Low-voltage power distribution network high-resistance fault diagnosis method and system, medium and product

Abstract

The application discloses a kind of low-voltage distribution network high resistance fault diagnosis method, system and medium, product, belong to distribution network fault diagnosis technical field.The method first constructs the measurement uncertainty comprehensive statistical model fusing amplitude phase error, time asynchronization error and communication packet loss, obtains measurement data with uncertainty;Then extract multi-dimensional node features and integrate uncertainty, construct feature vector;Then based on power grid topology, construct Bayesian graph convolutional neural network and train;Finally, output fault probability and cognitive uncertainty, combined with double threshold to complete fault determination.The application can effectively deal with smart meter measurement uncertainty problem, enhance the weak feature recognition ability of high resistance fault, reduce the false alarm and miss rate, and has excellent robustness and diagnostic accuracy under different fault resistance and noise level.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for low-voltage distribution networks, specifically to a method, system, medium, and product for diagnosing high-resistance faults in low-voltage distribution networks. Background Technology

[0002] With the rapid growth of distributed power sources and electricity loads, the operating environment of low-voltage distribution networks is becoming increasingly complex. High-impedance faults (HIFs), due to their small fault current amplitude and strong randomness, are difficult to be reliably detected by traditional protection devices, and have become a significant problem restricting the safe operation of distribution networks. In recent years, the large-scale deployment of smart meters has provided a new data foundation for the perception of distribution network operating status and fault diagnosis. However, limited by cost and communication conditions, the measurement accuracy of smart meters is limited and susceptible to factors such as transformer errors, clock asynchrony, and quantization noise, which further mask the weak characteristics of HIFs.

[0003] Existing research on high-resistance faults mostly employs data-driven methods, extracting harmonic features, sequence component features, or time-domain statistical features, and combining them with machine learning models for fault identification. While these methods have achieved some success under ideal measurement conditions, they typically assume accurate and synchronized measurement data, neglecting unavoidable uncertainties in low-cost smart meters such as transformer errors, clock drift, and quantization noise. In actual operating environments, when load fluctuations intensify or communication conditions are limited, measurement errors often exhibit random and cumulative characteristics, further masking the already subtle characteristics of high-resistance faults.

[0004] In summary, how to achieve robust and reliable diagnosis of high-resistance faults under conditions of measurement uncertainty remains a critical issue that urgently needs to be addressed in the field of smart grids. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, medium, and product for diagnosing high-resistance faults in low-voltage distribution networks, in order to solve the technical problem that high-resistance faults in low-voltage distribution networks are difficult to diagnose robustly and reliably under conditions of measurement uncertainty.

[0006] To achieve the objectives of this invention, the technical solution provided by this invention is as follows: First aspect This invention provides a method for diagnosing high-resistance faults in low-voltage distribution networks, comprising the following steps: Step S1: Quantitatively model the amplitude and phase measurement deviations to generate amplitude and phase measurement error vectors; model the uncertainty caused by clock asynchrony to obtain clock deviation error vectors; model the uncertainty of data loss and communication fluctuations to construct a random masking factor; integrate the amplitude and phase measurement error vectors, clock deviation error vectors, and random masking factor to establish a comprehensive statistical model of measurement uncertainty to characterize the deviation mapping relationship between the ideal real electrical state and the actual observation data. Step S2: Extract the initial voltage observation, initial current observation, and initial phase observation collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features; Step S3: Based on the observation dataset containing uncertainty features, mine the initial multidimensional features of the node dimension to construct a single time window feature vector; based on the single time window feature vector, calculate the multidimensional node feature descriptor that aggregates the measurement uncertainty information; the initial multidimensional features include: the proportion of higher harmonics, the variance of harmonic energy fluctuation, the proportion of positive sequence voltage and the proportion of negative sequence voltage, the variance of positive sequence components, the variance of negative sequence components, the kurtosis of the current sequence, and the skewness of the current sequence. Step S4: Retrieve the topology connection data and line impedance parameters of the low-voltage distribution network, and calculate and generate a normalized adjacency matrix; combine the multi-dimensional node feature descriptor that aggregates the metering uncertainty information with the normalized adjacency matrix to build a Bayesian graph convolutional deep network; perform parameter training and optimization on the Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model. Step S5: Using a well-trained Bayesian graph network model, infer and calculate the probability value of each smart meter node malfunctioning and the corresponding cognitive uncertainty of the smart meter node; combine the probability value of each smart meter node malfunctioning and the cognitive uncertainty, and finally evaluate and output the fault diagnosis result of each smart meter node.

[0007] Second aspect This invention provides a high-resistance fault diagnosis system for low-voltage distribution networks, used to execute the aforementioned high-resistance fault diagnosis method for low-voltage distribution networks. The system includes the following functional units: The measurement uncertainty comprehensive statistical model building unit is configured to perform quantitative modeling of amplitude and phase measurement deviations to generate amplitude and phase measurement error vectors; model the uncertainty caused by clock asynchrony to obtain clock deviation error vectors; model the uncertainty of data loss and communication fluctuations to construct random masking factors; and integrate the amplitude and phase measurement error vectors, clock deviation error vectors, and random masking factors to establish a measurement uncertainty comprehensive statistical model for characterizing the deviation mapping relationship between ideal real electrical state and actual observation data. An uncertain observation data generation unit is configured to extract the initial voltage observation, initial current observation, and initial phase observation values ​​collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features. A multidimensional node feature descriptor construction unit is configured to mine initial multidimensional features of the node dimension based on the observation dataset containing uncertain features, so as to construct a single time window feature vector; based on the single time window feature vector, a multidimensional node feature descriptor that aggregates measurement uncertainty information is calculated; the initial multidimensional features include: proportion of higher harmonics, variance of harmonic energy fluctuation, proportion of positive sequence voltage and proportion of negative sequence voltage, variance of positive sequence component, variance of negative sequence component, kurtosis of current sequence, and skewness of current sequence; A Bayesian graph network construction and training unit is configured to retrieve topology connection data and line impedance parameters of the low-voltage distribution network and calculate and generate a normalized adjacency matrix; a Bayesian graph convolutional deep network is constructed by combining the multi-dimensional node feature descriptor that aggregates metering uncertainty information with the normalized adjacency matrix; parameter training and optimization are performed on the Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model. The final fault diagnosis and judgment unit is configured to use a well-trained Bayesian graph network model to infer and calculate the probability value of each smart meter node's fault and the corresponding cognitive uncertainty of the smart meter node; by combining the probability value and cognitive uncertainty of each smart meter node's fault, the unit finally evaluates and outputs the fault diagnosis result of each smart meter node.

