A method for rapid location of leakage faults on the low-voltage side of a distribution transformer

Abstract

This invention belongs to the field of fault prediction and health management technology, specifically relating to a rapid method for locating leakage faults on the low-voltage side of a distribution transformer. The method includes: collecting zero-sequence currents at each node, performing variational mode decomposition, and using the kurtosis-to-energy entropy ratio for noise reduction to reconstruct the zero-sequence waveform; calculating the average mutation amount and subtracting background noise using a noise prediction model; eliminating outliers from historical data, correcting the 3σ threshold using a skewness correction coefficient, adjusting the sliding window length and trigger threshold according to the mode completeness factor, and using the mutation amount to accumulate over-limit triggering criteria within the fault-tolerant sliding window; calculating the Pearson correlation coefficient of the reconstructed waveform, and locating the fault in the polarity reversal section. This invention can effectively eliminate background noise and adaptively adjust the judgment parameters, reducing the false tripping and refusal-to-operate rate, and achieving accurate and rapid location of leakage faults.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology. More specifically, this invention relates to a method for rapid location of leakage faults on the low-voltage side of a distribution transformer. Background Technology

[0002] The low-voltage side of distribution transformers has a large number of users, a complex network topology, and numerous branch lines. Furthermore, the operating environment is susceptible to factors such as humidity, external damage, and equipment aging, leading to frequent leakage current faults. Leakage current faults not only cause energy loss but can also lead to electric shock injuries and electrical fires, seriously threatening power distribution safety. The initial zero-sequence current signal of a fault is usually extremely weak and easily masked by the inherent zero-sequence current generated by three-phase imbalance, frequent fluctuations in user-side load, and strong ambient noise. Traditional low-voltage leakage current protection devices mostly only achieve over-limit tripping at the main outlet or coarse area isolation, suffering from low sensitivity, high false tripping rate, and insufficient location accuracy, making it difficult to accurately identify weak leakage current characteristics in complex, high-noise environments.

[0003] Currently, mode decomposition of zero-sequence current signals and extraction of transient signals from the decomposed characteristic components for fault identification has become a common method. Chinese patent document CN111413588B discloses a method for selecting fault lines in a distribution network with single-phase grounding faults. This method utilizes variational mode decomposition of the zero-sequence current to obtain multiple modes, calculates kurtosis values, and selects the mode with the largest kurtosis value for first-order difference operations to determine singular points. Then, it calculates the comprehensive inner product between the faulty feeder and the healthy feeder at the singular point, performs standardization processing, and compares it with a threshold. Based on these techniques, fault characteristics can be measured from the perspective of microscopic features, improving the accuracy of fault line selection results and avoiding misjudgments under extreme fault conditions.

[0004] However, the above-mentioned technical solutions can, to some extent, reduce noise through modal decomposition and utilize feature point calculations to achieve fault diagnosis in power distribution networks. However, existing technologies still have significant shortcomings in rapidly locating weak leakage faults on the low-voltage side of distribution transformers. In the mode screening stage, relying solely on single indicators such as kurtosis makes it difficult to simultaneously preserve transient impact characteristics and suppress random environmental noise, easily leading to the loss of effective high-frequency transient characteristics or the retention of low-frequency noise, causing weak leakage characteristics to be easily submerged by strong noise. In the abrupt change extraction stage, simple differential or fixed historical benchmarks are typically used, failing to consider the real-time impact of dynamic factors on background noise, resulting in insufficient consideration of load fluctuations, temperature and humidity changes, and other operating conditions and environmental variations in abrupt change calculation. In the fault trigger determination stage, fixed judgment thresholds or conventional Gaussian criteria are often used, lacking correction mechanisms for the skewed distribution characteristics of long-tailed noise in the field and the lack of mode completeness under extremely strong noise, resulting in poor adaptability of the trigger threshold and a high rate of false tripping or refusal to trip. In the fault section location stage, reliance on amplitude comparison or inner product operations at a single singular point is often insufficient, failing to effectively utilize the correlation and polarity reversal characteristics of reconstructed waveforms within continuous time windows, leading to insufficient location accuracy for multi-branch sections in complex distribution networks. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as weak leakage current characteristics being easily drowned out by strong noise, insufficient consideration of changes in operating conditions and environment in the calculation of abrupt changes, and poor adaptability of trigger thresholds, which lead to high false tripping and refusal to trip rates and insufficient positioning accuracy in multi-branch sections, this invention proposes a rapid positioning method for leakage current faults on the low-voltage side of distribution transformers.

[0006] This invention provides a method for rapid location of leakage faults on the low-voltage side of a distribution transformer, comprising: S1, collecting zero-sequence current signals from each node on the low-voltage side of the distribution transformer, obtaining intrinsic mode components using variational mode decomposition, using the ratio of kurtosis to energy entropy of each component as a screening factor to remove noise and reconstruct the zero-sequence waveform; calculating the difference between sampling points corresponding to adjacent power frequency cycles to obtain a mutation sequence from the reconstructed zero-sequence waveform, calculating the average mutation value according to the initial sliding window, establishing a background noise prediction model based on load fluctuation and temperature and humidity data, obtaining the current background noise reference value, and subtracting the current background noise reference value from the average mutation value to obtain the mutation value; S2, using a box plot to remove outliers from historical mutation values, calculating the mean and standard deviation, and using skewness correction... The positive coefficient is used to correct the 3σ threshold to obtain the trigger threshold. The ratio of the number of retained components to the preset number is defined as the modal completeness factor. The initial sliding window length is adjusted so that the lower the modal completeness factor, the longer the initial sliding window length. The updated mutation amount is calculated according to the adjusted initial sliding window length. The trigger threshold is corrected using the modal completeness factor. When the updated mutation amount of any node accumulates to a set number of sampling points exceeding the corrected trigger threshold within the preset sliding time window, the trigger criterion is triggered. The set number is determined by the modal completeness factor. S3. The Pearson correlation coefficient is calculated for the reconstructed zero-sequence waveform of the trigger node and adjacent nodes. The segment between adjacent nodes with reversed polarity of the Pearson correlation coefficient is taken as the fault segment and the location result is output.

[0007] This invention can effectively eliminate complex background noise interference to reconstruct weak leakage signals with high signal-to-noise ratio. At the same time, by dynamically and adaptively adjusting the judgment threshold, smoothing scale, and cumulative over-limit point confirmation conditions, it overcomes the masking effect of long-tailed noise distribution and high impedance conditions on feature detection, significantly reducing the false triggering rate and rejection rate of fault judgment. Finally, through the linkage of correlation and polarity reversal, it can quickly and accurately locate the leakage fault section in the multi-branch network.

[0008] Preferably, the step of using the ratio of kurtosis to energy entropy of each component as a screening factor to remove noise and reconstruct the zero-sequence waveform includes: calculating the kurtosis value and energy entropy of each intrinsic mode component after variational mode decomposition; dividing the kurtosis value of each component by the corresponding energy entropy to obtain the screening factor of each component; setting a screening threshold; determining intrinsic mode components with screening factors less than the screening threshold as noise-dominant components and removing them; and linearly superimposing the remaining intrinsic mode components after removing the noise-dominant components to reconstruct the denoised zero-sequence waveform.

