Active distribution network fault section location method based on topology coding and graph propagation decoding

Abstract

The active distribution network fault section positioning method based on topology coding and graph propagation decoding strictly aligns the physical propagation path of the injected signal with the decoding message transmission path through the topology adjacency matrix, and eliminates the information loss caused by the misalignment of the path. The steps include: taking the bus as the root, performing breadth-first search on the topology adjacency matrix to obtain the topology distance of each grid-connected inverter, assigning differentiated characteristic frequencies according to the increasing distance to form topology coding, and controlling the inverter to inject corresponding signals; each feeder terminal unit collects responses and separates interference, constructs a topology propagation judgment matrix and performs physical consistency check; input the effective judgment matrix and the topology adjacency matrix into the graph propagation decoding model, and the message is transmitted along the topology edge and then the fault probability of each branch is output. The application utilizes the frequency correlation characteristics of line impedance and the uniqueness of topology path to realize path distinguishability, and significantly improves the accuracy and distinguishability of fault section positioning in the scene of high transition resistance and high penetration of distributed photovoltaic.

Description

Technical Field

[0001] This invention relates to the field of active distribution network fault location technology, specifically to an active distribution network fault section location method based on topology coding and graph propagation decoding. Background Technology

[0002] In current technologies, the penetration rate of distributed photovoltaic (PV) power in distribution networks continues to increase, and traditional passive distribution networks are rapidly evolving into active distribution networks with bidirectional power flow. In power systems, over 85% of faults occur in distribution networks, with single-phase grounding faults accounting for over 80%. Quickly and accurately locating faulty sections is directly related to power supply reliability and the capacity for renewable energy absorption.

[0003] At present, the fault location technology of distribution network is mainly divided into two categories: passive and active. The passive method relies on the electrical characteristic quantity generated by the fault itself for discrimination. Related studies have shown that the location method based on transient zero-sequence current similarity recorded in reference [1] or transient reactive power direction recorded in reference [2] performs well under low and medium resistance grounding conditions. However, under the neutral point non-effective grounding mode of distribution network, when a high transition resistance grounding fault occurs, the fault current amplitude may be as low as less than 1% of the rated current of the system. Especially in the case where the transition resistance exceeds 1500Ω, the fault characteristics are extremely weak and difficult to be effectively distinguished from the background noise, which leads to a significant deterioration in the location accuracy of such methods. Reference [3] points out that the location success rate of the method based on phase current mutation and zero-sequence power direction has dropped to less than 70% when the transition resistance exceeds 1000Ω. The root cause is that the performance of the passive method is limited by the physical conditions of the signal source. Simply improving the algorithm level is not enough to fundamentally break through this bottleneck.

[0004] To enhance fault characteristics, an active signal injection method has been proposed, which achieves location by injecting characteristic signals into the distribution network. Reference [4] proposes a single-phase ground fault location scheme based on the S-signal injection method, which uses the injected AC signal for path detection. However, existing active injection schemes still have some shortcomings, which restrict the further improvement of location performance: First, the injection frequency is usually a fixed value, without considering the frequency correlation characteristics of the line impedance. Reference [5] proposes to inject a specific frequency current at the neutral point for fault selection and location, but the injection frequency used is a single fixed value. When signals propagating on different topological paths reach the common measurement point, since the frequencies are completely consistent, it is difficult to effectively distinguish the signal source path based solely on amplitude and phase information, resulting in insufficient distinguishability of path identification. Reference [6] also points out that the fixed frequency injection method faces the problem of path aliasing in multi-path power grids, and the response modes of different paths are highly similar.

[0005] Secondly, the injection strategy and the localization algorithm are loosely coupled. The injection signal is physically propagated along the actual topological edges of the power distribution network, but the information aggregation path of the localization algorithm is not along these same topological edges. In the existing literature, the active injection localization method described in reference [4] mostly uses threshold comparison, the fully connected neural network described in reference [7], or the convolutional neural network described in reference [8] for feature discrimination at the decoding end.

[0006] In the above methods, the threshold comparison only uses single-node information and completely ignores the topological adjacency structure of the distribution network; the fully connected network flattens the features of each measurement point into a one-dimensional vector for processing, and the topological adjacency relationship between nodes is discarded in this process; the convolutional neural network relies on a regular grid structure to define the sliding path of the convolution kernel, which is difficult to adapt to the tree-like connection characteristics of the non-Euclidean topology of the distribution network. Reference [9] clearly pointed out in the analysis of the deep learning-based distribution network fault location method that ignoring topological constraints will lead to serious performance degradation of the model when facing topological changes. This inconsistency between the physical propagation path and the computational aggregation path in the topological structure leads to the spatial propagation features obtained by the physical layer being discarded or distorted in the computational layer, resulting in the loss of key discrimination information.

[0007] Third, the existing schemes do not establish a structured association between the injection coding rules and the localization model, which makes the injection end and the decoding end designed independently and lacks collaborative optimization. Reference

[10] reviews the distribution network fault location technology based on active injection, and points out that the current research mainly focuses on the waveform design of the injected signal and the signal detection algorithm, but there is a lack of joint design between the injection strategy and the localization identification model at the topology level. When the distribution network topology or line parameters change, after the injection end parameters are adjusted, the localization model needs to be retrained or the threshold needs to be manually adjusted, and the two ends cannot be linked to adapt.

[0008] Fourth, existing solutions often fail to adequately address engineering constraints such as limited control authority, communication delays, and heterogeneous equipment. Reference

[11] points out that distributed inverters in actual distribution networks suffer from diverse manufacturer protocols, heterogeneous communication methods, and inconsistent scheduling authority, making it difficult to ensure that all inverters in the network have remote control capabilities. Reference

[12] also points out in its analysis of the current status of distribution IoT terminal access that existing FTU devices have significant differences in sampling rate, analog-to-digital converter bit width, and communication protocol. If the active injection scheme does not consider the compatibility issues of such devices, it will be difficult to deploy on a large scale in actual engineering projects. When the distribution network topology changes due to switching operations or other reasons, existing injection strategies and positioning models both need to be manually reconfigured, lacking the ability to adapt to topology changes.

[0009] In summary, the fundamental flaw of existing active injection positioning schemes lies in the topological misalignment between the physical propagation path of the injected signal and the information aggregation path of the positioning algorithm. The signal propagates along the topological edge of the power grid, but the positioning algorithm does not aggregate information along the same topological edge. How to construct a positioning system that strictly aligns the signal propagation path with the computational aggregation path in the topological structure, thereby fundamentally eliminating the information loss caused by path misalignment, is a pressing technical problem that needs to be solved in this field.

[0010] References: Reference [1]: Xue Yongduan, Li Juan, Xu Bingyin. A method for detecting small current grounding faults using transient zero-sequence current characteristics [J]. Automation of Electric Power Systems, 2017. Reference [2]: Zhang Baohui, Yin Xianggen. Power System Relay Protection [M]. China Electric Power Press, 2019. Reference [3]: Li Yongli, Song Jinzhao, Li Botong. A review of single-phase grounding fault location methods in distribution networks with distributed power sources [J]. Automation of Electric Power Systems, 2020. Reference [4]: ​​Liu Jian, Zhang Zhihua, Zhang Xiaoqing. Discussion on several issues of distribution network fault handling [J]. Automation of Electric Power Systems, 2012. Reference [5]: Guo Junhong, Chen Gang, Wang Baohua. Research on single-phase grounding fault selection in resonant grounding system based on signal injection method [J]. Power System Protection and Control, 2015. Reference [6]: Cai Xinjing, Zhang Fan. A review of fault location methods for resonant grounding systems in distribution networks [J]. Journal of Electric Power Systems and Automation, 2018. Reference [7]: Qi Zhong, Shi Ye, Li Feng. Intelligent fault diagnosis method for distribution network based on deep learning [J]. Electric Power Engineering Technology, 2021. Reference [8]: Yang Sha, Gao Zhanjun. A method for locating fault sections in power distribution networks based on convolutional neural networks [J]. Power System Protection and Control, 2020. Reference [9]: Chen Ying, Guan Huizhe, Huang Shaowei. Transient stability assessment of power system based on graph deep learning [J]. Proceedings of the CSEE, 2019. Reference

[10] : Jia Ke, Shi Zhiming, Bi Tianshu. A review of single-phase grounding fault location technology in distribution network based on active injection [J]. Proceedings of the CSEE, 2018. Reference

[11] : Liu Dong, Sheng Wanxing, Wang Yun. Practical Technology of Distribution Automation System [M]. China Electric Power Press, 2019. Reference

[12] : Sheng Wanxing, Liu Keyan, Meng Xiaoli. Terminal access technology and engineering practice of distribution Internet of Things [J]. Automation of Electric Power Systems, 2020. Summary of the Invention To address the aforementioned technical problems, this invention provides an active distribution network fault location method based on topology coding and graph propagation decoding. This method utilizes a tightly coupled design of topology coding and graph propagation decoding, ensuring strict alignment between the physical propagation path of the injected signal and the information aggregation path of the location algorithm on the distribution network topology adjacency matrix. This overcomes the information loss caused by path misalignment in traditional methods. This method significantly improves the accuracy and distinguishability of fault location in scenarios with high transition resistance and high penetration of distributed photovoltaic power.

