Fault detection method for ring main unit

Abstract

The invention relates to a power system fault detection technology, in particular to a fault detection method of a ring main unit, and aims to solve the problems that an existing detection technology based on a fundamental component or a high-frequency signal is insufficient in stability under a fluctuation condition, misjudgment is prone to being generated, the power supply direction can be dynamically adjusted due to power output change or load switching, and the reliability is poor. Under a low-load or no-load working condition, a traditional phase judgment method has a detection blind area, and a detection strategy cannot be adaptively corrected according to a system state. According to the method, the three-dimensional feature set is extracted through nonlinear and non-stationary signal processing, and fault detection is carried out in combination with the deep learning model, so that the problems of weak signals and waveform distortion caused by output fluctuation and harmonic interference of the distributed power supply are effectively solved, and the method has the advantage of improving the accuracy and reliability of high-resistance grounding fault detection.

Description

Technical Field

[0001] This application relates to power system fault detection technology, and more specifically, to a fault detection method for a ring main unit. Background Technology

[0002] With the large-scale application of distributed power sources such as photovoltaic and wind power in resonant grounding systems, the detection of high-resistance single-phase grounding faults in ring main units faces severe challenges. High-resistance grounding faults typically manifest as a state of high transition resistance, resulting in extremely weak zero-sequence current signals. At the same time, harmonic interference generated during the operation of distributed power sources further masks the fault characteristics, making it difficult for traditional signal processing methods to effectively extract key information.

[0003] Furthermore, the output of distributed power sources fluctuates frequently due to natural conditions, causing significant distortion in the zero-sequence current waveform. Existing detection technologies based on fundamental components or high-frequency signals lack stability under fluctuating conditions and are prone to misjudgment. The power supply direction also dynamically adjusts due to changes in power output or load switching. Under low-load or no-load conditions, traditional phase judgment methods have detection blind spots and cannot adaptively correct detection strategies according to system status.

[0004] These problems severely restrict the power supply reliability of the distribution network. In actual operation, high-resistance grounding faults account for a relatively high proportion of single-phase grounding faults, and fault misjudgment occurs frequently, often leading to the expansion of the fault range and the extension of power outage time, posing a major threat to the safe and stable operation of the power system. Summary of the Invention

[0005] (a) Technical problems to be solved The purpose of this application is to provide a fault detection method, electronic equipment, and computer-readable storage medium for ring main units, which has the advantages of improving the accuracy and reliability of high-resistance grounding fault detection.

[0006] (II) Technical Solution This application provides a fault detection method for ring main units, the technical solution of which is as follows: Acquire zero-sequence current signals from multiple monitoring points in the ring main unit; Nonlinear nonstationary signal processing is performed on the zero-sequence current signal to extract a three-dimensional feature set for characterizing high-resistance grounding faults. The three-dimensional feature set includes fundamental component features, instantaneous frequency features of high-frequency components, and energy features of characteristic frequency bands. The three-dimensional feature set is input into a trained deep learning model, which outputs the fault probability and candidate fault segments. Based on the fault probability and candidate fault sections, the fault detection results of the ring main unit are output.

[0007] Furthermore, this application also proposes to perform nonlinear, non-stationary signal processing on the zero-sequence current signal to extract a three-dimensional feature set, including: Time-frequency analysis was used to process zero-sequence current signals in order to extract fundamental component characteristics, instantaneous frequency characteristics of high-frequency components, and energy characteristics of characteristic frequency bands.

[0008] Furthermore, this application proposes that the time-frequency analysis method be the Hilbert-Huang transform (HHT) method.

[0009] Furthermore, this application also proposes that processing zero-sequence current signals using the Hilbert-Huang transform (HHT) method includes: The zero-sequence current signal is decomposed into multiple intrinsic mode function (IMF) components using ensemble empirical mode decomposition (EEMD). Effective IMF components are selected from multiple IMF components, and the effective IMF components are determined by adaptive noise threshold selection. Hilbert transform is applied to the effective IMF components to construct a three-dimensional feature set.

[0010] Furthermore, this application also proposes a deep learning model that is a convolutional neural network-long short-term memory network (CNN-LSTM) model with an enhanced attention mechanism.

[0011] Furthermore, this application also proposes to obtain output prediction information of distributed power sources; Based on power output prediction information, the power supply direction judgment strategy is adaptively adjusted. In this process, when the three-dimensional feature set is input into the deep learning model, fault detection is performed in conjunction with the adjusted power supply direction judgment strategy.

[0012] Furthermore, this application also proposes a verification step before outputting the fault detection results: Zero-sequence current conservation verification based on Kirchhoff's current law (KCL) is performed on candidate fault sections. And / or, The consistency of fault detection results with the output fluctuation trend of distributed power sources is verified.

[0013] Furthermore, this application also proposes that the zero-sequence current conservation verification of KCL includes: If the calculated zero-sequence current conservation error is greater than the preset error threshold, then the fault detection will be re-performed or the output verification failure flag will be triggered.

[0014] Furthermore, this application also proposes an electronic device comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the above-mentioned fault detection method for ring main units.

[0015] Furthermore, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned fault detection method for the ring main unit.

