Power distribution network high resistance fault identification method and system based on physical feature reconstruction
By constructing a physical knowledge graph and extracting features using wavelet convolutional layers, combined with graph convolutional networks and dual-view fusion, the problem of incomplete data in high-resistivity fault diagnosis of power distribution networks is solved, achieving highly reliable fault identification and improving diagnostic accuracy under extreme operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HUNAN ELECTRIC POWER CO LTD ELECTRIC POWER SCI RES INST
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-10
AI Technical Summary
Existing high-resistivity fault diagnosis technologies for distribution networks suffer from issues such as feature distortion, false artifacts in generative models, and lack of physical interpretability when faced with incomplete data, leading to misjudgments and failure to operate.
By constructing a physical knowledge graph, extracting time-frequency features using shared continuous wavelet convolutional layers, combining graph convolutional networks for feature inference and reconstruction, introducing a physical consistency loss function, performing dual-view feature fusion, and employing a knowledge distillation training strategy to generate highly recognizable fault feature vectors.
When sensors fail or data transmission is interrupted, high-resistance fault identification with high reliability is achieved, improving recall and operational safety, and increasing accuracy by 19.4%.
Smart Images
Figure CN122365053A_ABST
Abstract
Description
Technical Field
[0001] This invention mainly relates to the field of power distribution network technology, specifically to a method and system for identifying high-resistance faults in power distribution networks based on physical characteristics. Background Technology
[0002] As the final link in the power system, the distribution network operates in a complex environment and is prone to frequent faults. High-impedance faults (HIFs), in particular, are typically caused by broken conductors coming into contact with high-impedance media such as trees, gravel surfaces, or asphalt roads. Unlike metallic short circuits, HIFs generate extremely weak fault currents (usually less than 10A), and their transient characteristics are easily masked by normal load fluctuations or background noise. If such faults are not cleared promptly and accurately, they can lead to continuous arcing to ground, easily igniting surrounding dry branches and leaves, causing forest fires, and may also pose a fatal electric shock threat to pedestrians due to live conductors falling to the ground.
[0003] Existing detection technologies mainly rely on the integrity and synchronization of multi-source monitoring data (such as three-phase voltage, current, and zero-sequence components). However, in the complex actual power grid operation environment, due to limitations such as sensor aging, communication link packet loss, electromagnetic interference, or measurement equipment failure, the data received by the control center often faces the severe challenge of "sensor loss" (i.e., some channel data is lost or invalid).
[0004] Existing high-resistance fault diagnosis technologies for power distribution networks suffer from the following three main insurmountable drawbacks when dealing with incomplete data: 1. Feature distortion in traditional statistical imputation methods: Existing zero-filling or mean-filling methods merely handle missing values from a statistical perspective. For high-impedance faults, these methods directly smooth out their unique weak high-frequency transient characteristics (such as zero-filling, waveform distortion, and random spikes). For example, mean-filling replaces fluctuating fault currents with flat lines, resulting in the complete loss of frequency domain information. The feature extraction layer cannot capture effective fault fingerprints, thus triggering protection failure.
[0005] 2. Risk of “physical illusion” in generative models: While data restoration methods based on GANs (Generative Adversarial Networks) or VAEs (Variational Autoencoders) can generate realistic waveforms, the generated signals often contain "artifacts" that violate physical laws due to a lack of understanding of the physical topology of the power grid (such as line impedance, phase relationships, and Kirchhoff's current law). For example, the generated current waveform may be out of phase with the voltage or violate the law of conservation of node current. These non-physical noises can be misinterpreted as fault features by subsequent diagnostic models, leading to system misjudgments.
[0006] 3. Lack of physical interpretability and robustness: Existing fusion diagnostic methods are often purely data-driven "black box" models that ignore the physical interpretability of features. When data is missing, the model cannot effectively deduce from the "mechanism" level (such as the inevitable downstream response caused by a sudden change in current at an upstream node). In addition, existing multi-source fusion methods are mostly simple feature splicing, lacking in-depth exploration of the complementarity and redundancy between different sensors, resulting in a precipitous drop in overall diagnostic performance when some sensors fail. Summary of the Invention
[0007] To address the technical problems existing in the prior art, this invention provides a highly reliable diagnostic method and system for identifying high-resistance faults in distribution networks based on physical feature reconfiguration.
