A power distribution network multi-modal current fault identification method and system

Abstract

The application discloses a power distribution network multi-modal current fault identification method and system, and relates to the technical field of fault prediction and health management. The method comprises the following steps: acquiring a zero sequence voltage signal and a zero sequence current signal of a power distribution network; when the zero sequence voltage signal exceeds a preset starting threshold, extracting high-frequency transient pulses from the zero sequence current signal, constructing a pulse time interval sequence, and extracting pulse interval periodicity features; extracting a zero sequence voltage amplitude feature from the zero sequence voltage signal and a phase domain imbalance feature from three-phase current signals of the power distribution network; calculating a dynamic integration coefficient according to the pulse interval periodicity features and the zero sequence voltage amplitude feature; calling a corresponding number of fault identification models from a pre-constructed fault identification model array, outputting fault determination sub-results, and performing majority consistent determination on each fault determination sub-result to obtain a fault determination result. The application effectively improves the detection sensitivity and determination reliability of high-impedance grounding faults under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, specifically to a method and system for identifying multimodal current faults in power distribution networks. Background Technology

[0002] Existing single-phase grounding fault detection technologies in distribution networks typically involve configuring voltage and current transformers at line monitoring points to collect zero-sequence voltage and zero-sequence current signals in real time. When the zero-sequence voltage exceeds the activation threshold, fault identification is initiated. The measured zero-sequence current amplitude is compared with a preset action threshold; if it exceeds the limit, the line is identified as faulty and an alarm is issued.

[0003] However, when a high-impedance ground fault occurs in the distribution network, there is a large impedance between the fault point and the ground, resulting in an extremely low fault current amplitude, which cannot be detected by traditional overcurrent protection methods based on current amplitude thresholds. Although the fault current contains weak high-frequency transient pulses generated by intermittent arcs, the regularity of the pulse interval is affected by various factors such as power grid background noise and arc stability. A single fault identification model is difficult to accurately determine under all operating conditions, and is prone to misjudgment or missed judgment. Summary of the Invention

[0004] This invention provides a method and system for identifying multi-mode current faults in power distribution networks, aiming to solve the technical problems in the prior art where traditional overcurrent protection is difficult to detect high-impedance grounding faults and single fault identification models are prone to misjudgment and omission.

[0005] In view of the above problems, the present invention provides a method and system for identifying multi-mode current faults in power distribution networks.

[0006] In a first aspect, the present invention provides a method for identifying multi-mode current faults in a power distribution network, comprising: The zero-sequence voltage signal and zero-sequence current signal of the distribution network are acquired. When the zero-sequence voltage signal exceeds a preset start-up threshold, a high-frequency transient pulse is extracted from the zero-sequence current signal, a pulse time interval sequence is constructed, and the periodicity feature of the pulse interval is extracted. Extract the zero-sequence voltage amplitude feature from the zero-sequence voltage signal, and extract the phase domain imbalance feature from the three-phase current signal of the distribution network; Calculate the dynamic integration coefficient based on the periodicity of the pulse interval and the amplitude characteristics of the zero-sequence voltage; Based on the dynamic integration coefficient, a corresponding number of fault identification models are called from the pre-constructed fault identification model array. Each called fault identification model outputs a fault determination sub-result based on the pulse interval periodicity characteristic, the zero-sequence voltage amplitude characteristic, and the phase domain imbalance characteristic. The majority consensus judgment is performed on each fault determination sub-result to obtain the determination result of whether the distribution network has a high impedance grounding fault.

[0007] Secondly, the present invention provides a multi-mode current fault identification system for power distribution networks, comprising: The pulse period feature extraction module is used to acquire the zero-sequence voltage signal and the zero-sequence current signal of the distribution network. When the zero-sequence voltage signal exceeds the preset start-up threshold, the module extracts high-frequency transient pulses from the zero-sequence current signal, constructs a pulse time interval sequence, and extracts the periodic features of the pulse interval. The steady-state electrical feature extraction module is used to extract the zero-sequence voltage amplitude feature from the zero-sequence voltage signal and extract the phase domain imbalance feature from the three-phase current signal of the distribution network. The dynamic integration coefficient calculation module is used to calculate the dynamic integration coefficient based on the periodicity characteristics of the pulse interval and the amplitude characteristics of the zero-sequence voltage. The fault determination result output module is used to call a corresponding number of fault identification models from the pre-constructed fault identification model array according to the dynamic integration coefficient. Each called fault identification model outputs a fault determination sub-result based on the pulse interval periodicity feature, the zero-sequence voltage amplitude feature, and the phase domain imbalance feature. The module then performs a majority consensus determination on each fault determination sub-result to obtain a determination result on whether the distribution network has a high impedance grounding fault.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a method and system for identifying multi-mode current faults in power distribution networks. First, the identification process is initiated using a zero-sequence voltage signal. High-frequency transient pulses are extracted from the zero-sequence current signal, and a pulse time interval sequence is constructed to quantify the periodicity of the pulse intervals. Simultaneously, the amplitude characteristics of the zero-sequence voltage and the phase-domain imbalance characteristics of the three-phase currents are extracted to comprehensively characterize the electrical characteristics of the fault from both transient and steady-state dimensions. Based on this, dynamic integration coefficients are calculated according to the periodicity of the pulse intervals and the amplitude characteristics of the zero-sequence voltage. These dynamic integration coefficients can assess the clarity of the current fault event's characteristic information in real time, thereby adaptively calling a corresponding number of fault identification models from a pre-constructed fault identification model array for parallel judgment. By performing majority consensus judgment on the fault judgment sub-results output by multiple models, the generalization advantage of multi-model integration is fully utilized, effectively overcoming the limitations of fixed decision boundaries of single models, and achieving accurate and reliable identification of high-impedance grounding faults under complex conditions such as strong background noise and random arc disturbances. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a method for identifying multi-mode current faults in a power distribution network, provided in an embodiment of the present invention; Figure 2This is an overall logic diagram of a method for identifying multi-mode current faults in a power distribution network, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a multi-mode current fault identification system for a power distribution network provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: Pulse period feature extraction module 11, steady-state electrical feature extraction module 12, dynamic integration coefficient calculation module 13, and fault judgment result output module 14. Detailed Implementation

[0010] This invention provides a method and system for identifying multi-mode current faults in power distribution networks, which addresses the technical problems of traditional overcurrent protection being difficult to detect under high-impedance grounding faults and the ease with which single fault identification models can misjudge or miss faults.

[0011] Example 1, as Figure 1 , Figure 2 As shown, the present invention provides a method for identifying multi-mode current faults in a power distribution network, the method comprising: S100: Acquire the zero-sequence voltage signal and zero-sequence current signal of the distribution network. When the zero-sequence voltage signal exceeds the preset start-up threshold, extract high-frequency transient pulses from the zero-sequence current signal, construct a pulse time interval sequence, and extract the periodic characteristics of the pulse interval.

[0012] When a high-impedance grounding fault occurs in a distribution network, the transition resistance between the fault point and the ground is large, and the amplitude of the power frequency component of the fault current is significantly suppressed, making traditional detection methods based on current amplitude thresholds ineffective. However, the intermittent arc during a high-impedance grounding fault generates high-frequency transient pulses, and the timing of these pulses exhibits a regularity related to the arc extinction cycle. To extract high-frequency transient pulses from strong background noise and the power frequency component and quantify their periodicity, time-frequency transformation and pulse peak detection of the zero-sequence current signal are required to construct a pulse time interval sequence and calculate the periodic characteristics of the pulse interval.

[0013] Step S100 in the method provided in this embodiment of the invention includes: The acquisition of zero-sequence voltage and zero-sequence current signals from the distribution network includes: Zero-sequence voltage signals are collected at a preset sampling frequency by using zero-sequence voltage transformers deployed on the distribution network bus. Zero-sequence current transformers deployed on each feeder of the distribution network are used to collect the zero-sequence current signal of each feeder at a preset sampling frequency.

[0014] First, zero-sequence voltage signals are acquired at a preset sampling frequency using zero-sequence voltage transformers deployed on the distribution network bus. A zero-sequence voltage transformer is an electrical device used to measure the three-phase voltage vector and the corresponding zero-sequence component voltage in a distribution network; its secondary-side output voltage signal is proportional to the primary-side zero-sequence voltage. The preset sampling frequency is a pre-defined number of signal data points collected per second, and its value is no less than twice the highest frequency component of the acquired signal.

[0015] Secondly, zero-sequence current transformers deployed on each feeder of the distribution network are used to collect the zero-sequence current signal of each feeder at a preset sampling frequency. A zero-sequence current transformer is an electrical device used to measure the three-phase current vector and the corresponding zero-sequence component current of the distribution network. Its secondary side output current signal is proportional to the primary side zero-sequence current.

[0016] Specifically, a zero-sequence voltage transformer is installed at the distribution network bus, and its secondary output is connected to the first analog input channel of the data acquisition device. A zero-sequence current transformer is installed at the outgoing line of each feeder in the distribution network, and the secondary output of each zero-sequence current transformer is connected to the corresponding analog input channel of the data acquisition device. The sampling frequency of the data acquisition device is set to a preset frequency, and the device is started to synchronously acquire the voltage and current signals of each channel, obtaining discrete sampling sequences of the zero-sequence voltage signal and the zero-sequence current signal with time as the independent variable.