[0016] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts a three-dimensional feature set through nonlinear non-stationary signal processing and combines it with a deep learning model for fault detection. This effectively overcomes the problems of weak signal and waveform distortion caused by power fluctuations and harmonic interference in distributed power sources, and has the advantage of improving the accuracy and reliability of high-resistance grounding fault detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall structure of a fault detection method for ring main units. Detailed Implementation

[0019] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] Example 1 like Figure 1As shown in the figure, this application proposes a fault detection method for a ring main unit, including the following steps: S100: Acquire zero-sequence current signals from multiple monitoring points in the ring main unit; S200. Perform nonlinear non-stationary signal processing on the zero-sequence current signal to extract a three-dimensional feature set for characterizing high-resistance grounding faults. The three-dimensional feature set includes fundamental component features, high-frequency component instantaneous frequency features, and characteristic frequency band energy features. S300: Input the three-dimensional feature set into the trained deep learning model, and the deep learning model outputs the fault probability and candidate fault segments; S400: Based on the fault probability and candidate fault sections, output the fault detection results of the ring main unit.

[0022] In practical applications, acquiring zero-sequence current signals from multiple monitoring points in a ring main unit can be achieved by installing current sensors at key nodes of the ring main unit. For example, Hall effect sensors or Rogowski coil sensors can be used. The Hall effect sensor is set with a measurement range of 0~5A, an accuracy class of 0.1, and a sampling frequency of 2kHz to cover the high-frequency components that may be generated by high-resistance faults. The Rogowski coil is set with a bandwidth of 0~10kHz and an output sensitivity of 1V / A. Its main purpose is to collect zero-sequence current signals as the basis for subsequent analysis. Specifically, 3~5 monitoring points are set up and deployed at one loop-in end, one loop-out end, and 1~3 outgoing ends of the ring main unit to ensure coverage of all sections where faults may occur and to eliminate monitoring blind spots.

[0023] Furthermore, nonlinear nonstationary signal processing of zero-sequence current signals can be understood as a process of decomposing, filtering, and extracting features from the signal using a specific algorithm. Specifically, the Hilbert-Huang Transform (HHT) method is used. This method decomposes the zero-sequence current signal into multiple intrinsic mode function (IMF) components through ensemble empirical mode decomposition (EEMD). During decomposition, the amplitude of the Gaussian white noise superimposed on the original signal is 0.2 times the amplitude of the original signal, and the number of decompositions is set to 200 to effectively suppress the mode aliasing phenomenon in traditional empirical mode decomposition.

[0024] Furthermore, effective IMF components are selected from multiple IMF components. These effective IMF components are determined through an adaptive noise threshold, which is generated by calculating the signal-to-noise ratio (SNR) of each IMF component. The specific formula is as follows: ,in, The signal power in the IMF component, i.e., the component signal energy, can be obtained by integrating the square of the component signal in the time domain. , The time-domain signal of the IMF component. , For signal analysis time window; This represents the noise power in the IMF component, i.e., the component noise energy. , This represents the noise signal in the IMF component.

[0025] Components with an SNR > 15dB are retained as effective IMF components. If the SNR of all components is ≤ 15dB, the top three components in terms of energy percentage are retained. Finally, a Hilbert transform is performed on the effective IMF components to construct a three-dimensional feature set, thereby extracting the fundamental component features, the instantaneous frequency features of the high-frequency components, and the energy features of the characteristic frequency band. The fundamental component features are extracted from the first two low-frequency components in the effective IMF components by taking the mean instantaneous amplitude of the 50Hz fundamental wave. The instantaneous frequency features of the high-frequency components are extracted from the last one or two high-frequency components in the effective IMF components by taking the mean and standard deviation of the frequency band from 500Hz to 2kHz. The energy features of the characteristic frequency band are obtained by calculating the proportion of the energy in the 50Hz to 2kHz frequency band to the total energy of the zero-sequence current signal.

[0026] These features provide complementary information from the perspectives of low-frequency stability, high-frequency dynamic changes, and energy distribution, respectively, to address the problem of weak signals and harmonic interference under high-impedance grounding faults.

[0027] Specifically, inputting the three-dimensional feature set into the trained deep learning model can be achieved using various neural network architectures, such as a hybrid model combining convolutional neural networks (CNN) and gated recurrent units (GRU), or a feature enhancement model based on autoencoders. The main purpose is to use the nonlinear modeling capabilities of deep learning to intelligently analyze complex signals, thereby outputting fault probabilities and candidate fault sections.

[0028] The innovation of this application lies in solving the core challenge of high-resistance grounding fault detection in ring main units after distributed power supply integration by combining nonlinear non-stationary signal processing with a deep learning model. Specifically, the nonlinear non-stationary signal processing extracts multi-dimensional features, avoiding the limitations of traditional single-feature extraction methods in the face of signal distortion; the deep learning model can adapt to signal distortion caused by power output fluctuations in distributed power sources, reducing the possibility of misjudgment. The overall solution achieves effective detection of high-resistance grounding faults through systematic design, overcoming the shortcomings of existing technologies in scenarios with weak signals, harmonic interference, and power output fluctuations.

[0029] The working principle of this application embodiment is as follows: First, by acquiring the zero-sequence current signals from multiple monitoring points in the ring main unit, raw data support is provided for subsequent analysis, ensuring the basic reliability of fault detection. Further, nonlinear non-stationary signal processing is performed on the zero-sequence current signals to extract a three-dimensional feature set for characterizing high-resistance grounding faults. This three-dimensional feature set includes fundamental component characteristics, instantaneous frequency characteristics of high-frequency components, and energy characteristics of characteristic frequency bands.

