Short-circuit sensing processing method driven by the electrical characteristics of low-impedance voltage transformers

Abstract

This application provides a short-circuit sensing processing method based on the electrical characteristics of a low-impedance voltage transformer, belonging to the field of short-circuit sensing technology. The method includes: interacting with a target voltage transformer to acquire transient voltage fluctuation data; acquiring a list of feature dimensions, and performing multi-dimensional feature extraction based on the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features; constructing a fused short-circuit discrimination model; inputting the multi-dimensional voltage fluctuation features and the list of feature dimensions into the fused short-circuit discrimination model to obtain short-circuit sensing categories and their corresponding category confidence scores; based on a sensing confidence score threshold, filtering category confidence scores to determine confident short-circuit sensing categories, normalizing the category confidence scores of multiple confident short-circuit sensing categories, and outputting them as short-circuit sensing information for short-circuit processing. This method solves the technical problems of untimely and inaccurate short-circuit sensing caused by the dual issues of hardware response lag and weak anti-interference due to single criteria in existing short-circuit sensing methods.

Description

Technical Field

[0001] This invention relates to the field of short-circuit sensing, and more particularly to a short-circuit sensing processing method based on the electrical characteristics of a low-impedance voltage transformer. Background Technology

[0002] In power systems, the instantaneous large current and sudden voltage drop during a short circuit can directly impact critical equipment such as transformers and circuit breakers. Therefore, timely detection of short circuit faults can effectively curb the spread of faults, reduce economic losses, and ensure the continuity of power supply and equipment safety.

[0003] However, traditional short-circuit sensing methods typically rely on high-impedance voltage transformers to collect voltage signals. These transformers have high internal resistance, resulting in a delayed response to microsecond-level transient voltage fluctuations at the initial stage of a short circuit. This often leads to missed detection of critical features within 5-10 μs after a fault occurs, causing a delay in short-circuit detection. Furthermore, existing discrimination methods often use single voltage features to construct criteria, failing to consider interference from complex power grid conditions, leading to frequent false positives and false negatives. Therefore, traditional short-circuit sensing methods suffer from both delayed hardware response and weak interference immunity due to their reliance on single criteria, resulting in untimely and inaccurate short-circuit sensing.

[0004] Therefore, there is an urgent need for an intelligent discrimination processing method that combines sensitive hardware characteristics with multi-dimensional feature fusion. Summary of the Invention

[0005] This invention addresses the technical problems of untimely and inaccurate short-circuit sensing in existing short-circuit sensing methods due to both hardware response lag and weak interference immunity caused by single criteria. It provides a short-circuit sensing processing method based on the electrical characteristics driven by a low-impedance voltage transformer.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer, comprising:

[0008] Interact with the target voltage transformer to acquire transient voltage fluctuation data, wherein the transient voltage fluctuation data includes primary side and secondary side;

[0009] Obtain a list of feature dimensions and perform multi-dimensional feature extraction in conjunction with the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions.

[0010] A fusion short-circuit discrimination model is constructed, wherein the fusion short-circuit discrimination model includes multiple univariate short-circuit sensing channels and is integrated and output through an integration layer;

[0011] Input the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence.

[0012] Based on the perception confidence threshold, the category confidence is filtered to determine the confidence short-circuit perception category. The category confidence of multiple confidence short-circuit perception categories is normalized and output as short-circuit perception information for short-circuit processing.

[0013] The beneficial effects of this invention are:

[0014] Compared to existing technologies, this application first interacts with the target voltage transformer to acquire transient voltage fluctuation data, leveraging the transient sensitivity advantage of the low-impedance voltage transformer and providing a reliable data foundation for subsequent multi-dimensional feature extraction and fusion model discrimination. Secondly, it obtains a list of feature dimensions and performs multi-dimensional feature extraction in conjunction with the transient voltage fluctuation data, transforming the raw transient voltage fluctuation data into a feature language understandable by the model, providing reliable data input for subsequent steps. Thirdly, it constructs a fusion short-circuit discrimination model, which includes multiple univariate short-circuit sensing channels and integrates the outputs through an integration layer, resulting in a fusion short-circuit discrimination model that integrates the advantages of multiple features and has strong anti-interference capabilities, providing reliable support for subsequent accurate short-circuit discrimination. Furthermore, it inputs the multi-dimensional voltage fluctuation features and the list of feature dimensions into the fusion short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence, transforming the multi-dimensional voltage fluctuation features into interpretable and high-precision short-circuit discrimination conclusions, solving the pain points of poor anti-interference and low accuracy of traditional single-feature discrimination. Finally, based on the perception confidence threshold, the category confidence is filtered to determine the confidence short-circuit perception category. The category confidence of multiple confidence short-circuit perception categories is normalized and output as short-circuit perception information. Short-circuit processing is performed to obtain reliable and interpretable short-circuit perception information, avoid misjudgment caused by low confidence results, and provide a clear basis for short-circuit processing.

[0015] Through the above technical solution, this application first obtains transient voltage fluctuation data of the primary and secondary sides from the target voltage transformer to ensure that key fault characteristics are captured in a timely manner. Then, it constructs a fusion short-circuit discrimination model composed of multiple univariate short-circuit sensing channels and an integrated layer to avoid the risk of misjudgment by a single channel. Next, it inputs the multi-dimensional voltage fluctuation characteristics and feature dimension list into the model to obtain the short-circuit sensing category and the corresponding category confidence level. Finally, it filters and normalizes the sensing confidence level threshold to form short-circuit sensing information and performs short-circuit processing. In this way, it effectively solves the problems of untimely and inaccurate sensing caused by the lag in hardware response and the single criterion in traditional methods, and realizes rapid, accurate and reliable handling of short-circuit faults. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating the short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer provided by the present invention.

[0017] Figure 2 This is a flowchart illustrating the process of obtaining a list of feature dimensions in the short-circuit sensing processing method based on the electrical characteristics of a low-impedance voltage transformer provided by the present invention. Detailed Implementation

[0018] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0021] Examples, such as Figure 1 As shown, this embodiment of the invention provides a short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer, including:

[0022] S10: Interact with the target voltage transformer to obtain transient voltage fluctuation data, wherein the transient voltage fluctuation data includes the primary side and the secondary side.

[0023] Low-impedance voltage transformers (PTs) have extremely low internal resistance and are close to no-load conditions during normal operation. However, when a short circuit occurs, the secondary current increases sharply. This characteristic makes them extremely sensitive to transient voltage fluctuations and can quickly capture microsecond-level signal changes at the initial stage of a short circuit.

[0024] To address the aforementioned issues, this application uses an interactive target voltage transformer to acquire transient voltage fluctuation data, wherein the transient voltage fluctuation data includes primary side and secondary side.

[0025] Specifically, step S10 in the method includes:

[0026] Obtain the original short-circuit sensing data of the series to which the target voltage transformer belongs, and analyze and determine the transient voltage sampling parameters, including the sampling frequency and sampling duration;

[0027] Based on the transient voltage sampling parameters, synchronous voltage signals are acquired at the primary and secondary terminals of the target voltage transformer to obtain the transient voltage fluctuation data.

