mutation rate. A vector decomposition algorithm is used to calculate the proportion of the zero-sequence component. A frequency-domain transformation algorithm is used to calculate the second harmonic ratio. Using the Z-score normalization algorithm and sigmoid nonlinear mapping, the current mutation rate, the proportion of the zero-sequence component, and the second harmonic ratio corresponding to each sample are uniformly mapped to a value range of 0 to 1. All samples containing the normalized current mutation rate, the proportion of the zero-sequence component, and the second harmonic ratio are summarized to form a feature set.

[0028] In this embodiment, the sampling frequency is set to 10kHz. In other embodiments, the implementer can adjust the sampling frequency according to the system voltage level and the complexity of the fault characteristics. For example, when the system harmonic content is high, the sampling frequency can be appropriately increased; when the system is running stably, the sampling frequency can be decreased.

[0029] In this embodiment, the sliding window length is set to 10ms. It can be modified according to the actual situation. For example, when the system harmonic content is high, the sliding window length can be appropriately shortened; when the system is running stably, the sliding window length can be extended.

[0030] In this embodiment, the time-domain analysis algorithm is the cycle difference algorithm, the vector decomposition algorithm is the symmetric component method, and the frequency domain transformation algorithm is the fast Fourier transform algorithm. These can be modified according to the actual situation.

[0031] It should be added that the current surge refers to the non-periodic jump in the current waveform that occurs in a very short time. It can effectively filter out the load current during normal system operation and reflects the disturbance component generated by the fault. The zero-sequence component refers to the component generated when the vector sum of the three-phase currents is not zero. It is a characteristic physical quantity for judging whether a ground fault has occurred in the power system. The second harmonic ratio is the ratio of the amplitude of the second harmonic in the current to the amplitude of the fundamental wave. Due to the asymmetry of the saturation characteristics of the transformer core, the inrush current contains a large number of second harmonics, while the fault current is mainly the fundamental wave. Therefore, this indicator is a key criterion for distinguishing between faults and inrush currents.

[0032] S2. Calculate the local sparsity of the samples to be analyzed in the feature set based on the K-nearest neighbor algorithm.

[0033] It should be noted that the characteristic impedance of a high-impedance grounding fault is higher than the normal level, which reduces the rate of change of the fault current amplitude and makes its distribution sparse, causing the relay protection device to be unable to effectively identify the fault, resulting in missed detection. Therefore, this invention measures the sample dispersion by calculating local sparsity, providing a basis for separating weak fault modes from strong noise.

[0034] Specifically, the nearest neighbor parameter k is obtained by multiplying the preset sampling frequency by the sliding window length and taking the square root. The standard K-nearest neighbor algorithm is used to traverse the feature set, calculating the Euclidean distance between each sample to be analyzed and all other samples. The k nearest samples are selected to form the nearest neighbor set of the sample to be analyzed. The Euclidean distance between the sample to be analyzed and its k-th nearest neighbor is obtained and defined as the k-distance index of the sample to be analyzed. Simultaneously, the nearest neighbor set is traversed, and for each sample in the set, its corresponding k-th nearest neighbor is determined, and the Euclidean distance between them is obtained and defined as the k-distance index of the neighbor sample. Based on the k-distance index of the sample to be analyzed and the k-distance index corresponding to each sample in the set, the local sparsity of the sample to be analyzed is calculated.

[0035] Specifically, the local sparsity satisfies the following relationship: ; In the formula, Let be the local sparsity of the i-th sample to be analyzed; Let be the Euclidean distance between the i-th sample to be analyzed and its k-th nearest neighbor sample; Let be the number of samples in the nearest neighbor set of the i-th sample to be analyzed; Let k be the k-distance index of the j-th neighbor sample in the nearest neighbor set of the i-th sample to be analyzed; This represents the operation of an exponential function with the natural constant e as the base. To prevent constants with a denominator of zero; in this embodiment, It is 0.001, and can be modified later according to the actual situation.