[0011] The technical solution adopted in this invention is as follows: The active distribution network fault location method based on topology coding and graph propagation decoding includes the following steps: Step 1: Fault Startup and Data Acquisition Real-time monitoring of zero-sequence voltage at each node of the distribution network; when the zero-sequence voltage at any node exceeds the start-up threshold, a ground fault is determined and the location process is initiated; the transition resistance range is estimated simultaneously and a waveform recording call command is sent to each feeder terminal unit (FTU). Step 2: Topological Encoding and Feature Injection: The master station obtains the topology adjacency matrix of the target distribution network. Using the bus node as the root in the topological adjacency matrix The breadth-first search is performed to obtain the topological distance from each grid-connected inverter to the bus. Increasing characteristic frequencies are assigned to each grid-connected inverter according to the topological distance from near to far to form topological coding rules. After a fault, each grid-connected inverter is controlled to inject the corresponding characteristic frequency injection signal according to the topological coding rules. Step 3: Discriminant matrix construction and physical consistency verification: Each feeder terminal unit (FTU) acquires transient waveforms and separates the interference sources at fault points. The net injection response generated solely by active injection is obtained through vector difference calculation. The amplitude and phase of the net injection response at each characteristic frequency are extracted. A matrix is ​​constructed with characteristic frequencies as rows and feeder terminal units (FTUs) as columns, arranged in order with the topological adjacency matrix. The topology propagation discriminant matrix with consistent node order ; Topological propagation discriminant matrix Execute based on topological adjacency matrix The physical consistency check will be performed by using the topology propagation discrimination matrix of the check. Marked as a valid discrimination matrix; Step 4: Graph propagation decoding and localization: The effective discriminant matrix and the topological adjacency matrix The input graph propagation decoding model is used together with the topological adjacency matrix. Define message passing paths; messages are only sent through the topological adjacency matrix. The fault probability of each branch of the distribution network is output after passing between adjacent nodes, node feature aggregation and edge embedding, and the branch with the highest fault probability and exceeding the preset confidence threshold is selected as the fault section location result.

[0012] In step 2, the topological adjacency matrix satisfy If and only if node i and j There are line connections between them, that is, A is a symmetric matrix, which corresponds to the bidirectional signal propagation characteristics of the distribution network lines.

[0013] Using the bus node as the root in the topological adjacency matrix The topological distance from each grid-connected inverter to the bus is obtained by performing a breadth-first search. The specific process is as follows: Set bus node Define topological distance Establish a queue ; If Q is not empty, pop the head of the queue u and iterate through all queues that satisfy the condition. That is, if there is a line connection between nodes u and v, and a neighbor v that has not been visited is enqueued and recorded. ,in: Represents a node Topological distance to the bus node; After the search is completed, the topological distance of each grid-connected inverter k is... .

[0014] In step 2, the incremental characteristic frequencies are allocated to each grid-connected inverter according to the topological distance from near to far, specifically including: S2.1: Set of available feature frequencies Sort by numerical value in ascending order, and then sort the grid-connected inverters by topological distance. After sorting in ascending order, the available feature frequency sets are allocated sequentially. The corresponding frequencies are used to form an inverter-frequency mapping table. ; S2.2: Path Separability Verification: For each pair of inverters , Traverse all FTU nodes ,like and All injected signals can reach the same FTU node. Then check and At this node Does the set of injected feature frequencies at a given location have an intersection? If an intersection exists, prioritize moving the overlapping feature frequencies with the larger topological distance to the available feature frequency set. The next unused feature frequency in the middle ensures that the sets of injected feature frequencies at the FTUs that any two topological paths pass through are not exactly the same; S2.3: Propagation Coverage Verification: Inverter-Frequency Mapping Table For each assigned characteristic frequency, the signal attenuation to the farthest FTU is estimated based on the line length and unit impedance parameters. If the signal-to-noise ratio after attenuation is lower than the preset detection threshold, the characteristic frequency is removed and adjusted to the next available characteristic frequency for the corresponding grid-connected inverter. S2.4: Spectrum Gap Compliance Check: Verify Inverter-Frequency Mapping Table If all characteristic frequencies fall within the spectral gap between integer harmonics of the power frequency and the difference between any two characteristic frequencies is not less than the preset frequency interval, then adjust to the nearest compliant frequency point.

[0015] In step 2, after a fault occurs, each grid-connected inverter is controlled to inject an injection signal at the corresponding characteristic frequency according to the topology coding rules; specifically: The main site uses the mapping table To the k The inverter issues a command carrying a unified time tag. The injected signal adopts the form of a negative sequence component, and its time-domain expression is as follows: ; in: The instantaneous value of the negative sequence signal injected into the k-th inverter; The injection amplitude is the same as the current amplitude of the grid-connected inverter. The voltage amplitude of the grid-connected inverter shall not exceed 2% of the rated current. The voltage shall not exceed 1% of the rated voltage; The characteristic frequency allocated to the k-th inverter satisfies ; This is the initial phase for synchronization.

[0016] In step 3, the interference source at the fault point is separated, and the net injection response generated solely by active injection is obtained through vector difference calculation. Specifically, this includes: S3.1: Before feature injection is initiated, capture the waveform for at least one power frequency cycle and extract the background negative sequence phasor at each feature frequency point. During feature injection, the negative-order phasors corresponding to the feature frequencies are extracted simultaneously. ; S3.2: Calculate the elements of the net injection response matrix generated solely by active injection using the following formula: ; in: During feature injection, FTU nodes j At characteristic frequency Negative sequence voltage phasor (complex number) at the location; Before feature injection is initiated, the FTU node j At characteristic frequency Background negative sequence voltage phasor (complex number) at the location. The net injected response phasor, whose magnitude and argument are the th element in the topological propagation discriminant matrix M. i Line 1 j The element values ​​of the column; F is the total number of characteristic frequencies; N is the total number of FTU nodes.

[0017] Topological propagation discriminant matrix That is to The amplitude and phase are used as element values.

[0018] In step 3, the net injection response phasors obtained in step S3.2 are... The characteristic frequencies are arranged in rows, FTU nodes in columns, and the column order is aligned with the topological adjacency matrix. The node indices are arranged in a strictly consistent manner to form a complex matrix. : ; Where the i-th row corresponds to the characteristic frequency The j-th column corresponds to FTU node j. element ; F represents the total number of characteristic frequencies, and N represents the total number of FTUs. To ensure the effectiveness of phase difference extraction, all FTU column data must be collected within the same BeiDou / GPS synchronization time window.

[0019] In step 3, the physical consistency verification includes a trend verification, which includes: a1: During normal system operation, the impedance parameters of each line segment at the characteristic frequency are obtained based on online parameter identification, and the impedance parameters of each adjacent node are pre-calculated. For characteristic frequencies under fault-free operating conditions Expected decay ratio The details are as follows: ; in: Representing the characteristic frequencies of nodes i and j respectively. Voltage phasors under these conditions; This is the voltage amplitude attenuation ratio estimated based on the impedance model, i.e., the desired attenuation ratio. ; For line segment series impedance, For line resistance, For line inductance; The equivalent impedance seen from node j into the network except in the direction of i, including load, distributed capacitance and downstream network, is uniformly provided by online parameter identification.

[0020] The mean value is calculated based on the historical sequence of expected decay ratios identified during continuous T days (e.g., 30 days) of normal operation. and standard deviation The tolerance range is determined by the following formula: [ ; in: Characteristic frequency The tolerance range for the lower expected attenuation ratio. The lower limit of the interval, The upper limit of the interval; The confidence coefficient is 3 for overhead lines and 2.5 for cables, corresponding to a confidence level of approximately 95% to 99.7%.

[0021] The above expected attenuation ratio The definition is based on signals from nodes To the node Dissemination, that is Comparison Closer to the injection source. Online parameter identification provides... For the node To remove The equivalent impedance seen from the rest of the network outside the specified direction. During verification, the positive direction of signal propagation needs to be determined based on the injection source location, and a reference attenuation ratio for that direction should be selected; if the signal flow on that side is... Then its expected attenuation ratio should be used .

[0022] a2: During fault location, for each controlled injection source, along the topological adjacency matrix Given a defined propagation path, calculate the measured attenuation ratio between adjacent nodes: ; in To at characteristic frequency The next signal comes from the node. i propagation to nodes j The measured voltage attenuation ratio; For matrix row k, row The magnitude of the column element corresponds to the characteristic frequency. At the node i The net injection response amplitude after compensation at the location.

[0023] For matrix row k, row The magnitude of the column element corresponds to the characteristic frequency. At the node j The net injection response amplitude after compensation at the location.

[0024] a3: Comparison and : when and If the deviation is within the tolerance range, it is determined to be normal attenuation and passes the verification. when If the deviation exceeds the lower limit of the tolerance range, it is marked as abnormal attenuation and prior weighting information is generated for use in step S4; when Furthermore, if the deviation exceeds the upper limit of the tolerance range or an amplitude increase occurs, it is marked as an abnormal enhancement; verification is deemed a failure and a backup positioning mode is triggered only when a step attenuation without physical basis occurs. For example, in a single node pair or between two adjacent measurement points, the measured attenuation ratio suddenly drops by more than 50% and cannot be explained by changes in switch status, sudden changes in line parameters, or adjustments to known operating modes.

