Power system fault location method and device, storage medium and electronic equipment

Abstract

The application discloses a power system fault positioning method and device, a storage medium and an electronic device. The method comprises the following steps: acquiring real-time time sequence measurement point data of a power line; performing parallel processing on the real-time time sequence measurement point data through at least two filtering channels with different memory attenuation rates, extracting feature components corresponding to each filtering channel from the real-time time sequence measurement point data; performing frequency spectrum analysis on the real-time time sequence measurement point data to obtain frequency domain energy distribution information, determining the output weight of each filtering channel based on the proportion of the energy of each frequency band in the frequency domain energy distribution information and the frequency response characteristic corresponding to the memory attenuation rate of each filtering channel; performing weighted calculation on the feature components corresponding to each filtering channel according to the output weight to obtain a waveform sequence; when the amplitude change rate of the waveform sequence exceeds a preset mutation threshold, it is indicated that a fault occurs in the power line, the arrival time of a transient traveling wave front is identified based on the waveform sequence, and the fault occurrence position of the power line is determined according to the arrival time.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a power system fault location method, device, storage medium and electronic equipment. Background Technology

[0002] Existing power system fault location schemes generally rely on fixed-parameter filters to perform post-processing noise reduction on time-series electrical quantities acquired on-site. These traditional filtering schemes lock the passband and stopband parameters during the factory commissioning phase, and their design logic is based on the premise of long-term stable operation of the power grid. During normal power grid operation, the signal is primarily a 50Hz power frequency waveform, and various random measurement interferences are high-frequency noise. Therefore, the filter sets the low-frequency power frequency range as the passband that allows the signal to pass through, and designates the entire high-frequency range as the stopband where noise needs to be filtered out, thereby eliminating noise and extracting a clean steady-state power frequency waveform.

[0003] However, the power grid is not always in a stable state; it is a non-stationary system whose operating state changes constantly. When a short circuit or lightning strike occurs, the fault point will instantly generate a high-frequency transient traveling wave. This high-frequency traveling wave carries all the key information for determining the location of the fault. Comparing the preset frequency band division of the filter, it can be found that the high-frequency range where the traveling wave is located is exactly within the filter's stopband. At the same time, the passband and stopband parameters of traditional filters are fixed and cannot be automatically adjusted according to the frequency composition of the current real-time signal during equipment operation. This will cause the high-frequency traveling wave of the fault to be treated as useless noise and its signal energy to be greatly weakened when filtered. The original traveling wave initial waveform (traveling wave front), which has a rapidly changing amplitude and clear boundaries, will be smoothed and blunted by the filtering operation.

[0004] The core calculation logic of traveling wave ranging relies on the time difference between the wavefronts captured by measuring points at both ends of the line to deduce the fault distance. Once the wavefront characteristics are smoothed out, the equipment cannot accurately identify the precise time of the wavefront's arrival, resulting in a significant deviation in the time difference value. Ultimately, the error in the calculated fault location increases dramatically, making it difficult to meet the actual work requirements of quickly and accurately locating fault sections in power grid operation and maintenance. Summary of the Invention

[0005] In view of the above problems, this application provides a method, apparatus, storage medium and electronic device for locating faults in a power system.

[0006] To solve the above-mentioned technical problems, this application proposes the following solution: In a first aspect, this application provides a method for locating faults in a power system. The method includes: acquiring real-time time-series measurement data of a power line; processing the real-time time-series measurement data in parallel using at least two filter channels with different memory decay rates, and extracting the feature components corresponding to each filter channel from the real-time time-series measurement data; performing spectral analysis on the real-time time-series measurement data to obtain frequency domain energy distribution information, and determining the output weight of each filter channel based on the proportion of energy in each frequency band in the frequency domain energy distribution information and the frequency response characteristics corresponding to the memory decay rate of each filter channel; performing weighted calculation on the feature components corresponding to each filter channel according to the output weight to obtain a waveform sequence; indicating a fault in the power line when the amplitude change rate of the waveform sequence exceeds a preset abrupt change threshold; identifying the arrival time of the transient traveling wave front based on the waveform sequence; and determining the location of the fault in the power line based on the arrival time.

[0007] This application extracts timing features in parallel through multiple filter channels with different memory decay rates, and then dynamically adjusts the fusion ratio of the two types of components in combination with the real-time signal spectrum. When the power grid is running smoothly, the steady-state waveform is retained first, and high-frequency fault information is automatically amplified when a fault occurs. The waveform retains the steep leading edge of the traveling wave completely and there is no wavefront distortion. This solves the defect of traditional filters in smoothing out transient signals from the signal processing level, and ultimately improves the accuracy of fault location.

[0008] In conjunction with the first aspect, in one possible implementation, the state transition matrix and input matrix of the filtering channel are discretized according to the sampling step size of the real-time time-series measurement data, respectively, to obtain the discretized state transition matrix and input matrix. The discretized state transition matrix and input matrix are substituted into the discrete state-space model to obtain the discrete recursive equation. The unit impulse response of the filtering channel is obtained by solving the discrete recursive equation. The unit impulse response is used to characterize the memory attenuation law of the power line electrical quantity of the filtering channel. The unit impulse response is truncated according to the memory attenuation rate of the filtering channel to generate the filtering coefficient sequence. The filtering coefficient sequence and the real-time time-series measurement data are subjected to fast Fourier transform, respectively. The two sets of frequency domain data obtained by the transform are multiplied point-to-point in the frequency domain. The multiplication result is subjected to inverse fast Fourier transform to obtain the characteristic components corresponding to the filtering channel.

[0009] This application first discretizes the matrix to obtain an iterative recursive formula, and then generates a dedicated filter coefficient sequence based on the memory characteristics of each channel. Directly performing convolution operations in the time domain involves enormous computational costs, making it difficult to meet the real-time processing requirements of substation edge equipment. By using FFT to transform the data to the frequency domain and perform multiplication operations, computational overhead can be significantly reduced. This approach accurately distinguishes between two types of time-series characteristics: slow power flow and transient faults, while adapting to the limitations of on-site hardware computing power, ensuring that real-time data processing is not delayed.

[0010] In conjunction with the first aspect, in one possible implementation, a short-time Fourier transform is performed on the real-time time-series measurement data to obtain the energy spectral density sequence corresponding to each frequency point; the squared amplitude values ​​of each frequency point in the energy spectral density sequence are extracted, and the squared amplitude values ​​of each frequency point are concatenated in frequency order to obtain the frequency domain energy distribution information.

[0011] This application extracts the signal energy spectrum through short-time Fourier transform and then splices the squares of the amplitudes at each frequency point to obtain the complete frequency domain energy distribution. It can simultaneously take into account both time domain and frequency domain information, accurately capture the instantaneous spectrum change of the fault, and intuitively distinguish between the steady-state low-frequency energy of the power grid and the high-frequency energy of the fault.

[0012] In conjunction with the first aspect, in one possible implementation, the frequency domain energy distribution information is aggregated along the frequency dimension to obtain global energy feature components. For each filtering channel, the inner product of the preset coefficient corresponding to the filtering channel and the global energy feature components is calculated. The result of the inner product is added to the bias amount corresponding to the filtering channel to obtain the pre-activation value of the filtering channel. The preset coefficient is used to characterize the correlation between the energy transmittance of the filtering channel in its corresponding passband frequency band and the corresponding frequency band energy value of the global energy feature components. The bias amount is used to adjust the numerical range of the pre-activation value to avoid the pre-activation value falling into the saturation region of the nonlinear function. The pre-activation value is subjected to a nonlinear transformation to obtain the initial weight value of the filtering channel. The initial weight values ​​of all filtering channels are normalized to obtain the output weight of each filtering channel.

[0013] This application relies on real-time frequency domain energy characteristics, combined with the frequency band adaptation coefficients of each filter channel, and the offset to generate dynamic output weights through nonlinear transformation and normalization. When the power grid is running smoothly, the weight of the low-frequency slow-attenuation channel is automatically increased to suppress high-frequency noise. When a fault occurs, the weight of the high-frequency fast-attenuation channel is automatically increased to highlight the transient characteristics of the traveling wave. This enables adaptive weighted fusion of multi-scale components under different operating conditions, significantly improving the fidelity of fault waveforms.

[0014] In conjunction with the first aspect, in one possible implementation, the reconstruction error is determined based on the waveform sequence and a preset power frequency reference signal. The reconstruction error is used to characterize the temporal consistency between the rated power frequency component in the waveform sequence and the preset power frequency reference signal. The frequency domain constraint error is determined based on the waveform sequence and real-time time-series measurement data. The frequency domain constraint error is used to characterize the consistency between the waveform sequence and the real-time time-series measurement data in terms of spectral amplitude and phase. The decoupling error is determined based on the characteristic components output by each filter channel. The decoupling error is used to characterize the degree of feature redundancy between the characteristic components output by different filter channels. The reconstruction error, frequency domain constraint error, and decoupling error are weighted and summed according to preset weights to obtain the total error. The output weights of each filter channel are updated based on the total error.