[0009] Preferably, the step of calculating the difference between corresponding sampling points of adjacent power frequency cycles to obtain the mutation amount sequence for the reconstructed zero-sequence waveform, and calculating the average mutation amount by calculating the average value according to the initial sliding window, includes: acquiring the reconstructed zero-sequence waveform, extracting the amplitude of each sampling point of the current power frequency cycle according to the sampling rate, and subtracting it from the amplitude of the corresponding sampling point of the previous adjacent power frequency cycle, calculating the absolute value of the difference, and forming the mutation amount sequence; setting the length of the initial sliding window to the number of sampling points corresponding to half a power frequency cycle, sliding the initial sliding window point by point on the mutation amount sequence, and calculating the average value of the absolute values ​​of the mutation amounts of all sampling points in the initial sliding window as the average mutation amount corresponding to the sliding window time.

[0010] Preferably, the step of establishing a background noise prediction model based on load fluctuation and temperature and humidity data, obtaining the current background noise baseline value, and subtracting the current background noise baseline value from the average mutation amount to obtain the mutation amount includes: extracting low-frequency load fluctuation feature sequences and corresponding environmental temperature and humidity time series data from historical operating data to construct an input feature set; constructing a long short-term memory network as a background noise prediction model, and training the model using the input feature set and the average mutation amount under historical normal conditions; inputting the real-time load fluctuation component and real-time temperature and humidity data at the current moment into the trained background noise prediction model, and outputting the background noise baseline value at the current moment; subtracting the current background noise baseline value from the average mutation amount calculated by the current initial sliding window, and setting the result to zero if the subtraction result is less than zero, to obtain the mutation amount after de-baselineing.

[0011] This invention utilizes a long short-term memory network to construct a background noise prediction model, which can accurately predict and deduct dynamic background disturbances caused by temperature and humidity changes and random changes in low-frequency loads in the transformer area, thus avoiding false alarms caused by weak leakage current detection due to simple load changes or temperature and humidity changes.

[0012] Preferably, the step of correcting the 3σ threshold using a skewness correction coefficient to obtain the trigger threshold includes: calculating the skewness coefficient of the cleaned data; multiplying the absolute value of the skewness coefficient by a preset constant weight coefficient, and adding the resulting product to 3 to obtain the threshold multiple after skewness correction; multiplying the threshold multiple by the standard deviation of the cleaned data, and adding the result to the mean of the cleaned data to obtain the trigger threshold.

[0013] This invention employs a threshold calculation scheme that incorporates a skewness correction mechanism, effectively overcoming the shortcomings of the traditional Gaussian 3σ criterion, which uses a fixed multiplier of 3 in noisy environments with typical right-skewed long tail characteristics, resulting in excessively low calculation thresholds. This enables the system to maintain high detection stability in complex power distribution scenarios.

[0014] Preferably, the step of adjusting the initial sliding window length, calculating the updated mutation amount according to the adjusted initial sliding window length, and correcting the trigger threshold using the modal completeness factor, with the set quantity determined by the modal completeness factor, includes: dividing the initial sliding window length by the modal completeness factor and rounding up to update the initial sliding window length; recalculating the average mutation amount and mutation amount according to the updated initial sliding window length; multiplying the trigger threshold by the reciprocal of the modal completeness factor to achieve joint correction of the trigger threshold; and multiplying the base number by the reciprocal of the modal completeness factor and rounding up to obtain the set quantity.

[0015] This invention utilizes a modal completeness factor to dynamically adjust the sliding window length, trigger threshold, and cumulative number of over-limit judgment points, thereby mitigating the risk of feature loss caused by extreme white noise coverage and reducing the risk of incorrect localization due to insufficient effective features.

[0016] Preferably, the step of taking the segment between adjacent nodes with polarity reversal of Pearson correlation coefficient as the fault segment and outputting the location result includes: extracting the reconstructed zero-sequence waveform signal sequence of the trigger node and its upstream and downstream adjacent nodes within a preset time window before and after the triggering time; calculating the Pearson correlation coefficient between the trigger node and the upstream node, and between the trigger node and the downstream node, respectively; if the Pearson correlation coefficient between the trigger node and an adjacent node is less than zero, it is determined that polarity reversal has occurred between the two nodes, and the segment between the adjacent node pair with polarity reversal is taken as the location segment of the leakage fault, generating a location command containing node number and topology segment information and uploading it to the monitoring platform.

[0017] Preferably, the step of extracting the low-frequency load fluctuation feature sequence and the corresponding ambient temperature and humidity time series data from historical operating data to construct the input feature set includes: extracting and synchronously aligning three types of time series data, including the three-phase low-frequency load active power fluctuation feature sequence after low-pass filtering, and the ambient temperature sequence and relative humidity sequence; before inputting the three-dimensional feature quantities into the network, uniformly performing normalization mapping on the three-dimensional feature quantities to a set closed interval as a standardized input feature set.

[0018] Preferably, the step of calculating the mean and standard deviation after removing outliers from historical mutation data using box plots includes: calculating the first quartile, the third quartile, and the interquartile range using the box plot interquartile range method on the historical mutation data sequence; calculating the difference between the first quartile and the interquartile range at a preset multiple as the lower limit, and calculating the sum of the third quartile and the interquartile range at a preset multiple as the upper limit; removing outliers exceeding the upper and lower limits, and calculating the mean and standard deviation of the cleaned data.

[0019] Preferably, the step of calculating the Pearson correlation coefficients of the reconstructed zero-sequence waveform signal sequences between the trigger node and the upstream node, and between the trigger node and the downstream node, includes: for any two nodes for which the Pearson correlation coefficient needs to be calculated, performing a mean-removal operation on their reconstructed zero-sequence waveform signal sequences, calculating the covariance of the two sequences after mean removal; calculating the standard deviation of each of the two sequences, and calculating the product of the two standard deviations; calculating the ratio of the covariance to the product to obtain the dimensionless Pearson correlation coefficient.

[0020] This invention uses the ratio of the product of the mean-removed covariance and the standard deviation to calculate the Pearson correlation coefficient, eliminating the direct influence of the transition resistance value on the determination of the absolute amplitude of the waveform, accurately capturing the opposite trend of the reconstructed zero-sequence waveform on both sides of the fault point, and shortening the location and addressing time.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention collects zero-sequence current signals for variational mode decomposition, combines the ratio of component kurtosis to energy entropy to remove noise and reconstruct waveforms, thereby improving the purity of characteristic signals. It also integrates load fluctuation data and temperature and humidity data to establish a background noise prediction model to deduct background noise, obtains the abrupt change reflecting leakage current abnormalities, and weakens the interference of environmental factors and operating conditions on the signal.