[0008] Third aspect The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the various steps of the high-resistance fault diagnosis method for low-voltage distribution networks.

[0009] Fourth aspect This invention provides a computer program product, including a computer program that, when executed by a processor, can implement the various steps of the high-resistance fault diagnosis method for low-voltage distribution networks.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: The technical solution provided by this invention, firstly, addresses the common problems of amplitude error, phase deviation, time asynchrony, and communication instability in low-cost smart meters by establishing a unified statistical model for measurement uncertainty, providing a foundation for subsequent feature extraction and probabilistic inference. Secondly, it integrates multi-dimensional information such as high-order harmonic ratios, sequence components, and statistical characteristics at the edge, effectively enhancing the separability of weak features of high-resistivity faults and reducing the interference of load fluctuations on diagnostic results. Finally, it introduces a Bayesian graph network in the cloud, explicitly incorporating measurement uncertainty into the model parameter distribution, and achieves a joint output of fault probability and cognitive uncertainty through variational inference and Monte Carlo prediction.

[0011] Multiple experimental verifications based on standard IEEE power distribution system test networks and actual field datasets demonstrate that the technical solution proposed in this invention exhibits excellent diagnostic performance under different fault resistances and measurement noise levels. Simultaneously, the uncertainty-aware alarm mechanism significantly reduces the false alarm rate and false negative rate, making the fault diagnosis results more reliable and robust. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the low-voltage distribution network high-resistance fault diagnosis method provided in an embodiment of the present invention; Figure 2 A schematic diagram showing the accuracy verification results of the simulation platform data under different fault resistances; Figure 3 A schematic diagram showing the accuracy verification results of the simulation platform data under different fault resistances; Figure 4 This is a schematic diagram illustrating the recall rate verification results under different fault resistances in the simulation platform data. Figure 5 This is a schematic diagram illustrating the F1 score verification results under different fault resistors in the simulation platform data. Figure 6 This is a schematic diagram showing the accuracy verification results of the simulation platform data under different measurement noise levels. Figure 7 This is a schematic diagram showing the accuracy verification results of the simulation platform data under different measurement noise levels. Figure 8 This is a schematic diagram illustrating the recall verification results under different noise levels in the simulation platform data. Figure 9 This is a schematic diagram showing the F1 score verification results under different measurement noise levels of the simulation platform data. Figure 10 A schematic diagram showing the accuracy verification results for different fault resistances based on actual field data. Figure 11 A schematic diagram showing the accuracy verification results under different fault resistances based on actual field data. Figure 12 This is a schematic diagram illustrating the recall rate verification results under different fault resistances based on actual field data. Figure 13 This is a schematic diagram illustrating the F1 score verification results under different fault resistors based on actual field data. Figure 14 This is a schematic diagram showing the accuracy verification results of actual field data under different measurement noise levels. Figure 15 This is a schematic diagram illustrating the accuracy verification results of actual field data under different measurement noise levels. Figure 16 This is a schematic diagram illustrating the recall rate verification results under different noise levels in actual field data. Figure 17 This is a schematic diagram showing the F1 score verification results under different noise levels measured in actual field data. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] It should be noted that the acquisition of data and collection of information in this application are legal, compliant, or obtained with the consent of the subject of the data collection.

[0015] like Figure 1 As shown in the figure, this embodiment provides a method for diagnosing high-resistance faults in low-voltage distribution networks, including the following steps: Step S1: Quantitatively model the amplitude and phase measurement deviations to generate amplitude and phase measurement error vectors; model the uncertainty caused by clock asynchrony to obtain clock deviation error vectors; model the uncertainty of data loss and communication fluctuations to construct a random masking factor; integrate the amplitude and phase measurement error vectors, clock deviation error vectors, and random masking factor to establish a comprehensive statistical model of measurement uncertainty to characterize the deviation mapping relationship between the ideal real electrical state and the actual observation data. Step S1 specifically includes the following: Step S1.1: Mathematically abstract the deviations existing in the amplitude and phase observation process, and construct the amplitude and phase measurement error vector. The specific derivation process is as follows: The observed perturbations of amplitude and phase are approximated to follow a normal distribution with zero mean, expressed as: ; In the formula, Characterizes the properties of a normal distribution; Sampling time; The covariance matrix representing the corresponding measurement bias has the following structure: ; In the formula, These refer to the statistical variances of voltage observation deviation, current observation deviation, and phase angle observation deviation, respectively; from this, the perturbed voltage observation value can be obtained. Current observation value With phase angle observations The mapping relationship is as follows: ; ; ; in, Refers to the actual voltage parameter under ideal, interference-free conditions; Refers to the actual current parameter under ideal, interference-free conditions; The actual phase angle parameter of the agent under interference-free conditions; The measurement deviation component in the voltage dimension has a zero-mean normal distribution property; The measurement deviation component in the current dimension has the property of zero-mean normal distribution; The measurement deviation component in the phase angle dimension also possesses the property of a zero-mean normal distribution; By introducing the aforementioned normal distribution to quantify random disturbances that are difficult to physically eliminate, such uncertainties are subordinate to the background noise inherent in the data, which cannot be completely eliminated even with an infinitely expanded sample size.