[0015] This application introduces three types of indicators: reconstruction error, frequency domain constraint error, and decoupling error. Reconstruction error controls the steady-state waveform restoration effect, frequency domain constraint error retains fault spectrum information, and decoupling error prevents the two types of signals from overlapping. The three types of errors are weighted and merged into a unified loss value, which is then used to iteratively update the preset coefficients and bias. Faced with operating condition deviations such as line aging and long-term changes in renewable energy output, the parameters can continuously self-correct, ensuring long-term stability and accuracy in fault waveform extraction and fault location.

[0016] In conjunction with the first aspect, in one possible implementation, fast Fourier transform is performed on the waveform sequence and real-time time-series measurement data respectively to extract the corresponding spectral amplitude sequence and phase sequence; the sum of errors of the spectral amplitude sequence is calculated to obtain the amplitude constraint error; the cosine similarity of the phase sequence is calculated, and the complement of the cosine similarity is used as the phase constraint error; the amplitude constraint error and the phase constraint error are added together to obtain the frequency domain constraint error.

[0017] This application constructs frequency domain constraint error by simultaneously constraining the spectral amplitude error and phase error of the waveform sequence and the original measurement point. This not only preserves the high-frequency energy of the fault traveling wave without attenuation, but also maintains the phase synchronization of the signals at each measurement point, avoiding deviation in traveling wave front identification caused by spectral distortion, and improving the authenticity of fault waveform restoration and the accuracy of fault location.

[0018] In conjunction with the first aspect, in one possible implementation, the cross-correlation value between the characteristic components of any two filter channel outputs is calculated; the sum of the squares of all cross-correlation values ​​is taken as the decoupling error.

[0019] This application calculates the cross-correlation between the characteristic components of each filter channel and takes the sum of squares as the decoupling error. This effectively constrains the redundant correlation of the output characteristics of different channels, allowing the slow-attenuation and fast-attenuation filter channels to perform their respective functions and independently extract steady-state power flow and fault traveling wave characteristics. This avoids the mutual superposition and interference of multi-channel characteristics, improves the purity of the fused waveform, and ensures the accuracy of fault waveform extraction and fault location in the long term.

[0020] Secondly, this application provides a power system fault location device, which includes: The acquisition module is used to acquire real-time time-series measurement data of power lines; The extraction module is used to process real-time time series measurement data in parallel through at least two filter channels with different memory decay rates, and extract the feature components corresponding to each filter channel from the real-time time series measurement data. The first determining module is used to perform spectrum analysis on real-time time-series measurement data to obtain frequency domain energy distribution information. Based on the proportion of energy in each frequency band in the frequency domain energy distribution information and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel, the output weight of each filter channel is determined. The second determining module is used to perform weighted calculations on the feature components corresponding to each filter channel according to the output weights to obtain the waveform sequence. The fault location module is used to indicate a fault in the power line when the amplitude change rate of the waveform sequence exceeds a preset sudden change threshold. It identifies the arrival time of the transient traveling wave front based on the waveform sequence and determines the location of the fault in the power line based on the arrival time.

[0021] To achieve the above objectives, according to a third aspect of this application, a storage medium is provided, the storage medium including a stored program, wherein, when the program is running, the device where the storage medium is located is controlled to execute the power system fault location method of the first aspect.

[0022] To achieve the above objectives, according to a fourth aspect of this application, an electronic device is provided, the device including at least one processor, and at least one memory and bus connected to the processor; wherein the processor and memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the power system fault location method of the first aspect described above.

[0023] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a power system fault location method provided in an embodiment of this application is shown. Figure 2 A schematic diagram of a process for extracting feature components from a filter channel according to an embodiment of this application is shown; Figure 3 This illustration shows a flowchart of a method for determining frequency domain energy distribution information according to an embodiment of this application. Figure 4 A schematic diagram of a weight calculation process provided in an embodiment of this application is shown; Figure 5 A flowchart illustrating a multi-error joint optimization parameter provided in an embodiment of this application is shown; Figure 6This paper shows a schematic diagram of the structure of a power system fault location device provided in an embodiment of this application; Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0025] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0026] In the embodiments of this application, the terms "first," "second," etc., do not have a logical or temporal dependency, nor do they limit the quantity or execution order. It should also be understood that although the following description uses the terms "first," "second," etc., to describe various elements, these elements should not be limited by the terms. These terms are merely used to distinguish one element from another.

[0027] In this application, the term "at least one" means one or more, and the term "multiple" means two or more.

[0028] It should also be understood that the term “if” can be interpreted as “when” or “upon”, or “in response to determination” or “in response to detection”. Similarly, depending on the context, the phrase “if determination…” or “if detection [the stated condition or event]” can be interpreted as “when determination…” or “in response to determination…” or “when detection [the stated condition or event]” or “in response to detection [the stated condition or event]”.

[0029] Power system fault location typically relies on post-filtering of real-time time-series measurement data from power lines. Limited by the architecture of traditional digital filters, this processing often employs a single fixed-coefficient infinite impulse response (IR) or finite impulse response (FIR) filter link. This link is designed based on the assumption that the power grid signal is in a stable state, with preset fixed passband and stopband parameters, aiming to obtain a stable power frequency background signal by filtering out broadband measurement noise. However, at the moment a power grid fault occurs, the real-time time-series measurement data will be superimposed with millisecond-level high-frequency transient components rich in fault information. The characteristic frequencies of these non-stationary signals often fall within the stopband range of the aforementioned fixed parameters. Because the frequency response characteristics of this filter link cannot be dynamically adjusted according to operating conditions after deployment, the high-frequency transient components are significantly attenuated during transmission, even causing the loss of the traveling wave front steepness characterizing the fault location. Ultimately, this leads to deviations in fault location results based on traveling wave theory, making it difficult to meet the application requirements for high-precision fault location.

[0030] Based on this, this application provides a method for locating faults in a power system. The method for locating faults in a power system is described in detail below with reference to the accompanying drawings. Figure 1 This application provides a flowchart illustrating a power system fault location method. Specifically, it includes the following steps: Step 110: Obtain real-time time-series measurement data of the power line.

[0031] In this embodiment, the executing entity can be an edge computing device deployed at a substation or a remote power dispatching master station. The process of acquiring real-time time-series measurement data is as follows.

[0032] First, the primary electrical signals of the power line are sensed by voltage and current transformers pre-installed at the beginning and end of the power line. The voltage and current transformers typically employ electromagnetic or capacitive voltage divider structures, and their secondary outputs are analog signals with a fixed transformation ratio to the primary voltage and current.

[0033] Next, the acquired analog signal is input to the analog input module. The analog input module has a built-in anti-aliasing filter, whose cutoff frequency is set to less than half the sampling frequency to prevent spectral aliasing. Subsequently, the filtered analog signal is discretized and sampled by a high-precision analog-to-digital converter. In this embodiment, the sampling frequency of the analog-to-digital converter is configured to 10kHz, the sampling bit depth is 16 bits, and synchronous sampling is used, i.e., three-phase voltage and three-phase current signals are acquired simultaneously to ensure phase consistency between the phase signals.

[0034] Next, the sampled digital signal is sent to a data buffer. In the data buffer, preliminary preprocessing operations are performed on the raw sampled data. These preprocessing operations include: removing obvious outlier data points, such as abnormal spikes exceeding the normal measurement range caused by transformer saturation or communication interference; and normalizing the data, mapping the sampled values ​​to the standard data range of [-1, 1].

[0035] Finally, the preprocessed data is organized into a real-time time-series measurement point data sequence according to time order. The real-time time-series measurement point data sequence is stored in memory in a first-in-first-out queue. Each set of data contains corresponding timestamp information, which is provided by the Global Positioning System or BeiDou Navigation Satellite System module deployed in the device, ensuring that the time synchronization accuracy of multi-terminal data is better than 1 microsecond.

[0036] After the above processing, real-time time-series measurement data is obtained. The real-time time-series measurement data includes at least a sequence of instantaneous three-phase voltage values ​​and a sequence of instantaneous three-phase current values, serving as the data source for subsequent multi-scale filtering and feature extraction.

[0037] Step 120: Perform parallel processing on the real-time time series measurement data through at least two filter channels with different memory decay rates, and extract the feature components corresponding to each filter channel from the real-time time series measurement data.

[0038] The goal of this step is to separate features at different time scales from real-time time-series measurement data containing multi-scale dynamic characteristics through multiple sets of parallel, customized filtering paths. At least one slow-decay filtering channel is included to specifically retain slow-varying power flow trends lasting from several seconds to several hours, outputting long-time memory components. These long-time memory components mainly represent steady-state power frequency and slow-varying power flow. Steady-state power frequency is the basic rated AC electrical signal under normal operating conditions of the power system. This signal waveform changes periodically and smoothly, and is a fundamental component present for most of the power grid's operation, determining the reference amplitude of line voltage and current. Power flow represents the active and reactive power transmitted on transmission lines. Slow-varying power flow refers to the voltage and current components generated by normal operating disturbances such as load switching and changes in renewable energy output. These components change over hundreds of milliseconds to several hours, with a low rate of change and no abrupt jumps at the millisecond level. They only characterize the slow migration process of power under normal grid operating conditions and are not fault disturbance signals. It includes at least one set of fast-attenuation filter channels specifically designed to capture high-frequency fault traveling waves lasting only a few milliseconds, outputting short-time memory components. These components primarily reflect the high-frequency transient traveling waves excited by the fault. When a transmission line experiences a short circuit, lightning strike, or other fault, the voltage and current at the fault point change abruptly. According to the superposition principle, a sudden step excitation is equivalently superimposed at the fault point, thereby generating voltage and current electromagnetic waves that propagate along the conductor to both ends of the line—the high-frequency transient traveling wave. This signal lasts only a few milliseconds, possesses a steeply abrupt wavefront, contains abundant high-frequency components, and propagates at near the speed of light in overhead lines. In practical applications, more sets of medium-attenuation channels with memory attenuation rates between these two can be added according to monitoring needs, adapting to the extraction requirements of medium-duration dynamic characteristics (such as load steps, intermittent new energy output fluctuations, etc.). To achieve this goal, this application first customizes differentiated signal processing rules for each filter channel, and then implements these rules into an actual filtering process that can run in parallel on edge computing devices. The specific process is as follows.