[0023] Furthermore, outliers in historical data are removed using box plots, the basic trigger threshold is corrected using skewness correction coefficients, and the trigger threshold, sliding window length, and number of cumulative sampling points required for judgment are adjusted according to modal completeness factor to construct trigger criteria, thereby reducing the probability of false alarms and missed alarms in fault judgment. By calculating the Pearson correlation coefficient of waveforms of adjacent nodes and identifying polarity reversal sections, the location of leakage fault sections is realized, ensuring the safe and stable operation of the distribution network. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to the present invention; Figure 2 This is a schematic diagram comparing the absolute value sequence of mutations with the sliding window average smoothed envelope sequence in this invention; Figure 3 This is a comparative schematic diagram showing the effect of rapid location of leakage faults on the low-voltage side of the distribution transformer in this invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a method for rapid location of leakage faults on the low-voltage side of a distribution transformer, referring to... Figure 1 This includes steps S1-S3: S1. Acquire zero-sequence current signal and remove noise to extract mutation amount.

[0027] In an optional embodiment, zero-sequence current signals from each node on the low-voltage side of the power distribution are collected, and intrinsic mode components are obtained using a variational mode decomposition algorithm. Noise is eliminated and the zero-sequence waveform is reconstructed using the ratio of the kurtosis to the energy entropy of each component as a screening factor. The difference between the sampling points corresponding to adjacent power frequency cycles is calculated on the reconstructed zero-sequence waveform to obtain the mutation sequence, and the average mutation is obtained by calculating the average value according to the initial sliding window. A background noise prediction model is established based on load fluctuation and temperature and humidity data, the current background noise reference value is obtained, and the mutation is obtained by subtracting the current background noise reference value from the average mutation.

[0028] Specifically, analog current signals are collected using zero-sequence current transformers installed at various nodes on the low-voltage side of the power distribution system. These signals are then discretized into a digital sequence using an analog-to-digital converter at a specific sampling rate, serving as the original zero-sequence current signal. It should be noted that in actual engineering development and deployment, a cloud-edge collaborative architecture is adopted. The underlying node acquisition terminals are only responsible for high-frequency sampling and timestamping the data. The original digital sequence is then uploaded via a high-speed industrial bus to a high-performance edge computing gateway deployed on the power distribution transformer side. This edge computing gateway has a built-in Advanced Reduced Instruction Set Computing (ARPC) processor with a hardware floating-point unit and a neural network acceleration module, providing ample memory and computing power.

[0029] After receiving the raw data, the edge computing gateway inputs the raw zero-sequence current signal into the variational mode decomposition algorithm. By setting the penalty parameter and the number of decomposition layers, the raw signal is decomposed into multiple intrinsic mode components with different center frequencies. The floating-point unit inside the edge computing gateway calculates the kurtosis and energy entropy of each intrinsic mode component, calculates the normalized value of the fourth-order central moment to obtain the kurtosis, calculates the square of each component and its normalized probability distribution, and obtains the energy entropy of each intrinsic mode component using the Shannon information entropy formula. Dividing the calculated kurtosis by the energy entropy yields the selection factor for each intrinsic mode component. A predetermined selection threshold is set, and intrinsic mode components with selection factors less than the predetermined threshold are removed. The remaining intrinsic mode components are retained as preferred components, and the preferred components are accumulated point by point to obtain the reconstructed zero-sequence waveform. For the reconstructed zero-sequence waveform, the value of the current sampling point is subtracted from the value of the corresponding sampling point in the previous power frequency cycle, and the absolute value of the difference is taken to form a mutation sequence.

[0030] An initial sliding window length is set, and a moving average method is used to slide along the mutation sequence, calculating the mean within each window to generate an average mutation sequence. Load fluctuation characteristics, ambient temperature, and relative humidity data of synchronous distribution nodes are collected and used as input features to construct a Long Short-Term Memory (LSTM) network as a background noise prediction model, trained using feature data collected under historical fault-free conditions. The current load fluctuation characteristics and temperature and humidity data are input into the trained LSM network to predict the current background noise baseline value. Each element in the average mutation sequence is subtracted from the corresponding current background noise baseline value; if the difference is less than zero, it is set to zero, resulting in the de-baselined mutation sequence.

[0031] The Long Short-Term Memory (LSTM) network processes time series data. The input layer receives a time series composed of multi-dimensional feature vectors, including synchronously collected load fluctuation characteristics, ambient temperature, and relative humidity. A mathematical relationship for updating time series information is established within the hidden layer. The output layer generates the background noise baseline value for the corresponding time. During model training, it is considered that the inherent leakage current of insulation materials under normal fault-free conditions will generate dynamic background noise fluctuations with changes in ambient temperature, humidity, and load. This inherent leakage current drift caused by insulation aging is a long-term physical phenomenon in the distribution network, constituting a dynamic background noise source during the detection process, which is fundamentally different from sudden leakage faults in the time domain. Subsequently, load fluctuation and temperature / humidity data under historical fault-free conditions are selected as input samples. The background noise level obtained from the average mutation rate at the same time is used as the target label. After normalizing the input features, the mean squared error loss function and the Adam optimization algorithm are used to update each weight matrix and bias term until the loss function converges. This allows the trained LSM network to accurately predict the background noise baseline value under fault-free conditions based on the current operating conditions, thus providing a benchmark for subsequent accurate removal of background disturbances and highlighting of true leakage faults.

[0032] In some embodiments, the ratio of kurtosis to energy entropy of each component is used as a screening factor to remove noise and reconstruct the zero-sequence waveform. Specifically, the kurtosis value and energy entropy of each intrinsic mode component after variational mode decomposition are calculated. The kurtosis value of each component is divided by the corresponding energy entropy to obtain the screening factor of each component. A screening threshold is set, and intrinsic mode components with screening factors less than the screening threshold are identified as noise-dominant components and removed. The remaining intrinsic mode components after removing the noise-dominant components are linearly superimposed to reconstruct the denoised zero-sequence waveform.

[0033] When performing variational mode decomposition on the zero-sequence current signal on the low-voltage side of the power distribution, it is preferable to set the number of decomposition modes K to a range of 5 to 8, and the second-order penalty factor parameter to a range of 1500 to 2500. Taking K equal to 6 and the penalty factor as an example, the signal is decomposed in parallel into 6 independent eigenmode components in each frequency band. to For each intrinsic mode component, the kurtosis value in the time domain and the energy entropy in the frequency domain are extracted. The kurtosis value is mainly used to represent the high-frequency impact and non-Gaussian abrupt change characteristics generated when a zero-sequence current signal experiences a leakage fault, while the energy entropy is used to represent the randomness and complexity of the signal in each frequency band. By dividing the kurtosis value of a single mode component by the energy entropy, a screening factor that can amplify fault characteristics and suppress noise characteristics is obtained.