[0016] Step S1.2: Analyze and abstract the uncertainty caused by clock asynchrony, and derive the clock deviation error vector. : Regarding the first Each smart meter node has its corresponding actual sampling trigger time. Satisfying Relationship: ; Furthermore, the drift at that moment approximately conforms to a normal distribution with zero mean: ; In the formula, The statistical variance characterizing the clock drift variable; Corresponding to the The specific clock offset values ​​of each smart meter node; under the boundary assumption of small clock drift, the real electrical parameter signals are approximately linearized using a first-order Taylor expansion, as follows: ; ; In the formula, for Ideal, real electrical parameters that are undisturbed at all times for Ideal, real electrical parameters that are undisturbed at all times; based on the above deduction logic, the asynchronous problem at the clock level can be equivalently transformed into uncertainty at the data measurement level. This effectively avoids the complex assumption problems brought about by high-order nonlinear modeling, and thus provides a mathematical expression with clear analytical meaning for the subsequent uncertainty propagation.

[0017] Step S1.3: Perform quantitative modeling to address data gaps and communication link uncertainties, and derive the random masking factor. ; The random masking factor is defined as a discrete random variable that follows a Bernoulli distribution: ; The specific value selection logic of this random variable is as follows: (This indicates that the data has been successfully uploaded to the central server.) (This indicates that data at this sampling point has been lost); Its probability distribution function satisfies: and ,in, For the first Data communication success rate metric for each smart meter node; Step S1.4: Perform linear fusion of the error terms: ; By systematically modeling multiple uncertainties such as amplitude and phase deviation, time synchronization disturbance, and network communication packet loss, the reliability of the measurement sequence can be explicitly measured and quantified, thereby defining the credible boundary of the observation data. The model can deeply characterize the statistical evolution law between the field measured data and the ideal electrical truth value, providing a data foundation with confidence support for subsequent feature space construction and fault logic reasoning. Step S1.5: Establish the comprehensive statistical model for measurement uncertainty, whose mathematical expression is as follows: ; ; In the formula, Indicates the first Each node Comprehensive observational data generated in real time and incorporating characteristics of uncertainty; Represents the Hadamard product. This is the combined measurement error term for amplitude and phase; This is the error term caused by clock drift; A random masking factor that reflects the state of data packets; Step S2: Extract the initial voltage observation, initial current observation, and initial phase observation collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features; Step S3: Based on the observation dataset containing uncertainty features, mine the initial multidimensional features of the node dimension to construct a single time window feature vector; based on the single time window feature vector, calculate the multidimensional node feature descriptor that aggregates the measurement uncertainty information; the initial multidimensional features include: the proportion of higher harmonics, the variance of harmonic energy fluctuation, the proportion of positive sequence voltage and the proportion of negative sequence voltage, the variance of positive sequence component, the variance of negative sequence component, the kurtosis of the current sequence, and the skewness of the current sequence.

[0018] Specifically, it includes the following: Step S3.1: For the observation dataset with uncertain characteristics and a total observation duration of , divide it into continuous and non-overlapping observation sub-intervals at equal intervals; Step S3.2: Extract initial multidimensional features for each observation sub-interval; the initial multidimensional features include: proportion of higher harmonics. Harmonic energy fluctuation variance Positive sequence voltage ratio Negative sequence voltage ratio , positive order component variance negative order component variance kurtosis of current sequence and current sequence skewness ; Step S3.3: Integrate the initial multidimensional features extracted above to construct a single-time-window feature vector for the m-th observation sub-interval. The expression is as follows: ; Among them, higher harmonics account for a certain percentage. The quantification formula is: ; In the formula, Refers to the first Each node The time measured at the first The amplitude of the second harmonic current; and The lowest and highest order boundaries for higher harmonic analysis were defined respectively; This represents the amplitude of the fundamental current at the corresponding moment; Among them, the variance of harmonic energy distribution The measurement formula is: ; ; In the formula, Indicates the first Each node The corresponding time Subharmonic energy accumulation; The sliding integral variable within the sampling time window; For variance operators; Refers to the first Each node The time measured at the first The amplitude of the second harmonic current; Among them, the positive sequence voltage ratio The derivation formula is as follows: ; In the formula, Corresponding positive sequence voltage component amplitude; Corresponding to the amplitude of the negative sequence voltage component; Among them, the negative sequence voltage ratio The derivation formula is as follows: ; In the formula, Represents the magnitude of the zero-sequence voltage component; Among them, the variance of the positive order components The calculation relationship is as follows: ; Among them, the variance of the negative order component The calculation relationship is as follows: ; Among them, the kurtosis of the current waveform The evaluation formula is as follows: ; ; In the formula, and They depicted the first The statistical standard deviation and mathematical expectation of the current observation sequence corresponding to each node; That is, the current observation sequence values ​​in the time dimension; Represents the kurtosis operation function; Represents the expected value in statistical mathematics; Among them, current waveform skewness The evaluation formula is as follows: ; ; In the formula, Refers to the skewness operation function; Step S3.4: Perform uncertainty fusion statistical calculations on the single-time-window feature vector sequence, using the following formula: ; ; In the formula, Characterizing the first The characteristic mathematical expectation vector (i.e., mean vector) of a smart meter node. Characterizing the first The feature standard deviation vector of a smart meter node.

[0019] Step S3.5: Combine the feature expectation vector and the feature standard deviation vector to generate a multi-dimensional node feature descriptor that aggregates quantitative uncertainty information. The mapping relationship is as follows:

[0020] ; The core efficiency of this step lies in transforming the noisy observation sequences captured by front-end sensors into robust, highly refined, and physically clear node feature representations. By extracting discrimination information such as the proportion of higher-order harmonics, order component evolution features, and statistical moments at the edge, and combining this with time window statistics to encode the uncertainties at the metrological level, this module significantly reduces cloud-edge communication overhead while greatly mitigating the interference of random load fluctuations and environmental background noise. The resulting feature descriptors not only fully preserve the essential characteristics of high-resistance faults but also provide a high-confidence data foundation for probabilistic inference based on graph deep learning models in the cloud.