[0036] in, Characterizes the trend of local density changes. The larger the value, the greater the density difference between the i-th sample to be analyzed and its neighboring samples in the set, indicating that the i-th sample to be analyzed is more likely to be located on the edge of the transition from the dense area of ​​fault features to the sparse area of ​​background noise. The larger the exponent value, the better the effect on the density difference between the i-th sample to be analyzed and its neighboring samples in the set. The stronger the amplification effect, the greater the local sparsity of the i-th sample to be analyzed. The larger the value, the more the i-th sample to be analyzed can be numerically distinguished from the uniform noise points inside, thus preventing it from being misjudged as background noise due to weak features.

[0037] The smaller the value, the smaller the density difference between the i-th sample to be analyzed and its neighboring samples in the set. This means the i-th sample is more likely to be located within a uniformly distributed noise region, making it difficult to distinguish from background noise. A smaller exponential term output value indicates a better understanding of the relationship between the i-th sample and its neighboring samples. The smaller the amplification effect, the lower the local sparsity. The smaller the value, the more likely the sample belongs to background noise or a typical failure mode.

[0038] S3. Based on logarithmic transformation and activation threshold, a density adjustment factor is constructed to realize the mapping of the density of the environment.

[0039] It should be noted that the effective value of metallic short-circuit fault current is high and concentrated, while the effective value of transitional fault current is low and dispersed. Background noise is also dispersed, making it impossible for a fixed threshold to adapt to these differences, leading to misjudgments by the relay protector when identifying dispersed faults. Therefore, this invention constructs a density adjustment factor that integrates sparsity bias and energy gating. This factor adaptively adjusts the analysis scale while suppressing low-energy noise, providing a foundation for identifying loose fault modes.

[0040] Specifically, samples containing only those from the fault disturbance period are extracted from the historical fault sample database, and their mean local sparsity is calculated as the historical baseline mean. The logarithmic difference between the local sparsity of the i-th sample to be analyzed and the historical baseline mean is calculated. A Heaviside step function is introduced, using the difference between the current mutation and a preset start-up threshold as input to construct an energy gating term. The density adjustment factor of the i-th sample to be analyzed is obtained based on the difference between the unnormalized current mutation and the start-up threshold, and the logarithmic difference between the local sparsity and the historical baseline mean for the samples to be analyzed in the feature set.

[0041] Specifically, the density adjustment factor satisfies the following relationship: ; In the formula, For the first Density adjustment factor for each sample to be analyzed; For shape parameters; For the first Local sparsity of the sample to be analyzed; This is the historical benchmark average. Let Heaviside be the unit step function; Let be the current mutation amount of the i-th sample to be analyzed; This is the activation threshold; It is the natural logarithm function.

[0042] in, This is an energy-gated term. In this embodiment, the Heaviside function is defined as follows: the function value is 1 when the input variable is strictly greater than 0, otherwise the function value is 0; when the current changes abruptly... Greater than the startup threshold When the function output is 1, the gating is enabled, allowing the density adjustment factor to adaptively adjust according to the sparsity; when... Less than or equal to When the function output is 0, the gating is forcibly closed, and the density adjustment factor is directly set to zero.

[0043] in, Let be the logarithmic difference between the local sparsity of the i-th sample to be analyzed and the historical benchmark mean. The larger the value, the higher the sample dispersion characteristic, and the more it conforms to atypical malfunction patterns. The closer it is to 0, the more it makes It approaches 1. Based on this, if the gating is enabled, the density adjustment factor... The value approaches 1, thereby driving the subsequent search radius to expand so that it can cover sparsely distributed fault edge points and prevent missed detections. The smaller the value, the more likely the sample conforms to normal operation or typical fault characteristics. Approaching 0, density adjustment factor The value approaches 0, thus limiting the search radius to stay near the base value, avoiding the incorrect absorption of background noise due to an excessively large search range, and ensuring the purity of the cluster boundaries.

[0044] In this embodiment, the shape parameter determines the shape of the Sigmoid deformation function, with an empirical value range of [1,3], and the shape parameter is set to 2. In other embodiments, the implementer can adjust the shape parameter according to the system's operating conditions. For example, when the system is in a scenario with high penetration of new energy sources, the shape parameter can be appropriately increased; when the system is running at a stable power frequency, the shape parameter can be appropriately decreased.