[0008] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for identifying high-resistivity faults in a distribution network based on physical characteristics includes the following steps: S1: Constructing a physical knowledge graph; S2: Convolve the original electrical quantity data using a shared continuous wavelet convolutional layer to extract time-frequency feature maps; S3: Input the time-frequency feature map and physical knowledge graph into the graph convolutional network, use the complete features of physical topological neighbors to infer and reconstruct the features of missing data nodes, and constrain the reconstructed features to conform to the laws of electrical physics through the physical consistency loss function to generate a complete feature matrix. S4: Input the complete feature matrix into the dual-view fusion module, extract features from the spatial domain and the temporal domain respectively and fuse them to generate a fault feature vector; S5: A dual-path knowledge distillation training architecture is adopted, which uses the distribution of fault feature vectors under complete data paths as soft labels to constrain the model learning of incomplete data paths, so as to optimize the parameters of graph convolutional network and dual-view fusion module. S6: Input the fault feature vector into the classifier and output the fault type identification result.
[0009] Preferably, in step S1, the physical knowledge graph Specifically:
[0010] Among them, node definition Mapping measurement units in the distribution network as graph nodes; Edge definition Establish edges based on the actual electrical connection relationship; if there is a direct connection between two monitoring points via a power transmission line, then establish a connection edge. Physical weighted adjacency matrix Construct a weighted matrix to characterize the physical propagation attenuation of fault traveling waves; edge weights. It not only represents the connection, but also initializes it according to the line length and impedance, thereby simulating the attenuation characteristics of the signal during transmission.
[0011] Preferably, each node This includes a multidimensional electrical quantity sequence collected at this location, including three-phase current. Zero-sequence current and three-phase voltage .
[0012] Preferably, the wavelet basis functions in step S2 that share the continuous wavelet convolutional layer for:
[0013] in, This is the amplitude normalization coefficient; It is the attenuation factor; It is the angular frequency of the oscillation; It is a unit step function.
[0014] Preferably, the physical consistency loss function in step S3 for:
[0015] in, The total number of nodes; The set of edges in the physical graph; and They are nodes and nodes The reconstructed feature vector; The weight is the physical connection weight between the two nodes.
[0016] Preferably, in step S4, in the spatial domain, the confidence threshold value for each channel is calculated. :
[0017] in, For the first Confidence threshold values for each channel; Use the Sigmoid activation function; and These represent the weights and biases of the gated network, respectively. For the first Input characteristics of each channel; Confidence threshold Element-wise multiplication with the features of the corresponding channels yields the spatially fused feature vector. :
[0018] Here, ⊙ represents element-wise multiplication.
[0019] Preferably, in step S4, in the time domain, a multi-head attention mechanism is introduced to capture the long-distance dependency relationship spanning multiple power frequency cycles, targeting the timing characteristics of the intermittent re-ignition of the high-resistivity fault arc.
[0020] Preferably, in step S4, the features in the spatial domain and the time domain are weighted and integrated, and residual connections are introduced to output the final high-discrimination fault feature vector.
[0021] Preferably, in step S5, during the training phase, the complete samples are randomly masked to form pairs of complete-incomplete input training sets; and KL divergence is introduced as distillation loss to force the feature distribution of the intermediate layer under the incomplete path to approximate the feature distribution under the complete path.
[0022] The present invention also discloses a high-resistance fault identification system for power distribution network based on physical characteristics, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when the processor runs the computer program.
[0023] Compared with the prior art, the advantages of the present invention are as follows: The method of this invention aims to solve the problem of false operation or failure to operate caused by feature extraction failure or abnormal input dimension when existing fault diagnosis technologies rely on complete data and face sensor failure or data transmission interruption. Its core significance lies in the creative proposal of a diagnostic paradigm of "physical feature collaborative reconstruction": using a complete physical topology knowledge graph of the distribution network as a priori constraint, and using electrical correlation information of channels without missing channels (such as current coupling under Kirchhoff's laws) to intelligently infer and fill in the missing key fault features in the feature domain, so as to achieve a highly reliable diagnosis that is "not missing features even if data is missing".