[0021] Step S4: The physical topology data and line impedance parameters of the low-voltage distribution network are used to calculate and generate a normalized adjacency matrix. A Bayesian graph convolutional deep network is constructed by combining the aforementioned multi-dimensional node feature descriptors that aggregate metering uncertainty information with the normalized adjacency matrix. The parameters of the Bayesian graph convolutional deep network are iteratively optimized to obtain a well-trained Bayesian graph network model. It should be noted that in this step, the multi-dimensional feature matrix extracted from the edge side is transmitted to the cloud processing center in real time. The Bayesian graph convolutional network (BYGNet) is used to perform probabilistic modeling and feature fusion on the edge features, outputting fault judgment conclusions and quantifying the cognitive uncertainty of the prediction results in real time. This suppresses false alarms at the source and achieves highly reliable fault diagnosis.

[0022] The specific execution process is as follows: Step S4.1: Based on the physical topology data and line impedance parameters of the low-voltage distribution network, calculate and generate the normalized adjacency matrix. The detailed process includes: modeling the low-voltage distribution network as a graph structure. ; In the formula, This is a vertex set, corresponding to all smart meter nodes within the system; Let edge set be used to characterize the electrical connection state between nodes; the weighted adjacency matrix of the graph structure is defined as follows: Its internal elements The logic for determining the value is as follows: ; In the formula, Representation Nodes With nodes The equivalent impedance of the lines between them; the weighted adjacency matrix is ​​standardized to generate a normalized adjacency matrix. : ; In the formula, Let be the corresponding degree matrix, whose diagonal elements satisfy... ; It is the inverse square root of the degree matrix; Step S4.2: Joint multi-dimensional node feature descriptors With normalized adjacency matrix A Bayesian graph convolutional deep network is constructed; the construction process of the Bayesian graph convolutional deep network includes: The first step is to calculate the feature representation matrix of the first-layer hidden state through the feature initialization module. The calculation formula is as follows: ; In the formula, Refers to a non-linear activation function; Corresponding to the weight matrix of the first layer network; The second step involves performing in-depth node feature aggregation calculations using the topological information transmission layer. The iterative formula is as follows: ; In the formula, For the first The latent state feature representation matrix of order; This is the index of the current layer in the Bayesian graph convolutional deep network; For the first The latent state feature representation matrix of order; For the first The connection weight matrix of the hierarchy; The third step involves using the Bayesian weight modeling module to assign probability distribution constraints to the network weights at each level, as shown in the following expression: ; In the formula, For the first The mathematical expectation vector of the weight distribution; For the first The diagonal covariance matrix of the weight distribution; It represents a normal distribution model; it is worth emphasizing that, unlike conventional deterministic graph convolutional networks, BYGNet uses the weight matrix of each layer... Treating them as random variables gives the model the ability to predict probability distributions (rather than a single numerical result).

[0023] Step S4.3: Perform iterative parameter optimization on the constructed Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model; Specifically, in the forward propagation stage of this Bayesian graph convolutional deep network, the connection weights are computed as random variables. Variational inference techniques are used to approximate and optimize the posterior distribution of the model weights, with maximizing the variational evidence lower bound (ELBO) set as the optimization criterion. Its objective loss function... The structure is as follows: ; In the formula, This refers to the variational approximation of the posterior distribution of network weights. Statistical expectation under constraints; Refers to the initial prior distribution of network weights; It describes the conditional likelihood probability of fault prediction given input features and weights; This is a relative entropy penalty term used to measure the divergence between the variational posterior and prior distributions. In this loss function, the first term aims to ensure the model's ability to predict and reconstruct faults, while the tail term, relative entropy, serves as a regularization penalty factor to constrain the model's topological complexity and prevent overfitting.

[0024] Step S5: Using a well-trained Bayesian graph network model, the probability value of each smart meter node malfunctioning and the corresponding cognitive uncertainty of the smart meter node are calculated. Combining the probability value and cognitive uncertainty of each smart meter node malfunctioning, the fault diagnosis result of each smart meter node is finally evaluated and output. The specific execution logic of the above steps is as follows: Step S5.1: For a well-trained Bayesian graphical network model, independently extract from the posterior probability distribution of its parameters. Group network weight samples, and the aforementioned multidimensional node feature descriptors that aggregate econometric uncertainty information. Feeding into the network model, forward inference is performed sequentially using each set of extracted network weight samples to obtain the single prediction probability value for each sampling round. The mapping relationship is as follows: ; In the formula, The forward mapping function represents a well-trained Bayesian graph network model; Indicates the first The network weight matrix obtained from each sampling round; Characterizing the first The predicted probability value obtained from the inference. , To determine the maximum number of inferences, ; Step S5.2: For The arithmetic mean of the single prediction probability values ​​is used as the criterion for evaluation. Expected probability of a smart meter node failing : ; Step S5.3: Calculate the cognitive uncertainty of the corresponding smart meter node. The calculation formula is as follows: ; Step S5.4: Combine the expected probabilities With cognitive uncertainty Output the final fault diagnosis results for each smart meter node: like If so, a fault diagnosis alarm command will be triggered; like and If so, it is judged as a suspected fault state and marked as requiring continuous monitoring; like If so, the node is determined to be operating normally; In the formula, and These represent the critical threshold for probability determination and the uncertainty tolerance threshold, respectively, determined by combining historical operational records and prior expert knowledge.

[0025] It should be further explained that this invention utilizes a Bayesian graph convolutional network in a cloud processing center to achieve probabilistic inference and diagnosis of edge-side features. Under the constraints of the distribution network's physical topology, the invention leverages graph convolution mechanisms to deeply mine the spatial correlation characteristics between nodes and creatively sets the network weight parameters as randomly distributed variables, thereby explicitly characterizing the model's cognitive blind spots (i.e., cognitive uncertainty). By employing Monte Carlo multiple sampling techniques to perform iterative forward inference, the model can not only output the expected probability of fault occurrence for each physical node but also quantitatively assess the reliability boundary of this prediction. In harsh operating conditions where high-resistance fault electrical characteristics are extremely weak and on-site measurement noise interference is strong, this mechanism can effectively filter false alarms induced by uncertain measurements, thus providing highly confident diagnostic decision support for the safe and stable operation of the distribution network.