[0039] First, continuous-time signal processing rules are constructed for each filter channel. The real-time time-series measurement data output in step 110 is a discrete sampling sequence, while various dynamic processes of the power grid belong to continuous electromagnetic dynamic processes, which need to be characterized by a continuous state-space model. The model definitions for each filter channel are unified as follows: In the formula, each variable represents an inherent parameter of the k-th filter channel (k is the channel index, k=1,2,…,K, and K is the total number of filter channels). ), This is a state vector used to store the memory content of the k-th channel for historical input signals. In engineering implementations, N is often set to 64, which is a standard configuration in the field of digital signal processing. In this embodiment, this state vector specifically represents the evolution characteristics of electrical quantities in the power grid (such as voltage and current) at different time scales. The state derivative vector is used to represent the state vector. The instantaneous rate of change with time t. For the state transition matrix, the absolute value of the real part of its eigenvalues ​​determines the retention time of historical information of the state; the smaller the value, the longer the information is retained. The input matrix is ​​used to map real-time time-series measurement data to a state vector. The state space in which it resides. This is the output matrix, used to convert the state vector... Mapping to the output space yields the feature components of the corresponding filter channel. As a through matrix, it is often configured as a zero matrix in engineering implementation to avoid interference from the original input signal on the output characteristics. This represents the real-time time-series measurement point data sequence output in step 110. These are the feature components extracted by the filtering channels. The output of the slow decay filtering channel (with a relatively low memory decay rate and corresponding to a longer time scale) is the long-term memory component, while the output of the fast decay filtering channel (with a relatively high memory decay rate and corresponding to a shorter time scale) is the short-term memory component.

[0040] Secondly, differentiated memory decay rates are configured for each filtering channel to achieve the basic capability of multi-scale feature separation. For each filtering channel, the eigenvalues ​​of the state transition matrix are configured according to its preset memory characteristic requirements. For the slow decay filtering channel used to extract long-term memory components, the absolute value of the real part of its state transition matrix eigenvalue is set to a first value, ranging from 0.001 to 0.1, preferably 0.01, corresponding to the slow decay characteristic, which can retain historical input information for up to hundreds of seconds. For the fast decay filtering channel used to extract short-term memory components, the absolute value of the real part of its state transition matrix eigenvalue is set to a second value, ranging from 1 to 50, preferably 10.0. The second value is greater than the first value, corresponding to the fast decay characteristic, which only retains historical input information for the most recent few milliseconds. For the extended medium decay channel, the absolute value of the real part of its state transition matrix eigenvalue is set to an intermediate value between the first and second values, suitable for the feature extraction requirements of the corresponding duration.

[0041] The matrix parameters of each filter channel are independently initialized using a high-order polynomial projection operator method to ensure that the memory capacity of each channel reaches the theoretical optimal level.

[0042] First, construct the reference continuous state transition matrix. An N*N dimensional reference continuous state transition matrix is ​​constructed based on the Legendre orthogonal polynomial basis. Its matrix elements Assign values ​​according to the following rules. hour, ;when hour, ;when hour, .

[0043] Next, calculate the input matrix and the output matrix. Input matrix Given an N*1 dimensional vector, its nth element Output matrix Configured as an N*1 dimensional identity matrix, used to represent the state vector The mapping is applied to the output characteristics of the corresponding filter channel. The pass-through matrix is ​​configured as a 1*1 zero matrix to avoid interference from the original input signal on the output characteristics.

[0044] Finally, based on the memory rate requirement of the k-th channel, the reference matrix is... Multiply by the corresponding value (first value, second value, or intermediate value) to obtain the final state transition matrix of the channel. At this point, the continuous domain signal processing rules for each filtering channel have been configured, and the system is now capable of separating multi-scale features.

[0045] However, the aforementioned continuous domain rules only exist at the mathematical level and cannot be directly run in the edge computing devices of the power system. This application further transforms the rules into discrete filtering processes that can be executed independently by each channel. All filtering channels complete the engineering implementation according to the following process, and finally realize the extraction of feature components.

[0046] Figure 2 This document illustrates a flowchart of a filter channel feature component extraction method according to an embodiment of this application. Figure 2 As shown.

[0047] Step 210: According to the sampling step size of the real-time time series measurement point data, the state transition matrix and input matrix of the filter channel are discretized respectively to obtain the discretized state transition matrix and input matrix. Substitute the discretized state transition matrix and input matrix into the discrete state space model to obtain the discrete recursive equation.

[0048] Edge computing devices can only perform discrete addition, subtraction, multiplication, and division iterative operations and cannot handle continuous differential equations. Therefore, this application converts continuous rules into channel-specific recursive formulas through discretization. The sampling step size of the real-time time-series measurement data corresponds to the sampling frequency in step 110; when the sampling frequency is 10kHz, the sampling step size is 100μs. Considering that the power grid signal contains broadband components ranging from power frequency to megahertz, using simple discretization methods such as the forward Euler method can easily introduce high-frequency aliasing or low-frequency distortion, leading to the failure of subsequent traveling wave feature identification. Therefore, this application uses the bilinear transform method for discretization. This method approximates the differential operation through trapezoidal numerical integration, which can map the poles of the left half-plane of the continuous system to the unit circle of the z-plane of the discrete system one by one, thereby maintaining the frequency domain amplitude and phase characteristics of the continuous system unchanged during the discretization process, adapting to the modeling requirements of broadband power grid signals.

[0049] The discretization transformation formula is as follows Where s is the complex variable of the Laplace transform, For the Z-transform complex variable, Let be the sampling step size. Substituting this relationship into the continuous state-space model of the k-th channel, the coefficient matrix of the discrete recursive equation for that channel is derived: , In the formula, I is an N-order identity matrix. The discrete state transition matrix, Let be the discrete input matrix. Based on the above coefficient matrix, we obtain the discrete recursive equation applicable to the cycle-by-cycle iteration of a digital signal processor: in, Let k be the state vector of the k-th filter channel at the (n-1)th sampling point. This is the input data for the nth sampling point. This discretization process transforms the continuous dynamic evolution process into discrete matrix iterative operations. While preserving the excellent memory characteristics of continuous systems, it also meets the fixed-point operation requirements of embedded devices, making it a key step in achieving real-time processing at the edge.

[0050] Step 220: Obtain the unit impulse response of the filter channel by solving the discrete recursive equation.

[0051] The unit impulse response is the response of a linear time-invariant system to a unit impulse sequence. The response trajectory (i.e., the input is 1 when n=0 and 0 at other times) fully characterizes the filter's ability to retain the weighted input signal at different times and is the basis for generating the filter's core parameters.

[0052] The process of obtaining the unit impulse response of the filter channel based on solving the discrete recursive equation is as follows. Initial state is set. (With zero initial conditions), iteratively calculate the state response at different sampling points n. When n=0, the input... The initial state is calculated. ;when When, input The state response satisfies the recursive relationship .

[0053] The final output of the filter channel (i.e., the long-time / short-time memory component) is obtained by mapping the memory state through the output matrix. Therefore, the value of the unit impulse response at sampling point n is... n=0,1,…,L k -1, where L k This represents the effective cutoff length of the unit impulse response for this filter channel. Channels with different decay rates exhibit differentiated impulse response characteristics. The state transition matrix eigenvalue of the slow-decay filter channel is close to 1, with extremely slow power-law decay. Its unit impulse response exhibits a long tail characteristic, providing high retention of early input signals and making it suitable for extracting slowly changing power flows. The state transition matrix eigenvalue of the fast-decay filter channel is much less than 1, decaying rapidly to 0 with a power-law decay. Its unit impulse response exhibits a short and sharp characteristic, retaining only recent input signals and making it suitable for extracting rapidly changing fault traveling waves. The characteristics of the medium-decay channel fall between the two, suitable for feature extraction requirements with corresponding durations.

[0054] Step 230: Extract the unit impulse response based on the memory decay rate of the filter channel to generate a filter coefficient sequence.

[0055] Theoretically, the unit impulse response of a slow-decaying filter channel decays extremely slowly, with an effective length approaching infinity. Directly storing the entire response sequence and performing calculations would consume significant storage and computing power from edge devices, failing to meet real-time processing requirements. Therefore, this application independently truncates the unit impulse response of each channel based on energy convergence characteristics. A truncation threshold is set. This threshold is a standard signal truncation criterion in digital signal processing, which can significantly reduce hardware overhead while ensuring feature extraction accuracy. The specific decision logic is as follows: calculate the response energy at the current truncation length L. ,like If the energy of the subsequent response is negligible, then the extension of the response sequence is stopped.