[0030] Specifically, the fundamental frequency component characteristics are used to capture fundamental frequency information under fault conditions, thereby stably characterizing low-frequency characteristics; the instantaneous frequency characteristics of the high-frequency components dynamically reflect the instantaneous changes of harmonic interference, enhancing anti-interference capabilities; and the characteristic frequency band energy characteristics quantify the energy distribution of key frequency bands to identify fault-specific modes.

[0031] Thus, these three elements work together to form a robust feature set, overcoming the limitations of traditional methods in feature extraction under signal distortion. Next, the three-dimensional feature set is input into a trained deep learning model, which outputs the fault probability and candidate fault segments. As a preferred implementation, the deep learning model utilizes its nonlinear modeling capabilities to intelligently analyze the fault probability and locate fault segments based on the three-dimensional feature set, effectively adapting to signal distortion caused by distributed power source output fluctuations and avoiding misjudgments caused by simple threshold judgments.

[0032] Finally, based on the fault probability and candidate fault sections, the fault detection results of the ring main unit are output. The final detection decision is formed through the comprehensive model output, ensuring the reliability and pertinence of the results. Overall, signal acquisition provides the data foundation, nonlinear and non-stationary signal processing extracts key features, and deep learning models realize intelligent judgment. These three aspects are interconnected and jointly solve the detection challenges caused by weak signals, interference, and power fluctuations under high-resistance faults.

[0033] This application further proposes to perform nonlinear nonstationary signal processing on zero-sequence current signals to extract a three-dimensional feature set, including: using time-frequency analysis methods to process the zero-sequence current signal to extract fundamental component features, high-frequency component instantaneous frequency features, and characteristic frequency band energy features.

[0034] Specifically, time-frequency analysis is a tool that can simultaneously analyze the dynamic changes of signals in the time and frequency domains. It can be implemented using short-time Fourier transform, wavelet transform, or Hilbert-Huang transform, among others. The purpose of introducing time-frequency analysis is to overcome the limitations of traditional fixed-window analysis, thereby more accurately capturing the nonlinear and non-stationary characteristics of zero-sequence current signals.

[0035] The fundamental frequency component characteristic refers to the characteristics of the zero-sequence current signal related to the fundamental frequency component of the fault. It can be obtained by extracting the dominant frequency component or low-frequency component of the signal, with the aim of providing stable basic characteristics for subsequent fault detection. The high-frequency component instantaneous frequency characteristic refers to the frequency characteristics of the high-frequency part of the zero-sequence current signal as a function of time. It can be obtained by tracking the instantaneous frequency change curve of the signal, with the aim of reflecting the rapid electromagnetic transient process near the fault point.

[0036] The characteristic frequency band energy characteristic refers to the energy distribution characteristics of a zero-sequence current signal within a specific frequency band. It can be achieved by dividing the signal into frequency bands and calculating the energy proportion of each frequency band, with the aim of quantifying the key distribution patterns of fault energy.

[0037] Specifically, this scheme processes zero-sequence current signals using time-frequency analysis, which can dynamically capture the joint changes of the signal in the time and frequency domains, avoiding the limitations of traditional fixed-window analysis methods in high-resistance grounding fault scenarios.

[0038] Based on this, extracting fundamental component features can effectively resist signal distortion caused by power output fluctuations of distributed power sources and ensure the stability of low-frequency fundamental features; extracting instantaneous frequency features of high-frequency components can accurately distinguish the weak high-frequency components unique to high-impedance faults from random noise interference, thereby improving the reliability of high-frequency features; extracting energy features of characteristic frequency bands relies on the ability of time-frequency analysis to focus on specific frequency bands, avoiding the noise accumulation problem caused by full-band analysis, and significantly enhancing the sensitivity and robustness of features to fault conditions.

[0039] Overall, by introducing time-frequency analysis, this scheme enables the feature extraction process to closely match the actual dynamic characteristics of the zero-sequence current signal, thus solving the problem that general signal processing methods are susceptible to harmonic interference and power fluctuations in high-impedance fault detection, leading to feature distortion.

[0040] Furthermore, this solution, combined with techniques for acquiring zero-sequence current signals from multiple monitoring points in the ring main unit and subsequently inputting the three-dimensional feature set into a deep learning model, forms a complete technical chain from signal acquisition to feature extraction to fault detection. This combination not only improves the accuracy of feature extraction but also provides high-quality input data for the subsequent deep learning model's output of fault probability and candidate fault sections, thereby further enhancing the overall performance of fault detection.

[0041] This application further proposes a time-frequency analysis method, namely the Hilbert-Huang transform (HHT) method.

[0042] The Hilbert-Huang Transform (HHT) method is a time-frequency analysis technique specifically designed for processing nonlinear and non-stationary signals. It can be implemented using ensemble empirical mode decomposition (EEMD) combined with adaptive noise thresholding. In practical applications, this method generates intrinsic mode functions (IMF) components through an adaptive decomposition mechanism, avoiding the dependence of traditional time-frequency analysis methods on fixed basis functions or preset windows, thus enabling accurate extraction of fault characteristics from zero-sequence current signals.