[0028] In this embodiment, the original short-circuit sensing data of the target voltage transformer series is first obtained, and the transient voltage sampling parameters are analyzed and determined. These transient voltage sampling parameters include the sampling frequency and sampling duration. The original short-circuit sensing data refers to historical or laboratory short-circuit condition data of the target voltage transformer series (such as those of the same model and design standard). For example, records of the primary and secondary voltage waveforms of voltage transformers of the same model as the target voltage transformer in past power grid short-circuit events can be obtained, along with a complete process including at least the fault type and the pre-fault steady-state, fault transient, and post-fault transient states. Obtaining the original short-circuit sensing data of the target voltage transformer series is crucial because different series of low-impedance voltage transformers exhibit differences in transient response speed (e.g., response time, bandwidth) and anti-saturation capability. However, voltage transformers of the same series have almost identical winding structures, core materials, and equivalent impedances, resulting in highly similar transient voltage transmission characteristics (e.g., transient response time, high-frequency signal attenuation rate). Historical data accurately reflects the fault signal performance of the target voltage transformer, providing a basis for determining the sampling parameters. For example, low-impedance voltage transformers are sensitive to microsecond-level voltage fluctuations. When a short circuit occurs, the grid voltage will drop sharply within 10-50μs, accompanied by high-frequency harmonics (such as transient components of 1kHz-1MHz). If the sampling frequency is too low, high-frequency signal aliasing and distortion will occur due to the Nyquist sampling theorem (in order to reconstruct the original continuous signal without distortion from the sampled discrete signal, the sampling frequency must be greater than or equal to twice the highest frequency of the original continuous signal), resulting in the loss of key fault characteristics. Therefore, the sampling frequency can extract the highest frequency component of the short-circuit transient from the original short-circuit sensing data. For example, through Fourier transform analysis, it is found that the highest short-circuit transient frequency of the voltage transformers in the series to which the target voltage transformer belongs is 500kHz. Then, according to the Nyquist criterion that the sampling frequency is ≥ 2 times the highest frequency, and a safety margin of 1.2-1.5 times is added to avoid edge frequency distortion, the sampling frequency is finally determined to be 500kHz × 2 × 1.2 = 1.2MHz. For example, determining a reasonable sampling duration is to cover the complete cycle of the short-circuit transient, which includes the three stages of fault triggering → transient oscillation → new steady state. The sampling duration can be determined by statistically analyzing the longest duration of the short-circuit transient of the target voltage transformer series from the original short-circuit sensing data. For example, if it takes 8ms for the voltage to oscillate to the new steady state after a fault, a redundancy duration of 2-3ms is added to ensure coverage of the steady-state stage before the fault. The final sampling duration is thus determined to be 8 + 2 = 10ms. In this way, by integrating the sampling frequency and sampling duration of the target voltage transformer, the transient voltage sampling parameters are obtained.

[0029] Secondly, based on the transient voltage sampling parameters, synchronous voltage signal acquisition is performed on the primary and secondary terminals of the target voltage transformer to obtain transient voltage fluctuation data. The primary side of the target voltage transformer is directly connected to the high-voltage grid, reflecting the actual fault state of the grid; the secondary side is a low-voltage signal after transformation, reflecting the voltage transformer's transmission characteristics of fault signals. Only synchronous acquisition can eliminate misjudgments caused by faults in the target voltage transformer itself through the correspondence between the two sides' signals. For example, a dual-channel synchronous ADC (analog-to-digital converter) can be used, sharing the same high-precision clock source (e.g., a 10MHz crystal oscillator) to ensure that the sampling trigger time and sampling interval of the primary and secondary side signals are completely consistent, with a time deviation controlled within 1μs. The primary side is connected to the ADC through a high-voltage divider module (adapted to the rated voltage of the primary side of the voltage transformer, e.g., 110kV), and the secondary side is directly connected to the ADC (adapted to the rated voltage of the secondary side of the voltage transformer, e.g., 100V), avoiding asynchrony caused by wiring delays. The timestamp of each voltage data point is recorded to obtain transient voltage fluctuation data. For example, primary side data: [(t0,U 1-t0 (t1,U) 1-t1 ),...,(t) n U 1-tn ], where U 1-tn The primary side voltage value at time t; secondary side data: [(t0, U 2-t0 (t1,U) 2-t1 ),...,(t) n U 2-tn ], where U 2-tn The value is the secondary voltage at time t; the timestamps of the primary and secondary data correspond one-to-one and can be directly used for subsequent voltage fluctuation feature extraction.

[0030] In summary, compared to existing technologies, this application uses an interactive target voltage transformer to acquire transient voltage fluctuation data, wherein the transient voltage fluctuation data includes both primary and secondary side data. This leverages the transient sensitivity advantage of low-impedance voltage transformers, providing a reliable data foundation for subsequent multi-dimensional feature extraction and fusion model discrimination.

[0031] S20: Obtain a list of feature dimensions and perform multi-dimensional feature extraction in conjunction with the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions.

[0032] Since there are many feature dimensions that can be used to determine the normal and short-circuit states of voltage transformers, the importance and reliability of different features vary significantly. This means that if all features are directly included in the analysis, not only will redundant information be introduced, increasing computational complexity, but the interference of low-reliability features may also blur key fault signals, ultimately reducing the accuracy and stability of the short-circuit discrimination model, or even causing misjudgment or omission.

[0033] To address the aforementioned issues, this application obtains a list of feature dimensions and performs multi-dimensional feature extraction in conjunction with the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions.

[0034] Specifically, such as Figure 2 As shown, step S20 in the method includes:

[0035] Based on the original short-circuit sensing data, statistical analysis is performed to determine the correlation strength between the candidate feature dimensions and the short-circuit event occurrence state, and the output is the analysis correlation degree;

[0036] Based on the analytical relevance, a threshold filter is performed to determine the list of feature dimensions;

[0037] The analytical relevance of multiple feature dimension items in the feature dimension list is normalized and stored as the dimension contribution.

[0038] The candidate feature dimensions include at least the root mean square rate of change of voltage, spectral entropy, waveform matching coefficient, skewness, abundance, trend slope, and rate of change of trend slope.

[0039] In this embodiment, the correlation strength between candidate feature dimensions and the short-circuit event occurrence state is first determined by statistical analysis based on the original short-circuit sensing data, and the output is the correlation score. The candidate feature dimensions are pre-defined voltage characteristic indicators that may reflect the short-circuit state, including at least the root mean square rate of change of voltage, spectral entropy, waveform matching coefficient, skewness, abundance, trend slope, and rate of change of trend slope. Specifically:

[0040] The root mean square rate of change is the magnitude of change of the effective voltage (RMS) per unit time. For example, the rate at which the RMS drops from 220V to 50V within 1ms. Due to the sudden drop in voltage during a short circuit, the root mean square rate of change will suddenly spike, more than 10 times the normal value.