[0026] Simulation platform data verification To verify the effectiveness of the present invention, the IEEE 33-bus was used as a test platform, and simulation data was used to verify the solution. The fault resistance range is... The fault location is randomly selected from line nodes. Smart meters are deployed at each node, with a sampling frequency of 6.4kHz. The measured variables include the voltage value. Current measurement value Phase angle measurement value Furthermore, amplitude noise, phase noise, time asynchrony, and random communication packet loss are introduced to simulate the real measurement environment of low-cost smart meters. The measurement metrics used to evaluate the method's performance are as follows: accuracy (…). ), recall rate ( ), accuracy ( ) and F1 score ( ): ; ; ; ; in, and It is the number of correctly predicted samples. and It represents the number of samples that were predicted incorrectly.

[0027] 1) Robustness verification under different fault resistors Set the fault resistors to respectively 500 test samples were generated for each type of resistor. The BYGNet provided by this invention was compared with traditional methods that extract artificial features and use Support Vector Machines (SVM), Convolutional Neural Networks (CNN), Generative Networks (GCN), and Attention Graph Neural Networks (GAT).

[0028] The robustness verification results of the simulation platform data under different fault resistances are shown in Table 1 and Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown: Table 1

[0029] It can be seen that as the fault resistance increases, the electrical disturbances caused by high-resistance faults gradually weaken, and the recognition performance of all methods shows a downward trend. Traditional SVM and CNN, due to their reliance on local features or temporal patterns, struggle to effectively distinguish between faults and normal operating conditions under high fault resistance. GCN and GAT significantly improve overall performance by introducing distribution network topology information, with GAT outperforming GCN under medium resistance conditions, indicating that the attention mechanism can alleviate the feature weakening problem to some extent. However, in high-resistance regions (… The performance degradation of deterministic models remains quite significant.

[0030] In contrast, the BYGNet proposed in this invention maintains the highest accuracy and F1 score under all fault resistance conditions, indicating that it can more effectively suppress the adverse effects of weak features through uncertainty modeling.

[0031] 2) Stability verification experiment under different measurement noise levels Set the standard deviation of measurement noise To verify the stability of the proposed method under different measurement noise levels, BYGNet was compared with SVM, CNN, GCN, and GAT. The comparison results are shown in Table 2 and... Figure 6 , Figure 7 , Figure 8 , Figure 9 As shown: Table 2

[0032] It can be seen that the fault identification performance of all methods decreases to varying degrees with the increase of measurement noise level. Among them, SVM and CNN are the most sensitive to noise and their performance degrades rapidly. GCN and GAT, by introducing the distribution network topology, show better stability under low to medium noise conditions, but are still inevitably affected under high noise levels. Compared with GCN, GAT achieves better accuracy and F1 scores under all noise levels, indicating that the attention mechanism can adaptively adjust the importance of neighboring nodes to a certain extent, thereby reducing the impact of local noise propagation. However, since the model parameters are still deterministic, its performance still declines significantly under high noise conditions. The proposed BYGNet maintains the highest identification performance under different noise levels, especially under high noise conditions (5%), its F1 score is still about 6.1 percentage points higher than GAT. This shows that by introducing uncertainty modeling at the model parameter level, the negative impact of measurement noise on feature propagation and decision-making processes can be effectively mitigated, thereby significantly improving the stability and reliability of the model under complex operating conditions.

[0033] 3) The impact of uncertainty-aware alarms on false alarms and missed alarms Using new data, we compared the BYGNet model and the GAT algorithm (which does not have uncertainty-aware alarms) with BGCNW, and added evaluation metrics such as False Positive Rate (FPR) and False Negative Rate (FNR): ; ; The experimental results are shown in Table 3 below: Table 3

[0034] The results show that GAT (Graphical Anomaly Test) is prone to misclassifying normal samples as faults under measurement uncertainty, with an FPR (Fault Rate) as high as 9.23%. BGCNW (Graphical Network-Channeled Network) slightly reduces the false alarm rate to 7.82% by modeling with a graph structure. After introducing prediction uncertainty, BYGNet can filter low-confidence samples, reducing the FPR to 5.92%, significantly reducing the probability of false alarms. GAT is affected by measurement errors, and some fault samples are missed, with an FNR of 9.64%. BGCNW reduces this to 8.13%, indicating that graph structure and feature extraction help capture more fault features. Bayesian uncertainty perception further reduces the FNR to 6.43%, indicating that the ability to identify weak fault signals is enhanced. Introducing Bayesian uncertainty not only improves accuracy and F1 score, but also effectively reduces false alarm and missed alarm rates, making fault diagnosis more reliable and suitable for engineering deployment.

[0035] Actual field data verification To verify the generalization ability and practical application performance of the proposed solution, on-site data collection was conducted using a smart meter in a city's low-voltage distribution network. The sampling frequency was 2kHz, covering 33 distribution nodes, with a collection period of two consecutive weeks, collecting data 24 hours daily. The measured variables included voltage values. Current measurement value Phase angle measurement value The smart meter measurements exhibit actual errors: transformer error ±0.5%, clock asynchrony ±1 ms, significant load fluctuations, and large seasonal / time-period differences. The field data experimental design remains consistent with the simulation experimental data described above.

[0036] 1) Robustness verification under different fault resistors Set the fault resistors to respectively The method BYGNet provided in this invention is compared with traditional SVM CNN, GCN, and GAT. The comparison results are shown in Table 4. Figure 10 , Figure 11 , Figure 12 , Figure 13 As shown: Table 4

[0037] The comparison results show that as the fault resistance increases, the high-resistance fault features weaken, and the performance of SVM and CNN declines significantly. GCN utilizes topology to improve recognition capabilities, while GAT's attention mechanism further enhances robustness under moderate fault resistance. BYGNet maintains high performance; compared to GAT, BYGNet shows a significant performance improvement (approximately 3-5%) under high-resistance conditions, demonstrating its ability to suppress the influence of weak features through uncertainty modeling.