[0043] Specifically, the above scheme solves the problem of feature extraction under high-impedance grounding faults by introducing an improved Hilbert-Huang transform (HHT) method, which addresses the weakness of the zero-sequence current signal and its interference from distributed source harmonics. First, relying on ensemble empirical mode decomposition (EEMD), the zero-sequence current signal is decomposed into multiple intrinsic mode functions (IMF) components. This process is entirely based on the signal's own local dynamic characteristics, requiring no preset basis functions or fixed windows, perfectly adapting to the nonlinear and non-stationary nature of high-impedance fault signals.

[0044] Subsequently, invalid components are eliminated through adaptive noise thresholding, retaining the effective IMF components carrying fundamental wave characteristics and high-frequency transient information. Finally, three-dimensional feature extraction is completed through Hilbert transform. This improved HHT method can control the fundamental wave component feature extraction error within 5% when signal distortion is caused by power fluctuations. The frequency drift of the instantaneous frequency characteristics of high-frequency components is reduced by 80% compared with traditional methods, and the energy identification signal-to-noise ratio of the energy characteristics in the characteristic frequency band is improved to over 25dB.

[0045] Building upon this, the aforementioned scheme works organically with the methods described above for acquiring zero-sequence current signals, extracting three-dimensional feature sets, and serving as input for subsequent deep learning models. By employing an improved HHT method, the overall accuracy of three-dimensional feature set extraction is significantly improved. Even in high-resistance scenarios with transition resistances of 1500-3000Ω and extreme conditions where distributed power supply output fluctuations exceed 30%, the feature extraction stability remains above 90%, providing high-quality input for subsequent fault detection and directly increasing the fault detection accuracy from 65% in existing technologies to over 92%.

[0046] This application further proposes a method for processing zero-sequence current signals using the Hilbert-Huang Transform (HHT) method, which includes: decomposing the zero-sequence current signal into multiple Intrinsic Mode Function (IMF) components using ensemble empirical mode decomposition (EEMD); selecting effective IMF components from the multiple IMF components, wherein the effective IMF components are determined by adaptive noise thresholding; and performing a Hilbert transform on the effective IMF components to construct a three-dimensional feature set.

[0047] Among them, ensemble empirical mode decomposition (EEMD) is an improved empirical mode decomposition method. It can be achieved by superimposing Gaussian white noise on the original signal and decomposing it multiple times and taking the average. The purpose is to suppress the mode aliasing phenomenon in traditional empirical mode decomposition.

[0048] Adaptive noise threshold refers to a selection criterion that is dynamically adjusted based on signal characteristics. It can be generated by calculating the energy proportion of each component and the statistical characteristics of noise, with the aim of accurately identifying and eliminating invalid components subject to interference. Hilbert transform is an instantaneous signal analysis tool that can extract instantaneous frequency and amplitude information through analytical operations on the signal. Its purpose is to accurately capture the instantaneous frequency fluctuation characteristics of high-frequency components and the energy distribution patterns of characteristic frequency bands.

[0049] Specifically, this scheme first decomposes the zero-sequence current signal into multiple independent Intrinsic Mode Function (IMF) components through Empirical Mode Decomposition (EEMD). This process utilizes the statistical properties of Gaussian white noise to break the coupling relationship between components with similar frequencies in the signal, thereby effectively separating the fundamental component from the high-frequency fault component. Next, effective IMF components are screened based on an adaptive noise threshold. This threshold is dynamically generated according to the energy proportion of each component and the noise statistical characteristics, ensuring that the screened components only contain high-impedance fault-related features.

[0050] Finally, a Hilbert transform is performed on the selected effective IMF components. Leveraging the instantaneous analytical capability of this transform, the amplitude stability characteristics of the fundamental component, the instantaneous frequency trajectory of the high-frequency components, and the energy distribution density of the characteristic frequency bands are extracted simultaneously, thereby constructing a robust three-dimensional feature set. The above steps are optimized to address the characteristics of weak zero-sequence current signals under high-impedance faults and their susceptibility to harmonic interference from distributed power sources, significantly improving the accuracy and reliability of feature extraction.

[0051] Through the above technical solution, even in extreme high resistance scenarios with a transition resistance of 3000Ω and a distributed power supply output fluctuation of 40%, the accuracy of three-dimensional features in representing faults remains above 95%. Compared with the solution of directly applying HHT, the fault misjudgment rate is reduced by 70%, providing high-quality input for subsequent deep learning models and directly supporting the high reliability of fault detection.

[0052] This application further proposes a deep learning model that enhances the attention mechanism of a convolutional neural network-long short-term memory network (CNN-LSTM).

[0053] Specifically, the attention mechanism refers to a technical means that can dynamically allocate feature weights. It can be implemented by designing weight allocation logic based on fault discrimination. The purpose is to focus on the weak features unique to high-impedance faults and suppress invalid information caused by harmonic interference.

[0054] Convolutional Neural Networks (CNNs) are neural network structures with local receptive fields and weight sharing characteristics. They can extract spatial features through a combination of multiple convolutional and pooling layers, making them particularly suitable for processing nonlinear and non-stationary signal features generated by time-frequency analysis. Long Short-Term Memory (LSTM) networks are recurrent neural network structures that include input gates, forget gates, and output gates. They capture time-series dependencies through gating mechanisms, effectively adapting to dynamic signal changes caused by fluctuations in distributed power supply output.