[0017] For example, the turns ratio of the zero-sequence voltage transformer installed at the distribution network bus. That is, a primary line voltage of 10kV corresponds to a secondary line voltage of 100V, and the secondary output of the zero-sequence voltage transformer is connected to a data acquisition device. A zero-sequence current transformer with a transformation ratio of 100A / 5A is installed at the outgoing line of a feeder in the distribution network, and its secondary output is connected to a data acquisition device. The preset sampling frequency is set to 10kHz, meaning 10,000 data points are collected per second. After the data acquisition device is started, the secondary output signal of the zero-sequence voltage transformer is converted into a digital quantity through the acquisition channel to obtain the zero-sequence voltage signal u0(t); the secondary output signal of the zero-sequence current transformer is converted into a digital quantity through the acquisition channel to obtain the zero-sequence current signal i0(t) of the feeder.

[0018] Specifically, when the zero-sequence voltage signal exceeds a preset start-up threshold, a high-frequency transient pulse is extracted from the zero-sequence current signal to construct a pulse time interval sequence, and the periodic characteristics of the pulse interval are extracted, including: Wavelet transform is performed on the zero-sequence current signal to obtain wavelet coefficients. Wavelet coefficients within a preset high-frequency characteristic band are extracted from the wavelet coefficients. The preset high-frequency characteristic band is the frequency range corresponding to the high-frequency pulses generated by the intermittent arc in the high-impedance grounding fault. The modulus maxima of wavelet coefficients are detected, and the modulus maxima with amplitude exceeding the preset pulse extraction threshold are taken as the pulse peak time. Arrange the peak times of each pulse in chronological order, calculate the time interval between adjacent peak times of pulses, and form a pulse time interval sequence. Calculate the coefficient of variation of the pulse time interval sequence, and use the reciprocal of the coefficient of variation as the periodicity characteristic of the pulse interval.

[0019] The preset start-up threshold is a pre-defined threshold value for the effective value of the zero-sequence voltage. This threshold is set based on the balance between the maximum zero-sequence unbalanced voltage caused by the asymmetry of three-phase parameters during normal operation of the distribution network and the minimum amplitude at which the zero-sequence voltage can reliably rise when there is high impedance grounding. When the zero-sequence voltage signal exceeds the preset start-up threshold, high-frequency transient pulses are extracted from the zero-sequence current signal to construct a pulse time interval sequence, and the periodic characteristics of the pulse intervals are extracted.

[0020] First, wavelet transform is performed on the zero-sequence current signal to obtain wavelet coefficients. Wavelet coefficients within a preset high-frequency characteristic band are then extracted from the wavelet coefficients. The preset high-frequency characteristic band is the frequency range corresponding to the high-frequency pulses generated by intermittent arcs in high-impedance grounding faults.

[0021] Wavelet transform is a mathematical tool for time-frequency analysis. It decomposes a one-dimensional time series signal into wavelet coefficients at different scales and time positions by performing an inner product operation between the signal and wavelet basis functions that have undergone scale scaling and time shifting. The wavelet coefficients, obtained after wavelet transform, reflect the similarity between the signal and the wavelet basis functions at that scale and time position. The preset high-frequency characteristic band is the frequency range corresponding to the high-frequency pulses generated by intermittent arcing in high-impedance grounding faults. This frequency range is predetermined based on the intermittent arc's extinguishing transition time and the actual noise spectrum distribution of the distribution network.

[0022] Specifically, when the effective value of the zero-sequence voltage signal at power frequency exceeds a preset trigger threshold, the processing flow for the corresponding feeder zero-sequence current signal is activated. A suitable wavelet basis function for detecting transient impact signals is selected, and multi-scale wavelet decomposition is performed on the discrete sampling sequence of the zero-sequence current signal to obtain wavelet coefficient matrices at multiple scales. Each row corresponds to a decomposition scale, and each column corresponds to a time sampling point. Based on the upper and lower boundary frequencies of a preset high-frequency characteristic band, the wavelet decomposition scale layer corresponding to that band is determined. From all wavelet coefficients, wavelet coefficients at all time positions within that scale layer are extracted, while wavelet coefficients from other scale layers are discarded.

[0023] For example, suppose the effective value of the zero-sequence voltage signal u0(t) at the power frequency exceeds a preset start threshold of 100V, and the processing flow for the zero-sequence current signal i0(t) is initiated. The db4 wavelet from the Daubechies wavelet basis functions is selected as the wavelet basis, and multi-scale wavelet decomposition is performed on the discrete sampling sequence of i0(t). Assuming the preset high-frequency characteristic band has a frequency range of 4kHz to 10kHz, the first and second scales of the wavelet decomposition corresponding to 4kHz to 10kHz are determined based on the preset sampling frequency of 10kHz and the frequency range corresponding to each layer of the wavelet decomposition. Wavelet coefficients at all time positions on the first and second scales are extracted from the wavelet coefficient matrix to form the high-frequency characteristic band wavelet coefficient sequence W(n), where n is the time sampling point number.

[0024] Secondly, the modulus maxima of the wavelet coefficients are detected, and the modulus maxima whose amplitude exceeds a preset pulse extraction threshold are taken as the pulse peak time. A modulus maxima refers to a point in the wavelet coefficient sequence whose absolute value is greater than the absolute value of the wavelet coefficient at its adjacent time position; this point corresponds to a local abrupt change in the signal at that scale. The preset pulse extraction threshold is a pre-set amplitude threshold value used to determine the pulse peak time.

[0025] The preset pulse extraction threshold is based on the statistical distribution of the amplitude of the modulus maxima of the background noise in the high-frequency characteristic band of the zero-sequence current signal. It is set to a value higher than the amplitude of the maximum modulus maxima of the background noise, so as to retain the real high-frequency transient pulses generated by the intermittent arc as much as possible while effectively filtering out noise interference.

[0026] Specifically, the absolute value of each point in the wavelet coefficient sequence W(n) of the high-frequency characteristic band is compared with the absolute values ​​of the points before and after that point at each time position. If the absolute value of W(n0) at a certain time position n0 is greater than the absolute value of W(n0-1) and greater than the absolute value of W(n0+1), then time position n0 is determined to be a modulus maxima. All modulus maxima points and their corresponding time positions and absolute amplitude values ​​are recorded. The absolute amplitude values ​​of each modulus maxima point are compared with a preset pulse extraction threshold, and modulus maxima points whose absolute amplitude values ​​exceed the preset pulse extraction threshold are retained. The time position corresponding to that point is taken as the pulse peak time.

[0027] For example, the wavelet coefficient sequence W(n) of the high-frequency characteristic band has 10,000 time sampling points. Comparing each time sampling point from the 2nd to the 9999th time sampling point, let the absolute value of W(100) at the 100th time sampling point be 0.35, the absolute value of W(99) be 0.12, and the absolute value of W(101) be 0.08. Since 0.35 is greater than both 0.12 and 0.08, the time position n=100 is determined to be a modulus maxima. Assuming a preset pulse extraction threshold of 0.2, the absolute value of the amplitude at the 100th time sampling point, 0.35, is greater than 0.2, so this time position is retained as the pulse peak moment. After completing the above comparison for all time sampling points, a total of M pulse peak moments satisfying the conditions are detected, denoted as t1, t2, ..., t... M .

[0028] Next, the peak times of each pulse are arranged in chronological order, and the time interval between adjacent peak times is calculated to form a pulse time interval sequence. The pulse time interval sequence is a sequence composed of the differences between the peak times of each adjacent pulse, where each element in the sequence is the difference between a peak time of a pulse and the peak time of the preceding pulse.

[0029] Specifically, the peak times of the M pulses are arranged in chronological order to obtain an ordered sequence t1. <t2<......<t M Starting from the peak time of the second pulse, the difference between the peak time of each pulse and the peak time of the previous pulse is calculated sequentially to obtain the time interval between adjacent pulses. , ,......, The obtained M-1 time interval values ​​are arranged in the order of their corresponding pulse occurrences to form a pulse time interval sequence. , ,......, }

[0030] For example, suppose a total of 5 pulse peak times are detected, arranged in chronological order as follows: , , , , Calculate the time interval between adjacent pulses: , , , The pulse time interval sequence is {0.025s, 0.020s, 0.018s, 0.022s}.

[0031] Finally, the coefficient of variation (COP) of the pulse time interval sequence is calculated, and its reciprocal is used as the periodicity characteristic of the pulse intervals. The COP is the ratio of the standard deviation to the mean, used to measure the dispersion of a set of data. A larger COP indicates more dispersed data and weaker regularity, while a smaller COP indicates more concentrated data and stronger regularity. The periodicity characteristic of the pulse intervals is the reciprocal of the COP; a larger COP value indicates stronger periodicity of the pulse time interval sequence.

[0032] Specifically, the average value of the pulse time interval sequence is calculated by summing all time interval values ​​and dividing by the total number of time intervals. The standard deviation of the pulse time interval sequence is calculated using the standard deviation formula. Dividing the standard deviation by the average value yields the coefficient of variation. Dividing the value 1 by this coefficient of variation gives the result as the periodicity characteristic of the pulse intervals.

[0033] For example, the pulse time interval sequence is {0.025s, 0.020s, 0.018s, 0.022s}, with a total of four values. Average value. Standard deviation Coefficient of variation Periodicity of pulse intervals .

[0034] In this embodiment of the invention, the zero-sequence current signal is transformed from the time domain to the time-frequency domain using wavelet transform, and wavelet coefficients are extracted within a preset high-frequency characteristic band, effectively filtering out power frequency components and low-frequency noise interference. By detecting modulus maxima and setting a pulse extraction threshold, the time position of high-frequency transient pulses generated by intermittent electric arcs is accurately captured. By constructing a pulse time interval sequence and calculating the reciprocal of the coefficient of variation, the regularity of pulse occurrence is quantified into a dimensionless feature value, namely the periodicity feature of pulse intervals, providing a quantitative input for subsequent fault identification that characterizes the intermittent periodicity of the electric arc.