[0024] This invention overcomes feature distortion caused by simple interpolation and noise interference introduced by generative models by constructing a collaborative mechanism of physical wavelet feature extraction, feature inference based on topological graphs, and dual-view spatiotemporal fusion. It is of great significance for ensuring the sensing capability of distribution networks under extreme communication conditions, improving the recall rate of high-resistance faults, and enhancing operation and maintenance safety. Attached Figure Description
[0025] Figure 1 This is a flowchart of the high-resistivity fault identification method for power distribution network based on physical feature reconstruction according to an embodiment of the present invention. Detailed Implementation
[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0027] like Figure 1 As shown in the figure, the high-resistivity fault identification method for distribution networks based on physical feature reconstruction provided by this invention includes the following steps: S1: Distribution Network Physical Topology Modeling and Graph Construction First, the electrical quantities of the distribution network are acquired, and physical topology modeling and graph construction are performed. The purpose is to endow subsequent deep learning models with the physical cognitive ability to understand the power grid structure. In this step, to give the algorithm the ability to "understand the power grid structure," a physical knowledge graph is first constructed based on the single-line diagram of the distribution network. ; Among them, node definition ( ): Mapping measurement units (such as FTUs, smart meters, and substation outgoing lines) in the distribution network as graph nodes; each node Includes a multidimensional electrical quantity sequence (three-phase current) collected at this location. Zero-sequence current Three-phase voltage ); Edge definition ( ): Establish edges based on the actual electrical connection relationship; if there is a direct connection between two monitoring points via a transmission line, then establish a connection edge; Physical weighted adjacency matrix ( ): Construct a weighted matrix to characterize the physical propagation attenuation law of fault traveling waves; edge weights It not only represents the connection, but also initializes based on the line length and impedance to simulate the attenuation characteristics of the signal during transmission; the calculation formula is as follows:
[0028] in For nodes and Line impedance between is the attenuation constant.
[0029] This physics knowledge map This will serve as a guide for subsequent feature inference in graph neural networks.
[0030] S2: Shared Physical Wavelet Time-Frequency Feature Extraction (SCWConv) To address the non-stationary and nonlinear characteristics of high-impedance fault signals, a Shared Continuous Wavelet Conv (SCWConv) layer is designed to replace the traditional CNN, extracting time-frequency features with clear physical meaning and high sensitivity to high-impedance faults from the original waveform.
[0031] Physical wavelet kernel construction: A shared continuous wavelet convolutional layer is designed, which uses the Laplace wavelet as the basis function for the convolution kernel. Its time-domain waveform of unilateral exponentially decaying oscillations closely matches the transient impact response (Arcing Impulse) of a high-resistivity fault arc. Its mathematical expression is:
[0032] in, For a moment The wavelet basis function values; This is the amplitude normalization coefficient; As an attenuation factor, it determines the rate at which the impact signal fades away; This is the oscillation angular frequency, corresponding to the dominant frequency of the fault signal; It is a unit step function.
[0033] Multi-channel shared convolution: using the same set of wavelet kernel parameters for data from all sensor channels. Perform convolution operations to generate physical feature maps containing rich time-frequency information. Specifically:
[0034] in As a scale factor, This is the translation factor.
[0035] This step ensures that the extracted features are physically interpretable and can keenly capture the faint high-frequency components of the electric arc.
[0036] S3: Reconstruction based on missing features of topological neighbors When partial node data is detected as missing (due to the mask matrix) When the label is applied, the graph convolution-based feature inference module is launched. The input to this step is the feature map output from step S2 and the physical knowledge graph constructed in step S1. The goal is to intelligently infer and complete the features of missing nodes using complete topological association information.
[0037] Missing data detection: First, generate a binary mask vector. In this context, "1" represents valid data and "0" represents missing data.
[0038] Graph Convolutional Feature Inference: This utilizes Graph Convolutional Networks (GCNs) to propagate information across physical graphs. Even the central node... Even with missing features, GCN can still aggregate features from its first- and second-order "physical neighbors" (NeighborNodes) to infer nodes. Status:
[0039] in For the first The feature matrix of the layer, To incorporate the physical adjacency matrix of self-loops, These are learnable weights.
[0040] The underlying physical principle is Kirchhoff's Current Law (KCL), which states that the current flowing into a node equals the sum of the currents flowing out. Current fluctuations in upstream nodes will inevitably produce a response in downstream nodes. The algorithm utilizes this deterministic electrical coupling relationship to "infer" the characteristics of missing nodes from complete neighboring nodes.
[0041] Physical consistency constraints To prevent the reconstructed features from violating physical laws, a physical consistency loss function based on graph Laplacian regularization is introduced:
[0042] in, The total number of nodes; The set of edges in the physical graph; and They are nodes and nodes The reconstructed feature vector; The weight is the physical connection weight between the two nodes.
[0043] This formula constrains physically connected nodes, requiring their characteristic changes to conform to the smoothness of the power grid topology, penalizing non-physical abrupt changes, and ensuring that the reconstructed waveform conforms to the power grid topology constraints.