[0041] Among them, spectral entropy refers to the degree of disorder of voltage signal in the frequency domain. The higher the spectral entropy, the more chaotic the frequency distribution. Normal voltage frequency is stable (mainly the 50 / 60Hz fundamental wave) and the spectral entropy is low. However, a short circuit will generate a large number of harmonics (such as the second and third harmonics), which will cause the spectral entropy to rise sharply.

[0042] The waveform matching coefficient refers to the similarity between the real-time voltage waveform and the normal voltage waveform. A waveform matching coefficient of 1 represents a perfect match, while a waveform matching coefficient of 0 represents a complete mismatch. Under normal conditions, the waveform matching coefficient is close to 1. Under short-circuit conditions, the voltage waveform will be distorted (such as the appearance of spikes or dips), and the waveform matching coefficient will drop rapidly to below 0.5.

[0043] Skewness refers to the degree of asymmetry in the distribution of voltage data. Skewness = 0 represents symmetry, skewness > 0 represents right skewness, and skewness < 0 represents left skewness. Normal voltage distribution is close to symmetry (skewness ≈ 0). During a short circuit, the voltage will shift to the lower value end, and the skewness will become negative.

[0044] Abundance refers to the proportion of a specific frequency component (such as fundamental frequency and harmonics) in the total voltage signal. Under normal conditions, the fundamental frequency abundance is >90%. During a short circuit, the fundamental frequency abundance drops sharply, while the fault harmonics (such as the third harmonic) increase sharply. For example, during a short circuit, the abundance increases from 1% to more than 30%.

[0045] The trend slope refers to the overall trend of voltage change over time. For example, the slope obtained by linear fitting is close to 0 when the voltage is stable under normal conditions, but the voltage drops rapidly during a short circuit and the trend slope becomes significantly negative.

[0046] Among them, the trend slope change rate refers to the rate of change of the trend slope itself. When a short circuit occurs, it is a sudden event, and the trend slope change rate will reach its peak instantaneously. Under normal circumstances, the slope changes slowly, and the trend slope change rate approaches 0.

[0047] For example, correlation strength refers to the degree of correlation between a candidate feature and the state of a short circuit event, quantified by statistical methods. The higher the correlation strength, the better the candidate feature can distinguish between normal and short circuit. For instance, the values ​​of all candidate feature dimensions can be extracted from the original short circuit sensing data, and statistical correlation analysis can be performed between these candidate feature dimension values ​​and the short circuit state label. Preferably, the linear correlation between the candidate feature and the short circuit state can be determined by the Pearson correlation coefficient. Each candidate feature dimension will have a correlation strength, such as 0.9 representing a strong correlation, 0.2 representing a weak correlation, etc., which can be used to analyze the correlation. In this way, potentially useful features can be initially screened out, while redundant features unrelated to the short circuit state can be eliminated.

[0048] Secondly, a threshold screening is performed based on the analytical relevance to determine the feature dimension list. For example, an analytical relevance threshold can be pre-set based on empirical values ​​or experimental data. Preferably, this application recommends setting the analytical relevance threshold to 0.6 based on empirical values ​​and a large amount of experimental data. Those skilled in the art can dynamically adjust it according to actual circumstances. If the analytical relevance of a candidate feature dimension is less than the analytical relevance threshold, it indicates that the feature dimension has a weak ability to distinguish between normal and short-circuit conditions, and it is eliminated. Conversely, if the analytical relevance of a candidate feature dimension is greater than or equal to the analytical relevance threshold, it indicates that the feature dimension has a strong ability to distinguish between normal and short-circuit conditions, and it is retained. All retained feature dimensions together form the feature dimension list. In this way, weakly correlated feature dimensions can be effectively eliminated, and strongly correlated feature dimensions can be retained, avoiding subsequent computational complexity and decreased discrimination accuracy due to redundant features.

[0049] Finally, the analytical relevance of multiple feature dimension items in the normalized feature dimension list is stored as a dimensional contribution. Specifically, the analytical relevance of multiple feature dimension items in the feature dimension list is converted into a value between 0 and 1 to eliminate the influence of dimensions, allowing the importance of each feature to be directly compared. The normalized analytical relevance is then used as the dimensional contribution of that feature in short-circuit discrimination; the higher the dimensional contribution, the greater the contribution. For example, if the feature dimension list contains three features: root mean square rate of change of voltage, spectral entropy, and waveform matching coefficient, and the analytical relevances are 0.9, 0.7, and 0.6 respectively, then linear normalization can be used to calculate the dimension contribution = analytical relevance / sum of all analytical relevances, i.e., 0.9 / (0.9+0.7+0.6)=0.45, 0.7 / (0.9+0.7+0.6)=0.32, and 0.6 / (0.9+0.7+0.6)=0.27. Then, the feature dimension items are bound to the corresponding dimension contribution, such as root mean square rate of change of voltage with 0.45, spectral entropy with 0.32, and waveform matching coefficient with 0.27. In this way, the importance difference of different feature dimension items is quantified, and the dimension contribution can be used to determine the priority of each feature dimension item.

[0050] Furthermore, based on the feature dimension list and combined with transient voltage fluctuation data, multidimensional feature extraction is performed to obtain multidimensional voltage fluctuation features. Specifically, for the feature dimensions already selected in the feature dimension list, according to the definition and calculation logic, the corresponding feature values ​​are calculated from the transient voltage fluctuation data to form multidimensional voltage fluctuation features. For example, if the feature dimension list includes the root mean square rate of change of voltage, spectral entropy, and trend slope, the root mean square rate of change of voltage can be calculated first, and then the rate of change of RMS over time can be calculated using ΔRMS / Δt to obtain the root mean square rate of change of voltage, where ΔRMS is the change in RMS over time Δt. Similarly, according to the definition and calculation logic of the feature dimensions, the specific feature values ​​of other feature dimensions are calculated to form multidimensional voltage fluctuation features. Multidimensional voltage fluctuation features are quantitative and physically meaningful indicators that can directly reflect voltage anomalies during short circuits (such as sudden drops, spectral instability, etc.). In this way, the original transient voltage fluctuation data is transformed into a feature language that the model can understand.

[0051] In summary, compared to existing technologies, this application obtains a list of feature dimensions and performs multi-dimensional feature extraction based on the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions. In this way, the original transient voltage fluctuation data is transformed into a feature language that the model can understand, providing reliable data input for subsequent steps.

[0052] S30: Construct a fusion short-circuit discrimination model, wherein the fusion short-circuit discrimination model includes multiple univariate short-circuit sensing channels and is integrated and output through an integration layer.

[0053] Traditional short-circuit detection typically employs a fixed threshold comparison method based on a single electrical characteristic, such as a sudden drop in voltage amplitude or a sudden change in current value. This method has limitations in adapting to complex power grid conditions and capturing transient fault characteristics incompletely, resulting in detection accuracy being constrained by the scenario and a high rate of false positives and false negatives.

[0054] To address the aforementioned issues, this application constructs a fusion short-circuit discrimination model, wherein the fusion short-circuit discrimination model includes multiple univariate short-circuit sensing channels, and the output is integrated through an integration layer.