[0055] In detail, this technical solution addresses the problems of weak zero-sequence current signals, severe harmonic interference, and signal distortion caused by power output fluctuations in distributed power supply environments by constructing an attention mechanism and an overall architecture of CNN-LSTM. The attention mechanism dynamically allocates weights based on fault discrimination, assigning high weights to key features such as minute fluctuations in the instantaneous frequency of high-frequency components and abnormal peaks in characteristic frequency bands, while reducing the weights of features corresponding to fixed-frequency harmonics of the distributed power supply. This allows for the suppression of noise while concentrating computing power on effective fault information.

[0056] Convolutional neural networks (CNNs) utilize their local perception capabilities to accurately extract spatial correlation patterns from three-dimensional feature sets, such as the coupling relationship between the fundamental frequency and high-frequency features, and the distribution pattern of energy in feature frequency bands after time-frequency analysis. Long Short-Term Memory (LSTM) networks capture the long-term temporal dependence of zero-sequence current signals through gating mechanisms, adapting to the dynamic evolution of features caused by power output fluctuations in distributed power sources. These three components work together to form a complete feature processing chain: CNNs are responsible for spatial feature extraction, LSTMs for temporal dynamic capture, and attention mechanisms for focusing key information, ultimately achieving accurate identification of fault features under complex operating conditions.

[0057] Building upon this foundation, the technical solution is closely integrated with the aforementioned steps of acquiring zero-sequence current signals and extracting three-dimensional feature sets. The three-dimensional feature set obtained through nonlinear, non-stationary signal processing of the zero-sequence current signal fully leverages the advantages of the attention-enhanced CNN-LSTM model, enabling it to exhibit stronger adaptability and accuracy in handling weak fault signals and complex interference scenarios. This effectively solves the technical challenge of detecting high-resistance grounding faults in distributed power supply environments.

[0058] This application further proposes the following technical solutions: it also includes obtaining output prediction information of distributed power sources; adaptively adjusting the power supply direction judgment strategy based on the output prediction information; wherein, when the three-dimensional feature set is input into the deep learning model, fault detection is performed in combination with the adjusted power supply direction judgment strategy.

[0059] Specifically, power output prediction information refers to a data set generated by analyzing the power output trend of distributed power sources over a future period. It can be achieved by using a short-term rolling prediction model (such as a 5-minute time window) combined with historical power output data and weather prediction information (such as solar irradiance prediction for photovoltaic power or wind speed prediction for wind power). The prediction model uses an LSTM neural network. The input features include power output data from the past hour and solar irradiance / wind speed prediction data for the next hour. The output is the predicted value of power output fluctuation amplitude for the next 5 minutes. The purpose is to provide a forward-looking basis for power supply direction judgment strategy and avoid misjudgment caused by the lag in real-time data response.

[0060] The adaptive adjustment strategy for determining the power supply direction refers to the process of dynamically optimizing the rules for determining the power supply direction based on the output fluctuation characteristics of distributed power sources. This can be achieved by setting a dynamic adjustment mechanism for the phase judgment threshold or by introducing high-frequency characteristic band energy to assist in the judgment. When the predicted output fluctuation amplitude is greater than 20%, the phase judgment threshold is set to the basic threshold × (1 + output fluctuation amplitude × 0.5), where the basic threshold is set to 0.3A. When the predicted low load (output < 20% of rated output) or no load (output < 5% of rated output) scenario is entered, high-frequency characteristic band energy is used to assist in the judgment, forming a dual-criteria mode of fundamental phase and high-frequency energy. At this time, the proportion of high-frequency characteristic band energy ≥ 8% can be determined as a fault-related feature. The purpose is to solve the detection dead zone problem in low load or no load scenarios.

[0061] Fault detection combined with the adjusted power supply direction judgment strategy refers to embedding dynamically adjusted power supply direction judgment rules into the deep learning model, enabling the model to accurately align fault probability calculation and section location logic based on the real-time power supply direction. This can be achieved by embedding a prior weight calibration mechanism related to the power supply direction in the model feature processing layer. When the power supply is in the forward direction, the feature weight of the loop-in cabinet is set to 0.6 and the feature weight of the outgoing cabinet is set to 0.4. When the power supply is in the reverse direction, the feature weight of the loop-out cabinet is set to 0.6 and the feature weight of the outgoing cabinet is set to 0.4. The purpose is to improve the robustness of detection under complex working conditions.

[0062] Specifically, this solution addresses the core pain points of frequent power supply direction switching caused by output fluctuations in distributed power sources and the detection dead zone of traditional fixed strategies in low-load / no-load scenarios by constructing a closed-loop mechanism of "output prediction - strategy adaptation - model collaborative detection".

[0063] First, obtain short-term output forecast information of distributed power sources and combine it with real-time output data to form dual data support, predict the output fluctuation range in advance, and provide a forward-looking basis for strategy adjustment.