[0038] 2) Stability verification under different measurement noise levels Consistent with simulation data verification, the standard deviation of measurement noise was set. To verify the stability of the proposed method under different measurement noise levels, BYGNet was compared with SVM, CNN, GCN, and GAT. The comparison results are shown in Table 5. Figure 14 , Figure 15 , Figure 16 , Figure 17 As shown: Table 5

[0039] The results show that the performance of all methods decreases with increasing measurement noise, with SVM and CNN being particularly sensitive to noise. GCN and GAT mitigate some of the noise impact through graph structures, with GAT slightly outperforming GCN under low-noise conditions. At different noise levels, BYGNet's performance degradation is significantly less than that of CNN and GCN, indicating that Bayesian modeling effectively mitigates the impact of measurement noise. Even under 5% high noise conditions, BYGNet maintains nearly 96% accuracy, demonstrating strong stability and robustness.

[0040] 3) The impact of uncertainty-aware alarms on false alarms and missed alarms The BYGNet model and the GAT algorithm without uncertainty-aware alarms are compared with BGCNW. The comparison results are shown in Table 6.

[0041] Table 6

[0042] The results show that for deterministic methods like GAT and BGCNW, the models typically make fault judgments directly based on a single forward inference result. When measurement data is affected by transformer errors, communication noise, or load fluctuations, the models are prone to overconfident predictions of boundary samples, leading to a simultaneous increase in both false alarm and false negative rates. In contrast, the uncertainty-aware BYGNet, which incorporates Bayesian modeling, not only outputs fault judgment results but also quantifies prediction uncertainty. When the cognitive uncertainty corresponding to the prediction result is high, the system can mark the sample as a "low-confidence alarm" or delay the decision, effectively avoiding erroneous triggering caused by measurement uncertainty. Experimental results show that, at the same recall level, BYGNet's false alarm rate is significantly lower than that of non-Bayesian methods, indicating that it can effectively suppress false alarms under normal operating conditions.

[0043] On the other hand, traditional methods are more prone to false negatives in cases of high-resistance faults or weak signal characteristics. Uncertainty perception mechanisms, however, can identify potentially high-risk samples by statistically analyzing multiple Monte Carlo predictions. Even if the confidence of a single prediction is insufficient, uncertainty aggregation can increase the probability of fault detection. Therefore, BYGNet also shows a significantly lower false negative rate on real-world datasets, demonstrating its greater sensitivity to weak fault scenarios.

[0044] In addition, this embodiment provides a high-resistance fault diagnosis system for low-voltage distribution networks, used to execute the aforementioned high-resistance fault diagnosis method for low-voltage distribution networks. This system includes the following functional units: The measurement uncertainty comprehensive statistical model building unit is configured to perform quantitative modeling of amplitude and phase measurement deviations to generate amplitude and phase measurement error vectors; model the uncertainty caused by clock asynchrony to obtain clock deviation error vectors; model the uncertainty of data loss and communication fluctuations to construct random masking factors; and integrate the amplitude and phase measurement error vectors, clock deviation error vectors, and random masking factors to establish a measurement uncertainty comprehensive statistical model for characterizing the deviation mapping relationship between ideal real electrical state and actual observation data. An uncertain observation data generation unit is configured to extract the initial voltage observation, initial current observation, and initial phase observation values ​​collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features. A multidimensional node feature descriptor construction unit is configured to mine initial multidimensional features of the node dimension based on the observation dataset containing uncertain features, so as to construct a single time window feature vector; based on the single time window feature vector, a multidimensional node feature descriptor that aggregates measurement uncertainty information is calculated; the initial multidimensional features include: proportion of higher harmonics, variance of harmonic energy fluctuation, proportion of positive sequence voltage and proportion of negative sequence voltage, variance of positive sequence component, variance of negative sequence component, kurtosis of current sequence, and skewness of current sequence; A Bayesian graph network construction and training unit is configured to retrieve topology connection data and line impedance parameters of the low-voltage distribution network and calculate and generate a normalized adjacency matrix; a Bayesian graph convolutional deep network is constructed by combining the multi-dimensional node feature descriptor that aggregates metering uncertainty information with the normalized adjacency matrix; parameter training and optimization are performed on the Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model. The final fault diagnosis and judgment unit is configured to use a well-trained Bayesian graph network model to infer and calculate the probability value of each smart meter node's fault and the corresponding cognitive uncertainty of the smart meter node; by combining the probability value and cognitive uncertainty of each smart meter node's fault, the unit finally evaluates and outputs the fault diagnosis result of each smart meter node.

[0045] The mathematical expression of the comprehensive statistical model for measurement uncertainty is as follows: ; ; in, Indicates the first Each smart meter node at the sampling time Observational data containing uncertainties; For Hadamah accumulation; This is the amplitude and phase measurement error vector; This is the clock skew error vector; This represents the random masking factor used to characterize missing information; These are the ideal, true electrical parameters that have not been disturbed. The multi-dimensional node feature descriptor construction unit is specifically configured to perform the following: Step S3.1: For a total observation duration of... The observation dataset containing uncertainties is divided into equal intervals. A continuous and non-overlapping observation sub-interval; Step S3.2: Extract initial multidimensional features for each observation sub-interval; the initial multidimensional features include: proportion of higher harmonics. Harmonic energy fluctuation variance Positive sequence voltage ratio Negative sequence voltage ratio , positive order component variance negative order component variance kurtosis of current sequence and current sequence skewness ; Step S3.3: Integrate the initial multidimensional features extracted above to construct a single-time-window feature vector for the m-th observation sub-interval. The expression is as follows: ; Step S3.4: Perform uncertainty fusion statistical calculation on the single-time-window feature vector, using the following formula: ; ; In the formula, Characterizing the first The characteristic mathematical expectation vector of a smart meter node; Characterizing the first The characteristic standard deviation vector of a smart meter node

[0046] Step S3.5: Combine the feature expectation vector and the feature standard deviation vector to generate a multi-dimensional node feature descriptor that aggregates quantitative uncertainty information. The mapping relationship is as follows: .