[0055] Specifically, step S30 in the method includes:

[0056] Acquire raw short-circuit sensing data;

[0057] Based on the feature dimension list, the original short-circuit sensing data is reconstructed to obtain standard short-circuit sensing data, wherein the standard short-circuit sensing data includes multiple univariate data groups, and each univariate data group corresponds to one feature dimension item.

[0058] The standard short-circuit sensing data is traversed to establish multiple univariate short-circuit sensing channels, wherein each univariate short-circuit sensing channel takes the feature dimension as input and the predefined category probability of the short-circuit sensing category as output.

[0059] Construct the integration layer and initialize the integration weights to equal values, connect the outputs of multiple single-variable short-circuit sensing channels to the input of the integration layer, and obtain the fused short-circuit discrimination model.

[0060] In this embodiment, the original short-circuit sensing data is first obtained. The original short-circuit sensing data refers to the historical or experimental data of the series to which the target voltage transformer belongs, including transient voltage fluctuation data and its corresponding real operating condition labels. The real operating condition labels are predefined short-circuit sensing categories, such as normal = 0, single-phase ground fault = 1, three-phase short circuit = 2, etc.

[0061] Secondly, the original short-circuit sensing data is reconstructed based on the feature dimension list to obtain standard short-circuit sensing data. This standard short-circuit sensing data includes multiple univariate data sets, each corresponding to a feature dimension item. Specifically, data reconstruction transforms the original short-circuit sensing data into a standardized format adapted to the univariate short-circuit sensing channel. For example, if the feature dimension list contains three features: root mean square rate of change of voltage, spectral entropy, and waveform matching coefficient, data reconstruction yields three univariate data sets: a root mean square rate of change data set containing the root mean square rate of change feature values ​​for all samples and their corresponding operating condition labels; a spectral entropy data set containing the spectral entropy feature values ​​for all samples and their corresponding operating condition labels; and a waveform matching coefficient data set containing the waveform matching coefficient feature values ​​for all samples and their corresponding operating condition labels.

[0062] Next, the standard short-circuit sensing data is traversed to establish multiple univariate short-circuit sensing channels. Each univariate short-circuit sensing channel takes a feature dimension as input and outputs the class probability of a predefined short-circuit sensing category. Each univariate short-circuit sensing channel is equivalent to a lightweight classifier that focuses on short-circuit discrimination of only one feature dimension, thus avoiding interference from multiple feature cross-interference. For example, by traversing the standard short-circuit sensing data, a short-circuit sensing channel is trained separately for each reconstructed univariate data set. For instance, a univariate short-circuit sensing channel can be constructed based on the C4.5 decision tree algorithm, mainly consisting of an input layer, a splitting node layer, and a leaf node layer. The input layer receives the value of a single feature dimension, such as the feature value of the root mean square rate of change. The splitting node layer determines the optimal splitting threshold by recursively calculating the information gain ratio of the feature. For example, when the root mean square rate of change is >3% / ms, the sample is more likely to belong to a single-phase short circuit, forming a multi-branch splitting path. The leaf node layer stores the class distribution of all training samples under this path. For example, a certain leaf node contains 80 three-phase short-circuit samples and 20 normal samples, and the sample proportion is used as the output class probability, that is, the leaf node outputs three-phase short circuit 0.8 and normal 0.2. During training, using the feature dimension as input and the predefined class probability of the short-circuit sensing category as output, iterative optimization can be achieved through the following steps: Initialize the decision tree depth to 1, calculate the information gain ratio of the feature starting from the root node, and select the threshold of the maximum gain ratio for node splitting; recursively split layer by layer until the number of node samples is less than a preset threshold (e.g., 50 samples), the information gain ratio is lower than the minimum threshold (e.g., 0.01), or the tree depth reaches the upper limit (e.g., 8 layers), at which point convergence is considered achieved, splitting is stopped, and the class probability distribution of the leaf nodes is fixed. In this way, a univariate short-circuit sensing channel is obtained, which can accurately capture the nonlinear correlation between a single feature and the short-circuit category. For example, a spectral entropy in the range of 0.3 to 0.5 corresponds to normal, while >0.7 corresponds to a three-phase short circuit. Each univariate short-circuit sensing channel focuses on the discriminative advantage of its feature dimension, maximizing the feature's accuracy in distinguishing the short-circuit category.

[0063] Finally, an ensemble layer is constructed and its ensemble weights are initialized to equal values. The outputs of multiple univariate short-circuit sensing channels are connected to the input of the ensemble layer to obtain a fused short-circuit discrimination model. The core function of the ensemble layer is to integrate the outputs of all univariate short-circuit sensing channels and obtain the final short-circuit discrimination conclusion through weighted calculation, avoiding misjudgments caused by the feature limitations of a single channel. For example, the ensemble layer is essentially a weighted fusion module, taking the class probabilities of all univariate short-circuit sensing channels as input and outputting the final short-circuit sensing class and total confidence score. For example, when the integration layer is first built, the integration weights of all univariate short-circuit sensing channels are set to the same value. For instance, if there are 3 univariate short-circuit sensing channels, the integration weight of each univariate short-circuit sensing channel is 1 / 3. Then, the output of each univariate short-circuit sensing channel is connected to the input of the integration layer, forming a complete model architecture with parallel input of multiple univariate short-circuit sensing channels and aggregated output of the integration layer. This results in a fused short-circuit discrimination model. The fused short-circuit discrimination model has the ability to perform parallel analysis of multi-dimensional features and weighted integrated decision-making, providing a reliable analysis model for subsequent input of real-time data for short-circuit discrimination.

[0064] Specifically, the step of "reconstructing the original short-circuit sensing data based on the feature dimension list to obtain standard short-circuit sensing data" includes:

[0065] Based on the aforementioned feature dimension list, the missing feature dimension items in the original short-circuit sensing data are calculated and the calculation results are added to the original short-circuit sensing data.

[0066] The original short-circuit sensing data is cleaned using the feature dimension list as an index.

[0067] The data cleaning results are divided into multiple data groups, each of which is suitable for the prediction of different feature dimensions, to obtain the standard short-circuit sensing data.

[0068] In this embodiment, the missing feature dimension items in the original short-circuit sensing data are first calculated by iterating through the feature dimension list, and the calculation results are added to the original short-circuit sensing data. Specifically, since the original short-circuit sensing data is short-circuit condition data, it usually only records voltage amplitude time series, etc., while the feature dimension list includes derived features such as the root mean square rate of change of voltage, spectral entropy, and waveform matching coefficient, which need to be calculated. Therefore, there is a problem that the original short-circuit sensing data is missing feature dimension items. For example, the original short-circuit sensing data is traversed, and the missing feature dimension items are determined by referring to the feature dimension list. For example, the root mean square rate of change and spectral entropy are missing. Then, the missing feature dimension items are calculated according to the definition and formula of the feature dimension items in step S20: If the root mean square rate of change is missing, the root mean square (RMS) of the voltage is calculated first, and then the rate of change of RMS over time is calculated by ΔRMS / Δt to obtain the root mean square rate of change of voltage, where ΔRMS is the change of RMS over time Δt; if the spectral entropy is missing, the voltage sequence is subjected to FFT to obtain the spectrum, and then the entropy value of the spectrum is calculated; in this way, the calculated feature values ​​are added to the original short-circuit sensing data to form a complete data sample containing all feature dimension items.