[0025] In step 3, the physical consistency check includes a phase check, which includes: b1: For the topological adjacency matrix Each pair of adjacent nodes , ( ), characteristic frequency The measured phase shift is defined as follows: ; The results were normalized to .

[0026] in: Indicates frequency The measured voltage phase shift from node i to node j; Represents the principal argument of a complex number; and These are the compensated net injected response phasors for nodes j and i, respectively.

[0027] And based on online parameter identification provided and The reference phase shift is calculated using the following formula: ; in: The reference phase shift is obtained based on the impedance model; The series impedance of the line segment; This is the downstream equivalent impedance.

[0028] This formula shares the same set of equivalent impedance parameters with the desired attenuation ratio, ensuring physical consistency.

[0029] The tolerance range is determined by the following formula: ; in: Phase tolerance (degrees); As the baseline tolerance value, typically taken as... L represents the actual length of the line segment (km). The unit length is 1 km; For the accuracy coefficient of the current transformer: 0.5 for level 0.5 and 1.0 for level 1. The lower the accuracy, the larger the coefficient and the wider the tolerance range, which is in line with engineering practice.

[0030] b2: Construct differential phase features for characteristic frequencies , Calculate the difference in measured phase shifts on the same adjacent node pair (i,j): ; in: For frequency and The difference in measured phase shift on edge (i,j) is called the differential phase characteristic; , This represents two different injected feature frequencies; For frequency The measured phase shift below is defined the same as b1.

[0031] Although most of the synchronization deviation has been compensated for by global synchronization delay estimation before phasor construction, residual small synchronization errors may still introduce phase shift differences between adjacent node pairs that are approximately proportional to the frequency. Calculating the phase shift difference between two different characteristic frequencies of the node pair can further weaken this residual effect and improve the robustness of phase verification. If the measured phase shift or differential phase shift exceeds the corresponding tolerance range, the phase verification is deemed to have failed.

[0032] In step 4, the graph propagation decoding model uses a topological adjacency matrix. Define message passing paths; messages are only sent through the topological adjacency matrix. The transmission between adjacent nodes includes: 4.1: For FTU nodes After decomposing all the characteristic frequency responses of the corresponding column in matrix M, an initial eigenvector is constructed: ; in: and They are characteristic frequencies The net injected response amplitude and phase at node j; F represents the total number of characteristic frequencies; The dimension is 3F.

[0033] 4.2: The graph propagation decoding model adopts a graph network structure based on a message-passing mechanism. The message passing at the k-th layer is performed according to the following formula. OK: ; in, Let i be the feature vector of node i after aggregation and nonlinear transformation at the (k+1)th layer; It is a non-linear activation function; This indicates that the node itself is also included in the aggregation scope; Let A be the set of neighbors of node i. In this invention, the power distribution line is regarded as an undirected edge, A is a symmetric matrix, and messages can be transmitted bidirectionally. Let j be the feature representation of node j at the k-th layer; Let be the learnable weight matrix for the k-th layer.

[0034] 4.3: The number of model layers is set to the maximum topology depth from the furthest grid-connected inverter to the bus, so that the [number of layers]... Layer message passing corresponding injection signal from Information is physically propagated to neighboring nodes; 4.4: For the branch connecting nodes i and j ( The edge embedding vector is constructed as follows: ; in: The high-level features of nodes i and j after passing through an L-layer graph network; It is a feature difference vector; For the splicing operation, the embedded dimension after splicing is 192. Through classification layers (192→32→1) and Output .

[0035] The graph propagation decoding model adopts a multi-task learning architecture, including a branch classification main path and an auxiliary amplitude reconstruction path. The following total loss function is used for optimization during model training: ; in: For branch classification, the binary cross-entropy loss, For physical regularization loss term, These are the weighting coefficients; Physical regularization loss term Defined as: ; in: Let be the set of all upstream → downstream edges in the topology. And i is closer to the injection source; The frequencies of nodes i and j predicted by the auxiliary reconstruction head are respectively... The noiseless voltage amplitude; The desired voltage attenuation ratio is defined in the same trend verification definition. This represents the total number of characteristic frequencies.

[0036] The loss forces the graph network to learn a decay pattern that conforms to circuit laws, which is a form of physical information regularization.

[0037] Also includes 5: Topology adaptation: When the distribution network topology changes, the topology adjacency matrix is ​​automatically updated. This involves regenerating the topology encoding rules and triggering a lightweight adaptation of the graph propagation decoding model, specifically including: During system operation, the weights of the message passing body (GAT / GCN layer) of the graph propagation decoding model are determined. Keep it fixed, and only update the following three learnable parameters: 1) Input encoding layer: weights and bias; 2) Auxiliary Reconstruction Head: Weights and bias; 3) Classification layer: weights And bias.

[0038] The adaptation process involves fine-tuning a small number of new topology simulation samples (approximately 50-100 sets) for 5-10 rounds, with a learning rate of... The update can be completed within 1 to 2 minutes.

[0039] It also includes time synchronization fault tolerance and node data missing fault tolerance mechanisms; The time synchronization fault tolerance includes: after using BeiDou / GPS second pulses to achieve coarse synchronization of the sampling clocks of each FTU, the residual synchronization deviation is eliminated by using the global synchronization delay estimation method as described in step S3 data preprocessing.

[0040] The node data missing fault tolerance includes: When individual FTU node n is missing data, all elements in the nth column of matrix M are forcibly set to zero: ; Make the initial feature vector corresponding to this node It degenerates into a zero vector.

[0041] In the graph propagation decoding model, during message passing, although the initial feature of the missing node n is zero, its healthy neighbors... Still aggregate messages to n, let Integrate direct neighbor information; after L-layer iteration, By fusing the attenuation and phase shift characteristics of healthy nodes within the L-hop neighborhood, the state of the signal flowing through that node can be indirectly inferred. (Edge embedding) This inference feature is used to calculate the probability of branch faults, thereby achieving "soft recovery" of the state of the missing area.

[0042] This invention discloses an active distribution network fault location method based on topology coding and graph propagation decoding, with the following technical advantages: 1) This invention utilizes a topological adjacency matrix. The sharing ensures that the physical propagation path of the injected signal is strictly aligned with the message passing path of the graph propagation decoding model in terms of topology. The spatial attenuation pattern generated by the injected signal propagating along the topological edge can be completely extracted by the localization algorithm along the same topological edge, eliminating the information loss caused by path misalignment in traditional schemes.

[0043] 2) This invention is based on a topological adjacency matrix. The topological distance between each grid-connected inverter is calculated, and differentiated characteristic frequencies are allocated in ascending order of distance. This allows inverters closer to the bus to achieve wide-area coverage with lower frequencies, while those farther away use higher frequencies to limit local impact. Combined with path separability verification steps, it is ensured that the injection characteristic frequency sets of any two topological paths at the shared FTU are not completely identical. This mechanism utilizes the frequency dependence characteristics of line impedance, enabling different paths to form distinguishable multi-frequency response modes at the shared measurement node, fundamentally solving the path aliasing problem faced by fixed-frequency injection methods.

[0044] 3) This invention uses a topological adjacency matrix To ensure a unified foundation, both the topology encoding rules at the encoding end and the message passing paths at the decoding end are based on the topology adjacency matrix. Generation. When the distribution network topology changes due to switching operations or other reasons, the system automatically updates the topology adjacency matrix. The system regenerates topology coding rules and triggers a lightweight adaptation of the graph propagation decoding model, updating only the parameters of the input coding layer, auxiliary reconstruction head, and classification layer, while maintaining the main structure of message transmission in the graph network. In typical topology change scenarios such as branch line switching, tie switch switching, and feeder reconfiguration, the positioning accuracy recovers to over 90% within 30 to 120 seconds after lightweight adaptation, overcoming the drawback of requiring manual reconfiguration in traditional solutions.

[0045] 4) Before the discrimination matrix is ​​input to the graph propagation decoding model, this invention performs trend verification, phase verification, and low correlation verification based on the topological adjacency matrix A. Trend verification identifies abnormal attenuation or abnormal enhancement by comparing the deviation between the measured attenuation ratio and the reference attenuation ratio and generates prior weighted information; phase verification effectively reduces the sensitivity to synchronization error by utilizing differential phase features and performs accurate phase compensation through pre-emptive global synchronization delay estimation to ensure high-fidelity input of phase information; low correlation verification ensures that the identification information provided by each frequency dimension is not redundant, effectively suppressing misjudgments caused by reduced signal-to-noise ratio under high impedance conditions.

[0046] 5) This invention fully considers engineering constraints such as limited inverter control authority, heterogeneous communication conditions, and performance differences of existing FTU equipment in actual power distribution networks. The compatible deployment mode allows feature injection to be performed only on inverters with remote controllability, and the master station adaptively adjusts the topology coding rules according to the topology distribution of the actual controllable injection sources; for existing FTU equipment with low sampling rates, it can automatically switch to envelope detection mode or group time-division injection strategy. Attached Figure Description

[0047] The present invention will be further described below with reference to the accompanying drawings and examples; Figure 1 This is a diagram illustrating the architecture of an active distribution network fault location system provided in an embodiment of the present invention.

[0048] Figure 2 The flowchart shows the overall process of the fault segment localization algorithm based on topology coding and graph propagation decoding provided in this embodiment of the invention.