[0047] Furthermore, embodiments of the present invention also provide a computer-readable storage medium that stores a computer program internally. When the computer program is loaded and executed by a processor, it can implement each step in the aforementioned method for diagnosing high-resistance faults in low-voltage distribution networks.

[0048] In addition, this embodiment of the invention also provides a computer program product, which includes a computer program. When the computer program is loaded and executed by a processor, it can also realize the various steps in the aforementioned low-voltage distribution network high-resistance fault diagnosis method.

[0049] Finally, it should be noted that the foregoing specific embodiments are merely examples and explanations of the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention to the specific embodiments listed. Those skilled in the art should understand that, without departing from the core design concept of the present invention, various equivalent substitutions, modifications, and alterations can be made to the above embodiments, and all such evolutions and extensions based on the essential spirit of the present invention should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for diagnosing high-resistance faults in low-voltage distribution networks, characterized in that, Includes the following steps: Step S1: Perform quantization modeling for amplitude and phase measurement deviations to generate amplitude and phase measurement error vectors; Model the uncertainty caused by clock asynchrony to obtain the clock deviation error vector; To address the uncertainty of data loss and communication fluctuations, a random masking factor is constructed. By integrating the amplitude and phase measurement error vector, clock deviation error vector, and random masking factor, a comprehensive statistical model for measurement uncertainty is established to characterize the deviation mapping relationship between the ideal real electrical state and the actual observed data. Step S2: Extract the initial voltage observation, initial current observation, and initial phase observation collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features; Step S3: Based on the observation dataset containing uncertain features, mine the initial multidimensional features of the node dimension to construct a single time window feature vector; Based on this single-time-window feature vector, a multi-dimensional node feature descriptor that aggregates metrological uncertainty information is calculated; The initial multidimensional features include: the proportion of higher harmonics, the variance of harmonic energy fluctuation, the proportion of positive-sequence voltage and the proportion of negative-sequence voltage, the variance of positive-sequence components, the variance of negative-sequence components, the kurtosis of the current sequence, and the skewness of the current sequence. Step S4: Retrieve the topology connection data and line impedance parameters of the low-voltage distribution network, and calculate and generate a normalized adjacency matrix; combine the multi-dimensional node feature descriptor that aggregates the metering uncertainty information with the normalized adjacency matrix to build a Bayesian graph convolutional deep network; perform parameter training and optimization on the Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model. Step S5: Using a well-trained Bayesian graph network model, infer and calculate the probability value of each smart meter node malfunctioning and the corresponding cognitive uncertainty of the smart meter node; combine the probability value of each smart meter node malfunctioning and the cognitive uncertainty, and finally evaluate and output the fault diagnosis result of each smart meter node.

2. The method for diagnosing high-resistance faults in low-voltage distribution networks according to claim 1, characterized in that, In step S1, the mathematical expression of the comprehensive statistical model for measurement uncertainty is as follows: ； ； in, Indicates the first Each smart meter node at the sampling time Observational data containing uncertainties; For Hadamah accumulation; This is the amplitude and phase measurement error vector; This is the clock skew error vector; This represents the random masking factor used to characterize missing information; These are the ideal, true electrical parameters that have not been disturbed.

3. The method for diagnosing high-resistance faults in low-voltage distribution networks according to claim 2, characterized in that, Step S3 includes the following: Step S3.1: For a total observation duration of... The observation dataset containing uncertainties is divided into equal intervals. A continuous and non-overlapping observation sub-interval; Step S3.2: Extract initial multidimensional features for each observation sub-interval; The initial multidimensional features include: the proportion of higher harmonics. Harmonic energy fluctuation variance Positive sequence voltage ratio Negative sequence voltage ratio , positive order component variance negative order component variance kurtosis of current sequence and current sequence skewness ; Step S3.3: Integrate the initial multidimensional features extracted above to construct a single-time-window feature vector for the m-th observation sub-interval. The expression is as follows: ； Step S3.4: Perform uncertainty fusion statistical calculation on the single-time-window feature vector, using the following formula: ； ； In the formula, Characterizing the first The characteristic mathematical expectation vector of a smart meter node; Characterizing the first The feature standard deviation vector of a smart meter node; Step S3.5: Combine the feature expectation vector and the feature standard deviation vector to generate a multi-dimensional node feature descriptor that aggregates quantitative uncertainty information. The mapping relationship is as follows: 。 4. The method for diagnosing high-resistance faults in low-voltage distribution networks according to claim 3, characterized in that, Step S4 specifically includes the following: Step S4.1: Retrieve the topology connection data and line impedance parameters of the low-voltage distribution network, and calculate and generate the normalized adjacency matrix. ; Step S4.2: Combine the multi-dimensional node feature descriptor that aggregates measurement uncertainty information. With normalized adjacency matrix A Bayesian graph convolutional deep network is constructed; the specific process of constructing the Bayesian graph convolutional deep network is as follows: The first step involves calculating the feature representation matrix of the first-layer hidden state through the feature initialization module. The mapping relationship is calculated as follows: ; In the formula, It is a non-linear activation function; It is the inverse square root of the degree matrix; This is the weight matrix of the first layer network; The second step involves using the topological information transmission layer to perform node feature aggregation calculations. The iterative formula is as follows: ; In the formula, For the first The latent state feature representation matrix of order; This is the index of the current layer in the Bayesian graph convolutional deep network; For the first The latent state feature representation matrix of order; For the first The connection weight matrix of the hierarchy; The third step is to use the Bayesian parameter estimation module to assign probability distribution constraints to the weights of each network layer, as shown in the following expression: ; In the formula, For the first Mathematical expectation vector of weight distribution in a network of order Corresponding to the The diagonal covariance matrix of the weight distribution of an order-order network; Characterizing the normal distribution model; Step S4.3: Perform iterative parameter optimization on the constructed Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model.