[0045] In this embodiment, the activation threshold determines the sensitivity of the energy gating. The empirical range of the activation threshold is [0.1, 0.3]. In this embodiment, the activation threshold is set to 0.2 times the rated current. In other embodiments, the implementer can adjust the activation threshold according to the background noise intensity. For example, when the background noise is strong, the activation threshold can be appropriately increased; when the background environment is clean, the activation threshold can be appropriately decreased.

[0046] S4. Utilize the density adjustment factor to generate the dynamic radius and introduce shared nearest neighbor verification to establish the search threshold.

[0047] It should be noted that, due to the extremely sparse data distribution of weak fault characteristics such as high-resistance grounding, the search range must be expanded to extend the detection boundary in order to effectively capture such faults. However, simply expanding the search range will inevitably include non-fault noise samples that are also discretely distributed, leading to a blurring of the physical boundary between faults and interference, thereby inducing maloperation of relay protection devices. Therefore, this invention constructs an improved DBSCAN algorithm, which uses dynamic radius expansion to capture sparse faults and combines shared nearest neighbor verification to eliminate false noise, thereby accurately defining the fault range.

[0048] Specifically, the arithmetic mean of the density adjustment factors of all samples to be analyzed in the feature set is calculated; based on the density adjustment factor of the i-th sample to be analyzed and the arithmetic mean of the density adjustment factors of all samples to be analyzed in the feature set, the adaptive gain coefficient of the i-th sample to be analyzed is obtained.

[0049] Specifically, the adaptive gain coefficients satisfy the following relationship: ; In the formula, Let be the adaptive gain coefficient for the i-th sample to be analyzed; For the first Density adjustment factor for each sample to be analyzed; This is the arithmetic mean of the density adjustment factors for all samples in the feature set; is the basis smoothing constant. In this embodiment, It is 0.001, and can be modified later according to the actual situation.

[0050] in, The numerical value reflects whether the i-th sample to be analyzed is worth clustering expansion in the current environmental context. A larger value indicates that the i-th sample to be analyzed is more likely to be in a working condition with a clean background but extremely discrete individual components, which is a typical characteristic of high-resistance grounding faults, and the probability of it being a fault is higher. In this case, the sample should be given greater search authority to break the spatial limitations of sparsity, actively trace potential similar fault points, and thus prevent missed detections.

[0051] when The smaller the value, the more likely the i-th sample to be analyzed is in a chaotic background with generally discrete noise, which is a typical environmental interference and has an extremely low fault confidence. In this case, the search behavior of the sample should be strictly limited and shielded, forcing it to remain silent. Regardless of its individual characteristics, it should not be allowed to blindly expand its influence range, thereby preventing accidental activation.

[0052] Furthermore, the dynamic radius of the i-th sample to be analyzed is calculated based on the preset base radius, the density adjustment factor of the i-th sample to be analyzed, and the adaptive gain coefficient.

[0053] Specifically, the dynamic radius satisfies the following relationship: ); In the formula, For the first The dynamic radius of the sample to be analyzed; Base radius; Let be the density adjustment factor for the i-th sample to be analyzed; Let be the adaptive gain coefficient for the i-th sample to be analyzed; It is the natural logarithm function.

[0054] in, The larger the value, the greater the probability that the i-th sample to be analyzed is in a highly discrete atypical fault condition. The larger the value, the greater the dynamic radius of the i-th sample to be analyzed. The larger the stretch, the larger the search radius according to the density clustering principle. The more easily the i-th sample to be analyzed can capture similar samples that are far apart in space, thus better preventing missed detections caused by the discrete and sparse fault features. The smaller the value, the greater the probability that the i-th sample to be analyzed is in a typical fault condition or normal operating state with low dispersion. The smaller the value, the smaller the dynamic radius of the i-th sample to be analyzed. The smaller the stretch, the smaller the search range, which helps to avoid misjudging normal operating conditions while maintaining the ability to identify typical fault modes and improving the accuracy of malfunction identification.

[0055] The larger the value, the stronger the discrete significance of the sample relative to the global background, and the stronger its significance relative to the background. The greater the magnification, the more the first... The dynamic radius of the sample to be analyzed The more significant the expansion. The smaller the value, the more deeply the sample is submerged by global background noise, and the more effective it is at responding to background noise. The stronger the inhibitory effect, the better.