[0049] Figure 3 A schematic diagram of topology coding rules and spatial frequency reuse logic provided for embodiments of the present invention.

[0050] Figure 4 The flowchart for constructing the discrimination matrix and verifying constraints is provided for embodiments of the present invention.

[0051] Figure 5 This is a diagram of a multi-task graph propagation decoding network architecture provided in an embodiment of the present invention.

[0052] Figure 6 This is a schematic diagram illustrating the fault-tolerance mechanism under extreme operating conditions provided in an embodiment of the present invention.

[0053] Figure 7 This is a schematic diagram of topology coding rules and spatial frequency reuse logic.

[0054] Figure 8 This is a schematic diagram of the physical consistency verification of the topology propagation discrimination matrix.

[0055] Figure 9 A heatmap of the multi-task graph propagation decoding network and localization results. Detailed Implementation

[0056] like Figure 1 As shown, the active distribution network fault location system provided in this embodiment of the invention includes a master station controller, multiple feeder terminal units (FTUs), a communication network, and grid-connected inverters. The master station controller is responsible for maintaining the topology adjacency matrix, generating topology coding rules, verifying the physical consistency of the topology propagation discrimination matrix, decoding and locating the fault in the graph propagation matrix, and performing topology adaptive control. Each FTU is responsible for synchronously acquiring transient waveforms and uploading them to the master station. The communication network enables data transmission and control command issuance between the master station, FTUs, and grid-connected inverters. The grid-connected inverters inject injection signals of corresponding characteristic frequencies according to the topology coding rules issued by the master station.

[0057] like Figure 2 As shown, the overall algorithm flow of the method of the present invention covers five stages from step S1 to step S5. The following describes each step in detail with reference to specific embodiments, and elaborates on the preferred embodiments.

[0058] Step S1: Fault Initiation and Data Acquisition Real-time monitoring of zero-sequence voltage at each node of the distribution network. In a preferred embodiment, an adaptive threshold strategy is adopted: the master station continuously calculates the mean and standard deviation of the zero-sequence voltage imbalance during normal operation of each FTU using a sliding time window, and then sums the real-time statistical mean values. The standard deviation is used as the floating start threshold under the current operating conditions. When the neutral grounding method or system operating mode changes, causing a change in the zero-sequence voltage distribution, this floating start threshold is automatically updated. When the zero-sequence voltage of any FTU exceeds this floating start threshold, a ground fault is determined to have occurred in the system, and the location process is initiated.

[0059] After the fault is initiated, the master station synchronously completes the fault phase identification and fault feeder screening, and uses the zero-sequence voltage to zero-sequence current ratio method to initially estimate the transition resistance range. Then, it sends a waveform recording call command to all FTUs in the network to collect the voltage and current data of each FTU before and after the fault.

[0060] Step S2: Topological Encoding and Feature Injection: This step uses a topological adjacency matrix. To establish a unified foundation, a correspondence is constructed between the physical propagation path of the injected signal and the message passing path of the graph propagation decoding model.

[0061] 1. Topological adjacency matrix Real-time generation: The master station obtains the real-time open / close status of each switch and feeder log information from the distribution automation system. Using bus nodes, FTU installation nodes, and grid-connected inverter connection points in the distribution network as vertices, and the distribution lines connecting these vertices as edges, it constructs an undirected graph. For undirected graphs Perform radial verification: Using the bus node as the root, perform a depth-first search to check for loop edges. If a loop edge exists, disconnect the loop according to the switch status, and then perform a radial verification of the undirected graph. Convert to a radial tree. After successful validation, export the topological adjacency matrix. If node With nodes If there is a line connection, then ,otherwise Topological adjacency matrix It can be stored using a sparse matrix format. This invention is also compatible with topologically equivalent structures such as adjacency lists.

[0062] 2. Formation of topological coding rules: like Figure 3 As shown on the left, the main station takes the bus node as the root and is in the topological adjacency matrix. A breadth-first search is performed to calculate the topological distance from each grid-connected inverter to the bus. The specific search process is as follows: Assume the bus node... , ; Create a queue When Q is not empty, pop the head of the queue u, and iterate through the neighbors v. If not visited, enqueue and set aside. After the search is completed, the topological distance of each inverter k is calculated. Based on the topological distance from nearest to farthest, incremental characteristic frequencies are assigned to each grid-connected inverter. This allocation method allows grid-connected inverters closer to the bus to cover a wider area with a lower characteristic frequency, while grid-connected inverters farther away limit their impact to a local area with a higher characteristic frequency.

[0063] Frequency allocation must satisfy the path uniqueness mapping condition: for any two different topological paths flowing through the same FTU, their injected characteristic frequency sets must differ at least in one characteristic frequency. The execution steps of the frequency allocation algorithm are as follows: S2.1: Set of available feature frequencies Sort by numerical value in ascending order, and then sort the grid-connected inverters by topological distance. After sorting in ascending order, allocate in sequence. The corresponding frequencies are used to form an inverter-frequency mapping table. ; S2.2: Path separability verification, for each pair of grid-connected inverters , Traverse all FTU nodes ,like and All injected signals can reach the same FTU node. ,Right now exist To the busbar On the path to the busbar, check and At this FTU node Does the set of injected feature frequencies at a given location have an overlap? If there is an overlap, prioritize moving the overlapping feature frequencies with the larger topological distance to the next location. The next unused feature frequency is used to ensure that the injected feature frequency sets at the FTUs traversed by any two topological paths are not completely identical. When the available feature frequency resources are insufficient to achieve complete non-intersection, for path pairs with frequency overlap, a stricter correlation coefficient threshold is used in the low correlation check of subsequent physical consistency verification, and post-processing based on independent component analysis based on blind source separation is enabled. S2.3: Propagation coverage verification, for For each assigned characteristic frequency, the signal attenuation to the farthest FTU is estimated based on the line length and unit impedance parameters. If the signal-to-noise ratio after attenuation is lower than the preset detection threshold, the characteristic frequency is removed and adjusted to the next available characteristic frequency for the corresponding grid-connected inverter. S2.4: Spectrum Gap Compliance Check: Verification If all characteristic frequencies fall within the spectral gap between integer harmonics of the power frequency and the difference between any two characteristic frequencies is not less than the preset frequency interval, then adjust to the nearest compliant frequency point.

[0064] All characteristic frequencies are selected from the spectral gaps between integer harmonics of the power frequency. During steady-state operation of the distribution network, background harmonic energy is highly concentrated in integer multiples of the power frequency, while the frequency bands between integer harmonics have lower energy. Selecting these gap frequency bands can avoid the inherent strong interference of the power grid. To cope with fluctuations in the fundamental frequency of the power grid, the grid-connected inverter uses a phase-locked loop to track the fundamental frequency in real time. Each characteristic frequency is defined as a fixed offset relative to the fundamental frequency. When the fundamental frequency of the power grid changes, the absolute value of the injected frequency is adjusted synchronously to keep this fixed offset constant.

[0065] 3. Implementation of signal injection control: After a fault occurs, the master station generates injection control commands according to the topology coding rules. The injected signals uniformly use negative-sequence components, and the time-domain reference waveform is... ; in: To inject amplitude, , To synchronize the initial phase. The instructions include... The time stamp and other information are transmitted via a communication channel. Specific control implementations vary depending on the inverter type, as follows.

[0066] For grid-connected inverters, a synchronous rotating coordinate system (dq system) or a stationary coordinate system (dq system) is typically used. The current control strategy of the system. Taking proportional resonance (PR) control in a coordinate system as an example, the corresponding characteristic frequency is superimposed in the current control loop. Negative sequence current reference component: ; in: The injected current amplitude shall not exceed 2% of the rated output current of the grid-connected inverter. The PR controller operates at its characteristic frequency. A resonant term is added to achieve zero steady-state error tracking of this frequency component. The amplitude of the negative-sequence voltage response excited by the grid-connected inverter depends on the system equivalent impedance at the connection point. When the short-circuit capacity at the connection point is large, the voltage response excited by the rated injection current may not be sufficient to meet the detection sensitivity of the remote FTU. During the topology coding stage, the system pre-assesses the injection effectiveness based on the short-circuit capacity of each grid-connected inverter's connection point: if the signal-to-noise ratio of the expected excitation voltage of a grid-connected inverter at the farthest FTU is lower than the preset detection threshold, then the grid-connected inverter is marked as a "weak injection source" and is not included in the frequency allocation sequence of the topology coding, or its characteristic frequency is adjusted to the same characteristic frequency of its downstream adjacent controllable grid-connected inverter to enhance the signal strength on that path.

[0067] For grid-connected inverters, droop control or virtual synchronous machine strategies are employed, operating in voltage source mode. A negative-sequence voltage reference component, after coordinate transformation, is superimposed on the dq-axis reference voltage generation stage of its voltage control loop. The characteristic frequency is then... After the negative-sequence voltage phasor is transformed to the dq coordinate system, the corresponding AC components are superimposed on the voltage reference values ​​of the d-axis and q-axis, respectively. The grid-type inverter generates the corresponding negative-sequence voltage response by tracking the reference value through the voltage closed loop. The injected voltage amplitude does not exceed 1% of the rated voltage.