[0064] Based on the power output prediction information, the power supply direction judgment strategy is dynamically adjusted. When the predicted power output fluctuation is greater than 20%, the phase judgment threshold is dynamically adjusted to avoid misjudgment of direction caused by distorted current. When the prediction enters the low load or no load scenario, the high frequency characteristic frequency band energy is automatically enabled to assist in the judgment, forming a dual judgment mode of fundamental phase and high frequency energy. At the same time, the feature weights of the ring-in cabinet and ring-out cabinet are pre-calibrated according to the predicted power supply direction to provide prior guidance for model detection.

[0065] Finally, when inputting the 3D feature set into the deep learning model, the adjusted power supply direction judgment rule is embedded into the feature processing layer of the model. When the power supply is in the forward direction, the focus is on extracting the feature correlation between the loop-in cabinet and the outgoing cabinet. When the power supply is in the reverse direction, the focus is on the feature matching degree between the loop-out cabinet and the outgoing cabinet, so as to ensure that the fault probability calculation and section location are accurately aligned with the real-time power supply direction.

[0066] The above solution not only effectively addresses the detection challenges caused by fluctuations in the output of distributed power sources, but also significantly improves the detection accuracy and robustness under complex operating conditions such as high-resistance grounding faults.

[0067] This application further proposes a verification step before outputting the fault detection results: performing zero-sequence current conservation verification on the candidate fault section based on Kirchhoff's current law (KCL); and / or verifying the consistency between the fault detection results and the output fluctuation trend of the distributed power source.

[0068] Specifically, zero-sequence current conservation verification refers to analyzing the zero-sequence current distribution of candidate fault sections to verify whether it conforms to the basic principles of Kirchhoff's Current Law (KCL). In practical applications, this can be achieved by calculating the vector sum of the inflow and outflow zero-sequence currents in the candidate fault section and combining it with a preset error threshold. If the calculation error exceeds the threshold, the candidate section is determined to have signal distortion or location error. The purpose of this step is to physically eliminate erroneous candidate sections caused by harmonic interference, thereby improving the reliability of the detection results.

[0069] The consistency verification of output fluctuation trends of distributed generation can be understood as a soft verification mechanism. Its core is to compare the logical matching between fault characteristics and real-time output fluctuations of the distributed generation. Specifically, by acquiring output prediction information and real-time data from the distributed generation, the fluctuation level can be determined, such as slight fluctuation, moderate fluctuation, or severe fluctuation. This level is then correlated with the fundamental amplitude mutation and high-frequency instantaneous frequency drift in the three-dimensional feature set. The purpose of this verification is to correct misjudgments caused by weak or distorted signals, ensuring that the detection results accurately reflect the real-time operating status of the power grid.

[0070] In detail, the above verification steps establish a closed-loop guarantee before the fault detection results are output by constructing a dual verification mechanism. The first verification is based on Kirchhoff's Current Law (KCL), which accurately calculates the zero-sequence current conservation error for the topological characteristics of the candidate fault section. For example, in a forward power supply scenario, the rationality of the candidate section is judged by combining the zero-sequence current vector sum of the incoming cabinet, outgoing cabinet, and other ring network cabinets. If the error exceeds the adaptive threshold, it triggers the re-extraction of three-dimensional features or adjustment of deep learning model parameters, thereby effectively eliminating erroneous candidate sections caused by harmonic interference.

[0071] The second layer of verification identifies logical inconsistencies by comparing the degree of matching between distributed power source output fluctuations and fault characteristics. For example, when output fluctuates slightly but high-frequency characteristic distortion is severe, or when output fluctuates drastically but characteristics show no obvious abnormalities, both are judged as logical inconsistencies, thus correcting misjudgments. The two verifications do not function independently but complement each other: KCL verification focuses on physical-level current balance, resolving positioning errors; output fluctuation consistency verification focuses on operational-level logical matching, resolving characteristic misjudgments, thereby covering the main risk points of signal distortion and harmonic interference.

[0072] Furthermore, the aforementioned verification mechanism complements the deep learning model and power supply direction adaptive strategy described above. For example, by combining the fault probability and candidate fault segments output by the convolutional neural network-long short-term memory network (CNN-LSTM) model enhanced with a trained attention mechanism, and the power supply direction judgment strategy adjusted based on distributed power source output prediction information, dual verification can more accurately eliminate erroneous candidate segments, ensuring the accuracy and robustness of the detection results. This multi-layered technical combination significantly improves the reliability of high-resistance grounding fault detection, especially demonstrating excellent performance in dynamic grid environments caused by distributed power source integration.

[0073] This application further proposes to perform zero-sequence current conservation verification of KCL, including: if the calculated zero-sequence current conservation error is greater than a preset error threshold, then triggering a re-fault detection or outputting a verification failure flag.

[0074] In practical applications, zero-sequence current conservation error refers to the quantification result of the deviation of the zero-sequence current inflow and outflow vector sum of the candidate fault section based on Kirchhoff's current law. It can be achieved by dynamic weight adjustment. Especially when the output of distributed power sources fluctuates drastically, the interference of distortion current on error calculation can be reduced by introducing dynamic weights.

[0075] The preset error threshold is not a fixed value, but is adaptively adjusted according to the system capacitor current, the accuracy of the zero-sequence current transformer, and the output fluctuation level of the distributed power source. For example, it is set to 0.5A when the output fluctuation is ≤10%, and relaxed to 0.8A when the fluctuation is >30%, so as to adapt to different scenario requirements.