[0044] S4: Dual-view spatiotemporal feature fusion After obtaining the reconstructed complete features, a dual-view spatiotemporal feature fusion is performed. A dual-view fusion module is designed to extract discriminative features from the spatial and temporal domains respectively, addressing potential errors in the reconstructed features. The input to this step is the complete features output from step S3, and the goal is to extract discriminative features from the spatial and temporal domains respectively, and finally fuse them into a highly recognizable fault feature vector.
[0045] Completeness Gating (Viewpoint 1: Spatial Domain): A gating mechanism is used to calculate weights based on the consistency of features across channels. For channels with missing data, their weight in the final decision is adaptively reduced to suppress potential reconstruction errors.
[0046] Specifically, the confidence threshold value for each channel is calculated. , confidence level gate value Element-wise multiplication with the features of the corresponding channels yields the spatially fused feature vector. .
[0047] The corresponding calculation formula is:
[0048] in, For the first Confidence threshold values for each channel; Use the Sigmoid activation function; and These represent the weights and biases of the gated network, respectively. For the first The input features of each channel; ⊙ represents element-wise multiplication; This is the feature vector after spatial domain fusion.
[0049] Multi-Head Attention Fusion (Viewpoint 2: Time Domain): Addressing the temporal characteristics of the "intermittent reignition" of high-resistivity fault arcs, a multi-head attention mechanism is introduced to capture long-distance dependencies spanning multiple power frequency cycles. This helps to identify periodic electric arc impact characteristics and overcome the problem of limited receptive field in convolutional layers.
[0050] Feature ensemble: The fusion results of the two views are weighted and integrated, and a residual connection is introduced to output the final high-discrimination fault feature vector.
[0051] S5: Dual-path training based on knowledge distillation To enable the above model to learn "how to extrapolate", a dual-path training architecture is adopted.
[0052] Random masking strategy: During the training phase, complete samples are randomly masked to artificially simulate various sensor failure modes, forming a pair of "complete input-incomplete input" training sets.
[0053] Knowledge distillation constraint: Kullback-Leibler divergence is introduced as a distillation loss to force the feature distribution of intermediate layers under incomplete paths to approximate the feature distribution under complete paths. This ensures that the model can effectively "transfer" the physical laws of faults learned under complete data to the working conditions with missing data.
[0054]
[0055] in, This is due to distillation losses; The feature probability distribution under complete data input; This refers to the feature probability distribution under incomplete data input.
[0056] Joint optimization: End-to-end training using a joint loss function:
[0057] in Cross-entropy classification loss, The reconstruction error (MSE) is the error that is created in the first place. For physical consistency loss, This represents the balance coefficient for each loss term.
[0058] S6: Fault Identification and Decision Making The fused high-dimensional feature vector is input into a fully connected layer classifier (Softmax), which outputs a fault probability distribution (e.g., high-resistance grounding, metallic grounding, normal disturbance, etc.). If a high-resistance grounding fault is identified, an alarm signal is triggered and the faulty line number is output.
[0059] This invention presents a method for identifying high-resistivity faults in distribution networks based on incomplete data and driven feature reconstruction using physical knowledge. This method addresses the technical challenge of traditional fault detection techniques, which heavily rely on data completeness and are prone to missed or false positives when sensors fail or communication packets are lost. The core of this method lies in introducing a physical topological knowledge graph as a priori constraint and intelligently inferring missing key features using topological association information from channels that are not missing. Specifically, it extracts fault features with clear mechanistic meanings through physical wavelet transform and constructs a feature reconstruction module based on a graph convolutional network (GCN). Combined with a dual-view spatiotemporal fusion and knowledge distillation training strategy, it overcomes the feature ambiguity of single interpolation methods and the noise interference of generative models.
[0060] Simulation results show that even under extreme conditions where the sensor data loss rate is as high as 50%, this method can still maintain a recognition accuracy of 92.4%, which is 19.4% higher than traditional pure data-driven methods (such as CNN). This effectively improves the robustness and accuracy of high-resistance grounding fault diagnosis under extreme conditions of sensor loss.
[0061] The method of this invention aims to solve the problem of false operation or failure to operate caused by feature extraction failure or abnormal input dimension when existing fault diagnosis technologies rely on complete data and face sensor failure or data transmission interruption. Its core significance lies in the creative proposal of a diagnostic paradigm of "physical feature collaborative reconstruction": using a complete physical topology knowledge graph of the distribution network as a priori constraint, and using electrical correlation information of channels without missing channels (such as current coupling under Kirchhoff's laws) to intelligently infer and fill in the missing key fault features in the feature domain, so as to achieve a highly reliable diagnosis that is "not missing features even if data is missing".