[0035] S200: Extracts zero-sequence voltage amplitude characteristics from the zero-sequence voltage signal and extracts phase domain imbalance characteristics from the three-phase current signal of the distribution network.

[0036] After a high-impedance ground fault occurs, the zero-sequence voltage amplitude increases with the fault, and its effective value reflects the severity of the fault and the overall response of the distribution network to changes in ground parameters. Meanwhile, during normal operation, the three-phase currents of the distribution network are basically balanced, but a single-phase ground fault will cause a significant difference in the current amplitudes of the faulty and non-faulty phases. The degree of imbalance between the three-phase current amplitudes can serve as another electrical criterion for identifying whether a fault has occurred. Relying solely on the periodic characteristics of high-frequency transient pulses may lead to missed detections when the arc is unstable or the pulse is weak. Therefore, it is necessary to extract power frequency steady-state characteristics from the zero-sequence voltage and three-phase currents to provide supplementary information for subsequent multi-model integrated judgment.

[0037] Step S200 in the method provided in this embodiment of the invention includes: Calculate the effective value of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds the preset start-up threshold, and use it as the zero-sequence voltage amplitude characteristic; Obtain the three-phase current signals of the distribution network and calculate the effective values ​​of phase A current, phase B current, and phase C current respectively; The maximum value among the effective values ​​of phase A current, phase B current, and phase C current is divided by the minimum value to obtain the three-phase current amplitude ratio, which is then used as a characteristic of phase domain imbalance.

[0038] First, the effective voltage value of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds a preset threshold is calculated, serving as the zero-sequence voltage amplitude characteristic. The effective voltage value is the square root of the average of the squares of the instantaneous values ​​of the AC voltage signal, used to characterize the energy magnitude of the AC voltage. The period exceeding the preset threshold is defined as the time from when the power frequency effective value of the zero-sequence voltage signal first exceeds the preset threshold. The moment From the beginning, until the power frequency effective value of the zero-sequence voltage signal is less than or equal to The moment The continuous time period constituted by [the end] , The zero-sequence voltage amplitude characteristic is defined as the result of calculating the effective value of all sampling points of the zero-sequence voltage signal within this time period.

[0039] Specifically, according to the preset sampling frequency and power frequency Determine the number of sampling points N within one power frequency cycle. The zero-sequence voltage signal is continuously acquired at a preset sampling frequency. The effective value of the zero-sequence voltage within a sliding power frequency cycle ending at the current time is calculated in real time. This effective value is then compared with a preset start-up threshold. A comparison is performed. When the valid value is detected to be greater than [a certain value] for the first time... At that time, record the current time as And begin storing subsequent zero-sequence voltage sampling data; when the effective value is detected to be less than or equal to... At that time, record the current time as Stop storing. Capture. to The discrete sampling sequence of the zero-sequence voltage signal within a time period, assuming there are K sampling points within this time period, the sequence is denoted as... Let k = 1, 2, ..., K. Calculate the effective value of the zero-sequence voltage during this time period using the root mean square formula for effective values. The root mean square formula for effective values ​​is: ,in, Let K be the instantaneous voltage value at the k-th sampling point; K is the total number of sampling points in this time period. The calculated zero-sequence voltage amplitude characteristics.

[0040] For example, a preset sampling frequency power frequency Then the sampling point One. Set a preset startup threshold. The effective value of the zero-sequence voltage is calculated in real time using a sliding motion method. When the effective value is first detected to be greater than 100V at 0.01s, it is recorded. If an effective value less than or equal to 100V is detected at 0.06s, record it. Zero-sequence voltage sampling data was extracted during the period from 0.01s to 0.06s, which lasted 0.05s. There are sampling points, and the sequence is denoted as . to Calculated using the root mean square formula for effective value, assuming the result after substituting the instantaneous values ​​at each sampling point is 124.5V, then this value is the zero-sequence voltage amplitude characteristic.

[0041] Next, the three-phase current signals of the distribution network are acquired, and the effective values ​​of the A-phase current, B-phase current, and C-phase current are calculated respectively. The three-phase current signals are the current signals on the three phase lines of the distribution network (A-phase, B-phase, and C-phase), and are obtained from the secondary side output of the current transformers installed on the corresponding phase lines. The effective value of the current is the square root of the average of the squares of the instantaneous values ​​of the AC current signal, used to characterize the energy magnitude of the AC current.

[0042] Specifically, using the A-phase current transformer, B-phase current transformer, and C-phase current transformer deployed at each feeder of the distribution network, at the same preset sampling frequency as S100. Discrete sampling sequences of phase A, phase B, and phase C current signals are acquired simultaneously. These are respectively... to Discrete sampling data of the three-phase currents are extracted within a time period. Let the sampling sequence of phase A current within this time period be... The B-phase current sampling sequence is The C-phase current sampling sequence is Let k = 1, 2, ..., K, where K is the total number of sampling points within the time period. Calculate the effective current values ​​of phases A, B, and C respectively using the root mean square formula for effective values. The root mean square formula for effective values ​​is: ; ; ;in, , , These are the instantaneous current values ​​of phase A, phase B, and phase C at the k-th sampling point, respectively. , , These are the calculated effective values ​​of the currents in phases A, B, and C, respectively.

[0043] For example, in to During the time period, there are a total of K=500 sampling points. The current transformers for phase A, phase B, and phase C installed at this feeder in the distribution network all have a transformation ratio of 200A / 5A. The three-phase current signals are synchronously acquired during this time period at a sampling frequency of 10kHz, resulting in the phase A current sampling sequence. to B-phase current sampling sequence to C-phase current sampling sequence to The effective value of the phase A current is calculated to be 2.3A, the effective value of the phase B current is 1.8A, and the effective value of the phase C current is 4.5A, according to the root mean square formula for effective values.

[0044] Next, the maximum value among the effective values ​​of phase A, phase B, and phase C currents is divided by the minimum value to obtain the three-phase current amplitude ratio. This ratio is used as a characteristic of phase imbalance. The three-phase current amplitude ratio is the ratio of the maximum to the minimum effective value of the three-phase currents. This three-phase current amplitude ratio quantifies the degree of asymmetry in the amplitude distribution of the three-phase currents in the distribution network; a larger ratio indicates a greater imbalance in the three-phase currents.

[0045] Specifically, from , , Find the maximum value among the three values. and minimum value Three-phase current amplitude ratio The calculation formula is: The calculated three-phase current amplitude ratio As a characteristic of phase-domain imbalance. For example, the effective value of phase A current. RMS value of phase B current C-phase current RMS value The maximum value among the three values. minimum value Three-phase current amplitude ratio: The value of 2.5 is taken as the phase domain imbalance characteristic, which reflects the degree of imbalance in the three-phase current amplitude distribution of the feeder during the fault period.

[0046] In this embodiment of the invention, the overall amplitude level of the zero-sequence voltage during the fault period is extracted by calculating the effective value of the zero-sequence voltage within the time period exceeding the threshold. The amplitude characteristics of the zero-sequence voltage are closely related to the magnitude of the fault transition resistance and the distribution network-to-ground parameters, and can reflect the severity of the fault. By calculating the effective values ​​of the three-phase currents separately and taking the ratio of the maximum to the minimum, a phase domain imbalance feature is constructed. The phase domain imbalance feature has a clear response capability to the three-phase current amplitude asymmetry caused by a single-phase ground fault. The above two power frequency steady-state features and the high-frequency transient features extracted in S100 belong to different feature domains. They complement each other and together constitute multi-mode feature input, providing more complete input data for subsequent multi-model integrated judgment.

[0047] S300: Calculate the dynamic integration coefficient based on the periodic characteristics of the pulse interval and the amplitude characteristics of the zero-sequence voltage.

[0048] In this embodiment of the invention, the pulse interval periodicity feature extracted in S100 reflects the regularity of the intermittent extinguishing of the fault arc, and the zero-sequence voltage amplitude feature extracted in S200 reflects the overall amplitude level of the zero-sequence voltage during the fault. However, the stability and reliability of these two features differ in specific fault scenarios. When the pulse interval periodicity feature value is large, it indicates that the high-frequency transient pulses are regular, and the confidence level of the judgment based on this feature is high; conversely, when the pulse interval periodicity feature value is small, it indicates that the pulse regularity is weak, and the reliability of this feature decreases. Similarly, if the amplitude of the zero-sequence voltage signal fluctuates drastically during the fault period, the reliability of the judgment based on a single effective value decreases. Therefore, it is necessary to comprehensively consider the pulse interval periodicity feature and the stability of the zero-sequence voltage amplitude change to quantitatively calculate a dynamic integration coefficient that can characterize the clarity of the current fault event feature information, providing a basis for the adaptive adjustment of the number of subsequent model calls.

[0049] Step S300 in the method provided in this embodiment of the invention includes: Calculate the voltage fluctuation rate of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds the preset start-up threshold, and use the reciprocal of the voltage fluctuation rate as the voltage stability. The pulse interval periodicity and voltage stability are normalized to the 0 to 1 range to obtain the normalized periodicity and normalized stability. The dynamic integration coefficient is obtained by weighted summation of the normalized periodicity and the normalized stability. The dynamic integration coefficient ranges from 0 to 1. The larger the dynamic integration coefficient, the less clear the characteristics of the current fault event are, and the more fault identification models need to be called.

[0050] First, calculate the voltage fluctuation rate of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds a preset threshold, and use the reciprocal of the voltage fluctuation rate as the voltage stability. The voltage fluctuation rate is the percentage of the zero-sequence voltage signal during the period when it exceeds the preset threshold. , Voltage stability is a quantitative indicator of the degree of drastic change in instantaneous amplitude over time; a larger value indicates a more drastic change in voltage amplitude over time. Voltage stability is the reciprocal of voltage fluctuation rate; a larger value indicates a more stable zero-sequence voltage signal during that period.