[0076] Specifically, triggering a re-detection of faults means that when the error is slightly out of range, the feature re-extraction process is initiated first and the data is re-input into the deep learning model for analysis, thereby correcting the initial misjudgment; while outputting a verification failure flag means that when the error is seriously out of range, the candidate fault segment is directly locked and a clear fault indication signal is provided, which facilitates subsequent manual intervention.

[0077] Specifically, this solution addresses the core issue of lack of automatic response when verification fails in complex scenarios by constructing a verification mechanism of "error judgment - graded response - closed-loop error correction". First, based on the pre-determined power supply direction and candidate fault section topology, the zero-sequence current conservation error is accurately calculated, and the error judgment benchmark is dynamically adjusted in combination with the output fluctuation characteristics of distributed power sources to ensure that the error can truly reflect the degree of verification failure.

[0078] Secondly, when the error exceeds the preset threshold, a graded response is triggered based on the severity of the error: minor errors trigger the re-detection process first, using the adaptive denoising function of the improved HHT to recover the three-dimensional features, and then re-inputting them into the attention CNN-LSTM model for analysis, thereby correcting the initial misjudgment caused by signal distortion; severe errors directly output a verification failure flag, while locking the candidate fault section to prevent erroneous results from being transmitted to the main station or maintenance process.

[0079] Furthermore, this solution organically combines KCL verification with preceding technical features, such as improved HHT feature extraction and attention-based CNN-LSTM models, forming a closed-loop design that significantly enhances adaptive error correction capabilities under complex operating conditions. Through this technical solution, not only is the issue of missing response when verification fails resolved, but the false positive rate and fault handling delay are also significantly reduced. Especially in scenarios where distributed power supply access leads to zero-sequence current distortion, the accuracy of verification anomaly identification is significantly improved.

[0080] Example 2

[0081] In another embodiment, this application also discloses an electronic device, including: one or more processors; a memory for storing one or more programs; when one or more programs are executed by one or more processors, the one or more processors implement the fault detection method for the ring main unit as described above.

[0082] The core innovation of this embodiment lies in its synergistic combination of nonlinear non-stationary signal processing and deep learning models. This effectively solves the problem of misjudgment caused by weak zero-sequence current signals, harmonic interference, difficulty in feature extraction, and frequent switching of power supply direction when detecting high-resistance single-phase grounding faults in ring main units of distributed power sources connected to resonant grounding systems. Specifically, this scheme utilizes nonlinear non-stationary signal processing technology to extract multi-dimensional feature sets, avoiding the limitations of traditional single feature extraction methods when dealing with signal distortion. Simultaneously, the nonlinear modeling capability of the deep learning model adapts to signal distortion caused by power output fluctuations in distributed power sources, significantly reducing the possibility of misjudgment and improving the accuracy and reliability of fault detection.

[0083] Furthermore, this electronic device, through a hardware and software collaborative architecture, provides a reliable execution platform for high-resistance grounding fault detection in ring main units. Specifically, the electronic device, as the overall carrier, constructs the physical foundation for the fault detection method; one or more processors are responsible for executing computational instructions, processing monitoring data in real time, and running the fault detection logic; the memory is used to store the fault detection program, ensuring the integrity and callability of the method steps. When the program is executed by the processor, it drives the processor to implement the fault detection method, thereby integrating signal processing, feature extraction, and model judgment into the device. These technical features work together to ensure accurate output of fault detection results even under conditions of weak signals, severe interference, and dynamic changes in power supply direction, avoiding the misjudgment defects of traditional methods.

[0084] Example 3

[0085] In another embodiment, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned fault detection method for the ring main unit.

[0086] The core innovation of this application lies in combining the three-dimensional feature set extracted by nonlinear non-stationary signal processing with a deep learning model in an intelligent analysis manner, thereby effectively solving the core problem of high-resistance single-phase grounding fault detection in ring main units after distributed power sources are connected to a resonant grounding system. Specifically, nonlinear non-stationary signal processing specifically extracts the fundamental component features, the instantaneous frequency features of high-frequency components, and the energy features of characteristic frequency bands, avoiding the limitations of traditional single feature extraction methods when signal distortion occurs. At the same time, the deep learning model utilizes its nonlinear modeling capabilities to intelligently analyze the fault probability and locate the fault segment based on the three-dimensional feature set, significantly reducing the possibility of misjudgment caused by power output fluctuations of distributed power sources. The overall solution achieves effective detection of high-resistance grounding faults through systematic design, overcoming the shortcomings of existing technologies in scenarios with weak signals, harmonic interference, and power output fluctuations.

[0087] In practical applications, acquiring zero-sequence current signals from multiple monitoring points in a ring main unit can be achieved using current sensors installed at key nodes of the unit, such as Hall effect sensors or Rogowski coil sensors. The primary purpose is to collect zero-sequence current signals as the basis for subsequent analysis. Furthermore, nonlinear, non-stationary signal processing of the zero-sequence current signal can be understood as a process of decomposing, filtering, and extracting features from the signal using specific algorithms. Specifically, empirical mode decomposition combined with wavelet transform, or variational mode decomposition (VMD) techniques can be used to process the signal, thereby extracting fundamental component features, instantaneous frequency features of high-frequency components, and energy features of characteristic frequency bands. These features provide complementary information from the perspectives of low-frequency stability, high-frequency dynamic changes, and energy distribution, addressing the problem of weak signals and harmonic interference under high-resistance grounding faults.