[0069] Secondly, the original short-circuit sensing data is cleaned using a list of feature dimensions as an index to remove invalid and abnormal data, thus avoiding noise interference with model training. For example, data cleaning can be achieved through the following technical path: using the list of feature dimensions as an index, only the feature dimension items in the list are retained, and redundant data unrelated to short-circuit detection is deleted; then, data can be filtered through box plot analysis or based on the physical reasonable range of the feature dimension items. For example, based on industry experience, the normal range of the root mean square rate of change of voltage is 0~0.5% / ms, with a maximum of 10% / ms during a short circuit. If a certain root mean square rate of change is 50% / ms, it is judged as an outlier and removed. In this way, the feature data input to the model can truly reflect the short-circuit characteristics, avoiding outliers from causing incorrect judgments in model learning.

[0070] Finally, the cleaned data is divided into multiple data groups, each suitable for prediction of different feature dimensions, thus obtaining standard short-circuit sensing data. For example, according to each feature dimension in the feature dimension list, the cleaned data is split into multiple independent data groups. Each data group contains only all feature values ​​of one feature dimension and their corresponding operating condition labels. For instance, if the cleaned complete data contains 1000 samples, each sample has three features (root mean square rate of change, spectral entropy, and waveform matching coefficient) and one corresponding operating condition label, then after dividing by feature dimension, we get: Data group 1 contains the root mean square rate of change feature values ​​of 1000 samples and their corresponding 1000 operating condition labels; Data group 2 contains the spectral entropy feature values ​​of 1000 samples and their corresponding 1000 operating condition labels; and Data group 3 contains the waveform matching coefficient feature values ​​of 1000 samples and their corresponding 1000 operating condition labels. In this way, standard short-circuit sensing data is obtained, which can adapt to the training needs of multiple univariate short-circuit sensing channels, with each short-circuit sensing channel requiring only training data for one feature dimension.

[0071] In summary, compared to existing technologies, this application constructs a fusion short-circuit discrimination model, which includes multiple univariate short-circuit sensing channels and integrates them through an integration layer. Thus, by using the architecture of multiple univariate short-circuit sensing channels and an integration layer, a fusion short-circuit discrimination model is obtained. This leverages the targeted discrimination advantages of single features while overcoming the limitations of single features through integration, resulting in a fusion short-circuit discrimination model that integrates the advantages of multiple features and has strong anti-interference capabilities, providing reliable support for subsequent accurate short-circuit discrimination.

[0072] S40: Input the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence.

[0073] The aforementioned steps yielded the fusion short-circuit discrimination model and dimensional contribution, which can be used to perform short-circuit sensing and discrimination of multi-dimensional voltage fluctuation characteristics.

[0074] To address the aforementioned issues, this application inputs the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence.

[0075] Specifically, step S40 in the method includes:

[0076] Construct a mapping relationship between the feature dimension terms and multiple univariate short-circuit sensing channels in the fusion short-circuit discrimination model;

[0077] According to the mapping relationship, the multidimensional voltage fluctuation features are split and input into multiple single-variable short-circuit sensing channels of the fused short-circuit discrimination model to obtain multiple sets of channel sensing results. Each set of channel sensing results includes the channel sensing category and the corresponding channel category probability.

[0078] The integration weights of the integration layer are updated based on the dimension contribution of multiple feature dimension items in the feature dimension list.

[0079] The multiple sets of channel perception results are input to the integration layer, and the multiple channel perception categories are weighted and integrated to obtain the confidence scores of the multiple categories.

[0080] The channel sensing category and the category confidence level are correlated and output to obtain the short-circuit sensing category and the corresponding category confidence level.

[0081] In this embodiment, a mapping relationship is first constructed between feature dimension items and multiple univariate short-circuit sensing channels in the fused short-circuit discrimination model. Specifically, a mapping relationship is established between each feature dimension item in the feature dimension list and each univariate short-circuit sensing channel in the fused short-circuit discrimination model. For example, if the feature dimension list includes root mean square rate of change, spectral entropy, and waveform matching coefficient, and in the fused short-circuit discrimination model, channel 1 is trained based on a univariate data set with root mean square rate of change, channel 2 is trained based on a univariate data set with spectral entropy, and channel 3 is trained based on a univariate data set with waveform matching coefficient, then a mapping relationship is established between root mean square rate of change and channel 1, spectral entropy and channel 2, and waveform matching coefficient and channel 3. This ensures that each feature dimension item can be analyzed by the short-circuit sensing channel that is best suited to handle it, avoiding discrimination bias caused by mismatch between feature dimension items and univariate short-circuit sensing channels.

[0082] Secondly, based on the mapping relationship, the multidimensional voltage fluctuation features are split and input into multiple univariate short-circuit sensing channels of the fused short-circuit discrimination model to obtain multiple sets of channel sensing results. Each set of channel sensing results includes the channel sensing category and the corresponding channel category probability. The channel sensing category refers to the most likely short-circuit category judged by the channel, and the channel category probability refers to the confidence distribution of each short-circuit category of the channel (with a value range of 0 to 1). For example, the multidimensional voltage fluctuation characteristics are decomposed according to the mapping relationship into: root mean square rate of change = 5% / ms, spectral entropy = 0.8, waveform matching coefficient = 0.3, which are respectively input into channel 1, channel 2, and channel 3 of the fused short-circuit discrimination model. After analyzing the input feature values, each univariate short-circuit sensing channel outputs the channel sensing category and the corresponding channel category probability: Channel 1 outputs channel sensing category = single-phase short circuit, channel category probability = normal 0.1, single-phase short circuit 0.8, three-phase short circuit 0.1 (total probability is 1); Channel 2 outputs channel sensing category = three-phase short circuit, channel category probability = normal 0.1, single-phase short circuit 0.2, three-phase short circuit 0.7 (total probability is 1); Channel 3 outputs channel sensing category = single-phase short circuit, channel category probability = normal 0.1, single-phase short circuit 0.6, three-phase short circuit 0.3 (total probability is 1); in this way, multiple sets of channel sensing results are obtained.

[0083] Next, the integration weights of the integration layer are updated based on the dimensional contribution of multiple feature dimension items in the feature dimension list. For example, based on the dimensional contribution of the root mean square rate of change of voltage, spectral entropy, and waveform matching coefficient in the feature dimension list: 0.45, 0.32, and 0.27, the initial equal integration weights of the integration layer are updated to 0.45, 0.32, and 0.27, respectively. In this way, the integration layer focuses on important feature dimension items, improving integration accuracy.