[0068] Multiple grid-connected inverters synchronously initiate injection based on a unified time stamp issued by the master station. Injection commands are sent via communication channels, and the time stamp is generated based on BeiDou / GPS or Network Time Protocol (NTP). The injection duration is no less than 10 power frequency cycles to meet the frequency resolution requirements of frequency domain analysis. If a current voltage drop has triggered a low-voltage crossover of the mandatory reactive power support threshold, reactive power support obligations are prioritized, and characteristic injection is initiated only after voltage recovery. If the available injection time window is insufficient to meet the minimum frequency resolution requirements, the system automatically switches to standby positioning mode.

[0069] Step S3: Discriminant matrix construction and physical consistency verification: This step processes the transient waveforms acquired by the FTU, constructs a topology propagation discrimination matrix, and performs physical consistency checks. For example... Figure 4 As shown, the topology propagation discrimination matrix must pass trend verification, phase verification, and low correlation verification before it is input to the graph propagation decoding model.

[0070] 1. Online parameter identification and reference attenuation ratio establishment: During normal operation, the system periodically performs online identification of line parameters. The master station, either within a preset low-interference time window or triggered by system commands, controls each grid-connected inverter to inject a set of low-amplitude (not exceeding 0.5% of the rated voltage) characteristic detection signals. The frequency of these characteristic detection signals can be the same as or a subset of the characteristic frequency used for fault injection. Each FTU uploads the collected steady-state response data to the master station, which uses a topology adjacency matrix... To constrain this, establish the following for each section of the line: Equivalent circuit parameter identification model. Line series impedance. Parallel admittance and downstream equivalent impedance Obtained within the same identification framework, it satisfies the inherent consistency of circuit laws and avoids parameter conflicts. A recursive least squares algorithm with a variable forgetting factor is employed to track the slow changes in line impedance caused by factors such as temperature and aging in real time. The identified series impedance and parallel admittance parameters are stored in the main station database. The identification results are used after being filtered by moving average and anomaly detection, and the reference attenuation ratio in the auxiliary reconstruction head part of the graph propagation decoding model is updated synchronously.

[0071] 2. Data Acquisition and Preprocessing: Each FTU synchronously records the transient voltage waveforms before and after a fault at a sampling frequency that meets the Nyquist criterion. The analog-to-digital converter resolution of the FTU measurement unit is no less than 16 bits, and the noise floor of the analog front-end in the characteristic frequency passband is no higher than -80dB (based on the rated voltage). For existing FTU equipment with low sampling rates, the system can automatically switch to envelope detection mode, and meet the frequency resolution requirements by extending the injection duration and sampling window; alternatively, a grouped time-division injection strategy can be adopted, dividing all characteristic frequencies into several groups, with each group injected and acquired sequentially in different sub-time windows, and each group only needs to distinguish the frequencies within the group within a shorter sampling window.

[0072] Data preprocessing first involves separating the interference sources at the fault points: before feature injection begins, waveforms of at least one power frequency cycle are captured, and the background negative-sequence phasors at each characteristic frequency point are extracted through frequency domain analysis. After injection is initiated, the negative-order phasors of the corresponding characteristic frequencies during the injection period are extracted synchronously. The net injected response phasor generated solely by active injection is obtained through vector difference operations: ; Where i is the characteristic frequency number ( ), j is the FTU number ( ), and These are the negative sequence voltage phasors during and before injection, respectively.

[0073] The extraction of the net injected response is based on the negative-sequence component. Each FTU synchronously acquires the instantaneous values ​​of the three-phase voltage and current, and calculates the negative-sequence voltage and current components in the frequency or time domain using the symmetrical component method. For the net injected response, a windowed Fast Fourier Transform (FFT) is used to extract the amplitude and phase at each characteristic frequency point. Before extraction, the original signal is filtered using a bandpass filter bank centered at each characteristic frequency, complementing the spectral gap design. When conventional FFT cannot meet the frequency resolution requirements due to sampling window limitations, a subspace-based high-resolution frequency estimation algorithm can be used instead.

[0074] Before constructing the matrix using the aforementioned phasors, to ensure strict synchronization of the phase information of each FTU, the following steps are performed: and Global synchronization delay compensation is performed on the argument angle. Specifically, for each FTU, the characteristic frequencies are extracted. Measured phase below And unwind; calculate the reference phase based on the circuit model provided by online parameter identification. ;make The median is taken as the synchronization delay estimate for the FTU. The compensated phase is Update in this way and The phase angle is then used to perform vector difference calculations to obtain the final net injected response phasor, denoted as . The superscript 'c' indicates that synchronization error has been compensated.

[0075] Each compensated net injection response phasor Arrange them into a complex matrix with characteristic frequencies as rows, FTU nodes as columns, and the column order strictly consistent with the node indices of the topological adjacency matrix A. .element That is, the characteristic frequency. exist Amplitude and phase at the point 。

[0076] 3. Physical consistency verification: Physical consistency verification is based entirely on the net injection response of the injected characteristic signals. The expected value of each verification quantity is determined based on the impedance parameters of the distribution network lines at the corresponding characteristic frequencies, decoupling from the power distribution and flow direction of the fundamental frequency. Therefore, this verification system is inherently compatible with bidirectional power flow conditions at the power frequency, including distributed photovoltaic systems. It is designed to be compatible with bidirectional power flow conditions at the power frequency.

[0077] (1) Trend verification: The system uses online parameter identification to pre-establish a reference attenuation ratio matrix for the entire network. The reference attenuation ratio is calculated using the following formula: ; This is the series impedance of the line. This is the downstream equivalent impedance. The reference attenuation ratio tolerance range is determined through statistical analysis of 30 consecutive days of normal operation data: the mean of historical data is calculated. and standard deviation ,Pick (Overhead section) or (Cable section), the maximum daily attenuation fluctuation of overhead line sections shall not exceed The cable section does not exceed Unified This tolerance range serves as a margin for future engineering work. This tolerance range can be dynamically adjusted based on online updates of historical operational data statistical characteristics.

[0078] After a fault, for each controlled injection source, along the topological adjacency matrix Given a defined propagation path, calculate the measured attenuation ratio between adjacent nodes: ; in For matrix The magnitude of the element in the k-th row and i-th column. Compare. and : ①. Normal attenuation: and The deviation is within the tolerance range, and the verification is passed.

[0079] ②. Abnormal attenuation: Furthermore, if the deviation exceeds the lower limit of the tolerance range, it indicates that the signal has experienced additional attenuation beyond the normal line consumption. The verification passes, but an abnormal attenuation marker is generated synchronously, which is converted into the prior attention weighting of the corresponding branch during the graph propagation decoding stage.

[0080] ④. Abnormal enhancement: If the deviation exceeds the upper limit of the tolerance range or the amplitude increases in reverse, the verification passes and an abnormal enhancement prompt mark is generated.

[0081] ⑤. Verification Failure: Verification failure is determined and backup positioning mode is triggered only when there is a step decay or continuous violent fluctuation without physical basis.

[0082] When frequency reuse exists, each injection source is verified along its respective topology path. For grid-connected inverters marked as "weak injection sources", the attenuation monotonicity constraint of their corresponding characteristic frequency row is relaxed.

[0083] (2) Phase verification: The equivalent admittance of parallel loads distributed along the distribution network and the line distributed capacitance have a shunt effect on the injected signal, causing the node voltage phase to shift along the propagation direction. For each pair of adjacent nodes (i,j) in the topological adjacency matrix A, calculate the characteristic frequencies. The measured phase shift is as follows: Reference phase shift: The tolerance range is determined by the following formula: ; in: The baseline tolerance value (e.g.) L is the line length. The error coefficient for the current transformer is 0.5 for level 0.5 and 1.0 for level 1.

[0084] Simultaneously construct differential phase features: This differential operation can further reduce the impact of the small synchronization deviation that may still remain after synchronization compensation on the phase shift criterion, and improve the robustness of phase verification.

[0085] (3) Low correlation check: Within the typical distribution network parameter range, injected signals of different characteristic frequencies exhibit low correlation during propagation. The topology propagation discrimination matrix is ​​calculated. The Pearson correlation coefficient between any two response vectors in the matrix. The correlation coefficient is low between response vectors of different characteristic frequencies on non-faulty paths, and slightly increases on faulty paths due to a common attenuation trend. The correlation coefficient threshold is dynamically set based on the sufficiency of available characteristic frequency resources: when available frequency resources are sufficient to allocate non-overlapping frequency combinations to each path according to S2.2 requirements, the threshold is set to 0.4; if frequency resources are scarce, resulting in frequency overlap in some paths, the threshold is tightened to 0.2 to ensure that the identification information provided by each frequency dimension is not redundant. The topology propagation discrimination matrix that satisfies all constraints. This constitutes an effective discrimination space; if any constraint is not satisfied, the backup positioning mode is triggered and an alarm flag is generated simultaneously.

[0086] Step S4: Output of graph propagation decoding and localization results: In this step, the message passing path of the graph propagation decoding model is determined by the topological adjacency matrix. Defined as sharing the same topology as the physical path along which the injected signal propagates along the topological edges.

[0087] In a distribution network with multiple branches, injected signals can reach the same FTU node along multiple paths. The message passing in the graph propagation decoding model follows the topological adjacency matrix. The defined tree-like connection corresponds to the shortest path component of the injected signal propagated along the distribution line, which contributes the main energy to the arrival response under most operating conditions. Multipath aliasing components introduced by branch node reflections or circulating paths will be identified as abnormal fluctuations and marked as such during trend and phase checks because they deviate from the monotonic attenuation and phase shift relationship along the shortest path.