[0034] The screening threshold is calculated using statistical methods. The arithmetic mean and standard deviation of the screening factors for all components obtained from the current decomposition are calculated. The arithmetic mean minus 0.5 times the standard deviation is set as the floating baseline for the threshold. Simultaneously, 20% of the largest screening factor among all components is used as a hard error prevention limit. The larger value between the floating baseline and the hard error prevention limit is taken as the screening threshold. During the comparison operation, if the calculated screening factor for a component is 1.2, which is lower than the currently determined screening threshold of 1.8, the current component is characterized as an invalid component dominated by Gaussian white noise or environmental steady-state interference and is removed from the reconstructed sequence. The remaining components are then... to The intrinsic mode components with screening factors greater than a threshold are subjected to equal-weighted, point-to-point linear summation on the time-domain sequence, thereby reconstructing a high signal-to-noise ratio zero-sequence waveform that filters out noise-dominated interference and retains transient characteristics.

[0035] In some embodiments, the difference between corresponding sampling points of adjacent power frequency cycles is calculated to obtain a mutation rate sequence for the reconstructed zero-sequence waveform, and the average mutation rate is obtained by calculating the average value according to the initial sliding window. Specifically, the reconstructed zero-sequence waveform is obtained, the amplitude of each sampling point of the current power frequency cycle is extracted according to the sampling rate, and the difference is calculated with the amplitude of the corresponding sampling point of the previous adjacent power frequency cycle. The absolute value of the difference is calculated to form a zero-sequence current mutation rate sequence. The length of the initial sliding window is set to the number of sampling points corresponding to half a power frequency cycle. The initial sliding window is slid point by point on the zero-sequence current mutation rate sequence, and the average value of the absolute values ​​of the mutation rates of all sampling points within the initial sliding window is calculated as the average mutation rate corresponding to the sliding window moment.

[0036] The power frequency standard for the power distribution network system is constant at 50Hz, and the duration of a single power frequency cycle is 20ms. In actual engineering data acquisition, to accurately detect the high-frequency transient details of leakage current, the sampling rate is preferably set to 10kHz, with each complete 20ms power frequency cycle containing 200 discrete sampling points. When constructing the abrupt change sequence, the current amplitude of the i-th sampling point in the current power frequency cycle is extracted, and the amplitude of the i-th minus 200th sampling point in the corresponding phase of the previous cycle is found by traversing along the time axis. The values ​​of the two corresponding points are subtracted and the absolute value is taken to reduce the component of steady-state power frequency periodicity. For example, if the amplitude of the reconstructed waveform at the current point is 1.5A, and the amplitude at the corresponding point 20ms ago is 1.2A, then the absolute value of the abrupt change at the corresponding time is recorded as 0.3A. All absolute values ​​are arranged sequentially to form a one-dimensional zero-sequence current abrupt change sequence.

[0037] Based on a sampling frequency of 10kHz and a half-cycle duration of 10ms, the initial sliding window length is set to 100 sampling points. The 100-data-point sliding window continuously slides across a continuously extending sequence of zero-sequence current abrupt changes, with a fixed step size of one data point. Within each specific interval of the sliding window, the microprocessor accumulates the absolute values ​​of the abrupt changes at all 100 sampling points within the window and divides them by 100. The resulting arithmetic mean is assigned to the end or center point of the current sliding window as the corresponding average abrupt change output, achieving smooth low-pass filtering in the time domain and preserving the macroscopic envelope characteristics of the abrupt change energy at the moment of low-voltage side leakage. To demonstrate the effectiveness of abrupt change extraction and sliding window smoothing, a comparison is given between the absolute value sequence of abrupt changes and the average smoothed envelope sequence of the sliding window under typical operating conditions.

[0038] Reference Figure 2 The scatter plot representing the absolute value of the mutation exhibits violent high-frequency oscillations and numerous spikes throughout the entire sampling point index range, especially in the range of sampling point index positions 100 to 200, where transient fluctuation values ​​jump sharply and reach high amplitudes. In contrast, the smooth curve representing the envelope sequence after sliding window averaging clearly filters out the high-frequency spikes and random jumps in the scatter plot, closely following the overall trend of the scatter plot and forming a smooth upward and downward trajectory. Based on the smooth curve's ability to remove high-frequency oscillations while retaining the overall trend, it is proven that using a sliding average method to calculate along the mutation sequence can effectively achieve low-pass filtering and retain the macroscopic envelope characteristics of the instantaneous mutation energy during leakage.

[0039] In some embodiments, a noise prediction model is established based on load fluctuation and temperature and humidity data to obtain the current background noise baseline value. The mutation amount is obtained by subtracting the current background noise baseline value from the average mutation amount. Specifically, low-frequency load fluctuation feature sequences and corresponding environmental temperature and humidity time series data are extracted from historical operating data to construct an input feature set. A long short-term memory network is constructed as a background noise prediction model. The model is trained using the input feature set and the average mutation amount under historical normal conditions. The real-time load fluctuation component and real-time temperature and humidity data at the current moment are input into the trained background noise prediction model, and the current background noise baseline value is output. The average mutation amount calculated by the initial sliding window is subtracted from the current background noise baseline value. If the result of the subtraction is less than zero, it is set to zero to obtain the de-baselined mutation amount.

[0040] Multi-dimensional operational monitoring data from the past 30 to 90 days were extracted from the historical database of the target distribution substation. To construct a multi-dimensional input feature set, three types of time-series data were extracted and synchronously aligned: a three-phase low-frequency load active power fluctuation feature sequence (kW / s) after low-pass filtering from 0 to 5 Hz; an ambient temperature sequence (-20℃ to 60℃) and a relative humidity sequence (0%RH to 100%RH) provided by environmental monitors deployed in the distribution cabinets. Before inputting the three-dimensional features into the network, the three-dimensional features were uniformly normalized to a closed interval of 0 to 1 using Min-Max normalization. The inference operations of the Long Short-Term Memory (LSTM) network were performed using the neural network processing unit of a high-performance edge computing gateway, effectively avoiding the problem of insufficient memory space at the underlying acquisition nodes.

[0041] This paper establishes a mathematical relationship for updating time-series information based on Long Short-Term Memory (LSTM) networks. The LTM network model takes a standardized time-series vector composed of low-frequency load fluctuation characteristics, ambient temperature, and relative humidity as input. The input vector for each time step is denoted as... The hidden state of the previous time step is recorded as The cell state at the previous time step is recorded as Long Short-Term Memory (LSTM) networks update temporal information through input gates, forget gates, output gates, and candidate cell states. The input gate is calculated as follows:

[0042] in, For input gates; For activation functions; The input weight matrix; The input vector; This is a cyclic weight matrix; This is the hidden state from the previous time step; This is a bias term.

[0043] Computational forget gate:

[0044] in, Forgotten Gate; For activation functions; The input weight matrix; The input vector; This is a cyclic weight matrix; This is the hidden state from the previous time step; This is a bias term.

[0045] Calculate candidate cell states:

[0046] in, Candidate cell state; It is the hyperbolic tangent function; The input weight matrix; The input vector; This is a cyclic weight matrix; This is the hidden state from the previous time step; This is a bias term.

[0047] Calculate cell state:

[0048] in, In cellular state; Forgotten Gate; It is the element-wise multiplication operator; This represents the cell state at the previous time step; For input gates; This represents the candidate cell state.