5. The method for diagnosing high-resistance faults in low-voltage distribution networks according to claim 4, characterized in that, In step S4.3, during the forward propagation stage of the Bayesian graph convolutional deep network, the connection weights are incorporated into the computation as randomly sampled variables. Variational inference techniques are used to approximate and optimize the posterior distribution of the model weights, setting the maximization of the variational evidence lower bound (ELBO) as the optimization criterion. The objective loss function is... The structure is as follows: ; In the formula, For the variational posterior probability distribution of network weights Statistical expectation under constraints; This is the initial prior distribution function of the network weights; To predict the likelihood probability of faults given input features and weights; It is a relative entropy penalty term used to measure the divergence between the variational posterior distribution and the prior distribution.

6. The method for diagnosing high-resistance faults in low-voltage distribution networks according to claim 5, characterized in that, Step S5 includes the following: Step S5.1: For a well-trained Bayesian graphical network model, independently extract from the posterior probability distribution of its parameters. Group network weight samples, and the aforementioned multidimensional node feature descriptors that aggregate econometric uncertainty information. Feeding the data into a well-trained Bayesian graphical network model, forward inference is performed sequentially using each set of extracted network weight samples to obtain the single-prediction probability value for each sampling round. The mapping relationship is as follows: ; In the formula, The forward mapping function for training mature Bayesian graph network models; For the first The network weight matrix obtained from each sampling round; For the first The predicted probability value obtained from the inference. , To determine the maximum number of inferences, ; Step S5.2: For The arithmetic mean of the single prediction probability values ​​is used as the criterion for evaluation. Expected probability of a smart meter node failing : ; Step S5.3: Calculate the cognitive uncertainty of the corresponding smart meter node. The calculation formula is as follows: ; Step S5.4: Synthesis and Output the final fault diagnosis results for each smart meter node: like If so, a fault diagnosis alarm command will be triggered; like and If so, it is determined to be a suspected fault state and marked as requiring continuous tracking and monitoring; like If so, the node is determined to be operating normally; In the formula, and These represent the critical threshold for probability determination and the uncertainty tolerance threshold, respectively, determined by combining historical operational records and prior expert knowledge.

7. A high-resistance fault diagnosis system for low-voltage distribution networks, used to execute the high-resistance fault diagnosis method for low-voltage distribution networks as described in any one of claims 1-6, characterized in that, The system includes the following functional units: The measurement uncertainty comprehensive statistical model building unit is configured to perform quantitative modeling of amplitude and phase measurement deviations and generate amplitude and phase measurement error vectors. Model the uncertainty caused by clock asynchrony to obtain the clock deviation error vector; To address the uncertainty of data loss and communication fluctuations, a random masking factor is constructed. By integrating the amplitude and phase measurement error vector, clock deviation error vector, and random masking factor, a comprehensive statistical model for measurement uncertainty is established to characterize the deviation mapping relationship between the ideal real electrical state and the actual observed data. An uncertain observation data generation unit is configured to extract the initial voltage observation, initial current observation, and initial phase observation values ​​collected at each smart meter node, and combine them with the measurement uncertainty comprehensive statistical model to generate an observation dataset containing uncertainty features. A multidimensional node feature descriptor building unit is configured to mine initial multidimensional features of the node dimension based on the observation dataset containing uncertain features, so as to construct a single time window feature vector; Based on this single-time-window feature vector, a multi-dimensional node feature descriptor that aggregates metrological uncertainty information is calculated; The initial multidimensional features include: the proportion of higher harmonics, the variance of harmonic energy fluctuation, the proportion of positive-sequence voltage and the proportion of negative-sequence voltage, the variance of positive-sequence components, the variance of negative-sequence components, the kurtosis of the current sequence, and the skewness of the current sequence. A Bayesian graph network construction and training unit is configured to retrieve topology connection data and line impedance parameters of the low-voltage distribution network and calculate and generate a normalized adjacency matrix; a Bayesian graph convolutional deep network is constructed by combining the multi-dimensional node feature descriptor that aggregates metering uncertainty information with the normalized adjacency matrix; parameter training and optimization are performed on the Bayesian graph convolutional deep network to obtain a mature Bayesian graph network model. The final fault diagnosis and judgment unit is configured to use a well-trained Bayesian graph network model to infer and calculate the probability value of each smart meter node's fault and the corresponding cognitive uncertainty of the smart meter node; by combining the probability value and cognitive uncertainty of each smart meter node's fault, the unit finally evaluates and outputs the fault diagnosis result of each smart meter node.

8. A high-resistance fault diagnosis system for low-voltage distribution networks according to claim 7, characterized in that, The mathematical expression of the comprehensive statistical model for measurement uncertainty is as follows: ； ； in, Indicates the first Each smart meter node at the sampling time Observational data containing uncertainties; For Hadamah accumulation; This is the amplitude and phase measurement error vector; This is the clock skew error vector; This represents the random masking factor used to characterize missing information; These are the ideal, true electrical parameters that have not been disturbed. The multi-dimensional node feature descriptor construction unit is specifically configured to perform the following: Step S3.1: For a total observation duration of... The observation dataset containing uncertainties is divided into equal intervals. A continuous and non-overlapping observation sub-interval; Step S3.2: Extract initial multidimensional features for each observation sub-interval; the initial multidimensional features include: proportion of higher harmonics. Harmonic energy fluctuation variance Positive sequence voltage ratio Negative sequence voltage ratio , positive order component variance negative order component variance kurtosis of current sequence and current sequence skewness ; Step S3.3: Integrate the initial multidimensional features extracted above to construct a single-time-window feature vector for the m-th observation sub-interval. The expression is as follows: ； Step S3.4: Perform uncertainty fusion statistical calculation on the single-time-window feature vector, using the following formula: ； ； In the formula, Characterizing the first The characteristic mathematical expectation vector of a smart meter node; Characterizing the first The characteristic standard deviation vector of a smart meter node Step S3.5: Combine the feature expectation vector and the feature standard deviation vector to generate a multi-dimensional node feature descriptor that aggregates quantitative uncertainty information. The mapping relationship is as follows: 。 9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it can implement the various steps of the low-voltage distribution network high-resistance fault diagnosis method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it can implement the various steps of the low-voltage distribution network high-resistance fault diagnosis method as described in any one of claims 1-6.

Images