[0051] Specifically, the time period exceeding the preset start-up threshold is selected from S200. , The zero-sequence voltage discrete sampling sequence within ] k = 1, 2, ..., K, where K is the total number of sampling points within the time period. First, calculate the average value of the zero-sequence voltage samples within this time period. ,average value The calculation method is to sum the instantaneous values ​​of all sampling points and then divide by the total number of sampling points K. Then, the standard deviation of the zero-sequence voltage sampling values ​​within that time period is calculated. Standard deviation The standard deviation is calculated by taking the square root of the sum of the squares of the differences between the instantaneous values ​​and the average value at each sampling point. Divide by average The resulting ratio is used as the voltage fluctuation rate. ,Right now Dividing the value 1 by the voltage fluctuation rate yields the voltage stability. ,Right now .

[0052] For example, the period exceeding the preset start-up threshold is from 0.01s to 0.06s, during which K=500 sampling points are collected, and the zero-sequence voltage sampling sequence is denoted as... to The average value of the zero-sequence voltage during this period. The standard deviation of the zero-sequence voltage during this period Then the voltage fluctuation rate Voltage stability .

[0053] Secondly, the pulse interval periodicity and voltage stability are normalized to the 0-1 interval, respectively, to obtain normalized periodicity and normalized stability. Normalization is a data processing method that maps raw data with different dimensions or numerical ranges to the same specified numerical interval through linear transformation. The normalized periodicity is the value in the 0-1 interval obtained after normalizing the pulse interval periodicity; the larger the value, the stronger the pulse regularity. The normalized stability is the value in the 0-1 interval obtained after normalizing the voltage stability; the larger the value, the more stable the zero-sequence voltage signal.

[0054] Specifically, the Min-Max normalization method is used to analyze the pulse interval periodicity characteristic P and voltage stability. Normalization was performed separately. The distribution range of the pulse interval periodicity characteristics from a large number of historical fault events was pre-statistically analyzed to determine its theoretical minimum value. and theoretical maximum value The normalized formula for the periodicity of the pulse interval is: ,in To normalize the periodicity characteristic, the distribution range of voltage stability in a large number of historical fault events is statistically analyzed in advance to determine its theoretical minimum value. and theoretical maximum value The normalized formula for voltage stability is: ,in This represents the normalized stability. After the above normalization process, and All are located in the range of 0 to 1.

[0055] For example, the pulse interval periodicity characteristic P = 8.22. The theoretical minimum value of the pulse interval periodicity characteristic is determined by pre-statistically analyzing a large number of historical fault events. Theoretical maximum value Then the normalized periodicity characteristic Voltage stability The theoretical minimum value of voltage stability is determined by pre-statistically analyzing a large number of historical fault events. Theoretical maximum value Then the normalized stability .

[0056] Next, the normalized periodicity and normalized stability are weighted and summed to obtain the dynamic integration coefficient. The dynamic integration coefficient ranges from 0 to 1; a larger coefficient indicates less clear characteristics of the current fault event, requiring more fault identification models to be invoked. The dynamic integration coefficient is a comprehensive index obtained by weighting and summing the normalized periodicity and normalized stability, with a value ranging from 0 to 1. A larger value indicates less clear characteristic information of the current fault event, requiring more fault identification models to be invoked.

[0057] Specifically, the weights of the normalized periodic features are set as follows: The weights for normalized stability are ,and The formula for calculating the dynamic integration coefficient D is as follows: .in, This value characterizes the degree of uncertainty in the regularity of pulses; the lower the periodicity of the pulse interval, the larger the value. This value characterizes the degree of uncertainty in voltage stability; the lower the voltage stability, the larger the value. Weighting and The dynamic integration coefficient D is pre-set based on the confidence contribution of the two features in fault identification. The value of D ranges from 0 to 1. The larger the value of D, the less clear the current feature information is, and more fault identification models need to be called to improve the reliability of the judgment.

[0058] For example, let normalized periodicity characteristics be... Normalized stability Pre-set the weights of the normalized periodic features. Weights of normalized stability Calculate the dynamic integration coefficient. A dynamic integration coefficient greater than 0.5 indicates that the clarity of the feature information of the current fault event is low, and subsequent steps require calling more fault identification models for comprehensive judgment.

[0059] In this embodiment of the invention, the voltage stability is obtained by calculating the voltage fluctuation rate of the zero-sequence voltage signal during the period exceeding the preset start-up threshold and taking its reciprocal, thus quantifying the stability of the zero-sequence voltage amplitude during the fault period. By normalizing, the periodicity of the pulse interval and the voltage stability are mapped to the interval between 0 and 1, respectively, eliminating the differences in the dimensions and numerical ranges of the two features and making them additive. By weighted summing of the normalized periodicity and the normalized stability, a dynamic integration coefficient is constructed. The dynamic integration coefficient comprehensively reflects the clarity of the regularity of the high-frequency transient pulse and the stability of the zero-sequence voltage amplitude, providing a quantitative basis for the subsequent adaptive selection of the number of fault identification models, avoiding the problems of computational redundancy when the features are clear and insufficient judgment basis when the features are ambiguous when the number of models is fixed.

[0060] S400: Based on the dynamic integration coefficient, the corresponding number of fault identification models are called from the pre-built fault identification model array. Each fault identification model outputs a fault judgment sub-result based on the pulse interval periodicity characteristics, zero-sequence voltage amplitude characteristics, and phase domain imbalance characteristics. The majority consensus judgment is performed on each fault judgment sub-result to obtain the judgment result of whether there is a high impedance grounding fault in the distribution network.

[0061] In this embodiment of the invention, the dynamic integration coefficient reflects the clarity of the current fault event's feature information. A larger dynamic integration coefficient indicates less clear features, lower confidence of a single model's judgment, and a need for more models to participate in the integrated decision-making process to reduce the risk of misjudgment. Conversely, a smaller dynamic integration coefficient indicates clearer features, where a small number of high-performance models can ensure accurate judgment. Furthermore, due to differences in initial training data partitioning and network parameter initialization, different fault identification models in the fault identification model array exhibit diverse decision boundaries. By prioritizing models based on accuracy and employing majority consensus, the random bias of individual models can be effectively suppressed, improving the robustness of the overall judgment. In addition, the fault identification model array needs to be pre-constructed using historical samples and possess incremental update capabilities to adapt to the continuously accumulating new fault samples during actual operation.

[0062] Step S400 in the method provided in this embodiment of the invention includes: Multiply the dynamic integration coefficient by the total number of models N in the pre-built fault identification model array, and round up the product to obtain the number of models K to be called in this identification. Here, N is an odd number greater than or equal to 5. When K is even, add 1 to K. When K is greater than N, set K to N to ensure that K is an odd number greater than or equal to 1 and less than or equal to N. Based on the current recognition accuracy of each fault recognition model in descending order, select the top K fault recognition models from the fault recognition model array, where the current recognition accuracy is the initial recognition accuracy or the recognition accuracy after the most recent incremental update. The pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature are concatenated into an input feature vector. The input feature vector is then input into the first K selected fault identification models. Each fault identification model outputs a fault determination sub-result, which indicates whether a high-impedance grounding fault exists or not. The number of high-impedance grounding faults in each fault determination sub-result is counted. When the number exceeds half of K, the result of determining that the distribution network has a high-impedance grounding fault is output. When the number does not exceed half of K, the result of determining that the distribution network does not have a high-impedance grounding fault is output.

[0063] The pre-construction process of the fault identification model array includes: The pulse interval periodicity, zero-sequence voltage amplitude, and phase domain imbalance characteristics of historical fault events and historical non-fault events are obtained. Each historical fault event is labeled with a fault label, and each historical non-fault event is labeled with a non-fault label. A model training sample set is constructed. Divide the model training sample set into an initial validation subset, and then divide the remaining sample data in the model training sample set into N mutually exclusive sample subsets. During each training session, N minus one sample subset is taken as the training subset and the remaining sample subset is taken as the validation subset. This process is repeated N times, with a different sample subset used as the validation subset each time, resulting in N fault identification models. These N fault identification models form a fault identification model array. The accuracy of N fault identification models in the fault identification model array is evaluated using an initial validation subset to obtain the initial identification accuracy of each fault identification model.

[0064] First, the pulse interval periodicity, zero-sequence voltage amplitude, and phase domain imbalance characteristics of historical fault events and historical non-fault events are obtained. Each historical fault event is labeled with a fault label, and each historical non-fault event is labeled with a non-fault label, thus constructing a model training sample set. Historical fault events are records of electrical events where a high-impedance grounding fault in the distribution network has been confirmed to have occurred. Historical non-fault events are records of normal operation or non-fault disturbance events where a high-impedance grounding fault has not been confirmed to have occurred. The fault label is a marker value identifying the sample category; a value of 1 is assigned to historical fault events, and a value of 0 is assigned to historical non-fault events. The model training sample set is a collection of feature data for all historical events and their corresponding labels.

[0065] Specifically, from the historical waveform database of the distribution network or field test records, events clearly identified as high-impedance grounding faults and non-fault events are selected. For each event, pulse interval periodicity features, zero-sequence voltage amplitude features, and phase domain imbalance features are extracted according to steps S100 to S200, and these three features are used as input features for the sample. Each sample is assigned a label value: 1 for fault events and 0 for non-fault events. All samples are aggregated to form the model training sample set.

[0066] For example, 1000 event samples are obtained from historical data, of which 500 are high-impedance grounding fault events and 500 are non-fault events. For each event, pulse interval periodicity features, zero-sequence voltage amplitude features, and phase domain imbalance features are extracted and labeled to form a training sample set containing 1000 samples.