[0049] Calculate the output gate:

[0050] in, For output gate; For activation functions; The input weight matrix; The input vector; This is a cyclic weight matrix; This is the hidden state from the previous time step; This is a bias term.

[0051] Calculate the hidden state:

[0052] in, It is in a hidden state; For output gate; It is the element-wise multiplication operator; It is the hyperbolic tangent function; It is in the cellular state.

[0053] The hidden state of the last time step is input into the linear output layer to obtain the baseline value of the background noise at the current time step. During model training, the average mutation amount under historical fault-free states is used as the target label. The mean squared error loss function and the Adam optimization algorithm are used to update each weight matrix and bias term until the loss function converges.

[0054] In actual online monitoring, if the current ambient temperature is 35℃ and the relative humidity is 85%, and there is an instantaneous load fluctuation of 0.5kW / s, the Long Short-Term Memory (LSTM) network model outputs a predicted value of 0.025A for the current background noise baseline through forward inference. If the actual average mutation amount obtained from the sliding window calculation is 0.06A, a de-benchmarking operation is performed, i.e., 0.06A minus 0.025A, resulting in a mutation amount of 0.035A. If the result of the subtraction operation is negative, the difference is truncated and assigned a value of 0, thus avoiding false alarms caused by simple load changes or drastic temperature and humidity fluctuations due to weak leakage current detection.

[0055] S2. Calculate the trigger threshold and adjust the sliding window to trigger the criterion.

[0056] In an optional embodiment, the mean and standard deviation of historical mutations are calculated after outlier removal using a box plot. The skewness correction coefficient is used to correct the 3σ threshold to obtain the trigger threshold. The ratio of the number of retained components to a preset number is defined as the modal completeness factor. The initial sliding window length is adjusted so that the lower the modal completeness factor, the longer the initial sliding window length. The updated mutation is calculated according to the adjusted initial sliding window length. The trigger threshold is then corrected using the modal completeness factor. When the updated mutation of any node accumulates to a set number of sampling points exceeding the trigger threshold within a preset sliding time window, the trigger criterion is triggered. The set number is determined by the modal completeness factor.

[0057] Collect mutation sequence data recorded under historical fault-free conditions. Calculate the first quartile corresponding to the 25% threshold and the third quartile corresponding to the 75% threshold. Define the difference between the first and third quartiles as the interquartile range. Remove outliers that are less than the first quartile minus 1.5 times the interquartile range, or greater than the third quartile plus 1.5 times the interquartile range. Calculate the mean and standard deviation of the data after outlier removal. Calculate the skewness value of the historical mutation data. Multiply the preset constant weighting coefficient by the absolute value of the skewness coefficient and add it to 3 to obtain the skewness-corrected warning factor. Add the product of the warning factor and the standard deviation to the mean of the cleaned data to obtain the basic trigger threshold. Count the number of retained components not removed in the previous steps. Divide the number of components by the preset total number of decomposition layers in the variational mode decomposition to obtain the mode completeness factor. A function inversely proportional to the modal completeness factor is constructed for the initial sliding window length. The closer the modal completeness factor is to zero, the larger the calculated initial sliding window length value is. The calculated initial sliding window length value is rounded up and the number of sample points in the sliding window is updated to adjust the initial sliding window length to smooth the mutation amount. The average mutation amount is recalculated according to the updated sliding window. The recalculated average mutation amount is subtracted from the current background noise baseline value at the corresponding time. If the result of the subtraction is less than zero, it is set to zero to obtain the updated mutation amount sequence.

[0058] The final trigger threshold is obtained by multiplying the base trigger threshold by the reciprocal of the modal completeness factor. An inverse proportional relationship is established between the set number and the modal completeness factor, such that a smaller modal completeness factor corresponds to a larger set number, rounded up to an integer number of sampling points. The updated mutation sequence of each node is compared in real-time with the calculated trigger threshold. When the updated mutation value of any node exceeds the trigger threshold, and the cumulative number of sampling points satisfying the condition of exceeding the trigger threshold reaches a set number within a preset sliding time window, a Boolean truth value trigger fault criterion is output.

[0059] In some embodiments, after removing outliers using a box plot, the mean and standard deviation of historical mutation data are calculated. A skewness correction coefficient is then used to correct the 3σ threshold to obtain the trigger threshold. Specifically, the first quartile, third quartile, and interquartile range are calculated using a box plot quartile method on the historical mutation data sequence. Outliers exceeding the upper and lower boundaries are removed. The mean and standard deviation of the cleaned data are calculated, and the skewness coefficient of the data distribution is calculated. When calculating the trigger threshold, the absolute value of the preset constant weighting coefficient and the skewness coefficient are multiplied, and the resulting product is added to 3 to obtain the skewness-corrected threshold multiple. Then, the threshold multiple is multiplied by the standard deviation of the cleaned data, and the result is added to the mean of the cleaned data to obtain the trigger threshold.

[0060] For the in-depth cleaning and parameter tuning of the background data, the continuous abrupt change time series recorded under steady-state operating conditions over the past 24 hours for the target transformer area was extracted. This sample set of continuous abrupt change time series contains tens of thousands of data points. The dataset of continuous abrupt change time series was sorted in ascending order of numerical value, and based on mathematical statistics, the first quartile located at the 25% cumulative probability distribution position was calculated using the quantile function. And the third quartile located at the 75th percentile And find the first quartile. With the third quartile The difference is used as the interquartile range The specific calculation method is as follows:

[0061] in, Interquartile range; It is the third quartile; It is the first quartile.

[0062] To filter out atypical transient pulses such as those caused by circuit breaker tripping, a reasonable data reception interval was defined. All values ​​falling outside the closed data reception interval were identified as invalid outliers and deleted, thereby obtaining a high-confidence cleaned sequence that meets statistical significance.

[0063] Specifically, calculate the upper and lower limits of the receiving interval: ;

[0064] in, This is the upper limit value; This is the lower limit value; It is the third quartile; It is the first quartile; It is the interquartile range.

[0065] The cleaned sequence is subjected to mathematical statistical characterization, and the expected mean and standard deviation represented by the square root of the variance are calculated. The skewness coefficient representing the distribution symmetry is then determined based on the third-order central moment formula. Due to the nonlinear cumulative characteristics of capacitive leakage current in field equipment, the data distribution during normal operation typically exhibits a right-skewed long-tail characteristic. When calculating the trigger threshold, a preset constant weighting coefficient is input, preferably set within the range of 1.5 to 2.5; here, 2 is used as an example.

[0066] Based on the skewness correction mechanism, a formula for calculating the trigger threshold is established. The basic trigger threshold is calculated as follows: ;

[0067] in, This is the warning factor after skewness correction; These are preset constant weighting coefficients; This is the skewness coefficient; Based on the trigger threshold; This represents the mean of the cleaned data. This represents the standard deviation of the cleaned data.