[0056] In practical implementation, for filter channels with different attenuation rates, the effective truncation length L of its response sequence is... k Configure independently as needed. L of the slow-attenuation filter channel. k Using 2000-5000 sampling points, covering a memory duration of hundreds of milliseconds to several seconds, is sufficient to capture slowly changing currents. The L-channel of the fast-attenuation filter... k Using 32-128 sampling points, matched with a rapid attenuation characteristic of a few milliseconds, is sufficient to capture fault traveling waves; the L of the medium attenuation channel kA selection of 128-2000 sampling points, falling between these two ranges, is chosen to suit medium-duration feature extraction needs. The truncated finite-length response sequence (filter coefficient sequence) is the final executable parameter for the k-th channel, and its attenuation characteristics directly determine the filtering characteristics of that channel.

[0057] Step 240: Perform Fast Fourier Transform on the filter coefficient sequence and the real-time time series measurement data respectively. Multiply the two sets of frequency domain data obtained by the transform point by point in the frequency domain. Perform Inverse Fast Fourier Transform on the multiplication result to obtain the feature components corresponding to the filter channel.

[0058] If convolution is performed directly in the time domain, the computational complexity is O(n log n). O( ) Where M is the length of the input data. The length of the filter coefficient sequence obtained in step 230. Processing long data sequences is extremely inefficient and cannot meet the real-time requirements of substation edge devices. Therefore, the computation needs to be converted to the frequency domain based on the convolution theorem. Time-domain linear filtering is equivalent to point-by-point multiplication in the frequency domain, which can significantly reduce computational complexity. Furthermore, the frequency domain operations of each channel can be performed in parallel, further improving the efficiency of multi-channel processing.

[0059] The specific execution process is as follows: For each filter channel, a Fast Fourier Transform (FFT) is first performed on its filter coefficient sequence and real-time time-series measurement data. To avoid frequency domain aliasing, the number of FFT points is... greater than or equal to The most recent power of 2, i.e. ( This is the floor function (rounding up). This value maximizes the use of the butterfly operation structure of the FFT algorithm, improving computational efficiency. Next, the transformed frequency domain sequences are multiplied point by point (the operator is denoted as...). The computational complexity of this step is . The frequency response is significantly lower than that of time-domain convolution. Finally, an inverse fast Fourier transform is performed on the multiplication result to restore the signal from the frequency domain to the time domain, yielding the output sequence y of the k-th channel. k .

[0060] After the above process, the output sequences of all slow attenuation filter channels are summed to obtain the long-time memory component, the output sequences of all fast attenuation filter channels are summed to obtain the short-time memory component, and the output of the medium attenuation channel can be used as an auxiliary feature in subsequent processing.

[0061] Step 130: Perform spectrum analysis on the real-time time-series measurement data to obtain frequency domain energy distribution information. Based on the proportion of energy in each frequency band and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel in the frequency domain energy distribution information, determine the output weight of each filter channel.

[0062] Step 120 outputs long-time memory components, short-time memory components, and an optional medium-time memory component through filter channels with different memory decay rates. The slow-decay filter channel has a filter coefficient sequence length of 2000-5000 sampling points, and its frequency response is low-pass, capable of filtering out high-frequency noise while retaining the slow-changing power flow trend lasting from hundreds of milliseconds to several seconds. The fast-decay filter channel has a filter coefficient sequence of only 32-128 sampling points, and its frequency response is high-pass, capable of capturing millisecond-level fault traveling waves but simultaneously amplifying high-frequency components such as white noise introduced by the sampling link and residual interference from switching operations.

[0063] Using fixed weights to fuse the two types of components will trigger two inherent defects. During steady-state operation, the power grid signal only contains the 50Hz power frequency and a slowly changing power flow trend. The high-frequency noise output from the fast-attenuation filter channel will be superimposed on the fusion result, causing numerous irregular spikes in the power flow trend curve. This makes it impossible to accurately identify slow-changing characteristics such as load ramp-up and frequency drift; in other words, high-frequency noise contaminates the power flow trend. When a fault occurs, the fault traveling wave is an electromagnetic wave generated at the fault point and propagating along the line. Its core identification feature is that the wavefront rises sharply from the steady-state amplitude to several times the peak value within microseconds. The starting point of this rise is the arrival time of the traveling wave. Existing traveling wave identification algorithms generally use threshold detection (e.g., signal amplitude exceeding 3 times the steady-state standard deviation) or slope detection (e.g., the amplitude change rate of 3 consecutive sampling points exceeds a preset threshold) to determine the arrival time. The filter coefficient sequence length of the slow-attenuation filter channel corresponds to a time window of 200~500ms, which essentially performs a weighted average of the input signal within the window. Assuming the peak value of the traveling wave front reaches 10 times the steady-state value, and the wavefront energy is only distributed within 100 sampling points (10ms), while the remaining 1900 sampling points are steady-state values, after long-time channel convolution, the output peak value will be diluted to twice the steady-state value. The steep slope of the original wavefront increasing by 9 times the steady-state value every 100μs will also be smoothed to an increase of 0.45 times the steady-state value every 100μs. At this point, the fusion result cannot reach the amplitude threshold for traveling wave recognition, nor can it meet the slope detection requirements, leading to missed or false detections in subsequent step 150, i.e., the smoothing effect of the long-time channel erases the traveling wave front.

[0064] Therefore, this step dynamically adjusts the weights of each channel based on the real-time spectrum of the real-time time-series measurement data to adapt to the millisecond-level switching requirements of the power grid. The spectrum analysis object is selected from the real-time time-series measurement data output in step 110, rather than any filtered output in step 120, because the filtered output has been colored by the frequency response of the corresponding channel, and performing spectrum analysis based on the colored signal would cause distortion in the gating network perception. The specific execution is carried out as follows.

[0065] Figure 3 This application provides a schematic diagram of a process for determining frequency domain energy distribution information according to an embodiment of the present application. Figure 3 As shown.

[0066] Step 310: Perform a short-time Fourier transform on the real-time time series measurement data to obtain the energy spectral density sequence corresponding to each frequency point.

[0067] Considering that the power grid signal exhibits significant non-stationary characteristics before and after a fault, and that the global Fourier transform cannot capture the transient characteristics of the spectrum over time, this application adopts the short-time Fourier transform, which uses a sliding time window to extract local signals for spectrum analysis, while retaining the two-dimensional characteristics of the signal in terms of time and frequency, thus meeting the monitoring needs of the transient process of power grid faults.

[0068] In the specific implementation, the Hanning window is selected as the sliding window function, with a window length of 256 sampling points, corresponding to an analysis time of 25.6ms at a sampling frequency of 10kHz, slightly longer than two power frequency cycles. This ensures both frequency resolution (approximately 19.5Hz, sufficient to distinguish power frequency and harmonics) and time resolution (6.4ms, sufficient to capture the abrupt changes in millisecond-level fault traveling waves). The window shift is set to 64 sampling points, meaning adjacent windows overlap by 75%, avoiding spectral leakage caused by signal abrupt changes between windows. A 512-point Fast Fourier Transform is performed on the signal within each window to obtain the energy spectral density sequence of the corresponding window position. The calculation formula is as follows: In the formula It is a real-time time-series measurement point data sequence. Here, is the Hanning window function, j is the imaginary unit, W is the window length (value 256), and f is the frequency index. This formula calculates the energy spectral density value corresponding to each frequency point, reflecting the distribution of signal energy across different frequency components at the current moment.

[0069] Step 320: Extract the squared amplitude values ​​of each frequency point in the energy spectral density sequence, and concatenate the squared amplitude values ​​of each frequency point in frequency order to obtain the frequency domain energy distribution information.

[0070] The energy spectral density sequence calculated in step 310 is essentially the result of summing the squares of the real and imaginary parts of the complex spectrum, i.e. This calculation process has already removed the phase information and directly outputs the energy intensity value corresponding to each frequency point. Therefore, this step does not need to repeat the modulus squaring operation, but directly extracts the sequence. The index contains the energy values ​​corresponding to 512 frequency points from 0 to 511.

[0071] Subsequently, these energy values ​​are arranged and concatenated in ascending order of frequency index to construct a one-dimensional vector of length 512. This vector represents the frequency domain energy distribution information, fully depicting the energy spectrum profile of the signal within the 0Hz to 5kHz frequency band in the current analysis window. The lower index region at the beginning of the vector (corresponding to 0~100Hz) has higher values, representing the power frequency and harmonic energy of the power grid in steady-state operation. If a sudden increase in value occurs in the higher index region at the end of the vector (corresponding to 1kHz~5kHz), it represents transient energy such as fault traveling waves or switching operations.