[0088] like Figure 5 As shown, the graph propagation decoding model adopts a multi-task learning architecture, the specific structure of which is as follows: S4.1: Initial Feature Encoding of Nodes: For FTU Nodes Its topological propagation discriminant matrix The corresponding columns are the characteristic frequencies. Net injection response amplitude and phase To avoid periodic jumps in phase angle values, the phase is decomposed into two channels: sine and cosine, and initial feature vectors for each node are constructed. ; The feature dimension is 3 ,in: The total number of characteristic frequencies, indicated by the superscript. This indicates transpose. It passes through a fully connected coding layer (input 3...). The output is 64, and the ReLU activation function maps it to a 64-dimensional initial node feature vector. .

[0089] S4.2: Graph Network Main Structure: A stacked structure of two layers of Graph Attention Network (GAT) and one layer of Graph Convolutional Network (GCN) is employed. First GAT layer: Input dimension 64, output dimension 64, 4 attention heads, concatenated dimension 256, linearly projected to 64 dimensions; Second GAT layer: Input dimension 64, output dimension 64, 4 attention heads, concatenated and projected to 64 dimensions; GCN layer: Input dimension 64, output dimension 64, ReLU activation function. Message passing occurs only along the topological adjacency matrix. The edge with a median value of 1 is processed. The layer message passing formula is: ; in, For nodes exist The set of neighboring nodes on, For nodes In the Layer feature representation, For the first The learnable weight matrix of the layer, It is a non-linear activation function.

[0090] S4.3: The number of model layers should not be less than the maximum topology depth from the furthest grid-connected inverter to the bus in the distribution network, to ensure that message transmission can cover the complete path from the furthest injection point to the root node. In the embodiment, the maximum topology depth is usually 3, so the total number of graph network layers is set to 3; if used for deeper networks, the number of layers can be increased for flexible adaptation.

[0091] S4.4: Edge Embedding and Branch Classification: For Connecting FTU Nodes and The branch is used to generate the failure probability through edge embedding. The edge embedding vector is constructed as follows: ; in: , They are nodes , High-level feature representation after the graph network body. This represents a vector concatenation operation. Edge embedding vectors. The dimension is 192. After passing through two fully connected classification layers (192→32→1, with ReLU activation function in the intermediate layer), the output logits are transformed into the probability that the branch is a faulty segment by the Sigmoid function. .

[0092] S4.5: Auxiliary Reconstruction Header: High-Level Features of Each Node The output, via a fully connected layer (64→F), provides predicted values ​​for the node response amplitudes at each characteristic frequency. Based on topological adjacency matrix Given the upstream and downstream relationships determined by the injection source location, the difference between the predicted attenuation ratio between adjacent nodes and the reference attenuation ratio estimated based on line impedance parameters is calculated to construct a physical regularization loss term. .

[0093] S4.6: Model Training. The total loss function is: ; in: For branch classification, the binary cross-entropy loss, The weight coefficient is set to 0.05. Training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a maximum of 200 training epochs. An early stopping mechanism is implemented, terminating training if the loss does not decrease for 20 consecutive epochs. The model training dataset covers different photovoltaic penetration rates (0% to 80%), transition resistances (0 to 3000Ω), fault locations, and topological conditions. A domain randomization strategy is employed to enhance generalization ability.

[0094] Selecting the probability of failure The branch that is the largest and exceeds the preset confidence threshold is used as the fault location result. The preset confidence threshold typically ranges from 0.6 to 0.8.

[0095] Step S5: Topology Adaptation: The master station monitors the status of distribution network switches in real time. When a topology change is detected, the topology adjacency matrix is ​​automatically updated. The topology distances of each grid-connected inverter are recalculated, and new topology coding rules are generated. These topology coding rules are based on the topology adjacency matrix. The design and topology changes directly trigger a linked update of the topology encoding rules. Simultaneously, a lightweight adaptation of the graph propagation decoding model is triggered: only the input encoding layer parameters, auxiliary reconstruction head parameters, and classification layer parameters are updated; the message passing portion of the graph network entity and its corresponding network weights remain fixed. Adaptation samples can be selected from the historical simulation database based on topology similarity. During the lightweight adaptation transition period, the system switches to a backup positioning mode to ensure uninterrupted positioning.

[0096] During system operation, the message passing structure of the graph propagation decoding model, including the network weights of the graph convolutional layer and graph attention layer, remains fixed and is not modified online due to topology changes or data distribution variations. Only the input encoding layer parameters, as well as the auxiliary reconstruction head and classification layer parameters, can be used for lightweight adaptation. This mechanism maintains the stability of the model's core inference logic during long-term operation and reduces the risks introduced by online updates to the model's main body.

[0097] Preferred Implementation Method 1: Compatible Deployment Mode: Considering the heterogeneity of grid-connected inverters in practical engineering in terms of manufacturer protocols, communication methods, and scheduling permissions, this invention supports a compatible deployment mode: feature injection is performed only on grid-connected inverters with remote controllability, while the remaining grid-connected inverters maintain normal operation. The master station adaptively adjusts the topology coding rules according to the topology distribution of the actual controllable injection sources: if all grid-connected inverters on a certain branch are uncontrollable, an injection source is selected from an adjacent controllable branch, and its wide coverage characteristic of low-frequency injection signals is used to excite the branch across regions; if controllable injection sources are sparse in a certain area, the path uniqueness constraint of frequency allocation in that area is relaxed, and a higher local regularization coefficient is assigned to the edge set corresponding to the injection sources in that area during the lightweight adaptation stage.

[0098] Preferred Implementation Method Two: Spatial Frequency Reuse Mechanism When the number of grid-connected inverters exceeds the total number of available characteristic frequencies, the electrical isolation threshold is determined through pre-simulation—based on a signal attenuation of no less than 15dB between two grid-connected inverters. For example... Figure 3 As shown on the right, a conflict graph is constructed with grid-connected inverters as vertices and node pairs with electrical distances less than the isolation threshold as edges. This conflict graph is then colored to achieve frequency reuse. For grid-connected inverters reusing the same characteristic frequency, a time-division multiplexing method is further adopted, specifying non-overlapping injection time micro-windows. Each FTU extracts the corresponding component in segments according to the time window sequence. The injection timing of the time-division multiplexing mechanism is dynamically allocated by the master station based on the distribution network protection action time margin.

[0099] Preferred Implementation Method 3: Time Synchronization Fault Tolerance This invention requires that the coarse synchronization error of each FTU sampling clock does not exceed Based on coarse synchronization, a global synchronization delay estimation method as described in step S3 (data preprocessing) is used to further eliminate residual bias: for each FTU, the equivalent delay is calculated using the difference between the measured phase and the reference phase at multiple characteristic frequencies, the median is taken as the synchronization delay estimate, and compensation is applied to the phase at each frequency. The compensated phase is used to construct the topology propagation discrimination matrix M, thereby ensuring that the input for graph propagation decoding is not affected by synchronization errors.

[0100] Preferred Implementation Method 4: Graph Node Masking and Two-End Missing Bidirectional Reasoning like Figure 6 As shown in subgraph b on the right, when individual FTU node n has missing or invalid data, all elements in the nth column of the topology propagation discrimination matrix M are set to zero: ( In the graph propagation decoding model, the topological adjacency matrix is ​​used during message passing. The corresponding edge weights are automatically weakened because the input features are zero. The model uses the high-level features of the remaining healthy FTUs to infer the propagation state of the missing node through graph convolution. In the extreme scenario where data loss occurs simultaneously at both ends of the faulty section's FTUs, the injected signal from the grid-connected inverter propagates along the topology path, inevitably passing through the upstream and downstream branch nodes of that branch. The response data of these adjacent nodes contains information about signal attenuation and phase shift as it traverses the branch. The graph propagation decoding model performs bidirectional reasoning combining forward aggregation and backward propagation during message passing, comprehensively inferring the fault probability of the branch by integrating the propagation patterns of adjacent regions. Specifically, the neighbors of the missing node still aggregate messages towards it, and its high-level features gradually integrate the attenuation and phase shift patterns of neighboring healthy nodes, thereby achieving soft recovery of the node's propagation state.

[0101] Preferred Implementation Method 5: Degradation Location When Insufficient Injection Source When the number of available injection sources falls below a preset lower limit, and the topology propagation discrimination matrix... If any physical consistency constraint is not met and cannot be recovered through resampling, or if the lightweight adaptation of the model after topology change is not yet complete, the main station automatically switches to the degraded positioning implementation method. This degraded scheme adopts a dual-modal gated fusion network. The input consists of two modes: the incomplete topology propagation discrimination matrix M' (the incomplete part is set to zero according to the graph node masking mechanism) and the power frequency negative sequence voltage and negative sequence current phasors extracted synchronously by each FTU. Normalized modal weight coefficients are generated by a learnable modal quality evaluator. The two latent representations are weighted and summed, and then fed into the topology adjacency matrix shared with the main model. However, the lightweight graph propagation decoder with half the number of layers outputs the failure probability of each branch. When the standby mode is activated, the system synchronously generates alarm flags and informs maintenance personnel of the currently available modal combinations and expected positioning accuracy.