[0068] Assuming the measured mean after cleaning is 0.035A, the standard deviation is 0.012A, and the skewness coefficient is 1.5, the calculated basic trigger threshold is 0.107A. This threshold calculation scheme, incorporating a skewness correction mechanism, overcomes the deficiency of excessively low thresholds in long-tailed noise environments under the traditional Gaussian 3σ criterion with a fixed multiplier of 3, enabling the system to maintain high detection stability in complex power distribution scenarios.

[0069] In some embodiments, the ratio of the number of retained components to a preset number is defined as the modal completeness factor. The initial sliding window length is adjusted so that the lower the modal completeness factor, the longer the initial sliding window length. The updated mutation amount is calculated according to the adjusted initial sliding window length. The trigger threshold is then corrected using the modal completeness factor. When the updated mutation amount of any node accumulates to a set number of sampling points exceeding the trigger threshold within a preset sliding time window, the trigger criterion is triggered. The set number is determined by the modal completeness factor. Specifically, the number of intrinsic modal components retained during reconstruction is obtained. If the number of components is less than 1, it is taken as 1; otherwise, the number of intrinsic modal components obtained is taken. The obtained value is divided by the preset value of the total number of variational mode decomposition layers to obtain the modal completeness factor. The initial sliding window length is divided by the modal completeness factor and rounded up to update the initial sliding window length. The average mutation amount is recalculated according to the updated initial sliding window length. The updated mutation amount is obtained by de-benchmarking. The updated mutation amount is used as the comparison object of the trigger criterion. The trigger threshold is multiplied by the reciprocal of the modal completeness factor to achieve joint correction of the trigger threshold. The base number is multiplied by the reciprocal of the modal completeness factor and rounded up to obtain the set number.

[0070] Based on the modal completeness factor, establish the mathematical relationships for parameter tuning. Calculate the updated initial sliding window length, final trigger threshold, and set quantity. ; ;

[0071] in, This is the updated initial sliding window length; The rounding up symbol; This is the initial sliding window length; This is the modal completeness factor; This is the final trigger threshold; Based on the trigger threshold; To set the quantity; The base points.

[0072] The actual number of intrinsic modal components retained after threshold screening during the reconstruction process is counted. To ensure the mathematical continuity of subsequent parameter adjustment calculation logic and prevent abnormal calculations due to division by zero, a minimum constraint mechanism is set. When the number of retained components is detected to be less than 1, the number of retained components is forced to be equal to 1; if the number of retained components is greater than or equal to 1, the original count value is maintained. Assuming that the preset total number of variational mode decomposition layers is fixed at 6 layers during initialization, and in a certain leakage transient analysis, only 3 components that can represent the characteristics remain, the obtained value of 3 is divided by the total number of layers 6 to calculate the current modal completeness factor value as 0.5. Since a smaller modal completeness factor indicates a lower degree of effective mode retention, the sliding window length, trigger threshold, and cumulative number of out-of-limit points are then adjusted synchronously.

[0073] The first step in parameter tuning based on the obtained completeness factor is to update the scale of the data smoothing window. The initial sliding window, originally set to 100 sampling points, is divided by a factor of 0.5 and rounded up, instantly increasing the new sliding window length to 200 sampling points. A wider integration interval is used to smoothly calculate the new average mutation rate sequence. The new average mutation rate sequence is then subtracted from the current background noise baseline value at the corresponding time. If the result is less than zero, it is set to zero, resulting in the updated mutation rate sequence, which serves as the input for subsequent limit comparisons. The basic trigger threshold obtained in the aforementioned process, for example 0.107A, is multiplied by the reciprocal of the factor by 2, adjusting the final trigger threshold to 0.214A. The fault tolerance limit confirmation condition is adjusted, setting the basic number of points required to determine a fault to 15. The basic number of points is multiplied by the reciprocal of 2 and rounded up, expanding the required cumulative limit exceedance within the sliding window to 30 points. When the available modal frequency band is scarce due to strong environmental interference, the leakage current waveform is required to have a higher amplitude and a longer duration before an alarm command is issued, in order to reduce the risk of fault mislocation due to insufficient characteristics.

[0074] S3. Calculate the correlation coefficient to locate the faulty section.

[0075] In an optional embodiment, the Pearson correlation coefficient is calculated on the reconstructed zero-sequence waveforms of the trigger node and adjacent nodes. The segment between adjacent nodes where the polarity of the Pearson correlation coefficient is reversed is taken as the fault segment and the location result is output.

[0076] Specifically, after the fault criterion is triggered, under the conditions that the sampling time of each node is synchronized and the installation polarity direction of the zero-sequence current transformer is uniform, the reconstructed zero-sequence waveform data of the trigger node and the upstream and downstream nodes connected to the trigger node are extracted according to the electrical topology of the distribution network. The data time window length for correlation analysis is set to a preset time window before and after the trigger moment. The Pearson correlation coefficient between the reconstructed zero-sequence waveform of the trigger node and the reconstructed zero-sequence waveform of the upstream node, and the reconstructed zero-sequence waveform of the trigger node and the downstream node are calculated respectively within the time window.

[0077] Based on the calculation of sequence correlation in mathematical statistics, a Pearson correlation coefficient is established. The mean is removed from two time-synchronized reconstructed zero-sequence waveform signal sequences. The ratio of the covariance of the two mean-removed sequences to the product of their standard deviations is calculated to obtain the dimensionless Pearson correlation coefficient.

[0078] in, The Pearson correlation coefficient; Let covariance be the variance of the two reconstructed zero-sequence waveform signal sequences; The standard deviation of the first reconstructed zero-sequence waveform signal sequence; The standard deviation of the second reconstructed zero-sequence waveform signal sequence.

[0079] The polarity of the calculated Pearson correlation coefficients is determined. If the correlation coefficient between the triggering node and the upstream node is greater than zero and the correlation coefficient between the triggering node and the downstream node is less than zero, it indicates that the polarity has reversed between the triggering node and the downstream node, and the power distribution line section between the triggering node and the downstream node is determined to be the faulty section. If the correlation coefficient between the triggering node and the upstream node is less than zero, it indicates that the polarity has reversed between the triggering node and the upstream node, and the power distribution line section between the upstream node and the triggering node is determined to be the faulty section. After determining the faulty section, the device numbers of the first and last nodes of the faulty section and the latitude and longitude information in the geographic coordinate system are packaged and printed as characters through the standard output interface of the console, or sent to the human-machine interface front end to render the line segment as a specific graphical highlight display, thereby completing the output operation of the positioning result.

[0080] In some embodiments, Pearson correlation coefficients are calculated for the reconstructed zero-sequence waveforms of the trigger node and adjacent nodes. The segment between adjacent nodes where the polarity of the Pearson correlation coefficient is reversed is taken as the fault segment, and the location result is output. Specifically, when the updated mutation amount of any node accumulates to a set number of sampling points exceeding the trigger threshold within a preset sliding time window, the reconstructed zero-sequence waveform signal sequences of the trigger node and its upstream and downstream adjacent nodes within the preset time window before and after the trigger time are extracted. The Pearson correlation coefficients of the reconstructed zero-sequence waveform signal sequences between the trigger node and the upstream node, and between the trigger node and the downstream node are calculated respectively. If the Pearson correlation coefficient between the trigger node and an adjacent node is less than zero, it is determined that a polarity reversal has occurred between the two nodes. The segment between the adjacent node pairs where the polarity reversal has occurred is taken as the location segment of the leakage fault. A location command containing the node number and topology segment information is generated and uploaded to the monitoring platform.