[0072] Since the frequency domain energy distribution information vector is as high as 512 dimensions, directly inputting it into the subsequent decision network would lead to a surge in computational load, exceeding the processing capacity of the substation's edge computing equipment. Furthermore, there is significant energy redundancy among the 512 frequency points (for example, frequencies near the power frequency of 50Hz all have high energy levels, essentially reflecting the same physical phenomenon). Therefore, this application further compresses and refines the high-dimensional frequency domain energy distribution information to extract the core features that best characterize the power grid's operating state. Figure 4 A flowchart illustrating a weight calculation method provided in an embodiment of this application is shown, with the specific steps as follows.

[0073] Step 410: Aggregate the frequency domain energy distribution information along the frequency dimension to obtain the global energy feature components.

[0074] The frequency domain energy distribution information is a high-dimensional vector of length 512. Directly using it for weight calculation would introduce excessive computational overhead. Therefore, this application aggregates along the frequency dimension to reduce the data dimension while retaining the energy characteristics of the core frequency band.

[0075] This application preferably employs a segmented average pooling aggregation method. The frequency range of 0–5 kHz is divided into three frequency bands: 0–100 Hz (low frequency), 100–1000 Hz (mid frequency), and 1 kHz–5 kHz (high frequency). The arithmetic mean of the squared amplitude values ​​within each frequency band is calculated to obtain the three-dimensional global energy feature components. ,in The average energy in the low-frequency band. The average energy in the mid-frequency band. The aggregation method uses high-frequency average energy. It effectively highlights the core frequency band characteristics of the power grid signal: low-frequency energy corresponds to slow-changing power flow trends, high-frequency energy corresponds to fast-changing fault traveling waves, and mid-frequency energy corresponds to medium-duration dynamics such as load steps and intermittent new energy output fluctuations. Simultaneously, it compresses the data dimension from 512 to 3, significantly reducing subsequent computational load and adapting to the computing power constraints of edge computing devices. Besides segmented average pooling, max pooling or single-layer linear mapping can also be used to achieve aggregation, but segmented average pooling requires no training parameters, has the lowest hardware implementation cost, and is the preferred solution for edge deployment.

[0076] Step 420: For each filter channel, calculate the inner product of the preset coefficient corresponding to the filter channel and the global energy feature component, and add the result of the inner product to the bias amount corresponding to the filter channel to obtain the pre-activation value of the filter channel.

[0077] The preset coefficients for each filter channel are pre-set during the initialization phase based on the channel's memory decay rate. Slow-decay filter channels have higher preset coefficients in the low-frequency range and lower values ​​in the high-frequency range; fast-decay filter channels have higher preset coefficients in the high-frequency range and lower values ​​in the low-frequency range; and medium-decay filter channels have higher preset coefficients in the mid-frequency range. This ensures that the inner product result directly reflects the degree of matching between the current spectral characteristics and the memory characteristics of each channel. The preset coefficients and bias values ​​can be specifically adjusted according to the power grid topology type (such as overhead lines, cable lines, and hybrid lines) to adapt to the differences in fault characteristics under different scenarios.

[0078] In this embodiment, a multi-scale orthogonal spectrum matching calibration method is used to complete the offline accurate initialization of preset coefficients, instead of conventional random assignment or fixed constant assignment. First, a standard sample set is constructed offline, consisting of five typical power grid time-series signals: no-load steady state, heavy-load steady state, line lightning fault, metallic short circuit, and new energy power fluctuation. The sampling length of each sample is uniformly set. The sampling frequency is 10kHz. For any type of standard sample time-series data, the corresponding three-dimensional global energy feature components are calculated according to the short-time Fourier transform and frequency band mean aggregation method described in steps 310 to 320 of this application. The feature matrix is ​​formed by combining all the feature components of the samples. Then, the eigencomponents within the matrix are orthogonally decoupled using the Schmitt orthogonalization algorithm. The orthogonalization iterative calculation relationship satisfies: ,in, Let be the k-th orthogonalized basis vector (i.e., the orthogonal standard basis). Let be the k-th eigenvector in the original eigenma matrix that has not been orthogonalized, and j be the summation index, representing the cumulative projection onto the first k-1 orthogonalized bases. For vectors On the orthogonalized base The projection coefficients (i.e., the projection weights in Schmidt orthogonalization) are used to calculate the mutually orthogonal low-frequency, mid-frequency, and high-frequency standard basis vectors. , , Orthogonal basis is used to eliminate data redundancy interference between energy characteristics of different frequency bands.

[0079] For any k-th filter channel, retrieve the filter coefficient sequence generated by truncating the filter channel according to its own memory decay rate. Performing a 512-point Fast Fourier Transform on the filter coefficient sequence yields the complete frequency domain transfer function of the filter channel. The 0–5 kHz monitoring frequency band is divided into a low-frequency band (0–100 Hz), a mid-frequency band (100–1000 Hz), and a high-frequency band (1 kHz–5 kHz). The amplitude of the transfer function within each of the three frequency bands is integrated. The integration formula is as follows: , This represents the cumulative amplitude of the transfer function of the k-th channel in the 0~100Hz low-frequency band, reflecting the channel's transmittance of low-frequency signals. This represents the cumulative amplitude of the transfer function of the k-th channel in the 100-1000Hz mid-frequency band, reflecting the channel's transmittance of mid-frequency signals. Let be the cumulative amplitude of the transfer function of the k-th channel in the 1kHz~5kHz high-frequency band, reflecting the channel's transmittance of high-frequency signals. Based on the three-segment integral results, normalization is performed to obtain a ternary weighted reassembly that characterizes the channel's natural frequency band adaptability. The normalized calculation formula is: In the formula, X represents the three frequency bands: LF, MF, and HF.

[0080] Combining the absolute value of the real part of the eigenvalues ​​of the state transition matrix of the k-th channel Constructing a memory decay correction factor The correction factor calculation relationship is as follows: The slow decay filter channel corresponds to The numerical value is relatively small, and the correction factor is close to 1, which can fully preserve the low-frequency steady-state matching weight. Fast attenuation filter channel. If the values ​​are too large, the correction factor automatically compresses the low-frequency weight ratio to adapt to the high-frequency fault feature extraction requirements. The inner product operation is performed on the three-element weighted recombination of the channel frequency band and the aforementioned orthogonal basis, and then an attenuation correction factor is introduced to complete the scaling, resulting in coarse preset coefficients that have not been finely tuned. The calculation expression is: .

[0081] To further optimize the adaptability of coarse preset coefficients to different operating conditions, two types of quantitative evaluation loss functions were built offline for grid optimization and fine-tuning. The first type is steady-state fidelity loss. The mean square error between the long-term memory component after standard steady-state sample input and the preset 50Hz power frequency reference signal is used as the evaluation index, and the calculation method is as follows: The second category is fault characteristic gain loss. The optimization objective is constructed by taking the negative sum of the squared amplitudes of the short-time memory components corresponding to the standard fault samples. The expression is written as follows: A unified offline optimization objective is constructed by combining the two types of losses. In this embodiment, the perturbation boundary δ is set to 0.05. The optimal correction amount is searched within the small perturbation range using a three-dimensional mesh traversal method. The first-layer preset coefficients used by the channel are obtained by superimposing the coarse preset coefficients with the optimal correction amount. .

[0082] For the k-th filter channel, its first-layer preset coefficients are a three-dimensional vector. The bias is The result of the first-level linear mapping The calculation formula is: In the formula Global energy characteristic components The transpose of . The inner product operation is essentially a weighted matching process. If the current signal has a high proportion of high-frequency energy ( (For larger values), the linear mapping result of the fast attenuation filter channel is... A larger value indicates a higher degree of matching between the channel's memory characteristics and the current transient signal. First layer bias. This is used to calibrate the baseline output of each channel, preventing the inner product result from becoming negative or too small due to spectral fluctuations, and ensuring that the linear mapping result is numerically reasonable. The first-layer linear mapping of all filter channels is implemented using a parallel computing architecture, adapted to... The multi-channel expansion requirement eliminates the need for sequential processing of each channel, significantly improving the computing efficiency of edge devices.

[0083] Step 430: Perform a nonlinear transformation on the pre-activation value to obtain the initial weight value of the filter channel.

[0084] The first-level linear mapping relies solely on linear operations to match the spectrum and channels, lacking nonlinear filtering capabilities. Measurement noise and small load disturbances mixed in the measurement point spectrum can easily lead to negative values ​​in the mapped output or an overall imbalance in the numerical distribution, making it difficult to clearly distinguish the signal intervals corresponding to steady-state and fault conditions. Therefore, this application introduces the GeLU function to perform a nonlinear transformation on the pre-activation value of the linear output to obtain the intermediate characteristics of the channels. By constructing nonlinear distinguishing boundaries, the recognition and resolution of complex mixed spectra can be improved.

[0085] The GeLU function is selected to adapt to the low computing power constraints of substation edge computing equipment. This function has lightweight computational logic, maintaining real-time computing speed while ensuring nonlinear fitting effects. GeLU can automatically weaken the channel weight contribution corresponding to negative mapping results, achieving adaptive masking of invalid channel outputs. This aligns with the operating conditions where power grid faults occur only momentarily and operate stably for most of the time: During the steady-state operation of the power grid, the pre-activation values ​​corresponding to the fast attenuation filter channel are mostly negative, and after transformation, the intermediate feature values ​​approach 0, automatically reducing the proportion of high-frequency component fusion. At the instant of a short circuit or lightning strike fault, the high-frequency energy of the signal increases sharply, the pre-activation value of the fast attenuation filter channel turns positive, and the intermediate feature values ​​rise synchronously, automatically amplifying the channel contribution corresponding to the fault traveling wave.