[0067] Secondly, an initial validation subset is created from the model training sample set. The remaining sample data in the model training sample set, excluding the initial validation subset, is divided into N mutually exclusive sample subsets. The initial validation subset consists of a portion of samples randomly selected from the training sample set, used for subsequent evaluation of the initial recognition accuracy of each model; this subset does not participate in model training. The mutually exclusive sample subsets refer to dividing the remaining samples into N disjoint sample sets, where each sample appears in only one subset.

[0068] Specifically, 20% of the samples are randomly selected from the training sample set as the initial validation subset, and the remaining 80% are used as training samples. The training samples are then randomly shuffled and divided into N equal parts, each containing approximately the same number of samples that are mutually exclusive (no duplicate samples). This results in N mutually exclusive sample subsets, denoted as . , ,..., For example, 200 samples are randomly selected from 1000 samples as the initial validation subset, leaving 800 samples. These 800 samples are then randomly divided into 7 parts, i.e., N=7, with approximately 114 samples in each part, and they do not overlap, resulting in 7 mutually exclusive subsets.

[0069] Furthermore, during each training iteration, one sample subset (N minus one) is taken as the training subset, and the remaining sample subset is taken as the validation subset. This process is repeated N times, with each iteration using a different sample subset as the validation subset, resulting in N fault identification models. These N fault identification models then form a fault identification model array.

[0070] In each training iteration, N minus one sample subset is used as the training subset, and the remaining sample subset is used as the validation subset. This training yields a fault identification model, which includes: An initial fault identification model is constructed, which includes a pulse feature branch, an electrical feature branch, a feature fusion layer, and a classification output layer. Using the pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature in the training subset of the current round as input features, and the corresponding fault label or non-fault label as supervision label, the network parameters of the initial fault identification model are updated with the goal of minimizing the focus loss between the prediction result output by the initial fault identification model and the supervision label, until convergence is verified on the validation subset of the current round, thus obtaining a fault identification model.

[0071] First, an initial fault identification model is constructed, comprising a pulse feature branch, an electrical feature branch, a feature fusion layer, and a classification output layer. The initial fault identification model is a neural network structure containing two independent branches. The pulse feature branch receives the periodic features of the pulse interval, while the electrical feature branch receives the zero-sequence voltage amplitude features and phase imbalance features. The feature fusion layer concatenates or adds the feature vectors extracted from the two branches. The classification output layer is a fully connected layer with a softmax activation function, outputting binary classification probabilities.

[0072] Specifically, the pulse feature branch is set as a multilayer perceptron with an input dimension of 1, corresponding to the periodic features of the pulse interval. It contains several fully connected layers and a ReLU activation function, and outputs a feature vector of fixed dimension. The electrical feature branch is set as another multilayer perceptron with an input dimension of 2, corresponding to the zero-sequence voltage amplitude features and phase imbalance features. It contains several fully connected layers and a ReLU activation function, and outputs a feature vector of the same dimension as the pulse feature branch. A feature fusion layer is set up to add or concatenate the output vectors of the two branches element-wise to obtain the fused features. A classification output layer is set up to map the fused features to 2D through a fully connected layer, and then convert them into fault or non-fault probabilities through a softmax function. All network parameters are initialized to random small values.

[0073] For example, the impulse feature branch takes a 1-dimensional input, passes through two fully connected layers (4 neurons in the first layer and 4 neurons in the second layer) to output a 4-dimensional feature. The electrical feature branch takes a 2-dimensional input, also passes through two fully connected layers (4 neurons in the first layer and 4 neurons in the second layer) to output a 4-dimensional feature. The feature fusion layer adds the two 4-dimensional vectors element-wise to obtain a 4-dimensional fused feature. The classification output layer maps the 4-dimensional features to 2-dimensionality through fully connected layers, and the softmax function outputs the probability of fault or non-fault.

[0074] Secondly, the pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature in the training subset of the current round are used as input features, and the corresponding fault label or non-fault label is used as supervision label. The training objective is to minimize the focus loss between the prediction result output by the initial fault identification model and the supervision label. The network parameters of the initial fault identification model are updated until convergence is verified on the validation subset of the current round, thus obtaining a fault identification model.

[0075] The focus loss is a loss function used for classification tasks. It introduces a modulation factor to reduce the loss contribution of easily classified samples, making the model focus more on difficult samples. The supervision label is the true class (0 or 1) labeled in the training samples. The network parameters include the weights and biases of each layer.

[0076] Specifically, a training batch is first formed by randomly selecting several samples from the current training subset according to the batch size (Batches). Each sample contains three input features: pulse interval periodicity, zero-sequence voltage amplitude, and phase imbalance. These three features of each sample in the batch are then input into the initial fault identification model. After forward computation through the pulse feature branch, electrical feature branch, feature fusion layer, and classification output layer, the model outputs the predicted probability p of the high-impedance grounding fault category corresponding to the sample. The true label y of the sample is obtained, where y=1 indicates the presence of a high-impedance grounding fault, and y=0 indicates the absence of a high-impedance grounding fault. The FocalLoss function is used as the training objective, and its calculation formula is as follows: Where α is the class balance coefficient, ranging from 0 to 1, used to adjust the weight ratio of positive and negative samples; γ is the focusing parameter, a real number greater than or equal to 0, used to control the modulation factor. The intensity of ; p is the predicted probability output by the model; y is the true label; α and γ are pre-defined hyperparameters.

[0077] Secondly, the backpropagation algorithm is used to calculate the gradient of the focal loss function value with respect to each network parameter in the initial fault identification model. The Adam optimizer is then used to update each network parameter based on the calculated gradient, with the learning rate of the Adam optimizer preset to 0.001. The above-mentioned forward calculation, loss calculation, gradient calculation, and parameter update process is performed after each training batch. After a preset number of training batches, such as every 100 batches, the samples in the current round's validation subset are input into the current model, the average focal loss value on the validation subset is calculated, and this validation loss value is recorded. When the validation loss value recorded for 5 consecutive times is not lower than the previously recorded value, the model is considered to have converged, training is stopped, the current network parameters are saved, and a fault identification model is obtained. If the validation loss still does not meet the above convergence condition after the total number of training batches reaches 10,000, training is stopped, the current network parameters are saved, and a fault identification model is obtained.

[0078] For example, let N=7, the total number of training samples is 800, and each training iteration uses 6 subsets as the training subset and 1 subset as the validation subset. Let the training subset have 686 samples and the validation subset have 114 samples. Let the batch size be 32, meaning each training batch randomly selects 32 samples from the training subset. Let the Adam optimizer learning rate be 0.001, the focus loss hyperparameters be α=0.75, and γ=2. After every 100 batches of parameter updates, the average focus loss value is calculated on the validation subset. If the validation loss value is not lower than the previous recorded value for 5 consecutive iterations, convergence is considered, training stops, and the model is saved. If convergence is not achieved after 10,000 training iterations, the current model is forcibly stopped and saved. After training, a fault recognition model is obtained.

[0079] Furthermore, the above training process is repeated N times, with a different sample subset used as the validation subset each time, resulting in N fault identification models. These N fault identification models form a fault identification model array. For the i-th round (i=1,2,...,N), the subset is removed. The remaining N-1 subsets are combined to form the training subset for the current round. This serves as the validation subset for the current round. Using this training subset, a fault identification model is trained following the steps described above, and the validation subset is used to monitor convergence. This process is repeated for round i, and trained N times in total, resulting in N fault identification models. , ,..., Construct a fault identification model array using N models. For example, if N=7, the first round selects... to As a training subset Training as a validation subset Round 2 , to As a training subset Training as a validation subset And so on, until the 7th round. The seven models together constitute the fault identification model array.

[0080] Subsequently, the accuracy of N fault identification models in the fault identification model array is evaluated using the initial validation subset, yielding the initial identification accuracy of each model. The initial identification accuracy is the proportion of correctly predicted samples out of the total number of samples in the initial validation subset after training.

[0081] Specifically, the three features of each sample in the initial validation subset are input into each fault identification model in the fault identification model array to obtain the predicted category for each model. The prediction result of each model is compared with the true label of the validation subset, and the number of correctly predicted samples is counted. The initial recognition accuracy of each model is obtained by dividing the number of correct predictions by the total number of samples in the validation subset, and this is stored as the current recognition accuracy of each model. For example, if the initial validation subset contains 200 samples, ... Conduct the test, if If 186 samples were correctly predicted, then... The initial recognition accuracy was 186 / 200 = 93.0%. Similarly, the calculation... to The initial recognition accuracy.

[0082] Based on this, the dynamic integration coefficient is multiplied by the total number of models N in the pre-built fault identification model array, and the product result is rounded up to obtain the number of models K to be called in this identification. Here, N is an odd number greater than or equal to 5. When K is an even number, K is incremented by 1, and when K is greater than N, K is set to N to ensure that K is an odd number greater than or equal to 1 and less than or equal to N.

[0083] The fault identification model array is a set of N fault identification models, where N is an odd number greater than or equal to 5. K is the number of fault identification models actually called in this identification, which must be an odd number greater than or equal to 1 and less than or equal to N to ensure that the majority consensus decision can uniquely determine the decision result.

[0084] Specifically, the dynamic integration coefficient D calculated in S300 is multiplied by the total number of fault identification model arrays N, resulting in a product D×N. This product is rounded up to the nearest integer, denoted as K'. The parity of K' is then determined: if K' is even, it is incremented by 1 to obtain K; if K' is odd, K=K'. Further checks are made to see if K is greater than N. If K is greater than N, K is reset to N. After these steps, K is an odd number greater than or equal to 1 and less than or equal to N. For example, if the dynamic integration coefficient D=0.6348 and the total number of fault identification model arrays N=7, D×N=0.6348×7=4.4436, rounding up gives K'=5. Since K' is odd, K=5. K=5 is not greater than N=7, therefore the final number of models called in this operation is determined to be K=5.