[0056] In this embodiment, the base radius determines the benchmark for the dynamic radius, with an empirical value range of [0.05, 0.2], and the base radius is set to 0.1. In other embodiments, the implementer can adjust the base radius according to the sparsity of the fault characteristics. For example, when identifying a high-resistance ground fault, the base radius can be appropriately reduced; when identifying a metallic short circuit, the base radius can be appropriately increased.

[0057] Furthermore, the DBSCAN algorithm is improved by replacing the fixed search threshold with the dynamic radius corresponding to the samples in the feature set. The improved DBSCAN clustering algorithm is then used to cluster the feature set to obtain an initial cluster set. A shared nearest neighbor check is applied to the clustering results. Sample pairs with density-directed reach relationships are traversed in the initial cluster set, and the number of shared nearest neighbors between each pair is calculated. If this number does not exceed a preset connectivity threshold, the pair is considered pseudo-connected, and the association is forcibly severed. Isolated samples in the initial set that have lost connectivity due to association severance are removed, thus completing the construction of the temporal clusters.

[0058] In this embodiment, the connectivity threshold determines the density of the topological structure of samples within a cluster, and is used to determine whether there is a real physical connection between samples. The empirical range is [10, 30], and the connectivity threshold is set to 20. In other embodiments, the implementer can adjust the connectivity threshold according to the environmental noise level and recognition accuracy requirements.

[0059] It should be noted that the standard DBSCAN algorithm relies on a globally fixed neighborhood radius parameter as a unified benchmark to measure the physical density between samples, used to determine core points and drive cluster growth. However, a fixed radius has limitations when dealing with the multi-density distribution of power grid faults, easily leading to the contradiction of large radii introducing noise and small radii missing faults. This invention, on the other hand, constructs a sample-specific dynamic radius to achieve adaptive adjustment of the search range: that is, automatically reducing the radius in feature-dense areas to improve identification accuracy, and automatically expanding the radius in feature-sparse areas to improve capture sensitivity.

[0060] S5. Perform time-series multi-modal clustering analysis and make a decision based on the full-time correlation.

[0061] It should be noted that transient disturbances such as inrush current are extremely similar to real fault characteristics in their initial stages, and it is difficult to accurately distinguish them by simply relying on instantaneous sampling features, which poses a risk of misjudgment. To address this, this invention utilizes the constructed time series cluster to establish a full-time correlation, and by analyzing the evolution of feature quantities along the complete time axis, it mines the differences between disturbance signals and real faults from a dynamic trend perspective.

[0062] Specifically, the proportion of samples in a time series cluster whose unnormalized second harmonic ratio exceeds the inrush current blocking threshold is calculated. If the proportion exceeds a preset confidence threshold, the time series cluster is determined to be an excitation inrush current interference. A blocking control signal is then generated to shield the trip command output of the relay protection device, terminate the current process, and block the fault alarm. For time series clusters that do not meet the excitation inrush current determination conditions, if the average unnormalized current mutation amount of all samples in the time series cluster exceeds a preset short-circuit mutation threshold, it is determined to be a metallic fault. If the average unnormalized current mutation amount of all samples in the time series cluster is less than the short-circuit mutation threshold, but the average proportion of the unnormalized zero-sequence component exceeds a preset high-resistance grounding threshold, and the duration of the time series cluster on the time axis exceeds a preset continuity determination threshold, it is determined to be a high-resistance grounding fault.

[0063] In this embodiment, the inrush blockage threshold determines the identification threshold, with an empirical range of [0.1, 0.2], and is set to 0.15 in this embodiment; the confidence threshold determines the reliability of the judgment, with a range of [70%, 90%], and is set to 80% in this embodiment. When the system fault characteristics are weak, the above thresholds can be appropriately reduced to enhance the identification sensitivity.

[0064] In this embodiment, the short-circuit sudden change threshold determines the identification strength of metallic faults, ranging from [1.0, 1.5] times the rated current; in this embodiment, 1.2 times the rated current is used. In directly grounded systems, this threshold can be appropriately increased to avoid load fluctuation interference.