[0084] Furthermore, multiple sets of channel sensing results are input to the integration layer, and the multiple channel sensing categories are weighted and integrated to obtain multiple category confidence scores. For example, if the updated integration weights of channel 1, channel 2, and channel 3 are 0.45, 0.32, and 0.27 respectively, then the category confidence score for a single-phase short circuit is 0.45×0.8+0.32×0.2+0.27×0.6=0.586, the category confidence score for a three-phase short circuit is 0.45×0.1+0.32×0.7+0.27×0.3=0.35, and the category confidence score for normal is 0.45×0.1+0.32×0.1+0.27×0.1=0.104. In this way, by fusing the multi-channel discrimination conclusions and weighting the integration based on the dimensional contribution, the discrimination conclusions of important feature dimensions have a higher proportion in the final result, avoiding misjudgment of a single channel.

[0085] Finally, the output channel sensing category and category confidence score are correlated to obtain the short-circuit sensing category and its corresponding category confidence score. For example, all channel sensing categories are correlated with their corresponding category confidence scores to form the short-circuit sensing category and its corresponding category confidence score, such as [normal: 0.104, single-phase short circuit: 0.586, three-phase short circuit: 0.35]. This provides a reliable quantitative basis for subsequent steps.

[0086] In summary, compared to existing technologies, this application inputs the multi-dimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence. In this way, the multi-dimensional voltage fluctuation features are transformed into interpretable, high-precision short-circuit discrimination conclusions, solving the pain points of poor anti-interference and low accuracy of traditional single-feature discrimination.

[0087] S50: Based on the perception confidence threshold, filter the category confidence to determine the confidence short-circuit perception category, normalize the category confidence of multiple confidence short-circuit perception categories, and output them as short-circuit perception information for short-circuit processing.

[0088] Because the characteristic signal strengths of different short-circuit sensing categories are fundamentally different, and complex operating conditions such as resonance may occur during power grid operation, the discrimination value of the same feature changes dynamically under different operating conditions. Therefore, the sensing confidence threshold is not a fixed value, but a dynamic threshold that sets a specific benchmark for different short-circuit categories and can be adaptively adjusted according to real-time operating conditions, thereby adapting to specific application scenarios.

[0089] To address the aforementioned issues, this application filters the category confidence based on a perception confidence threshold to determine a confidence short-circuit perception category, normalizes the category confidence of multiple confidence short-circuit perception categories, and outputs them as short-circuit perception information for short-circuit processing.

[0090] Specifically, step S50 in the method includes:

[0091] The original short-circuit sensing data is input into the fused short-circuit discrimination model to obtain the verification discrimination input;

[0092] Based on the preset confidence constraints and the verification and discrimination input, statistical analysis is performed with the category confidence as the target to determine the perception confidence threshold corresponding to each short-circuit perception category;

[0093] The confidence scores of the categories are traversed and threshold filtering is performed. The short-circuit sensing categories that are greater than the confidence score threshold are output as the confident short-circuit sensing categories.

[0094] Normalize the category confidence of multiple confidence short-circuit sensing categories, obtain the category probability distribution, and merge the multiple confidence short-circuit sensing categories with the category probability distribution to obtain the short-circuit sensing information.

[0095] In this embodiment, the original short-circuit sensing data is first input into the fused short-circuit discrimination model to obtain the verification discrimination input. The original short-circuit sensing data serves as prior data of known operating conditions for the target voltage transformer series, and therefore can be used to verify the effectiveness of the fused short-circuit discrimination model's output. Specifically, a data point from the original short-circuit sensing data with a true label of single-phase short circuit can be selected. Its multi-dimensional features are extracted according to the process and then input into the fused short-circuit discrimination model. The output includes a category confidence distribution containing the probabilities of each category, for example: single-phase short circuit 0.75, three-phase short circuit 0.2, normal 0.05. This result serves as the verification discrimination input. By comparing the verification input with the actual operating condition labels (such as single-phase short circuit) of the original short circuit sensing data one by one, the discrimination performance of the fusion short circuit discrimination model for different short circuit categories can be statistically obtained. This includes quantitative indicators such as true positive rate (such as the correct identification ratio of single-phase short circuit), false positive rate (such as the proportion of normal operating conditions misclassified as short circuit), and the mean and fluctuation range of confidence scores for each category. These indicators can accurately reflect the reliability of the model in discriminating different fault types.

[0096] Secondly, based on preset confidence constraints and verification inputs, statistical analysis is performed with category confidence as the target to determine the perception confidence threshold corresponding to each short-circuit sensing category. The preset confidence constraints are quantitative performance indicators set for different short-circuit sensing categories based on the actual reliability requirements of power system short-circuit protection (such as equipment tolerance and grid power supply continuity requirements) and industry standards. These indicators include a false positive rate of ≤5% under normal operating conditions, a single-phase short-circuit missed detection rate of ≤8%, and a three-phase short-circuit misjudgment rate of ≤3%. Those skilled in the art can dynamically adjust these indicators according to actual scenarios (such as grid voltage levels, key equipment parameters, and load types). Ideally, for high-voltage transmission networks, due to the high value of equipment and the wide range of fault impacts, the three-phase short-circuit misjudgment rate can be tightened to ≤1%; for distribution networks, due to large load fluctuations and complex fault characteristics, the single-phase short-circuit missed detection rate can be relaxed to ≤10%. For example, based on preset confidence constraints, such as a false positive rate ≤5% under normal operating conditions, a single-phase short-circuit false negative rate ≤8% and a three-phase short-circuit false negative rate ≤3%, combined with the verification and discrimination input, the sensing confidence threshold can be determined through the following process: First, the verification and discrimination input is statistically analyzed by grouping according to the short-circuit sensing category (e.g., normal, single-phase short circuit, three-phase short circuit). For each group, a threshold-performance index curve is plotted, for example, the threshold-false positive rate curve for the normal category and the threshold-false negative rate curve for the single-phase short circuit. Then, the minimum threshold point that satisfies the preset confidence constraints is located on the curve. For example, for the normal category, the false positive rate curve needs to be found. The minimum threshold for a positive rate ≤ 5% is 0.5; for a single-phase short circuit, the minimum threshold for a false negative rate ≤ 8% is 0.55; and for a three-phase short circuit, the minimum threshold for a false positive rate ≤ 3% is 0.75. This yields the perception confidence threshold for each short circuit perception category: 0.5 for normal perception, 0.55 for single-phase short circuit, and 0.75 for three-phase short circuit. This approach adapts the perception confidence threshold to the characteristics of different short circuit categories, better reflecting the actual performance of different fault types in the power system and improving the accuracy of threshold selection.

[0097] Next, the category confidence scores are iterated through and thresholded, and short-circuit sensing categories with confidence scores greater than the perceived confidence threshold are output as confident short-circuit sensing categories. For example, if the perceived confidence threshold is 0.5 for normal, 0.55 for single-phase short circuit, and 0.75 for three-phase short circuit, when the category confidence score for normal is 0.104, the category confidence score for single-phase short circuit is 0.586, and the category confidence score for three-phase short circuit is 0.35, then normal and three-phase short circuit are removed, and single-phase short circuit is retained as the final confident short-circuit sensing category. In this way, low-confidence results are filtered out.