[0085] Secondly, based on the current recognition accuracy of each fault identification model, from highest to lowest, the top K fault identification models are selected from the fault identification model array. The current recognition accuracy is either the initial recognition accuracy or the recognition accuracy after the most recent incremental update. The current recognition accuracy is the proportion of correctly classified samples in the most recent evaluation of each fault identification model in the array, expressed as a percentage or decimal. The initial value is the initial recognition accuracy obtained after the model pre-construction using the initial validation subset evaluation. Subsequent incremental updates will update it to the recognition accuracy obtained after the most recent incremental update. Selecting the top K models means choosing the top K models after sorting them in descending order of accuracy.

[0086] Specifically, the current recognition accuracy values ​​stored in each of the N fault recognition models in the fault recognition model array are read. These N accuracy values ​​are sorted from largest to smallest to obtain a descending list of models. The top K models are then selected from this list as the model set for this recognition operation. If there are ties for the same accuracy, models are selected arbitrarily according to their model numbers or training order, but the final number of selected models must be K.

[0087] For example, N=7, K=5. The current recognition accuracies of the 7 models in the fault recognition model array are as follows: 95.2% 93.8% 96.1%, 91.5% 94.7% 92.3%, The percentage is 90.8%. The order after sorting in descending order is: (96.1%) (95.2%) (94.7%) (93.8%) (92.3%) (91.5%) (90.8%). Select the first 5, i.e., choose... , , , , This serves as the fault identification model used in this call.

[0088] Next, the pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature are concatenated into an input feature vector. The input feature vector is then input into the first K selected fault identification models. Each fault identification model outputs a fault determination sub-result, which indicates whether a high-impedance grounding fault exists or not.

[0089] The input feature vector consists of the pulse interval periodicity feature P obtained from S100 and the zero-sequence voltage amplitude feature obtained from S200. Phase domain imbalance characteristics A one-dimensional vector is formed by concatenating three numerical values ​​in a fixed order. The fault determination sub-result is a binary classification result independently output by each fault identification model based on the input feature vector, which indicates either the presence or absence of a high-impedance grounding fault.

[0090] Specifically, the three characteristic values ​​are arranged in a preset order: pulse interval periodicity characteristic, zero-sequence voltage amplitude characteristic, and phase domain imbalance characteristic, forming a three-dimensional feature vector. .Will Input the selected K fault identification models, and perform forward inference computation on each model, outputting a binary classification result. Record the fault determination sub-results output by each model, and label them as follows: , ,..., Each The value can be 1 or 0. 1 indicates the presence of a high-impedance grounding fault, and 0 indicates the absence of a high-impedance grounding fault.

[0091] For example, P=8.22, , Concatenate them into the input feature vector. Input this vector into the five selected models respectively. Output 1, Output 1, Output 0, Output 1, Output 1. This will yield 5 fault determination sub-results: , , , , .

[0092] Finally, the number of high-impedance grounding faults in each fault determination sub-result is counted. When the number exceeds half of K, the output result indicates that a high-impedance grounding fault exists in the distribution network; when the number does not exceed half of K, the output result indicates that a high-impedance grounding fault does not exist in the distribution network. The majority consensus decision involves a majority vote on the binary sub-results of the K fault identification models. Specifically, the number of outputs indicating a high-impedance grounding fault is counted. If this number is greater than K / 2, the final determination is that a fault exists; otherwise, it is determined that no fault exists.

[0093] Specifically, count the number of values ​​of 1 among the K sub-results, denoted as Calculate half of K, which is K / 2. If The final determination is that there is a high-impedance grounding fault in the distribution network; if Therefore, the final determination is that there is no high-impedance grounding fault in the distribution network. Since K is an odd number and K / 2 is a decimal, there are no equal cases, so there is no need to process the tie. For example, if K=5, there are 4 out of the 5 sub-results that take the value 1, i.e. . Since 4 > 2.5, it was ultimately determined that there was a high impedance grounding fault in the distribution network.

[0094] Specifically, the top K fault identification models are selected from the fault identification model array in descending order of their current identification accuracy, including: After each fault identification is completed, when the final judgment result is subsequently confirmed, the pulse interval periodicity characteristics, zero-sequence voltage amplitude characteristics, and phase domain imbalance characteristics of this fault event are used as incremental samples and added to the incremental sample set. When the cumulative number of updates to the incremental sample set reaches the preset update threshold, a subset of incremental samples is randomly sampled from the incremental sample set. The incremental sample subset is used to evaluate the accuracy of all fault identification models in the fault identification model array to obtain the identification accuracy of each fault identification model. The N fault identification models are sorted in descending order according to their recognition accuracy. The M fault identification models at the bottom of the sort are taken and retrained using an incremental sample set to obtain the M retrained fault identification models. Here, M is one-third of N rounded down. Replace the last M fault identification models in the original fault identification model array with the M retrained fault identification models to obtain the incrementally updated fault identification model array. Re-verify the accuracy of N fault identification models in the incrementally updated fault identification model array, and use the verified recognition accuracy as the recognition accuracy of the N fault identification models after the most recent incremental update.

[0095] First, after each fault identification is completed, when the final judgment result is subsequently confirmed, the pulse interval periodicity characteristic, zero-sequence voltage amplitude characteristic, and phase domain imbalance characteristic of this fault event are added to the incremental sample set as incremental samples. Subsequent confirmation refers to verifying the correctness of the judgment result through on-site inspection, maintenance, or other independent verification methods. The incremental sample is a confirmed complete event sample, containing the three features and their corresponding true labels. The incremental sample set is the cumulative set of all incremental samples.

[0096] Specifically, after each final judgment result is output, the result is awaited for verification by distribution network maintenance personnel or an automatic confirmation mechanism. If the confirmation result is correct, the three characteristic data and the true label of the event are added as an incremental sample to a pre-set incremental sample set buffer. For example, if the current judgment indicates a fault, and the maintenance personnel confirm it as a real fault on-site, then the P=8.22 of this event will be set. , And label 1 is added to the incremental sample set.

[0097] Secondly, when the cumulative number of updates to the incremental sample set reaches a preset update threshold, a subset of incremental samples is randomly sampled from the incremental sample set. This subset is then used to evaluate the accuracy of all fault identification models in the fault identification model array, yielding the identification accuracy of each model. The preset update threshold is the minimum number of new samples required to trigger an incremental update, for example, 50. The preset update threshold is set based on the following principle: achieving a balance between accumulating a sufficient number of incremental samples to ensure the statistical validity of model parameter updates and responding promptly to changes in distribution network conditions to ensure model timeliness. The incremental sample subset is a portion of samples randomly selected from the cumulative incremental sample set and is used to evaluate the current performance of the model.

[0098] Specifically, after each incremental sample is added, the total number of incremental samples is counted. When the total number reaches a preset update threshold, the update process is triggered. 80% of the incremental sample set is randomly selected as an evaluation subset, and the remaining 20% ​​is reserved for subsequent training. Each sample in the evaluation subset is input into each model in the fault identification model array, and the prediction accuracy of each model is calculated as the current identification accuracy of each model. For example, if the incremental sample set accumulates to 50 samples, 40 are randomly selected as the evaluation subset. The accuracy of each of the seven models on these 40 samples is calculated to obtain the current identification accuracy of each model.

[0099] Furthermore, the N fault identification models are sorted in descending order of their recognition accuracy. The bottom M fault identification models are then retrained using an incremental sample set, resulting in M ​​retrained fault identification models, where M is the floor(N / 3) rounded down from one-third of N. The bottom M models represent the M models with the lowest accuracy in the current evaluation. Retraining refers to updating the model parameters using the incremental sample set as training data.

[0100] Specifically, the N recognition accuracies are sorted from highest to lowest, and the last M models in the sort are assigned numbers. These last M models are then selected, and their current parameters are used as initial parameters. Using the entire incremental sample set as the training set, training continues for several rounds with the objective of minimizing the focus loss, until the loss converges on the remaining validation portion. After training, the model parameters are updated, and the recognition accuracy is improved. For example, if N=7, M=floor(7 / 3)=2, and the last two numbers in the recognition accuracy ranking are... and Use all 50 incremental samples as the training set, and 10 of them as the validation set, to retrain. and .

[0101] Finally, the M fault identification models at the bottom of the original fault identification model array are replaced with the M retrained fault identification models to obtain the incrementally updated fault identification model array. The accuracy of the N fault identification models in the incrementally updated array is then re-validated, and the resulting accuracy is used as the most recent incrementally updated accuracy of the N fault identification models. Replacement refers to overwriting the original model parameters with the newly trained model parameters, keeping the model numbers unchanged or updating them uniformly. Re-validating the accuracy involves re-evaluating all N models using a separate validation set.

[0102] Specifically, the retrained M models are used to replace the corresponding positions in the fault identification model array. A set of validation data not used in this retraining process is selected, such as the 20% reserved in the incremental samples. All N models are tested on this validation data, and the accuracy of each model is calculated. The updated accuracy is stored as the current recognition accuracy of each model for sorting during subsequent calls. The incremental sample set is cleared or used samples are removed, awaiting the next round of accumulation. For example, the retrained models are used... and Replace the original and The updated array is obtained. Using the reserved 10 incremental samples as a validation set, the accuracy of the 7 models is recalculated, and the stored current recognition accuracy is updated. The incremental sample set is cleared to prepare for the next round of incremental processing.