[0065] In this embodiment, the high-impedance grounding threshold determines the threshold for detecting weak faults, ranging from [0.2, 0.4], and is set to 0.3 in this embodiment; the continuity determination threshold aims to eliminate short-term random disturbances, ranging from [20ms, 60ms], and is set to 40ms in this embodiment. In strong interference environments, the continuity determination threshold can be appropriately increased to prevent false triggering.

[0066] It should be added that inrush current often occurs when the transformer is closed under no-load or when the voltage recovers. Its amplitude can reach 6-8 times the rated current, and the initial peak shape is extremely similar to that of a short-circuit fault, which can easily cause protection malfunction.

[0067] For example, Figure 2The image shows the clustering result of a fixed-radius clustering algorithm. In the three-dimensional feature space composed of current abrupt changes, zero-sequence components, and second harmonic ratios, the samples to be analyzed exhibit a non-uniform discrete distribution. The search range for each sample is defined by a fixed radius. Because this fixed-scale search range cannot cover the Euclidean distance between the samples, the search ranges of adjacent samples do not overlap or connect. This causes samples belonging to the same fault process to be incorrectly fragmented into unrelated clusters or misclassified as discrete points, disrupting the temporal integrity of the fault characteristics and preventing relay protection devices from accurately identifying faults.

[0068] Figure 3 This is a clustering result diagram of the dynamic radius clustering algorithm, in conjunction with... Figure 2 Within the same feature space, the search range corresponding to each sample to be analyzed is defined by a dynamic radius. In sparse regions, the search range expands significantly and overlaps with and connects with the search ranges of neighboring samples to be analyzed, forming continuous density channels; in dense regions, the search range automatically converges, thus ensuring that samples to be analyzed for the same fault process are all labeled into the same cluster. This search range constructed by the dynamic radius successfully establishes the temporal correlation between sparse and dense samples, aggregating them into complete temporal clusters, verifying the adaptive capture capability of this invention for non-uniformly distributed fault features.

Claims

1. A method for recognizing malfunction patterns in relay protection based on cluster analysis, characterized in that, include: The sample to be analyzed is obtained by slicing the waveform data using a sliding window; The current abrupt change in the sample to be analyzed is obtained using time-domain analysis algorithms; The proportion of zero-sequence components of the sample to be analyzed is obtained by using the vector decomposition algorithm; the second harmonic ratio of the sample to be analyzed is obtained by using the frequency domain transformation algorithm; and the current abrupt change, the proportion of zero-sequence components and the second harmonic ratio of each sample to be analyzed are normalized and then constructed as a feature set. The KNN algorithm is used to determine the nearest neighbor set of the features of the sample to be analyzed; the local sparsity of the sample to be analyzed is obtained based on the distance difference between the features of the sample to be analyzed and its neighboring samples in the nearest neighbor set; the density adjustment factor is obtained based on the difference between the current mutation amount and the start-up threshold of the sample to be analyzed, and the logarithmic difference between the local sparsity and the historical benchmark mean. The adaptive gain coefficient is obtained based on the density adjustment factor; the dynamic radius is generated based on the preset base radius, the density adjustment factor, and the adaptive gain coefficient. The DBSCAN clustering algorithm is improved by replacing the fixed search threshold in the DBSCAN clustering algorithm with a dynamic radius. The improved DBSCAN clustering algorithm is used to cluster the feature set, and the shared nearest neighbor verification is applied to the clustering results to complete the construction of time-series clusters. Based on the second harmonic ratio, current mutation amount and zero-sequence component ratio in the time-series clusters, the malfunction pattern recognition results are determined and output.

2. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The process of normalizing the current mutation rate, zero-sequence component ratio, and second harmonic ratio of each sample to be analyzed and constructing a feature set includes: using the Z-score normalization algorithm and Sigmoid nonlinear mapping to uniformly map the current mutation rate, zero-sequence component ratio, and second harmonic ratio of each sample to be analyzed to a numerical range of 0 to 1; and constructing a feature set from the current mutation rate, zero-sequence component ratio, and second harmonic ratio of each sample to be analyzed.

3. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The step of obtaining the local sparsity of the sample to be analyzed based on the distance difference between the features of the sample to be analyzed and its neighboring samples in the nearest neighbor set includes: traversing the feature set using the standard K-nearest neighbor algorithm, calculating the Euclidean distance of each sample to be analyzed to all other samples, and selecting the k nearest samples to form the nearest neighbor set of the sample to be analyzed; obtaining the Euclidean distance between the sample to be analyzed and the k-th nearest neighbor sample, and defining it as the k-distance index of the sample to be analyzed; simultaneously traversing the nearest neighbor set, for each sample in the set, determining its corresponding k-th nearest neighbor sample and obtaining the Euclidean distance between them, and defining it as the k-distance index of the neighbor sample; and calculating the local sparsity of the sample to be analyzed based on the k-distance index of the sample to be analyzed and the k-distance index corresponding to each sample in the set.

4. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 3, characterized in that, The local sparsity satisfies the following relationship: ； In the formula, Let be the local sparsity of the i-th sample to be analyzed; Let be the Euclidean distance between the i-th sample to be analyzed and its k-th nearest neighbor sample; Let be the number of samples in the nearest neighbor set of the i-th sample to be analyzed; Let be the Euclidean distance between the j-th neighbor sample and the k-th nearest neighbor sample in the nearest neighbor set of the i-th sample to be analyzed; This represents the operation of an exponential function with the natural constant e as the base. To prevent constants with a denominator of zero.

5. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The density adjustment factor satisfies the following relationship: ； In the formula, Let be the density adjustment factor for the i-th sample to be analyzed; For shape parameters; Let be the local sparsity of the i-th sample to be analyzed; This is the historical benchmark average. Let Heaviside be the unit step function; Let be the current mutation amount of the i-th sample to be analyzed; This is the activation threshold; It is the natural logarithm function.

6. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The adaptive gain coefficients satisfy the following relationship: ； In the formula, Let be the adaptive gain coefficient for the i-th sample to be analyzed; For the first Density adjustment factor for each sample to be analyzed; This is the arithmetic mean of the density adjustment factors for all samples in the feature set; is the basis smoothing constant.

7. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The dynamic radius satisfies the following relationship: )； In the formula, Let be the dynamic radius of the i-th sample to be analyzed; Base radius; Let be the density adjustment factor for the i-th sample to be analyzed; Let be the adaptive gain coefficient for the i-th sample to be analyzed; It is the natural logarithm function.

8. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The process of applying shared nearest neighbor verification to the clustering results to complete the construction of time-series clusters includes: applying shared nearest neighbor verification to the clustering results, traversing sample pairs with density direct access relationships in the initial cluster set, calculating the number of shared nearest neighbors between each pair of samples, and if the number does not exceed a preset connectivity threshold, determining it as pseudo-connectivity and forcibly severing the association, removing isolated samples in the initial set that have lost connectivity due to the association being severed, and completing the construction of time-series clusters.

9. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The method for determining and outputting malfunction pattern recognition results based on the second harmonic ratio, current surge, and zero-sequence component proportion in a time series cluster includes: calculating the proportion of samples in the time series cluster whose unnormalized second harmonic ratio exceeds the inrush current blocking threshold; if the proportion exceeds a preset confidence threshold, the time series cluster is determined to be an excitation inrush current interference; for time series clusters that do not meet the excitation inrush current interference criteria, if the average unnormalized current surge exceeds a preset short-circuit surge threshold, the time series cluster is determined to be a metallic fault; if the average unnormalized current surge is less than the short-circuit surge threshold, but the average proportion of its unnormalized zero-sequence component exceeds a preset high-resistance grounding threshold, and the duration exceeds a preset continuity determination threshold, the time series cluster is determined to be a high-resistance grounding fault.

10. The method for relay protection malfunction pattern recognition based on cluster analysis according to claim 1, characterized in that, The current mutation amount of the sample to be analyzed is obtained by using a time-domain analysis algorithm; The proportion of zero-sequence components of the sample to be analyzed is obtained using a vector decomposition algorithm; the second harmonic ratio of the sample to be analyzed is obtained using a frequency domain transformation algorithm, including: the time domain analysis algorithm is the cycle difference algorithm, the vector decomposition algorithm is the symmetric component method, and the frequency domain transformation algorithm is the fast Fourier transform algorithm.

Images