[0102] Technical effectiveness verification: Figure 7 correspond Figure 3 The aim is to demonstrate how to assign incremental characteristic frequencies to inverters based on topological distance, and how to perform spatial frequency reuse. Figure 7A simplified three-branch distribution network model visually illustrates the core coding concept of this invention. The diagram labels the bus, FTU, and inverter nodes. Based on the topological distance calculated using breadth-first search, each inverter (red circle) is assigned an increasing characteristic frequency (marked below the node). INV1, closer to the bus, is assigned a lower frequency (e.g., 25Hz) to utilize the low attenuation and wide coverage of low frequencies; INV4, farther away, is assigned a higher frequency (e.g., 55Hz) to regionalize its signal impact. The gray dashed line further illustrates the spatial frequency reuse mechanism: when two inverters (e.g., INV3 and INV1) are electrically far enough apart and their signal attenuation exceeds 15dB, they can reuse the same characteristic frequency without causing path confusion. This effectively solves the bottleneck problem when the number of inverters exceeds the available frequency resources.

[0103] Figure 8 correspond Figure 4 This is used to demonstrate how signals acquired at the FTU are extracted and verified through net injection response to ensure the physical reliability of the discrimination matrix. Figure 8 This section demonstrates how physical consistency verification ensures the data quality of the discrimination matrix through three dimensions. The left subplot (trend verification) shows the voltage amplitude at downstream node j at three characteristic frequencies: 25Hz, 35Hz, and 45Hz. The solid blue line represents the normal operating condition, where the amplitude monotonically decreases with increasing frequency (consistent with the frequency characteristics of line impedance). The dashed red line represents the fault condition, where the fault point abruptly bypasses or blocks the signal, causing a step-like abnormal voltage decrease at the downstream node, far below the tolerance range set based on historical parameters (the dashed black line). The system marks this as "abnormal attenuation" and generates prior weighted information. The middle subplot (phase verification) compares the voltage phase shift under normal and abnormal operating conditions. Under abnormal conditions, the phase shift deviation reaches as high as 15°, far exceeding the 5° tolerance. Differential phase characteristics can further filter out residual synchronization errors, ensuring the reliability of phase information. The right subplot (low correlation verification) compares the correlation coefficients of the response vectors at different frequencies under normal and fault paths. Along the fault path, since all signals experience severe attenuation at the fault point, the response patterns of different frequencies tend to be consistent, resulting in a significant increase in the correlation coefficient and exceeding the 0.4 threshold. This statistically confirms the path anomaly and provides a high-quality input matrix for subsequent graph propagation decoding.

[0104] Figure 9 correspond Figure 5 This is used to demonstrate how the graph propagation decoding model transmits information along topological edges and outputs the branch failure probability.

[0105] Figure 9The left subgraph represents a simplified structure of the graph propagation decoding network. The net injected response amplitude and phase collected by the FTU nodes are encoded to form an initial feature vector X, which is then input into a "message passing layer" composed of a multi-layer graph attention network (GAT) and a graph convolutional network (GCN). The key to this architecture is that messages are only transmitted and aggregated along the distribution network topology edges defined by the topological adjacency matrix A, ensuring strict alignment between the information aggregation path and the physical propagation path of the injected signal. Finally, a branch feature vector e{ij} is generated through edge embedding operations, and the fault probability is output via a classification layer. Simultaneously, a parallel "auxiliary reconstruction head" predicts the voltage amplitude based on the node's higher-level features, using the minimization of the difference between the predicted attenuation and the reference attenuation as a physical regularization term to guide the graph network to learn a propagation pattern that conforms to circuit laws.

[0106] Figure 9 The right subplot is a visual heatmap representation of the fault location results. The color depth and line width of the branches in the right subplot represent the probability of a fault occurring in that branch. As can be seen, the branches connecting FTU2 and FTU3 (numbered 2-3) are highlighted in red, with a probability of 0.95, far exceeding other normal branches (blue / green, with a probability below 0.15). Therefore, the system identifies this branch as a faulty section. This visualization result verifies that the graph propagation decoding model can effectively aggregate multi-frequency response features propagated along the topology, accurately locate fault positions, and perform excellently even under high-resistance (1500Ω) conditions.

[0107] The method of this invention was validated on an improved distribution network topology model based on the IEEE 34-node standard. The test conditions were: 3 main branches, 6 FTU installation nodes, 60% photovoltaic penetration, 15 distributed photovoltaic inverters (80% grid-connected, 20% network-connected), single-phase ground fault transition resistance range of 0 to 3000 Ω, 8 available characteristic frequencies, FTU sampling rate of 10 kHz, analog-to-digital converter resolution of 16 bits, median signal-to-noise ratio of 5.5 dB in the characteristic frequency band, and 500 independent test samples. Experimental results show that the fault location accuracy of the method of this invention is [insert accuracy here]. The accuracy rate is 97.1% under transition resistance conditions of 0 to 500Ω, 95.8% under 500 to 1500Ω, and approximately 90% under 1500 to 3000Ω, all significantly better than passive schemes based on the same graph network body but without topology coding and physical consistency checks. ) and the conventional active scheme using fixed-frequency negative-order injection and threshold discrimination ( In the ablation experiment, keeping the topology coding rules unchanged, after replacing the message passing adjacency matrix of the graph propagation decoding model with a random adjacency matrix and a fully connected adjacency matrix respectively, the localization accuracy decreased from 96.2% to 62.1% and 58.5%, respectively, verifying the key role of aligning the message passing path and the physical propagation path in the topology for localization performance.

[0108] The aforementioned technical effects do not stem from the individual use of graph networks or active injection, but rather from the specific topology encoding rules and graph propagation decoding model in the topology adjacency matrix. Coordination and cooperation at all levels.

Claims

1. An active distribution network fault location method based on topology coding and graph propagation decoding, characterized in that... Includes the following steps: Step 1: Monitor the zero-sequence voltage of each node in the distribution network in real time. When the zero-sequence voltage of any node exceeds the start-up threshold, it is determined that a ground fault has occurred and the location process is initiated. Simultaneously estimate the range of transition resistance and send a waveform recording call command to each feeder terminal unit (FTU). Step 2: The master station obtains the topology adjacency matrix of the target distribution network. Using the bus node as the root in the topological adjacency matrix The breadth-first search is performed to obtain the topological distance from each grid-connected inverter to the bus. Increasing characteristic frequencies are assigned to each grid-connected inverter according to the topological distance from near to far to form topological coding rules. After a fault, each grid-connected inverter is controlled to inject the corresponding characteristic frequency injection signal according to the topological coding rules. Step 3: Each feeder terminal unit (FTU) acquires transient waveforms and separates the interference sources at fault points. The net injection response generated solely by active injection is obtained through vector difference calculation. The amplitude and phase of the net injection response at each characteristic frequency are extracted. A matrix is ​​constructed with characteristic frequencies as rows and feeder terminal units (FTUs) as columns, arranged in order with the topological adjacency matrix. The topology propagation discriminant matrix with consistent node order ; Topological propagation discriminant matrix Execute based on topological adjacency matrix The physical consistency check will be performed through the topology propagation discrimination matrix of the check. Marked as a valid discrimination matrix; Step 4: Combine the effective discriminant matrix with the topological adjacency matrix The input graph propagation decoding model is used together with the topological adjacency matrix. Define message passing paths; messages are only sent through the topological adjacency matrix. The fault probability of each branch of the distribution network is output after passing between adjacent nodes, node feature aggregation and edge embedding, and the branch with the highest fault probability and exceeding the preset confidence threshold is selected as the fault section location result.

2. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 1, characterized in that: In step 2, the topological adjacency matrix satisfy If and only if node i and j There are line connections between them, that is, A is a symmetric matrix, which corresponds to the bidirectional signal propagation characteristics of the distribution network line; Using the bus node as the root in the topological adjacency matrix The topological distance from each grid-connected inverter to the bus is obtained by performing a breadth-first search. The specific process is as follows: Set bus node Define topological distance Establish a queue ; If Q is not empty, pop the head of the queue u and iterate through all queues that satisfy the condition. That is, if there is a line connection between nodes u and v, and a neighbor v that has not been visited is enqueued and recorded. ,in: Represents a node Topological distance to the bus node; After the search is completed, the topological distance of each grid-connected inverter k is... .

3. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 2, characterized in that: In step 2, the incremental characteristic frequencies are allocated to each grid-connected inverter according to the topological distance from near to far, specifically including: S2.1: Set of available feature frequencies Sort by numerical value in ascending order, and then sort the grid-connected inverters by topological distance. After sorting in ascending order, the available feature frequency sets are allocated sequentially. The corresponding frequencies are used to form an inverter-frequency mapping table. ; S2.2: Path Separability Verification: For each pair of inverters , Traverse all FTU nodes ,like and All injected signals can reach the same FTU node. Then check and At this node Does the set of injected feature frequencies at a given location have an intersection? If an intersection exists, prioritize moving the overlapping feature frequencies with the larger topological distance to the available feature frequency set. The next unused feature frequency in the middle ensures that the sets of injected feature frequencies at the FTUs that any two topological paths pass through are not exactly the same; S2.3: Propagation Coverage Verification: Inverter-Frequency Mapping Table For each assigned characteristic frequency, the signal attenuation to the farthest FTU is estimated based on the line length and unit impedance parameters. If the signal-to-noise ratio after attenuation is lower than the preset detection threshold, the characteristic frequency is removed and adjusted to the next available characteristic frequency for the corresponding grid-connected inverter. S2.4: Spectrum Gap Compliance Check: Verify Inverter-Frequency Mapping Table If all characteristic frequencies fall within the spectral gap between integer harmonics of the power frequency and the difference between any two characteristic frequencies is not less than the preset frequency interval, then adjust to the nearest compliant frequency point.

4. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 3, characterized in that: In step 2, after a fault occurs, each grid-connected inverter is controlled to inject an injection signal at the corresponding characteristic frequency according to the topology coding rules; specifically: The main site uses the mapping table To the k The inverter issues a command carrying a unified time tag. The injected signal adopts the form of a negative sequence component, and its time-domain expression is as follows: ； in: The instantaneous value of the negative sequence signal injected into the k-th inverter; To inject amplitude, The characteristic frequency allocated to the k-th inverter satisfies ; This is the initial phase for synchronization.

5. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 4, characterized in that: In step 3, the interference source at the fault point is separated, and the net injection response generated solely by active injection is obtained through vector difference calculation. Specifically, this includes: S3.1: Before feature injection is initiated, capture the waveform for at least one power frequency cycle and extract the background negative sequence phasor at each feature frequency point. During feature injection, the negative-order phasors corresponding to the feature frequencies are extracted simultaneously. ; S3.2: Calculate the elements of the net injection response matrix generated solely by active injection using the following formula: ； in: During feature injection, FTU nodes j At characteristic frequency Negative sequence voltage phasor at the location; Before feature injection is initiated, the FTU node j At characteristic frequency Background negative sequence voltage phasor at the location; The net injected response phasor, whose magnitude and argument are the th element in the topological propagation discriminant matrix M. i Line 1 j The element values ​​of the column; F is the total number of characteristic frequencies; N is the total number of FTU nodes; Topological propagation discriminant matrix That is to The amplitude and phase are used as element values.

6. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 5, characterized in that: In step 3, the net injection response phasors obtained in step S3.2 are... The characteristic frequencies are arranged in rows, FTU nodes in columns, and the column order is aligned with the topological adjacency matrix. The node indices are arranged in a strictly consistent manner to form a complex matrix. : ； Where the i-th row corresponds to the characteristic frequency The j-th column corresponds to FTU node j. element F represents the total number of characteristic frequencies, and N represents the total number of FTUs.

7. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 6, characterized in that: In step 3, the physical consistency verification includes a trend verification, which includes: a1: During normal system operation, the impedance parameters of each line segment at the characteristic frequency are obtained based on online parameter identification, and the impedance parameters of each adjacent node are pre-calculated. For characteristic frequencies under fault-free operating conditions Expected decay ratio The details are as follows: ； in: Representing the characteristic frequencies of nodes i and j respectively. Voltage phasors under these conditions; This is the voltage amplitude attenuation ratio estimated based on the impedance model, i.e., the desired attenuation ratio. ; For line segment series impedance, For line resistance, For line inductance; The equivalent impedance seen from node j into the rest of the network except for the direction i; The mean value is calculated based on the historical sequence of expected decay ratios identified during continuous T days of normal operation. and standard deviation The tolerance range is determined by the following formula: [ ； in: Characteristic frequency The tolerance range for the lower expected attenuation ratio. The lower limit of the interval, The upper limit of the interval; Confidence coefficient; The above expected attenuation ratio The definition is based on signals from nodes To the node Dissemination, that is Comparison Closer to the injection source; online parameter identification provides For the node To remove The equivalent impedance seen from the network outside the direction; during verification, the positive direction of signal propagation needs to be determined based on the injection source location, and a reference attenuation ratio for the corresponding direction should be selected; if the signal flow on this side is Then its expected attenuation ratio should be used ; a2: During fault location, for each controlled injection source, along the topological adjacency matrix Given a defined propagation path, calculate the measured attenuation ratio between adjacent nodes: ； in To at characteristic frequency The next signal comes from the node. i propagation to nodes j The measured voltage attenuation ratio; For matrix row k, row The magnitude of the column element corresponds to the characteristic frequency. At the node i Net injection response amplitude after compensation at the location; For matrix row k, row The magnitude of the column element corresponds to the characteristic frequency. At the node j Net injection response amplitude after compensation at the location; a3: Comparison and : when and If the deviation is within the tolerance range, it is determined to be normal attenuation and passes the verification. when If the deviation exceeds the lower limit of the tolerance range, it is marked as abnormal attenuation and prior weighting information is generated for use in step S4; when Furthermore, if the deviation exceeds the upper limit of the tolerance range or an amplitude increases in reverse, it is marked as an abnormal enhancement; only when a step decay without physical basis occurs is the verification determined to be a failure and the backup positioning mode is triggered.

8. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 7, characterized in that: In step 3, the physical consistency check includes a phase check, which includes: b1: For the topological adjacency matrix Each pair of adjacent nodes , ( ), characteristic frequency The measured phase shift is defined as follows: ； The results were normalized to ; in: Indicates frequency The measured voltage phase shift from node i to node j; Represents the principal argument of a complex number; and These are the compensated net injected response phasors for nodes j and i, respectively. And based on online parameter identification provided and The reference phase shift is calculated using the following formula: ； in: The reference phase shift is obtained based on the impedance model; The series impedance of the line segment; The downstream equivalent impedance; The tolerance range is determined by the following formula: ； in: For phase tolerance; The baseline tolerance value is L; L is the actual length of the line segment. Unit length; For mutual inductor accuracy coefficient: the lower the accuracy, the larger the coefficient, and the wider the tolerance range; b2: Construct differential phase features for characteristic frequencies , Calculate the difference in measured phase shifts on the same adjacent node pair (i,j): ； in: For frequency and The difference in measured phase shift on edge (i,j) is called the differential phase characteristic; , This represents two different injected feature frequencies; For frequency The measured phase shift is shown below.

9. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 8, characterized in that: In step 4, the graph propagation decoding model uses a topological adjacency matrix. Define message passing paths; messages are only sent through the topological adjacency matrix. The transmission between adjacent nodes includes: 4.1: For FTU nodes After decomposing all the characteristic frequency responses of the corresponding column in matrix M, an initial eigenvector is constructed: ； in: and They are characteristic frequencies The net injected response amplitude and phase at node j; F represents the total number of characteristic frequencies; The dimension is 3F; 4.2: The graph propagation decoding model adopts a graph network structure based on a message-passing mechanism. The message passing at the k-th layer is performed according to the following formula. OK: ; in, Let i be the feature vector of node i after aggregation and nonlinear transformation at the (k+1)th layer; It is a non-linear activation function; This indicates that the node itself is also included in the aggregation scope; Let A be the set of neighbors of node i, treat the power distribution lines as undirected edges, and let A be a symmetric matrix. Messages can be transmitted bidirectionally. Let j be the feature representation of node j at the k-th layer; Let be the learnable weight matrix of the k-th layer; 4.3: The number of model layers is set to the maximum topology depth from the furthest grid-connected inverter to the bus, so that the [number of layers]... Layer message passing corresponding injection signal from Information is physically propagated to neighboring nodes; 4.4: For the branch connecting nodes i and j ( The edge embedding vector is constructed as follows: ； in: The high-level features of nodes i and j after passing through an L-layer graph network; It is a feature difference vector; For the splicing operation, After classification layer and Output ; The graph propagation decoding model adopts a multi-task learning architecture, including a branch classification main path and an auxiliary amplitude reconstruction path. The following total loss function is used for optimization during model training: ； in: For branch classification, the binary cross-entropy loss, For physical regularization loss term, These are the weighting coefficients; Physical regularization loss term Defined as: ； in: Let be the set of all upstream → downstream edges in the topology. And i is closer to the injection source; The frequencies of nodes i and j predicted by the auxiliary reconstruction head are respectively... The noiseless voltage amplitude; For the desired voltage attenuation ratio, The total number of characteristic frequencies; The loss forces the graph network to learn a decay pattern that conforms to circuit laws, which is a form of physical information regularization.

10. The active distribution network fault location method based on topology coding and graph propagation decoding according to claim 9, characterized in that: It also includes time synchronization fault tolerance and node data missing fault tolerance mechanisms; The time synchronization fault tolerance includes: after using BeiDou / GPS second pulses to achieve coarse synchronization of the sampling clocks of each FTU, the residual synchronization deviation is eliminated by using the global synchronization delay estimation method as described in step S3 data preprocessing; The node data missing fault tolerance includes: When individual FTU node n data is missing, all elements in the nth column of matrix M are forcibly set to zero: ； Make the initial feature vector corresponding to this node Degenerates into a zero vector; In the graph propagation decoding model, during message passing, although the initial feature of the missing node n is zero, its healthy neighbors... Still aggregate messages to n, making Integrate direct neighbor information; after L-layer iteration, By fusing the attenuation and phase shift characteristics of healthy nodes within the L-hop neighborhood, the state of the signal flowing through that node can be indirectly inferred; edge embedding This inference feature is used to calculate the probability of branch faults, so as to achieve "soft recovery" of the state of the missing area.

Images