[0062] This invention overcomes feature distortion caused by simple interpolation and noise interference introduced by generative models by constructing a collaborative mechanism of physical wavelet feature extraction, feature inference based on topological graphs, and dual-view spatiotemporal fusion. It is of great significance for ensuring the sensing capability of distribution networks under extreme communication conditions, improving the recall rate of high-resistance faults, and enhancing operation and maintenance safety.
[0063] This invention also discloses a high-resistance fault identification system for power distribution networks based on physical feature reconstruction, comprising an interconnected memory and a processor. The memory stores a computer program, which, when run by the processor, executes the steps of the method described above. The system of this invention corresponds to the method described above and also possesses the advantages described therein.
[0064] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.
[0065] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A method for identifying high-resistivity faults in a distribution network based on physical characteristics, characterized in that, Includes the following steps: S1: Constructing a physical knowledge graph; S2: Convolve the original electrical quantity data using a shared continuous wavelet convolutional layer to extract time-frequency feature maps; S3: Input the time-frequency feature map and physical knowledge graph into the graph convolutional network, use the complete features of physical topological neighbors to infer and reconstruct the features of missing data nodes, and constrain the reconstructed features to conform to the laws of electrical physics through the physical consistency loss function to generate a complete feature matrix. S4: Input the complete feature matrix into the dual-view fusion module, extract features from the spatial domain and the temporal domain respectively and fuse them to generate a fault feature vector; S5: A dual-path knowledge distillation training architecture is adopted, which uses the distribution of fault feature vectors under complete data paths as soft labels to constrain the model learning of incomplete data paths, so as to optimize the parameters of graph convolutional network and dual-view fusion module. S6: Input the fault feature vector into the classifier and output the fault type identification result.
2. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 1, characterized in that, In step S1, the physical knowledge graph Specifically: Among them, node definition Mapping measurement units in the distribution network as graph nodes; Edge definition Establish edges based on the actual electrical connection relationship; if there is a direct connection between two monitoring points via a power transmission line, then establish a connection edge. Physical weighted adjacency matrix Construct a weighted matrix to characterize the physical propagation attenuation of fault traveling waves; edge weights. It not only represents the connection, but also initializes it according to the line length and impedance, thereby simulating the attenuation characteristics of the signal during transmission.
3. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 2, characterized in that, Each node This includes a multidimensional electrical quantity sequence collected at this location, including three-phase current. Zero-sequence current and three-phase voltage .
4. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 1, 2, or 3, characterized in that, Wavelet basis functions shared in step S2 of continuous wavelet convolutional layers for: in, This is the amplitude normalization coefficient; It is the attenuation factor; It is the angular frequency of the oscillation; It is a unit step function.
5. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 1, 2, or 3, characterized in that, Physical consistency loss function in step S3 for: in, The total number of nodes; The set of edges in the physical graph; and They are nodes and nodes The reconstructed feature vector; The weight is the physical connection weight between the two nodes.
6. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 1, 2, or 3, characterized in that, In step S4, in the spatial domain, the confidence threshold value for each channel is calculated. : in, For the first Confidence threshold values for each channel; Use the Sigmoid activation function; and These represent the weights and biases of the gated network, respectively. For the first Input characteristics of each channel; Confidence threshold Element-wise multiplication with the features of the corresponding channels yields the spatially fused feature vector. : Here, ⊙ represents element-wise multiplication.
7. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 6, characterized in that, In step S4, in the time domain, a multi-head attention mechanism is introduced to capture the long-distance dependency relationship spanning multiple power frequency cycles, targeting the timing characteristics of intermittent re-ignition of high-resistivity fault arcs.
8. The method for identifying high-resistance faults in distribution networks based on physical feature reconstruction according to claim 7, characterized in that, In step S4, the features in the spatial domain and the time domain are weighted and integrated, and residual connections are introduced to output the final high-discrimination fault feature vector.
9. The method for identifying high-resistivity faults in distribution networks based on physical feature reconstruction according to claim 1, 2, or 3, characterized in that, In step S5, during the training phase, the complete samples are randomly masked to form pairs of complete-incomplete input training sets. Furthermore, KL divergence is introduced as a distillation loss to force the intermediate layer feature distribution under incomplete paths to approximate the feature distribution under complete paths.
10. A high-resistance fault identification system for a distribution network based on physical characteristics, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that... The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-9.