[0086] Obtain the nonlinear intermediate characteristics of each channel Then, a second layer of preset coefficients is introduced. With the second layer bias Perform a quadratic linear operation, the operation relation is: Calculated This is the original score value corresponding to the channel. The value directly represents the degree of matching between the spectrum of the current measurement point and the memory decay characteristics of this channel. The higher the original score value, the higher the weight that the output component of this channel should be assigned in the subsequent waveform fusion process.

[0087] Step 440: Normalize the initial weight values ​​of all filter channels to obtain the output weights of each filter channel.

[0088] The raw scores calculated independently for each channel have unrestricted ranges and their sum is not equal to 1, making them unsuitable for direct weighted summation of multi-channel convolution outputs. Therefore, the Softmax function is used for normalization to generate a K-dimensional dynamic weight vector. After normalization, the sum of all channel weights equals 1. The higher the original score of a single channel, the larger the corresponding output weight, and the greater the influence of that channel in the final fused features.

[0089] The formula for calculating the dynamic output weight corresponding to the kth filter channel is: Softmax normalization maps all weights to the (0,1) interval, simultaneously amplifying the weight advantage of high-resolution channels. Under steady-state conditions, the output weights of the slow-decaying filter channels approach 1, and the fusion features are dominated by long-term power flow components. During fault-prone instantaneous faults, the weights of the fast-decaying filter channels rapidly increase, and the fusion features highlight short-term fault traveling wave components, achieving adaptive fusion of multi-scale spectral features.

[0090] Step 140: Calculate the weighted feature components corresponding to each filter channel according to the output weights to obtain the waveform sequence.

[0091] After obtaining the output weights of each filter channel through normalization in step 440, hierarchical weighted fusion is performed based on the output components corresponding to each group of filter channels to remove the steady-state power frequency power flow background and output a pure waveform sequence. The specific process is as follows.

[0092] When there are more than two filter channels, firstly, all filter channels are divided into slow-attenuation filter channel groups and fast-attenuation filter channel groups. If medium-attenuation channels are configured, an intermediate-frequency channel group is added. For the slow-attenuation filter channel groups, the long-time memory components within each group are weighted and summed using the corresponding output weights to obtain the steady-state power flow reference components. For the fast-decaying filter channel group, the short-time memory components within each group are weighted and summed using the corresponding output weights to obtain the transient fundamental components. If a moderate attenuation channel exists, the intermediate frequency auxiliary component can be obtained by weighting in the same way. .

[0093] Next, the total output weights of each group of channels are calculated, and global waveform fusion is performed. The output weights of the slow attenuation, fast attenuation, and selectable medium attenuation channels are summed separately to obtain the total weights for each group. , , The sum of the total weights of each group is 1. Based on the total weights of each group, the grouped fusion components are linearly superimposed to obtain the global fusion waveform. .

[0094] Next, waveform sequences are extracted through differential decoupling. The globally fused waveform contains both steady-state power flow and fault transient components. By subtracting the steady-state power flow reference component from the globally fused waveform and eliminating slowly changing power frequency background signals, only the characteristics of abrupt faults are retained, resulting in the waveform sequence. .

[0095] Finally, boundary correction is performed on the waveform sequence and the output is buffered. Due to the difference in truncation length of the filter coefficient sequences of each channel, there are distorted sampling points at the beginning and end of the sequence. The maximum truncation length of all filter coefficient sequences is taken. Before and after the resection sequence One distortion sampling point is used to complete waveform boundary correction.

[0096] Under steady-state conditions, the proportion of high-frequency energy in the signal is extremely low, the total weight of the fast-attenuation filter channel approaches 0, and the amplitude of the decoupled waveform sequence is close to the baseline noise level. At the moment of fault occurrence, the high-frequency energy increases sharply, the weight of the fast-attenuation filter channel increases significantly, the waveform sequence retains the steep traveling wave front completely, and there is no waveform smoothing distortion.

[0097] To continuously optimize the gated network parameters such as preset coefficients and bias values ​​of each filter channel, improve the adaptive fusion effect of multi-scale components, and reduce the coupling interference between steady-state power flow and fault transient components, this application constructs a multi-dimensional joint error constraint system and supervises model training through multi-layer error index joint supervision. Figure 5 A flowchart illustrating a multi-error joint optimization parameter according to an embodiment of this application is shown. The specific execution flow is shown in steps 510 to 550.

[0098] Step 510: Determine the reconstruction error based on the waveform sequence and the preset power frequency reference signal.

[0099] First, a reconstruction error is constructed to constrain the steady-state power flow restoration effect. This application pre-defines a 50Hz power frequency reference signal, which is generated by an ideal sine function, expressed as follows: ,in, Here, n is the preset power frequency reference signal frequency, and n is the sampling point index. Let be the sampling step size, and the total length of the reference signal sequence be exactly the same as the extracted long-time memory component. Let the waveform sequence corresponding to the nth sampling point be... The reconstruction error is obtained by calculating the difference between the waveform sequence and the preset power frequency reference signal point by point, and then calculating the square mean of the differences of all sampling points. The calculation formula is as follows: .

[0100] This calculation method uses the average of the squared differences of all sampling points. First, the deviation of the waveform sequence from the power frequency reference at each sampling point is squared to eliminate the influence of the deviation direction. This ensures that the contribution of a single-point deviation is limited to a reasonable range, preventing individual extreme values ​​from dominating the overall error. Then, the arithmetic mean of the squared deviations of all N sampling points is calculated. This effectively distributes the large error caused by a single-point mutation evenly across all sampling points in the entire time series, avoiding a situation where the overall error is determined solely by a few mutation sampling points. This method can reflect the overall deviation of the waveform from the reference over the entire time period in a balanced way. If methods such as taking only the maximum single-point deviation or summing within local intervals were used, a small instantaneous disturbance at one or two sampling points (such as transformer noise or residual interference from switching operations) would significantly increase the overall error value, leading to a biased evaluation result. The average method, however, integrates the amplitude differences across all time points, objectively measuring the level of the waveform sequence's adherence to the standard 50Hz power frequency throughout the entire duration.

[0101] Step 520: Determine the frequency domain constraint error based on the waveform sequence and real-time time-series measurement data.

[0102] In one implementation, the waveform sequence and real-time time-series measurement data are subjected to Fast Fourier Transform to extract the corresponding spectral amplitude sequence and phase sequence; the sum of errors of the spectral amplitude sequence is calculated to obtain the amplitude constraint error; the cosine similarity of the phase sequence is calculated, and the complement of the cosine similarity is used as the phase constraint error; the amplitude constraint error and the phase constraint error are added together to obtain the frequency domain constraint error.

[0103] The frequency domain constraint error is constructed here to ensure that the spectral characteristics of the fault waveform and the original measurement point signal remain consistent. Specifically, the waveform sequences are respectively... The real-time time series measurement data is used to perform a Fast Fourier Transform (FFT). After unifying the lengths of the two sets of sequences and filling in the missing sampling points, the corresponding spectral amplitude sequences are extracted from each set. , and phase sequence , Using f as the frequency index and F as the total number of frequency points, first calculate the sum of the absolute errors of the amplitudes of the two sets of spectra point by point as the amplitude constraint error. The formula for calculating amplitude constraint error is as follows: Next, calculate the cosine similarity between the two phase sequences, and take the similarity complement as the phase constraint error. The formula for calculating the phase constraint error is: The frequency domain constraint error is obtained by adding the two errors together. The formula for calculating the frequency domain constraint error is as follows: .

[0104] Frequency domain constraint error constrains the fault waveform spectrum distribution to remain consistent with the original measurement point in terms of both amplitude and phase, thus avoiding the loss of high-frequency fault characteristics and synchronization phase distortion.

[0105] Step 530: Determine the decoupling error based on the characteristic components output by each filter channel.

[0106] In one implementation, the cross-correlation value between the characteristic components output by any two filter channels is calculated; the sum of the squares of all cross-correlation values ​​is taken as the decoupling error. Specifically, let the total number of filter channels be K, and the time sequence of the characteristic components output by the k-th filter channel be... The timing length of each channel is uniformly N. First, for any two different groups of channels k and j (k The zero-delay normalized cross-correlation value is obtained from the eigencomponents of j). Iterate through all cases that satisfy k. The channel combination of j yields all cross-correlation values. The squares of each cross-correlation value are then summed, and the result is the decoupling error. The calculation formula is: .

[0107] The decoupling error is used to characterize the degree of feature redundancy between the output feature components of different filtering channels. The stronger the correlation between channel features, the larger the decoupling error value. It is then used in the weighted calculation of the total error and the preset coefficients and bias are updated in reverse iteration to constrain the output features of each channel to decouple from each other and reduce information redundancy.

[0108] Step 540: The reconstruction error, frequency domain constraint error, and decoupling error are weighted and summed according to preset weights to obtain the total error.