[0081] In a distributed synchronous measurement network comprised of multiple intelligent terminals in a distribution station area, to meet the computational power requirements of large-scale matrix and correlation analysis, the calculation of the Pearson correlation coefficient and topology addressing are centrally processed by the upper-layer IoT cloud platform or the main station-level main control chip computing pool. When it is detected that in the updated mutation sequence generated by a core node, such as node B, 30 data points (i.e., the corrected set number of data points) have amplitudes exceeding the corrected trigger threshold of 0.214A within a preset sliding time window, a microsecond-level synchronization timestamp will be added at the point of exceeding the limit. Using this high-precision time anchor as the core reference, the time span is truncated forward and backward to extract a preset time window sequence containing the fault development process, preferably extracting a data segment from 20ms ahead of the trigger time to 40ms behind the trigger time. At a sampling frequency of 10kHz, a reconstructed one-dimensional array of transient zero-sequence current containing 600 discrete numerical sequence points is assembled. Simultaneously, the power distribution IoT platform utilizes a geographic information topology model for rapid address lookup, retrieving the upstream power supply node (node ​​A) and the downstream load-side node (node ​​C) adjacent to node B along the power feeder. Under strict time synchronization alignment, the zero-sequence waveforms of nodes A and C, each containing a corresponding time window of 600 points, are extracted. The reconstructed zero-sequence waveform signal sequences of nodes B and A are mean-triggered, and the ratio of the covariance to the product of the standard deviations of the two mean-triggered sequences is calculated to obtain the Pearson correlation coefficient between nodes B and A. The same steps are then used to calculate the Pearson correlation coefficient between nodes B and C.

[0082] Based on the principle that a ground leakage path is formed at the fault point when leakage occurs, altering the zero-sequence current distribution, under conditions of uniform measurement polarity and determined topology direction, the reconstructed zero-sequence waveforms on both sides of the fault section may exhibit opposite trends. If the calculated correlation coefficient between node B and its superior node A is 0.92, it indicates that node B and its superior node A have the same direction and polarity. However, the calculated value between node B and its subordinate node C is -0.85, satisfying the condition of being less than zero. This indicates that the current polarity reverses between node B and node C. The line section between node B and node C is then identified as the location of the leakage fault. A formatted location command message is generated according to the IEC 61850 protocol. The location command message contains detailed trigger times and a textualized topology section from distribution transformer branch B cabinet to terminal meter box C. This message is then reported to the power company's main station via an encrypted 4G channel or dedicated fiber optic channel to guide emergency repairs.

[0083] During the R&D phase, a 10kV to 0.4kV low-voltage distribution network simulation and verification system was built using high-performance edge computing devices and intelligent sensors. Five test nodes and ten distribution feeders were deployed, with the underlying data sampling frequency set at 10000Hz. The core algorithms in the system all run on edge architecture servers with ample computing resources. Under simulated ambient temperatures of 35℃ and relative humidity of 85%, a continuous injection of 0.5kW of active load fluctuations per second and strong background white noise with a signal-to-noise ratio of 5dB was injected into the network. During a 30-day continuous online monitoring test, the system applied 500 instances of single-phase high-resistance weak leakage faults with a transition resistance as high as 1500 ohms at different spatial topology locations and random time points. Large nonlinear electrical equipment was used to simulate and record 2000 strong interference events that could induce transient overvoltage and current fluctuations, thereby constructing a comprehensive test sample library.

[0084] Control group 1 employs conventional empirical mode decomposition (EMD) denoising techniques combined with a traditional fault triggering determination scheme using fixed mean and fixed 3σ threshold. Control group 2 utilizes variational mode decomposition and kurtosis to energy entropy ratio filtering denoising techniques, but completely eliminates the long short-term memory network background noise prediction model and the mode completeness factor co-tuning mechanism. Experimental group 3 employs a complete, fully modular, interconnected low-voltage side zero-sequence current leakage fault location scheme. The three different schemes are deployed in parallel within a unified main control chip computing pool, independently processing reconstructed zero-sequence waveforms of the same time dimension and outputting topology location logs.

[0085] Experimental group 3 accurately located 492 real leakage faults, achieving a fault detection accuracy of 98.4%. In 2000 high-frequency nonlinear load interference tests, it only produced 3 false alarms, a false alarm rate of 0.15%, with an average network topology location time of 45ms. Control group 1 accurately located faults only 410 times, achieving a fault detection accuracy of 82%, but produced as many as 156 false alarms, a false alarm rate of 7.8%, with an average location time of 72ms. Control group 2 accurately located faults 455 times, increasing the accuracy to 91%. However, due to the lack of a temperature and humidity-based prediction subtraction and adjustment mechanism, it accumulated 42 false alarms in the range of frequent and sudden load changes, a false alarm rate of 2.1%, and reduced the average location time to 55ms.

[0086] Reference Figure 3 The bars representing the accurate detection and positioning ratio show an increasing trend in the conventional empirical algorithm group, the truncated comparison control group, and the linkage experimental test group, with the value of the bar representing the accurate detection and positioning ratio in the linkage experimental test group approaching 100%. Simultaneously, the bars representing the proportion of failures induced by high impedance and strong noise environments show a decreasing trend in the conventional empirical algorithm group, the truncated comparison control group, and the linkage experimental test group, with the value of the bar representing the proportion of failures induced by high impedance and strong noise environments in the linkage experimental test group approaching zero. Based on the data trends showing the highest accurate detection and positioning ratio and the lowest proportion of failures induced by high impedance and strong noise environments in the linkage experimental test group, it is demonstrated that the complete full-module linkage power distribution low-voltage side zero-sequence current leakage fault location scheme exhibits high stability and good fault identification effect under high impedance and background noise interference environments.

[0087] The Long Short-Term Memory (LSTM) network prediction model can predict and subtract background disturbances caused by temperature and humidity fluctuations and random changes in low-frequency loads in the distribution area. Combined with a non-Gaussian skewness threshold, the false alarm rate under strong interference is reduced compared to conventional fixed methods. Based on a mechanism that links the modal completeness factor to increase the smoothing window scale and the cumulative number of points exceeding the limit required for fault confirmation, the risk of feature loss caused by extreme white noise coverage is mitigated to some extent, and the risk of incorrect location due to insufficient effective features is reduced. The polarity discrimination method using the Pearson correlation coefficient shortens the addressing time and improves the sensitivity of leakage fault detection and location in the low-voltage distribution network.