[0088] Specifically, inputting the three-dimensional feature set into the trained deep learning model can be achieved using various neural network architectures, such as a hybrid model combining convolutional neural networks with gated recurrent units, or a feature enhancement model based on autoencoders. The main purpose is to use the nonlinear modeling capabilities of deep learning to intelligently analyze complex signals, thereby outputting fault probabilities and candidate fault segments.

[0089] The working principle of this application is as follows: First, by acquiring the zero-sequence current signals from multiple monitoring points in the ring main unit, raw data support is provided for subsequent analysis, ensuring the basic reliability of fault detection.

[0090] Furthermore, nonlinear nonstationary signal processing is performed on the zero-sequence current signal to extract a three-dimensional feature set for characterizing high-resistance grounding faults. The three-dimensional feature set includes fundamental component features, instantaneous frequency features of high-frequency components, and energy features of characteristic frequency bands.

[0091] Specifically, the fundamental frequency component characteristics are used to capture fundamental frequency information under fault conditions, thereby stably characterizing low-frequency characteristics; the instantaneous frequency characteristics of the high-frequency components dynamically reflect the instantaneous changes of harmonic interference, enhancing anti-interference capabilities; and the characteristic frequency band energy characteristics quantify the energy distribution of key frequency bands to identify fault-specific modes.

[0092] Thus, the three elements work together to form a robust feature set, overcoming the limitations of traditional methods in feature extraction under signal distortion. Next, the three-dimensional feature set is input into a trained deep learning model, which outputs the fault probability and candidate fault segments.

[0093] As a preferred implementation method, deep learning models utilize their nonlinear modeling capabilities to intelligently analyze fault probability and locate fault sections based on three-dimensional feature sets, effectively adapting to signal distortion caused by power output fluctuations of distributed power sources and avoiding misjudgments caused by simple threshold judgments.

[0094] Finally, based on the fault probability and candidate fault sections, the fault detection results of the ring main unit are output. The final detection decision is formed through the comprehensive model output, ensuring the reliability and pertinence of the results. Overall, signal acquisition provides the data foundation, nonlinear and non-stationary signal processing extracts key features, and deep learning models realize intelligent judgment. These three aspects are interconnected and jointly solve the detection challenges caused by weak signals, interference, and power fluctuations under high-resistance faults.

[0095] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A fault detection method for a ring main unit, characterized in that, include: Acquire zero-sequence current signals from multiple monitoring points in the ring main unit; The zero-sequence current signal is subjected to nonlinear non-stationary signal processing to extract a three-dimensional feature set for characterizing high-resistance grounding faults. The three-dimensional feature set includes fundamental component features, high-frequency component instantaneous frequency features, and characteristic frequency band energy features. The three-dimensional feature set is input into a trained deep learning model, which outputs the fault probability and candidate fault segments. Based on the fault probability and candidate fault sections, the fault detection results of the ring main unit are output.

2. The fault detection method for ring main unit according to claim 1, characterized in that, The nonlinear, non-stationary signal processing of the zero-sequence current signal to extract a three-dimensional feature set includes: The zero-sequence current signal is processed using time-frequency analysis methods to extract the fundamental component characteristics, the instantaneous frequency characteristics of the high-frequency components, and the energy characteristics of the characteristic frequency bands.

3. The fault detection method for ring main unit according to claim 2, characterized in that, The time-frequency analysis method is the Hilbert-Huang transform (HHT) method.

4. The fault detection method for ring main unit according to claim 3, characterized in that, The process of processing zero-sequence current signals using the Hilbert-Huang Transform (HHT) method includes: The zero-sequence current signal was decomposed into multiple intrinsic mode function (IMF) components using ensemble empirical mode decomposition (EEMD). Valid IMF components are selected from the plurality of IMF components, and the valid IMF components are determined by adaptive noise threshold filtering. The effective IMF components are subjected to Hilbert transform to construct the three-dimensional feature set.

5. The fault detection method for ring main unit according to claim 1, characterized in that, The deep learning model is an attention-enhanced convolutional neural network-long short-term memory network (CNN-LSTM) model.

6. The fault detection method for ring main unit according to claim 1, characterized in that, Also includes: Obtain output prediction information for distributed power sources; Based on the power output prediction information, the power supply direction determination strategy is adaptively adjusted. Specifically, when the three-dimensional feature set is input into the deep learning model, fault detection is performed in conjunction with the adjusted power supply direction judgment strategy.

7. The fault detection method for ring main unit according to claim 1, characterized in that, Before outputting the fault detection results, a verification step is also included: The candidate fault section is subjected to zero-sequence current conservation verification based on Kirchhoff's current law (KCL). And / or, The consistency of the fault detection results with the output fluctuation trend of the distributed power source is verified.

8. The fault detection method for ring main unit according to claim 7, characterized in that, The zero-sequence current conservation check of KCL includes: If the calculated zero-sequence current conservation error is greater than the preset error threshold, then the fault detection will be re-performed or the output verification failure flag will be triggered.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the fault detection method for the ring main unit as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the fault detection method for the ring main unit as described in any one of claims 1-8.

Images