[0103] In this embodiment of the invention, the number of models to be called is adaptively determined by dynamic integration coefficients. This reduces computational redundancy when features are clear and increases model diversity to improve reliability when features are ambiguous. By prioritizing models based on accuracy, it is ensured that the models participating in the majority consensus determination are the best performing models at present. The majority consensus determination mechanism effectively suppresses misjudgments by a single model and improves the overall determination accuracy. The pre-built cross-validation fault identification model array ensures the diversity and generalization ability of the initial models. The incremental update mechanism enables the fault identification model array to continuously adapt to changes in the operating conditions of the distribution network and maintain long-term stable identification performance.

[0104] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a method and system for identifying multi-mode current faults in distribution networks. By extracting the periodic characteristics of high-frequency transient pulses from zero-sequence current, and simultaneously extracting zero-sequence voltage amplitude characteristics and phase imbalance characteristics from zero-sequence voltage and three-phase current, it effectively integrates transient high-frequency information with power frequency steady-state information, overcoming the deficiency of insufficient discrimination capability of single features under complex operating conditions. Based on this, dynamic integration coefficients are dynamically calculated according to the clarity of feature information, adaptively determining the number of models to be called. This reduces computational overhead when features are clear, and enhances the reliability of judgment through multi-model integration when features are ambiguous. Through a mechanism of selecting the best model based on recognition accuracy and a majority consensus judgment mechanism, the occasional deviation of single models is effectively suppressed, improving the recognition accuracy of high-impedance grounding faults. Furthermore, by combining a pre-constructed cross-validation fault identification model array with an incremental update strategy, the fault identification model array can continuously adapt to changes in distribution network operating conditions, achieving high-reliability fault judgment under long-term stable operation.

[0105] Example 2, as Figure 3 As shown, the present invention provides a multi-mode current fault identification system for power distribution networks, the system comprising: The pulse period feature extraction module 11 is used to acquire the zero-sequence voltage signal and zero-sequence current signal of the distribution network. When the zero-sequence voltage signal exceeds the preset start threshold, the module extracts high-frequency transient pulses from the zero-sequence current signal, constructs a pulse time interval sequence, and extracts the periodic features of the pulse interval. The steady-state electrical feature extraction module 12 is used to extract the zero-sequence voltage amplitude feature from the zero-sequence voltage signal and extract the phase domain imbalance feature from the three-phase current signal of the distribution network. The dynamic integration coefficient calculation module 13 is used to calculate the dynamic integration coefficient based on the periodicity characteristics of the pulse interval and the amplitude characteristics of the zero-sequence voltage. The fault determination result output module 14 is used to call a corresponding number of fault identification models from the pre-built fault identification model array according to the dynamic integration coefficient. Each called fault identification model outputs a fault determination sub-result based on the pulse interval periodicity characteristics, zero-sequence voltage amplitude characteristics, and phase domain imbalance characteristics. The majority consensus judgment is performed on each fault determination sub-result to obtain the determination result of whether there is a high impedance grounding fault in the distribution network.

[0106] In one embodiment, the pulse period feature extraction module 11 is further configured to: The acquisition of zero-sequence voltage and zero-sequence current signals from the distribution network includes: Zero-sequence voltage signals are collected at a preset sampling frequency by using zero-sequence voltage transformers deployed on the distribution network bus. Zero-sequence current transformers deployed on each feeder of the distribution network are used to collect the zero-sequence current signal of each feeder at a preset sampling frequency.

[0107] Specifically, when the zero-sequence voltage signal exceeds a preset start-up threshold, a high-frequency transient pulse is extracted from the zero-sequence current signal to construct a pulse time interval sequence, and the periodic characteristics of the pulse interval are extracted, including: Wavelet transform is performed on the zero-sequence current signal to obtain wavelet coefficients. Wavelet coefficients within a preset high-frequency characteristic band are extracted from the wavelet coefficients. The preset high-frequency characteristic band is the frequency range corresponding to the high-frequency pulses generated by the intermittent arc in the high-impedance grounding fault. The modulus maxima of wavelet coefficients are detected, and the modulus maxima with amplitudes exceeding a preset pulse extraction threshold are taken as the pulse peak moments. Arrange the peak times of each pulse in chronological order, calculate the time interval between adjacent peak times of pulses, and form a pulse time interval sequence. Calculate the coefficient of variation of the pulse time interval sequence, and use the reciprocal of the coefficient of variation as the periodicity characteristic of the pulse interval.

[0108] In one embodiment, the steady-state electrical feature extraction module 12 is further configured to: Calculate the effective value of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds the preset start-up threshold, and use it as the zero-sequence voltage amplitude characteristic; Obtain the three-phase current signals of the distribution network and calculate the effective values ​​of phase A current, phase B current, and phase C current respectively; The maximum value among the effective values ​​of phase A current, phase B current, and phase C current is divided by the minimum value to obtain the three-phase current amplitude ratio, which is then used as a characteristic of phase domain imbalance.

[0109] In one embodiment, the dynamic integration coefficient calculation module 13 is further configured to: Calculate the voltage fluctuation rate of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds the preset start-up threshold, and use the reciprocal of the voltage fluctuation rate as the voltage stability. The pulse interval periodicity and voltage stability are normalized to the 0 to 1 range to obtain the normalized periodicity and normalized stability. The dynamic integration coefficient is obtained by weighted summation of the normalized periodicity and the normalized stability. The dynamic integration coefficient ranges from 0 to 1. The larger the dynamic integration coefficient, the less clear the characteristics of the current fault event are, and the more fault identification models need to be called.

[0110] In one embodiment, the fault determination result output module 14 is further configured to: Multiply the dynamic integration coefficient by the total number of models N in the pre-built fault identification model array, and round up the product to obtain the number of models K to be called in this identification. Here, N is an odd number greater than or equal to 5. When K is even, add 1 to K. When K is greater than N, set K to N to ensure that K is an odd number greater than or equal to 1 and less than or equal to N. Based on the current recognition accuracy of each fault recognition model in descending order, select the top K fault recognition models from the fault recognition model array, where the current recognition accuracy is the initial recognition accuracy or the recognition accuracy after the most recent incremental update. The pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature are concatenated into an input feature vector. The input feature vector is then input into the first K selected fault identification models. Each fault identification model outputs a fault determination sub-result, which indicates whether a high-impedance grounding fault exists or not. The number of high-impedance grounding faults in each fault determination sub-result is counted. When the number exceeds half of K, the result of determining that the distribution network has a high-impedance grounding fault is output. When the number does not exceed half of K, the result of determining that the distribution network does not have a high-impedance grounding fault is output.

[0111] The pre-construction process of the fault identification model array includes: The pulse interval periodicity, zero-sequence voltage amplitude, and phase domain imbalance characteristics of historical fault events and historical non-fault events are obtained. Each historical fault event is labeled with a fault label, and each historical non-fault event is labeled with a non-fault label. A model training sample set is constructed. Divide the model training sample set into an initial validation subset, and then divide the remaining sample data in the model training sample set into N mutually exclusive sample subsets. During each training session, N minus one sample subset is taken as the training subset and the remaining sample subset is taken as the validation subset. This process is repeated N times, with a different sample subset used as the validation subset each time, resulting in N fault identification models. These N fault identification models form a fault identification model array. The accuracy of N fault identification models in the fault identification model array is evaluated using an initial validation subset to obtain the initial identification accuracy of each fault identification model.

[0112] In each training iteration, N minus one sample subset is used as the training subset, and the remaining sample subset is used as the validation subset. This training yields a fault identification model, which includes: An initial fault identification model is constructed, which includes a pulse feature branch, an electrical feature branch, a feature fusion layer, and a classification output layer. Using the pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature in the training subset of the current round as input features, and the corresponding fault label or non-fault label as supervision label, the network parameters of the initial fault identification model are updated with the goal of minimizing the focus loss between the prediction result output by the initial fault identification model and the supervision label, until convergence is verified on the validation subset of the current round, thus obtaining a fault identification model.

[0113] Specifically, the top K fault identification models are selected from the fault identification model array in descending order of their current identification accuracy, including: After each fault identification is completed, when the final judgment result is subsequently confirmed, the pulse interval periodicity characteristics, zero-sequence voltage amplitude characteristics, and phase domain imbalance characteristics of this fault event are used as incremental samples and added to the incremental sample set. When the cumulative number of updates to the incremental sample set reaches the preset update threshold, a subset of incremental samples is randomly sampled from the incremental sample set. The incremental sample subset is used to evaluate the accuracy of all fault identification models in the fault identification model array to obtain the identification accuracy of each fault identification model. The N fault identification models are sorted in descending order according to their recognition accuracy. The M fault identification models at the bottom of the sort are taken and retrained using an incremental sample set to obtain the M retrained fault identification models. Here, M is one-third of N rounded down. Replace the last M fault identification models in the original fault identification model array with the M retrained fault identification models to obtain the incrementally updated fault identification model array. Re-verify the accuracy of N fault identification models in the incrementally updated fault identification model array, and use the verified recognition accuracy as the recognition accuracy of the N fault identification models after the most recent incremental update.