[0088] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for rapid location of leakage faults on the low-voltage side of a distribution transformer, characterized in that, include: S1. Collect zero-sequence current signals from each node on the low-voltage side of the power distribution system. Use variational mode decomposition to obtain intrinsic mode components. Use the ratio of kurtosis to energy entropy of each component as a screening factor to remove noise and reconstruct the zero-sequence waveform. Calculate the difference between sampling points corresponding to adjacent power frequency cycles on the reconstructed zero-sequence waveform to obtain the mutation sequence. Calculate the average mutation value using the initial sliding window. Establish a background noise prediction model based on load fluctuations and temperature / humidity data. Obtain the current background noise baseline value. Subtract the current background noise baseline value from the average mutation value to obtain the mutation amount. S2. Use a box plot to remove outliers from historical mutation amounts and calculate the mean and standard deviation. Use a skewness correction coefficient to correct the 3σ threshold to obtain the trigger threshold. The ratio of the number of retained components to the preset number is defined as the modal completeness factor. The initial sliding window length is adjusted so that the lower the modal completeness factor, the longer the initial sliding window length. The updated mutation amount is calculated according to the adjusted initial sliding window length. The trigger threshold is corrected using the modal completeness factor. When the updated mutation amount of any node accumulates to a set number of sampling points exceeding the corrected trigger threshold within the preset sliding time window, the trigger criterion is triggered. The set number is determined by the modal completeness factor. S3. The Pearson correlation coefficient is calculated for the reconstructed zero-sequence waveform of the trigger node and adjacent nodes. The segment between adjacent nodes with reversed polarity of the Pearson correlation coefficient is taken as the fault segment and the location result is output.

2. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The step of using the ratio of kurtosis to energy entropy of each component as a screening factor to remove noise and reconstruct the zero-sequence waveform includes: calculating the kurtosis value and energy entropy of each intrinsic mode component after variational mode decomposition; dividing the kurtosis value of each component by the corresponding energy entropy to obtain the screening factor of each component; setting a screening threshold; determining intrinsic mode components with screening factors less than the screening threshold as noise-dominant components and removing them; and linearly superimposing the remaining intrinsic mode components after removing the noise-dominant components to reconstruct the denoised zero-sequence waveform.

3. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The step of calculating the difference between corresponding sampling points of adjacent power frequency cycles to obtain the mutation amount sequence and calculating the average mutation amount according to the initial sliding window includes: acquiring the reconstructed zero-sequence waveform, extracting the amplitude of each sampling point of the current power frequency cycle according to the sampling rate, and subtracting it from the amplitude of the corresponding sampling point of the previous adjacent power frequency cycle, calculating the absolute value of the difference to form the mutation amount sequence; setting the length of the initial sliding window to the number of sampling points corresponding to half a power frequency cycle, sliding the initial sliding window point by point on the mutation amount sequence, and calculating the average value of the absolute value of the mutation amount of all sampling points in the initial sliding window as the average mutation amount corresponding to the sliding window time.

4. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The process of establishing a background noise prediction model based on load fluctuation and temperature and humidity data, obtaining the current background noise baseline value, and subtracting the current background noise baseline value from the average mutation value to obtain the mutation value includes: extracting low-frequency load fluctuation feature sequences and corresponding environmental temperature and humidity time series data from historical operating data to construct an input feature set; constructing a long short-term memory network as a background noise prediction model, and training the model using the input feature set and the average mutation value under historical normal conditions; inputting the real-time load fluctuation component and real-time temperature and humidity data at the current moment into the trained background noise prediction model, and outputting the current background noise baseline value; subtracting the current background noise baseline value from the average mutation value calculated by the initial sliding window, and setting the result to zero if the subtraction result is less than zero, to obtain the mutation value after de-baselineing.

5. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The step of using a skewness correction coefficient to correct the 3σ threshold to obtain the trigger threshold includes: calculating the skewness coefficient of the cleaned data; multiplying the absolute value of the skewness coefficient by a preset constant weight coefficient, and adding the resulting product to 3 to obtain the skewness-corrected threshold multiple; multiplying the skewness-corrected threshold multiple by the standard deviation of the cleaned data, and adding the result to the mean of the cleaned data to obtain the trigger threshold.

6. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The process of adjusting the initial sliding window length, calculating the updated mutation amount based on the adjusted initial sliding window length, and correcting the trigger threshold using the modal completeness factor, with the set quantity determined by the modal completeness factor, includes: dividing the initial sliding window length by the modal completeness factor and rounding up to update the initial sliding window length; recalculating the average mutation amount and mutation amount based on the updated initial sliding window length; multiplying the trigger threshold by the reciprocal of the modal completeness factor to achieve joint correction of the trigger threshold; and multiplying the base number by the reciprocal of the modal completeness factor and rounding up to obtain the set quantity.

7. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The step of taking the segment between adjacent nodes with reversed Pearson correlation coefficients as the fault segment and outputting the location result includes: extracting the reconstructed zero-sequence waveform signal sequence of the trigger node and its upstream and downstream adjacent nodes within a preset time window before and after the triggering time; calculating the Pearson correlation coefficients between the trigger node and the upstream node, and between the trigger node and the downstream node, respectively; if the Pearson correlation coefficient between the trigger node and an adjacent node is less than zero, it is determined that a polarity reversal has occurred between the two nodes, and the segment between the adjacent node pairs with reversed polarity is taken as the location segment of the leakage fault, generating a location command containing node number and topology segment information and uploading it to the monitoring platform.

8. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 4, characterized in that, The process of extracting low-frequency load fluctuation feature sequences and corresponding ambient temperature and humidity time series data from historical operating data to construct an input feature set includes: extracting and synchronously aligning three types of time series data, including the three-phase low-frequency load active power fluctuation feature sequence after low-pass filtering, as well as the ambient temperature sequence and relative humidity sequence; before inputting the three-dimensional feature quantities into the network, the three-dimensional feature quantities are uniformly normalized and mapped to a set closed interval as a standardized input feature set.

9. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 1, characterized in that, The calculation of the mean and standard deviation after removing outliers using box plots for historical mutation data includes: calculating the first quartile, third quartile, and interquartile range using the box plot interquartile range method for the historical mutation data sequence; calculating the difference between the first quartile and the interquartile range at a preset multiple as the lower limit, and calculating the sum of the third quartile and the interquartile range at a preset multiple as the upper limit; removing outliers exceeding the upper and lower limits, and calculating the mean and standard deviation of the cleaned data.

10. The method for rapid location of leakage faults on the low-voltage side of a distribution transformer according to claim 7, characterized in that, The step of calculating the Pearson correlation coefficients between the trigger node and the upstream node, and between the trigger node and the downstream node for the reconstructed zero-sequence waveform signal sequences, includes: for any two nodes for which the Pearson correlation coefficient needs to be calculated, performing a mean-removal operation on their reconstructed zero-sequence waveform signal sequences, and calculating the covariance of the two sequences after mean removal; calculating the standard deviation of each of the two sequences, and calculating the product of the two standard deviations; and calculating the ratio of the covariance to the product to obtain the dimensionless Pearson correlation coefficient.

Images