[0109] After obtaining the three categories of errors—reconstruction error, frequency domain constraint error, and decoupling error—these three errors are weighted and fused to obtain a unified total error. The weighting coefficients for the three types of errors are pre-calibrated offline according to different power grid topologies such as overhead lines, cables, and hybrid lines. , The total error, λ3, is obtained through weighted fusion and is in a unified scalar form. The calculation formula is as follows: .

[0110] The three types of errors control the fidelity of steady-state waveforms, the integrity of fault spectrum, and the independence of multi-scale components, respectively. A single error can only achieve local optimization. After weighted fusion, the weight ratio of each constraint can be flexibly adjusted according to the power grid scenario. For example, the frequency domain error weight can be increased for cable lines, and the decoupling error constraint can be strengthened for overhead lines to adapt to the inherent characteristics of different power grid signals.

[0111] Step 550: Update the output weights of each filter channel based on the total error.

[0112] Finally, using the total error as the loss function, backpropagation is used to update the preset coefficients and biases of the gated network. A fixed learning rate η is set, and the total error is used as the loss function. As the global loss function, the gradient is calculated layer by layer along the gated network to obtain the preset coefficients of each filtering channel. Bias The gradient value is used to perform iterative parameter updates using the gradient descent method, and the update formula is: .

[0113] By relying on the unified total error to complete the gradient update of the entire set of gating parameters, there is no need to optimize different parameters separately in batches. After the update is completed, the preset coefficients and biases are stored in the local dual-port RAM and can be directly called in the next round of time-series spectrum weight calculation. This enables continuous adaptive optimization as long-term operating conditions change, such as line aging and fluctuations in new energy output, and steadily reduces fault location errors caused by steady-state noise, waveform distortion, and component aliasing.

[0114] Step 150: When the amplitude change rate of the waveform sequence exceeds the preset sudden change threshold, it indicates that a fault has occurred in the power line. Based on the waveform sequence, the arrival time of the transient traveling wave front is identified, and the location of the power line fault is determined according to the arrival time.

[0115] The waveform sequence after boundary correction fully preserves the high-frequency transient traveling wave pulses generated by the fault. The waveform amplitude change rate directly reflects the degree of instantaneous abrupt change in the signal. Under normal steady-state conditions, only small noise fluctuations exist, and the amplitude change rate is relatively small. Once a fault such as a short circuit or lightning strike occurs on the line, the transient traveling wave excited at the fault point will cause the waveform to experience a sudden and sharp rise / fall, with a significant increase in the amplitude change rate. Therefore, a sudden change threshold is preset as a discrimination criterion. When the amplitude change rate at any position in the waveform sequence exceeds the preset sudden change threshold, it can be determined that there is a sudden disturbance in the line, directly indicating that a fault has occurred in the power line. Based on this, the wavefront time is captured by relying on the steep abrupt change characteristics of the waveform, and the fault distance is calculated by combining the line wave velocity and the time difference of the measurement point. The specific calculation process is as follows.

[0116] Let the corrected waveform sequence be... The sequence sampling interval is The corresponding sampling point timestamp is First, calculate the slope of the change in amplitude for each sampling point. Preset traveling wave steep slope threshold Iterate through all sampling points and select the first one that satisfies the condition. The timestamp corresponding to the sampling point is recorded as the arrival time of the transient traveling wavefront. If measuring points are deployed at both ends of the line and fault waveforms are collected synchronously, the timing of the first wavefront at each end can be obtained. Terminal wavefront time Let v be the speed of the traveling wave propagating along the line, L be the total length of the line, and the time difference of the traveling wave between the two measuring points be v. The formula for calculating the fault distance at one end of the measuring point is: .

[0117] Identifying the traveling wave front using slope thresholding effectively filters out minute waveform jitter caused by low-frequency noise, locking only the steep pulse leading edge generated by the fault and ensuring accurate wave front acquisition. Dual-end traveling wave time-difference ranging fully avoids the difficulty of wave reflection identification in single-end ranging. Combined with microsecond-level timestamps synchronized with BeiDou / GPS, it can significantly reduce fault location errors, accurately outputting the line fault location and providing accurate coordinate data for maintenance and inspection.

[0118] It is understood that, in order to achieve the functions in the above embodiments, the computer device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and method steps described in conjunction with the embodiments disclosed in this application, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application scenario and design constraints of the technical solution.

[0119] Furthermore, as a response to the above Figure 1 The implementation of the method embodiment shown in this application provides a power system fault location device. This device embodiment corresponds to the foregoing method embodiments. For ease of reading, this embodiment will not repeat the details of the foregoing method embodiments one by one, but it should be understood that the device in this embodiment can correspondingly implement all the contents of the foregoing method embodiments. Specifically, as shown... Figure 6 As shown, the power system fault location device 600 includes: The acquisition module 610 is used to acquire real-time time-series measurement data of power lines; The extraction module 620 is used to process real-time time series measurement data in parallel through at least two filter channels with different memory decay rates, and extract the feature components corresponding to each filter channel from the real-time time series measurement data. The first determining module 630 is used to perform spectrum analysis on real-time time-series measurement point data to obtain frequency domain energy distribution information. Based on the proportion of energy in each frequency band in the frequency domain energy distribution information and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel, the output weight of each filter channel is determined. The second determining module 640 is used to perform weighted calculations on the feature components corresponding to each filter channel according to the output weights to obtain a waveform sequence. The fault location module 650 is used to indicate that a power line has a fault when the amplitude change rate of the waveform sequence exceeds a preset sudden change threshold. It identifies the arrival time of the transient traveling wave front based on the waveform sequence and determines the location of the power line fault based on the arrival time.

[0120] Furthermore, such as Figure 6 As shown, the extraction module 620 is specifically used to discretize the state transition matrix and input matrix of the filter channel according to the sampling step size of the real-time time-series measurement data, respectively, to obtain the discretized state transition matrix and input matrix. The discretized state transition matrix and input matrix are substituted into the discrete state-space model to obtain the discrete recursive equation. The unit impulse response of the filter channel is obtained by solving the discrete recursive equation. The unit impulse response is used to characterize the memory attenuation law of the power line electrical quantity of the filter channel. The unit impulse response is truncated according to the memory attenuation rate of the filter channel to generate the filter coefficient sequence. The filter coefficient sequence and the real-time time-series measurement data are subjected to fast Fourier transform respectively. The two sets of frequency domain data obtained by the transform are multiplied point-to-point in the frequency domain. The multiplication result is subjected to inverse fast Fourier transform to obtain the characteristic components corresponding to the filter channel.

[0121] Furthermore, such as Figure 6As shown, the first determining module 630 is specifically used to perform short-time Fourier transform on the real-time time series measurement data to obtain the energy spectral density sequence corresponding to each frequency point; extract the amplitude square value of each frequency point in the energy spectral density sequence, and splice the amplitude square values ​​of each frequency point in frequency order to obtain the frequency domain energy distribution information.

[0122] Furthermore, such as Figure 6 As shown, the first determining module 630 is specifically used to aggregate the frequency domain energy distribution information along the frequency dimension to obtain global energy feature components; for each filtering channel, it calculates the inner product of the preset coefficient corresponding to the filtering channel and the global energy feature components, adds the result of the inner product to the bias amount corresponding to the filtering channel to obtain the pre-activation value of the filtering channel. The preset coefficient is used to characterize the correlation between the energy transmittance of the filtering channel in its corresponding passband frequency band and the corresponding frequency band energy value of the global energy feature components. The bias amount is used to adjust the numerical range of the pre-activation value to avoid the pre-activation value falling into the saturation region of the nonlinear function; it performs a nonlinear transformation on the pre-activation value to obtain the initial weight value of the filtering channel; and it normalizes the initial weight values ​​of all filtering channels to obtain the output weight of each filtering channel.

[0123] Furthermore, such as Figure 6 As shown, the second determining module 640 is further configured to: determine the reconstruction error based on the waveform sequence and the preset power frequency reference signal, wherein the reconstruction error is used to characterize the time-domain consistency between the rated power frequency component in the waveform sequence and the preset power frequency reference signal; determine the frequency domain constraint error based on the waveform sequence and real-time time-series measurement point data, wherein the frequency domain constraint error is used to characterize the consistency between the waveform sequence and the real-time time-series measurement point data in terms of spectral amplitude and phase; determine the decoupling error based on the characteristic components output by each filter channel, wherein the decoupling error is used to characterize the degree of characteristic redundancy between the characteristic components output by different filter channels; perform a weighted summation of the reconstruction error, frequency domain constraint error, and decoupling error according to preset weights to obtain the total error; and update the output weights of each filter channel based on the total error.

[0124] Furthermore, such as Figure 6 As shown, the second determining module 640 is also used to perform fast Fourier transform on the waveform sequence and real-time time-series measurement point data respectively, extract the corresponding spectral amplitude sequence and phase sequence; calculate the sum of errors of the spectral amplitude sequence to obtain the amplitude constraint error; calculate the cosine similarity of the phase sequence, and use the complement of the cosine similarity as the phase constraint error; add the amplitude constraint error and the phase constraint error to obtain the frequency domain constraint error.

[0125] Furthermore, such as Figure 6 As shown, the second determining module 640 is also used to calculate the cross-correlation value between the characteristic components of any two filter channel outputs; and to take the sum of the squares of all cross-correlation values ​​as the decoupling error.