[0098] Finally, the category confidence scores of multiple confidence short-circuit sensing categories are normalized to obtain the category probability distribution. These multiple confidence short-circuit sensing categories and their category probability distributions are then merged to obtain short-circuit sensing information. For example, if multiple confidence short-circuit sensing categories are retained after filtering, such as when single-phase short circuit is 0.7 and three-phase short circuit is 0.68, both thresholds are greater than the sensing confidence threshold. Then, the category confidence scores of the multiple confidence short-circuit sensing categories are converted into a probability distribution with a sum of 1. Optionally, the normalization formula is: Normalized probability of a certain confidence short-circuit sensing category = Category confidence score of that confidence short-circuit sensing category / Sum of category confidence scores of all confidence short-circuit sensing categories. For example, if single-phase short circuit is 0.7 and three-phase short circuit is 0.68, both thresholds are greater than the sensing confidence threshold. If the probability is 0.68, then the normalized probability of a single-phase short circuit is 0.7 / (0.7+0.68)≈0.507, and the normalized probability of a three-phase short circuit is 0.68 / (0.7+0.68)≈0.493. Then, multiple confidence short circuit sensing categories and category probability distributions are merged to obtain short circuit sensing information: 0.507 for single-phase short circuit and 0.493 for three-phase short circuit. The short circuit sensing information can be directly transmitted to the power system protection device to trigger targeted short circuit processing, ensuring the accuracy and flexibility of the processing action.

[0099] Furthermore, because the core magnetic circuit of a low-impedance voltage transformer will exhibit nonlinear distortion during grid resonance (such as ferroresonance or series resonance), causing the secondary-side output voltage signal to be abnormally amplified or suppressed / attenuated, for example, 100V may be abnormally amplified to 120V or reduced to 80V. This distortion may cause the fused short-circuit discrimination model to mistakenly identify a normal signal as a short circuit or a suppressed short-circuit signal as normal. Therefore, step S50 in the method further includes:

[0100] Based on experimental methods, the resonant sensitivity of the target voltage transformer was evaluated, and a fault amplification effect model was established, wherein the fault amplification effect model is a mathematical model.

[0101] Real-time operating parameters of the target voltage transformer are acquired, and the amplification effect coefficient is obtained by combining the real-time operating parameters with the fault amplification effect model.

[0102] The perceived confidence threshold is adaptively adjusted based on the amplification effect coefficient.

[0103] In this embodiment, the resonant sensitivity of the target voltage transformer is first evaluated based on experimental methods, and a fault amplification effect model is established, wherein the fault amplification effect model is a mathematical model. For example, the resonant environment of the target voltage transformer under different operating conditions is simulated in the laboratory, such as adjusting the grid frequency to 50-60Hz and changing the load impedance. The output signal changes of the target voltage transformer under different resonant conditions are tested, and the resonant parameters including frequency and load impedance and the corresponding signal amplification factors are recorded. Then, several resonant parameters and their corresponding signal amplification factors are fitted into a mathematical formula to quantify the relationship between the resonant parameters and the signal amplification factors, which serves as the fault amplification effect model, for example, K=a×f. 2 +b×Z+c, where a, b, and c are experimental fitting coefficients, f is the frequency, Z is the load impedance, and K is the amplification effect coefficient. K>1 indicates that the signal is amplified, and K<1 indicates that the signal is suppressed. In this way, the influence of resonance on the target voltage transformer signal is transformed into a calculable quantization coefficient, providing a mathematical basis for subsequent threshold correction.

[0104] Secondly, the real-time operating parameters of the target voltage transformer are acquired, and the amplification effect coefficient is obtained by combining the real-time operating parameters with the fault amplification effect model. For example, if the fault amplification effect model is: K = 0.001 × f 2 +0.002×Z+0.05, when the real-time operating parameters of the target voltage transformer are f=30Hz and Z=110Ω, K=0.001×30×30+0.002×110+0.05=1.17, indicating that the current signal is amplified by 1.17 times, which is easy to misjudge.

[0105] Finally, the sensing confidence threshold is adaptively adjusted based on the amplification effect coefficient. Specifically, when the amplification effect coefficient > 1, it indicates that the signal is amplified, which is prone to being misjudged as a short circuit. Therefore, the sensing confidence threshold needs to be increased, for example, by multiplying the amplification effect coefficient by the sensing confidence threshold to obtain the corrected sensing confidence threshold. When the amplification effect coefficient < 1, it indicates that the signal is suppressed, which is prone to missing short circuits. Therefore, the sensing confidence threshold needs to be decreased, for example, by multiplying the amplification effect coefficient by the sensing confidence threshold to obtain the corrected sensing confidence threshold. When the amplification effect coefficient ≈ 1, it indicates that there is no significant amplification / suppression, and the sensing confidence threshold remains unchanged. For example, if the sensing confidence threshold for a single-phase short circuit is 0.55, when the amplification effect coefficient K = 1.17, the sensing confidence threshold is corrected to 0.55 × 1.17 = 0.64. This avoids the interference of resonance on short circuit detection, making the short circuit sensing method adaptable to more power grid scenarios.

[0106] In summary, compared to existing technologies, this application uses a perception confidence threshold to filter the category confidence to determine the confidence short-circuit perception category. It then normalizes the category confidence of multiple confidence short-circuit perception categories and outputs the results as short-circuit perception information for short-circuit processing. This obtains reliable and interpretable short-circuit perception information, avoids misjudgments caused by low-confidence results, and provides a clear basis for short-circuit processing.

[0107] In summary, the embodiments of this application have at least the following technical effects:

[0108] Compared to existing technologies, this application first interacts with the target voltage transformer to acquire transient voltage fluctuation data, which includes both primary and secondary voltage fluctuation data. This leverages the transient sensitivity advantage of the low-impedance voltage transformer, providing a reliable data foundation for subsequent multi-dimensional feature extraction and fusion model discrimination.

[0109] Secondly, this application obtains a list of feature dimensions and performs multi-dimensional feature extraction based on the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions. In this way, the original transient voltage fluctuation data is transformed into a feature language that the model can understand, providing reliable data input for subsequent steps.

[0110] Furthermore, this application constructs a fusion short-circuit discrimination model, which includes multiple univariate short-circuit sensing channels and is integrated and output through an integration layer. Thus, through the architecture of multiple univariate short-circuit sensing channels and an integration layer, a fusion short-circuit discrimination model is obtained. This leverages the targeted discrimination advantages of single features while avoiding the limitations of single features through integration, resulting in a fusion short-circuit discrimination model that can integrate the advantages of multiple features and has strong anti-interference capabilities, providing reliable support for subsequent accurate short-circuit discrimination.

[0111] Furthermore, this application inputs the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence. In this way, the multidimensional voltage fluctuation features are transformed into interpretable, high-precision short-circuit discrimination conclusions, solving the pain points of poor anti-interference and low accuracy of traditional single-feature discrimination.