[0114] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying multi-mode current faults in a power distribution network, characterized in that, The method includes: The zero-sequence voltage signal and zero-sequence current signal of the distribution network are acquired. When the zero-sequence voltage signal exceeds a preset start-up threshold, a high-frequency transient pulse is extracted from the zero-sequence current signal, a pulse time interval sequence is constructed, and the periodicity feature of the pulse interval is extracted. Extract the zero-sequence voltage amplitude feature from the zero-sequence voltage signal, and extract the phase domain imbalance feature from the three-phase current signal of the distribution network; Based on the periodicity of the pulse interval and the amplitude of the zero-sequence voltage, a dynamic integration coefficient is calculated, which is used to characterize the clarity of the current fault event feature information. Based on the dynamic integration coefficient, a corresponding number of fault identification models are called from the pre-constructed fault identification model array. The fault identification model array consists of N fault identification models, where N is an odd number greater than or equal to 5. Each called fault identification model outputs a fault determination sub-result based on the pulse interval periodicity characteristic, the zero-sequence voltage amplitude characteristic, and the phase domain imbalance characteristic. The majority consensus judgment is performed on each fault determination sub-result to obtain the determination result of whether the distribution network has a high impedance grounding fault.

2. The method for identifying multi-mode current faults in distribution networks according to claim 1, characterized in that, Acquiring the zero-sequence voltage and zero-sequence current signals of the distribution network, including: Zero-sequence voltage signals are collected at a preset sampling frequency by using zero-sequence voltage transformers deployed on the distribution network bus. Zero-sequence current transformers deployed on each feeder of the distribution network are used to collect the zero-sequence current signal of each feeder at the preset sampling frequency.

3. The method for identifying multi-mode current faults in a distribution network according to claim 2, characterized in that, When the zero-sequence voltage signal exceeds a preset start-up threshold, a high-frequency transient pulse is extracted from the zero-sequence current signal to construct a pulse time interval sequence, and the periodic characteristics of the pulse interval are extracted, including: Wavelet transform is performed on the zero-sequence current signal to obtain wavelet coefficients. Wavelet coefficients within a preset high-frequency characteristic band are extracted from the wavelet coefficients. The preset high-frequency characteristic band is the frequency range corresponding to the high-frequency pulses generated by intermittent electric arcs in high-impedance grounding faults. The modulus maxima of the wavelet coefficients are detected, and the modulus maxima with amplitudes exceeding a preset pulse extraction threshold are taken as the pulse peak moments. Arrange the peak times of each pulse in chronological order, calculate the time interval between adjacent peak times of pulses, and form a pulse time interval sequence. Calculate the coefficient of variation of the pulse time interval sequence, and use the reciprocal of the coefficient of variation as the periodicity characteristic of the pulse interval.

4. The method for identifying multi-mode current faults in distribution networks according to claim 1, characterized in that, Extracting zero-sequence voltage amplitude characteristics from the zero-sequence voltage signal and extracting phase domain imbalance characteristics from the three-phase current signal of the distribution network includes: The effective value of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds the preset start-up threshold is calculated as the zero-sequence voltage amplitude feature; Obtain the three-phase current signals of the distribution network and calculate the effective values ​​of phase A current, phase B current, and phase C current respectively; The maximum value among the effective values ​​of phase A current, phase B current, and phase C current is divided by the minimum value to obtain the three-phase current amplitude ratio, which is then used as a characteristic of phase domain imbalance.

5. The method for identifying multi-mode current faults in a distribution network according to claim 1, characterized in that, Based on the periodic characteristics of the pulse interval and the amplitude characteristics of the zero-sequence voltage, the dynamic integration coefficient is calculated, including: Calculate the voltage fluctuation rate of the zero-sequence voltage signal during the period when the zero-sequence voltage signal exceeds a preset start-up threshold, and use the reciprocal of the voltage fluctuation rate as the voltage stability. The pulse interval periodicity and the voltage stability are normalized to the 0 to 1 range to obtain the normalized periodicity and normalized stability. The normalized periodicity and the normalized stability are weighted and summed to obtain the dynamic integration coefficient, wherein the value of the dynamic integration coefficient ranges from 0 to 1. The larger the dynamic integration coefficient, the less clear the characteristics of the current fault event, and the more fault identification models need to be called.

6. The method for identifying multi-mode current faults in a distribution network according to claim 1, characterized in that, Based on the dynamic integration coefficient, a corresponding number of fault identification models are called from the pre-constructed fault identification model array. Each called fault identification model outputs a fault determination sub-result based on the pulse interval periodicity characteristic, the zero-sequence voltage amplitude characteristic, and the phase domain imbalance characteristic. A majority consensus determination is performed on each fault determination sub-result to obtain a determination result on whether a high-impedance grounding fault exists in the distribution network, including: Multiply the dynamic integration coefficient by the total number of models N in the pre-built fault identification model array, and round up the product to obtain the number of models K to be called in this identification. Here, N is an odd number greater than or equal to 5. When K is even, add 1 to K. When K is greater than N, set K to N to ensure that K is an odd number greater than or equal to 1 and less than or equal to N. The first K fault identification models are selected from the fault identification model array in descending order of their current identification accuracy, wherein the current identification accuracy is the initial identification accuracy or the identification accuracy after the most recent incremental update. The pulse interval periodicity feature, the zero-sequence voltage amplitude feature, and the phase domain imbalance feature are concatenated into an input feature vector. The input feature vector is then input into the first K selected fault identification models. Each fault identification model outputs a fault determination sub-result, which indicates whether a high-impedance grounding fault exists or not. The number of high-impedance grounding faults in each fault determination sub-result is counted. When the number exceeds half of K, the determination result that the distribution network has a high-impedance grounding fault is output. When the number does not exceed half of K, the determination result that the distribution network does not have a high-impedance grounding fault is output.

7. The method for identifying multi-mode current faults in a distribution network according to claim 6, characterized in that, The pre-construction process of the fault identification model array includes: The pulse interval periodicity, zero-sequence voltage amplitude, and phase domain imbalance characteristics of historical fault events and historical non-fault events are obtained. Each historical fault event is labeled with a fault label, and each historical non-fault event is labeled with a non-fault label. A model training sample set is constructed. Divide the model training sample set into an initial validation subset, and then divide the remaining sample data in the model training sample set, excluding the initial validation subset, into N mutually exclusive sample subsets. During each training session, N minus one sample subset is taken as the training subset and the remaining sample subset is taken as the validation subset to train a fault identification model. This process is repeated N times, with a different sample subset used as the validation subset each time, to obtain N fault identification models. The N fault identification models form a fault identification model array. The initial verification subset is used to evaluate the accuracy of N fault identification models in the fault identification model array to obtain the initial identification accuracy of each fault identification model.

8. The method for identifying multi-mode current faults in a distribution network according to claim 7, characterized in that, During each training iteration, N minus one sample subset is used as the training subset, and the remaining sample subset is used as the validation subset. This training yields a fault identification model, which includes: An initial fault identification model is constructed, which includes a pulse feature branch, an electrical feature branch, a feature fusion layer, and a classification output layer. Using the pulse interval periodicity feature, zero-sequence voltage amplitude feature, and phase domain imbalance feature in the training subset of the current round as input features, and the corresponding fault label or non-fault label as supervision label, the network parameters of the initial fault identification model are updated with the training objective of minimizing the focus loss between the prediction result output by the initial fault identification model and the supervision label, until convergence is verified on the validation subset of the current round, thus obtaining a fault identification model.

9. The method for identifying multi-mode current faults in a distribution network according to claim 6, characterized in that, Based on the current recognition accuracy of each fault recognition model, from highest to lowest, the top K fault recognition models are selected from the fault recognition model array, including: After each fault identification is completed, when the final judgment result is subsequently confirmed, the pulse interval periodicity characteristics, zero-sequence voltage amplitude characteristics, and phase domain imbalance characteristics of this fault event are used as incremental samples and added to the incremental sample set. When the cumulative number of updates in the incremental sample set reaches a preset update threshold, a subset of incremental samples is randomly sampled from the incremental sample set, and the accuracy of all fault identification models in the fault identification model array is evaluated using the subset of incremental samples to obtain the identification accuracy of each fault identification model. The N fault identification models are sorted in descending order according to the identification accuracy. The last M fault identification models are taken and retrained using the incremental sample set to obtain the retrained M fault identification models, where M is one-third of N rounded down. The M fault identification models at the bottom of the original sorting in the fault identification model array are replaced with the M retrained fault identification models to obtain the incrementally updated fault identification model array. The accuracy of the N fault identification models in the incrementally updated fault identification model array is re-verified, and the recognition accuracy obtained from the verification is taken as the recognition accuracy of the N fault identification models after the most recent incremental update.

10. A multi-mode current fault identification system for power distribution networks, characterized in that, The system is used to implement the multi-mode current fault identification method for a distribution network according to any one of claims 1 to 9, the system comprising: The pulse period feature extraction module is used to acquire the zero-sequence voltage signal and the zero-sequence current signal of the distribution network. When the zero-sequence voltage signal exceeds the preset start-up threshold, the module extracts high-frequency transient pulses from the zero-sequence current signal, constructs a pulse time interval sequence, and extracts the periodic features of the pulse interval. The steady-state electrical feature extraction module is used to extract the zero-sequence voltage amplitude feature from the zero-sequence voltage signal and extract the phase domain imbalance feature from the three-phase current signal of the distribution network. The dynamic integration coefficient calculation module is used to calculate the dynamic integration coefficient based on the periodicity characteristics of the pulse interval and the amplitude characteristics of the zero-sequence voltage. The dynamic integration coefficient is used to characterize the clarity of the current fault event feature information. The fault determination result output module is used to call a corresponding number of fault identification models from the pre-constructed fault identification model array according to the dynamic integration coefficient. The fault identification model array consists of N fault identification models, where N is an odd number greater than or equal to 5. Each called fault identification model outputs a fault determination sub-result according to the pulse interval periodicity feature, the zero-sequence voltage amplitude feature, and the phase domain imbalance feature. The majority consensus judgment is performed on each fault determination sub-result to obtain the determination result of whether there is a high impedance grounding fault in the distribution network.

Images