[0126] Optionally, the power system fault location device may be an electronic device with data processing capabilities, or a functional module within such electronic device, without limitation.

[0127] For example, the electronic device can be a server, which can be a single server or a server cluster consisting of multiple servers. As another example, the electronic device can be a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, ultra-mobile personal computer (UMPC), netbook, as well as cellular phones, personal digital assistants (PDAs), augmented reality (AR) devices, virtual reality (VR) devices, and other terminal devices. As yet another example, the electronic device can also be a recording device, video surveillance equipment, etc. This application does not impose any special limitations on the specific form of the electronic device.

[0128] The following example uses an electronic device as the fault location device in a power system. Figure 7 As shown, Figure 7 The hardware structure of an electronic device 700 provided in this application.

[0129] like Figure 7 As shown, the electronic device 700 includes a processor 710, a communication line 720, and a communication interface 730.

[0130] Optionally, the electronic device 700 may also include a memory 740. The processor 710, memory 740, and communication interface 730 can be connected via a communication line 720.

[0131] The processor 710 can be a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. The processor 710 can also be any other device with processing capabilities, such as a circuit, device, or software module, without limitation.

[0132] In one example, processor 710 may include one or more CPUs, for example Figure 7 CPU0 and CPU1 in the CPU.

[0133] As an optional implementation, the electronic device 700 includes multiple processors; for example, in addition to processor 710, it may also include processor 770. A communication line 720 is used to transmit information between the components included in the electronic device 700.

[0134] Communication interface 730 is used for communicating with other devices or other communication networks. These other communication networks can be Ethernet, Radio Access Network (RAN), Wireless Local Area Networks (WLAN), etc. Communication interface 730 can be a module, circuit, transceiver, or any device capable of enabling communication.

[0135] Memory 740 is used to store instructions. These instructions can be computer programs.

[0136] The memory 740 can be a read-only memory (ROM) or other type of static storage device that can store static information and / or instructions; it can also be a random access memory (RAM) or other type of dynamic storage device that can store information and / or instructions; it can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, etc., without limitation.

[0137] It should be noted that the memory 740 can exist independently of the processor 710, or it can be integrated with the processor 710. The memory 740 can be used to store instructions, program code, or some data, etc. The memory 740 can be located inside or outside the electronic device 700, without restriction.

[0138] The processor 710 is configured to execute instructions stored in the memory 740 to implement the communication method provided in the following embodiments of this application. For example, when the electronic device 700 is a terminal or a chip in a terminal, the processor 710 can execute instructions stored in the memory 740 to implement the steps performed by the transmitting end in the following embodiments of this application.

[0139] As an optional implementation, the electronic device 700 also includes an output device 750 and an input device 760. The output device 750 can be a display screen, speaker, or other device capable of outputting data from the electronic device 700 to the user. The input device 760 can be a keyboard, mouse, microphone, joystick, or other device capable of inputting data into the electronic device 700.

[0140] It should be pointed out that, Figure 7 The structure shown does not constitute a limitation on the electronic device, except... Figure 7 In addition to the components shown, the electronic device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0141] The power system fault location device and application scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of power system fault location devices and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0142] This application provides a storage medium storing a program that, when executed by a processor, implements the power system fault location method.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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 processor, 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0145] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for locating faults in a power system, characterized in that, The method includes: Acquire real-time time-series measurement data of power lines; The real-time time series measurement data is processed in parallel by at least two filter channels with different memory decay rates, and the feature components corresponding to each filter channel are extracted from the real-time time series measurement data. Spectral analysis is performed on the real-time time-series measurement data to obtain frequency domain energy distribution information. Based on the proportion of energy in each frequency band in the frequency domain energy distribution information and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel, the output weight of each filter channel is determined. The characteristic components corresponding to each filter channel are weighted according to the output weights to obtain the waveform sequence. When the amplitude change rate of the waveform sequence exceeds a preset abrupt change threshold, it indicates that a fault has occurred in the power line. The arrival time of the transient traveling wave front is identified based on the waveform sequence, and the location of the fault in the power line is determined according to the arrival time.

2. The method according to claim 1, characterized in that, The real-time time-series measurement data is processed in parallel using at least two filter channels with different memory decay rates. Feature components corresponding to each filter channel are extracted from the real-time time-series measurement data, including: According to the sampling step size of the real-time time series measurement point data, the state transition matrix and input matrix of the filter channel are discretized respectively to obtain the discretized state transition matrix and input matrix. The discretized state transition matrix and input matrix are substituted into the discrete state space model to obtain the discrete recursive equation. The unit impulse response of the filter channel is obtained by solving the discrete recursive equation. The unit impulse response is used to characterize the memory attenuation law of the power line electrical quantity of the filter channel. The unit impulse response is extracted based on the memory decay rate of the filter channel to generate a filter coefficient sequence; The filter coefficient sequence and the real-time time-series measurement data are subjected to Fast Fourier Transform (FFT) respectively. The two sets of frequency domain data obtained by the transform are multiplied point-to-point in the frequency domain. The multiplication result is subjected to Inverse Fast Fourier Transform (IFFT) to obtain the feature components corresponding to the filter channel.

3. The method according to any one of claims 1-2, characterized in that, Spectral analysis is performed on the real-time time-series measurement data to obtain frequency domain energy distribution information, including: The real-time time-series measurement data is subjected to a short-time Fourier transform to obtain the energy spectral density sequence corresponding to each frequency point; The amplitude squared value of each frequency point in the energy spectral density sequence is extracted, and the amplitude squared values ​​of each frequency point are spliced ​​together in frequency order to obtain the frequency domain energy distribution information.

4. The method according to claim 3, characterized in that, Based on the proportion of energy in each frequency band in the frequency domain energy distribution information and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel, the output weight of each filter channel is determined, including: The frequency domain energy distribution information is aggregated along the frequency dimension to obtain global energy feature components; For each filter channel, the inner product of the preset coefficient corresponding to the filter channel and the global energy feature component is calculated. The result of the inner product is added to the bias amount corresponding to the filter channel to obtain the pre-activation value of the filter channel. The preset coefficient is used to characterize the correlation between the energy transmittance of the filter channel in its corresponding passband frequency band and the corresponding frequency band energy value of the global energy feature component. The bias amount is used to adjust the numerical range of the pre-activation value to avoid the pre-activation value falling into the saturation region of the nonlinear function. The pre-activation value is subjected to a nonlinear transformation to obtain the initial weight value of the filter channel; The initial weight values ​​of all filter channels are normalized to obtain the output weights of each filter channel.

5. The method according to claim 1, characterized in that, After obtaining the waveform sequence, the method further includes: Based on the waveform sequence and the preset power frequency reference signal, the reconstruction error is determined. The reconstruction error is used to characterize the time domain consistency between the rated power frequency component in the waveform sequence and the preset power frequency reference signal. Based on the waveform sequence and the real-time time-series measurement data, a frequency domain constraint error is determined. The frequency domain constraint error is used to characterize the consistency between the waveform sequence and the real-time time-series measurement data in terms of spectral amplitude and phase. Based on the characteristic components output by each filter channel, the decoupling error is determined. The decoupling error is used to characterize the degree of feature redundancy between the characteristic components output by different filter channels. The reconstruction error, the frequency domain constraint error, and the decoupling error are weighted and summed according to preset weights to obtain the total error. The output weights of each filter channel are updated based on the total error.

6. The method according to claim 5, characterized in that, Based on the waveform sequence and the real-time time-series measurement data, the frequency domain constraint error is determined, including: Fast Fourier transform is performed on the waveform sequence and the real-time time-series measurement data respectively to extract the corresponding spectral amplitude sequence and phase sequence; The sum of errors in the spectral amplitude sequence is calculated to obtain the amplitude constraint error; Calculate the cosine similarity of the phase sequence, and use the complement of the cosine similarity as the phase constraint error; The amplitude constraint error is added to the phase constraint error to obtain the frequency domain constraint error.

7. The method according to claim 5, characterized in that, Based on the characteristic components of the output of each filter channel, the decoupling error is determined, including: Calculate the cross-correlation value between the characteristic components of any two filter channel outputs; The sum of squares of all cross-correlation values ​​is taken as the decoupling error.

8. A power system fault location device, characterized in that, The device includes: The acquisition module is used to acquire real-time time-series measurement data of power lines; The extraction module is used to process the real-time time series measurement data in parallel through at least two filter channels with different memory decay rates, and extract the feature components corresponding to each filter channel from the real-time time series measurement data. The first determining module is used to perform spectrum analysis on the real-time time-series measurement data to obtain frequency domain energy distribution information, and determine the output weight of each filter channel based on the proportion of energy in each frequency band and the frequency response characteristics corresponding to the memory attenuation rate of each filter channel in the frequency domain energy distribution information. The second determining module is used to perform weighted calculations on the feature components corresponding to each filter channel according to the output weights to obtain a waveform sequence; The fault location module is used to indicate that a fault has occurred in the power line when the amplitude change rate of the waveform sequence exceeds a preset sudden change threshold, identify the arrival time of the transient traveling wave front based on the waveform sequence, and determine the location of the fault in the power line based on the arrival time.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to perform the power system fault location method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the power system fault location method as described in any one of claims 1-7.

Images