[0112] Finally, based on the perceived confidence threshold, this application filters the category confidence to determine the confidence short-circuit perception category, normalizes the category confidence of multiple confidence short-circuit perception categories, and outputs it as short-circuit perception information for short-circuit processing. In this way, reliable and interpretable short-circuit perception information is obtained, avoiding misjudgments caused by low-confidence results and providing a clear basis for short-circuit processing.

[0113] Through the above technical solution, this application first obtains transient voltage fluctuation data of the primary and secondary sides from the target voltage transformer to ensure that key fault characteristics are captured in a timely manner. Then, it constructs a fusion short-circuit discrimination model composed of multiple univariate short-circuit sensing channels and an integrated layer to avoid the risk of misjudgment by a single channel. Next, it inputs the multi-dimensional voltage fluctuation characteristics and feature dimension list into the model to obtain the short-circuit sensing category and the corresponding category confidence level. Finally, it filters and normalizes the sensing confidence level threshold to form short-circuit sensing information and performs short-circuit processing. In this way, it effectively solves the problems of untimely and inaccurate sensing caused by the lag in hardware response and the single criterion in traditional methods, and realizes rapid, accurate and reliable handling of short-circuit faults.

[0114] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0115] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0120] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer, characterized in that, include: Interact with the target voltage transformer to acquire transient voltage fluctuation data, wherein the transient voltage fluctuation data includes primary side and secondary side; Obtain a list of feature dimensions and perform multi-dimensional feature extraction in conjunction with the transient voltage fluctuation data to obtain multi-dimensional voltage fluctuation features. The list of feature dimensions includes feature dimension items of multiple orders and their corresponding dimensional contributions. A fusion short-circuit discrimination model is constructed, wherein the fusion short-circuit discrimination model includes multiple univariate short-circuit sensing channels and is integrated and output through an integration layer; Input the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence. Based on the perception confidence threshold, the category confidence is filtered to determine the confidence short-circuit perception category. The category confidence of multiple confidence short-circuit perception categories is normalized and output as short-circuit perception information for short-circuit processing.

2. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 1, characterized in that, Interacting with the target voltage transformer to acquire transient voltage fluctuation data, wherein the transient voltage fluctuation data includes primary and secondary sides, including: Obtain the original short-circuit sensing data of the series to which the target voltage transformer belongs, and analyze and determine the transient voltage sampling parameters, including the sampling frequency and sampling duration; Based on the transient voltage sampling parameters, synchronous voltage signals are acquired at the primary and secondary terminals of the target voltage transformer to obtain the transient voltage fluctuation data.

3. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 2, characterized in that, Obtain a list of feature dimensions, including: Based on the original short-circuit sensing data, statistical analysis is performed to determine the correlation strength between the candidate feature dimensions and the short-circuit event occurrence state, and the output is the analysis correlation degree; Based on the analytical relevance, a threshold filter is performed to determine the list of feature dimensions; The analytical relevance of multiple feature dimension items in the feature dimension list is normalized and stored as the dimension contribution. The candidate feature dimensions include at least the root mean square rate of change of voltage, spectral entropy, waveform matching coefficient, skewness, abundance, trend slope, and rate of change of trend slope.

4. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 3, characterized in that, A fusion short-circuit discrimination model is constructed, wherein the fusion short-circuit discrimination model includes multiple univariate short-circuit sensing channels, and the output is integrated through an integration layer, including: Acquire raw short-circuit sensing data; Based on the feature dimension list, the original short-circuit sensing data is reconstructed to obtain standard short-circuit sensing data, wherein the standard short-circuit sensing data includes multiple univariate data groups, and each univariate data group corresponds to one feature dimension item. The standard short-circuit sensing data is traversed to establish multiple univariate short-circuit sensing channels, wherein each univariate short-circuit sensing channel takes the feature dimension as input and the predefined category probability of the short-circuit sensing category as output. Construct the integration layer and initialize the integration weights to equal values, connect the outputs of multiple single-variable short-circuit sensing channels to the input of the integration layer, and obtain the fused short-circuit discrimination model.

5. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 4, characterized in that, Based on the feature dimension list, the original short-circuit sensing data is reconstructed to obtain standard short-circuit sensing data, including: Based on the aforementioned feature dimension list, the missing feature dimension items in the original short-circuit sensing data are calculated and the calculation results are added to the original short-circuit sensing data. The original short-circuit sensing data is cleaned using the feature dimension list as an index. The data cleaning results are divided into multiple data groups, each of which is suitable for the prediction of different feature dimensions, to obtain the standard short-circuit sensing data.

6. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 5, characterized in that, Inputting the multidimensional voltage fluctuation features and the feature dimension list into the fused short-circuit discrimination model to obtain the short-circuit sensing category and the corresponding category confidence score, the method further includes: Construct a mapping relationship between the feature dimension terms and multiple univariate short-circuit sensing channels in the fusion short-circuit discrimination model; According to the mapping relationship, the multidimensional voltage fluctuation features are split and input into multiple single-variable short-circuit sensing channels of the fused short-circuit discrimination model to obtain multiple sets of channel sensing results. Each set of channel sensing results includes the channel sensing category and the corresponding channel category probability. The integration weights of the integration layer are updated based on the dimension contribution of multiple feature dimension items in the feature dimension list. The multiple sets of channel perception results are input to the integration layer, and the multiple channel perception categories are weighted and integrated to obtain the confidence scores of the multiple categories. The channel sensing category and the category confidence level are correlated and output to obtain the short-circuit sensing category and the corresponding category confidence level.

7. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 6, characterized in that, Based on the perceived confidence threshold, the category confidence is filtered to determine the confidence short-circuit sensing category. The category confidence of multiple confidence short-circuit sensing categories is normalized and output as short-circuit sensing information for short-circuit processing. The method also includes: The original short-circuit sensing data is input into the fused short-circuit discrimination model to obtain the verification discrimination input; Based on the preset confidence constraints and the verification and discrimination input, statistical analysis is performed with the category confidence as the target to determine the perception confidence threshold corresponding to each short-circuit perception category; The confidence scores of the categories are traversed and threshold filtering is performed. The short-circuit sensing categories that are greater than the confidence score threshold are output as the confident short-circuit sensing categories. Normalize the category confidence of multiple confidence short-circuit sensing categories, obtain the category probability distribution, and merge the multiple confidence short-circuit sensing categories with the category probability distribution to obtain the short-circuit sensing information.

8. The short-circuit sensing and processing method based on the electrical characteristics of a low-impedance voltage transformer as described in claim 1, characterized in that, Also includes: Based on experimental methods, the resonant sensitivity of the target voltage transformer was evaluated, and a fault amplification effect model was established, wherein the fault amplification effect model is a mathematical model. Real-time operating parameters of the target voltage transformer are acquired, and the amplification effect coefficient is obtained by combining the real-time operating parameters with the fault amplification effect model. The perceived confidence threshold is adaptively adjusted based on the amplification effect coefficient.

Images